LCOV - code coverage report
Current view: top level - lib/nvmf - rdma.c (source / functions) Hit Total Coverage
Test: ut_cov_unit.info Lines: 667 2395 27.8 %
Date: 2024-07-11 22:20:29 Functions: 30 119 25.2 %

          Line data    Source code
       1             : /*   SPDX-License-Identifier: BSD-3-Clause
       2             :  *   Copyright (C) 2016 Intel Corporation. All rights reserved.
       3             :  *   Copyright (c) 2019-2021 Mellanox Technologies LTD. All rights reserved.
       4             :  *   Copyright (c) 2021-2024 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
       5             :  */
       6             : 
       7             : #include "spdk/stdinc.h"
       8             : 
       9             : #include "spdk/config.h"
      10             : #include "spdk/thread.h"
      11             : #include "spdk/likely.h"
      12             : #include "spdk/nvmf_transport.h"
      13             : #include "spdk/string.h"
      14             : #include "spdk/trace.h"
      15             : #include "spdk/tree.h"
      16             : #include "spdk/util.h"
      17             : 
      18             : #include "spdk_internal/assert.h"
      19             : #include "spdk/log.h"
      20             : #include "spdk_internal/rdma.h"
      21             : 
      22             : #include "nvmf_internal.h"
      23             : #include "transport.h"
      24             : 
      25             : #include "spdk_internal/trace_defs.h"
      26             : 
      27             : struct spdk_nvme_rdma_hooks g_nvmf_hooks = {};
      28             : const struct spdk_nvmf_transport_ops spdk_nvmf_transport_rdma;
      29             : 
      30             : /*
      31             :  RDMA Connection Resource Defaults
      32             :  */
      33             : #define NVMF_DEFAULT_MSDBD              16
      34             : #define NVMF_DEFAULT_TX_SGE             SPDK_NVMF_MAX_SGL_ENTRIES
      35             : #define NVMF_DEFAULT_RSP_SGE            1
      36             : #define NVMF_DEFAULT_RX_SGE             2
      37             : 
      38             : #define NVMF_RDMA_MAX_EVENTS_PER_POLL   32
      39             : 
      40             : SPDK_STATIC_ASSERT(NVMF_DEFAULT_MSDBD <= SPDK_NVMF_MAX_SGL_ENTRIES,
      41             :                    "MSDBD must not exceed SPDK_NVMF_MAX_SGL_ENTRIES");
      42             : 
      43             : /* The RDMA completion queue size */
      44             : #define DEFAULT_NVMF_RDMA_CQ_SIZE       4096
      45             : #define MAX_WR_PER_QP(queue_depth)      (queue_depth * 3 + 2)
      46             : 
      47             : enum spdk_nvmf_rdma_request_state {
      48             :         /* The request is not currently in use */
      49             :         RDMA_REQUEST_STATE_FREE = 0,
      50             : 
      51             :         /* Initial state when request first received */
      52             :         RDMA_REQUEST_STATE_NEW,
      53             : 
      54             :         /* The request is queued until a data buffer is available. */
      55             :         RDMA_REQUEST_STATE_NEED_BUFFER,
      56             : 
      57             :         /* The request is waiting on RDMA queue depth availability
      58             :          * to transfer data from the host to the controller.
      59             :          */
      60             :         RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING,
      61             : 
      62             :         /* The request is currently transferring data from the host to the controller. */
      63             :         RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER,
      64             : 
      65             :         /* The request is ready to execute at the block device */
      66             :         RDMA_REQUEST_STATE_READY_TO_EXECUTE,
      67             : 
      68             :         /* The request is currently executing at the block device */
      69             :         RDMA_REQUEST_STATE_EXECUTING,
      70             : 
      71             :         /* The request finished executing at the block device */
      72             :         RDMA_REQUEST_STATE_EXECUTED,
      73             : 
      74             :         /* The request is waiting on RDMA queue depth availability
      75             :          * to transfer data from the controller to the host.
      76             :          */
      77             :         RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING,
      78             : 
      79             :         /* The request is waiting on RDMA queue depth availability
      80             :          * to send response to the host.
      81             :          */
      82             :         RDMA_REQUEST_STATE_READY_TO_COMPLETE_PENDING,
      83             : 
      84             :         /* The request is ready to send a completion */
      85             :         RDMA_REQUEST_STATE_READY_TO_COMPLETE,
      86             : 
      87             :         /* The request is currently transferring data from the controller to the host. */
      88             :         RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST,
      89             : 
      90             :         /* The request currently has an outstanding completion without an
      91             :          * associated data transfer.
      92             :          */
      93             :         RDMA_REQUEST_STATE_COMPLETING,
      94             : 
      95             :         /* The request completed and can be marked free. */
      96             :         RDMA_REQUEST_STATE_COMPLETED,
      97             : 
      98             :         /* Terminator */
      99             :         RDMA_REQUEST_NUM_STATES,
     100             : };
     101             : 
     102           2 : SPDK_TRACE_REGISTER_FN(nvmf_trace, "nvmf_rdma", TRACE_GROUP_NVMF_RDMA)
     103             : {
     104           0 :         spdk_trace_register_object(OBJECT_NVMF_RDMA_IO, 'r');
     105             : 
     106           0 :         struct spdk_trace_tpoint_opts opts[] = {
     107             :                 {
     108             :                         "RDMA_REQ_NEW", TRACE_RDMA_REQUEST_STATE_NEW,
     109             :                         OWNER_TYPE_NONE, OBJECT_NVMF_RDMA_IO, 1,
     110             :                         {
     111             :                                 { "qpair", SPDK_TRACE_ARG_TYPE_PTR, 8 },
     112             :                                 { "qd", SPDK_TRACE_ARG_TYPE_INT, 4 }
     113             :                         }
     114             :                 },
     115             :                 {
     116             :                         "RDMA_REQ_COMPLETED", TRACE_RDMA_REQUEST_STATE_COMPLETED,
     117             :                         OWNER_TYPE_NONE, OBJECT_NVMF_RDMA_IO, 0,
     118             :                         {
     119             :                                 { "qpair", SPDK_TRACE_ARG_TYPE_PTR, 8 },
     120             :                                 { "qd", SPDK_TRACE_ARG_TYPE_INT, 4 }
     121             :                         }
     122             :                 },
     123             :         };
     124             : 
     125           0 :         spdk_trace_register_description_ext(opts, SPDK_COUNTOF(opts));
     126           0 :         spdk_trace_register_description("RDMA_REQ_NEED_BUFFER", TRACE_RDMA_REQUEST_STATE_NEED_BUFFER,
     127             :                                         OWNER_TYPE_NONE, OBJECT_NVMF_RDMA_IO, 0,
     128             :                                         SPDK_TRACE_ARG_TYPE_PTR, "qpair");
     129           0 :         spdk_trace_register_description("RDMA_REQ_TX_PENDING_C2H",
     130             :                                         TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING,
     131             :                                         OWNER_TYPE_NONE, OBJECT_NVMF_RDMA_IO, 0,
     132             :                                         SPDK_TRACE_ARG_TYPE_PTR, "qpair");
     133           0 :         spdk_trace_register_description("RDMA_REQ_TX_PENDING_H2C",
     134             :                                         TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING,
     135             :                                         OWNER_TYPE_NONE, OBJECT_NVMF_RDMA_IO, 0,
     136             :                                         SPDK_TRACE_ARG_TYPE_PTR, "qpair");
     137           0 :         spdk_trace_register_description("RDMA_REQ_TX_H2C",
     138             :                                         TRACE_RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER,
     139             :                                         OWNER_TYPE_NONE, OBJECT_NVMF_RDMA_IO, 0,
     140             :                                         SPDK_TRACE_ARG_TYPE_PTR, "qpair");
     141           0 :         spdk_trace_register_description("RDMA_REQ_RDY_TO_EXECUTE",
     142             :                                         TRACE_RDMA_REQUEST_STATE_READY_TO_EXECUTE,
     143             :                                         OWNER_TYPE_NONE, OBJECT_NVMF_RDMA_IO, 0,
     144             :                                         SPDK_TRACE_ARG_TYPE_PTR, "qpair");
     145           0 :         spdk_trace_register_description("RDMA_REQ_EXECUTING",
     146             :                                         TRACE_RDMA_REQUEST_STATE_EXECUTING,
     147             :                                         OWNER_TYPE_NONE, OBJECT_NVMF_RDMA_IO, 0,
     148             :                                         SPDK_TRACE_ARG_TYPE_PTR, "qpair");
     149           0 :         spdk_trace_register_description("RDMA_REQ_EXECUTED",
     150             :                                         TRACE_RDMA_REQUEST_STATE_EXECUTED,
     151             :                                         OWNER_TYPE_NONE, OBJECT_NVMF_RDMA_IO, 0,
     152             :                                         SPDK_TRACE_ARG_TYPE_PTR, "qpair");
     153           0 :         spdk_trace_register_description("RDMA_REQ_RDY2COMPL_PEND",
     154             :                                         TRACE_RDMA_REQUEST_STATE_READY_TO_COMPLETE_PENDING,
     155             :                                         OWNER_TYPE_NONE, OBJECT_NVMF_RDMA_IO, 0,
     156             :                                         SPDK_TRACE_ARG_TYPE_PTR, "qpair");
     157           0 :         spdk_trace_register_description("RDMA_REQ_RDY_TO_COMPL",
     158             :                                         TRACE_RDMA_REQUEST_STATE_READY_TO_COMPLETE,
     159             :                                         OWNER_TYPE_NONE, OBJECT_NVMF_RDMA_IO, 0,
     160             :                                         SPDK_TRACE_ARG_TYPE_PTR, "qpair");
     161           0 :         spdk_trace_register_description("RDMA_REQ_COMPLETING_C2H",
     162             :                                         TRACE_RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST,
     163             :                                         OWNER_TYPE_NONE, OBJECT_NVMF_RDMA_IO, 0,
     164             :                                         SPDK_TRACE_ARG_TYPE_PTR, "qpair");
     165           0 :         spdk_trace_register_description("RDMA_REQ_COMPLETING",
     166             :                                         TRACE_RDMA_REQUEST_STATE_COMPLETING,
     167             :                                         OWNER_TYPE_NONE, OBJECT_NVMF_RDMA_IO, 0,
     168             :                                         SPDK_TRACE_ARG_TYPE_PTR, "qpair");
     169             : 
     170           0 :         spdk_trace_register_description("RDMA_QP_CREATE", TRACE_RDMA_QP_CREATE,
     171             :                                         OWNER_TYPE_NONE, OBJECT_NONE, 0,
     172             :                                         SPDK_TRACE_ARG_TYPE_INT, "");
     173           0 :         spdk_trace_register_description("RDMA_IBV_ASYNC_EVENT", TRACE_RDMA_IBV_ASYNC_EVENT,
     174             :                                         OWNER_TYPE_NONE, OBJECT_NONE, 0,
     175             :                                         SPDK_TRACE_ARG_TYPE_INT, "type");
     176           0 :         spdk_trace_register_description("RDMA_CM_ASYNC_EVENT", TRACE_RDMA_CM_ASYNC_EVENT,
     177             :                                         OWNER_TYPE_NONE, OBJECT_NONE, 0,
     178             :                                         SPDK_TRACE_ARG_TYPE_INT, "type");
     179           0 :         spdk_trace_register_description("RDMA_QP_DISCONNECT", TRACE_RDMA_QP_DISCONNECT,
     180             :                                         OWNER_TYPE_NONE, OBJECT_NONE, 0,
     181             :                                         SPDK_TRACE_ARG_TYPE_INT, "");
     182           0 :         spdk_trace_register_description("RDMA_QP_DESTROY", TRACE_RDMA_QP_DESTROY,
     183             :                                         OWNER_TYPE_NONE, OBJECT_NONE, 0,
     184             :                                         SPDK_TRACE_ARG_TYPE_INT, "");
     185           0 : }
     186             : 
     187             : enum spdk_nvmf_rdma_wr_type {
     188             :         RDMA_WR_TYPE_RECV,
     189             :         RDMA_WR_TYPE_SEND,
     190             :         RDMA_WR_TYPE_DATA,
     191             : };
     192             : 
     193             : struct spdk_nvmf_rdma_wr {
     194             :         /* Uses enum spdk_nvmf_rdma_wr_type */
     195             :         uint8_t type;
     196             : };
     197             : 
     198             : /* This structure holds commands as they are received off the wire.
     199             :  * It must be dynamically paired with a full request object
     200             :  * (spdk_nvmf_rdma_request) to service a request. It is separate
     201             :  * from the request because RDMA does not appear to order
     202             :  * completions, so occasionally we'll get a new incoming
     203             :  * command when there aren't any free request objects.
     204             :  */
     205             : struct spdk_nvmf_rdma_recv {
     206             :         struct ibv_recv_wr                      wr;
     207             :         struct ibv_sge                          sgl[NVMF_DEFAULT_RX_SGE];
     208             : 
     209             :         struct spdk_nvmf_rdma_qpair             *qpair;
     210             : 
     211             :         /* In-capsule data buffer */
     212             :         uint8_t                                 *buf;
     213             : 
     214             :         struct spdk_nvmf_rdma_wr                rdma_wr;
     215             :         uint64_t                                receive_tsc;
     216             : 
     217             :         STAILQ_ENTRY(spdk_nvmf_rdma_recv)       link;
     218             : };
     219             : 
     220             : struct spdk_nvmf_rdma_request_data {
     221             :         struct ibv_send_wr              wr;
     222             :         struct ibv_sge                  sgl[SPDK_NVMF_MAX_SGL_ENTRIES];
     223             : };
     224             : 
     225             : struct spdk_nvmf_rdma_request {
     226             :         struct spdk_nvmf_request                req;
     227             : 
     228             :         bool                                    fused_failed;
     229             : 
     230             :         struct spdk_nvmf_rdma_wr                data_wr;
     231             :         struct spdk_nvmf_rdma_wr                rsp_wr;
     232             : 
     233             :         /* Uses enum spdk_nvmf_rdma_request_state */
     234             :         uint8_t                                 state;
     235             : 
     236             :         /* Data offset in req.iov */
     237             :         uint32_t                                offset;
     238             : 
     239             :         struct spdk_nvmf_rdma_recv              *recv;
     240             : 
     241             :         struct {
     242             :                 struct  ibv_send_wr             wr;
     243             :                 struct  ibv_sge                 sgl[NVMF_DEFAULT_RSP_SGE];
     244             :         } rsp;
     245             : 
     246             :         uint16_t                                iovpos;
     247             :         uint16_t                                num_outstanding_data_wr;
     248             :         /* Used to split Write IO with multi SGL payload */
     249             :         uint16_t                                num_remaining_data_wr;
     250             :         uint64_t                                receive_tsc;
     251             :         struct spdk_nvmf_rdma_request           *fused_pair;
     252             :         STAILQ_ENTRY(spdk_nvmf_rdma_request)    state_link;
     253             :         struct ibv_send_wr                      *remaining_tranfer_in_wrs;
     254             :         struct ibv_send_wr                      *transfer_wr;
     255             :         struct spdk_nvmf_rdma_request_data      data;
     256             : };
     257             : 
     258             : struct spdk_nvmf_rdma_resource_opts {
     259             :         struct spdk_nvmf_rdma_qpair     *qpair;
     260             :         /* qp points either to an ibv_qp object or an ibv_srq object depending on the value of shared. */
     261             :         void                            *qp;
     262             :         struct spdk_rdma_mem_map        *map;
     263             :         uint32_t                        max_queue_depth;
     264             :         uint32_t                        in_capsule_data_size;
     265             :         bool                            shared;
     266             : };
     267             : 
     268             : struct spdk_nvmf_rdma_resources {
     269             :         /* Array of size "max_queue_depth" containing RDMA requests. */
     270             :         struct spdk_nvmf_rdma_request           *reqs;
     271             : 
     272             :         /* Array of size "max_queue_depth" containing RDMA recvs. */
     273             :         struct spdk_nvmf_rdma_recv              *recvs;
     274             : 
     275             :         /* Array of size "max_queue_depth" containing 64 byte capsules
     276             :          * used for receive.
     277             :          */
     278             :         union nvmf_h2c_msg                      *cmds;
     279             : 
     280             :         /* Array of size "max_queue_depth" containing 16 byte completions
     281             :          * to be sent back to the user.
     282             :          */
     283             :         union nvmf_c2h_msg                      *cpls;
     284             : 
     285             :         /* Array of size "max_queue_depth * InCapsuleDataSize" containing
     286             :          * buffers to be used for in capsule data.
     287             :          */
     288             :         void                                    *bufs;
     289             : 
     290             :         /* Receives that are waiting for a request object */
     291             :         STAILQ_HEAD(, spdk_nvmf_rdma_recv)      incoming_queue;
     292             : 
     293             :         /* Queue to track free requests */
     294             :         STAILQ_HEAD(, spdk_nvmf_rdma_request)   free_queue;
     295             : };
     296             : 
     297             : typedef void (*spdk_nvmf_rdma_qpair_ibv_event)(struct spdk_nvmf_rdma_qpair *rqpair);
     298             : 
     299             : typedef void (*spdk_poller_destroy_cb)(void *ctx);
     300             : 
     301             : struct spdk_nvmf_rdma_ibv_event_ctx {
     302             :         struct spdk_nvmf_rdma_qpair                     *rqpair;
     303             :         spdk_nvmf_rdma_qpair_ibv_event                  cb_fn;
     304             :         /* Link to other ibv events associated with this qpair */
     305             :         STAILQ_ENTRY(spdk_nvmf_rdma_ibv_event_ctx)      link;
     306             : };
     307             : 
     308             : struct spdk_nvmf_rdma_qpair {
     309             :         struct spdk_nvmf_qpair                  qpair;
     310             : 
     311             :         struct spdk_nvmf_rdma_device            *device;
     312             :         struct spdk_nvmf_rdma_poller            *poller;
     313             : 
     314             :         struct spdk_rdma_qp                     *rdma_qp;
     315             :         struct rdma_cm_id                       *cm_id;
     316             :         struct spdk_rdma_srq                    *srq;
     317             :         struct rdma_cm_id                       *listen_id;
     318             : 
     319             :         /* Cache the QP number to improve QP search by RB tree. */
     320             :         uint32_t                                qp_num;
     321             : 
     322             :         /* The maximum number of I/O outstanding on this connection at one time */
     323             :         uint16_t                                max_queue_depth;
     324             : 
     325             :         /* The maximum number of active RDMA READ and ATOMIC operations at one time */
     326             :         uint16_t                                max_read_depth;
     327             : 
     328             :         /* The maximum number of RDMA SEND operations at one time */
     329             :         uint32_t                                max_send_depth;
     330             : 
     331             :         /* The current number of outstanding WRs from this qpair's
     332             :          * recv queue. Should not exceed device->attr.max_queue_depth.
     333             :          */
     334             :         uint16_t                                current_recv_depth;
     335             : 
     336             :         /* The current number of active RDMA READ operations */
     337             :         uint16_t                                current_read_depth;
     338             : 
     339             :         /* The current number of posted WRs from this qpair's
     340             :          * send queue. Should not exceed max_send_depth.
     341             :          */
     342             :         uint32_t                                current_send_depth;
     343             : 
     344             :         /* The maximum number of SGEs per WR on the send queue */
     345             :         uint32_t                                max_send_sge;
     346             : 
     347             :         /* The maximum number of SGEs per WR on the recv queue */
     348             :         uint32_t                                max_recv_sge;
     349             : 
     350             :         struct spdk_nvmf_rdma_resources         *resources;
     351             : 
     352             :         STAILQ_HEAD(, spdk_nvmf_rdma_request)   pending_rdma_read_queue;
     353             : 
     354             :         STAILQ_HEAD(, spdk_nvmf_rdma_request)   pending_rdma_write_queue;
     355             : 
     356             :         STAILQ_HEAD(, spdk_nvmf_rdma_request)   pending_rdma_send_queue;
     357             : 
     358             :         /* Number of requests not in the free state */
     359             :         uint32_t                                qd;
     360             : 
     361             :         bool                                    ibv_in_error_state;
     362             : 
     363             :         RB_ENTRY(spdk_nvmf_rdma_qpair)          node;
     364             : 
     365             :         STAILQ_ENTRY(spdk_nvmf_rdma_qpair)      recv_link;
     366             : 
     367             :         STAILQ_ENTRY(spdk_nvmf_rdma_qpair)      send_link;
     368             : 
     369             :         /* Points to the a request that has fuse bits set to
     370             :          * SPDK_NVME_CMD_FUSE_FIRST, when the qpair is waiting
     371             :          * for the request that has SPDK_NVME_CMD_FUSE_SECOND.
     372             :          */
     373             :         struct spdk_nvmf_rdma_request           *fused_first;
     374             : 
     375             :         /*
     376             :          * io_channel which is used to destroy qpair when it is removed from poll group
     377             :          */
     378             :         struct spdk_io_channel          *destruct_channel;
     379             : 
     380             :         /* List of ibv async events */
     381             :         STAILQ_HEAD(, spdk_nvmf_rdma_ibv_event_ctx)     ibv_events;
     382             : 
     383             :         /* Lets us know that we have received the last_wqe event. */
     384             :         bool                                    last_wqe_reached;
     385             : 
     386             :         /* Indicate that nvmf_rdma_close_qpair is called */
     387             :         bool                                    to_close;
     388             : };
     389             : 
     390             : struct spdk_nvmf_rdma_poller_stat {
     391             :         uint64_t                                completions;
     392             :         uint64_t                                polls;
     393             :         uint64_t                                idle_polls;
     394             :         uint64_t                                requests;
     395             :         uint64_t                                request_latency;
     396             :         uint64_t                                pending_free_request;
     397             :         uint64_t                                pending_rdma_read;
     398             :         uint64_t                                pending_rdma_write;
     399             :         uint64_t                                pending_rdma_send;
     400             :         struct spdk_rdma_qp_stats               qp_stats;
     401             : };
     402             : 
     403             : struct spdk_nvmf_rdma_poller {
     404             :         struct spdk_nvmf_rdma_device            *device;
     405             :         struct spdk_nvmf_rdma_poll_group        *group;
     406             : 
     407             :         int                                     num_cqe;
     408             :         int                                     required_num_wr;
     409             :         struct ibv_cq                           *cq;
     410             : 
     411             :         /* The maximum number of I/O outstanding on the shared receive queue at one time */
     412             :         uint16_t                                max_srq_depth;
     413             :         bool                                    need_destroy;
     414             : 
     415             :         /* Shared receive queue */
     416             :         struct spdk_rdma_srq                    *srq;
     417             : 
     418             :         struct spdk_nvmf_rdma_resources         *resources;
     419             :         struct spdk_nvmf_rdma_poller_stat       stat;
     420             : 
     421             :         spdk_poller_destroy_cb                  destroy_cb;
     422             :         void                                    *destroy_cb_ctx;
     423             : 
     424             :         RB_HEAD(qpairs_tree, spdk_nvmf_rdma_qpair) qpairs;
     425             : 
     426             :         STAILQ_HEAD(, spdk_nvmf_rdma_qpair)     qpairs_pending_recv;
     427             : 
     428             :         STAILQ_HEAD(, spdk_nvmf_rdma_qpair)     qpairs_pending_send;
     429             : 
     430             :         TAILQ_ENTRY(spdk_nvmf_rdma_poller)      link;
     431             : };
     432             : 
     433             : struct spdk_nvmf_rdma_poll_group_stat {
     434             :         uint64_t                                pending_data_buffer;
     435             : };
     436             : 
     437             : struct spdk_nvmf_rdma_poll_group {
     438             :         struct spdk_nvmf_transport_poll_group           group;
     439             :         struct spdk_nvmf_rdma_poll_group_stat           stat;
     440             :         TAILQ_HEAD(, spdk_nvmf_rdma_poller)             pollers;
     441             :         TAILQ_ENTRY(spdk_nvmf_rdma_poll_group)          link;
     442             : };
     443             : 
     444             : struct spdk_nvmf_rdma_conn_sched {
     445             :         struct spdk_nvmf_rdma_poll_group *next_admin_pg;
     446             :         struct spdk_nvmf_rdma_poll_group *next_io_pg;
     447             : };
     448             : 
     449             : /* Assuming rdma_cm uses just one protection domain per ibv_context. */
     450             : struct spdk_nvmf_rdma_device {
     451             :         struct ibv_device_attr                  attr;
     452             :         struct ibv_context                      *context;
     453             : 
     454             :         struct spdk_rdma_mem_map                *map;
     455             :         struct ibv_pd                           *pd;
     456             : 
     457             :         int                                     num_srq;
     458             :         bool                                    need_destroy;
     459             :         bool                                    ready_to_destroy;
     460             :         bool                                    is_ready;
     461             : 
     462             :         TAILQ_ENTRY(spdk_nvmf_rdma_device)      link;
     463             : };
     464             : 
     465             : struct spdk_nvmf_rdma_port {
     466             :         const struct spdk_nvme_transport_id     *trid;
     467             :         struct rdma_cm_id                       *id;
     468             :         struct spdk_nvmf_rdma_device            *device;
     469             :         TAILQ_ENTRY(spdk_nvmf_rdma_port)        link;
     470             : };
     471             : 
     472             : struct rdma_transport_opts {
     473             :         int             num_cqe;
     474             :         uint32_t        max_srq_depth;
     475             :         bool            no_srq;
     476             :         bool            no_wr_batching;
     477             :         int             acceptor_backlog;
     478             : };
     479             : 
     480             : struct spdk_nvmf_rdma_transport {
     481             :         struct spdk_nvmf_transport      transport;
     482             :         struct rdma_transport_opts      rdma_opts;
     483             : 
     484             :         struct spdk_nvmf_rdma_conn_sched conn_sched;
     485             : 
     486             :         struct rdma_event_channel       *event_channel;
     487             : 
     488             :         struct spdk_mempool             *data_wr_pool;
     489             : 
     490             :         struct spdk_poller              *accept_poller;
     491             : 
     492             :         /* fields used to poll RDMA/IB events */
     493             :         nfds_t                  npoll_fds;
     494             :         struct pollfd           *poll_fds;
     495             : 
     496             :         TAILQ_HEAD(, spdk_nvmf_rdma_device)     devices;
     497             :         TAILQ_HEAD(, spdk_nvmf_rdma_port)       ports;
     498             :         TAILQ_HEAD(, spdk_nvmf_rdma_poll_group) poll_groups;
     499             : 
     500             :         /* ports that are removed unexpectedly and need retry listen */
     501             :         TAILQ_HEAD(, spdk_nvmf_rdma_port)               retry_ports;
     502             : };
     503             : 
     504             : struct poller_manage_ctx {
     505             :         struct spdk_nvmf_rdma_transport         *rtransport;
     506             :         struct spdk_nvmf_rdma_poll_group        *rgroup;
     507             :         struct spdk_nvmf_rdma_poller            *rpoller;
     508             :         struct spdk_nvmf_rdma_device            *device;
     509             : 
     510             :         struct spdk_thread                      *thread;
     511             :         volatile int                            *inflight_op_counter;
     512             : };
     513             : 
     514             : static const struct spdk_json_object_decoder rdma_transport_opts_decoder[] = {
     515             :         {
     516             :                 "num_cqe", offsetof(struct rdma_transport_opts, num_cqe),
     517             :                 spdk_json_decode_int32, true
     518             :         },
     519             :         {
     520             :                 "max_srq_depth", offsetof(struct rdma_transport_opts, max_srq_depth),
     521             :                 spdk_json_decode_uint32, true
     522             :         },
     523             :         {
     524             :                 "no_srq", offsetof(struct rdma_transport_opts, no_srq),
     525             :                 spdk_json_decode_bool, true
     526             :         },
     527             :         {
     528             :                 "no_wr_batching", offsetof(struct rdma_transport_opts, no_wr_batching),
     529             :                 spdk_json_decode_bool, true
     530             :         },
     531             :         {
     532             :                 "acceptor_backlog", offsetof(struct rdma_transport_opts, acceptor_backlog),
     533             :                 spdk_json_decode_int32, true
     534             :         },
     535             : };
     536             : 
     537             : static int
     538           2 : nvmf_rdma_qpair_compare(struct spdk_nvmf_rdma_qpair *rqpair1, struct spdk_nvmf_rdma_qpair *rqpair2)
     539             : {
     540           2 :         return rqpair1->qp_num < rqpair2->qp_num ? -1 : rqpair1->qp_num > rqpair2->qp_num;
     541             : }
     542             : 
     543           0 : RB_GENERATE_STATIC(qpairs_tree, spdk_nvmf_rdma_qpair, node, nvmf_rdma_qpair_compare);
     544             : 
     545             : static bool nvmf_rdma_request_process(struct spdk_nvmf_rdma_transport *rtransport,
     546             :                                       struct spdk_nvmf_rdma_request *rdma_req);
     547             : 
     548             : static void _poller_submit_sends(struct spdk_nvmf_rdma_transport *rtransport,
     549             :                                  struct spdk_nvmf_rdma_poller *rpoller);
     550             : 
     551             : static void _poller_submit_recvs(struct spdk_nvmf_rdma_transport *rtransport,
     552             :                                  struct spdk_nvmf_rdma_poller *rpoller);
     553             : 
     554             : static void _nvmf_rdma_remove_destroyed_device(void *c);
     555             : 
     556             : static inline enum spdk_nvme_media_error_status_code
     557           0 : nvmf_rdma_dif_error_to_compl_status(uint8_t err_type) {
     558             :         enum spdk_nvme_media_error_status_code result;
     559           0 :         switch (err_type)
     560             :         {
     561           0 :         case SPDK_DIF_REFTAG_ERROR:
     562           0 :                 result = SPDK_NVME_SC_REFERENCE_TAG_CHECK_ERROR;
     563           0 :                 break;
     564           0 :         case SPDK_DIF_APPTAG_ERROR:
     565           0 :                 result = SPDK_NVME_SC_APPLICATION_TAG_CHECK_ERROR;
     566           0 :                 break;
     567           0 :         case SPDK_DIF_GUARD_ERROR:
     568           0 :                 result = SPDK_NVME_SC_GUARD_CHECK_ERROR;
     569           0 :                 break;
     570           0 :         default:
     571           0 :                 SPDK_UNREACHABLE();
     572             :         }
     573             : 
     574           0 :         return result;
     575             : }
     576             : 
     577             : /*
     578             :  * Return data_wrs to pool starting from \b data_wr
     579             :  * Request's own response and data WR are excluded
     580             :  */
     581             : static void
     582           7 : _nvmf_rdma_request_free_data(struct spdk_nvmf_rdma_request *rdma_req,
     583             :                              struct ibv_send_wr *data_wr,
     584             :                              struct spdk_mempool *pool)
     585             : {
     586           7 :         struct spdk_nvmf_rdma_request_data      *work_requests[SPDK_NVMF_MAX_SGL_ENTRIES];
     587             :         struct spdk_nvmf_rdma_request_data      *nvmf_data;
     588             :         struct ibv_send_wr                      *next_send_wr;
     589           7 :         uint64_t                                req_wrid = (uint64_t)&rdma_req->data_wr;
     590           7 :         uint32_t                                num_wrs = 0;
     591             : 
     592          15 :         while (data_wr && data_wr->wr_id == req_wrid) {
     593           8 :                 nvmf_data = SPDK_CONTAINEROF(data_wr, struct spdk_nvmf_rdma_request_data, wr);
     594           8 :                 memset(nvmf_data->sgl, 0, sizeof(data_wr->sg_list[0]) * data_wr->num_sge);
     595           8 :                 data_wr->num_sge = 0;
     596           8 :                 next_send_wr = data_wr->next;
     597           8 :                 if (data_wr != &rdma_req->data.wr) {
     598           1 :                         data_wr->next = NULL;
     599           1 :                         assert(num_wrs < SPDK_NVMF_MAX_SGL_ENTRIES);
     600           1 :                         work_requests[num_wrs] = nvmf_data;
     601           1 :                         num_wrs++;
     602             :                 }
     603           8 :                 data_wr = (!next_send_wr || next_send_wr == &rdma_req->rsp.wr) ? NULL : next_send_wr;
     604             :         }
     605             : 
     606           7 :         if (num_wrs) {
     607           1 :                 spdk_mempool_put_bulk(pool, (void **) work_requests, num_wrs);
     608             :         }
     609           7 : }
     610             : 
     611             : static void
     612           7 : nvmf_rdma_request_free_data(struct spdk_nvmf_rdma_request *rdma_req,
     613             :                             struct spdk_nvmf_rdma_transport *rtransport)
     614             : {
     615           7 :         rdma_req->num_outstanding_data_wr = 0;
     616             : 
     617           7 :         _nvmf_rdma_request_free_data(rdma_req, rdma_req->transfer_wr, rtransport->data_wr_pool);
     618             : 
     619           7 :         if (rdma_req->remaining_tranfer_in_wrs) {
     620           0 :                 _nvmf_rdma_request_free_data(rdma_req, rdma_req->remaining_tranfer_in_wrs,
     621             :                                              rtransport->data_wr_pool);
     622           0 :                 rdma_req->remaining_tranfer_in_wrs = NULL;
     623             :         }
     624             : 
     625           7 :         rdma_req->data.wr.next = NULL;
     626           7 :         rdma_req->rsp.wr.next = NULL;
     627           7 : }
     628             : 
     629             : static void
     630           0 : nvmf_rdma_dump_request(struct spdk_nvmf_rdma_request *req)
     631             : {
     632           0 :         SPDK_ERRLOG("\t\tRequest Data From Pool: %d\n", req->req.data_from_pool);
     633           0 :         if (req->req.cmd) {
     634           0 :                 SPDK_ERRLOG("\t\tRequest opcode: %d\n", req->req.cmd->nvmf_cmd.opcode);
     635             :         }
     636           0 :         if (req->recv) {
     637           0 :                 SPDK_ERRLOG("\t\tRequest recv wr_id%lu\n", req->recv->wr.wr_id);
     638             :         }
     639           0 : }
     640             : 
     641             : static void
     642           0 : nvmf_rdma_dump_qpair_contents(struct spdk_nvmf_rdma_qpair *rqpair)
     643             : {
     644             :         int i;
     645             : 
     646           0 :         SPDK_ERRLOG("Dumping contents of queue pair (QID %d)\n", rqpair->qpair.qid);
     647           0 :         for (i = 0; i < rqpair->max_queue_depth; i++) {
     648           0 :                 if (rqpair->resources->reqs[i].state != RDMA_REQUEST_STATE_FREE) {
     649           0 :                         nvmf_rdma_dump_request(&rqpair->resources->reqs[i]);
     650             :                 }
     651             :         }
     652           0 : }
     653             : 
     654             : static void
     655           1 : nvmf_rdma_resources_destroy(struct spdk_nvmf_rdma_resources *resources)
     656             : {
     657           1 :         spdk_free(resources->cmds);
     658           1 :         spdk_free(resources->cpls);
     659           1 :         spdk_free(resources->bufs);
     660           1 :         spdk_free(resources->reqs);
     661           1 :         spdk_free(resources->recvs);
     662           1 :         free(resources);
     663           1 : }
     664             : 
     665             : 
     666             : static struct spdk_nvmf_rdma_resources *
     667           1 : nvmf_rdma_resources_create(struct spdk_nvmf_rdma_resource_opts *opts)
     668             : {
     669             :         struct spdk_nvmf_rdma_resources         *resources;
     670             :         struct spdk_nvmf_rdma_request           *rdma_req;
     671             :         struct spdk_nvmf_rdma_recv              *rdma_recv;
     672           1 :         struct spdk_rdma_qp                     *qp = NULL;
     673           1 :         struct spdk_rdma_srq                    *srq = NULL;
     674           1 :         struct ibv_recv_wr                      *bad_wr = NULL;
     675           1 :         struct spdk_rdma_memory_translation     translation;
     676             :         uint32_t                                i;
     677           1 :         int                                     rc = 0;
     678             : 
     679           1 :         resources = calloc(1, sizeof(struct spdk_nvmf_rdma_resources));
     680           1 :         if (!resources) {
     681           0 :                 SPDK_ERRLOG("Unable to allocate resources for receive queue.\n");
     682           0 :                 return NULL;
     683             :         }
     684             : 
     685           1 :         resources->reqs = spdk_zmalloc(opts->max_queue_depth * sizeof(*resources->reqs),
     686             :                                        0x1000, NULL, SPDK_ENV_LCORE_ID_ANY, SPDK_MALLOC_DMA);
     687           1 :         resources->recvs = spdk_zmalloc(opts->max_queue_depth * sizeof(*resources->recvs),
     688             :                                         0x1000, NULL, SPDK_ENV_LCORE_ID_ANY, SPDK_MALLOC_DMA);
     689           1 :         resources->cmds = spdk_zmalloc(opts->max_queue_depth * sizeof(*resources->cmds),
     690             :                                        0x1000, NULL, SPDK_ENV_LCORE_ID_ANY, SPDK_MALLOC_DMA);
     691           1 :         resources->cpls = spdk_zmalloc(opts->max_queue_depth * sizeof(*resources->cpls),
     692             :                                        0x1000, NULL, SPDK_ENV_LCORE_ID_ANY, SPDK_MALLOC_DMA);
     693             : 
     694           1 :         if (opts->in_capsule_data_size > 0) {
     695           1 :                 resources->bufs = spdk_zmalloc(opts->max_queue_depth * opts->in_capsule_data_size,
     696             :                                                0x1000, NULL, SPDK_ENV_LCORE_ID_ANY,
     697             :                                                SPDK_MALLOC_DMA);
     698             :         }
     699             : 
     700           1 :         if (!resources->reqs || !resources->recvs || !resources->cmds ||
     701           1 :             !resources->cpls || (opts->in_capsule_data_size && !resources->bufs)) {
     702           0 :                 SPDK_ERRLOG("Unable to allocate sufficient memory for RDMA queue.\n");
     703           0 :                 goto cleanup;
     704             :         }
     705             : 
     706           1 :         SPDK_DEBUGLOG(rdma, "Command Array: %p Length: %lx\n",
     707             :                       resources->cmds, opts->max_queue_depth * sizeof(*resources->cmds));
     708           1 :         SPDK_DEBUGLOG(rdma, "Completion Array: %p Length: %lx\n",
     709             :                       resources->cpls, opts->max_queue_depth * sizeof(*resources->cpls));
     710           1 :         if (resources->bufs) {
     711           1 :                 SPDK_DEBUGLOG(rdma, "In Capsule Data Array: %p Length: %x\n",
     712             :                               resources->bufs, opts->max_queue_depth *
     713             :                               opts->in_capsule_data_size);
     714             :         }
     715             : 
     716             :         /* Initialize queues */
     717           1 :         STAILQ_INIT(&resources->incoming_queue);
     718           1 :         STAILQ_INIT(&resources->free_queue);
     719             : 
     720           1 :         if (opts->shared) {
     721           1 :                 srq = (struct spdk_rdma_srq *)opts->qp;
     722             :         } else {
     723           0 :                 qp = (struct spdk_rdma_qp *)opts->qp;
     724             :         }
     725             : 
     726         129 :         for (i = 0; i < opts->max_queue_depth; i++) {
     727         128 :                 rdma_recv = &resources->recvs[i];
     728         128 :                 rdma_recv->qpair = opts->qpair;
     729             : 
     730             :                 /* Set up memory to receive commands */
     731         128 :                 if (resources->bufs) {
     732         128 :                         rdma_recv->buf = (void *)((uintptr_t)resources->bufs + (i *
     733         128 :                                                   opts->in_capsule_data_size));
     734             :                 }
     735             : 
     736         128 :                 rdma_recv->rdma_wr.type = RDMA_WR_TYPE_RECV;
     737             : 
     738         128 :                 rdma_recv->sgl[0].addr = (uintptr_t)&resources->cmds[i];
     739         128 :                 rdma_recv->sgl[0].length = sizeof(resources->cmds[i]);
     740         128 :                 rc = spdk_rdma_get_translation(opts->map, &resources->cmds[i], sizeof(resources->cmds[i]),
     741             :                                                &translation);
     742         128 :                 if (rc) {
     743           0 :                         goto cleanup;
     744             :                 }
     745         128 :                 rdma_recv->sgl[0].lkey = spdk_rdma_memory_translation_get_lkey(&translation);
     746         128 :                 rdma_recv->wr.num_sge = 1;
     747             : 
     748         128 :                 if (rdma_recv->buf) {
     749         128 :                         rdma_recv->sgl[1].addr = (uintptr_t)rdma_recv->buf;
     750         128 :                         rdma_recv->sgl[1].length = opts->in_capsule_data_size;
     751         128 :                         rc = spdk_rdma_get_translation(opts->map, rdma_recv->buf, opts->in_capsule_data_size, &translation);
     752         128 :                         if (rc) {
     753           0 :                                 goto cleanup;
     754             :                         }
     755         128 :                         rdma_recv->sgl[1].lkey = spdk_rdma_memory_translation_get_lkey(&translation);
     756         128 :                         rdma_recv->wr.num_sge++;
     757             :                 }
     758             : 
     759         128 :                 rdma_recv->wr.wr_id = (uintptr_t)&rdma_recv->rdma_wr;
     760         128 :                 rdma_recv->wr.sg_list = rdma_recv->sgl;
     761         128 :                 if (srq) {
     762           0 :                         spdk_rdma_srq_queue_recv_wrs(srq, &rdma_recv->wr);
     763             :                 } else {
     764         128 :                         spdk_rdma_qp_queue_recv_wrs(qp, &rdma_recv->wr);
     765             :                 }
     766             :         }
     767             : 
     768         129 :         for (i = 0; i < opts->max_queue_depth; i++) {
     769         128 :                 rdma_req = &resources->reqs[i];
     770             : 
     771         128 :                 if (opts->qpair != NULL) {
     772         128 :                         rdma_req->req.qpair = &opts->qpair->qpair;
     773             :                 } else {
     774           0 :                         rdma_req->req.qpair = NULL;
     775             :                 }
     776         128 :                 rdma_req->req.cmd = NULL;
     777         128 :                 rdma_req->req.iovcnt = 0;
     778         128 :                 rdma_req->req.stripped_data = NULL;
     779             : 
     780             :                 /* Set up memory to send responses */
     781         128 :                 rdma_req->req.rsp = &resources->cpls[i];
     782             : 
     783         128 :                 rdma_req->rsp.sgl[0].addr = (uintptr_t)&resources->cpls[i];
     784         128 :                 rdma_req->rsp.sgl[0].length = sizeof(resources->cpls[i]);
     785         128 :                 rc = spdk_rdma_get_translation(opts->map, &resources->cpls[i], sizeof(resources->cpls[i]),
     786             :                                                &translation);
     787         128 :                 if (rc) {
     788           0 :                         goto cleanup;
     789             :                 }
     790         128 :                 rdma_req->rsp.sgl[0].lkey = spdk_rdma_memory_translation_get_lkey(&translation);
     791             : 
     792         128 :                 rdma_req->rsp_wr.type = RDMA_WR_TYPE_SEND;
     793         128 :                 rdma_req->rsp.wr.wr_id = (uintptr_t)&rdma_req->rsp_wr;
     794         128 :                 rdma_req->rsp.wr.next = NULL;
     795         128 :                 rdma_req->rsp.wr.opcode = IBV_WR_SEND;
     796         128 :                 rdma_req->rsp.wr.send_flags = IBV_SEND_SIGNALED;
     797         128 :                 rdma_req->rsp.wr.sg_list = rdma_req->rsp.sgl;
     798         128 :                 rdma_req->rsp.wr.num_sge = SPDK_COUNTOF(rdma_req->rsp.sgl);
     799             : 
     800             :                 /* Set up memory for data buffers */
     801         128 :                 rdma_req->data_wr.type = RDMA_WR_TYPE_DATA;
     802         128 :                 rdma_req->data.wr.wr_id = (uintptr_t)&rdma_req->data_wr;
     803         128 :                 rdma_req->data.wr.next = NULL;
     804         128 :                 rdma_req->data.wr.send_flags = IBV_SEND_SIGNALED;
     805         128 :                 rdma_req->data.wr.sg_list = rdma_req->data.sgl;
     806         128 :                 rdma_req->data.wr.num_sge = SPDK_COUNTOF(rdma_req->data.sgl);
     807             : 
     808             :                 /* Initialize request state to FREE */
     809         128 :                 rdma_req->state = RDMA_REQUEST_STATE_FREE;
     810         128 :                 STAILQ_INSERT_TAIL(&resources->free_queue, rdma_req, state_link);
     811             :         }
     812             : 
     813           1 :         if (srq) {
     814           0 :                 rc = spdk_rdma_srq_flush_recv_wrs(srq, &bad_wr);
     815             :         } else {
     816           1 :                 rc = spdk_rdma_qp_flush_recv_wrs(qp, &bad_wr);
     817             :         }
     818             : 
     819           1 :         if (rc) {
     820           0 :                 goto cleanup;
     821             :         }
     822             : 
     823           1 :         return resources;
     824             : 
     825           0 : cleanup:
     826           0 :         nvmf_rdma_resources_destroy(resources);
     827           0 :         return NULL;
     828             : }
     829             : 
     830             : static void
     831           0 : nvmf_rdma_qpair_clean_ibv_events(struct spdk_nvmf_rdma_qpair *rqpair)
     832             : {
     833             :         struct spdk_nvmf_rdma_ibv_event_ctx *ctx, *tctx;
     834           0 :         STAILQ_FOREACH_SAFE(ctx, &rqpair->ibv_events, link, tctx) {
     835           0 :                 ctx->rqpair = NULL;
     836             :                 /* Memory allocated for ctx is freed in nvmf_rdma_qpair_process_ibv_event */
     837           0 :                 STAILQ_REMOVE(&rqpair->ibv_events, ctx, spdk_nvmf_rdma_ibv_event_ctx, link);
     838             :         }
     839           0 : }
     840             : 
     841             : static void nvmf_rdma_poller_destroy(struct spdk_nvmf_rdma_poller *poller);
     842             : 
     843             : static void
     844           0 : nvmf_rdma_qpair_destroy(struct spdk_nvmf_rdma_qpair *rqpair)
     845             : {
     846             :         struct spdk_nvmf_rdma_recv      *rdma_recv, *recv_tmp;
     847           0 :         struct ibv_recv_wr              *bad_recv_wr = NULL;
     848             :         int                             rc;
     849             : 
     850           0 :         spdk_trace_record(TRACE_RDMA_QP_DESTROY, 0, 0, (uintptr_t)rqpair);
     851             : 
     852           0 :         if (rqpair->qd != 0) {
     853           0 :                 struct spdk_nvmf_qpair *qpair = &rqpair->qpair;
     854           0 :                 struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(qpair->transport,
     855             :                                 struct spdk_nvmf_rdma_transport, transport);
     856             :                 struct spdk_nvmf_rdma_request *req;
     857           0 :                 uint32_t i, max_req_count = 0;
     858             : 
     859           0 :                 SPDK_WARNLOG("Destroying qpair when queue depth is %d\n", rqpair->qd);
     860             : 
     861           0 :                 if (rqpair->srq == NULL) {
     862           0 :                         nvmf_rdma_dump_qpair_contents(rqpair);
     863           0 :                         max_req_count = rqpair->max_queue_depth;
     864           0 :                 } else if (rqpair->poller && rqpair->resources) {
     865           0 :                         max_req_count = rqpair->poller->max_srq_depth;
     866             :                 }
     867             : 
     868           0 :                 SPDK_DEBUGLOG(rdma, "Release incomplete requests\n");
     869           0 :                 for (i = 0; i < max_req_count; i++) {
     870           0 :                         req = &rqpair->resources->reqs[i];
     871           0 :                         if (req->req.qpair == qpair && req->state != RDMA_REQUEST_STATE_FREE) {
     872             :                                 /* nvmf_rdma_request_process checks qpair ibv and internal state
     873             :                                  * and completes a request */
     874           0 :                                 nvmf_rdma_request_process(rtransport, req);
     875             :                         }
     876             :                 }
     877           0 :                 assert(rqpair->qd == 0);
     878             :         }
     879             : 
     880           0 :         if (rqpair->poller) {
     881           0 :                 RB_REMOVE(qpairs_tree, &rqpair->poller->qpairs, rqpair);
     882             : 
     883           0 :                 if (rqpair->srq != NULL && rqpair->resources != NULL) {
     884             :                         /* Drop all received but unprocessed commands for this queue and return them to SRQ */
     885           0 :                         STAILQ_FOREACH_SAFE(rdma_recv, &rqpair->resources->incoming_queue, link, recv_tmp) {
     886           0 :                                 if (rqpair == rdma_recv->qpair) {
     887           0 :                                         STAILQ_REMOVE(&rqpair->resources->incoming_queue, rdma_recv, spdk_nvmf_rdma_recv, link);
     888           0 :                                         spdk_rdma_srq_queue_recv_wrs(rqpair->srq, &rdma_recv->wr);
     889           0 :                                         rc = spdk_rdma_srq_flush_recv_wrs(rqpair->srq, &bad_recv_wr);
     890           0 :                                         if (rc) {
     891           0 :                                                 SPDK_ERRLOG("Unable to re-post rx descriptor\n");
     892             :                                         }
     893             :                                 }
     894             :                         }
     895             :                 }
     896             :         }
     897             : 
     898           0 :         if (rqpair->cm_id) {
     899           0 :                 if (rqpair->rdma_qp != NULL) {
     900           0 :                         spdk_rdma_qp_destroy(rqpair->rdma_qp);
     901           0 :                         rqpair->rdma_qp = NULL;
     902             :                 }
     903             : 
     904           0 :                 if (rqpair->poller != NULL && rqpair->srq == NULL) {
     905           0 :                         rqpair->poller->required_num_wr -= MAX_WR_PER_QP(rqpair->max_queue_depth);
     906             :                 }
     907             :         }
     908             : 
     909           0 :         if (rqpair->srq == NULL && rqpair->resources != NULL) {
     910           0 :                 nvmf_rdma_resources_destroy(rqpair->resources);
     911             :         }
     912             : 
     913           0 :         nvmf_rdma_qpair_clean_ibv_events(rqpair);
     914             : 
     915           0 :         if (rqpair->destruct_channel) {
     916           0 :                 spdk_put_io_channel(rqpair->destruct_channel);
     917           0 :                 rqpair->destruct_channel = NULL;
     918             :         }
     919             : 
     920           0 :         if (rqpair->poller && rqpair->poller->need_destroy && RB_EMPTY(&rqpair->poller->qpairs)) {
     921           0 :                 nvmf_rdma_poller_destroy(rqpair->poller);
     922             :         }
     923             : 
     924             :         /* destroy cm_id last so cma device will not be freed before we destroy the cq. */
     925           0 :         if (rqpair->cm_id) {
     926           0 :                 rdma_destroy_id(rqpair->cm_id);
     927             :         }
     928             : 
     929           0 :         free(rqpair);
     930           0 : }
     931             : 
     932             : static int
     933           5 : nvmf_rdma_resize_cq(struct spdk_nvmf_rdma_qpair *rqpair, struct spdk_nvmf_rdma_device *device)
     934             : {
     935             :         struct spdk_nvmf_rdma_poller    *rpoller;
     936             :         int                             rc, num_cqe, required_num_wr;
     937             : 
     938             :         /* Enlarge CQ size dynamically */
     939           5 :         rpoller = rqpair->poller;
     940           5 :         required_num_wr = rpoller->required_num_wr + MAX_WR_PER_QP(rqpair->max_queue_depth);
     941           5 :         num_cqe = rpoller->num_cqe;
     942           5 :         if (num_cqe < required_num_wr) {
     943           4 :                 num_cqe = spdk_max(num_cqe * 2, required_num_wr);
     944           4 :                 num_cqe = spdk_min(num_cqe, device->attr.max_cqe);
     945             :         }
     946             : 
     947           5 :         if (rpoller->num_cqe != num_cqe) {
     948           4 :                 if (device->context->device->transport_type == IBV_TRANSPORT_IWARP) {
     949           1 :                         SPDK_ERRLOG("iWARP doesn't support CQ resize. Current capacity %u, required %u\n"
     950             :                                     "Using CQ of insufficient size may lead to CQ overrun\n", rpoller->num_cqe, num_cqe);
     951           1 :                         return -1;
     952             :                 }
     953           3 :                 if (required_num_wr > device->attr.max_cqe) {
     954           1 :                         SPDK_ERRLOG("RDMA CQE requirement (%d) exceeds device max_cqe limitation (%d)\n",
     955             :                                     required_num_wr, device->attr.max_cqe);
     956           1 :                         return -1;
     957             :                 }
     958             : 
     959           2 :                 SPDK_DEBUGLOG(rdma, "Resize RDMA CQ from %d to %d\n", rpoller->num_cqe, num_cqe);
     960           2 :                 rc = ibv_resize_cq(rpoller->cq, num_cqe);
     961           2 :                 if (rc) {
     962           1 :                         SPDK_ERRLOG("RDMA CQ resize failed: errno %d: %s\n", errno, spdk_strerror(errno));
     963           1 :                         return -1;
     964             :                 }
     965             : 
     966           1 :                 rpoller->num_cqe = num_cqe;
     967             :         }
     968             : 
     969           2 :         rpoller->required_num_wr = required_num_wr;
     970           2 :         return 0;
     971             : }
     972             : 
     973             : static int
     974           0 : nvmf_rdma_qpair_initialize(struct spdk_nvmf_qpair *qpair)
     975             : {
     976             :         struct spdk_nvmf_rdma_qpair             *rqpair;
     977             :         struct spdk_nvmf_rdma_transport         *rtransport;
     978             :         struct spdk_nvmf_transport              *transport;
     979           0 :         struct spdk_nvmf_rdma_resource_opts     opts;
     980             :         struct spdk_nvmf_rdma_device            *device;
     981           0 :         struct spdk_rdma_qp_init_attr           qp_init_attr = {};
     982             : 
     983           0 :         rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair);
     984           0 :         device = rqpair->device;
     985             : 
     986           0 :         qp_init_attr.qp_context = rqpair;
     987           0 :         qp_init_attr.pd         = device->pd;
     988           0 :         qp_init_attr.send_cq    = rqpair->poller->cq;
     989           0 :         qp_init_attr.recv_cq    = rqpair->poller->cq;
     990             : 
     991           0 :         if (rqpair->srq) {
     992           0 :                 qp_init_attr.srq                = rqpair->srq->srq;
     993             :         } else {
     994           0 :                 qp_init_attr.cap.max_recv_wr    = rqpair->max_queue_depth;
     995             :         }
     996             : 
     997             :         /* SEND, READ, and WRITE operations */
     998           0 :         qp_init_attr.cap.max_send_wr    = (uint32_t)rqpair->max_queue_depth * 2;
     999           0 :         qp_init_attr.cap.max_send_sge   = spdk_min((uint32_t)device->attr.max_sge, NVMF_DEFAULT_TX_SGE);
    1000           0 :         qp_init_attr.cap.max_recv_sge   = spdk_min((uint32_t)device->attr.max_sge, NVMF_DEFAULT_RX_SGE);
    1001           0 :         qp_init_attr.stats              = &rqpair->poller->stat.qp_stats;
    1002             : 
    1003           0 :         if (rqpair->srq == NULL && nvmf_rdma_resize_cq(rqpair, device) < 0) {
    1004           0 :                 SPDK_ERRLOG("Failed to resize the completion queue. Cannot initialize qpair.\n");
    1005           0 :                 goto error;
    1006             :         }
    1007             : 
    1008           0 :         rqpair->rdma_qp = spdk_rdma_qp_create(rqpair->cm_id, &qp_init_attr);
    1009           0 :         if (!rqpair->rdma_qp) {
    1010           0 :                 goto error;
    1011             :         }
    1012             : 
    1013           0 :         rqpair->qp_num = rqpair->rdma_qp->qp->qp_num;
    1014             : 
    1015           0 :         rqpair->max_send_depth = spdk_min((uint32_t)(rqpair->max_queue_depth * 2),
    1016             :                                           qp_init_attr.cap.max_send_wr);
    1017           0 :         rqpair->max_send_sge = spdk_min(NVMF_DEFAULT_TX_SGE, qp_init_attr.cap.max_send_sge);
    1018           0 :         rqpair->max_recv_sge = spdk_min(NVMF_DEFAULT_RX_SGE, qp_init_attr.cap.max_recv_sge);
    1019           0 :         spdk_trace_record(TRACE_RDMA_QP_CREATE, 0, 0, (uintptr_t)rqpair);
    1020           0 :         SPDK_DEBUGLOG(rdma, "New RDMA Connection: %p\n", qpair);
    1021             : 
    1022           0 :         if (rqpair->poller->srq == NULL) {
    1023           0 :                 rtransport = SPDK_CONTAINEROF(qpair->transport, struct spdk_nvmf_rdma_transport, transport);
    1024           0 :                 transport = &rtransport->transport;
    1025             : 
    1026           0 :                 opts.qp = rqpair->rdma_qp;
    1027           0 :                 opts.map = device->map;
    1028           0 :                 opts.qpair = rqpair;
    1029           0 :                 opts.shared = false;
    1030           0 :                 opts.max_queue_depth = rqpair->max_queue_depth;
    1031           0 :                 opts.in_capsule_data_size = transport->opts.in_capsule_data_size;
    1032             : 
    1033           0 :                 rqpair->resources = nvmf_rdma_resources_create(&opts);
    1034             : 
    1035           0 :                 if (!rqpair->resources) {
    1036           0 :                         SPDK_ERRLOG("Unable to allocate resources for receive queue.\n");
    1037           0 :                         rdma_destroy_qp(rqpair->cm_id);
    1038           0 :                         goto error;
    1039             :                 }
    1040             :         } else {
    1041           0 :                 rqpair->resources = rqpair->poller->resources;
    1042             :         }
    1043             : 
    1044           0 :         rqpair->current_recv_depth = 0;
    1045           0 :         STAILQ_INIT(&rqpair->pending_rdma_read_queue);
    1046           0 :         STAILQ_INIT(&rqpair->pending_rdma_write_queue);
    1047           0 :         STAILQ_INIT(&rqpair->pending_rdma_send_queue);
    1048           0 :         rqpair->qpair.queue_depth = 0;
    1049             : 
    1050           0 :         return 0;
    1051             : 
    1052           0 : error:
    1053           0 :         rdma_destroy_id(rqpair->cm_id);
    1054           0 :         rqpair->cm_id = NULL;
    1055           0 :         return -1;
    1056             : }
    1057             : 
    1058             : /* Append the given recv wr structure to the resource structs outstanding recvs list. */
    1059             : /* This function accepts either a single wr or the first wr in a linked list. */
    1060             : static void
    1061           6 : nvmf_rdma_qpair_queue_recv_wrs(struct spdk_nvmf_rdma_qpair *rqpair, struct ibv_recv_wr *first)
    1062             : {
    1063           6 :         struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(rqpair->qpair.transport,
    1064             :                         struct spdk_nvmf_rdma_transport, transport);
    1065             : 
    1066           6 :         if (rqpair->srq != NULL) {
    1067           0 :                 spdk_rdma_srq_queue_recv_wrs(rqpair->srq, first);
    1068             :         } else {
    1069           6 :                 if (spdk_rdma_qp_queue_recv_wrs(rqpair->rdma_qp, first)) {
    1070           6 :                         STAILQ_INSERT_TAIL(&rqpair->poller->qpairs_pending_recv, rqpair, recv_link);
    1071             :                 }
    1072             :         }
    1073             : 
    1074           6 :         if (rtransport->rdma_opts.no_wr_batching) {
    1075           0 :                 _poller_submit_recvs(rtransport, rqpair->poller);
    1076             :         }
    1077           6 : }
    1078             : 
    1079             : static int
    1080           4 : request_transfer_in(struct spdk_nvmf_request *req)
    1081             : {
    1082             :         struct spdk_nvmf_rdma_request   *rdma_req;
    1083             :         struct spdk_nvmf_qpair          *qpair;
    1084             :         struct spdk_nvmf_rdma_qpair     *rqpair;
    1085             :         struct spdk_nvmf_rdma_transport *rtransport;
    1086             : 
    1087           4 :         qpair = req->qpair;
    1088           4 :         rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req);
    1089           4 :         rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair);
    1090           4 :         rtransport = SPDK_CONTAINEROF(rqpair->qpair.transport,
    1091             :                                       struct spdk_nvmf_rdma_transport, transport);
    1092             : 
    1093           4 :         assert(req->xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER);
    1094           4 :         assert(rdma_req != NULL);
    1095             : 
    1096           4 :         if (spdk_rdma_qp_queue_send_wrs(rqpair->rdma_qp, rdma_req->transfer_wr)) {
    1097           4 :                 STAILQ_INSERT_TAIL(&rqpair->poller->qpairs_pending_send, rqpair, send_link);
    1098             :         }
    1099           4 :         if (rtransport->rdma_opts.no_wr_batching) {
    1100           0 :                 _poller_submit_sends(rtransport, rqpair->poller);
    1101             :         }
    1102             : 
    1103           4 :         assert(rqpair->current_read_depth + rdma_req->num_outstanding_data_wr <= rqpair->max_read_depth);
    1104           4 :         rqpair->current_read_depth += rdma_req->num_outstanding_data_wr;
    1105           4 :         assert(rqpair->current_send_depth + rdma_req->num_outstanding_data_wr <= rqpair->max_send_depth);
    1106           4 :         rqpair->current_send_depth += rdma_req->num_outstanding_data_wr;
    1107           4 :         return 0;
    1108             : }
    1109             : 
    1110             : static inline void
    1111           0 : nvmf_rdma_request_reset_transfer_in(struct spdk_nvmf_rdma_request *rdma_req,
    1112             :                                     struct spdk_nvmf_rdma_transport *rtransport)
    1113             : {
    1114             :         /* Put completed WRs back to pool and move transfer_wr pointer */
    1115           0 :         _nvmf_rdma_request_free_data(rdma_req, rdma_req->transfer_wr, rtransport->data_wr_pool);
    1116           0 :         rdma_req->transfer_wr = rdma_req->remaining_tranfer_in_wrs;
    1117           0 :         rdma_req->remaining_tranfer_in_wrs = NULL;
    1118           0 :         rdma_req->num_outstanding_data_wr = rdma_req->num_remaining_data_wr;
    1119           0 :         rdma_req->num_remaining_data_wr = 0;
    1120           0 : }
    1121             : 
    1122             : static inline int
    1123           0 : request_prepare_transfer_in_part(struct spdk_nvmf_request *req, uint32_t num_reads_available)
    1124             : {
    1125             :         struct spdk_nvmf_rdma_request   *rdma_req;
    1126             :         struct ibv_send_wr              *wr;
    1127             :         uint32_t i;
    1128             : 
    1129           0 :         rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req);
    1130             : 
    1131           0 :         assert(req->xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER);
    1132           0 :         assert(rdma_req != NULL);
    1133           0 :         assert(num_reads_available > 0);
    1134           0 :         assert(rdma_req->num_outstanding_data_wr > num_reads_available);
    1135           0 :         wr = rdma_req->transfer_wr;
    1136             : 
    1137           0 :         for (i = 0; i < num_reads_available - 1; i++) {
    1138           0 :                 wr = wr->next;
    1139             :         }
    1140             : 
    1141           0 :         rdma_req->remaining_tranfer_in_wrs = wr->next;
    1142           0 :         rdma_req->num_remaining_data_wr = rdma_req->num_outstanding_data_wr - num_reads_available;
    1143           0 :         rdma_req->num_outstanding_data_wr = num_reads_available;
    1144             :         /* Break chain of WRs to send only part. Once this portion completes, we continue sending RDMA_READs */
    1145           0 :         wr->next = NULL;
    1146             : 
    1147           0 :         return 0;
    1148             : }
    1149             : 
    1150             : static int
    1151           6 : request_transfer_out(struct spdk_nvmf_request *req, int *data_posted)
    1152             : {
    1153           6 :         int                             num_outstanding_data_wr = 0;
    1154             :         struct spdk_nvmf_rdma_request   *rdma_req;
    1155             :         struct spdk_nvmf_qpair          *qpair;
    1156             :         struct spdk_nvmf_rdma_qpair     *rqpair;
    1157             :         struct spdk_nvme_cpl            *rsp;
    1158           6 :         struct ibv_send_wr              *first = NULL;
    1159             :         struct spdk_nvmf_rdma_transport *rtransport;
    1160             : 
    1161           6 :         *data_posted = 0;
    1162           6 :         qpair = req->qpair;
    1163           6 :         rsp = &req->rsp->nvme_cpl;
    1164           6 :         rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req);
    1165           6 :         rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair);
    1166           6 :         rtransport = SPDK_CONTAINEROF(rqpair->qpair.transport,
    1167             :                                       struct spdk_nvmf_rdma_transport, transport);
    1168             : 
    1169             :         /* Advance our sq_head pointer */
    1170           6 :         if (qpair->sq_head == qpair->sq_head_max) {
    1171           6 :                 qpair->sq_head = 0;
    1172             :         } else {
    1173           0 :                 qpair->sq_head++;
    1174             :         }
    1175           6 :         rsp->sqhd = qpair->sq_head;
    1176             : 
    1177             :         /* queue the capsule for the recv buffer */
    1178           6 :         assert(rdma_req->recv != NULL);
    1179             : 
    1180           6 :         nvmf_rdma_qpair_queue_recv_wrs(rqpair, &rdma_req->recv->wr);
    1181             : 
    1182           6 :         rdma_req->recv = NULL;
    1183           6 :         assert(rqpair->current_recv_depth > 0);
    1184           6 :         rqpair->current_recv_depth--;
    1185             : 
    1186             :         /* Build the response which consists of optional
    1187             :          * RDMA WRITEs to transfer data, plus an RDMA SEND
    1188             :          * containing the response.
    1189             :          */
    1190           6 :         first = &rdma_req->rsp.wr;
    1191             : 
    1192           6 :         if (spdk_unlikely(rsp->status.sc != SPDK_NVME_SC_SUCCESS)) {
    1193             :                 /* On failure, data was not read from the controller. So clear the
    1194             :                  * number of outstanding data WRs to zero.
    1195             :                  */
    1196           1 :                 rdma_req->num_outstanding_data_wr = 0;
    1197           5 :         } else if (req->xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) {
    1198           1 :                 first = rdma_req->transfer_wr;
    1199           1 :                 *data_posted = 1;
    1200           1 :                 num_outstanding_data_wr = rdma_req->num_outstanding_data_wr;
    1201             :         }
    1202           6 :         if (spdk_rdma_qp_queue_send_wrs(rqpair->rdma_qp, first)) {
    1203           6 :                 STAILQ_INSERT_TAIL(&rqpair->poller->qpairs_pending_send, rqpair, send_link);
    1204             :         }
    1205           6 :         if (rtransport->rdma_opts.no_wr_batching) {
    1206           0 :                 _poller_submit_sends(rtransport, rqpair->poller);
    1207             :         }
    1208             : 
    1209             :         /* +1 for the rsp wr */
    1210           6 :         assert(rqpair->current_send_depth + num_outstanding_data_wr + 1 <= rqpair->max_send_depth);
    1211           6 :         rqpair->current_send_depth += num_outstanding_data_wr + 1;
    1212             : 
    1213           6 :         return 0;
    1214             : }
    1215             : 
    1216             : static int
    1217           0 : nvmf_rdma_event_accept(struct rdma_cm_id *id, struct spdk_nvmf_rdma_qpair *rqpair)
    1218             : {
    1219           0 :         struct spdk_nvmf_rdma_accept_private_data       accept_data;
    1220           0 :         struct rdma_conn_param                          ctrlr_event_data = {};
    1221             :         int                                             rc;
    1222             : 
    1223           0 :         accept_data.recfmt = 0;
    1224           0 :         accept_data.crqsize = rqpair->max_queue_depth;
    1225             : 
    1226           0 :         ctrlr_event_data.private_data = &accept_data;
    1227           0 :         ctrlr_event_data.private_data_len = sizeof(accept_data);
    1228           0 :         if (id->ps == RDMA_PS_TCP) {
    1229           0 :                 ctrlr_event_data.responder_resources = 0; /* We accept 0 reads from the host */
    1230           0 :                 ctrlr_event_data.initiator_depth = rqpair->max_read_depth;
    1231             :         }
    1232             : 
    1233             :         /* Configure infinite retries for the initiator side qpair.
    1234             :          * We need to pass this value to the initiator to prevent the
    1235             :          * initiator side NIC from completing SEND requests back to the
    1236             :          * initiator with status rnr_retry_count_exceeded. */
    1237           0 :         ctrlr_event_data.rnr_retry_count = 0x7;
    1238             : 
    1239             :         /* When qpair is created without use of rdma cm API, an additional
    1240             :          * information must be provided to initiator in the connection response:
    1241             :          * whether qpair is using SRQ and its qp_num
    1242             :          * Fields below are ignored by rdma cm if qpair has been
    1243             :          * created using rdma cm API. */
    1244           0 :         ctrlr_event_data.srq = rqpair->srq ? 1 : 0;
    1245           0 :         ctrlr_event_data.qp_num = rqpair->qp_num;
    1246             : 
    1247           0 :         rc = spdk_rdma_qp_accept(rqpair->rdma_qp, &ctrlr_event_data);
    1248           0 :         if (rc) {
    1249           0 :                 SPDK_ERRLOG("Error %d on spdk_rdma_qp_accept\n", errno);
    1250             :         } else {
    1251           0 :                 SPDK_DEBUGLOG(rdma, "Sent back the accept\n");
    1252             :         }
    1253             : 
    1254           0 :         return rc;
    1255             : }
    1256             : 
    1257             : static void
    1258           0 : nvmf_rdma_event_reject(struct rdma_cm_id *id, enum spdk_nvmf_rdma_transport_error error)
    1259             : {
    1260           0 :         struct spdk_nvmf_rdma_reject_private_data       rej_data;
    1261             : 
    1262           0 :         rej_data.recfmt = 0;
    1263           0 :         rej_data.sts = error;
    1264             : 
    1265           0 :         rdma_reject(id, &rej_data, sizeof(rej_data));
    1266           0 : }
    1267             : 
    1268             : static int
    1269           0 : nvmf_rdma_connect(struct spdk_nvmf_transport *transport, struct rdma_cm_event *event)
    1270             : {
    1271             :         struct spdk_nvmf_rdma_transport *rtransport;
    1272           0 :         struct spdk_nvmf_rdma_qpair     *rqpair = NULL;
    1273             :         struct spdk_nvmf_rdma_port      *port;
    1274           0 :         struct rdma_conn_param          *rdma_param = NULL;
    1275           0 :         const struct spdk_nvmf_rdma_request_private_data *private_data = NULL;
    1276             :         uint16_t                        max_queue_depth;
    1277             :         uint16_t                        max_read_depth;
    1278             : 
    1279           0 :         rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport);
    1280             : 
    1281           0 :         assert(event->id != NULL); /* Impossible. Can't even reject the connection. */
    1282           0 :         assert(event->id->verbs != NULL); /* Impossible. No way to handle this. */
    1283             : 
    1284           0 :         rdma_param = &event->param.conn;
    1285           0 :         if (rdma_param->private_data == NULL ||
    1286           0 :             rdma_param->private_data_len < sizeof(struct spdk_nvmf_rdma_request_private_data)) {
    1287           0 :                 SPDK_ERRLOG("connect request: no private data provided\n");
    1288           0 :                 nvmf_rdma_event_reject(event->id, SPDK_NVMF_RDMA_ERROR_INVALID_PRIVATE_DATA_LENGTH);
    1289           0 :                 return -1;
    1290             :         }
    1291             : 
    1292           0 :         private_data = rdma_param->private_data;
    1293           0 :         if (private_data->recfmt != 0) {
    1294           0 :                 SPDK_ERRLOG("Received RDMA private data with RECFMT != 0\n");
    1295           0 :                 nvmf_rdma_event_reject(event->id, SPDK_NVMF_RDMA_ERROR_INVALID_RECFMT);
    1296           0 :                 return -1;
    1297             :         }
    1298             : 
    1299           0 :         SPDK_DEBUGLOG(rdma, "Connect Recv on fabric intf name %s, dev_name %s\n",
    1300             :                       event->id->verbs->device->name, event->id->verbs->device->dev_name);
    1301             : 
    1302           0 :         port = event->listen_id->context;
    1303           0 :         SPDK_DEBUGLOG(rdma, "Listen Id was %p with verbs %p. ListenAddr: %p\n",
    1304             :                       event->listen_id, event->listen_id->verbs, port);
    1305             : 
    1306             :         /* Figure out the supported queue depth. This is a multi-step process
    1307             :          * that takes into account hardware maximums, host provided values,
    1308             :          * and our target's internal memory limits */
    1309             : 
    1310           0 :         SPDK_DEBUGLOG(rdma, "Calculating Queue Depth\n");
    1311             : 
    1312             :         /* Start with the maximum queue depth allowed by the target */
    1313           0 :         max_queue_depth = rtransport->transport.opts.max_queue_depth;
    1314           0 :         max_read_depth = rtransport->transport.opts.max_queue_depth;
    1315           0 :         SPDK_DEBUGLOG(rdma, "Target Max Queue Depth: %d\n",
    1316             :                       rtransport->transport.opts.max_queue_depth);
    1317             : 
    1318             :         /* Next check the local NIC's hardware limitations */
    1319           0 :         SPDK_DEBUGLOG(rdma,
    1320             :                       "Local NIC Max Send/Recv Queue Depth: %d Max Read/Write Queue Depth: %d\n",
    1321             :                       port->device->attr.max_qp_wr, port->device->attr.max_qp_rd_atom);
    1322           0 :         max_queue_depth = spdk_min(max_queue_depth, port->device->attr.max_qp_wr);
    1323           0 :         max_read_depth = spdk_min(max_read_depth, port->device->attr.max_qp_init_rd_atom);
    1324             : 
    1325             :         /* Next check the remote NIC's hardware limitations */
    1326           0 :         SPDK_DEBUGLOG(rdma,
    1327             :                       "Host (Initiator) NIC Max Incoming RDMA R/W operations: %d Max Outgoing RDMA R/W operations: %d\n",
    1328             :                       rdma_param->initiator_depth, rdma_param->responder_resources);
    1329             :         /* from man3 rdma_get_cm_event
    1330             :          * responder_resources - Specifies the number of responder resources that is requested by the recipient.
    1331             :          * The responder_resources field must match the initiator depth specified by the remote node when running
    1332             :          * the rdma_connect and rdma_accept functions. */
    1333           0 :         if (rdma_param->responder_resources != 0) {
    1334           0 :                 if (private_data->qid) {
    1335           0 :                         SPDK_DEBUGLOG(rdma, "Host (Initiator) is not allowed to use RDMA operations,"
    1336             :                                       " responder_resources must be 0 but set to %u\n",
    1337             :                                       rdma_param->responder_resources);
    1338             :                 } else {
    1339           0 :                         SPDK_WARNLOG("Host (Initiator) is not allowed to use RDMA operations,"
    1340             :                                      " responder_resources must be 0 but set to %u\n",
    1341             :                                      rdma_param->responder_resources);
    1342             :                 }
    1343             :         }
    1344             :         /* from man3 rdma_get_cm_event
    1345             :          * initiator_depth - Specifies the maximum number of outstanding RDMA read operations that the recipient holds.
    1346             :          * The initiator_depth field must match the responder resources specified by the remote node when running
    1347             :          * the rdma_connect and rdma_accept functions. */
    1348           0 :         if (rdma_param->initiator_depth == 0) {
    1349           0 :                 SPDK_ERRLOG("Host (Initiator) doesn't support RDMA_READ or atomic operations\n");
    1350           0 :                 nvmf_rdma_event_reject(event->id, SPDK_NVMF_RDMA_ERROR_INVALID_IRD);
    1351           0 :                 return -1;
    1352             :         }
    1353           0 :         max_read_depth = spdk_min(max_read_depth, rdma_param->initiator_depth);
    1354             : 
    1355           0 :         SPDK_DEBUGLOG(rdma, "Host Receive Queue Size: %d\n", private_data->hrqsize);
    1356           0 :         SPDK_DEBUGLOG(rdma, "Host Send Queue Size: %d\n", private_data->hsqsize);
    1357           0 :         max_queue_depth = spdk_min(max_queue_depth, private_data->hrqsize);
    1358           0 :         max_queue_depth = spdk_min(max_queue_depth, private_data->hsqsize + 1);
    1359             : 
    1360           0 :         SPDK_DEBUGLOG(rdma, "Final Negotiated Queue Depth: %d R/W Depth: %d\n",
    1361             :                       max_queue_depth, max_read_depth);
    1362             : 
    1363           0 :         rqpair = calloc(1, sizeof(struct spdk_nvmf_rdma_qpair));
    1364           0 :         if (rqpair == NULL) {
    1365           0 :                 SPDK_ERRLOG("Could not allocate new connection.\n");
    1366           0 :                 nvmf_rdma_event_reject(event->id, SPDK_NVMF_RDMA_ERROR_NO_RESOURCES);
    1367           0 :                 return -1;
    1368             :         }
    1369             : 
    1370           0 :         rqpair->device = port->device;
    1371           0 :         rqpair->max_queue_depth = max_queue_depth;
    1372           0 :         rqpair->max_read_depth = max_read_depth;
    1373           0 :         rqpair->cm_id = event->id;
    1374           0 :         rqpair->listen_id = event->listen_id;
    1375           0 :         rqpair->qpair.transport = transport;
    1376           0 :         STAILQ_INIT(&rqpair->ibv_events);
    1377             :         /* use qid from the private data to determine the qpair type
    1378             :            qid will be set to the appropriate value when the controller is created */
    1379           0 :         rqpair->qpair.qid = private_data->qid;
    1380             : 
    1381           0 :         event->id->context = &rqpair->qpair;
    1382             : 
    1383           0 :         spdk_nvmf_tgt_new_qpair(transport->tgt, &rqpair->qpair);
    1384             : 
    1385           0 :         return 0;
    1386             : }
    1387             : 
    1388             : static inline void
    1389          28 : nvmf_rdma_setup_wr(struct ibv_send_wr *wr, struct ibv_send_wr *next,
    1390             :                    enum spdk_nvme_data_transfer xfer)
    1391             : {
    1392          28 :         if (xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) {
    1393          24 :                 wr->opcode = IBV_WR_RDMA_WRITE;
    1394          24 :                 wr->send_flags = 0;
    1395          24 :                 wr->next = next;
    1396           4 :         } else if (xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER) {
    1397           4 :                 wr->opcode = IBV_WR_RDMA_READ;
    1398           4 :                 wr->send_flags = IBV_SEND_SIGNALED;
    1399           4 :                 wr->next = NULL;
    1400             :         } else {
    1401           0 :                 assert(0);
    1402             :         }
    1403          28 : }
    1404             : 
    1405             : static int
    1406           6 : nvmf_request_alloc_wrs(struct spdk_nvmf_rdma_transport *rtransport,
    1407             :                        struct spdk_nvmf_rdma_request *rdma_req,
    1408             :                        uint32_t num_sgl_descriptors)
    1409             : {
    1410           6 :         struct spdk_nvmf_rdma_request_data      *work_requests[SPDK_NVMF_MAX_SGL_ENTRIES];
    1411             :         struct spdk_nvmf_rdma_request_data      *current_data_wr;
    1412             :         uint32_t                                i;
    1413             : 
    1414           6 :         if (spdk_unlikely(num_sgl_descriptors > SPDK_NVMF_MAX_SGL_ENTRIES)) {
    1415           0 :                 SPDK_ERRLOG("Requested too much entries (%u), the limit is %u\n",
    1416             :                             num_sgl_descriptors, SPDK_NVMF_MAX_SGL_ENTRIES);
    1417           0 :                 return -EINVAL;
    1418             :         }
    1419             : 
    1420           6 :         if (spdk_unlikely(spdk_mempool_get_bulk(rtransport->data_wr_pool, (void **)work_requests,
    1421             :                                                 num_sgl_descriptors))) {
    1422           0 :                 return -ENOMEM;
    1423             :         }
    1424             : 
    1425           6 :         current_data_wr = &rdma_req->data;
    1426             : 
    1427          12 :         for (i = 0; i < num_sgl_descriptors; i++) {
    1428           6 :                 nvmf_rdma_setup_wr(&current_data_wr->wr, &work_requests[i]->wr, rdma_req->req.xfer);
    1429           6 :                 current_data_wr->wr.next = &work_requests[i]->wr;
    1430           6 :                 current_data_wr = work_requests[i];
    1431           6 :                 current_data_wr->wr.sg_list = current_data_wr->sgl;
    1432           6 :                 current_data_wr->wr.wr_id = rdma_req->data.wr.wr_id;
    1433             :         }
    1434             : 
    1435           6 :         nvmf_rdma_setup_wr(&current_data_wr->wr, &rdma_req->rsp.wr, rdma_req->req.xfer);
    1436             : 
    1437           6 :         return 0;
    1438             : }
    1439             : 
    1440             : static inline void
    1441          16 : nvmf_rdma_setup_request(struct spdk_nvmf_rdma_request *rdma_req)
    1442             : {
    1443          16 :         struct ibv_send_wr              *wr = &rdma_req->data.wr;
    1444          16 :         struct spdk_nvme_sgl_descriptor *sgl = &rdma_req->req.cmd->nvme_cmd.dptr.sgl1;
    1445             : 
    1446          16 :         wr->wr.rdma.rkey = sgl->keyed.key;
    1447          16 :         wr->wr.rdma.remote_addr = sgl->address;
    1448          16 :         nvmf_rdma_setup_wr(wr, &rdma_req->rsp.wr, rdma_req->req.xfer);
    1449          16 : }
    1450             : 
    1451             : static inline void
    1452           1 : nvmf_rdma_update_remote_addr(struct spdk_nvmf_rdma_request *rdma_req, uint32_t num_wrs)
    1453             : {
    1454           1 :         struct ibv_send_wr              *wr = &rdma_req->data.wr;
    1455           1 :         struct spdk_nvme_sgl_descriptor *sgl = &rdma_req->req.cmd->nvme_cmd.dptr.sgl1;
    1456             :         uint32_t                        i;
    1457             :         int                             j;
    1458           1 :         uint64_t                        remote_addr_offset = 0;
    1459             : 
    1460           3 :         for (i = 0; i < num_wrs; ++i) {
    1461           2 :                 wr->wr.rdma.rkey = sgl->keyed.key;
    1462           2 :                 wr->wr.rdma.remote_addr = sgl->address + remote_addr_offset;
    1463          19 :                 for (j = 0; j < wr->num_sge; ++j) {
    1464          17 :                         remote_addr_offset += wr->sg_list[j].length;
    1465             :                 }
    1466           2 :                 wr = wr->next;
    1467             :         }
    1468           1 : }
    1469             : 
    1470             : static int
    1471          15 : nvmf_rdma_fill_wr_sgl(struct spdk_nvmf_rdma_device *device,
    1472             :                       struct spdk_nvmf_rdma_request *rdma_req,
    1473             :                       struct ibv_send_wr *wr,
    1474             :                       uint32_t total_length)
    1475             : {
    1476          15 :         struct spdk_rdma_memory_translation mem_translation;
    1477             :         struct ibv_sge  *sg_ele;
    1478             :         struct iovec *iov;
    1479             :         uint32_t lkey, remaining;
    1480             :         int rc;
    1481             : 
    1482          15 :         wr->num_sge = 0;
    1483             : 
    1484          74 :         while (total_length && wr->num_sge < SPDK_NVMF_MAX_SGL_ENTRIES) {
    1485          59 :                 iov = &rdma_req->req.iov[rdma_req->iovpos];
    1486          59 :                 rc = spdk_rdma_get_translation(device->map, iov->iov_base, iov->iov_len, &mem_translation);
    1487          59 :                 if (spdk_unlikely(rc)) {
    1488           0 :                         return rc;
    1489             :                 }
    1490             : 
    1491          59 :                 lkey = spdk_rdma_memory_translation_get_lkey(&mem_translation);
    1492          59 :                 sg_ele = &wr->sg_list[wr->num_sge];
    1493          59 :                 remaining = spdk_min((uint32_t)iov->iov_len - rdma_req->offset, total_length);
    1494             : 
    1495          59 :                 sg_ele->lkey = lkey;
    1496          59 :                 sg_ele->addr = (uintptr_t)iov->iov_base + rdma_req->offset;
    1497          59 :                 sg_ele->length = remaining;
    1498          59 :                 SPDK_DEBUGLOG(rdma, "sge[%d] %p addr 0x%"PRIx64", len %u\n", wr->num_sge, sg_ele, sg_ele->addr,
    1499             :                               sg_ele->length);
    1500          59 :                 rdma_req->offset += sg_ele->length;
    1501          59 :                 total_length -= sg_ele->length;
    1502          59 :                 wr->num_sge++;
    1503             : 
    1504          59 :                 if (rdma_req->offset == iov->iov_len) {
    1505          57 :                         rdma_req->offset = 0;
    1506          57 :                         rdma_req->iovpos++;
    1507             :                 }
    1508             :         }
    1509             : 
    1510          15 :         if (spdk_unlikely(total_length)) {
    1511           0 :                 SPDK_ERRLOG("Not enough SG entries to hold data buffer\n");
    1512           0 :                 return -EINVAL;
    1513             :         }
    1514             : 
    1515          15 :         return 0;
    1516             : }
    1517             : 
    1518             : static int
    1519          10 : nvmf_rdma_fill_wr_sgl_with_dif(struct spdk_nvmf_rdma_device *device,
    1520             :                                struct spdk_nvmf_rdma_request *rdma_req,
    1521             :                                struct ibv_send_wr *wr,
    1522             :                                uint32_t total_length,
    1523             :                                uint32_t num_extra_wrs)
    1524             : {
    1525          10 :         struct spdk_rdma_memory_translation mem_translation;
    1526          10 :         struct spdk_dif_ctx *dif_ctx = &rdma_req->req.dif.dif_ctx;
    1527             :         struct ibv_sge *sg_ele;
    1528             :         struct iovec *iov;
    1529             :         struct iovec *rdma_iov;
    1530             :         uint32_t lkey, remaining;
    1531             :         uint32_t remaining_data_block, data_block_size, md_size;
    1532             :         uint32_t sge_len;
    1533             :         int rc;
    1534             : 
    1535          10 :         data_block_size = dif_ctx->block_size - dif_ctx->md_size;
    1536             : 
    1537          10 :         if (spdk_likely(!rdma_req->req.stripped_data)) {
    1538           5 :                 rdma_iov = rdma_req->req.iov;
    1539           5 :                 remaining_data_block = data_block_size;
    1540           5 :                 md_size = dif_ctx->md_size;
    1541             :         } else {
    1542           5 :                 rdma_iov = rdma_req->req.stripped_data->iov;
    1543           5 :                 total_length = total_length / dif_ctx->block_size * data_block_size;
    1544           5 :                 remaining_data_block = total_length;
    1545           5 :                 md_size = 0;
    1546             :         }
    1547             : 
    1548          10 :         wr->num_sge = 0;
    1549             : 
    1550          25 :         while (total_length && (num_extra_wrs || wr->num_sge < SPDK_NVMF_MAX_SGL_ENTRIES)) {
    1551          15 :                 iov = rdma_iov + rdma_req->iovpos;
    1552          15 :                 rc = spdk_rdma_get_translation(device->map, iov->iov_base, iov->iov_len, &mem_translation);
    1553          15 :                 if (spdk_unlikely(rc)) {
    1554           0 :                         return rc;
    1555             :                 }
    1556             : 
    1557          15 :                 lkey = spdk_rdma_memory_translation_get_lkey(&mem_translation);
    1558          15 :                 sg_ele = &wr->sg_list[wr->num_sge];
    1559          15 :                 remaining = spdk_min((uint32_t)iov->iov_len - rdma_req->offset, total_length);
    1560             : 
    1561          53 :                 while (remaining) {
    1562          38 :                         if (wr->num_sge >= SPDK_NVMF_MAX_SGL_ENTRIES) {
    1563           1 :                                 if (num_extra_wrs > 0 && wr->next) {
    1564           1 :                                         wr = wr->next;
    1565           1 :                                         wr->num_sge = 0;
    1566           1 :                                         sg_ele = &wr->sg_list[wr->num_sge];
    1567           1 :                                         num_extra_wrs--;
    1568             :                                 } else {
    1569             :                                         break;
    1570             :                                 }
    1571             :                         }
    1572          38 :                         sg_ele->lkey = lkey;
    1573          38 :                         sg_ele->addr = (uintptr_t)((char *)iov->iov_base + rdma_req->offset);
    1574          38 :                         sge_len = spdk_min(remaining, remaining_data_block);
    1575          38 :                         sg_ele->length = sge_len;
    1576          38 :                         SPDK_DEBUGLOG(rdma, "sge[%d] %p addr 0x%"PRIx64", len %u\n", wr->num_sge, sg_ele,
    1577             :                                       sg_ele->addr, sg_ele->length);
    1578          38 :                         remaining -= sge_len;
    1579          38 :                         remaining_data_block -= sge_len;
    1580          38 :                         rdma_req->offset += sge_len;
    1581          38 :                         total_length -= sge_len;
    1582             : 
    1583          38 :                         sg_ele++;
    1584          38 :                         wr->num_sge++;
    1585             : 
    1586          38 :                         if (remaining_data_block == 0) {
    1587             :                                 /* skip metadata */
    1588          34 :                                 rdma_req->offset += md_size;
    1589          34 :                                 total_length -= md_size;
    1590             :                                 /* Metadata that do not fit this IO buffer will be included in the next IO buffer */
    1591          34 :                                 remaining -= spdk_min(remaining, md_size);
    1592          34 :                                 remaining_data_block = data_block_size;
    1593             :                         }
    1594             : 
    1595          38 :                         if (remaining == 0) {
    1596             :                                 /* By subtracting the size of the last IOV from the offset, we ensure that we skip
    1597             :                                    the remaining metadata bits at the beginning of the next buffer */
    1598          15 :                                 rdma_req->offset -= spdk_min(iov->iov_len, rdma_req->offset);
    1599          15 :                                 rdma_req->iovpos++;
    1600             :                         }
    1601             :                 }
    1602             :         }
    1603             : 
    1604          10 :         if (spdk_unlikely(total_length)) {
    1605           0 :                 SPDK_ERRLOG("Not enough SG entries to hold data buffer\n");
    1606           0 :                 return -EINVAL;
    1607             :         }
    1608             : 
    1609          10 :         return 0;
    1610             : }
    1611             : 
    1612             : static inline uint32_t
    1613           8 : nvmf_rdma_calc_num_wrs(uint32_t length, uint32_t io_unit_size, uint32_t block_size)
    1614             : {
    1615             :         /* estimate the number of SG entries and WRs needed to process the request */
    1616           8 :         uint32_t num_sge = 0;
    1617             :         uint32_t i;
    1618           8 :         uint32_t num_buffers = SPDK_CEIL_DIV(length, io_unit_size);
    1619             : 
    1620          23 :         for (i = 0; i < num_buffers && length > 0; i++) {
    1621          15 :                 uint32_t buffer_len = spdk_min(length, io_unit_size);
    1622          15 :                 uint32_t num_sge_in_block = SPDK_CEIL_DIV(buffer_len, block_size);
    1623             : 
    1624          15 :                 if (num_sge_in_block * block_size > buffer_len) {
    1625          11 :                         ++num_sge_in_block;
    1626             :                 }
    1627          15 :                 num_sge += num_sge_in_block;
    1628          15 :                 length -= buffer_len;
    1629             :         }
    1630           8 :         return SPDK_CEIL_DIV(num_sge, SPDK_NVMF_MAX_SGL_ENTRIES);
    1631             : }
    1632             : 
    1633             : static int
    1634          16 : nvmf_rdma_request_fill_iovs(struct spdk_nvmf_rdma_transport *rtransport,
    1635             :                             struct spdk_nvmf_rdma_device *device,
    1636             :                             struct spdk_nvmf_rdma_request *rdma_req)
    1637             : {
    1638             :         struct spdk_nvmf_rdma_qpair             *rqpair;
    1639             :         struct spdk_nvmf_rdma_poll_group        *rgroup;
    1640          16 :         struct spdk_nvmf_request                *req = &rdma_req->req;
    1641          16 :         struct ibv_send_wr                      *wr = &rdma_req->data.wr;
    1642             :         int                                     rc;
    1643          16 :         uint32_t                                num_wrs = 1;
    1644             :         uint32_t                                length;
    1645             : 
    1646          16 :         rqpair = SPDK_CONTAINEROF(req->qpair, struct spdk_nvmf_rdma_qpair, qpair);
    1647          16 :         rgroup = rqpair->poller->group;
    1648             : 
    1649             :         /* rdma wr specifics */
    1650          16 :         nvmf_rdma_setup_request(rdma_req);
    1651             : 
    1652          16 :         length = req->length;
    1653          16 :         if (spdk_unlikely(req->dif_enabled)) {
    1654           8 :                 req->dif.orig_length = length;
    1655           8 :                 length = spdk_dif_get_length_with_md(length, &req->dif.dif_ctx);
    1656           8 :                 req->dif.elba_length = length;
    1657             :         }
    1658             : 
    1659          16 :         rc = spdk_nvmf_request_get_buffers(req, &rgroup->group, &rtransport->transport,
    1660             :                                            length);
    1661          16 :         if (spdk_unlikely(rc != 0)) {
    1662           1 :                 return rc;
    1663             :         }
    1664             : 
    1665          15 :         assert(req->iovcnt <= rqpair->max_send_sge);
    1666             : 
    1667             :         /* When dif_insert_or_strip is true and the I/O data length is greater than one block,
    1668             :          * the stripped_buffers are got for DIF stripping. */
    1669          15 :         if (spdk_unlikely(req->dif_enabled && (req->xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST)
    1670             :                           && (req->dif.elba_length > req->dif.dif_ctx.block_size))) {
    1671           7 :                 rc = nvmf_request_get_stripped_buffers(req, &rgroup->group,
    1672             :                                                        &rtransport->transport, req->dif.orig_length);
    1673           7 :                 if (rc != 0) {
    1674           4 :                         SPDK_INFOLOG(rdma, "Get stripped buffers fail %d, fallback to req.iov.\n", rc);
    1675             :                 }
    1676             :         }
    1677             : 
    1678          15 :         rdma_req->iovpos = 0;
    1679             : 
    1680          15 :         if (spdk_unlikely(req->dif_enabled)) {
    1681           8 :                 num_wrs = nvmf_rdma_calc_num_wrs(length, rtransport->transport.opts.io_unit_size,
    1682             :                                                  req->dif.dif_ctx.block_size);
    1683           8 :                 if (num_wrs > 1) {
    1684           1 :                         rc = nvmf_request_alloc_wrs(rtransport, rdma_req, num_wrs - 1);
    1685           1 :                         if (spdk_unlikely(rc != 0)) {
    1686           0 :                                 goto err_exit;
    1687             :                         }
    1688             :                 }
    1689             : 
    1690           8 :                 rc = nvmf_rdma_fill_wr_sgl_with_dif(device, rdma_req, wr, length, num_wrs - 1);
    1691           8 :                 if (spdk_unlikely(rc != 0)) {
    1692           0 :                         goto err_exit;
    1693             :                 }
    1694             : 
    1695           8 :                 if (num_wrs > 1) {
    1696           1 :                         nvmf_rdma_update_remote_addr(rdma_req, num_wrs);
    1697             :                 }
    1698             :         } else {
    1699           7 :                 rc = nvmf_rdma_fill_wr_sgl(device, rdma_req, wr, length);
    1700           7 :                 if (spdk_unlikely(rc != 0)) {
    1701           0 :                         goto err_exit;
    1702             :                 }
    1703             :         }
    1704             : 
    1705             :         /* set the number of outstanding data WRs for this request. */
    1706          15 :         rdma_req->num_outstanding_data_wr = num_wrs;
    1707             : 
    1708          15 :         return rc;
    1709             : 
    1710           0 : err_exit:
    1711           0 :         spdk_nvmf_request_free_buffers(req, &rgroup->group, &rtransport->transport);
    1712           0 :         nvmf_rdma_request_free_data(rdma_req, rtransport);
    1713           0 :         req->iovcnt = 0;
    1714           0 :         return rc;
    1715             : }
    1716             : 
    1717             : static int
    1718           5 : nvmf_rdma_request_fill_iovs_multi_sgl(struct spdk_nvmf_rdma_transport *rtransport,
    1719             :                                       struct spdk_nvmf_rdma_device *device,
    1720             :                                       struct spdk_nvmf_rdma_request *rdma_req)
    1721             : {
    1722             :         struct spdk_nvmf_rdma_qpair             *rqpair;
    1723             :         struct spdk_nvmf_rdma_poll_group        *rgroup;
    1724             :         struct ibv_send_wr                      *current_wr;
    1725           5 :         struct spdk_nvmf_request                *req = &rdma_req->req;
    1726             :         struct spdk_nvme_sgl_descriptor         *inline_segment, *desc;
    1727             :         uint32_t                                num_sgl_descriptors;
    1728           5 :         uint32_t                                lengths[SPDK_NVMF_MAX_SGL_ENTRIES], total_length = 0;
    1729             :         uint32_t                                i;
    1730             :         int                                     rc;
    1731             : 
    1732           5 :         rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair);
    1733           5 :         rgroup = rqpair->poller->group;
    1734             : 
    1735           5 :         inline_segment = &req->cmd->nvme_cmd.dptr.sgl1;
    1736           5 :         assert(inline_segment->generic.type == SPDK_NVME_SGL_TYPE_LAST_SEGMENT);
    1737           5 :         assert(inline_segment->unkeyed.subtype == SPDK_NVME_SGL_SUBTYPE_OFFSET);
    1738             : 
    1739           5 :         num_sgl_descriptors = inline_segment->unkeyed.length / sizeof(struct spdk_nvme_sgl_descriptor);
    1740           5 :         assert(num_sgl_descriptors <= SPDK_NVMF_MAX_SGL_ENTRIES);
    1741             : 
    1742           5 :         desc = (struct spdk_nvme_sgl_descriptor *)rdma_req->recv->buf + inline_segment->address;
    1743          15 :         for (i = 0; i < num_sgl_descriptors; i++) {
    1744          10 :                 if (spdk_likely(!req->dif_enabled)) {
    1745           8 :                         lengths[i] = desc->keyed.length;
    1746             :                 } else {
    1747           2 :                         req->dif.orig_length += desc->keyed.length;
    1748           2 :                         lengths[i] = spdk_dif_get_length_with_md(desc->keyed.length, &req->dif.dif_ctx);
    1749           2 :                         req->dif.elba_length += lengths[i];
    1750             :                 }
    1751          10 :                 total_length += lengths[i];
    1752          10 :                 desc++;
    1753             :         }
    1754             : 
    1755           5 :         if (spdk_unlikely(total_length > rtransport->transport.opts.max_io_size)) {
    1756           0 :                 SPDK_ERRLOG("Multi SGL length 0x%x exceeds max io size 0x%x\n",
    1757             :                             total_length, rtransport->transport.opts.max_io_size);
    1758           0 :                 req->rsp->nvme_cpl.status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID;
    1759           0 :                 return -EINVAL;
    1760             :         }
    1761             : 
    1762           5 :         rc = nvmf_request_alloc_wrs(rtransport, rdma_req, num_sgl_descriptors - 1);
    1763           5 :         if (spdk_unlikely(rc != 0)) {
    1764           0 :                 return -ENOMEM;
    1765             :         }
    1766             : 
    1767           5 :         rc = spdk_nvmf_request_get_buffers(req, &rgroup->group, &rtransport->transport, total_length);
    1768           5 :         if (spdk_unlikely(rc != 0)) {
    1769           0 :                 nvmf_rdma_request_free_data(rdma_req, rtransport);
    1770           0 :                 return rc;
    1771             :         }
    1772             : 
    1773             :         /* When dif_insert_or_strip is true and the I/O data length is greater than one block,
    1774             :          * the stripped_buffers are got for DIF stripping. */
    1775           5 :         if (spdk_unlikely(req->dif_enabled && (req->xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST)
    1776             :                           && (req->dif.elba_length > req->dif.dif_ctx.block_size))) {
    1777           1 :                 rc = nvmf_request_get_stripped_buffers(req, &rgroup->group,
    1778             :                                                        &rtransport->transport, req->dif.orig_length);
    1779           1 :                 if (spdk_unlikely(rc != 0)) {
    1780           0 :                         SPDK_INFOLOG(rdma, "Get stripped buffers fail %d, fallback to req.iov.\n", rc);
    1781             :                 }
    1782             :         }
    1783             : 
    1784             :         /* The first WR must always be the embedded data WR. This is how we unwind them later. */
    1785           5 :         current_wr = &rdma_req->data.wr;
    1786           5 :         assert(current_wr != NULL);
    1787             : 
    1788           5 :         req->length = 0;
    1789           5 :         rdma_req->iovpos = 0;
    1790           5 :         desc = (struct spdk_nvme_sgl_descriptor *)rdma_req->recv->buf + inline_segment->address;
    1791          15 :         for (i = 0; i < num_sgl_descriptors; i++) {
    1792             :                 /* The descriptors must be keyed data block descriptors with an address, not an offset. */
    1793          10 :                 if (spdk_unlikely(desc->generic.type != SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK ||
    1794             :                                   desc->keyed.subtype != SPDK_NVME_SGL_SUBTYPE_ADDRESS)) {
    1795           0 :                         rc = -EINVAL;
    1796           0 :                         goto err_exit;
    1797             :                 }
    1798             : 
    1799          10 :                 if (spdk_likely(!req->dif_enabled)) {
    1800           8 :                         rc = nvmf_rdma_fill_wr_sgl(device, rdma_req, current_wr, lengths[i]);
    1801             :                 } else {
    1802           2 :                         rc = nvmf_rdma_fill_wr_sgl_with_dif(device, rdma_req, current_wr,
    1803             :                                                             lengths[i], 0);
    1804             :                 }
    1805          10 :                 if (spdk_unlikely(rc != 0)) {
    1806           0 :                         rc = -ENOMEM;
    1807           0 :                         goto err_exit;
    1808             :                 }
    1809             : 
    1810          10 :                 req->length += desc->keyed.length;
    1811          10 :                 current_wr->wr.rdma.rkey = desc->keyed.key;
    1812          10 :                 current_wr->wr.rdma.remote_addr = desc->address;
    1813          10 :                 current_wr = current_wr->next;
    1814          10 :                 desc++;
    1815             :         }
    1816             : 
    1817             : #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL
    1818             :         /* Go back to the last descriptor in the list. */
    1819           5 :         desc--;
    1820           5 :         if ((device->attr.device_cap_flags & IBV_DEVICE_MEM_MGT_EXTENSIONS) != 0) {
    1821           0 :                 if (desc->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_INVALIDATE_KEY) {
    1822           0 :                         rdma_req->rsp.wr.opcode = IBV_WR_SEND_WITH_INV;
    1823           0 :                         rdma_req->rsp.wr.imm_data = desc->keyed.key;
    1824             :                 }
    1825             :         }
    1826             : #endif
    1827             : 
    1828           5 :         rdma_req->num_outstanding_data_wr = num_sgl_descriptors;
    1829             : 
    1830           5 :         return 0;
    1831             : 
    1832           0 : err_exit:
    1833           0 :         spdk_nvmf_request_free_buffers(req, &rgroup->group, &rtransport->transport);
    1834           0 :         nvmf_rdma_request_free_data(rdma_req, rtransport);
    1835           0 :         return rc;
    1836             : }
    1837             : 
    1838             : static int
    1839          25 : nvmf_rdma_request_parse_sgl(struct spdk_nvmf_rdma_transport *rtransport,
    1840             :                             struct spdk_nvmf_rdma_device *device,
    1841             :                             struct spdk_nvmf_rdma_request *rdma_req)
    1842             : {
    1843          25 :         struct spdk_nvmf_request                *req = &rdma_req->req;
    1844             :         struct spdk_nvme_cpl                    *rsp;
    1845             :         struct spdk_nvme_sgl_descriptor         *sgl;
    1846             :         int                                     rc;
    1847             :         uint32_t                                length;
    1848             : 
    1849          25 :         rsp = &req->rsp->nvme_cpl;
    1850          25 :         sgl = &req->cmd->nvme_cmd.dptr.sgl1;
    1851             : 
    1852          25 :         if (sgl->generic.type == SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK &&
    1853          17 :             (sgl->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_ADDRESS ||
    1854           0 :              sgl->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_INVALIDATE_KEY)) {
    1855             : 
    1856          17 :                 length = sgl->keyed.length;
    1857          17 :                 if (spdk_unlikely(length > rtransport->transport.opts.max_io_size)) {
    1858           1 :                         SPDK_ERRLOG("SGL length 0x%x exceeds max io size 0x%x\n",
    1859             :                                     length, rtransport->transport.opts.max_io_size);
    1860           1 :                         rsp->status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID;
    1861           1 :                         return -1;
    1862             :                 }
    1863             : #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL
    1864          16 :                 if ((device->attr.device_cap_flags & IBV_DEVICE_MEM_MGT_EXTENSIONS) != 0) {
    1865           0 :                         if (sgl->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_INVALIDATE_KEY) {
    1866           0 :                                 rdma_req->rsp.wr.opcode = IBV_WR_SEND_WITH_INV;
    1867           0 :                                 rdma_req->rsp.wr.imm_data = sgl->keyed.key;
    1868             :                         }
    1869             :                 }
    1870             : #endif
    1871             : 
    1872             :                 /* fill request length and populate iovs */
    1873          16 :                 req->length = length;
    1874             : 
    1875          16 :                 rc = nvmf_rdma_request_fill_iovs(rtransport, device, rdma_req);
    1876          16 :                 if (spdk_unlikely(rc < 0)) {
    1877           1 :                         if (rc == -EINVAL) {
    1878           0 :                                 SPDK_ERRLOG("SGL length exceeds the max I/O size\n");
    1879           0 :                                 rsp->status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID;
    1880           0 :                                 return -1;
    1881             :                         }
    1882             :                         /* No available buffers. Queue this request up. */
    1883           1 :                         SPDK_DEBUGLOG(rdma, "No available large data buffers. Queueing request %p\n", rdma_req);
    1884           1 :                         return 0;
    1885             :                 }
    1886             : 
    1887          15 :                 SPDK_DEBUGLOG(rdma, "Request %p took %d buffer/s from central pool\n", rdma_req,
    1888             :                               req->iovcnt);
    1889             : 
    1890          15 :                 return 0;
    1891           8 :         } else if (sgl->generic.type == SPDK_NVME_SGL_TYPE_DATA_BLOCK &&
    1892           3 :                    sgl->unkeyed.subtype == SPDK_NVME_SGL_SUBTYPE_OFFSET) {
    1893           3 :                 uint64_t offset = sgl->address;
    1894           3 :                 uint32_t max_len = rtransport->transport.opts.in_capsule_data_size;
    1895             : 
    1896           3 :                 SPDK_DEBUGLOG(nvmf, "In-capsule data: offset 0x%" PRIx64 ", length 0x%x\n",
    1897             :                               offset, sgl->unkeyed.length);
    1898             : 
    1899           3 :                 if (spdk_unlikely(offset > max_len)) {
    1900           0 :                         SPDK_ERRLOG("In-capsule offset 0x%" PRIx64 " exceeds capsule length 0x%x\n",
    1901             :                                     offset, max_len);
    1902           0 :                         rsp->status.sc = SPDK_NVME_SC_INVALID_SGL_OFFSET;
    1903           0 :                         return -1;
    1904             :                 }
    1905           3 :                 max_len -= (uint32_t)offset;
    1906             : 
    1907           3 :                 if (spdk_unlikely(sgl->unkeyed.length > max_len)) {
    1908           2 :                         SPDK_ERRLOG("In-capsule data length 0x%x exceeds capsule length 0x%x\n",
    1909             :                                     sgl->unkeyed.length, max_len);
    1910           2 :                         rsp->status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID;
    1911           2 :                         return -1;
    1912             :                 }
    1913             : 
    1914           1 :                 rdma_req->num_outstanding_data_wr = 0;
    1915           1 :                 req->data_from_pool = false;
    1916           1 :                 req->length = sgl->unkeyed.length;
    1917             : 
    1918           1 :                 req->iov[0].iov_base = rdma_req->recv->buf + offset;
    1919           1 :                 req->iov[0].iov_len = req->length;
    1920           1 :                 req->iovcnt = 1;
    1921             : 
    1922           1 :                 return 0;
    1923           5 :         } else if (sgl->generic.type == SPDK_NVME_SGL_TYPE_LAST_SEGMENT &&
    1924           5 :                    sgl->unkeyed.subtype == SPDK_NVME_SGL_SUBTYPE_OFFSET) {
    1925             : 
    1926           5 :                 rc = nvmf_rdma_request_fill_iovs_multi_sgl(rtransport, device, rdma_req);
    1927           5 :                 if (spdk_unlikely(rc == -ENOMEM)) {
    1928           0 :                         SPDK_DEBUGLOG(rdma, "No available large data buffers. Queueing request %p\n", rdma_req);
    1929           0 :                         return 0;
    1930           5 :                 } else if (spdk_unlikely(rc == -EINVAL)) {
    1931           0 :                         SPDK_ERRLOG("Multi SGL element request length exceeds the max I/O size\n");
    1932           0 :                         rsp->status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID;
    1933           0 :                         return -1;
    1934             :                 }
    1935             : 
    1936           5 :                 SPDK_DEBUGLOG(rdma, "Request %p took %d buffer/s from central pool\n", rdma_req,
    1937             :                               req->iovcnt);
    1938             : 
    1939           5 :                 return 0;
    1940             :         }
    1941             : 
    1942           0 :         SPDK_ERRLOG("Invalid NVMf I/O Command SGL:  Type 0x%x, Subtype 0x%x\n",
    1943             :                     sgl->generic.type, sgl->generic.subtype);
    1944           0 :         rsp->status.sc = SPDK_NVME_SC_SGL_DESCRIPTOR_TYPE_INVALID;
    1945           0 :         return -1;
    1946             : }
    1947             : 
    1948             : static void
    1949           6 : _nvmf_rdma_request_free(struct spdk_nvmf_rdma_request *rdma_req,
    1950             :                         struct spdk_nvmf_rdma_transport *rtransport)
    1951             : {
    1952             :         struct spdk_nvmf_rdma_qpair             *rqpair;
    1953             :         struct spdk_nvmf_rdma_poll_group        *rgroup;
    1954             : 
    1955           6 :         rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair);
    1956           6 :         if (rdma_req->req.data_from_pool) {
    1957           5 :                 rgroup = rqpair->poller->group;
    1958             : 
    1959           5 :                 spdk_nvmf_request_free_buffers(&rdma_req->req, &rgroup->group, &rtransport->transport);
    1960             :         }
    1961           6 :         if (rdma_req->req.stripped_data) {
    1962           0 :                 nvmf_request_free_stripped_buffers(&rdma_req->req,
    1963           0 :                                                    &rqpair->poller->group->group,
    1964             :                                                    &rtransport->transport);
    1965             :         }
    1966           6 :         nvmf_rdma_request_free_data(rdma_req, rtransport);
    1967           6 :         rdma_req->req.length = 0;
    1968           6 :         rdma_req->req.iovcnt = 0;
    1969           6 :         rdma_req->offset = 0;
    1970           6 :         rdma_req->req.dif_enabled = false;
    1971           6 :         rdma_req->fused_failed = false;
    1972           6 :         rdma_req->transfer_wr = NULL;
    1973           6 :         if (rdma_req->fused_pair) {
    1974             :                 /* This req was part of a valid fused pair, but failed before it got to
    1975             :                  * READ_TO_EXECUTE state.  This means we need to fail the other request
    1976             :                  * in the pair, because it is no longer part of a valid pair.  If the pair
    1977             :                  * already reached READY_TO_EXECUTE state, we need to kick it.
    1978             :                  */
    1979           0 :                 rdma_req->fused_pair->fused_failed = true;
    1980           0 :                 if (rdma_req->fused_pair->state == RDMA_REQUEST_STATE_READY_TO_EXECUTE) {
    1981           0 :                         nvmf_rdma_request_process(rtransport, rdma_req->fused_pair);
    1982             :                 }
    1983           0 :                 rdma_req->fused_pair = NULL;
    1984             :         }
    1985           6 :         memset(&rdma_req->req.dif, 0, sizeof(rdma_req->req.dif));
    1986           6 :         rqpair->qd--;
    1987             : 
    1988           6 :         STAILQ_INSERT_HEAD(&rqpair->resources->free_queue, rdma_req, state_link);
    1989           6 :         rqpair->qpair.queue_depth--;
    1990           6 :         rdma_req->state = RDMA_REQUEST_STATE_FREE;
    1991           6 : }
    1992             : 
    1993             : static void
    1994           6 : nvmf_rdma_check_fused_ordering(struct spdk_nvmf_rdma_transport *rtransport,
    1995             :                                struct spdk_nvmf_rdma_qpair *rqpair,
    1996             :                                struct spdk_nvmf_rdma_request *rdma_req)
    1997             : {
    1998             :         enum spdk_nvme_cmd_fuse last, next;
    1999             : 
    2000           6 :         last = rqpair->fused_first ? rqpair->fused_first->req.cmd->nvme_cmd.fuse : SPDK_NVME_CMD_FUSE_NONE;
    2001           6 :         next = rdma_req->req.cmd->nvme_cmd.fuse;
    2002             : 
    2003           6 :         assert(last != SPDK_NVME_CMD_FUSE_SECOND);
    2004             : 
    2005           6 :         if (spdk_likely(last == SPDK_NVME_CMD_FUSE_NONE && next == SPDK_NVME_CMD_FUSE_NONE)) {
    2006           6 :                 return;
    2007             :         }
    2008             : 
    2009           0 :         if (last == SPDK_NVME_CMD_FUSE_FIRST) {
    2010           0 :                 if (next == SPDK_NVME_CMD_FUSE_SECOND) {
    2011             :                         /* This is a valid pair of fused commands.  Point them at each other
    2012             :                          * so they can be submitted consecutively once ready to be executed.
    2013             :                          */
    2014           0 :                         rqpair->fused_first->fused_pair = rdma_req;
    2015           0 :                         rdma_req->fused_pair = rqpair->fused_first;
    2016           0 :                         rqpair->fused_first = NULL;
    2017           0 :                         return;
    2018             :                 } else {
    2019             :                         /* Mark the last req as failed since it wasn't followed by a SECOND. */
    2020           0 :                         rqpair->fused_first->fused_failed = true;
    2021             : 
    2022             :                         /* If the last req is in READY_TO_EXECUTE state, then call
    2023             :                          * nvmf_rdma_request_process(), otherwise nothing else will kick it.
    2024             :                          */
    2025           0 :                         if (rqpair->fused_first->state == RDMA_REQUEST_STATE_READY_TO_EXECUTE) {
    2026           0 :                                 nvmf_rdma_request_process(rtransport, rqpair->fused_first);
    2027             :                         }
    2028             : 
    2029           0 :                         rqpair->fused_first = NULL;
    2030             :                 }
    2031             :         }
    2032             : 
    2033           0 :         if (next == SPDK_NVME_CMD_FUSE_FIRST) {
    2034             :                 /* Set rqpair->fused_first here so that we know to check that the next request
    2035             :                  * is a SECOND (and to fail this one if it isn't).
    2036             :                  */
    2037           0 :                 rqpair->fused_first = rdma_req;
    2038           0 :         } else if (next == SPDK_NVME_CMD_FUSE_SECOND) {
    2039             :                 /* Mark this req failed since it ia SECOND and the last one was not a FIRST. */
    2040           0 :                 rdma_req->fused_failed = true;
    2041             :         }
    2042             : }
    2043             : 
    2044             : bool
    2045          23 : nvmf_rdma_request_process(struct spdk_nvmf_rdma_transport *rtransport,
    2046             :                           struct spdk_nvmf_rdma_request *rdma_req)
    2047             : {
    2048             :         struct spdk_nvmf_rdma_qpair     *rqpair;
    2049             :         struct spdk_nvmf_rdma_device    *device;
    2050             :         struct spdk_nvmf_rdma_poll_group *rgroup;
    2051          23 :         struct spdk_nvme_cpl            *rsp = &rdma_req->req.rsp->nvme_cpl;
    2052             :         int                             rc;
    2053             :         struct spdk_nvmf_rdma_recv      *rdma_recv;
    2054             :         enum spdk_nvmf_rdma_request_state prev_state;
    2055          23 :         bool                            progress = false;
    2056          23 :         int                             data_posted;
    2057             :         uint32_t                        num_blocks, num_rdma_reads_available, qdepth;
    2058             : 
    2059          23 :         rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair);
    2060          23 :         device = rqpair->device;
    2061          23 :         rgroup = rqpair->poller->group;
    2062             : 
    2063          23 :         assert(rdma_req->state != RDMA_REQUEST_STATE_FREE);
    2064             : 
    2065             :         /* If the queue pair is in an error state, force the request to the completed state
    2066             :          * to release resources. */
    2067          23 :         if (spdk_unlikely(rqpair->ibv_in_error_state || !spdk_nvmf_qpair_is_active(&rqpair->qpair))) {
    2068           0 :                 switch (rdma_req->state) {
    2069           0 :                 case RDMA_REQUEST_STATE_NEED_BUFFER:
    2070           0 :                         STAILQ_REMOVE(&rgroup->group.pending_buf_queue, &rdma_req->req, spdk_nvmf_request, buf_link);
    2071           0 :                         break;
    2072           0 :                 case RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING:
    2073           0 :                         STAILQ_REMOVE(&rqpair->pending_rdma_read_queue, rdma_req, spdk_nvmf_rdma_request, state_link);
    2074           0 :                         break;
    2075           0 :                 case RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING:
    2076           0 :                         STAILQ_REMOVE(&rqpair->pending_rdma_write_queue, rdma_req, spdk_nvmf_rdma_request, state_link);
    2077           0 :                         break;
    2078           0 :                 case RDMA_REQUEST_STATE_READY_TO_COMPLETE_PENDING:
    2079           0 :                         STAILQ_REMOVE(&rqpair->pending_rdma_send_queue, rdma_req, spdk_nvmf_rdma_request, state_link);
    2080           0 :                         break;
    2081           0 :                 default:
    2082           0 :                         break;
    2083             :                 }
    2084           0 :                 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED;
    2085             :         }
    2086             : 
    2087             :         /* The loop here is to allow for several back-to-back state changes. */
    2088             :         do {
    2089          66 :                 prev_state = rdma_req->state;
    2090             : 
    2091          66 :                 SPDK_DEBUGLOG(rdma, "Request %p entering state %d\n", rdma_req, prev_state);
    2092             : 
    2093          66 :                 switch (rdma_req->state) {
    2094           6 :                 case RDMA_REQUEST_STATE_FREE:
    2095             :                         /* Some external code must kick a request into RDMA_REQUEST_STATE_NEW
    2096             :                          * to escape this state. */
    2097           6 :                         break;
    2098           6 :                 case RDMA_REQUEST_STATE_NEW:
    2099           6 :                         spdk_trace_record(TRACE_RDMA_REQUEST_STATE_NEW, 0, 0,
    2100             :                                           (uintptr_t)rdma_req, (uintptr_t)rqpair, rqpair->qpair.queue_depth);
    2101           6 :                         rdma_recv = rdma_req->recv;
    2102             : 
    2103             :                         /* The first element of the SGL is the NVMe command */
    2104           6 :                         rdma_req->req.cmd = (union nvmf_h2c_msg *)rdma_recv->sgl[0].addr;
    2105           6 :                         memset(rdma_req->req.rsp, 0, sizeof(*rdma_req->req.rsp));
    2106           6 :                         rdma_req->transfer_wr = &rdma_req->data.wr;
    2107             : 
    2108           6 :                         if (spdk_unlikely(rqpair->ibv_in_error_state || !spdk_nvmf_qpair_is_active(&rqpair->qpair))) {
    2109           0 :                                 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED;
    2110           0 :                                 break;
    2111             :                         }
    2112             : 
    2113           6 :                         if (spdk_unlikely(spdk_nvmf_request_get_dif_ctx(&rdma_req->req, &rdma_req->req.dif.dif_ctx))) {
    2114           0 :                                 rdma_req->req.dif_enabled = true;
    2115             :                         }
    2116             : 
    2117           6 :                         nvmf_rdma_check_fused_ordering(rtransport, rqpair, rdma_req);
    2118             : 
    2119             : #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL
    2120           6 :                         rdma_req->rsp.wr.opcode = IBV_WR_SEND;
    2121           6 :                         rdma_req->rsp.wr.imm_data = 0;
    2122             : #endif
    2123             : 
    2124             :                         /* The next state transition depends on the data transfer needs of this request. */
    2125           6 :                         rdma_req->req.xfer = spdk_nvmf_req_get_xfer(&rdma_req->req);
    2126             : 
    2127           6 :                         if (spdk_unlikely(rdma_req->req.xfer == SPDK_NVME_DATA_BIDIRECTIONAL)) {
    2128           1 :                                 rsp->status.sct = SPDK_NVME_SCT_GENERIC;
    2129           1 :                                 rsp->status.sc = SPDK_NVME_SC_INVALID_OPCODE;
    2130           1 :                                 STAILQ_INSERT_TAIL(&rqpair->pending_rdma_send_queue, rdma_req, state_link);
    2131           1 :                                 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE_PENDING;
    2132           1 :                                 SPDK_DEBUGLOG(rdma, "Request %p: invalid xfer type (BIDIRECTIONAL)\n", rdma_req);
    2133           1 :                                 break;
    2134             :                         }
    2135             : 
    2136             :                         /* If no data to transfer, ready to execute. */
    2137           5 :                         if (rdma_req->req.xfer == SPDK_NVME_DATA_NONE) {
    2138           0 :                                 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_EXECUTE;
    2139           0 :                                 break;
    2140             :                         }
    2141             : 
    2142           5 :                         rdma_req->state = RDMA_REQUEST_STATE_NEED_BUFFER;
    2143           5 :                         STAILQ_INSERT_TAIL(&rgroup->group.pending_buf_queue, &rdma_req->req, buf_link);
    2144           5 :                         break;
    2145           5 :                 case RDMA_REQUEST_STATE_NEED_BUFFER:
    2146           5 :                         spdk_trace_record(TRACE_RDMA_REQUEST_STATE_NEED_BUFFER, 0, 0,
    2147             :                                           (uintptr_t)rdma_req, (uintptr_t)rqpair);
    2148             : 
    2149           5 :                         assert(rdma_req->req.xfer != SPDK_NVME_DATA_NONE);
    2150             : 
    2151           5 :                         if (&rdma_req->req != STAILQ_FIRST(&rgroup->group.pending_buf_queue)) {
    2152             :                                 /* This request needs to wait in line to obtain a buffer */
    2153           0 :                                 break;
    2154             :                         }
    2155             : 
    2156             :                         /* Try to get a data buffer */
    2157           5 :                         rc = nvmf_rdma_request_parse_sgl(rtransport, device, rdma_req);
    2158           5 :                         if (spdk_unlikely(rc < 0)) {
    2159           0 :                                 STAILQ_REMOVE_HEAD(&rgroup->group.pending_buf_queue, buf_link);
    2160           0 :                                 STAILQ_INSERT_TAIL(&rqpair->pending_rdma_send_queue, rdma_req, state_link);
    2161           0 :                                 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE_PENDING;
    2162           0 :                                 break;
    2163             :                         }
    2164             : 
    2165           5 :                         if (rdma_req->req.iovcnt == 0) {
    2166             :                                 /* No buffers available. */
    2167           0 :                                 rgroup->stat.pending_data_buffer++;
    2168           0 :                                 break;
    2169             :                         }
    2170             : 
    2171           5 :                         STAILQ_REMOVE_HEAD(&rgroup->group.pending_buf_queue, buf_link);
    2172             : 
    2173             :                         /* If data is transferring from host to controller and the data didn't
    2174             :                          * arrive using in capsule data, we need to do a transfer from the host.
    2175             :                          */
    2176           5 :                         if (rdma_req->req.xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER &&
    2177             :                             rdma_req->req.data_from_pool) {
    2178           4 :                                 STAILQ_INSERT_TAIL(&rqpair->pending_rdma_read_queue, rdma_req, state_link);
    2179           4 :                                 rdma_req->state = RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING;
    2180           4 :                                 break;
    2181             :                         }
    2182             : 
    2183           1 :                         rdma_req->state = RDMA_REQUEST_STATE_READY_TO_EXECUTE;
    2184           1 :                         break;
    2185           4 :                 case RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING:
    2186           4 :                         spdk_trace_record(TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING, 0, 0,
    2187             :                                           (uintptr_t)rdma_req, (uintptr_t)rqpair);
    2188             : 
    2189           4 :                         if (rdma_req != STAILQ_FIRST(&rqpair->pending_rdma_read_queue)) {
    2190             :                                 /* This request needs to wait in line to perform RDMA */
    2191           0 :                                 break;
    2192             :                         }
    2193           4 :                         assert(rqpair->max_send_depth >= rqpair->current_send_depth);
    2194           4 :                         qdepth = rqpair->max_send_depth - rqpair->current_send_depth;
    2195           4 :                         assert(rqpair->max_read_depth >= rqpair->current_read_depth);
    2196           4 :                         num_rdma_reads_available = rqpair->max_read_depth - rqpair->current_read_depth;
    2197           4 :                         if (rdma_req->num_outstanding_data_wr > qdepth ||
    2198           4 :                             rdma_req->num_outstanding_data_wr > num_rdma_reads_available) {
    2199           0 :                                 if (num_rdma_reads_available && qdepth) {
    2200             :                                         /* Send as much as we can */
    2201           0 :                                         request_prepare_transfer_in_part(&rdma_req->req, spdk_min(num_rdma_reads_available, qdepth));
    2202             :                                 } else {
    2203             :                                         /* We can only have so many WRs outstanding. we have to wait until some finish. */
    2204           0 :                                         rqpair->poller->stat.pending_rdma_read++;
    2205           0 :                                         break;
    2206             :                                 }
    2207             :                         }
    2208             : 
    2209             :                         /* We have already verified that this request is the head of the queue. */
    2210           4 :                         if (rdma_req->num_remaining_data_wr == 0) {
    2211           4 :                                 STAILQ_REMOVE_HEAD(&rqpair->pending_rdma_read_queue, state_link);
    2212             :                         }
    2213             : 
    2214           4 :                         rc = request_transfer_in(&rdma_req->req);
    2215           4 :                         if (spdk_likely(rc == 0)) {
    2216           4 :                                 rdma_req->state = RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER;
    2217             :                         } else {
    2218           0 :                                 rsp->status.sc = SPDK_NVME_SC_INTERNAL_DEVICE_ERROR;
    2219           0 :                                 STAILQ_INSERT_TAIL(&rqpair->pending_rdma_send_queue, rdma_req, state_link);
    2220           0 :                                 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE_PENDING;
    2221             :                         }
    2222           4 :                         break;
    2223           4 :                 case RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER:
    2224           4 :                         spdk_trace_record(TRACE_RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER, 0, 0,
    2225             :                                           (uintptr_t)rdma_req, (uintptr_t)rqpair);
    2226             :                         /* Some external code must kick a request into RDMA_REQUEST_STATE_READY_TO_EXECUTE
    2227             :                          * to escape this state. */
    2228           4 :                         break;
    2229           5 :                 case RDMA_REQUEST_STATE_READY_TO_EXECUTE:
    2230           5 :                         spdk_trace_record(TRACE_RDMA_REQUEST_STATE_READY_TO_EXECUTE, 0, 0,
    2231             :                                           (uintptr_t)rdma_req, (uintptr_t)rqpair);
    2232             : 
    2233           5 :                         if (spdk_unlikely(rdma_req->req.dif_enabled)) {
    2234           0 :                                 if (rdma_req->req.xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER) {
    2235             :                                         /* generate DIF for write operation */
    2236           0 :                                         num_blocks = SPDK_CEIL_DIV(rdma_req->req.dif.elba_length, rdma_req->req.dif.dif_ctx.block_size);
    2237           0 :                                         assert(num_blocks > 0);
    2238             : 
    2239           0 :                                         rc = spdk_dif_generate(rdma_req->req.iov, rdma_req->req.iovcnt,
    2240           0 :                                                                num_blocks, &rdma_req->req.dif.dif_ctx);
    2241           0 :                                         if (rc != 0) {
    2242           0 :                                                 SPDK_ERRLOG("DIF generation failed\n");
    2243           0 :                                                 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED;
    2244           0 :                                                 spdk_nvmf_qpair_disconnect(&rqpair->qpair);
    2245           0 :                                                 break;
    2246             :                                         }
    2247             :                                 }
    2248             : 
    2249           0 :                                 assert(rdma_req->req.dif.elba_length >= rdma_req->req.length);
    2250             :                                 /* set extended length before IO operation */
    2251           0 :                                 rdma_req->req.length = rdma_req->req.dif.elba_length;
    2252             :                         }
    2253             : 
    2254           5 :                         if (rdma_req->req.cmd->nvme_cmd.fuse != SPDK_NVME_CMD_FUSE_NONE) {
    2255           0 :                                 if (rdma_req->fused_failed) {
    2256             :                                         /* This request failed FUSED semantics.  Fail it immediately, without
    2257             :                                          * even sending it to the target layer.
    2258             :                                          */
    2259           0 :                                         rsp->status.sct = SPDK_NVME_SCT_GENERIC;
    2260           0 :                                         rsp->status.sc = SPDK_NVME_SC_ABORTED_MISSING_FUSED;
    2261           0 :                                         STAILQ_INSERT_TAIL(&rqpair->pending_rdma_send_queue, rdma_req, state_link);
    2262           0 :                                         rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE_PENDING;
    2263           0 :                                         break;
    2264             :                                 }
    2265             : 
    2266           0 :                                 if (rdma_req->fused_pair == NULL ||
    2267           0 :                                     rdma_req->fused_pair->state != RDMA_REQUEST_STATE_READY_TO_EXECUTE) {
    2268             :                                         /* This request is ready to execute, but either we don't know yet if it's
    2269             :                                          * valid - i.e. this is a FIRST but we haven't received the next
    2270             :                                          * request yet or the other request of this fused pair isn't ready to
    2271             :                                          * execute.  So break here and this request will get processed later either
    2272             :                                          * when the other request is ready or we find that this request isn't valid.
    2273             :                                          */
    2274             :                                         break;
    2275             :                                 }
    2276             :                         }
    2277             : 
    2278             :                         /* If we get to this point, and this request is a fused command, we know that
    2279             :                          * it is part of valid sequence (FIRST followed by a SECOND) and that both
    2280             :                          * requests are READY_TO_EXECUTE. So call spdk_nvmf_request_exec() both on this
    2281             :                          * request, and the other request of the fused pair, in the correct order.
    2282             :                          * Also clear the ->fused_pair pointers on both requests, since after this point
    2283             :                          * we no longer need to maintain the relationship between these two requests.
    2284             :                          */
    2285           5 :                         if (rdma_req->req.cmd->nvme_cmd.fuse == SPDK_NVME_CMD_FUSE_SECOND) {
    2286           0 :                                 assert(rdma_req->fused_pair != NULL);
    2287           0 :                                 assert(rdma_req->fused_pair->fused_pair != NULL);
    2288           0 :                                 rdma_req->fused_pair->state = RDMA_REQUEST_STATE_EXECUTING;
    2289           0 :                                 spdk_nvmf_request_exec(&rdma_req->fused_pair->req);
    2290           0 :                                 rdma_req->fused_pair->fused_pair = NULL;
    2291           0 :                                 rdma_req->fused_pair = NULL;
    2292             :                         }
    2293           5 :                         rdma_req->state = RDMA_REQUEST_STATE_EXECUTING;
    2294           5 :                         spdk_nvmf_request_exec(&rdma_req->req);
    2295           5 :                         if (rdma_req->req.cmd->nvme_cmd.fuse == SPDK_NVME_CMD_FUSE_FIRST) {
    2296           0 :                                 assert(rdma_req->fused_pair != NULL);
    2297           0 :                                 assert(rdma_req->fused_pair->fused_pair != NULL);
    2298           0 :                                 rdma_req->fused_pair->state = RDMA_REQUEST_STATE_EXECUTING;
    2299           0 :                                 spdk_nvmf_request_exec(&rdma_req->fused_pair->req);
    2300           0 :                                 rdma_req->fused_pair->fused_pair = NULL;
    2301           0 :                                 rdma_req->fused_pair = NULL;
    2302             :                         }
    2303           5 :                         break;
    2304           5 :                 case RDMA_REQUEST_STATE_EXECUTING:
    2305           5 :                         spdk_trace_record(TRACE_RDMA_REQUEST_STATE_EXECUTING, 0, 0,
    2306             :                                           (uintptr_t)rdma_req, (uintptr_t)rqpair);
    2307             :                         /* Some external code must kick a request into RDMA_REQUEST_STATE_EXECUTED
    2308             :                          * to escape this state. */
    2309           5 :                         break;
    2310           5 :                 case RDMA_REQUEST_STATE_EXECUTED:
    2311           5 :                         spdk_trace_record(TRACE_RDMA_REQUEST_STATE_EXECUTED, 0, 0,
    2312             :                                           (uintptr_t)rdma_req, (uintptr_t)rqpair);
    2313           5 :                         if (rsp->status.sc == SPDK_NVME_SC_SUCCESS &&
    2314           5 :                             rdma_req->req.xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) {
    2315           1 :                                 STAILQ_INSERT_TAIL(&rqpair->pending_rdma_write_queue, rdma_req, state_link);
    2316           1 :                                 rdma_req->state = RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING;
    2317             :                         } else {
    2318           4 :                                 STAILQ_INSERT_TAIL(&rqpair->pending_rdma_send_queue, rdma_req, state_link);
    2319           4 :                                 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE_PENDING;
    2320             :                         }
    2321           5 :                         if (spdk_unlikely(rdma_req->req.dif_enabled)) {
    2322             :                                 /* restore the original length */
    2323           0 :                                 rdma_req->req.length = rdma_req->req.dif.orig_length;
    2324             : 
    2325           0 :                                 if (rdma_req->req.xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) {
    2326           0 :                                         struct spdk_dif_error error_blk;
    2327             : 
    2328           0 :                                         num_blocks = SPDK_CEIL_DIV(rdma_req->req.dif.elba_length, rdma_req->req.dif.dif_ctx.block_size);
    2329           0 :                                         if (!rdma_req->req.stripped_data) {
    2330           0 :                                                 rc = spdk_dif_verify(rdma_req->req.iov, rdma_req->req.iovcnt, num_blocks,
    2331           0 :                                                                      &rdma_req->req.dif.dif_ctx, &error_blk);
    2332             :                                         } else {
    2333           0 :                                                 rc = spdk_dif_verify_copy(rdma_req->req.stripped_data->iov,
    2334           0 :                                                                           rdma_req->req.stripped_data->iovcnt,
    2335           0 :                                                                           rdma_req->req.iov, rdma_req->req.iovcnt, num_blocks,
    2336           0 :                                                                           &rdma_req->req.dif.dif_ctx, &error_blk);
    2337             :                                         }
    2338           0 :                                         if (rc) {
    2339           0 :                                                 struct spdk_nvme_cpl *rsp = &rdma_req->req.rsp->nvme_cpl;
    2340             : 
    2341           0 :                                                 SPDK_ERRLOG("DIF error detected. type=%d, offset=%" PRIu32 "\n", error_blk.err_type,
    2342             :                                                             error_blk.err_offset);
    2343           0 :                                                 rsp->status.sct = SPDK_NVME_SCT_MEDIA_ERROR;
    2344           0 :                                                 rsp->status.sc = nvmf_rdma_dif_error_to_compl_status(error_blk.err_type);
    2345           0 :                                                 STAILQ_REMOVE(&rqpair->pending_rdma_write_queue, rdma_req, spdk_nvmf_rdma_request, state_link);
    2346           0 :                                                 STAILQ_INSERT_TAIL(&rqpair->pending_rdma_send_queue, rdma_req, state_link);
    2347           0 :                                                 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE_PENDING;
    2348             :                                         }
    2349             :                                 }
    2350             :                         }
    2351           5 :                         break;
    2352           1 :                 case RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING:
    2353           1 :                         spdk_trace_record(TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING, 0, 0,
    2354             :                                           (uintptr_t)rdma_req, (uintptr_t)rqpair);
    2355             : 
    2356           1 :                         if (rdma_req != STAILQ_FIRST(&rqpair->pending_rdma_write_queue)) {
    2357             :                                 /* This request needs to wait in line to perform RDMA */
    2358           0 :                                 break;
    2359             :                         }
    2360           1 :                         if ((rqpair->current_send_depth + rdma_req->num_outstanding_data_wr + 1) >
    2361           1 :                             rqpair->max_send_depth) {
    2362             :                                 /* We can only have so many WRs outstanding. we have to wait until some finish.
    2363             :                                  * +1 since each request has an additional wr in the resp. */
    2364           0 :                                 rqpair->poller->stat.pending_rdma_write++;
    2365           0 :                                 break;
    2366             :                         }
    2367             : 
    2368             :                         /* We have already verified that this request is the head of the queue. */
    2369           1 :                         STAILQ_REMOVE_HEAD(&rqpair->pending_rdma_write_queue, state_link);
    2370             : 
    2371             :                         /* The data transfer will be kicked off from
    2372             :                          * RDMA_REQUEST_STATE_READY_TO_COMPLETE state.
    2373             :                          * We verified that data + response fit into send queue, so we can go to the next state directly
    2374             :                          */
    2375           1 :                         rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE;
    2376           1 :                         break;
    2377           7 :                 case RDMA_REQUEST_STATE_READY_TO_COMPLETE_PENDING:
    2378           7 :                         spdk_trace_record(TRACE_RDMA_REQUEST_STATE_READY_TO_COMPLETE_PENDING, 0, 0,
    2379             :                                           (uintptr_t)rdma_req, (uintptr_t)rqpair);
    2380             : 
    2381           7 :                         if (rdma_req != STAILQ_FIRST(&rqpair->pending_rdma_send_queue)) {
    2382             :                                 /* This request needs to wait in line to send the completion */
    2383           0 :                                 break;
    2384             :                         }
    2385             : 
    2386           7 :                         assert(rqpair->current_send_depth <= rqpair->max_send_depth);
    2387           7 :                         if (rqpair->current_send_depth == rqpair->max_send_depth) {
    2388             :                                 /* We can only have so many WRs outstanding. we have to wait until some finish */
    2389           2 :                                 rqpair->poller->stat.pending_rdma_send++;
    2390           2 :                                 break;
    2391             :                         }
    2392             : 
    2393             :                         /* We have already verified that this request is the head of the queue. */
    2394           5 :                         STAILQ_REMOVE_HEAD(&rqpair->pending_rdma_send_queue, state_link);
    2395             : 
    2396             :                         /* The response sending will be kicked off from
    2397             :                          * RDMA_REQUEST_STATE_READY_TO_COMPLETE state.
    2398             :                          */
    2399           5 :                         rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE;
    2400           5 :                         break;
    2401           6 :                 case RDMA_REQUEST_STATE_READY_TO_COMPLETE:
    2402           6 :                         spdk_trace_record(TRACE_RDMA_REQUEST_STATE_READY_TO_COMPLETE, 0, 0,
    2403             :                                           (uintptr_t)rdma_req, (uintptr_t)rqpair);
    2404           6 :                         rc = request_transfer_out(&rdma_req->req, &data_posted);
    2405           6 :                         assert(rc == 0); /* No good way to handle this currently */
    2406           6 :                         if (spdk_unlikely(rc)) {
    2407           0 :                                 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED;
    2408             :                         } else {
    2409           6 :                                 rdma_req->state = data_posted ? RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST :
    2410             :                                                   RDMA_REQUEST_STATE_COMPLETING;
    2411             :                         }
    2412           6 :                         break;
    2413           1 :                 case RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST:
    2414           1 :                         spdk_trace_record(TRACE_RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST, 0, 0,
    2415             :                                           (uintptr_t)rdma_req, (uintptr_t)rqpair);
    2416             :                         /* Some external code must kick a request into RDMA_REQUEST_STATE_COMPLETED
    2417             :                          * to escape this state. */
    2418           1 :                         break;
    2419           5 :                 case RDMA_REQUEST_STATE_COMPLETING:
    2420           5 :                         spdk_trace_record(TRACE_RDMA_REQUEST_STATE_COMPLETING, 0, 0,
    2421             :                                           (uintptr_t)rdma_req, (uintptr_t)rqpair);
    2422             :                         /* Some external code must kick a request into RDMA_REQUEST_STATE_COMPLETED
    2423             :                          * to escape this state. */
    2424           5 :                         break;
    2425           6 :                 case RDMA_REQUEST_STATE_COMPLETED:
    2426           6 :                         spdk_trace_record(TRACE_RDMA_REQUEST_STATE_COMPLETED, 0, 0,
    2427             :                                           (uintptr_t)rdma_req, (uintptr_t)rqpair, rqpair->qpair.queue_depth);
    2428             : 
    2429           6 :                         rqpair->poller->stat.request_latency += spdk_get_ticks() - rdma_req->receive_tsc;
    2430           6 :                         _nvmf_rdma_request_free(rdma_req, rtransport);
    2431           6 :                         break;
    2432           0 :                 case RDMA_REQUEST_NUM_STATES:
    2433             :                 default:
    2434           0 :                         assert(0);
    2435             :                         break;
    2436             :                 }
    2437             : 
    2438          66 :                 if (rdma_req->state != prev_state) {
    2439          43 :                         progress = true;
    2440             :                 }
    2441          66 :         } while (rdma_req->state != prev_state);
    2442             : 
    2443          23 :         return progress;
    2444             : }
    2445             : 
    2446             : /* Public API callbacks begin here */
    2447             : 
    2448             : #define SPDK_NVMF_RDMA_DEFAULT_MAX_QUEUE_DEPTH 128
    2449             : #define SPDK_NVMF_RDMA_DEFAULT_AQ_DEPTH 128
    2450             : #define SPDK_NVMF_RDMA_DEFAULT_SRQ_DEPTH 4096
    2451             : #define SPDK_NVMF_RDMA_DEFAULT_MAX_QPAIRS_PER_CTRLR 128
    2452             : #define SPDK_NVMF_RDMA_DEFAULT_IN_CAPSULE_DATA_SIZE 4096
    2453             : #define SPDK_NVMF_RDMA_DEFAULT_MAX_IO_SIZE 131072
    2454             : #define SPDK_NVMF_RDMA_MIN_IO_BUFFER_SIZE (SPDK_NVMF_RDMA_DEFAULT_MAX_IO_SIZE / SPDK_NVMF_MAX_SGL_ENTRIES)
    2455             : #define SPDK_NVMF_RDMA_DEFAULT_NUM_SHARED_BUFFERS 4095
    2456             : #define SPDK_NVMF_RDMA_DEFAULT_BUFFER_CACHE_SIZE UINT32_MAX
    2457             : #define SPDK_NVMF_RDMA_DEFAULT_NO_SRQ false
    2458             : #define SPDK_NVMF_RDMA_DIF_INSERT_OR_STRIP false
    2459             : #define SPDK_NVMF_RDMA_ACCEPTOR_BACKLOG 100
    2460             : #define SPDK_NVMF_RDMA_DEFAULT_ABORT_TIMEOUT_SEC 1
    2461             : #define SPDK_NVMF_RDMA_DEFAULT_NO_WR_BATCHING false
    2462             : #define SPDK_NVMF_RDMA_DEFAULT_DATA_WR_POOL_SIZE 4095
    2463             : 
    2464             : static void
    2465           1 : nvmf_rdma_opts_init(struct spdk_nvmf_transport_opts *opts)
    2466             : {
    2467           1 :         opts->max_queue_depth =              SPDK_NVMF_RDMA_DEFAULT_MAX_QUEUE_DEPTH;
    2468           1 :         opts->max_qpairs_per_ctrlr = SPDK_NVMF_RDMA_DEFAULT_MAX_QPAIRS_PER_CTRLR;
    2469           1 :         opts->in_capsule_data_size = SPDK_NVMF_RDMA_DEFAULT_IN_CAPSULE_DATA_SIZE;
    2470           1 :         opts->max_io_size =          SPDK_NVMF_RDMA_DEFAULT_MAX_IO_SIZE;
    2471           1 :         opts->io_unit_size =         SPDK_NVMF_RDMA_MIN_IO_BUFFER_SIZE;
    2472           1 :         opts->max_aq_depth =         SPDK_NVMF_RDMA_DEFAULT_AQ_DEPTH;
    2473           1 :         opts->num_shared_buffers =   SPDK_NVMF_RDMA_DEFAULT_NUM_SHARED_BUFFERS;
    2474           1 :         opts->buf_cache_size =               SPDK_NVMF_RDMA_DEFAULT_BUFFER_CACHE_SIZE;
    2475           1 :         opts->dif_insert_or_strip =  SPDK_NVMF_RDMA_DIF_INSERT_OR_STRIP;
    2476           1 :         opts->abort_timeout_sec =    SPDK_NVMF_RDMA_DEFAULT_ABORT_TIMEOUT_SEC;
    2477           1 :         opts->transport_specific =      NULL;
    2478           1 :         opts->data_wr_pool_size      =       SPDK_NVMF_RDMA_DEFAULT_DATA_WR_POOL_SIZE;
    2479           1 : }
    2480             : 
    2481             : static int nvmf_rdma_destroy(struct spdk_nvmf_transport *transport,
    2482             :                              spdk_nvmf_transport_destroy_done_cb cb_fn, void *cb_arg);
    2483             : 
    2484             : static inline bool
    2485           0 : nvmf_rdma_is_rxe_device(struct spdk_nvmf_rdma_device *device)
    2486             : {
    2487           0 :         return device->attr.vendor_id == SPDK_RDMA_RXE_VENDOR_ID_OLD ||
    2488           0 :                device->attr.vendor_id == SPDK_RDMA_RXE_VENDOR_ID_NEW;
    2489             : }
    2490             : 
    2491             : static int nvmf_rdma_accept(void *ctx);
    2492             : static bool nvmf_rdma_retry_listen_port(struct spdk_nvmf_rdma_transport *rtransport);
    2493             : static void destroy_ib_device(struct spdk_nvmf_rdma_transport *rtransport,
    2494             :                               struct spdk_nvmf_rdma_device *device);
    2495             : 
    2496             : static int
    2497           0 : create_ib_device(struct spdk_nvmf_rdma_transport *rtransport, struct ibv_context *context,
    2498             :                  struct spdk_nvmf_rdma_device **new_device)
    2499             : {
    2500             :         struct spdk_nvmf_rdma_device    *device;
    2501           0 :         int                             flag = 0;
    2502           0 :         int                             rc = 0;
    2503             : 
    2504           0 :         device = calloc(1, sizeof(*device));
    2505           0 :         if (!device) {
    2506           0 :                 SPDK_ERRLOG("Unable to allocate memory for RDMA devices.\n");
    2507           0 :                 return -ENOMEM;
    2508             :         }
    2509           0 :         device->context = context;
    2510           0 :         rc = ibv_query_device(device->context, &device->attr);
    2511           0 :         if (rc < 0) {
    2512           0 :                 SPDK_ERRLOG("Failed to query RDMA device attributes.\n");
    2513           0 :                 free(device);
    2514           0 :                 return rc;
    2515             :         }
    2516             : 
    2517             : #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL
    2518           0 :         if ((device->attr.device_cap_flags & IBV_DEVICE_MEM_MGT_EXTENSIONS) == 0) {
    2519           0 :                 SPDK_WARNLOG("The libibverbs on this system supports SEND_WITH_INVALIDATE,");
    2520           0 :                 SPDK_WARNLOG("but the device with vendor ID %u does not.\n", device->attr.vendor_id);
    2521             :         }
    2522             : 
    2523             :         /**
    2524             :          * The vendor ID is assigned by the IEEE and an ID of 0 implies Soft-RoCE.
    2525             :          * The Soft-RoCE RXE driver does not currently support send with invalidate,
    2526             :          * but incorrectly reports that it does. There are changes making their way
    2527             :          * through the kernel now that will enable this feature. When they are merged,
    2528             :          * we can conditionally enable this feature.
    2529             :          *
    2530             :          * TODO: enable this for versions of the kernel rxe driver that support it.
    2531             :          */
    2532           0 :         if (nvmf_rdma_is_rxe_device(device)) {
    2533           0 :                 device->attr.device_cap_flags &= ~(IBV_DEVICE_MEM_MGT_EXTENSIONS);
    2534             :         }
    2535             : #endif
    2536             : 
    2537             :         /* set up device context async ev fd as NON_BLOCKING */
    2538           0 :         flag = fcntl(device->context->async_fd, F_GETFL);
    2539           0 :         rc = fcntl(device->context->async_fd, F_SETFL, flag | O_NONBLOCK);
    2540           0 :         if (rc < 0) {
    2541           0 :                 SPDK_ERRLOG("Failed to set context async fd to NONBLOCK.\n");
    2542           0 :                 free(device);
    2543           0 :                 return rc;
    2544             :         }
    2545             : 
    2546           0 :         TAILQ_INSERT_TAIL(&rtransport->devices, device, link);
    2547           0 :         SPDK_DEBUGLOG(rdma, "New device %p is added to RDMA trasport\n", device);
    2548             : 
    2549           0 :         if (g_nvmf_hooks.get_ibv_pd) {
    2550           0 :                 device->pd = g_nvmf_hooks.get_ibv_pd(NULL, device->context);
    2551             :         } else {
    2552           0 :                 device->pd = ibv_alloc_pd(device->context);
    2553             :         }
    2554             : 
    2555           0 :         if (!device->pd) {
    2556           0 :                 SPDK_ERRLOG("Unable to allocate protection domain.\n");
    2557           0 :                 destroy_ib_device(rtransport, device);
    2558           0 :                 return -ENOMEM;
    2559             :         }
    2560             : 
    2561           0 :         assert(device->map == NULL);
    2562             : 
    2563           0 :         device->map = spdk_rdma_create_mem_map(device->pd, &g_nvmf_hooks, SPDK_RDMA_MEMORY_MAP_ROLE_TARGET);
    2564           0 :         if (!device->map) {
    2565           0 :                 SPDK_ERRLOG("Unable to allocate memory map for listen address\n");
    2566           0 :                 destroy_ib_device(rtransport, device);
    2567           0 :                 return -ENOMEM;
    2568             :         }
    2569             : 
    2570           0 :         assert(device->map != NULL);
    2571           0 :         assert(device->pd != NULL);
    2572             : 
    2573           0 :         if (new_device) {
    2574           0 :                 *new_device = device;
    2575             :         }
    2576           0 :         SPDK_NOTICELOG("Create IB device %s(%p/%p) succeed.\n", ibv_get_device_name(context->device),
    2577             :                        device, context);
    2578             : 
    2579           0 :         return 0;
    2580             : }
    2581             : 
    2582             : static void
    2583           0 : free_poll_fds(struct spdk_nvmf_rdma_transport *rtransport)
    2584             : {
    2585           0 :         if (rtransport->poll_fds) {
    2586           0 :                 free(rtransport->poll_fds);
    2587           0 :                 rtransport->poll_fds = NULL;
    2588             :         }
    2589           0 :         rtransport->npoll_fds = 0;
    2590           0 : }
    2591             : 
    2592             : static int
    2593           0 : generate_poll_fds(struct spdk_nvmf_rdma_transport *rtransport)
    2594             : {
    2595             :         /* Set up poll descriptor array to monitor events from RDMA and IB
    2596             :          * in a single poll syscall
    2597             :          */
    2598           0 :         int device_count = 0;
    2599           0 :         int i = 0;
    2600             :         struct spdk_nvmf_rdma_device *device, *tmp;
    2601             : 
    2602           0 :         TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, tmp) {
    2603           0 :                 device_count++;
    2604             :         }
    2605             : 
    2606           0 :         rtransport->npoll_fds = device_count + 1;
    2607             : 
    2608           0 :         rtransport->poll_fds = calloc(rtransport->npoll_fds, sizeof(struct pollfd));
    2609           0 :         if (rtransport->poll_fds == NULL) {
    2610           0 :                 SPDK_ERRLOG("poll_fds allocation failed\n");
    2611           0 :                 return -ENOMEM;
    2612             :         }
    2613             : 
    2614           0 :         rtransport->poll_fds[i].fd = rtransport->event_channel->fd;
    2615           0 :         rtransport->poll_fds[i++].events = POLLIN;
    2616             : 
    2617           0 :         TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, tmp) {
    2618           0 :                 rtransport->poll_fds[i].fd = device->context->async_fd;
    2619           0 :                 rtransport->poll_fds[i++].events = POLLIN;
    2620             :         }
    2621             : 
    2622           0 :         return 0;
    2623             : }
    2624             : 
    2625             : static struct spdk_nvmf_transport *
    2626           0 : nvmf_rdma_create(struct spdk_nvmf_transport_opts *opts)
    2627             : {
    2628             :         int rc;
    2629             :         struct spdk_nvmf_rdma_transport *rtransport;
    2630           0 :         struct spdk_nvmf_rdma_device    *device;
    2631             :         struct ibv_context              **contexts;
    2632             :         size_t                          data_wr_pool_size;
    2633             :         uint32_t                        i;
    2634             :         int                             flag;
    2635             :         uint32_t                        sge_count;
    2636             :         uint32_t                        min_shared_buffers;
    2637             :         uint32_t                        min_in_capsule_data_size;
    2638           0 :         int                             max_device_sge = SPDK_NVMF_MAX_SGL_ENTRIES;
    2639             : 
    2640           0 :         rtransport = calloc(1, sizeof(*rtransport));
    2641           0 :         if (!rtransport) {
    2642           0 :                 return NULL;
    2643             :         }
    2644             : 
    2645           0 :         TAILQ_INIT(&rtransport->devices);
    2646           0 :         TAILQ_INIT(&rtransport->ports);
    2647           0 :         TAILQ_INIT(&rtransport->poll_groups);
    2648           0 :         TAILQ_INIT(&rtransport->retry_ports);
    2649             : 
    2650           0 :         rtransport->transport.ops = &spdk_nvmf_transport_rdma;
    2651           0 :         rtransport->rdma_opts.num_cqe = DEFAULT_NVMF_RDMA_CQ_SIZE;
    2652           0 :         rtransport->rdma_opts.max_srq_depth = SPDK_NVMF_RDMA_DEFAULT_SRQ_DEPTH;
    2653           0 :         rtransport->rdma_opts.no_srq = SPDK_NVMF_RDMA_DEFAULT_NO_SRQ;
    2654           0 :         rtransport->rdma_opts.acceptor_backlog = SPDK_NVMF_RDMA_ACCEPTOR_BACKLOG;
    2655           0 :         rtransport->rdma_opts.no_wr_batching = SPDK_NVMF_RDMA_DEFAULT_NO_WR_BATCHING;
    2656           0 :         if (opts->transport_specific != NULL &&
    2657           0 :             spdk_json_decode_object_relaxed(opts->transport_specific, rdma_transport_opts_decoder,
    2658             :                                             SPDK_COUNTOF(rdma_transport_opts_decoder),
    2659           0 :                                             &rtransport->rdma_opts)) {
    2660           0 :                 SPDK_ERRLOG("spdk_json_decode_object_relaxed failed\n");
    2661           0 :                 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL);
    2662           0 :                 return NULL;
    2663             :         }
    2664             : 
    2665           0 :         SPDK_INFOLOG(rdma, "*** RDMA Transport Init ***\n"
    2666             :                      "  Transport opts:  max_ioq_depth=%d, max_io_size=%d,\n"
    2667             :                      "  max_io_qpairs_per_ctrlr=%d, io_unit_size=%d,\n"
    2668             :                      "  in_capsule_data_size=%d, max_aq_depth=%d,\n"
    2669             :                      "  num_shared_buffers=%d, num_cqe=%d, max_srq_depth=%d, no_srq=%d,"
    2670             :                      "  acceptor_backlog=%d, no_wr_batching=%d abort_timeout_sec=%d\n",
    2671             :                      opts->max_queue_depth,
    2672             :                      opts->max_io_size,
    2673             :                      opts->max_qpairs_per_ctrlr - 1,
    2674             :                      opts->io_unit_size,
    2675             :                      opts->in_capsule_data_size,
    2676             :                      opts->max_aq_depth,
    2677             :                      opts->num_shared_buffers,
    2678             :                      rtransport->rdma_opts.num_cqe,
    2679             :                      rtransport->rdma_opts.max_srq_depth,
    2680             :                      rtransport->rdma_opts.no_srq,
    2681             :                      rtransport->rdma_opts.acceptor_backlog,
    2682             :                      rtransport->rdma_opts.no_wr_batching,
    2683             :                      opts->abort_timeout_sec);
    2684             : 
    2685             :         /* I/O unit size cannot be larger than max I/O size */
    2686           0 :         if (opts->io_unit_size > opts->max_io_size) {
    2687           0 :                 opts->io_unit_size = opts->max_io_size;
    2688             :         }
    2689             : 
    2690           0 :         if (rtransport->rdma_opts.acceptor_backlog <= 0) {
    2691           0 :                 SPDK_ERRLOG("The acceptor backlog cannot be less than 1, setting to the default value of (%d).\n",
    2692             :                             SPDK_NVMF_RDMA_ACCEPTOR_BACKLOG);
    2693           0 :                 rtransport->rdma_opts.acceptor_backlog = SPDK_NVMF_RDMA_ACCEPTOR_BACKLOG;
    2694             :         }
    2695             : 
    2696           0 :         if (opts->num_shared_buffers < (SPDK_NVMF_MAX_SGL_ENTRIES * 2)) {
    2697           0 :                 SPDK_ERRLOG("The number of shared data buffers (%d) is less than"
    2698             :                             "the minimum number required to guarantee that forward progress can be made (%d)\n",
    2699             :                             opts->num_shared_buffers, (SPDK_NVMF_MAX_SGL_ENTRIES * 2));
    2700           0 :                 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL);
    2701           0 :                 return NULL;
    2702             :         }
    2703             : 
    2704             :         /* If buf_cache_size == UINT32_MAX, we will dynamically pick a cache size later that we know will fit. */
    2705           0 :         if (opts->buf_cache_size < UINT32_MAX) {
    2706           0 :                 min_shared_buffers = spdk_env_get_core_count() * opts->buf_cache_size;
    2707           0 :                 if (min_shared_buffers > opts->num_shared_buffers) {
    2708           0 :                         SPDK_ERRLOG("There are not enough buffers to satisfy"
    2709             :                                     "per-poll group caches for each thread. (%" PRIu32 ")"
    2710             :                                     "supplied. (%" PRIu32 ") required\n", opts->num_shared_buffers, min_shared_buffers);
    2711           0 :                         SPDK_ERRLOG("Please specify a larger number of shared buffers\n");
    2712           0 :                         nvmf_rdma_destroy(&rtransport->transport, NULL, NULL);
    2713           0 :                         return NULL;
    2714             :                 }
    2715             :         }
    2716             : 
    2717           0 :         sge_count = opts->max_io_size / opts->io_unit_size;
    2718           0 :         if (sge_count > NVMF_DEFAULT_TX_SGE) {
    2719           0 :                 SPDK_ERRLOG("Unsupported IO Unit size specified, %d bytes\n", opts->io_unit_size);
    2720           0 :                 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL);
    2721           0 :                 return NULL;
    2722             :         }
    2723             : 
    2724           0 :         min_in_capsule_data_size = sizeof(struct spdk_nvme_sgl_descriptor) * SPDK_NVMF_MAX_SGL_ENTRIES;
    2725           0 :         if (opts->in_capsule_data_size < min_in_capsule_data_size) {
    2726           0 :                 SPDK_WARNLOG("In capsule data size is set to %u, this is minimum size required to support msdbd=16\n",
    2727             :                              min_in_capsule_data_size);
    2728           0 :                 opts->in_capsule_data_size = min_in_capsule_data_size;
    2729             :         }
    2730             : 
    2731           0 :         rtransport->event_channel = rdma_create_event_channel();
    2732           0 :         if (rtransport->event_channel == NULL) {
    2733           0 :                 SPDK_ERRLOG("rdma_create_event_channel() failed, %s\n", spdk_strerror(errno));
    2734           0 :                 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL);
    2735           0 :                 return NULL;
    2736             :         }
    2737             : 
    2738           0 :         flag = fcntl(rtransport->event_channel->fd, F_GETFL);
    2739           0 :         if (fcntl(rtransport->event_channel->fd, F_SETFL, flag | O_NONBLOCK) < 0) {
    2740           0 :                 SPDK_ERRLOG("fcntl can't set nonblocking mode for socket, fd: %d (%s)\n",
    2741             :                             rtransport->event_channel->fd, spdk_strerror(errno));
    2742           0 :                 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL);
    2743           0 :                 return NULL;
    2744             :         }
    2745             : 
    2746           0 :         data_wr_pool_size = opts->data_wr_pool_size;
    2747           0 :         if (data_wr_pool_size < SPDK_NVMF_MAX_SGL_ENTRIES * 2 * spdk_env_get_core_count()) {
    2748           0 :                 data_wr_pool_size = SPDK_NVMF_MAX_SGL_ENTRIES * 2 * spdk_env_get_core_count();
    2749           0 :                 SPDK_NOTICELOG("data_wr_pool_size is changed to %zu to guarantee enough cache for handling "
    2750             :                                "at least one IO in each core\n", data_wr_pool_size);
    2751             :         }
    2752           0 :         rtransport->data_wr_pool = spdk_mempool_create("spdk_nvmf_rdma_wr_data", data_wr_pool_size,
    2753             :                                    sizeof(struct spdk_nvmf_rdma_request_data), SPDK_MEMPOOL_DEFAULT_CACHE_SIZE,
    2754             :                                    SPDK_ENV_SOCKET_ID_ANY);
    2755           0 :         if (!rtransport->data_wr_pool) {
    2756           0 :                 if (spdk_mempool_lookup("spdk_nvmf_rdma_wr_data") != NULL) {
    2757           0 :                         SPDK_ERRLOG("Unable to allocate work request pool for poll group: already exists\n");
    2758           0 :                         SPDK_ERRLOG("Probably running in multiprocess environment, which is "
    2759             :                                     "unsupported by the nvmf library\n");
    2760             :                 } else {
    2761           0 :                         SPDK_ERRLOG("Unable to allocate work request pool for poll group\n");
    2762             :                 }
    2763           0 :                 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL);
    2764           0 :                 return NULL;
    2765             :         }
    2766             : 
    2767           0 :         contexts = rdma_get_devices(NULL);
    2768           0 :         if (contexts == NULL) {
    2769           0 :                 SPDK_ERRLOG("rdma_get_devices() failed: %s (%d)\n", spdk_strerror(errno), errno);
    2770           0 :                 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL);
    2771           0 :                 return NULL;
    2772             :         }
    2773             : 
    2774           0 :         i = 0;
    2775           0 :         rc = 0;
    2776           0 :         while (contexts[i] != NULL) {
    2777           0 :                 rc = create_ib_device(rtransport, contexts[i], &device);
    2778           0 :                 if (rc < 0) {
    2779           0 :                         break;
    2780             :                 }
    2781           0 :                 i++;
    2782           0 :                 max_device_sge = spdk_min(max_device_sge, device->attr.max_sge);
    2783           0 :                 device->is_ready = true;
    2784             :         }
    2785           0 :         rdma_free_devices(contexts);
    2786             : 
    2787           0 :         if (opts->io_unit_size * max_device_sge < opts->max_io_size) {
    2788             :                 /* divide and round up. */
    2789           0 :                 opts->io_unit_size = (opts->max_io_size + max_device_sge - 1) / max_device_sge;
    2790             : 
    2791             :                 /* round up to the nearest 4k. */
    2792           0 :                 opts->io_unit_size = (opts->io_unit_size + NVMF_DATA_BUFFER_ALIGNMENT - 1) & ~NVMF_DATA_BUFFER_MASK;
    2793             : 
    2794           0 :                 opts->io_unit_size = spdk_max(opts->io_unit_size, SPDK_NVMF_RDMA_MIN_IO_BUFFER_SIZE);
    2795           0 :                 SPDK_NOTICELOG("Adjusting the io unit size to fit the device's maximum I/O size. New I/O unit size %u\n",
    2796             :                                opts->io_unit_size);
    2797             :         }
    2798             : 
    2799           0 :         if (rc < 0) {
    2800           0 :                 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL);
    2801           0 :                 return NULL;
    2802             :         }
    2803             : 
    2804           0 :         rc = generate_poll_fds(rtransport);
    2805           0 :         if (rc < 0) {
    2806           0 :                 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL);
    2807           0 :                 return NULL;
    2808             :         }
    2809             : 
    2810           0 :         rtransport->accept_poller = SPDK_POLLER_REGISTER(nvmf_rdma_accept, &rtransport->transport,
    2811             :                                     opts->acceptor_poll_rate);
    2812           0 :         if (!rtransport->accept_poller) {
    2813           0 :                 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL);
    2814           0 :                 return NULL;
    2815             :         }
    2816             : 
    2817           0 :         return &rtransport->transport;
    2818             : }
    2819             : 
    2820             : static void
    2821           0 : destroy_ib_device(struct spdk_nvmf_rdma_transport *rtransport,
    2822             :                   struct spdk_nvmf_rdma_device *device)
    2823             : {
    2824           0 :         TAILQ_REMOVE(&rtransport->devices, device, link);
    2825           0 :         spdk_rdma_free_mem_map(&device->map);
    2826           0 :         if (device->pd) {
    2827           0 :                 if (!g_nvmf_hooks.get_ibv_pd) {
    2828           0 :                         ibv_dealloc_pd(device->pd);
    2829             :                 }
    2830             :         }
    2831           0 :         SPDK_DEBUGLOG(rdma, "IB device [%p] is destroyed.\n", device);
    2832           0 :         free(device);
    2833           0 : }
    2834             : 
    2835             : static void
    2836           0 : nvmf_rdma_dump_opts(struct spdk_nvmf_transport *transport, struct spdk_json_write_ctx *w)
    2837             : {
    2838             :         struct spdk_nvmf_rdma_transport *rtransport;
    2839           0 :         assert(w != NULL);
    2840             : 
    2841           0 :         rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport);
    2842           0 :         spdk_json_write_named_uint32(w, "max_srq_depth", rtransport->rdma_opts.max_srq_depth);
    2843           0 :         spdk_json_write_named_bool(w, "no_srq", rtransport->rdma_opts.no_srq);
    2844           0 :         if (rtransport->rdma_opts.no_srq == true) {
    2845           0 :                 spdk_json_write_named_int32(w, "num_cqe", rtransport->rdma_opts.num_cqe);
    2846             :         }
    2847           0 :         spdk_json_write_named_int32(w, "acceptor_backlog", rtransport->rdma_opts.acceptor_backlog);
    2848           0 :         spdk_json_write_named_bool(w, "no_wr_batching", rtransport->rdma_opts.no_wr_batching);
    2849           0 : }
    2850             : 
    2851             : static int
    2852           0 : nvmf_rdma_destroy(struct spdk_nvmf_transport *transport,
    2853             :                   spdk_nvmf_transport_destroy_done_cb cb_fn, void *cb_arg)
    2854             : {
    2855             :         struct spdk_nvmf_rdma_transport *rtransport;
    2856             :         struct spdk_nvmf_rdma_port      *port, *port_tmp;
    2857             :         struct spdk_nvmf_rdma_device    *device, *device_tmp;
    2858             : 
    2859           0 :         rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport);
    2860             : 
    2861           0 :         TAILQ_FOREACH_SAFE(port, &rtransport->retry_ports, link, port_tmp) {
    2862           0 :                 TAILQ_REMOVE(&rtransport->retry_ports, port, link);
    2863           0 :                 free(port);
    2864             :         }
    2865             : 
    2866           0 :         TAILQ_FOREACH_SAFE(port, &rtransport->ports, link, port_tmp) {
    2867           0 :                 TAILQ_REMOVE(&rtransport->ports, port, link);
    2868           0 :                 rdma_destroy_id(port->id);
    2869           0 :                 free(port);
    2870             :         }
    2871             : 
    2872           0 :         free_poll_fds(rtransport);
    2873             : 
    2874           0 :         if (rtransport->event_channel != NULL) {
    2875           0 :                 rdma_destroy_event_channel(rtransport->event_channel);
    2876             :         }
    2877             : 
    2878           0 :         TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, device_tmp) {
    2879           0 :                 destroy_ib_device(rtransport, device);
    2880             :         }
    2881             : 
    2882           0 :         if (rtransport->data_wr_pool != NULL) {
    2883           0 :                 if (spdk_mempool_count(rtransport->data_wr_pool) != transport->opts.data_wr_pool_size) {
    2884           0 :                         SPDK_ERRLOG("transport wr pool count is %zu but should be %u\n",
    2885             :                                     spdk_mempool_count(rtransport->data_wr_pool),
    2886             :                                     transport->opts.max_queue_depth * SPDK_NVMF_MAX_SGL_ENTRIES);
    2887             :                 }
    2888             :         }
    2889             : 
    2890           0 :         spdk_mempool_free(rtransport->data_wr_pool);
    2891             : 
    2892           0 :         spdk_poller_unregister(&rtransport->accept_poller);
    2893           0 :         free(rtransport);
    2894             : 
    2895           0 :         if (cb_fn) {
    2896           0 :                 cb_fn(cb_arg);
    2897             :         }
    2898           0 :         return 0;
    2899             : }
    2900             : 
    2901             : static int nvmf_rdma_trid_from_cm_id(struct rdma_cm_id *id,
    2902             :                                      struct spdk_nvme_transport_id *trid,
    2903             :                                      bool peer);
    2904             : 
    2905             : static bool nvmf_rdma_rescan_devices(struct spdk_nvmf_rdma_transport *rtransport);
    2906             : 
    2907             : static int
    2908           0 : nvmf_rdma_listen(struct spdk_nvmf_transport *transport, const struct spdk_nvme_transport_id *trid,
    2909             :                  struct spdk_nvmf_listen_opts *listen_opts)
    2910             : {
    2911             :         struct spdk_nvmf_rdma_transport *rtransport;
    2912             :         struct spdk_nvmf_rdma_device    *device;
    2913             :         struct spdk_nvmf_rdma_port      *port, *tmp_port;
    2914           0 :         struct addrinfo                 *res;
    2915           0 :         struct addrinfo                 hints;
    2916             :         int                             family;
    2917             :         int                             rc;
    2918             :         long int                        port_val;
    2919           0 :         bool                            is_retry = false;
    2920             : 
    2921           0 :         if (!strlen(trid->trsvcid)) {
    2922           0 :                 SPDK_ERRLOG("Service id is required\n");
    2923           0 :                 return -EINVAL;
    2924             :         }
    2925             : 
    2926           0 :         rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport);
    2927           0 :         assert(rtransport->event_channel != NULL);
    2928             : 
    2929           0 :         port = calloc(1, sizeof(*port));
    2930           0 :         if (!port) {
    2931           0 :                 SPDK_ERRLOG("Port allocation failed\n");
    2932           0 :                 return -ENOMEM;
    2933             :         }
    2934             : 
    2935           0 :         port->trid = trid;
    2936             : 
    2937           0 :         switch (trid->adrfam) {
    2938           0 :         case SPDK_NVMF_ADRFAM_IPV4:
    2939           0 :                 family = AF_INET;
    2940           0 :                 break;
    2941           0 :         case SPDK_NVMF_ADRFAM_IPV6:
    2942           0 :                 family = AF_INET6;
    2943           0 :                 break;
    2944           0 :         default:
    2945           0 :                 SPDK_ERRLOG("Unhandled ADRFAM %d\n", trid->adrfam);
    2946           0 :                 free(port);
    2947           0 :                 return -EINVAL;
    2948             :         }
    2949             : 
    2950           0 :         memset(&hints, 0, sizeof(hints));
    2951           0 :         hints.ai_family = family;
    2952           0 :         hints.ai_flags = AI_NUMERICSERV;
    2953           0 :         hints.ai_socktype = SOCK_STREAM;
    2954           0 :         hints.ai_protocol = 0;
    2955             : 
    2956             :         /* Range check the trsvcid. Fail in 3 cases:
    2957             :          * < 0: means that spdk_strtol hit an error
    2958             :          * 0: this results in ephemeral port which we don't want
    2959             :          * > 65535: port too high
    2960             :          */
    2961           0 :         port_val = spdk_strtol(trid->trsvcid, 10);
    2962           0 :         if (port_val <= 0 || port_val > 65535) {
    2963           0 :                 SPDK_ERRLOG("invalid trsvcid %s\n", trid->trsvcid);
    2964           0 :                 free(port);
    2965           0 :                 return -EINVAL;
    2966             :         }
    2967             : 
    2968           0 :         rc = getaddrinfo(trid->traddr, trid->trsvcid, &hints, &res);
    2969           0 :         if (rc) {
    2970           0 :                 SPDK_ERRLOG("getaddrinfo failed: %s (%d)\n", gai_strerror(rc), rc);
    2971           0 :                 free(port);
    2972           0 :                 return -(abs(rc));
    2973             :         }
    2974             : 
    2975           0 :         rc = rdma_create_id(rtransport->event_channel, &port->id, port, RDMA_PS_TCP);
    2976           0 :         if (rc < 0) {
    2977           0 :                 SPDK_ERRLOG("rdma_create_id() failed\n");
    2978           0 :                 freeaddrinfo(res);
    2979           0 :                 free(port);
    2980           0 :                 return rc;
    2981             :         }
    2982             : 
    2983           0 :         rc = rdma_bind_addr(port->id, res->ai_addr);
    2984           0 :         freeaddrinfo(res);
    2985             : 
    2986           0 :         if (rc < 0) {
    2987           0 :                 TAILQ_FOREACH(tmp_port, &rtransport->retry_ports, link) {
    2988           0 :                         if (spdk_nvme_transport_id_compare(tmp_port->trid, trid) == 0) {
    2989           0 :                                 is_retry = true;
    2990           0 :                                 break;
    2991             :                         }
    2992             :                 }
    2993           0 :                 if (!is_retry) {
    2994           0 :                         SPDK_ERRLOG("rdma_bind_addr() failed\n");
    2995             :                 }
    2996           0 :                 rdma_destroy_id(port->id);
    2997           0 :                 free(port);
    2998           0 :                 return rc;
    2999             :         }
    3000             : 
    3001           0 :         if (!port->id->verbs) {
    3002           0 :                 SPDK_ERRLOG("ibv_context is null\n");
    3003           0 :                 rdma_destroy_id(port->id);
    3004           0 :                 free(port);
    3005           0 :                 return -1;
    3006             :         }
    3007             : 
    3008           0 :         rc = rdma_listen(port->id, rtransport->rdma_opts.acceptor_backlog);
    3009           0 :         if (rc < 0) {
    3010           0 :                 SPDK_ERRLOG("rdma_listen() failed\n");
    3011           0 :                 rdma_destroy_id(port->id);
    3012           0 :                 free(port);
    3013           0 :                 return rc;
    3014             :         }
    3015             : 
    3016           0 :         TAILQ_FOREACH(device, &rtransport->devices, link) {
    3017           0 :                 if (device->context == port->id->verbs && device->is_ready) {
    3018           0 :                         port->device = device;
    3019           0 :                         break;
    3020             :                 }
    3021             :         }
    3022           0 :         if (!port->device) {
    3023           0 :                 SPDK_ERRLOG("Accepted a connection with verbs %p, but unable to find a corresponding device.\n",
    3024             :                             port->id->verbs);
    3025           0 :                 rdma_destroy_id(port->id);
    3026           0 :                 free(port);
    3027           0 :                 nvmf_rdma_rescan_devices(rtransport);
    3028           0 :                 return -EINVAL;
    3029             :         }
    3030             : 
    3031           0 :         SPDK_NOTICELOG("*** NVMe/RDMA Target Listening on %s port %s ***\n",
    3032             :                        trid->traddr, trid->trsvcid);
    3033             : 
    3034           0 :         TAILQ_INSERT_TAIL(&rtransport->ports, port, link);
    3035           0 :         return 0;
    3036             : }
    3037             : 
    3038             : static void
    3039           0 : nvmf_rdma_stop_listen_ex(struct spdk_nvmf_transport *transport,
    3040             :                          const struct spdk_nvme_transport_id *trid, bool need_retry)
    3041             : {
    3042             :         struct spdk_nvmf_rdma_transport *rtransport;
    3043             :         struct spdk_nvmf_rdma_port      *port, *tmp;
    3044             : 
    3045           0 :         rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport);
    3046             : 
    3047           0 :         if (!need_retry) {
    3048           0 :                 TAILQ_FOREACH_SAFE(port, &rtransport->retry_ports, link, tmp) {
    3049           0 :                         if (spdk_nvme_transport_id_compare(port->trid, trid) == 0) {
    3050           0 :                                 TAILQ_REMOVE(&rtransport->retry_ports, port, link);
    3051           0 :                                 free(port);
    3052             :                         }
    3053             :                 }
    3054             :         }
    3055             : 
    3056           0 :         TAILQ_FOREACH_SAFE(port, &rtransport->ports, link, tmp) {
    3057           0 :                 if (spdk_nvme_transport_id_compare(port->trid, trid) == 0) {
    3058           0 :                         SPDK_DEBUGLOG(rdma, "Port %s:%s removed. need retry: %d\n",
    3059             :                                       port->trid->traddr, port->trid->trsvcid, need_retry);
    3060           0 :                         TAILQ_REMOVE(&rtransport->ports, port, link);
    3061           0 :                         rdma_destroy_id(port->id);
    3062           0 :                         port->id = NULL;
    3063           0 :                         port->device = NULL;
    3064           0 :                         if (need_retry) {
    3065           0 :                                 TAILQ_INSERT_TAIL(&rtransport->retry_ports, port, link);
    3066             :                         } else {
    3067           0 :                                 free(port);
    3068             :                         }
    3069           0 :                         break;
    3070             :                 }
    3071             :         }
    3072           0 : }
    3073             : 
    3074             : static void
    3075           0 : nvmf_rdma_stop_listen(struct spdk_nvmf_transport *transport,
    3076             :                       const struct spdk_nvme_transport_id *trid)
    3077             : {
    3078           0 :         nvmf_rdma_stop_listen_ex(transport, trid, false);
    3079           0 : }
    3080             : 
    3081             : static void _nvmf_rdma_register_poller_in_group(void *c);
    3082             : static void _nvmf_rdma_remove_poller_in_group(void *c);
    3083             : 
    3084             : static bool
    3085           0 : nvmf_rdma_all_pollers_management_done(void *c)
    3086             : {
    3087           0 :         struct poller_manage_ctx        *ctx = c;
    3088             :         int                             counter;
    3089             : 
    3090           0 :         counter = __atomic_sub_fetch(ctx->inflight_op_counter, 1, __ATOMIC_SEQ_CST);
    3091           0 :         SPDK_DEBUGLOG(rdma, "nvmf_rdma_all_pollers_management_done called. counter: %d, poller: %p\n",
    3092             :                       counter, ctx->rpoller);
    3093             : 
    3094           0 :         if (counter == 0) {
    3095           0 :                 free((void *)ctx->inflight_op_counter);
    3096             :         }
    3097           0 :         free(ctx);
    3098             : 
    3099           0 :         return counter == 0;
    3100             : }
    3101             : 
    3102             : static int
    3103           0 : nvmf_rdma_manage_poller(struct spdk_nvmf_rdma_transport *rtransport,
    3104             :                         struct spdk_nvmf_rdma_device *device, bool *has_inflight, bool is_add)
    3105             : {
    3106             :         struct spdk_nvmf_rdma_poll_group        *rgroup;
    3107             :         struct spdk_nvmf_rdma_poller            *rpoller;
    3108             :         struct spdk_nvmf_poll_group             *poll_group;
    3109             :         struct poller_manage_ctx                *ctx;
    3110             :         bool                                    found;
    3111             :         int                                     *inflight_counter;
    3112             :         spdk_msg_fn                             do_fn;
    3113             : 
    3114           0 :         *has_inflight = false;
    3115           0 :         do_fn = is_add ? _nvmf_rdma_register_poller_in_group : _nvmf_rdma_remove_poller_in_group;
    3116           0 :         inflight_counter = calloc(1, sizeof(int));
    3117           0 :         if (!inflight_counter) {
    3118           0 :                 SPDK_ERRLOG("Failed to allocate inflight counter when removing pollers\n");
    3119           0 :                 return -ENOMEM;
    3120             :         }
    3121             : 
    3122           0 :         TAILQ_FOREACH(rgroup, &rtransport->poll_groups, link) {
    3123           0 :                 (*inflight_counter)++;
    3124             :         }
    3125             : 
    3126           0 :         TAILQ_FOREACH(rgroup, &rtransport->poll_groups, link) {
    3127           0 :                 found = false;
    3128           0 :                 TAILQ_FOREACH(rpoller, &rgroup->pollers, link) {
    3129           0 :                         if (rpoller->device == device) {
    3130           0 :                                 found = true;
    3131           0 :                                 break;
    3132             :                         }
    3133             :                 }
    3134           0 :                 if (found == is_add) {
    3135           0 :                         __atomic_fetch_sub(inflight_counter, 1, __ATOMIC_SEQ_CST);
    3136           0 :                         continue;
    3137             :                 }
    3138             : 
    3139           0 :                 ctx = calloc(1, sizeof(struct poller_manage_ctx));
    3140           0 :                 if (!ctx) {
    3141           0 :                         SPDK_ERRLOG("Failed to allocate poller_manage_ctx when removing pollers\n");
    3142           0 :                         if (!*has_inflight) {
    3143           0 :                                 free(inflight_counter);
    3144             :                         }
    3145           0 :                         return -ENOMEM;
    3146             :                 }
    3147             : 
    3148           0 :                 ctx->rtransport = rtransport;
    3149           0 :                 ctx->rgroup = rgroup;
    3150           0 :                 ctx->rpoller = rpoller;
    3151           0 :                 ctx->device = device;
    3152           0 :                 ctx->thread = spdk_get_thread();
    3153           0 :                 ctx->inflight_op_counter = inflight_counter;
    3154           0 :                 *has_inflight = true;
    3155             : 
    3156           0 :                 poll_group = rgroup->group.group;
    3157           0 :                 if (poll_group->thread != spdk_get_thread()) {
    3158           0 :                         spdk_thread_send_msg(poll_group->thread, do_fn, ctx);
    3159             :                 } else {
    3160           0 :                         do_fn(ctx);
    3161             :                 }
    3162             :         }
    3163             : 
    3164           0 :         if (!*has_inflight) {
    3165           0 :                 free(inflight_counter);
    3166             :         }
    3167             : 
    3168           0 :         return 0;
    3169             : }
    3170             : 
    3171             : static void nvmf_rdma_handle_device_removal(struct spdk_nvmf_rdma_transport *rtransport,
    3172             :                 struct spdk_nvmf_rdma_device *device);
    3173             : 
    3174             : static struct spdk_nvmf_rdma_device *
    3175           0 : nvmf_rdma_find_ib_device(struct spdk_nvmf_rdma_transport *rtransport,
    3176             :                          struct ibv_context *context)
    3177             : {
    3178             :         struct spdk_nvmf_rdma_device    *device, *tmp_device;
    3179             : 
    3180           0 :         TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, tmp_device) {
    3181           0 :                 if (device->need_destroy) {
    3182           0 :                         continue;
    3183             :                 }
    3184             : 
    3185           0 :                 if (strcmp(device->context->device->dev_name, context->device->dev_name) == 0) {
    3186           0 :                         return device;
    3187             :                 }
    3188             :         }
    3189             : 
    3190           0 :         return NULL;
    3191             : }
    3192             : 
    3193             : static bool
    3194           0 : nvmf_rdma_check_devices_context(struct spdk_nvmf_rdma_transport *rtransport,
    3195             :                                 struct ibv_context *context)
    3196             : {
    3197           0 :         struct spdk_nvmf_rdma_device    *old_device, *new_device;
    3198           0 :         int                             rc = 0;
    3199           0 :         bool                            has_inflight;
    3200             : 
    3201           0 :         old_device = nvmf_rdma_find_ib_device(rtransport, context);
    3202             : 
    3203           0 :         if (old_device) {
    3204           0 :                 if (old_device->context != context && !old_device->need_destroy && old_device->is_ready) {
    3205             :                         /* context may not have time to be cleaned when rescan. exactly one context
    3206             :                          * is valid for a device so this context must be invalid and just remove it. */
    3207           0 :                         SPDK_WARNLOG("Device %p has a invalid context %p\n", old_device, old_device->context);
    3208           0 :                         old_device->need_destroy = true;
    3209           0 :                         nvmf_rdma_handle_device_removal(rtransport, old_device);
    3210             :                 }
    3211           0 :                 return false;
    3212             :         }
    3213             : 
    3214           0 :         rc = create_ib_device(rtransport, context, &new_device);
    3215             :         /* TODO: update transport opts. */
    3216           0 :         if (rc < 0) {
    3217           0 :                 SPDK_ERRLOG("Failed to create ib device for context: %s(%p)\n",
    3218             :                             ibv_get_device_name(context->device), context);
    3219           0 :                 return false;
    3220             :         }
    3221             : 
    3222           0 :         rc = nvmf_rdma_manage_poller(rtransport, new_device, &has_inflight, true);
    3223           0 :         if (rc < 0) {
    3224           0 :                 SPDK_ERRLOG("Failed to add poller for device context: %s(%p)\n",
    3225             :                             ibv_get_device_name(context->device), context);
    3226           0 :                 return false;
    3227             :         }
    3228             : 
    3229           0 :         if (has_inflight) {
    3230           0 :                 new_device->is_ready = true;
    3231             :         }
    3232             : 
    3233           0 :         return true;
    3234             : }
    3235             : 
    3236             : static bool
    3237           0 : nvmf_rdma_rescan_devices(struct spdk_nvmf_rdma_transport *rtransport)
    3238             : {
    3239             :         struct spdk_nvmf_rdma_device    *device;
    3240           0 :         struct ibv_device               **ibv_device_list = NULL;
    3241           0 :         struct ibv_context              **contexts = NULL;
    3242           0 :         int                             i = 0;
    3243           0 :         int                             num_dev = 0;
    3244           0 :         bool                            new_create = false, has_new_device = false;
    3245           0 :         struct ibv_context              *tmp_verbs = NULL;
    3246             : 
    3247             :         /* do not rescan when any device is destroying, or context may be freed when
    3248             :          * regenerating the poll fds.
    3249             :          */
    3250           0 :         TAILQ_FOREACH(device, &rtransport->devices, link) {
    3251           0 :                 if (device->need_destroy) {
    3252           0 :                         return false;
    3253             :                 }
    3254             :         }
    3255             : 
    3256           0 :         ibv_device_list = ibv_get_device_list(&num_dev);
    3257             : 
    3258             :         /* There is a bug in librdmacm. If verbs init failed in rdma_get_devices, it'll be
    3259             :          * marked as dead verbs and never be init again. So we need to make sure the
    3260             :          * verbs is available before we call rdma_get_devices. */
    3261           0 :         if (num_dev >= 0) {
    3262           0 :                 for (i = 0; i < num_dev; i++) {
    3263           0 :                         tmp_verbs = ibv_open_device(ibv_device_list[i]);
    3264           0 :                         if (!tmp_verbs) {
    3265           0 :                                 SPDK_WARNLOG("Failed to init ibv device %p, err %d. Skip rescan.\n", ibv_device_list[i], errno);
    3266           0 :                                 break;
    3267             :                         }
    3268           0 :                         if (nvmf_rdma_find_ib_device(rtransport, tmp_verbs) == NULL) {
    3269           0 :                                 SPDK_DEBUGLOG(rdma, "Find new verbs init ibv device %p(%s).\n", ibv_device_list[i],
    3270             :                                               tmp_verbs->device->dev_name);
    3271           0 :                                 has_new_device = true;
    3272             :                         }
    3273           0 :                         ibv_close_device(tmp_verbs);
    3274             :                 }
    3275           0 :                 ibv_free_device_list(ibv_device_list);
    3276           0 :                 if (!tmp_verbs || !has_new_device) {
    3277           0 :                         return false;
    3278             :                 }
    3279             :         }
    3280             : 
    3281           0 :         contexts = rdma_get_devices(NULL);
    3282             : 
    3283           0 :         for (i = 0; contexts && contexts[i] != NULL; i++) {
    3284           0 :                 new_create |= nvmf_rdma_check_devices_context(rtransport, contexts[i]);
    3285             :         }
    3286             : 
    3287           0 :         if (new_create) {
    3288           0 :                 free_poll_fds(rtransport);
    3289           0 :                 generate_poll_fds(rtransport);
    3290             :         }
    3291             : 
    3292           0 :         if (contexts) {
    3293           0 :                 rdma_free_devices(contexts);
    3294             :         }
    3295             : 
    3296           0 :         return new_create;
    3297             : }
    3298             : 
    3299             : static bool
    3300           0 : nvmf_rdma_retry_listen_port(struct spdk_nvmf_rdma_transport *rtransport)
    3301             : {
    3302             :         struct spdk_nvmf_rdma_port      *port, *tmp_port;
    3303           0 :         int                             rc = 0;
    3304           0 :         bool                            new_create = false;
    3305             : 
    3306           0 :         if (TAILQ_EMPTY(&rtransport->retry_ports)) {
    3307           0 :                 return false;
    3308             :         }
    3309             : 
    3310           0 :         new_create = nvmf_rdma_rescan_devices(rtransport);
    3311             : 
    3312           0 :         TAILQ_FOREACH_SAFE(port, &rtransport->retry_ports, link, tmp_port) {
    3313           0 :                 rc = nvmf_rdma_listen(&rtransport->transport, port->trid, NULL);
    3314             : 
    3315           0 :                 TAILQ_REMOVE(&rtransport->retry_ports, port, link);
    3316           0 :                 if (rc) {
    3317           0 :                         if (new_create) {
    3318           0 :                                 SPDK_ERRLOG("Found new IB device but port %s:%s is still failed(%d) to listen.\n",
    3319             :                                             port->trid->traddr, port->trid->trsvcid, rc);
    3320             :                         }
    3321           0 :                         TAILQ_INSERT_TAIL(&rtransport->retry_ports, port, link);
    3322           0 :                         break;
    3323             :                 } else {
    3324           0 :                         SPDK_NOTICELOG("Port %s:%s come back\n", port->trid->traddr, port->trid->trsvcid);
    3325           0 :                         free(port);
    3326             :                 }
    3327             :         }
    3328             : 
    3329           0 :         return true;
    3330             : }
    3331             : 
    3332             : static void
    3333           0 : nvmf_rdma_qpair_process_pending(struct spdk_nvmf_rdma_transport *rtransport,
    3334             :                                 struct spdk_nvmf_rdma_qpair *rqpair, bool drain)
    3335             : {
    3336             :         struct spdk_nvmf_request *req, *tmp;
    3337             :         struct spdk_nvmf_rdma_request   *rdma_req, *req_tmp;
    3338             :         struct spdk_nvmf_rdma_resources *resources;
    3339             : 
    3340             :         /* First process requests which are waiting for response to be sent */
    3341           0 :         STAILQ_FOREACH_SAFE(rdma_req, &rqpair->pending_rdma_send_queue, state_link, req_tmp) {
    3342           0 :                 if (nvmf_rdma_request_process(rtransport, rdma_req) == false && drain == false) {
    3343           0 :                         break;
    3344             :                 }
    3345             :         }
    3346             : 
    3347             :         /* We process I/O in the data transfer pending queue at the highest priority. */
    3348           0 :         STAILQ_FOREACH_SAFE(rdma_req, &rqpair->pending_rdma_read_queue, state_link, req_tmp) {
    3349           0 :                 if (nvmf_rdma_request_process(rtransport, rdma_req) == false && drain == false) {
    3350           0 :                         break;
    3351             :                 }
    3352             :         }
    3353             : 
    3354             :         /* Then RDMA writes since reads have stronger restrictions than writes */
    3355           0 :         STAILQ_FOREACH_SAFE(rdma_req, &rqpair->pending_rdma_write_queue, state_link, req_tmp) {
    3356           0 :                 if (nvmf_rdma_request_process(rtransport, rdma_req) == false && drain == false) {
    3357           0 :                         break;
    3358             :                 }
    3359             :         }
    3360             : 
    3361             :         /* Then we handle request waiting on memory buffers. */
    3362           0 :         STAILQ_FOREACH_SAFE(req, &rqpair->poller->group->group.pending_buf_queue, buf_link, tmp) {
    3363           0 :                 rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req);
    3364           0 :                 if (nvmf_rdma_request_process(rtransport, rdma_req) == false && drain == false) {
    3365           0 :                         break;
    3366             :                 }
    3367             :         }
    3368             : 
    3369           0 :         resources = rqpair->resources;
    3370           0 :         while (!STAILQ_EMPTY(&resources->free_queue) && !STAILQ_EMPTY(&resources->incoming_queue)) {
    3371           0 :                 rdma_req = STAILQ_FIRST(&resources->free_queue);
    3372           0 :                 STAILQ_REMOVE_HEAD(&resources->free_queue, state_link);
    3373           0 :                 rdma_req->recv = STAILQ_FIRST(&resources->incoming_queue);
    3374           0 :                 STAILQ_REMOVE_HEAD(&resources->incoming_queue, link);
    3375             : 
    3376           0 :                 if (rqpair->srq != NULL) {
    3377           0 :                         rdma_req->req.qpair = &rdma_req->recv->qpair->qpair;
    3378           0 :                         rdma_req->recv->qpair->qd++;
    3379             :                 } else {
    3380           0 :                         rqpair->qd++;
    3381             :                 }
    3382             : 
    3383           0 :                 rdma_req->receive_tsc = rdma_req->recv->receive_tsc;
    3384           0 :                 rdma_req->state = RDMA_REQUEST_STATE_NEW;
    3385           0 :                 if (nvmf_rdma_request_process(rtransport, rdma_req) == false) {
    3386           0 :                         break;
    3387             :                 }
    3388             :         }
    3389           0 :         if (!STAILQ_EMPTY(&resources->incoming_queue) && STAILQ_EMPTY(&resources->free_queue)) {
    3390           0 :                 rqpair->poller->stat.pending_free_request++;
    3391             :         }
    3392           0 : }
    3393             : 
    3394             : static void
    3395           0 : nvmf_rdma_poller_process_pending_buf_queue(struct spdk_nvmf_rdma_transport *rtransport,
    3396             :                 struct spdk_nvmf_rdma_poller *rpoller)
    3397             : {
    3398             :         struct spdk_nvmf_request *req, *tmp;
    3399             :         struct spdk_nvmf_rdma_request *rdma_req;
    3400             : 
    3401           0 :         STAILQ_FOREACH_SAFE(req, &rpoller->group->group.pending_buf_queue, buf_link, tmp) {
    3402           0 :                 rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req);
    3403           0 :                 if (nvmf_rdma_request_process(rtransport, rdma_req) == false) {
    3404           0 :                         break;
    3405             :                 }
    3406             :         }
    3407           0 : }
    3408             : 
    3409             : static inline bool
    3410           0 : nvmf_rdma_can_ignore_last_wqe_reached(struct spdk_nvmf_rdma_device *device)
    3411             : {
    3412             :         /* iWARP transport and SoftRoCE driver don't support LAST_WQE_REACHED ibv async event */
    3413           0 :         return nvmf_rdma_is_rxe_device(device) ||
    3414           0 :                device->context->device->transport_type == IBV_TRANSPORT_IWARP;
    3415             : }
    3416             : 
    3417             : static void
    3418           0 : nvmf_rdma_destroy_drained_qpair(struct spdk_nvmf_rdma_qpair *rqpair)
    3419             : {
    3420           0 :         struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(rqpair->qpair.transport,
    3421             :                         struct spdk_nvmf_rdma_transport, transport);
    3422             : 
    3423           0 :         nvmf_rdma_qpair_process_pending(rtransport, rqpair, true);
    3424             : 
    3425             :         /* nvmf_rdma_close_qpair is not called */
    3426           0 :         if (!rqpair->to_close) {
    3427           0 :                 return;
    3428             :         }
    3429             : 
    3430             :         /* device is already destroyed and we should force destroy this qpair. */
    3431           0 :         if (rqpair->poller && rqpair->poller->need_destroy) {
    3432           0 :                 nvmf_rdma_qpair_destroy(rqpair);
    3433           0 :                 return;
    3434             :         }
    3435             : 
    3436             :         /* In non SRQ path, we will reach rqpair->max_queue_depth. In SRQ path, we will get the last_wqe event. */
    3437           0 :         if (rqpair->current_send_depth != 0) {
    3438           0 :                 return;
    3439             :         }
    3440             : 
    3441           0 :         if (rqpair->srq == NULL && rqpair->current_recv_depth != rqpair->max_queue_depth) {
    3442           0 :                 return;
    3443             :         }
    3444             : 
    3445           0 :         if (rqpair->srq != NULL && rqpair->last_wqe_reached == false &&
    3446           0 :             !nvmf_rdma_can_ignore_last_wqe_reached(rqpair->device)) {
    3447           0 :                 return;
    3448             :         }
    3449             : 
    3450           0 :         assert(rqpair->qpair.state == SPDK_NVMF_QPAIR_ERROR);
    3451             : 
    3452           0 :         nvmf_rdma_qpair_destroy(rqpair);
    3453             : }
    3454             : 
    3455             : static int
    3456           0 : nvmf_rdma_disconnect(struct rdma_cm_event *evt, bool *event_acked)
    3457             : {
    3458             :         struct spdk_nvmf_qpair          *qpair;
    3459             :         struct spdk_nvmf_rdma_qpair     *rqpair;
    3460             : 
    3461           0 :         if (evt->id == NULL) {
    3462           0 :                 SPDK_ERRLOG("disconnect request: missing cm_id\n");
    3463           0 :                 return -1;
    3464             :         }
    3465             : 
    3466           0 :         qpair = evt->id->context;
    3467           0 :         if (qpair == NULL) {
    3468           0 :                 SPDK_ERRLOG("disconnect request: no active connection\n");
    3469           0 :                 return -1;
    3470             :         }
    3471             : 
    3472           0 :         rdma_ack_cm_event(evt);
    3473           0 :         *event_acked = true;
    3474             : 
    3475           0 :         rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair);
    3476             : 
    3477           0 :         spdk_trace_record(TRACE_RDMA_QP_DISCONNECT, 0, 0, (uintptr_t)rqpair);
    3478             : 
    3479           0 :         spdk_nvmf_qpair_disconnect(&rqpair->qpair);
    3480             : 
    3481           0 :         return 0;
    3482             : }
    3483             : 
    3484             : #ifdef DEBUG
    3485             : static const char *CM_EVENT_STR[] = {
    3486             :         "RDMA_CM_EVENT_ADDR_RESOLVED",
    3487             :         "RDMA_CM_EVENT_ADDR_ERROR",
    3488             :         "RDMA_CM_EVENT_ROUTE_RESOLVED",
    3489             :         "RDMA_CM_EVENT_ROUTE_ERROR",
    3490             :         "RDMA_CM_EVENT_CONNECT_REQUEST",
    3491             :         "RDMA_CM_EVENT_CONNECT_RESPONSE",
    3492             :         "RDMA_CM_EVENT_CONNECT_ERROR",
    3493             :         "RDMA_CM_EVENT_UNREACHABLE",
    3494             :         "RDMA_CM_EVENT_REJECTED",
    3495             :         "RDMA_CM_EVENT_ESTABLISHED",
    3496             :         "RDMA_CM_EVENT_DISCONNECTED",
    3497             :         "RDMA_CM_EVENT_DEVICE_REMOVAL",
    3498             :         "RDMA_CM_EVENT_MULTICAST_JOIN",
    3499             :         "RDMA_CM_EVENT_MULTICAST_ERROR",
    3500             :         "RDMA_CM_EVENT_ADDR_CHANGE",
    3501             :         "RDMA_CM_EVENT_TIMEWAIT_EXIT"
    3502             : };
    3503             : #endif /* DEBUG */
    3504             : 
    3505             : static void
    3506           0 : nvmf_rdma_disconnect_qpairs_on_port(struct spdk_nvmf_rdma_transport *rtransport,
    3507             :                                     struct spdk_nvmf_rdma_port *port)
    3508             : {
    3509             :         struct spdk_nvmf_rdma_poll_group        *rgroup;
    3510             :         struct spdk_nvmf_rdma_poller            *rpoller;
    3511             :         struct spdk_nvmf_rdma_qpair             *rqpair;
    3512             : 
    3513           0 :         TAILQ_FOREACH(rgroup, &rtransport->poll_groups, link) {
    3514           0 :                 TAILQ_FOREACH(rpoller, &rgroup->pollers, link) {
    3515           0 :                         RB_FOREACH(rqpair, qpairs_tree, &rpoller->qpairs) {
    3516           0 :                                 if (rqpair->listen_id == port->id) {
    3517           0 :                                         spdk_nvmf_qpair_disconnect(&rqpair->qpair);
    3518             :                                 }
    3519             :                         }
    3520             :                 }
    3521             :         }
    3522           0 : }
    3523             : 
    3524             : static bool
    3525           0 : nvmf_rdma_handle_cm_event_addr_change(struct spdk_nvmf_transport *transport,
    3526             :                                       struct rdma_cm_event *event)
    3527             : {
    3528             :         const struct spdk_nvme_transport_id     *trid;
    3529             :         struct spdk_nvmf_rdma_port              *port;
    3530             :         struct spdk_nvmf_rdma_transport         *rtransport;
    3531           0 :         bool                                    event_acked = false;
    3532             : 
    3533           0 :         rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport);
    3534           0 :         TAILQ_FOREACH(port, &rtransport->ports, link) {
    3535           0 :                 if (port->id == event->id) {
    3536           0 :                         SPDK_ERRLOG("ADDR_CHANGE: IP %s:%s migrated\n", port->trid->traddr, port->trid->trsvcid);
    3537           0 :                         rdma_ack_cm_event(event);
    3538           0 :                         event_acked = true;
    3539           0 :                         trid = port->trid;
    3540           0 :                         break;
    3541             :                 }
    3542             :         }
    3543             : 
    3544           0 :         if (event_acked) {
    3545           0 :                 nvmf_rdma_disconnect_qpairs_on_port(rtransport, port);
    3546             : 
    3547           0 :                 nvmf_rdma_stop_listen(transport, trid);
    3548           0 :                 nvmf_rdma_listen(transport, trid, NULL);
    3549             :         }
    3550             : 
    3551           0 :         return event_acked;
    3552             : }
    3553             : 
    3554             : static void
    3555           0 : nvmf_rdma_handle_device_removal(struct spdk_nvmf_rdma_transport *rtransport,
    3556             :                                 struct spdk_nvmf_rdma_device *device)
    3557             : {
    3558             :         struct spdk_nvmf_rdma_port      *port, *port_tmp;
    3559             :         int                             rc;
    3560           0 :         bool                            has_inflight;
    3561             : 
    3562           0 :         rc = nvmf_rdma_manage_poller(rtransport, device, &has_inflight, false);
    3563           0 :         if (rc) {
    3564           0 :                 SPDK_ERRLOG("Failed to handle device removal, rc %d\n", rc);
    3565           0 :                 return;
    3566             :         }
    3567             : 
    3568           0 :         if (!has_inflight) {
    3569             :                 /* no pollers, destroy the device */
    3570           0 :                 device->ready_to_destroy = true;
    3571           0 :                 spdk_thread_send_msg(spdk_get_thread(), _nvmf_rdma_remove_destroyed_device, rtransport);
    3572             :         }
    3573             : 
    3574           0 :         TAILQ_FOREACH_SAFE(port, &rtransport->ports, link, port_tmp) {
    3575           0 :                 if (port->device == device) {
    3576           0 :                         SPDK_NOTICELOG("Port %s:%s on device %s is being removed.\n",
    3577             :                                        port->trid->traddr,
    3578             :                                        port->trid->trsvcid,
    3579             :                                        ibv_get_device_name(port->device->context->device));
    3580             : 
    3581             :                         /* keep NVMF listener and only destroy structures of the
    3582             :                          * RDMA transport. when the device comes back we can retry listening
    3583             :                          * and the application's workflow will not be interrupted.
    3584             :                          */
    3585           0 :                         nvmf_rdma_stop_listen_ex(&rtransport->transport, port->trid, true);
    3586             :                 }
    3587             :         }
    3588             : }
    3589             : 
    3590             : static void
    3591           0 : nvmf_rdma_handle_cm_event_port_removal(struct spdk_nvmf_transport *transport,
    3592             :                                        struct rdma_cm_event *event)
    3593             : {
    3594             :         struct spdk_nvmf_rdma_port              *port, *tmp_port;
    3595             :         struct spdk_nvmf_rdma_transport         *rtransport;
    3596             : 
    3597           0 :         port = event->id->context;
    3598           0 :         rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport);
    3599             : 
    3600           0 :         rdma_ack_cm_event(event);
    3601             : 
    3602             :         /* if device removal happens during ctrl qpair disconnecting, it's possible that we receive
    3603             :          * an DEVICE_REMOVAL event on qpair but the id->qp is just NULL. So we should make sure that
    3604             :          * we are handling a port event here.
    3605             :          */
    3606           0 :         TAILQ_FOREACH(tmp_port, &rtransport->ports, link) {
    3607           0 :                 if (port == tmp_port && port->device && !port->device->need_destroy) {
    3608           0 :                         port->device->need_destroy = true;
    3609           0 :                         nvmf_rdma_handle_device_removal(rtransport, port->device);
    3610             :                 }
    3611             :         }
    3612           0 : }
    3613             : 
    3614             : static void
    3615           0 : nvmf_process_cm_events(struct spdk_nvmf_transport *transport, uint32_t max_events)
    3616             : {
    3617             :         struct spdk_nvmf_rdma_transport *rtransport;
    3618           0 :         struct rdma_cm_event            *event;
    3619             :         uint32_t                        i;
    3620             :         int                             rc;
    3621           0 :         bool                            event_acked;
    3622             : 
    3623           0 :         rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport);
    3624             : 
    3625           0 :         if (rtransport->event_channel == NULL) {
    3626           0 :                 return;
    3627             :         }
    3628             : 
    3629           0 :         for (i = 0; i < max_events; i++) {
    3630           0 :                 event_acked = false;
    3631           0 :                 rc = rdma_get_cm_event(rtransport->event_channel, &event);
    3632           0 :                 if (rc) {
    3633           0 :                         if (errno != EAGAIN && errno != EWOULDBLOCK) {
    3634           0 :                                 SPDK_ERRLOG("Acceptor Event Error: %s\n", spdk_strerror(errno));
    3635             :                         }
    3636           0 :                         break;
    3637             :                 }
    3638             : 
    3639           0 :                 SPDK_DEBUGLOG(rdma, "Acceptor Event: %s\n", CM_EVENT_STR[event->event]);
    3640             : 
    3641           0 :                 spdk_trace_record(TRACE_RDMA_CM_ASYNC_EVENT, 0, 0, 0, event->event);
    3642             : 
    3643           0 :                 switch (event->event) {
    3644           0 :                 case RDMA_CM_EVENT_ADDR_RESOLVED:
    3645             :                 case RDMA_CM_EVENT_ADDR_ERROR:
    3646             :                 case RDMA_CM_EVENT_ROUTE_RESOLVED:
    3647             :                 case RDMA_CM_EVENT_ROUTE_ERROR:
    3648             :                         /* No action required. The target never attempts to resolve routes. */
    3649           0 :                         break;
    3650           0 :                 case RDMA_CM_EVENT_CONNECT_REQUEST:
    3651           0 :                         rc = nvmf_rdma_connect(transport, event);
    3652           0 :                         if (rc < 0) {
    3653           0 :                                 SPDK_ERRLOG("Unable to process connect event. rc: %d\n", rc);
    3654           0 :                                 break;
    3655             :                         }
    3656           0 :                         break;
    3657           0 :                 case RDMA_CM_EVENT_CONNECT_RESPONSE:
    3658             :                         /* The target never initiates a new connection. So this will not occur. */
    3659           0 :                         break;
    3660           0 :                 case RDMA_CM_EVENT_CONNECT_ERROR:
    3661             :                         /* Can this happen? The docs say it can, but not sure what causes it. */
    3662           0 :                         break;
    3663           0 :                 case RDMA_CM_EVENT_UNREACHABLE:
    3664             :                 case RDMA_CM_EVENT_REJECTED:
    3665             :                         /* These only occur on the client side. */
    3666           0 :                         break;
    3667           0 :                 case RDMA_CM_EVENT_ESTABLISHED:
    3668             :                         /* TODO: Should we be waiting for this event anywhere? */
    3669           0 :                         break;
    3670           0 :                 case RDMA_CM_EVENT_DISCONNECTED:
    3671           0 :                         rc = nvmf_rdma_disconnect(event, &event_acked);
    3672           0 :                         if (rc < 0) {
    3673           0 :                                 SPDK_ERRLOG("Unable to process disconnect event. rc: %d\n", rc);
    3674           0 :                                 break;
    3675             :                         }
    3676           0 :                         break;
    3677           0 :                 case RDMA_CM_EVENT_DEVICE_REMOVAL:
    3678             :                         /* In case of device removal, kernel IB part triggers IBV_EVENT_DEVICE_FATAL
    3679             :                          * which triggers RDMA_CM_EVENT_DEVICE_REMOVAL on all cma_id’s.
    3680             :                          * Once these events are sent to SPDK, we should release all IB resources and
    3681             :                          * don't make attempts to call any ibv_query/modify/create functions. We can only call
    3682             :                          * ibv_destroy* functions to release user space memory allocated by IB. All kernel
    3683             :                          * resources are already cleaned. */
    3684           0 :                         if (event->id->qp) {
    3685             :                                 /* If rdma_cm event has a valid `qp` pointer then the event refers to the
    3686             :                                  * corresponding qpair. Otherwise the event refers to a listening device. */
    3687           0 :                                 rc = nvmf_rdma_disconnect(event, &event_acked);
    3688           0 :                                 if (rc < 0) {
    3689           0 :                                         SPDK_ERRLOG("Unable to process disconnect event. rc: %d\n", rc);
    3690           0 :                                         break;
    3691             :                                 }
    3692             :                         } else {
    3693           0 :                                 nvmf_rdma_handle_cm_event_port_removal(transport, event);
    3694           0 :                                 event_acked = true;
    3695             :                         }
    3696           0 :                         break;
    3697           0 :                 case RDMA_CM_EVENT_MULTICAST_JOIN:
    3698             :                 case RDMA_CM_EVENT_MULTICAST_ERROR:
    3699             :                         /* Multicast is not used */
    3700           0 :                         break;
    3701           0 :                 case RDMA_CM_EVENT_ADDR_CHANGE:
    3702           0 :                         event_acked = nvmf_rdma_handle_cm_event_addr_change(transport, event);
    3703           0 :                         break;
    3704           0 :                 case RDMA_CM_EVENT_TIMEWAIT_EXIT:
    3705             :                         /* For now, do nothing. The target never re-uses queue pairs. */
    3706           0 :                         break;
    3707           0 :                 default:
    3708           0 :                         SPDK_ERRLOG("Unexpected Acceptor Event [%d]\n", event->event);
    3709           0 :                         break;
    3710             :                 }
    3711           0 :                 if (!event_acked) {
    3712           0 :                         rdma_ack_cm_event(event);
    3713             :                 }
    3714             :         }
    3715             : }
    3716             : 
    3717             : static void
    3718           0 : nvmf_rdma_handle_last_wqe_reached(struct spdk_nvmf_rdma_qpair *rqpair)
    3719             : {
    3720           0 :         rqpair->last_wqe_reached = true;
    3721           0 :         nvmf_rdma_destroy_drained_qpair(rqpair);
    3722           0 : }
    3723             : 
    3724             : static void
    3725           0 : nvmf_rdma_qpair_process_ibv_event(void *ctx)
    3726             : {
    3727           0 :         struct spdk_nvmf_rdma_ibv_event_ctx *event_ctx = ctx;
    3728             : 
    3729           0 :         if (event_ctx->rqpair) {
    3730           0 :                 STAILQ_REMOVE(&event_ctx->rqpair->ibv_events, event_ctx, spdk_nvmf_rdma_ibv_event_ctx, link);
    3731           0 :                 if (event_ctx->cb_fn) {
    3732           0 :                         event_ctx->cb_fn(event_ctx->rqpair);
    3733             :                 }
    3734             :         }
    3735           0 :         free(event_ctx);
    3736           0 : }
    3737             : 
    3738             : static int
    3739           0 : nvmf_rdma_send_qpair_async_event(struct spdk_nvmf_rdma_qpair *rqpair,
    3740             :                                  spdk_nvmf_rdma_qpair_ibv_event fn)
    3741             : {
    3742             :         struct spdk_nvmf_rdma_ibv_event_ctx *ctx;
    3743           0 :         struct spdk_thread *thr = NULL;
    3744             :         int rc;
    3745             : 
    3746           0 :         if (rqpair->qpair.group) {
    3747           0 :                 thr = rqpair->qpair.group->thread;
    3748           0 :         } else if (rqpair->destruct_channel) {
    3749           0 :                 thr = spdk_io_channel_get_thread(rqpair->destruct_channel);
    3750             :         }
    3751             : 
    3752           0 :         if (!thr) {
    3753           0 :                 SPDK_DEBUGLOG(rdma, "rqpair %p has no thread\n", rqpair);
    3754           0 :                 return -EINVAL;
    3755             :         }
    3756             : 
    3757           0 :         ctx = calloc(1, sizeof(*ctx));
    3758           0 :         if (!ctx) {
    3759           0 :                 return -ENOMEM;
    3760             :         }
    3761             : 
    3762           0 :         ctx->rqpair = rqpair;
    3763           0 :         ctx->cb_fn = fn;
    3764           0 :         STAILQ_INSERT_TAIL(&rqpair->ibv_events, ctx, link);
    3765             : 
    3766           0 :         rc = spdk_thread_send_msg(thr, nvmf_rdma_qpair_process_ibv_event, ctx);
    3767           0 :         if (rc) {
    3768           0 :                 STAILQ_REMOVE(&rqpair->ibv_events, ctx, spdk_nvmf_rdma_ibv_event_ctx, link);
    3769           0 :                 free(ctx);
    3770             :         }
    3771             : 
    3772           0 :         return rc;
    3773             : }
    3774             : 
    3775             : static int
    3776           0 : nvmf_process_ib_event(struct spdk_nvmf_rdma_device *device)
    3777             : {
    3778             :         int                             rc;
    3779           0 :         struct spdk_nvmf_rdma_qpair     *rqpair = NULL;
    3780           0 :         struct ibv_async_event          event;
    3781             : 
    3782           0 :         rc = ibv_get_async_event(device->context, &event);
    3783             : 
    3784           0 :         if (rc) {
    3785             :                 /* In non-blocking mode -1 means there are no events available */
    3786           0 :                 return rc;
    3787             :         }
    3788             : 
    3789           0 :         switch (event.event_type) {
    3790           0 :         case IBV_EVENT_QP_FATAL:
    3791             :         case IBV_EVENT_QP_LAST_WQE_REACHED:
    3792             :         case IBV_EVENT_QP_REQ_ERR:
    3793             :         case IBV_EVENT_QP_ACCESS_ERR:
    3794             :         case IBV_EVENT_COMM_EST:
    3795             :         case IBV_EVENT_PATH_MIG:
    3796             :         case IBV_EVENT_PATH_MIG_ERR:
    3797           0 :                 rqpair = event.element.qp->qp_context;
    3798           0 :                 if (!rqpair) {
    3799             :                         /* Any QP event for NVMe-RDMA initiator may be returned. */
    3800           0 :                         SPDK_NOTICELOG("Async QP event for unknown QP: %s\n",
    3801             :                                        ibv_event_type_str(event.event_type));
    3802           0 :                         break;
    3803             :                 }
    3804             : 
    3805           0 :                 switch (event.event_type) {
    3806           0 :                 case IBV_EVENT_QP_FATAL:
    3807           0 :                         SPDK_ERRLOG("Fatal event received for rqpair %p\n", rqpair);
    3808           0 :                         spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0,
    3809             :                                           (uintptr_t)rqpair, event.event_type);
    3810           0 :                         rqpair->ibv_in_error_state = true;
    3811           0 :                         spdk_nvmf_qpair_disconnect(&rqpair->qpair);
    3812           0 :                         break;
    3813           0 :                 case IBV_EVENT_QP_LAST_WQE_REACHED:
    3814             :                         /* This event only occurs for shared receive queues. */
    3815           0 :                         SPDK_DEBUGLOG(rdma, "Last WQE reached event received for rqpair %p\n", rqpair);
    3816           0 :                         rc = nvmf_rdma_send_qpair_async_event(rqpair, nvmf_rdma_handle_last_wqe_reached);
    3817           0 :                         if (rc) {
    3818           0 :                                 SPDK_WARNLOG("Failed to send LAST_WQE_REACHED event. rqpair %p, err %d\n", rqpair, rc);
    3819           0 :                                 rqpair->last_wqe_reached = true;
    3820             :                         }
    3821           0 :                         break;
    3822           0 :                 case IBV_EVENT_QP_REQ_ERR:
    3823             :                 case IBV_EVENT_QP_ACCESS_ERR:
    3824             :                 case IBV_EVENT_COMM_EST:
    3825             :                 case IBV_EVENT_PATH_MIG:
    3826             :                 case IBV_EVENT_PATH_MIG_ERR:
    3827           0 :                         SPDK_NOTICELOG("Async QP event: %s\n",
    3828             :                                        ibv_event_type_str(event.event_type));
    3829           0 :                         spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0,
    3830             :                                           (uintptr_t)rqpair, event.event_type);
    3831           0 :                         rqpair->ibv_in_error_state = true;
    3832           0 :                         break;
    3833           0 :                 default:
    3834           0 :                         break;
    3835             :                 }
    3836           0 :                 break;
    3837           0 :         case IBV_EVENT_DEVICE_FATAL:
    3838           0 :                 SPDK_ERRLOG("Device Fatal event[%s] received on %s. device: %p\n",
    3839             :                             ibv_event_type_str(event.event_type), ibv_get_device_name(device->context->device), device);
    3840           0 :                 device->need_destroy = true;
    3841           0 :                 break;
    3842           0 :         case IBV_EVENT_CQ_ERR:
    3843             :         case IBV_EVENT_PORT_ACTIVE:
    3844             :         case IBV_EVENT_PORT_ERR:
    3845             :         case IBV_EVENT_LID_CHANGE:
    3846             :         case IBV_EVENT_PKEY_CHANGE:
    3847             :         case IBV_EVENT_SM_CHANGE:
    3848             :         case IBV_EVENT_SRQ_ERR:
    3849             :         case IBV_EVENT_SRQ_LIMIT_REACHED:
    3850             :         case IBV_EVENT_CLIENT_REREGISTER:
    3851             :         case IBV_EVENT_GID_CHANGE:
    3852             :         case IBV_EVENT_SQ_DRAINED:
    3853             :         default:
    3854           0 :                 SPDK_NOTICELOG("Async event: %s\n",
    3855             :                                ibv_event_type_str(event.event_type));
    3856           0 :                 spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0, 0, event.event_type);
    3857           0 :                 break;
    3858             :         }
    3859           0 :         ibv_ack_async_event(&event);
    3860             : 
    3861           0 :         return 0;
    3862             : }
    3863             : 
    3864             : static void
    3865           0 : nvmf_process_ib_events(struct spdk_nvmf_rdma_device *device, uint32_t max_events)
    3866             : {
    3867           0 :         int rc = 0;
    3868           0 :         uint32_t i = 0;
    3869             : 
    3870           0 :         for (i = 0; i < max_events; i++) {
    3871           0 :                 rc = nvmf_process_ib_event(device);
    3872           0 :                 if (rc) {
    3873           0 :                         break;
    3874             :                 }
    3875             :         }
    3876             : 
    3877           0 :         SPDK_DEBUGLOG(rdma, "Device %s: %u events processed\n", device->context->device->name, i);
    3878           0 : }
    3879             : 
    3880             : static int
    3881           0 : nvmf_rdma_accept(void *ctx)
    3882             : {
    3883           0 :         int     nfds, i = 0;
    3884           0 :         struct spdk_nvmf_transport *transport = ctx;
    3885             :         struct spdk_nvmf_rdma_transport *rtransport;
    3886             :         struct spdk_nvmf_rdma_device *device, *tmp;
    3887             :         uint32_t count;
    3888             :         short revents;
    3889             :         bool do_retry;
    3890             : 
    3891           0 :         rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport);
    3892           0 :         do_retry = nvmf_rdma_retry_listen_port(rtransport);
    3893             : 
    3894           0 :         count = nfds = poll(rtransport->poll_fds, rtransport->npoll_fds, 0);
    3895             : 
    3896           0 :         if (nfds <= 0) {
    3897           0 :                 return do_retry ? SPDK_POLLER_BUSY : SPDK_POLLER_IDLE;
    3898             :         }
    3899             : 
    3900             :         /* The first poll descriptor is RDMA CM event */
    3901           0 :         if (rtransport->poll_fds[i++].revents & POLLIN) {
    3902           0 :                 nvmf_process_cm_events(transport, NVMF_RDMA_MAX_EVENTS_PER_POLL);
    3903           0 :                 nfds--;
    3904             :         }
    3905             : 
    3906           0 :         if (nfds == 0) {
    3907           0 :                 return SPDK_POLLER_BUSY;
    3908             :         }
    3909             : 
    3910             :         /* Second and subsequent poll descriptors are IB async events */
    3911           0 :         TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, tmp) {
    3912           0 :                 revents = rtransport->poll_fds[i++].revents;
    3913           0 :                 if (revents & POLLIN) {
    3914           0 :                         if (spdk_likely(!device->need_destroy)) {
    3915           0 :                                 nvmf_process_ib_events(device, NVMF_RDMA_MAX_EVENTS_PER_POLL);
    3916           0 :                                 if (spdk_unlikely(device->need_destroy)) {
    3917           0 :                                         nvmf_rdma_handle_device_removal(rtransport, device);
    3918             :                                 }
    3919             :                         }
    3920           0 :                         nfds--;
    3921           0 :                 } else if (revents & POLLNVAL || revents & POLLHUP) {
    3922           0 :                         SPDK_ERRLOG("Receive unknown revent %x on device %p\n", (int)revents, device);
    3923           0 :                         nfds--;
    3924             :                 }
    3925             :         }
    3926             :         /* check all flagged fd's have been served */
    3927           0 :         assert(nfds == 0);
    3928             : 
    3929           0 :         return count > 0 ? SPDK_POLLER_BUSY : SPDK_POLLER_IDLE;
    3930             : }
    3931             : 
    3932             : static void
    3933           0 : nvmf_rdma_cdata_init(struct spdk_nvmf_transport *transport, struct spdk_nvmf_subsystem *subsystem,
    3934             :                      struct spdk_nvmf_ctrlr_data *cdata)
    3935             : {
    3936           0 :         cdata->nvmf_specific.msdbd = NVMF_DEFAULT_MSDBD;
    3937             : 
    3938             :         /* Disable in-capsule data transfer for RDMA controller when dif_insert_or_strip is enabled
    3939             :         since in-capsule data only works with NVME drives that support SGL memory layout */
    3940           0 :         if (transport->opts.dif_insert_or_strip) {
    3941           0 :                 cdata->nvmf_specific.ioccsz = sizeof(struct spdk_nvme_cmd) / 16;
    3942             :         }
    3943             : 
    3944           0 :         if (cdata->nvmf_specific.ioccsz > ((sizeof(struct spdk_nvme_cmd) + 0x1000) / 16)) {
    3945           0 :                 SPDK_WARNLOG("RDMA is configured to support up to 16 SGL entries while in capsule"
    3946             :                              " data is greater than 4KiB.\n");
    3947           0 :                 SPDK_WARNLOG("When used in conjunction with the NVMe-oF initiator from the Linux "
    3948             :                              "kernel between versions 5.4 and 5.12 data corruption may occur for "
    3949             :                              "writes that are not a multiple of 4KiB in size.\n");
    3950             :         }
    3951           0 : }
    3952             : 
    3953             : static void
    3954           0 : nvmf_rdma_discover(struct spdk_nvmf_transport *transport,
    3955             :                    struct spdk_nvme_transport_id *trid,
    3956             :                    struct spdk_nvmf_discovery_log_page_entry *entry)
    3957             : {
    3958           0 :         entry->trtype = SPDK_NVMF_TRTYPE_RDMA;
    3959           0 :         entry->adrfam = trid->adrfam;
    3960           0 :         entry->treq.secure_channel = SPDK_NVMF_TREQ_SECURE_CHANNEL_NOT_REQUIRED;
    3961             : 
    3962           0 :         spdk_strcpy_pad(entry->trsvcid, trid->trsvcid, sizeof(entry->trsvcid), ' ');
    3963           0 :         spdk_strcpy_pad(entry->traddr, trid->traddr, sizeof(entry->traddr), ' ');
    3964             : 
    3965           0 :         entry->tsas.rdma.rdma_qptype = SPDK_NVMF_RDMA_QPTYPE_RELIABLE_CONNECTED;
    3966           0 :         entry->tsas.rdma.rdma_prtype = SPDK_NVMF_RDMA_PRTYPE_NONE;
    3967           0 :         entry->tsas.rdma.rdma_cms = SPDK_NVMF_RDMA_CMS_RDMA_CM;
    3968           0 : }
    3969             : 
    3970             : static int
    3971           0 : nvmf_rdma_poller_create(struct spdk_nvmf_rdma_transport *rtransport,
    3972             :                         struct spdk_nvmf_rdma_poll_group *rgroup, struct spdk_nvmf_rdma_device *device,
    3973             :                         struct spdk_nvmf_rdma_poller **out_poller)
    3974             : {
    3975             :         struct spdk_nvmf_rdma_poller            *poller;
    3976           0 :         struct spdk_rdma_srq_init_attr          srq_init_attr;
    3977           0 :         struct spdk_nvmf_rdma_resource_opts     opts;
    3978             :         int                                     num_cqe;
    3979             : 
    3980           0 :         poller = calloc(1, sizeof(*poller));
    3981           0 :         if (!poller) {
    3982           0 :                 SPDK_ERRLOG("Unable to allocate memory for new RDMA poller\n");
    3983           0 :                 return -1;
    3984             :         }
    3985             : 
    3986           0 :         poller->device = device;
    3987           0 :         poller->group = rgroup;
    3988           0 :         *out_poller = poller;
    3989             : 
    3990           0 :         RB_INIT(&poller->qpairs);
    3991           0 :         STAILQ_INIT(&poller->qpairs_pending_send);
    3992           0 :         STAILQ_INIT(&poller->qpairs_pending_recv);
    3993             : 
    3994           0 :         TAILQ_INSERT_TAIL(&rgroup->pollers, poller, link);
    3995           0 :         SPDK_DEBUGLOG(rdma, "Create poller %p on device %p in poll group %p.\n", poller, device, rgroup);
    3996           0 :         if (rtransport->rdma_opts.no_srq == false && device->num_srq < device->attr.max_srq) {
    3997           0 :                 if ((int)rtransport->rdma_opts.max_srq_depth > device->attr.max_srq_wr) {
    3998           0 :                         SPDK_WARNLOG("Requested SRQ depth %u, max supported by dev %s is %d\n",
    3999             :                                      rtransport->rdma_opts.max_srq_depth, device->context->device->name, device->attr.max_srq_wr);
    4000             :                 }
    4001           0 :                 poller->max_srq_depth = spdk_min((int)rtransport->rdma_opts.max_srq_depth, device->attr.max_srq_wr);
    4002             : 
    4003           0 :                 device->num_srq++;
    4004           0 :                 memset(&srq_init_attr, 0, sizeof(srq_init_attr));
    4005           0 :                 srq_init_attr.pd = device->pd;
    4006           0 :                 srq_init_attr.stats = &poller->stat.qp_stats.recv;
    4007           0 :                 srq_init_attr.srq_init_attr.attr.max_wr = poller->max_srq_depth;
    4008           0 :                 srq_init_attr.srq_init_attr.attr.max_sge = spdk_min(device->attr.max_sge, NVMF_DEFAULT_RX_SGE);
    4009           0 :                 poller->srq = spdk_rdma_srq_create(&srq_init_attr);
    4010           0 :                 if (!poller->srq) {
    4011           0 :                         SPDK_ERRLOG("Unable to create shared receive queue, errno %d\n", errno);
    4012           0 :                         return -1;
    4013             :                 }
    4014             : 
    4015           0 :                 opts.qp = poller->srq;
    4016           0 :                 opts.map = device->map;
    4017           0 :                 opts.qpair = NULL;
    4018           0 :                 opts.shared = true;
    4019           0 :                 opts.max_queue_depth = poller->max_srq_depth;
    4020           0 :                 opts.in_capsule_data_size = rtransport->transport.opts.in_capsule_data_size;
    4021             : 
    4022           0 :                 poller->resources = nvmf_rdma_resources_create(&opts);
    4023           0 :                 if (!poller->resources) {
    4024           0 :                         SPDK_ERRLOG("Unable to allocate resources for shared receive queue.\n");
    4025           0 :                         return -1;
    4026             :                 }
    4027             :         }
    4028             : 
    4029             :         /*
    4030             :          * When using an srq, we can limit the completion queue at startup.
    4031             :          * The following formula represents the calculation:
    4032             :          * num_cqe = num_recv + num_data_wr + num_send_wr.
    4033             :          * where num_recv=num_data_wr=and num_send_wr=poller->max_srq_depth
    4034             :          */
    4035           0 :         if (poller->srq) {
    4036           0 :                 num_cqe = poller->max_srq_depth * 3;
    4037             :         } else {
    4038           0 :                 num_cqe = rtransport->rdma_opts.num_cqe;
    4039             :         }
    4040             : 
    4041           0 :         poller->cq = ibv_create_cq(device->context, num_cqe, poller, NULL, 0);
    4042           0 :         if (!poller->cq) {
    4043           0 :                 SPDK_ERRLOG("Unable to create completion queue\n");
    4044           0 :                 return -1;
    4045             :         }
    4046           0 :         poller->num_cqe = num_cqe;
    4047           0 :         return 0;
    4048             : }
    4049             : 
    4050             : static void
    4051           0 : _nvmf_rdma_register_poller_in_group(void *c)
    4052             : {
    4053           0 :         struct spdk_nvmf_rdma_poller    *poller;
    4054           0 :         struct poller_manage_ctx        *ctx = c;
    4055             :         struct spdk_nvmf_rdma_device    *device;
    4056             :         int                             rc;
    4057             : 
    4058           0 :         rc = nvmf_rdma_poller_create(ctx->rtransport, ctx->rgroup, ctx->device, &poller);
    4059           0 :         if (rc < 0 && poller) {
    4060           0 :                 nvmf_rdma_poller_destroy(poller);
    4061             :         }
    4062             : 
    4063           0 :         device = ctx->device;
    4064           0 :         if (nvmf_rdma_all_pollers_management_done(ctx)) {
    4065           0 :                 device->is_ready = true;
    4066             :         }
    4067           0 : }
    4068             : 
    4069             : static void nvmf_rdma_poll_group_destroy(struct spdk_nvmf_transport_poll_group *group);
    4070             : 
    4071             : static struct spdk_nvmf_transport_poll_group *
    4072           5 : nvmf_rdma_poll_group_create(struct spdk_nvmf_transport *transport,
    4073             :                             struct spdk_nvmf_poll_group *group)
    4074             : {
    4075             :         struct spdk_nvmf_rdma_transport         *rtransport;
    4076             :         struct spdk_nvmf_rdma_poll_group        *rgroup;
    4077           5 :         struct spdk_nvmf_rdma_poller            *poller;
    4078             :         struct spdk_nvmf_rdma_device            *device;
    4079             :         int                                     rc;
    4080             : 
    4081           5 :         rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport);
    4082             : 
    4083           5 :         rgroup = calloc(1, sizeof(*rgroup));
    4084           5 :         if (!rgroup) {
    4085           0 :                 return NULL;
    4086             :         }
    4087             : 
    4088           5 :         TAILQ_INIT(&rgroup->pollers);
    4089             : 
    4090           5 :         TAILQ_FOREACH(device, &rtransport->devices, link) {
    4091           0 :                 rc = nvmf_rdma_poller_create(rtransport, rgroup, device, &poller);
    4092           0 :                 if (rc < 0) {
    4093           0 :                         nvmf_rdma_poll_group_destroy(&rgroup->group);
    4094           0 :                         return NULL;
    4095             :                 }
    4096             :         }
    4097             : 
    4098           5 :         TAILQ_INSERT_TAIL(&rtransport->poll_groups, rgroup, link);
    4099           5 :         if (rtransport->conn_sched.next_admin_pg == NULL) {
    4100           1 :                 rtransport->conn_sched.next_admin_pg = rgroup;
    4101           1 :                 rtransport->conn_sched.next_io_pg = rgroup;
    4102             :         }
    4103             : 
    4104           5 :         return &rgroup->group;
    4105             : }
    4106             : 
    4107             : static uint32_t
    4108          12 : nvmf_poll_group_get_io_qpair_count(struct spdk_nvmf_poll_group *pg)
    4109             : {
    4110             :         uint32_t count;
    4111             : 
    4112             :         /* Just assume that unassociated qpairs will eventually be io
    4113             :          * qpairs.  This is close enough for the use cases for this
    4114             :          * function.
    4115             :          */
    4116          12 :         pthread_mutex_lock(&pg->mutex);
    4117          12 :         count = pg->stat.current_io_qpairs + pg->current_unassociated_qpairs;
    4118          12 :         pthread_mutex_unlock(&pg->mutex);
    4119             : 
    4120          12 :         return count;
    4121             : }
    4122             : 
    4123             : static struct spdk_nvmf_transport_poll_group *
    4124          14 : nvmf_rdma_get_optimal_poll_group(struct spdk_nvmf_qpair *qpair)
    4125             : {
    4126             :         struct spdk_nvmf_rdma_transport *rtransport;
    4127             :         struct spdk_nvmf_rdma_poll_group **pg;
    4128             :         struct spdk_nvmf_transport_poll_group *result;
    4129             :         uint32_t count;
    4130             : 
    4131          14 :         rtransport = SPDK_CONTAINEROF(qpair->transport, struct spdk_nvmf_rdma_transport, transport);
    4132             : 
    4133          14 :         if (TAILQ_EMPTY(&rtransport->poll_groups)) {
    4134           2 :                 return NULL;
    4135             :         }
    4136             : 
    4137          12 :         if (qpair->qid == 0) {
    4138           6 :                 pg = &rtransport->conn_sched.next_admin_pg;
    4139             :         } else {
    4140             :                 struct spdk_nvmf_rdma_poll_group *pg_min, *pg_start, *pg_current;
    4141             :                 uint32_t min_value;
    4142             : 
    4143           6 :                 pg = &rtransport->conn_sched.next_io_pg;
    4144           6 :                 pg_min = *pg;
    4145           6 :                 pg_start = *pg;
    4146           6 :                 pg_current = *pg;
    4147           6 :                 min_value = nvmf_poll_group_get_io_qpair_count(pg_current->group.group);
    4148             : 
    4149             :                 while (1) {
    4150           6 :                         count = nvmf_poll_group_get_io_qpair_count(pg_current->group.group);
    4151             : 
    4152           6 :                         if (count < min_value) {
    4153           0 :                                 min_value = count;
    4154           0 :                                 pg_min = pg_current;
    4155             :                         }
    4156             : 
    4157           6 :                         pg_current = TAILQ_NEXT(pg_current, link);
    4158           6 :                         if (pg_current == NULL) {
    4159           2 :                                 pg_current = TAILQ_FIRST(&rtransport->poll_groups);
    4160             :                         }
    4161             : 
    4162           6 :                         if (pg_current == pg_start || min_value == 0) {
    4163             :                                 break;
    4164             :                         }
    4165             :                 }
    4166           6 :                 *pg = pg_min;
    4167             :         }
    4168             : 
    4169          12 :         assert(*pg != NULL);
    4170             : 
    4171          12 :         result = &(*pg)->group;
    4172             : 
    4173          12 :         *pg = TAILQ_NEXT(*pg, link);
    4174          12 :         if (*pg == NULL) {
    4175           4 :                 *pg = TAILQ_FIRST(&rtransport->poll_groups);
    4176             :         }
    4177             : 
    4178          12 :         return result;
    4179             : }
    4180             : 
    4181             : static void
    4182           0 : nvmf_rdma_poller_destroy(struct spdk_nvmf_rdma_poller *poller)
    4183             : {
    4184             :         struct spdk_nvmf_rdma_qpair     *qpair, *tmp_qpair;
    4185             :         int                             rc;
    4186             : 
    4187           0 :         TAILQ_REMOVE(&poller->group->pollers, poller, link);
    4188           0 :         RB_FOREACH_SAFE(qpair, qpairs_tree, &poller->qpairs, tmp_qpair) {
    4189           0 :                 nvmf_rdma_qpair_destroy(qpair);
    4190             :         }
    4191             : 
    4192           0 :         if (poller->srq) {
    4193           0 :                 if (poller->resources) {
    4194           0 :                         nvmf_rdma_resources_destroy(poller->resources);
    4195             :                 }
    4196           0 :                 spdk_rdma_srq_destroy(poller->srq);
    4197           0 :                 SPDK_DEBUGLOG(rdma, "Destroyed RDMA shared queue %p\n", poller->srq);
    4198             :         }
    4199             : 
    4200           0 :         if (poller->cq) {
    4201           0 :                 rc = ibv_destroy_cq(poller->cq);
    4202           0 :                 if (rc != 0) {
    4203           0 :                         SPDK_ERRLOG("Destroy cq return %d, error: %s\n", rc, strerror(errno));
    4204             :                 }
    4205             :         }
    4206             : 
    4207           0 :         if (poller->destroy_cb) {
    4208           0 :                 poller->destroy_cb(poller->destroy_cb_ctx);
    4209           0 :                 poller->destroy_cb = NULL;
    4210             :         }
    4211             : 
    4212           0 :         free(poller);
    4213           0 : }
    4214             : 
    4215             : static void
    4216           5 : nvmf_rdma_poll_group_destroy(struct spdk_nvmf_transport_poll_group *group)
    4217             : {
    4218             :         struct spdk_nvmf_rdma_poll_group        *rgroup, *next_rgroup;
    4219             :         struct spdk_nvmf_rdma_poller            *poller, *tmp;
    4220             :         struct spdk_nvmf_rdma_transport         *rtransport;
    4221             : 
    4222           5 :         rgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_rdma_poll_group, group);
    4223           5 :         if (!rgroup) {
    4224           0 :                 return;
    4225             :         }
    4226             : 
    4227           5 :         TAILQ_FOREACH_SAFE(poller, &rgroup->pollers, link, tmp) {
    4228           0 :                 nvmf_rdma_poller_destroy(poller);
    4229             :         }
    4230             : 
    4231           5 :         if (rgroup->group.transport == NULL) {
    4232             :                 /* Transport can be NULL when nvmf_rdma_poll_group_create()
    4233             :                  * calls this function directly in a failure path. */
    4234           0 :                 free(rgroup);
    4235           0 :                 return;
    4236             :         }
    4237             : 
    4238           5 :         rtransport = SPDK_CONTAINEROF(rgroup->group.transport, struct spdk_nvmf_rdma_transport, transport);
    4239             : 
    4240           5 :         next_rgroup = TAILQ_NEXT(rgroup, link);
    4241           5 :         TAILQ_REMOVE(&rtransport->poll_groups, rgroup, link);
    4242           5 :         if (next_rgroup == NULL) {
    4243           1 :                 next_rgroup = TAILQ_FIRST(&rtransport->poll_groups);
    4244             :         }
    4245           5 :         if (rtransport->conn_sched.next_admin_pg == rgroup) {
    4246           5 :                 rtransport->conn_sched.next_admin_pg = next_rgroup;
    4247             :         }
    4248           5 :         if (rtransport->conn_sched.next_io_pg == rgroup) {
    4249           5 :                 rtransport->conn_sched.next_io_pg = next_rgroup;
    4250             :         }
    4251             : 
    4252           5 :         free(rgroup);
    4253             : }
    4254             : 
    4255             : static void
    4256           0 : nvmf_rdma_qpair_reject_connection(struct spdk_nvmf_rdma_qpair *rqpair)
    4257             : {
    4258           0 :         if (rqpair->cm_id != NULL) {
    4259           0 :                 nvmf_rdma_event_reject(rqpair->cm_id, SPDK_NVMF_RDMA_ERROR_NO_RESOURCES);
    4260             :         }
    4261           0 : }
    4262             : 
    4263             : static int
    4264           0 : nvmf_rdma_poll_group_add(struct spdk_nvmf_transport_poll_group *group,
    4265             :                          struct spdk_nvmf_qpair *qpair)
    4266             : {
    4267             :         struct spdk_nvmf_rdma_poll_group        *rgroup;
    4268             :         struct spdk_nvmf_rdma_qpair             *rqpair;
    4269             :         struct spdk_nvmf_rdma_device            *device;
    4270             :         struct spdk_nvmf_rdma_poller            *poller;
    4271             :         int                                     rc;
    4272             : 
    4273           0 :         rgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_rdma_poll_group, group);
    4274           0 :         rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair);
    4275             : 
    4276           0 :         device = rqpair->device;
    4277             : 
    4278           0 :         TAILQ_FOREACH(poller, &rgroup->pollers, link) {
    4279           0 :                 if (poller->device == device) {
    4280           0 :                         break;
    4281             :                 }
    4282             :         }
    4283             : 
    4284           0 :         if (!poller) {
    4285           0 :                 SPDK_ERRLOG("No poller found for device.\n");
    4286           0 :                 return -1;
    4287             :         }
    4288             : 
    4289           0 :         if (poller->need_destroy) {
    4290           0 :                 SPDK_ERRLOG("Poller is destroying.\n");
    4291           0 :                 return -1;
    4292             :         }
    4293             : 
    4294           0 :         rqpair->poller = poller;
    4295           0 :         rqpair->srq = rqpair->poller->srq;
    4296             : 
    4297           0 :         rc = nvmf_rdma_qpair_initialize(qpair);
    4298           0 :         if (rc < 0) {
    4299           0 :                 SPDK_ERRLOG("Failed to initialize nvmf_rdma_qpair with qpair=%p\n", qpair);
    4300           0 :                 rqpair->poller = NULL;
    4301           0 :                 rqpair->srq = NULL;
    4302           0 :                 return -1;
    4303             :         }
    4304             : 
    4305           0 :         RB_INSERT(qpairs_tree, &poller->qpairs, rqpair);
    4306             : 
    4307           0 :         rc = nvmf_rdma_event_accept(rqpair->cm_id, rqpair);
    4308           0 :         if (rc) {
    4309             :                 /* Try to reject, but we probably can't */
    4310           0 :                 nvmf_rdma_qpair_reject_connection(rqpair);
    4311           0 :                 return -1;
    4312             :         }
    4313             : 
    4314           0 :         return 0;
    4315             : }
    4316             : 
    4317             : static int
    4318           0 : nvmf_rdma_poll_group_remove(struct spdk_nvmf_transport_poll_group *group,
    4319             :                             struct spdk_nvmf_qpair *qpair)
    4320             : {
    4321             :         struct spdk_nvmf_rdma_qpair             *rqpair;
    4322             : 
    4323           0 :         rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair);
    4324           0 :         assert(group->transport->tgt != NULL);
    4325             : 
    4326           0 :         rqpair->destruct_channel = spdk_get_io_channel(group->transport->tgt);
    4327             : 
    4328           0 :         if (!rqpair->destruct_channel) {
    4329           0 :                 SPDK_WARNLOG("failed to get io_channel, qpair %p\n", qpair);
    4330           0 :                 return 0;
    4331             :         }
    4332             : 
    4333             :         /* Sanity check that we get io_channel on the correct thread */
    4334           0 :         if (qpair->group) {
    4335           0 :                 assert(qpair->group->thread == spdk_io_channel_get_thread(rqpair->destruct_channel));
    4336             :         }
    4337             : 
    4338           0 :         return 0;
    4339             : }
    4340             : 
    4341             : static int
    4342           0 : nvmf_rdma_request_free(struct spdk_nvmf_request *req)
    4343             : {
    4344           0 :         struct spdk_nvmf_rdma_request   *rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req);
    4345           0 :         struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(req->qpair->transport,
    4346             :                         struct spdk_nvmf_rdma_transport, transport);
    4347           0 :         struct spdk_nvmf_rdma_qpair *rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair,
    4348             :                                               struct spdk_nvmf_rdma_qpair, qpair);
    4349             : 
    4350             :         /*
    4351             :          * AER requests are freed when a qpair is destroyed. The recv corresponding to that request
    4352             :          * needs to be returned to the shared receive queue or the poll group will eventually be
    4353             :          * starved of RECV structures.
    4354             :          */
    4355           0 :         if (rqpair->srq && rdma_req->recv) {
    4356             :                 int rc;
    4357           0 :                 struct ibv_recv_wr *bad_recv_wr;
    4358             : 
    4359           0 :                 spdk_rdma_srq_queue_recv_wrs(rqpair->srq, &rdma_req->recv->wr);
    4360           0 :                 rc = spdk_rdma_srq_flush_recv_wrs(rqpair->srq, &bad_recv_wr);
    4361           0 :                 if (rc) {
    4362           0 :                         SPDK_ERRLOG("Unable to re-post rx descriptor\n");
    4363             :                 }
    4364             :         }
    4365             : 
    4366           0 :         _nvmf_rdma_request_free(rdma_req, rtransport);
    4367           0 :         return 0;
    4368             : }
    4369             : 
    4370             : static int
    4371           0 : nvmf_rdma_request_complete(struct spdk_nvmf_request *req)
    4372             : {
    4373           0 :         struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(req->qpair->transport,
    4374             :                         struct spdk_nvmf_rdma_transport, transport);
    4375           0 :         struct spdk_nvmf_rdma_request   *rdma_req = SPDK_CONTAINEROF(req,
    4376             :                         struct spdk_nvmf_rdma_request, req);
    4377           0 :         struct spdk_nvmf_rdma_qpair     *rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair,
    4378             :                         struct spdk_nvmf_rdma_qpair, qpair);
    4379             : 
    4380           0 :         if (spdk_unlikely(rqpair->ibv_in_error_state)) {
    4381             :                 /* The connection is dead. Move the request directly to the completed state. */
    4382           0 :                 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED;
    4383             :         } else {
    4384             :                 /* The connection is alive, so process the request as normal */
    4385           0 :                 rdma_req->state = RDMA_REQUEST_STATE_EXECUTED;
    4386             :         }
    4387             : 
    4388           0 :         nvmf_rdma_request_process(rtransport, rdma_req);
    4389             : 
    4390           0 :         return 0;
    4391             : }
    4392             : 
    4393             : static void
    4394           0 : nvmf_rdma_close_qpair(struct spdk_nvmf_qpair *qpair,
    4395             :                       spdk_nvmf_transport_qpair_fini_cb cb_fn, void *cb_arg)
    4396             : {
    4397           0 :         struct spdk_nvmf_rdma_qpair *rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair);
    4398             : 
    4399           0 :         rqpair->to_close = true;
    4400             : 
    4401             :         /* This happens only when the qpair is disconnected before
    4402             :          * it is added to the poll group. Since there is no poll group,
    4403             :          * the RDMA qp has not been initialized yet and the RDMA CM
    4404             :          * event has not yet been acknowledged, so we need to reject it.
    4405             :          */
    4406           0 :         if (rqpair->qpair.state == SPDK_NVMF_QPAIR_UNINITIALIZED) {
    4407           0 :                 nvmf_rdma_qpair_reject_connection(rqpair);
    4408           0 :                 nvmf_rdma_qpair_destroy(rqpair);
    4409           0 :                 return;
    4410             :         }
    4411             : 
    4412           0 :         if (rqpair->rdma_qp) {
    4413           0 :                 spdk_rdma_qp_disconnect(rqpair->rdma_qp);
    4414             :         }
    4415             : 
    4416           0 :         nvmf_rdma_destroy_drained_qpair(rqpair);
    4417             : 
    4418           0 :         if (cb_fn) {
    4419           0 :                 cb_fn(cb_arg);
    4420             :         }
    4421             : }
    4422             : 
    4423             : static struct spdk_nvmf_rdma_qpair *
    4424           0 : get_rdma_qpair_from_wc(struct spdk_nvmf_rdma_poller *rpoller, struct ibv_wc *wc)
    4425             : {
    4426           0 :         struct spdk_nvmf_rdma_qpair find;
    4427             : 
    4428           0 :         find.qp_num = wc->qp_num;
    4429             : 
    4430           0 :         return RB_FIND(qpairs_tree, &rpoller->qpairs, &find);
    4431             : }
    4432             : 
    4433             : #ifdef DEBUG
    4434             : static int
    4435           0 : nvmf_rdma_req_is_completing(struct spdk_nvmf_rdma_request *rdma_req)
    4436             : {
    4437           0 :         return rdma_req->state == RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST ||
    4438           0 :                rdma_req->state == RDMA_REQUEST_STATE_COMPLETING;
    4439             : }
    4440             : #endif
    4441             : 
    4442             : static void
    4443           0 : _poller_reset_failed_recvs(struct spdk_nvmf_rdma_poller *rpoller, struct ibv_recv_wr *bad_recv_wr,
    4444             :                            int rc)
    4445             : {
    4446             :         struct spdk_nvmf_rdma_recv      *rdma_recv;
    4447             :         struct spdk_nvmf_rdma_wr        *bad_rdma_wr;
    4448             : 
    4449           0 :         SPDK_ERRLOG("Failed to post a recv for the poller %p with errno %d\n", rpoller, -rc);
    4450           0 :         while (bad_recv_wr != NULL) {
    4451           0 :                 bad_rdma_wr = (struct spdk_nvmf_rdma_wr *)bad_recv_wr->wr_id;
    4452           0 :                 rdma_recv = SPDK_CONTAINEROF(bad_rdma_wr, struct spdk_nvmf_rdma_recv, rdma_wr);
    4453             : 
    4454           0 :                 rdma_recv->qpair->current_recv_depth++;
    4455           0 :                 bad_recv_wr = bad_recv_wr->next;
    4456           0 :                 SPDK_ERRLOG("Failed to post a recv for the qpair %p with errno %d\n", rdma_recv->qpair, -rc);
    4457           0 :                 spdk_nvmf_qpair_disconnect(&rdma_recv->qpair->qpair);
    4458             :         }
    4459           0 : }
    4460             : 
    4461             : static void
    4462           0 : _qp_reset_failed_recvs(struct spdk_nvmf_rdma_qpair *rqpair, struct ibv_recv_wr *bad_recv_wr, int rc)
    4463             : {
    4464           0 :         SPDK_ERRLOG("Failed to post a recv for the qpair %p with errno %d\n", rqpair, -rc);
    4465           0 :         while (bad_recv_wr != NULL) {
    4466           0 :                 bad_recv_wr = bad_recv_wr->next;
    4467           0 :                 rqpair->current_recv_depth++;
    4468             :         }
    4469           0 :         spdk_nvmf_qpair_disconnect(&rqpair->qpair);
    4470           0 : }
    4471             : 
    4472             : static void
    4473           0 : _poller_submit_recvs(struct spdk_nvmf_rdma_transport *rtransport,
    4474             :                      struct spdk_nvmf_rdma_poller *rpoller)
    4475             : {
    4476             :         struct spdk_nvmf_rdma_qpair     *rqpair;
    4477           0 :         struct ibv_recv_wr              *bad_recv_wr;
    4478             :         int                             rc;
    4479             : 
    4480           0 :         if (rpoller->srq) {
    4481           0 :                 rc = spdk_rdma_srq_flush_recv_wrs(rpoller->srq, &bad_recv_wr);
    4482           0 :                 if (spdk_unlikely(rc)) {
    4483           0 :                         _poller_reset_failed_recvs(rpoller, bad_recv_wr, rc);
    4484             :                 }
    4485             :         } else {
    4486           0 :                 while (!STAILQ_EMPTY(&rpoller->qpairs_pending_recv)) {
    4487           0 :                         rqpair = STAILQ_FIRST(&rpoller->qpairs_pending_recv);
    4488           0 :                         rc = spdk_rdma_qp_flush_recv_wrs(rqpair->rdma_qp, &bad_recv_wr);
    4489           0 :                         if (spdk_unlikely(rc)) {
    4490           0 :                                 _qp_reset_failed_recvs(rqpair, bad_recv_wr, rc);
    4491             :                         }
    4492           0 :                         STAILQ_REMOVE_HEAD(&rpoller->qpairs_pending_recv, recv_link);
    4493             :                 }
    4494             :         }
    4495           0 : }
    4496             : 
    4497             : static void
    4498           0 : _qp_reset_failed_sends(struct spdk_nvmf_rdma_transport *rtransport,
    4499             :                        struct spdk_nvmf_rdma_qpair *rqpair, struct ibv_send_wr *bad_wr, int rc)
    4500             : {
    4501             :         struct spdk_nvmf_rdma_wr        *bad_rdma_wr;
    4502           0 :         struct spdk_nvmf_rdma_request   *prev_rdma_req = NULL, *cur_rdma_req = NULL;
    4503             : 
    4504           0 :         SPDK_ERRLOG("Failed to post a send for the qpair %p with errno %d\n", rqpair, -rc);
    4505           0 :         for (; bad_wr != NULL; bad_wr = bad_wr->next) {
    4506           0 :                 bad_rdma_wr = (struct spdk_nvmf_rdma_wr *)bad_wr->wr_id;
    4507           0 :                 assert(rqpair->current_send_depth > 0);
    4508           0 :                 rqpair->current_send_depth--;
    4509           0 :                 switch (bad_rdma_wr->type) {
    4510           0 :                 case RDMA_WR_TYPE_DATA:
    4511           0 :                         cur_rdma_req = SPDK_CONTAINEROF(bad_rdma_wr, struct spdk_nvmf_rdma_request, data_wr);
    4512           0 :                         if (bad_wr->opcode == IBV_WR_RDMA_READ) {
    4513           0 :                                 assert(rqpair->current_read_depth > 0);
    4514           0 :                                 rqpair->current_read_depth--;
    4515             :                         }
    4516           0 :                         break;
    4517           0 :                 case RDMA_WR_TYPE_SEND:
    4518           0 :                         cur_rdma_req = SPDK_CONTAINEROF(bad_rdma_wr, struct spdk_nvmf_rdma_request, rsp_wr);
    4519           0 :                         break;
    4520           0 :                 default:
    4521           0 :                         SPDK_ERRLOG("Found a RECV in the list of pending SEND requests for qpair %p\n", rqpair);
    4522           0 :                         prev_rdma_req = cur_rdma_req;
    4523           0 :                         continue;
    4524             :                 }
    4525             : 
    4526           0 :                 if (prev_rdma_req == cur_rdma_req) {
    4527             :                         /* this request was handled by an earlier wr. i.e. we were performing an nvme read. */
    4528             :                         /* We only have to check against prev_wr since each requests wrs are contiguous in this list. */
    4529           0 :                         continue;
    4530             :                 }
    4531             : 
    4532           0 :                 switch (cur_rdma_req->state) {
    4533           0 :                 case RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER:
    4534           0 :                         cur_rdma_req->req.rsp->nvme_cpl.status.sc = SPDK_NVME_SC_INTERNAL_DEVICE_ERROR;
    4535           0 :                         STAILQ_INSERT_TAIL(&rqpair->pending_rdma_send_queue, cur_rdma_req, state_link);
    4536           0 :                         cur_rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE_PENDING;
    4537           0 :                         break;
    4538           0 :                 case RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST:
    4539             :                 case RDMA_REQUEST_STATE_COMPLETING:
    4540           0 :                         cur_rdma_req->state = RDMA_REQUEST_STATE_COMPLETED;
    4541           0 :                         break;
    4542           0 :                 default:
    4543           0 :                         SPDK_ERRLOG("Found a request in a bad state %d when draining pending SEND requests for qpair %p\n",
    4544             :                                     cur_rdma_req->state, rqpair);
    4545           0 :                         continue;
    4546             :                 }
    4547             : 
    4548           0 :                 nvmf_rdma_request_process(rtransport, cur_rdma_req);
    4549           0 :                 prev_rdma_req = cur_rdma_req;
    4550             :         }
    4551             : 
    4552           0 :         if (spdk_nvmf_qpair_is_active(&rqpair->qpair)) {
    4553             :                 /* Disconnect the connection. */
    4554           0 :                 spdk_nvmf_qpair_disconnect(&rqpair->qpair);
    4555             :         }
    4556             : 
    4557           0 : }
    4558             : 
    4559             : static void
    4560           0 : _poller_submit_sends(struct spdk_nvmf_rdma_transport *rtransport,
    4561             :                      struct spdk_nvmf_rdma_poller *rpoller)
    4562             : {
    4563             :         struct spdk_nvmf_rdma_qpair     *rqpair;
    4564           0 :         struct ibv_send_wr              *bad_wr = NULL;
    4565             :         int                             rc;
    4566             : 
    4567           0 :         while (!STAILQ_EMPTY(&rpoller->qpairs_pending_send)) {
    4568           0 :                 rqpair = STAILQ_FIRST(&rpoller->qpairs_pending_send);
    4569           0 :                 rc = spdk_rdma_qp_flush_send_wrs(rqpair->rdma_qp, &bad_wr);
    4570             : 
    4571             :                 /* bad wr always points to the first wr that failed. */
    4572           0 :                 if (spdk_unlikely(rc)) {
    4573           0 :                         _qp_reset_failed_sends(rtransport, rqpair, bad_wr, rc);
    4574             :                 }
    4575           0 :                 STAILQ_REMOVE_HEAD(&rpoller->qpairs_pending_send, send_link);
    4576             :         }
    4577           0 : }
    4578             : 
    4579             : static const char *
    4580           0 : nvmf_rdma_wr_type_str(enum spdk_nvmf_rdma_wr_type wr_type)
    4581             : {
    4582           0 :         switch (wr_type) {
    4583           0 :         case RDMA_WR_TYPE_RECV:
    4584           0 :                 return "RECV";
    4585           0 :         case RDMA_WR_TYPE_SEND:
    4586           0 :                 return "SEND";
    4587           0 :         case RDMA_WR_TYPE_DATA:
    4588           0 :                 return "DATA";
    4589           0 :         default:
    4590           0 :                 SPDK_ERRLOG("Unknown WR type %d\n", wr_type);
    4591           0 :                 SPDK_UNREACHABLE();
    4592             :         }
    4593             : }
    4594             : 
    4595             : static inline void
    4596           0 : nvmf_rdma_log_wc_status(struct spdk_nvmf_rdma_qpair *rqpair, struct ibv_wc *wc)
    4597             : {
    4598           0 :         enum spdk_nvmf_rdma_wr_type wr_type = ((struct spdk_nvmf_rdma_wr *)wc->wr_id)->type;
    4599             : 
    4600           0 :         if (wc->status == IBV_WC_WR_FLUSH_ERR) {
    4601             :                 /* If qpair is in ERR state, we will receive completions for all posted and not completed
    4602             :                  * Work Requests with IBV_WC_WR_FLUSH_ERR status. Don't log an error in that case */
    4603           0 :                 SPDK_DEBUGLOG(rdma,
    4604             :                               "Error on CQ %p, (qp state %d, in_error %d) request 0x%lu, type %s, status: (%d): %s\n",
    4605             :                               rqpair->poller->cq, rqpair->qpair.state, rqpair->ibv_in_error_state, wc->wr_id,
    4606             :                               nvmf_rdma_wr_type_str(wr_type), wc->status, ibv_wc_status_str(wc->status));
    4607             :         } else {
    4608           0 :                 SPDK_ERRLOG("Error on CQ %p, (qp state %d, in_error %d) request 0x%lu, type %s, status: (%d): %s\n",
    4609             :                             rqpair->poller->cq, rqpair->qpair.state, rqpair->ibv_in_error_state, wc->wr_id,
    4610             :                             nvmf_rdma_wr_type_str(wr_type), wc->status, ibv_wc_status_str(wc->status));
    4611             :         }
    4612           0 : }
    4613             : 
    4614             : static int
    4615           0 : nvmf_rdma_poller_poll(struct spdk_nvmf_rdma_transport *rtransport,
    4616             :                       struct spdk_nvmf_rdma_poller *rpoller)
    4617             : {
    4618           0 :         struct ibv_wc wc[32];
    4619             :         struct spdk_nvmf_rdma_wr        *rdma_wr;
    4620             :         struct spdk_nvmf_rdma_request   *rdma_req;
    4621             :         struct spdk_nvmf_rdma_recv      *rdma_recv;
    4622             :         struct spdk_nvmf_rdma_qpair     *rqpair, *tmp_rqpair;
    4623             :         int reaped, i;
    4624           0 :         int count = 0;
    4625             :         int rc;
    4626           0 :         bool error = false;
    4627           0 :         uint64_t poll_tsc = spdk_get_ticks();
    4628             : 
    4629           0 :         if (spdk_unlikely(rpoller->need_destroy)) {
    4630             :                 /* If qpair is closed before poller destroy, nvmf_rdma_destroy_drained_qpair may not
    4631             :                  * be called because we cannot poll anything from cq. So we call that here to force
    4632             :                  * destroy the qpair after to_close turning true.
    4633             :                  */
    4634           0 :                 RB_FOREACH_SAFE(rqpair, qpairs_tree, &rpoller->qpairs, tmp_rqpair) {
    4635           0 :                         nvmf_rdma_destroy_drained_qpair(rqpair);
    4636             :                 }
    4637           0 :                 return 0;
    4638             :         }
    4639             : 
    4640             :         /* Poll for completing operations. */
    4641           0 :         reaped = ibv_poll_cq(rpoller->cq, 32, wc);
    4642           0 :         if (spdk_unlikely(reaped < 0)) {
    4643           0 :                 SPDK_ERRLOG("Error polling CQ! (%d): %s\n",
    4644             :                             errno, spdk_strerror(errno));
    4645           0 :                 return -1;
    4646           0 :         } else if (reaped == 0) {
    4647           0 :                 rpoller->stat.idle_polls++;
    4648             :         }
    4649             : 
    4650           0 :         rpoller->stat.polls++;
    4651           0 :         rpoller->stat.completions += reaped;
    4652             : 
    4653           0 :         for (i = 0; i < reaped; i++) {
    4654             : 
    4655           0 :                 rdma_wr = (struct spdk_nvmf_rdma_wr *)wc[i].wr_id;
    4656             : 
    4657           0 :                 switch (rdma_wr->type) {
    4658           0 :                 case RDMA_WR_TYPE_SEND:
    4659           0 :                         rdma_req = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvmf_rdma_request, rsp_wr);
    4660           0 :                         rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair);
    4661             : 
    4662           0 :                         if (spdk_likely(!wc[i].status)) {
    4663           0 :                                 count++;
    4664           0 :                                 assert(wc[i].opcode == IBV_WC_SEND);
    4665           0 :                                 assert(nvmf_rdma_req_is_completing(rdma_req));
    4666             :                         }
    4667             : 
    4668           0 :                         rdma_req->state = RDMA_REQUEST_STATE_COMPLETED;
    4669             :                         /* RDMA_WRITE operation completed. +1 since it was chained with rsp WR */
    4670           0 :                         assert(rqpair->current_send_depth >= (uint32_t)rdma_req->num_outstanding_data_wr + 1);
    4671           0 :                         rqpair->current_send_depth -= rdma_req->num_outstanding_data_wr + 1;
    4672           0 :                         rdma_req->num_outstanding_data_wr = 0;
    4673             : 
    4674           0 :                         nvmf_rdma_request_process(rtransport, rdma_req);
    4675           0 :                         break;
    4676           0 :                 case RDMA_WR_TYPE_RECV:
    4677             :                         /* rdma_recv->qpair will be invalid if using an SRQ.  In that case we have to get the qpair from the wc. */
    4678           0 :                         rdma_recv = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvmf_rdma_recv, rdma_wr);
    4679           0 :                         if (rpoller->srq != NULL) {
    4680           0 :                                 rdma_recv->qpair = get_rdma_qpair_from_wc(rpoller, &wc[i]);
    4681             :                                 /* It is possible that there are still some completions for destroyed QP
    4682             :                                  * associated with SRQ. We just ignore these late completions and re-post
    4683             :                                  * receive WRs back to SRQ.
    4684             :                                  */
    4685           0 :                                 if (spdk_unlikely(NULL == rdma_recv->qpair)) {
    4686           0 :                                         struct ibv_recv_wr *bad_wr;
    4687             : 
    4688           0 :                                         rdma_recv->wr.next = NULL;
    4689           0 :                                         spdk_rdma_srq_queue_recv_wrs(rpoller->srq, &rdma_recv->wr);
    4690           0 :                                         rc = spdk_rdma_srq_flush_recv_wrs(rpoller->srq, &bad_wr);
    4691           0 :                                         if (rc) {
    4692           0 :                                                 SPDK_ERRLOG("Failed to re-post recv WR to SRQ, err %d\n", rc);
    4693             :                                         }
    4694           0 :                                         continue;
    4695             :                                 }
    4696             :                         }
    4697           0 :                         rqpair = rdma_recv->qpair;
    4698             : 
    4699           0 :                         assert(rqpair != NULL);
    4700           0 :                         if (spdk_likely(!wc[i].status)) {
    4701           0 :                                 assert(wc[i].opcode == IBV_WC_RECV);
    4702           0 :                                 if (rqpair->current_recv_depth >= rqpair->max_queue_depth) {
    4703           0 :                                         spdk_nvmf_qpair_disconnect(&rqpair->qpair);
    4704           0 :                                         break;
    4705             :                                 }
    4706             :                         }
    4707             : 
    4708           0 :                         rdma_recv->wr.next = NULL;
    4709           0 :                         rqpair->current_recv_depth++;
    4710           0 :                         rdma_recv->receive_tsc = poll_tsc;
    4711           0 :                         rpoller->stat.requests++;
    4712           0 :                         STAILQ_INSERT_HEAD(&rqpair->resources->incoming_queue, rdma_recv, link);
    4713           0 :                         rqpair->qpair.queue_depth++;
    4714           0 :                         break;
    4715           0 :                 case RDMA_WR_TYPE_DATA:
    4716           0 :                         rdma_req = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvmf_rdma_request, data_wr);
    4717           0 :                         rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair);
    4718             : 
    4719           0 :                         assert(rdma_req->num_outstanding_data_wr > 0);
    4720             : 
    4721           0 :                         rqpair->current_send_depth--;
    4722           0 :                         rdma_req->num_outstanding_data_wr--;
    4723           0 :                         if (spdk_likely(!wc[i].status)) {
    4724           0 :                                 assert(wc[i].opcode == IBV_WC_RDMA_READ);
    4725           0 :                                 rqpair->current_read_depth--;
    4726             :                                 /* wait for all outstanding reads associated with the same rdma_req to complete before proceeding. */
    4727           0 :                                 if (rdma_req->num_outstanding_data_wr == 0) {
    4728           0 :                                         if (rdma_req->num_remaining_data_wr) {
    4729             :                                                 /* Only part of RDMA_READ operations was submitted, process the rest */
    4730           0 :                                                 nvmf_rdma_request_reset_transfer_in(rdma_req, rtransport);
    4731           0 :                                                 rdma_req->state = RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING;
    4732           0 :                                                 nvmf_rdma_request_process(rtransport, rdma_req);
    4733           0 :                                                 break;
    4734             :                                         }
    4735           0 :                                         rdma_req->state = RDMA_REQUEST_STATE_READY_TO_EXECUTE;
    4736           0 :                                         nvmf_rdma_request_process(rtransport, rdma_req);
    4737             :                                 }
    4738             :                         } else {
    4739             :                                 /* If the data transfer fails still force the queue into the error state,
    4740             :                                  * if we were performing an RDMA_READ, we need to force the request into a
    4741             :                                  * completed state since it wasn't linked to a send. However, in the RDMA_WRITE
    4742             :                                  * case, we should wait for the SEND to complete. */
    4743           0 :                                 if (rdma_req->data.wr.opcode == IBV_WR_RDMA_READ) {
    4744           0 :                                         rqpair->current_read_depth--;
    4745           0 :                                         if (rdma_req->num_outstanding_data_wr == 0) {
    4746           0 :                                                 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED;
    4747             :                                         }
    4748             :                                 }
    4749             :                         }
    4750           0 :                         break;
    4751           0 :                 default:
    4752           0 :                         SPDK_ERRLOG("Received an unknown opcode on the CQ: %d\n", wc[i].opcode);
    4753           0 :                         continue;
    4754             :                 }
    4755             : 
    4756             :                 /* Handle error conditions */
    4757           0 :                 if (spdk_unlikely(wc[i].status)) {
    4758           0 :                         rqpair->ibv_in_error_state = true;
    4759           0 :                         nvmf_rdma_log_wc_status(rqpair, &wc[i]);
    4760             : 
    4761           0 :                         error = true;
    4762             : 
    4763           0 :                         if (spdk_nvmf_qpair_is_active(&rqpair->qpair)) {
    4764             :                                 /* Disconnect the connection. */
    4765           0 :                                 spdk_nvmf_qpair_disconnect(&rqpair->qpair);
    4766             :                         } else {
    4767           0 :                                 nvmf_rdma_destroy_drained_qpair(rqpair);
    4768             :                         }
    4769           0 :                         continue;
    4770             :                 }
    4771             : 
    4772           0 :                 nvmf_rdma_qpair_process_pending(rtransport, rqpair, false);
    4773             : 
    4774           0 :                 if (spdk_unlikely(!spdk_nvmf_qpair_is_active(&rqpair->qpair))) {
    4775           0 :                         nvmf_rdma_destroy_drained_qpair(rqpair);
    4776             :                 }
    4777             :         }
    4778             : 
    4779           0 :         if (spdk_unlikely(error == true)) {
    4780           0 :                 return -1;
    4781             :         }
    4782             : 
    4783           0 :         if (reaped == 0) {
    4784             :                 /* In some cases we may not receive any CQE but we still may have pending IO requests waiting for
    4785             :                  * a resource (e.g. a WR from the data_wr_pool).
    4786             :                  * We need to start processing of such requests if no CQE reaped */
    4787           0 :                 nvmf_rdma_poller_process_pending_buf_queue(rtransport, rpoller);
    4788             :         }
    4789             : 
    4790             :         /* submit outstanding work requests. */
    4791           0 :         _poller_submit_recvs(rtransport, rpoller);
    4792           0 :         _poller_submit_sends(rtransport, rpoller);
    4793             : 
    4794           0 :         return count;
    4795             : }
    4796             : 
    4797             : static void
    4798           0 : _nvmf_rdma_remove_destroyed_device(void *c)
    4799             : {
    4800           0 :         struct spdk_nvmf_rdma_transport *rtransport = c;
    4801             :         struct spdk_nvmf_rdma_device    *device, *device_tmp;
    4802             :         int                             rc;
    4803             : 
    4804           0 :         TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, device_tmp) {
    4805           0 :                 if (device->ready_to_destroy) {
    4806           0 :                         destroy_ib_device(rtransport, device);
    4807             :                 }
    4808             :         }
    4809             : 
    4810           0 :         free_poll_fds(rtransport);
    4811           0 :         rc = generate_poll_fds(rtransport);
    4812             :         /* cannot handle fd allocation error here */
    4813           0 :         if (rc != 0) {
    4814           0 :                 SPDK_ERRLOG("Failed to generate poll fds after remove ib device.\n");
    4815             :         }
    4816           0 : }
    4817             : 
    4818             : static void
    4819           0 : _nvmf_rdma_remove_poller_in_group_cb(void *c)
    4820             : {
    4821           0 :         struct poller_manage_ctx        *ctx = c;
    4822           0 :         struct spdk_nvmf_rdma_transport *rtransport = ctx->rtransport;
    4823           0 :         struct spdk_nvmf_rdma_device    *device = ctx->device;
    4824           0 :         struct spdk_thread              *thread = ctx->thread;
    4825             : 
    4826           0 :         if (nvmf_rdma_all_pollers_management_done(c)) {
    4827             :                 /* destroy device when last poller is destroyed */
    4828           0 :                 device->ready_to_destroy = true;
    4829           0 :                 spdk_thread_send_msg(thread, _nvmf_rdma_remove_destroyed_device, rtransport);
    4830             :         }
    4831           0 : }
    4832             : 
    4833             : static void
    4834           0 : _nvmf_rdma_remove_poller_in_group(void *c)
    4835             : {
    4836           0 :         struct poller_manage_ctx                *ctx = c;
    4837             : 
    4838           0 :         ctx->rpoller->need_destroy = true;
    4839           0 :         ctx->rpoller->destroy_cb_ctx = ctx;
    4840           0 :         ctx->rpoller->destroy_cb = _nvmf_rdma_remove_poller_in_group_cb;
    4841             : 
    4842             :         /* qp will be disconnected after receiving a RDMA_CM_EVENT_DEVICE_REMOVAL event. */
    4843           0 :         if (RB_EMPTY(&ctx->rpoller->qpairs)) {
    4844           0 :                 nvmf_rdma_poller_destroy(ctx->rpoller);
    4845             :         }
    4846           0 : }
    4847             : 
    4848             : static int
    4849           0 : nvmf_rdma_poll_group_poll(struct spdk_nvmf_transport_poll_group *group)
    4850             : {
    4851             :         struct spdk_nvmf_rdma_transport *rtransport;
    4852             :         struct spdk_nvmf_rdma_poll_group *rgroup;
    4853             :         struct spdk_nvmf_rdma_poller    *rpoller, *tmp;
    4854           0 :         int                             count = 0, rc, rc2 = 0;
    4855             : 
    4856           0 :         rtransport = SPDK_CONTAINEROF(group->transport, struct spdk_nvmf_rdma_transport, transport);
    4857           0 :         rgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_rdma_poll_group, group);
    4858             : 
    4859           0 :         TAILQ_FOREACH_SAFE(rpoller, &rgroup->pollers, link, tmp) {
    4860           0 :                 rc = nvmf_rdma_poller_poll(rtransport, rpoller);
    4861           0 :                 if (spdk_unlikely(rc < 0)) {
    4862           0 :                         if (rc2 == 0) {
    4863           0 :                                 rc2 = rc;
    4864             :                         }
    4865           0 :                         continue;
    4866             :                 }
    4867           0 :                 count += rc;
    4868             :         }
    4869             : 
    4870           0 :         return rc2 ? rc2 : count;
    4871             : }
    4872             : 
    4873             : static int
    4874           0 : nvmf_rdma_trid_from_cm_id(struct rdma_cm_id *id,
    4875             :                           struct spdk_nvme_transport_id *trid,
    4876             :                           bool peer)
    4877             : {
    4878             :         struct sockaddr *saddr;
    4879             :         uint16_t port;
    4880             : 
    4881           0 :         spdk_nvme_trid_populate_transport(trid, SPDK_NVME_TRANSPORT_RDMA);
    4882             : 
    4883           0 :         if (peer) {
    4884           0 :                 saddr = rdma_get_peer_addr(id);
    4885             :         } else {
    4886           0 :                 saddr = rdma_get_local_addr(id);
    4887             :         }
    4888           0 :         switch (saddr->sa_family) {
    4889           0 :         case AF_INET: {
    4890           0 :                 struct sockaddr_in *saddr_in = (struct sockaddr_in *)saddr;
    4891             : 
    4892           0 :                 trid->adrfam = SPDK_NVMF_ADRFAM_IPV4;
    4893           0 :                 inet_ntop(AF_INET, &saddr_in->sin_addr,
    4894           0 :                           trid->traddr, sizeof(trid->traddr));
    4895           0 :                 if (peer) {
    4896           0 :                         port = ntohs(rdma_get_dst_port(id));
    4897             :                 } else {
    4898           0 :                         port = ntohs(rdma_get_src_port(id));
    4899             :                 }
    4900           0 :                 snprintf(trid->trsvcid, sizeof(trid->trsvcid), "%u", port);
    4901           0 :                 break;
    4902             :         }
    4903           0 :         case AF_INET6: {
    4904           0 :                 struct sockaddr_in6 *saddr_in = (struct sockaddr_in6 *)saddr;
    4905           0 :                 trid->adrfam = SPDK_NVMF_ADRFAM_IPV6;
    4906           0 :                 inet_ntop(AF_INET6, &saddr_in->sin6_addr,
    4907           0 :                           trid->traddr, sizeof(trid->traddr));
    4908           0 :                 if (peer) {
    4909           0 :                         port = ntohs(rdma_get_dst_port(id));
    4910             :                 } else {
    4911           0 :                         port = ntohs(rdma_get_src_port(id));
    4912             :                 }
    4913           0 :                 snprintf(trid->trsvcid, sizeof(trid->trsvcid), "%u", port);
    4914           0 :                 break;
    4915             :         }
    4916           0 :         default:
    4917           0 :                 return -1;
    4918             : 
    4919             :         }
    4920             : 
    4921           0 :         return 0;
    4922             : }
    4923             : 
    4924             : static int
    4925           0 : nvmf_rdma_qpair_get_peer_trid(struct spdk_nvmf_qpair *qpair,
    4926             :                               struct spdk_nvme_transport_id *trid)
    4927             : {
    4928             :         struct spdk_nvmf_rdma_qpair     *rqpair;
    4929             : 
    4930           0 :         rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair);
    4931             : 
    4932           0 :         return nvmf_rdma_trid_from_cm_id(rqpair->cm_id, trid, true);
    4933             : }
    4934             : 
    4935             : static int
    4936           0 : nvmf_rdma_qpair_get_local_trid(struct spdk_nvmf_qpair *qpair,
    4937             :                                struct spdk_nvme_transport_id *trid)
    4938             : {
    4939             :         struct spdk_nvmf_rdma_qpair     *rqpair;
    4940             : 
    4941           0 :         rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair);
    4942             : 
    4943           0 :         return nvmf_rdma_trid_from_cm_id(rqpair->cm_id, trid, false);
    4944             : }
    4945             : 
    4946             : static int
    4947           0 : nvmf_rdma_qpair_get_listen_trid(struct spdk_nvmf_qpair *qpair,
    4948             :                                 struct spdk_nvme_transport_id *trid)
    4949             : {
    4950             :         struct spdk_nvmf_rdma_qpair     *rqpair;
    4951             : 
    4952           0 :         rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair);
    4953             : 
    4954           0 :         return nvmf_rdma_trid_from_cm_id(rqpair->listen_id, trid, false);
    4955             : }
    4956             : 
    4957             : void
    4958           0 : spdk_nvmf_rdma_init_hooks(struct spdk_nvme_rdma_hooks *hooks)
    4959             : {
    4960           0 :         g_nvmf_hooks = *hooks;
    4961           0 : }
    4962             : 
    4963             : static void
    4964           0 : nvmf_rdma_request_set_abort_status(struct spdk_nvmf_request *req,
    4965             :                                    struct spdk_nvmf_rdma_request *rdma_req_to_abort,
    4966             :                                    struct spdk_nvmf_rdma_qpair *rqpair)
    4967             : {
    4968           0 :         rdma_req_to_abort->req.rsp->nvme_cpl.status.sct = SPDK_NVME_SCT_GENERIC;
    4969           0 :         rdma_req_to_abort->req.rsp->nvme_cpl.status.sc = SPDK_NVME_SC_ABORTED_BY_REQUEST;
    4970             : 
    4971           0 :         STAILQ_INSERT_TAIL(&rqpair->pending_rdma_send_queue, rdma_req_to_abort, state_link);
    4972           0 :         rdma_req_to_abort->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE_PENDING;
    4973             : 
    4974           0 :         req->rsp->nvme_cpl.cdw0 &= ~1U;       /* Command was successfully aborted. */
    4975           0 : }
    4976             : 
    4977             : static int
    4978           0 : _nvmf_rdma_qpair_abort_request(void *ctx)
    4979             : {
    4980           0 :         struct spdk_nvmf_request *req = ctx;
    4981           0 :         struct spdk_nvmf_rdma_request *rdma_req_to_abort = SPDK_CONTAINEROF(
    4982             :                                 req->req_to_abort, struct spdk_nvmf_rdma_request, req);
    4983           0 :         struct spdk_nvmf_rdma_qpair *rqpair = SPDK_CONTAINEROF(req->req_to_abort->qpair,
    4984             :                                               struct spdk_nvmf_rdma_qpair, qpair);
    4985             :         int rc;
    4986             : 
    4987           0 :         spdk_poller_unregister(&req->poller);
    4988             : 
    4989           0 :         switch (rdma_req_to_abort->state) {
    4990           0 :         case RDMA_REQUEST_STATE_EXECUTING:
    4991           0 :                 rc = nvmf_ctrlr_abort_request(req);
    4992           0 :                 if (rc == SPDK_NVMF_REQUEST_EXEC_STATUS_ASYNCHRONOUS) {
    4993           0 :                         return SPDK_POLLER_BUSY;
    4994             :                 }
    4995           0 :                 break;
    4996             : 
    4997           0 :         case RDMA_REQUEST_STATE_NEED_BUFFER:
    4998           0 :                 STAILQ_REMOVE(&rqpair->poller->group->group.pending_buf_queue,
    4999             :                               &rdma_req_to_abort->req, spdk_nvmf_request, buf_link);
    5000             : 
    5001           0 :                 nvmf_rdma_request_set_abort_status(req, rdma_req_to_abort, rqpair);
    5002           0 :                 break;
    5003             : 
    5004           0 :         case RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING:
    5005           0 :                 STAILQ_REMOVE(&rqpair->pending_rdma_read_queue, rdma_req_to_abort,
    5006             :                               spdk_nvmf_rdma_request, state_link);
    5007             : 
    5008           0 :                 nvmf_rdma_request_set_abort_status(req, rdma_req_to_abort, rqpair);
    5009           0 :                 break;
    5010             : 
    5011           0 :         case RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING:
    5012           0 :                 STAILQ_REMOVE(&rqpair->pending_rdma_write_queue, rdma_req_to_abort,
    5013             :                               spdk_nvmf_rdma_request, state_link);
    5014             : 
    5015           0 :                 nvmf_rdma_request_set_abort_status(req, rdma_req_to_abort, rqpair);
    5016           0 :                 break;
    5017             : 
    5018           0 :         case RDMA_REQUEST_STATE_READY_TO_COMPLETE_PENDING:
    5019             :                 /* Remove req from the list here to re-use common function */
    5020           0 :                 STAILQ_REMOVE(&rqpair->pending_rdma_send_queue, rdma_req_to_abort,
    5021             :                               spdk_nvmf_rdma_request, state_link);
    5022             : 
    5023           0 :                 nvmf_rdma_request_set_abort_status(req, rdma_req_to_abort, rqpair);
    5024           0 :                 break;
    5025             : 
    5026           0 :         case RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER:
    5027           0 :                 if (spdk_get_ticks() < req->timeout_tsc) {
    5028           0 :                         req->poller = SPDK_POLLER_REGISTER(_nvmf_rdma_qpair_abort_request, req, 0);
    5029           0 :                         return SPDK_POLLER_BUSY;
    5030             :                 }
    5031           0 :                 break;
    5032             : 
    5033           0 :         default:
    5034           0 :                 break;
    5035             :         }
    5036             : 
    5037           0 :         spdk_nvmf_request_complete(req);
    5038           0 :         return SPDK_POLLER_BUSY;
    5039             : }
    5040             : 
    5041             : static void
    5042           0 : nvmf_rdma_qpair_abort_request(struct spdk_nvmf_qpair *qpair,
    5043             :                               struct spdk_nvmf_request *req)
    5044             : {
    5045             :         struct spdk_nvmf_rdma_qpair *rqpair;
    5046             :         struct spdk_nvmf_rdma_transport *rtransport;
    5047             :         struct spdk_nvmf_transport *transport;
    5048             :         uint16_t cid;
    5049             :         uint32_t i, max_req_count;
    5050           0 :         struct spdk_nvmf_rdma_request *rdma_req_to_abort = NULL, *rdma_req;
    5051             : 
    5052           0 :         rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair);
    5053           0 :         rtransport = SPDK_CONTAINEROF(qpair->transport, struct spdk_nvmf_rdma_transport, transport);
    5054           0 :         transport = &rtransport->transport;
    5055             : 
    5056           0 :         cid = req->cmd->nvme_cmd.cdw10_bits.abort.cid;
    5057           0 :         max_req_count = rqpair->srq == NULL ? rqpair->max_queue_depth : rqpair->poller->max_srq_depth;
    5058             : 
    5059           0 :         for (i = 0; i < max_req_count; i++) {
    5060           0 :                 rdma_req = &rqpair->resources->reqs[i];
    5061             :                 /* When SRQ == NULL, rqpair has its own requests and req.qpair pointer always points to the qpair
    5062             :                  * When SRQ != NULL all rqpairs share common requests and qpair pointer is assigned when we start to
    5063             :                  * process a request. So in both cases all requests which are not in FREE state have valid qpair ptr */
    5064           0 :                 if (rdma_req->state != RDMA_REQUEST_STATE_FREE && rdma_req->req.cmd->nvme_cmd.cid == cid &&
    5065           0 :                     rdma_req->req.qpair == qpair) {
    5066           0 :                         rdma_req_to_abort = rdma_req;
    5067           0 :                         break;
    5068             :                 }
    5069             :         }
    5070             : 
    5071           0 :         if (rdma_req_to_abort == NULL) {
    5072           0 :                 spdk_nvmf_request_complete(req);
    5073           0 :                 return;
    5074             :         }
    5075             : 
    5076           0 :         req->req_to_abort = &rdma_req_to_abort->req;
    5077           0 :         req->timeout_tsc = spdk_get_ticks() +
    5078           0 :                            transport->opts.abort_timeout_sec * spdk_get_ticks_hz();
    5079           0 :         req->poller = NULL;
    5080             : 
    5081           0 :         _nvmf_rdma_qpair_abort_request(req);
    5082             : }
    5083             : 
    5084             : static void
    5085           0 : nvmf_rdma_poll_group_dump_stat(struct spdk_nvmf_transport_poll_group *group,
    5086             :                                struct spdk_json_write_ctx *w)
    5087             : {
    5088             :         struct spdk_nvmf_rdma_poll_group *rgroup;
    5089             :         struct spdk_nvmf_rdma_poller *rpoller;
    5090             : 
    5091           0 :         assert(w != NULL);
    5092             : 
    5093           0 :         rgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_rdma_poll_group, group);
    5094             : 
    5095           0 :         spdk_json_write_named_uint64(w, "pending_data_buffer", rgroup->stat.pending_data_buffer);
    5096             : 
    5097           0 :         spdk_json_write_named_array_begin(w, "devices");
    5098             : 
    5099           0 :         TAILQ_FOREACH(rpoller, &rgroup->pollers, link) {
    5100           0 :                 spdk_json_write_object_begin(w);
    5101           0 :                 spdk_json_write_named_string(w, "name",
    5102           0 :                                              ibv_get_device_name(rpoller->device->context->device));
    5103           0 :                 spdk_json_write_named_uint64(w, "polls",
    5104             :                                              rpoller->stat.polls);
    5105           0 :                 spdk_json_write_named_uint64(w, "idle_polls",
    5106             :                                              rpoller->stat.idle_polls);
    5107           0 :                 spdk_json_write_named_uint64(w, "completions",
    5108             :                                              rpoller->stat.completions);
    5109           0 :                 spdk_json_write_named_uint64(w, "requests",
    5110             :                                              rpoller->stat.requests);
    5111           0 :                 spdk_json_write_named_uint64(w, "request_latency",
    5112             :                                              rpoller->stat.request_latency);
    5113           0 :                 spdk_json_write_named_uint64(w, "pending_free_request",
    5114             :                                              rpoller->stat.pending_free_request);
    5115           0 :                 spdk_json_write_named_uint64(w, "pending_rdma_read",
    5116             :                                              rpoller->stat.pending_rdma_read);
    5117           0 :                 spdk_json_write_named_uint64(w, "pending_rdma_write",
    5118             :                                              rpoller->stat.pending_rdma_write);
    5119           0 :                 spdk_json_write_named_uint64(w, "pending_rdma_send",
    5120             :                                              rpoller->stat.pending_rdma_send);
    5121           0 :                 spdk_json_write_named_uint64(w, "total_send_wrs",
    5122             :                                              rpoller->stat.qp_stats.send.num_submitted_wrs);
    5123           0 :                 spdk_json_write_named_uint64(w, "send_doorbell_updates",
    5124             :                                              rpoller->stat.qp_stats.send.doorbell_updates);
    5125           0 :                 spdk_json_write_named_uint64(w, "total_recv_wrs",
    5126             :                                              rpoller->stat.qp_stats.recv.num_submitted_wrs);
    5127           0 :                 spdk_json_write_named_uint64(w, "recv_doorbell_updates",
    5128             :                                              rpoller->stat.qp_stats.recv.doorbell_updates);
    5129           0 :                 spdk_json_write_object_end(w);
    5130             :         }
    5131             : 
    5132           0 :         spdk_json_write_array_end(w);
    5133           0 : }
    5134             : 
    5135             : const struct spdk_nvmf_transport_ops spdk_nvmf_transport_rdma = {
    5136             :         .name = "RDMA",
    5137             :         .type = SPDK_NVME_TRANSPORT_RDMA,
    5138             :         .opts_init = nvmf_rdma_opts_init,
    5139             :         .create = nvmf_rdma_create,
    5140             :         .dump_opts = nvmf_rdma_dump_opts,
    5141             :         .destroy = nvmf_rdma_destroy,
    5142             : 
    5143             :         .listen = nvmf_rdma_listen,
    5144             :         .stop_listen = nvmf_rdma_stop_listen,
    5145             :         .cdata_init = nvmf_rdma_cdata_init,
    5146             : 
    5147             :         .listener_discover = nvmf_rdma_discover,
    5148             : 
    5149             :         .poll_group_create = nvmf_rdma_poll_group_create,
    5150             :         .get_optimal_poll_group = nvmf_rdma_get_optimal_poll_group,
    5151             :         .poll_group_destroy = nvmf_rdma_poll_group_destroy,
    5152             :         .poll_group_add = nvmf_rdma_poll_group_add,
    5153             :         .poll_group_remove = nvmf_rdma_poll_group_remove,
    5154             :         .poll_group_poll = nvmf_rdma_poll_group_poll,
    5155             : 
    5156             :         .req_free = nvmf_rdma_request_free,
    5157             :         .req_complete = nvmf_rdma_request_complete,
    5158             : 
    5159             :         .qpair_fini = nvmf_rdma_close_qpair,
    5160             :         .qpair_get_peer_trid = nvmf_rdma_qpair_get_peer_trid,
    5161             :         .qpair_get_local_trid = nvmf_rdma_qpair_get_local_trid,
    5162             :         .qpair_get_listen_trid = nvmf_rdma_qpair_get_listen_trid,
    5163             :         .qpair_abort_request = nvmf_rdma_qpair_abort_request,
    5164             : 
    5165             :         .poll_group_dump_stat = nvmf_rdma_poll_group_dump_stat,
    5166             : };
    5167             : 
    5168           2 : SPDK_NVMF_TRANSPORT_REGISTER(rdma, &spdk_nvmf_transport_rdma);
    5169           2 : SPDK_LOG_REGISTER_COMPONENT(rdma)

Generated by: LCOV version 1.15