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 : SPDK_STATIC_ASSERT(NVMF_DEFAULT_MSDBD <= SPDK_NVMF_MAX_SGL_ENTRIES,
39 : "MSDBD must not exceed SPDK_NVMF_MAX_SGL_ENTRIES");
40 :
41 : /* The RDMA completion queue size */
42 : #define DEFAULT_NVMF_RDMA_CQ_SIZE 4096
43 : #define MAX_WR_PER_QP(queue_depth) (queue_depth * 3 + 2)
44 :
45 : static int g_spdk_nvmf_ibv_query_mask =
46 : IBV_QP_STATE |
47 : IBV_QP_PKEY_INDEX |
48 : IBV_QP_PORT |
49 : IBV_QP_ACCESS_FLAGS |
50 : IBV_QP_AV |
51 : IBV_QP_PATH_MTU |
52 : IBV_QP_DEST_QPN |
53 : IBV_QP_RQ_PSN |
54 : IBV_QP_MAX_DEST_RD_ATOMIC |
55 : IBV_QP_MIN_RNR_TIMER |
56 : IBV_QP_SQ_PSN |
57 : IBV_QP_TIMEOUT |
58 : IBV_QP_RETRY_CNT |
59 : IBV_QP_RNR_RETRY |
60 : IBV_QP_MAX_QP_RD_ATOMIC;
61 :
62 : enum spdk_nvmf_rdma_request_state {
63 : /* The request is not currently in use */
64 : RDMA_REQUEST_STATE_FREE = 0,
65 :
66 : /* Initial state when request first received */
67 : RDMA_REQUEST_STATE_NEW,
68 :
69 : /* The request is queued until a data buffer is available. */
70 : RDMA_REQUEST_STATE_NEED_BUFFER,
71 :
72 : /* The request is waiting on RDMA queue depth availability
73 : * to transfer data from the host to the controller.
74 : */
75 : RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING,
76 :
77 : /* The request is currently transferring data from the host to the controller. */
78 : RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER,
79 :
80 : /* The request is ready to execute at the block device */
81 : RDMA_REQUEST_STATE_READY_TO_EXECUTE,
82 :
83 : /* The request is currently executing at the block device */
84 : RDMA_REQUEST_STATE_EXECUTING,
85 :
86 : /* The request finished executing at the block device */
87 : RDMA_REQUEST_STATE_EXECUTED,
88 :
89 : /* The request is waiting on RDMA queue depth availability
90 : * to transfer data from the controller to the host.
91 : */
92 : RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING,
93 :
94 : /* The request is waiting on RDMA queue depth availability
95 : * to send response to the host.
96 : */
97 : RDMA_REQUEST_STATE_READY_TO_COMPLETE_PENDING,
98 :
99 : /* The request is ready to send a completion */
100 : RDMA_REQUEST_STATE_READY_TO_COMPLETE,
101 :
102 : /* The request is currently transferring data from the controller to the host. */
103 : RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST,
104 :
105 : /* The request currently has an outstanding completion without an
106 : * associated data transfer.
107 : */
108 : RDMA_REQUEST_STATE_COMPLETING,
109 :
110 : /* The request completed and can be marked free. */
111 : RDMA_REQUEST_STATE_COMPLETED,
112 :
113 : /* Terminator */
114 : RDMA_REQUEST_NUM_STATES,
115 : };
116 :
117 2 : SPDK_TRACE_REGISTER_FN(nvmf_trace, "nvmf_rdma", TRACE_GROUP_NVMF_RDMA)
118 : {
119 0 : spdk_trace_register_object(OBJECT_NVMF_RDMA_IO, 'r');
120 0 : spdk_trace_register_description("RDMA_REQ_NEW", TRACE_RDMA_REQUEST_STATE_NEW,
121 : OWNER_NONE, OBJECT_NVMF_RDMA_IO, 1,
122 : SPDK_TRACE_ARG_TYPE_PTR, "qpair");
123 0 : spdk_trace_register_description("RDMA_REQ_NEED_BUFFER", TRACE_RDMA_REQUEST_STATE_NEED_BUFFER,
124 : OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0,
125 : SPDK_TRACE_ARG_TYPE_PTR, "qpair");
126 0 : spdk_trace_register_description("RDMA_REQ_TX_PENDING_C2H",
127 : TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING,
128 : OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0,
129 : SPDK_TRACE_ARG_TYPE_PTR, "qpair");
130 0 : spdk_trace_register_description("RDMA_REQ_TX_PENDING_H2C",
131 : TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING,
132 : OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0,
133 : SPDK_TRACE_ARG_TYPE_PTR, "qpair");
134 0 : spdk_trace_register_description("RDMA_REQ_TX_H2C",
135 : TRACE_RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER,
136 : OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0,
137 : SPDK_TRACE_ARG_TYPE_PTR, "qpair");
138 0 : spdk_trace_register_description("RDMA_REQ_RDY_TO_EXECUTE",
139 : TRACE_RDMA_REQUEST_STATE_READY_TO_EXECUTE,
140 : OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0,
141 : SPDK_TRACE_ARG_TYPE_PTR, "qpair");
142 0 : spdk_trace_register_description("RDMA_REQ_EXECUTING",
143 : TRACE_RDMA_REQUEST_STATE_EXECUTING,
144 : OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0,
145 : SPDK_TRACE_ARG_TYPE_PTR, "qpair");
146 0 : spdk_trace_register_description("RDMA_REQ_EXECUTED",
147 : TRACE_RDMA_REQUEST_STATE_EXECUTED,
148 : OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0,
149 : SPDK_TRACE_ARG_TYPE_PTR, "qpair");
150 0 : spdk_trace_register_description("RDMA_REQ_RDY_TO_COMPL_PEND",
151 : TRACE_RDMA_REQUEST_STATE_READY_TO_COMPLETE_PENDING,
152 : OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0,
153 : SPDK_TRACE_ARG_TYPE_PTR, "qpair");
154 0 : spdk_trace_register_description("RDMA_REQ_RDY_TO_COMPL",
155 : TRACE_RDMA_REQUEST_STATE_READY_TO_COMPLETE,
156 : OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0,
157 : SPDK_TRACE_ARG_TYPE_PTR, "qpair");
158 0 : spdk_trace_register_description("RDMA_REQ_COMPLETING_C2H",
159 : TRACE_RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST,
160 : OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0,
161 : SPDK_TRACE_ARG_TYPE_PTR, "qpair");
162 0 : spdk_trace_register_description("RDMA_REQ_COMPLETING",
163 : TRACE_RDMA_REQUEST_STATE_COMPLETING,
164 : OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0,
165 : SPDK_TRACE_ARG_TYPE_PTR, "qpair");
166 0 : spdk_trace_register_description("RDMA_REQ_COMPLETED",
167 : TRACE_RDMA_REQUEST_STATE_COMPLETED,
168 : OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0,
169 : SPDK_TRACE_ARG_TYPE_PTR, "qpair");
170 :
171 0 : spdk_trace_register_description("RDMA_QP_CREATE", TRACE_RDMA_QP_CREATE,
172 : OWNER_NONE, OBJECT_NONE, 0,
173 : SPDK_TRACE_ARG_TYPE_INT, "");
174 0 : spdk_trace_register_description("RDMA_IBV_ASYNC_EVENT", TRACE_RDMA_IBV_ASYNC_EVENT,
175 : OWNER_NONE, OBJECT_NONE, 0,
176 : SPDK_TRACE_ARG_TYPE_INT, "type");
177 0 : spdk_trace_register_description("RDMA_CM_ASYNC_EVENT", TRACE_RDMA_CM_ASYNC_EVENT,
178 : OWNER_NONE, OBJECT_NONE, 0,
179 : SPDK_TRACE_ARG_TYPE_INT, "type");
180 0 : spdk_trace_register_description("RDMA_QP_STATE_CHANGE", TRACE_RDMA_QP_STATE_CHANGE,
181 : OWNER_NONE, OBJECT_NONE, 0,
182 : SPDK_TRACE_ARG_TYPE_PTR, "state");
183 0 : spdk_trace_register_description("RDMA_QP_DISCONNECT", TRACE_RDMA_QP_DISCONNECT,
184 : OWNER_NONE, OBJECT_NONE, 0,
185 : SPDK_TRACE_ARG_TYPE_INT, "");
186 0 : spdk_trace_register_description("RDMA_QP_DESTROY", TRACE_RDMA_QP_DESTROY,
187 : OWNER_NONE, OBJECT_NONE, 0,
188 : SPDK_TRACE_ARG_TYPE_INT, "");
189 0 : }
190 :
191 : enum spdk_nvmf_rdma_wr_type {
192 : RDMA_WR_TYPE_RECV,
193 : RDMA_WR_TYPE_SEND,
194 : RDMA_WR_TYPE_DATA,
195 : };
196 :
197 : struct spdk_nvmf_rdma_wr {
198 : /* Uses enum spdk_nvmf_rdma_wr_type */
199 : uint8_t type;
200 : };
201 :
202 : /* This structure holds commands as they are received off the wire.
203 : * It must be dynamically paired with a full request object
204 : * (spdk_nvmf_rdma_request) to service a request. It is separate
205 : * from the request because RDMA does not appear to order
206 : * completions, so occasionally we'll get a new incoming
207 : * command when there aren't any free request objects.
208 : */
209 : struct spdk_nvmf_rdma_recv {
210 : struct ibv_recv_wr wr;
211 : struct ibv_sge sgl[NVMF_DEFAULT_RX_SGE];
212 :
213 : struct spdk_nvmf_rdma_qpair *qpair;
214 :
215 : /* In-capsule data buffer */
216 : uint8_t *buf;
217 :
218 : struct spdk_nvmf_rdma_wr rdma_wr;
219 : uint64_t receive_tsc;
220 :
221 : STAILQ_ENTRY(spdk_nvmf_rdma_recv) link;
222 : };
223 :
224 : struct spdk_nvmf_rdma_request_data {
225 : struct ibv_send_wr wr;
226 : struct ibv_sge sgl[SPDK_NVMF_MAX_SGL_ENTRIES];
227 : };
228 :
229 : struct spdk_nvmf_rdma_request {
230 : struct spdk_nvmf_request req;
231 :
232 : bool fused_failed;
233 :
234 : struct spdk_nvmf_rdma_wr data_wr;
235 : struct spdk_nvmf_rdma_wr rsp_wr;
236 :
237 : /* Uses enum spdk_nvmf_rdma_request_state */
238 : uint8_t state;
239 :
240 : /* Data offset in req.iov */
241 : uint32_t offset;
242 :
243 : struct spdk_nvmf_rdma_recv *recv;
244 :
245 : struct {
246 : struct ibv_send_wr wr;
247 : struct ibv_sge sgl[NVMF_DEFAULT_RSP_SGE];
248 : } rsp;
249 :
250 : uint16_t iovpos;
251 : uint16_t num_outstanding_data_wr;
252 : /* Used to split Write IO with multi SGL payload */
253 : uint16_t num_remaining_data_wr;
254 : uint64_t receive_tsc;
255 : struct spdk_nvmf_rdma_request *fused_pair;
256 : STAILQ_ENTRY(spdk_nvmf_rdma_request) state_link;
257 : struct ibv_send_wr *remaining_tranfer_in_wrs;
258 : struct ibv_send_wr *transfer_wr;
259 : struct spdk_nvmf_rdma_request_data data;
260 : };
261 :
262 : struct spdk_nvmf_rdma_resource_opts {
263 : struct spdk_nvmf_rdma_qpair *qpair;
264 : /* qp points either to an ibv_qp object or an ibv_srq object depending on the value of shared. */
265 : void *qp;
266 : struct spdk_rdma_mem_map *map;
267 : uint32_t max_queue_depth;
268 : uint32_t in_capsule_data_size;
269 : bool shared;
270 : };
271 :
272 : struct spdk_nvmf_rdma_resources {
273 : /* Array of size "max_queue_depth" containing RDMA requests. */
274 : struct spdk_nvmf_rdma_request *reqs;
275 :
276 : /* Array of size "max_queue_depth" containing RDMA recvs. */
277 : struct spdk_nvmf_rdma_recv *recvs;
278 :
279 : /* Array of size "max_queue_depth" containing 64 byte capsules
280 : * used for receive.
281 : */
282 : union nvmf_h2c_msg *cmds;
283 :
284 : /* Array of size "max_queue_depth" containing 16 byte completions
285 : * to be sent back to the user.
286 : */
287 : union nvmf_c2h_msg *cpls;
288 :
289 : /* Array of size "max_queue_depth * InCapsuleDataSize" containing
290 : * buffers to be used for in capsule data.
291 : */
292 : void *bufs;
293 :
294 : /* Receives that are waiting for a request object */
295 : STAILQ_HEAD(, spdk_nvmf_rdma_recv) incoming_queue;
296 :
297 : /* Queue to track free requests */
298 : STAILQ_HEAD(, spdk_nvmf_rdma_request) free_queue;
299 : };
300 :
301 : typedef void (*spdk_nvmf_rdma_qpair_ibv_event)(struct spdk_nvmf_rdma_qpair *rqpair);
302 :
303 : typedef void (*spdk_poller_destroy_cb)(void *ctx);
304 :
305 : struct spdk_nvmf_rdma_ibv_event_ctx {
306 : struct spdk_nvmf_rdma_qpair *rqpair;
307 : spdk_nvmf_rdma_qpair_ibv_event cb_fn;
308 : /* Link to other ibv events associated with this qpair */
309 : STAILQ_ENTRY(spdk_nvmf_rdma_ibv_event_ctx) link;
310 : };
311 :
312 : struct spdk_nvmf_rdma_qpair {
313 : struct spdk_nvmf_qpair qpair;
314 :
315 : struct spdk_nvmf_rdma_device *device;
316 : struct spdk_nvmf_rdma_poller *poller;
317 :
318 : struct spdk_rdma_qp *rdma_qp;
319 : struct rdma_cm_id *cm_id;
320 : struct spdk_rdma_srq *srq;
321 : struct rdma_cm_id *listen_id;
322 :
323 : /* Cache the QP number to improve QP search by RB tree. */
324 : uint32_t qp_num;
325 :
326 : /* The maximum number of I/O outstanding on this connection at one time */
327 : uint16_t max_queue_depth;
328 :
329 : /* The maximum number of active RDMA READ and ATOMIC operations at one time */
330 : uint16_t max_read_depth;
331 :
332 : /* The maximum number of RDMA SEND operations at one time */
333 : uint32_t max_send_depth;
334 :
335 : /* The current number of outstanding WRs from this qpair's
336 : * recv queue. Should not exceed device->attr.max_queue_depth.
337 : */
338 : uint16_t current_recv_depth;
339 :
340 : /* The current number of active RDMA READ operations */
341 : uint16_t current_read_depth;
342 :
343 : /* The current number of posted WRs from this qpair's
344 : * send queue. Should not exceed max_send_depth.
345 : */
346 : uint32_t current_send_depth;
347 :
348 : /* The maximum number of SGEs per WR on the send queue */
349 : uint32_t max_send_sge;
350 :
351 : /* The maximum number of SGEs per WR on the recv queue */
352 : uint32_t max_recv_sge;
353 :
354 : struct spdk_nvmf_rdma_resources *resources;
355 :
356 : STAILQ_HEAD(, spdk_nvmf_rdma_request) pending_rdma_read_queue;
357 :
358 : STAILQ_HEAD(, spdk_nvmf_rdma_request) pending_rdma_write_queue;
359 :
360 : STAILQ_HEAD(, spdk_nvmf_rdma_request) pending_rdma_send_queue;
361 :
362 : /* Number of requests not in the free state */
363 : uint32_t qd;
364 :
365 : RB_ENTRY(spdk_nvmf_rdma_qpair) node;
366 :
367 : STAILQ_ENTRY(spdk_nvmf_rdma_qpair) recv_link;
368 :
369 : STAILQ_ENTRY(spdk_nvmf_rdma_qpair) send_link;
370 :
371 : /* IBV queue pair attributes: they are used to manage
372 : * qp state and recover from errors.
373 : */
374 : enum ibv_qp_state ibv_state;
375 :
376 : /* Points to the a request that has fuse bits set to
377 : * SPDK_NVME_CMD_FUSE_FIRST, when the qpair is waiting
378 : * for the request that has SPDK_NVME_CMD_FUSE_SECOND.
379 : */
380 : struct spdk_nvmf_rdma_request *fused_first;
381 :
382 : /*
383 : * io_channel which is used to destroy qpair when it is removed from poll group
384 : */
385 : struct spdk_io_channel *destruct_channel;
386 :
387 : /* List of ibv async events */
388 : STAILQ_HEAD(, spdk_nvmf_rdma_ibv_event_ctx) ibv_events;
389 :
390 : /* Lets us know that we have received the last_wqe event. */
391 : bool last_wqe_reached;
392 :
393 : /* Indicate that nvmf_rdma_close_qpair is called */
394 : bool to_close;
395 : };
396 :
397 : struct spdk_nvmf_rdma_poller_stat {
398 : uint64_t completions;
399 : uint64_t polls;
400 : uint64_t idle_polls;
401 : uint64_t requests;
402 : uint64_t request_latency;
403 : uint64_t pending_free_request;
404 : uint64_t pending_rdma_read;
405 : uint64_t pending_rdma_write;
406 : uint64_t pending_rdma_send;
407 : struct spdk_rdma_qp_stats qp_stats;
408 : };
409 :
410 : struct spdk_nvmf_rdma_poller {
411 : struct spdk_nvmf_rdma_device *device;
412 : struct spdk_nvmf_rdma_poll_group *group;
413 :
414 : int num_cqe;
415 : int required_num_wr;
416 : struct ibv_cq *cq;
417 :
418 : /* The maximum number of I/O outstanding on the shared receive queue at one time */
419 : uint16_t max_srq_depth;
420 : bool need_destroy;
421 :
422 : /* Shared receive queue */
423 : struct spdk_rdma_srq *srq;
424 :
425 : struct spdk_nvmf_rdma_resources *resources;
426 : struct spdk_nvmf_rdma_poller_stat stat;
427 :
428 : spdk_poller_destroy_cb destroy_cb;
429 : void *destroy_cb_ctx;
430 :
431 : RB_HEAD(qpairs_tree, spdk_nvmf_rdma_qpair) qpairs;
432 :
433 : STAILQ_HEAD(, spdk_nvmf_rdma_qpair) qpairs_pending_recv;
434 :
435 : STAILQ_HEAD(, spdk_nvmf_rdma_qpair) qpairs_pending_send;
436 :
437 : TAILQ_ENTRY(spdk_nvmf_rdma_poller) link;
438 : };
439 :
440 : struct spdk_nvmf_rdma_poll_group_stat {
441 : uint64_t pending_data_buffer;
442 : };
443 :
444 : struct spdk_nvmf_rdma_poll_group {
445 : struct spdk_nvmf_transport_poll_group group;
446 : struct spdk_nvmf_rdma_poll_group_stat stat;
447 : TAILQ_HEAD(, spdk_nvmf_rdma_poller) pollers;
448 : TAILQ_ENTRY(spdk_nvmf_rdma_poll_group) link;
449 : };
450 :
451 : struct spdk_nvmf_rdma_conn_sched {
452 : struct spdk_nvmf_rdma_poll_group *next_admin_pg;
453 : struct spdk_nvmf_rdma_poll_group *next_io_pg;
454 : };
455 :
456 : /* Assuming rdma_cm uses just one protection domain per ibv_context. */
457 : struct spdk_nvmf_rdma_device {
458 : struct ibv_device_attr attr;
459 : struct ibv_context *context;
460 :
461 : struct spdk_rdma_mem_map *map;
462 : struct ibv_pd *pd;
463 :
464 : int num_srq;
465 : bool need_destroy;
466 : bool ready_to_destroy;
467 : bool is_ready;
468 :
469 : TAILQ_ENTRY(spdk_nvmf_rdma_device) link;
470 : };
471 :
472 : struct spdk_nvmf_rdma_port {
473 : const struct spdk_nvme_transport_id *trid;
474 : struct rdma_cm_id *id;
475 : struct spdk_nvmf_rdma_device *device;
476 : TAILQ_ENTRY(spdk_nvmf_rdma_port) link;
477 : };
478 :
479 : struct rdma_transport_opts {
480 : int num_cqe;
481 : uint32_t max_srq_depth;
482 : bool no_srq;
483 : bool no_wr_batching;
484 : int acceptor_backlog;
485 : };
486 :
487 : struct spdk_nvmf_rdma_transport {
488 : struct spdk_nvmf_transport transport;
489 : struct rdma_transport_opts rdma_opts;
490 :
491 : struct spdk_nvmf_rdma_conn_sched conn_sched;
492 :
493 : struct rdma_event_channel *event_channel;
494 :
495 : struct spdk_mempool *data_wr_pool;
496 :
497 : struct spdk_poller *accept_poller;
498 :
499 : /* fields used to poll RDMA/IB events */
500 : nfds_t npoll_fds;
501 : struct pollfd *poll_fds;
502 :
503 : TAILQ_HEAD(, spdk_nvmf_rdma_device) devices;
504 : TAILQ_HEAD(, spdk_nvmf_rdma_port) ports;
505 : TAILQ_HEAD(, spdk_nvmf_rdma_poll_group) poll_groups;
506 :
507 : /* ports that are removed unexpectedly and need retry listen */
508 : TAILQ_HEAD(, spdk_nvmf_rdma_port) retry_ports;
509 : };
510 :
511 : struct poller_manage_ctx {
512 : struct spdk_nvmf_rdma_transport *rtransport;
513 : struct spdk_nvmf_rdma_poll_group *rgroup;
514 : struct spdk_nvmf_rdma_poller *rpoller;
515 : struct spdk_nvmf_rdma_device *device;
516 :
517 : struct spdk_thread *thread;
518 : volatile int *inflight_op_counter;
519 : };
520 :
521 : static const struct spdk_json_object_decoder rdma_transport_opts_decoder[] = {
522 : {
523 : "num_cqe", offsetof(struct rdma_transport_opts, num_cqe),
524 : spdk_json_decode_int32, true
525 : },
526 : {
527 : "max_srq_depth", offsetof(struct rdma_transport_opts, max_srq_depth),
528 : spdk_json_decode_uint32, true
529 : },
530 : {
531 : "no_srq", offsetof(struct rdma_transport_opts, no_srq),
532 : spdk_json_decode_bool, true
533 : },
534 : {
535 : "no_wr_batching", offsetof(struct rdma_transport_opts, no_wr_batching),
536 : spdk_json_decode_bool, true
537 : },
538 : {
539 : "acceptor_backlog", offsetof(struct rdma_transport_opts, acceptor_backlog),
540 : spdk_json_decode_int32, true
541 : },
542 : };
543 :
544 : static int
545 2 : nvmf_rdma_qpair_compare(struct spdk_nvmf_rdma_qpair *rqpair1, struct spdk_nvmf_rdma_qpair *rqpair2)
546 : {
547 2 : return rqpair1->qp_num < rqpair2->qp_num ? -1 : rqpair1->qp_num > rqpair2->qp_num;
548 : }
549 :
550 0 : RB_GENERATE_STATIC(qpairs_tree, spdk_nvmf_rdma_qpair, node, nvmf_rdma_qpair_compare);
551 :
552 : static bool nvmf_rdma_request_process(struct spdk_nvmf_rdma_transport *rtransport,
553 : struct spdk_nvmf_rdma_request *rdma_req);
554 :
555 : static void _poller_submit_sends(struct spdk_nvmf_rdma_transport *rtransport,
556 : struct spdk_nvmf_rdma_poller *rpoller);
557 :
558 : static void _poller_submit_recvs(struct spdk_nvmf_rdma_transport *rtransport,
559 : struct spdk_nvmf_rdma_poller *rpoller);
560 :
561 : static void _nvmf_rdma_remove_destroyed_device(void *c);
562 :
563 : static inline int
564 2 : nvmf_rdma_check_ibv_state(enum ibv_qp_state state)
565 : {
566 2 : switch (state) {
567 1 : case IBV_QPS_RESET:
568 : case IBV_QPS_INIT:
569 : case IBV_QPS_RTR:
570 : case IBV_QPS_RTS:
571 : case IBV_QPS_SQD:
572 : case IBV_QPS_SQE:
573 : case IBV_QPS_ERR:
574 1 : return 0;
575 1 : default:
576 1 : return -1;
577 : }
578 : }
579 :
580 : static inline enum spdk_nvme_media_error_status_code
581 0 : nvmf_rdma_dif_error_to_compl_status(uint8_t err_type) {
582 : enum spdk_nvme_media_error_status_code result;
583 0 : switch (err_type)
584 : {
585 0 : case SPDK_DIF_REFTAG_ERROR:
586 0 : result = SPDK_NVME_SC_REFERENCE_TAG_CHECK_ERROR;
587 0 : break;
588 0 : case SPDK_DIF_APPTAG_ERROR:
589 0 : result = SPDK_NVME_SC_APPLICATION_TAG_CHECK_ERROR;
590 0 : break;
591 0 : case SPDK_DIF_GUARD_ERROR:
592 0 : result = SPDK_NVME_SC_GUARD_CHECK_ERROR;
593 0 : break;
594 0 : default:
595 0 : SPDK_UNREACHABLE();
596 : }
597 :
598 0 : return result;
599 : }
600 :
601 : static enum ibv_qp_state
602 3 : nvmf_rdma_update_ibv_state(struct spdk_nvmf_rdma_qpair *rqpair) {
603 : enum ibv_qp_state old_state, new_state;
604 3 : struct ibv_qp_attr qp_attr;
605 3 : struct ibv_qp_init_attr init_attr;
606 : int rc;
607 :
608 3 : old_state = rqpair->ibv_state;
609 3 : rc = ibv_query_qp(rqpair->rdma_qp->qp, &qp_attr,
610 : g_spdk_nvmf_ibv_query_mask, &init_attr);
611 :
612 3 : if (rc)
613 : {
614 1 : SPDK_ERRLOG("Failed to get updated RDMA queue pair state!\n");
615 1 : return IBV_QPS_ERR + 1;
616 : }
617 :
618 2 : new_state = qp_attr.qp_state;
619 2 : rqpair->ibv_state = new_state;
620 2 : qp_attr.ah_attr.port_num = qp_attr.port_num;
621 :
622 2 : rc = nvmf_rdma_check_ibv_state(new_state);
623 2 : if (rc)
624 : {
625 1 : SPDK_ERRLOG("QP#%d: bad state updated: %u, maybe hardware issue\n", rqpair->qpair.qid, new_state);
626 : /*
627 : * IBV_QPS_UNKNOWN undefined if lib version smaller than libibverbs-1.1.8
628 : * IBV_QPS_UNKNOWN is the enum element after IBV_QPS_ERR
629 : */
630 1 : return IBV_QPS_ERR + 1;
631 : }
632 :
633 1 : if (old_state != new_state)
634 : {
635 1 : spdk_trace_record(TRACE_RDMA_QP_STATE_CHANGE, 0, 0, (uintptr_t)rqpair, new_state);
636 : }
637 1 : return new_state;
638 : }
639 :
640 : /*
641 : * Return data_wrs to pool starting from \b data_wr
642 : * Request's own response and data WR are excluded
643 : */
644 : static void
645 7 : _nvmf_rdma_request_free_data(struct spdk_nvmf_rdma_request *rdma_req,
646 : struct ibv_send_wr *data_wr,
647 : struct spdk_mempool *pool)
648 : {
649 7 : struct spdk_nvmf_rdma_request_data *work_requests[SPDK_NVMF_MAX_SGL_ENTRIES];
650 : struct spdk_nvmf_rdma_request_data *nvmf_data;
651 : struct ibv_send_wr *next_send_wr;
652 7 : uint64_t req_wrid = (uint64_t)&rdma_req->data_wr;
653 7 : uint32_t num_wrs = 0;
654 :
655 15 : while (data_wr && data_wr->wr_id == req_wrid) {
656 8 : nvmf_data = SPDK_CONTAINEROF(data_wr, struct spdk_nvmf_rdma_request_data, wr);
657 8 : memset(nvmf_data->sgl, 0, sizeof(data_wr->sg_list[0]) * data_wr->num_sge);
658 8 : data_wr->num_sge = 0;
659 8 : next_send_wr = data_wr->next;
660 8 : if (data_wr != &rdma_req->data.wr) {
661 1 : data_wr->next = NULL;
662 1 : assert(num_wrs < SPDK_NVMF_MAX_SGL_ENTRIES);
663 1 : work_requests[num_wrs] = nvmf_data;
664 1 : num_wrs++;
665 : }
666 8 : data_wr = (!next_send_wr || next_send_wr == &rdma_req->rsp.wr) ? NULL : next_send_wr;
667 : }
668 :
669 7 : if (num_wrs) {
670 1 : spdk_mempool_put_bulk(pool, (void **) work_requests, num_wrs);
671 : }
672 7 : }
673 :
674 : static void
675 7 : nvmf_rdma_request_free_data(struct spdk_nvmf_rdma_request *rdma_req,
676 : struct spdk_nvmf_rdma_transport *rtransport)
677 : {
678 7 : rdma_req->num_outstanding_data_wr = 0;
679 :
680 7 : _nvmf_rdma_request_free_data(rdma_req, rdma_req->transfer_wr, rtransport->data_wr_pool);
681 :
682 7 : rdma_req->data.wr.next = NULL;
683 7 : rdma_req->rsp.wr.next = NULL;
684 7 : }
685 :
686 : static void
687 0 : nvmf_rdma_dump_request(struct spdk_nvmf_rdma_request *req)
688 : {
689 0 : SPDK_ERRLOG("\t\tRequest Data From Pool: %d\n", req->req.data_from_pool);
690 0 : if (req->req.cmd) {
691 0 : SPDK_ERRLOG("\t\tRequest opcode: %d\n", req->req.cmd->nvmf_cmd.opcode);
692 : }
693 0 : if (req->recv) {
694 0 : SPDK_ERRLOG("\t\tRequest recv wr_id%lu\n", req->recv->wr.wr_id);
695 : }
696 0 : }
697 :
698 : static void
699 0 : nvmf_rdma_dump_qpair_contents(struct spdk_nvmf_rdma_qpair *rqpair)
700 : {
701 : int i;
702 :
703 0 : SPDK_ERRLOG("Dumping contents of queue pair (QID %d)\n", rqpair->qpair.qid);
704 0 : for (i = 0; i < rqpair->max_queue_depth; i++) {
705 0 : if (rqpair->resources->reqs[i].state != RDMA_REQUEST_STATE_FREE) {
706 0 : nvmf_rdma_dump_request(&rqpair->resources->reqs[i]);
707 : }
708 : }
709 0 : }
710 :
711 : static void
712 1 : nvmf_rdma_resources_destroy(struct spdk_nvmf_rdma_resources *resources)
713 : {
714 1 : spdk_free(resources->cmds);
715 1 : spdk_free(resources->cpls);
716 1 : spdk_free(resources->bufs);
717 1 : spdk_free(resources->reqs);
718 1 : spdk_free(resources->recvs);
719 1 : free(resources);
720 1 : }
721 :
722 :
723 : static struct spdk_nvmf_rdma_resources *
724 1 : nvmf_rdma_resources_create(struct spdk_nvmf_rdma_resource_opts *opts)
725 : {
726 : struct spdk_nvmf_rdma_resources *resources;
727 : struct spdk_nvmf_rdma_request *rdma_req;
728 : struct spdk_nvmf_rdma_recv *rdma_recv;
729 1 : struct spdk_rdma_qp *qp = NULL;
730 1 : struct spdk_rdma_srq *srq = NULL;
731 1 : struct ibv_recv_wr *bad_wr = NULL;
732 1 : struct spdk_rdma_memory_translation translation;
733 : uint32_t i;
734 1 : int rc = 0;
735 :
736 1 : resources = calloc(1, sizeof(struct spdk_nvmf_rdma_resources));
737 1 : if (!resources) {
738 0 : SPDK_ERRLOG("Unable to allocate resources for receive queue.\n");
739 0 : return NULL;
740 : }
741 :
742 1 : resources->reqs = spdk_zmalloc(opts->max_queue_depth * sizeof(*resources->reqs),
743 : 0x1000, NULL, SPDK_ENV_LCORE_ID_ANY, SPDK_MALLOC_DMA);
744 1 : resources->recvs = spdk_zmalloc(opts->max_queue_depth * sizeof(*resources->recvs),
745 : 0x1000, NULL, SPDK_ENV_LCORE_ID_ANY, SPDK_MALLOC_DMA);
746 1 : resources->cmds = spdk_zmalloc(opts->max_queue_depth * sizeof(*resources->cmds),
747 : 0x1000, NULL, SPDK_ENV_LCORE_ID_ANY, SPDK_MALLOC_DMA);
748 1 : resources->cpls = spdk_zmalloc(opts->max_queue_depth * sizeof(*resources->cpls),
749 : 0x1000, NULL, SPDK_ENV_LCORE_ID_ANY, SPDK_MALLOC_DMA);
750 :
751 1 : if (opts->in_capsule_data_size > 0) {
752 1 : resources->bufs = spdk_zmalloc(opts->max_queue_depth * opts->in_capsule_data_size,
753 : 0x1000, NULL, SPDK_ENV_LCORE_ID_ANY,
754 : SPDK_MALLOC_DMA);
755 : }
756 :
757 1 : if (!resources->reqs || !resources->recvs || !resources->cmds ||
758 1 : !resources->cpls || (opts->in_capsule_data_size && !resources->bufs)) {
759 0 : SPDK_ERRLOG("Unable to allocate sufficient memory for RDMA queue.\n");
760 0 : goto cleanup;
761 : }
762 :
763 1 : SPDK_DEBUGLOG(rdma, "Command Array: %p Length: %lx\n",
764 : resources->cmds, opts->max_queue_depth * sizeof(*resources->cmds));
765 1 : SPDK_DEBUGLOG(rdma, "Completion Array: %p Length: %lx\n",
766 : resources->cpls, opts->max_queue_depth * sizeof(*resources->cpls));
767 1 : if (resources->bufs) {
768 1 : SPDK_DEBUGLOG(rdma, "In Capsule Data Array: %p Length: %x\n",
769 : resources->bufs, opts->max_queue_depth *
770 : opts->in_capsule_data_size);
771 : }
772 :
773 : /* Initialize queues */
774 1 : STAILQ_INIT(&resources->incoming_queue);
775 1 : STAILQ_INIT(&resources->free_queue);
776 :
777 1 : if (opts->shared) {
778 1 : srq = (struct spdk_rdma_srq *)opts->qp;
779 : } else {
780 0 : qp = (struct spdk_rdma_qp *)opts->qp;
781 : }
782 :
783 129 : for (i = 0; i < opts->max_queue_depth; i++) {
784 128 : rdma_recv = &resources->recvs[i];
785 128 : rdma_recv->qpair = opts->qpair;
786 :
787 : /* Set up memory to receive commands */
788 128 : if (resources->bufs) {
789 128 : rdma_recv->buf = (void *)((uintptr_t)resources->bufs + (i *
790 128 : opts->in_capsule_data_size));
791 : }
792 :
793 128 : rdma_recv->rdma_wr.type = RDMA_WR_TYPE_RECV;
794 :
795 128 : rdma_recv->sgl[0].addr = (uintptr_t)&resources->cmds[i];
796 128 : rdma_recv->sgl[0].length = sizeof(resources->cmds[i]);
797 128 : rc = spdk_rdma_get_translation(opts->map, &resources->cmds[i], sizeof(resources->cmds[i]),
798 : &translation);
799 128 : if (rc) {
800 0 : goto cleanup;
801 : }
802 128 : rdma_recv->sgl[0].lkey = spdk_rdma_memory_translation_get_lkey(&translation);
803 128 : rdma_recv->wr.num_sge = 1;
804 :
805 128 : if (rdma_recv->buf) {
806 128 : rdma_recv->sgl[1].addr = (uintptr_t)rdma_recv->buf;
807 128 : rdma_recv->sgl[1].length = opts->in_capsule_data_size;
808 128 : rc = spdk_rdma_get_translation(opts->map, rdma_recv->buf, opts->in_capsule_data_size, &translation);
809 128 : if (rc) {
810 0 : goto cleanup;
811 : }
812 128 : rdma_recv->sgl[1].lkey = spdk_rdma_memory_translation_get_lkey(&translation);
813 128 : rdma_recv->wr.num_sge++;
814 : }
815 :
816 128 : rdma_recv->wr.wr_id = (uintptr_t)&rdma_recv->rdma_wr;
817 128 : rdma_recv->wr.sg_list = rdma_recv->sgl;
818 128 : if (srq) {
819 0 : spdk_rdma_srq_queue_recv_wrs(srq, &rdma_recv->wr);
820 : } else {
821 128 : spdk_rdma_qp_queue_recv_wrs(qp, &rdma_recv->wr);
822 : }
823 : }
824 :
825 129 : for (i = 0; i < opts->max_queue_depth; i++) {
826 128 : rdma_req = &resources->reqs[i];
827 :
828 128 : if (opts->qpair != NULL) {
829 128 : rdma_req->req.qpair = &opts->qpair->qpair;
830 : } else {
831 0 : rdma_req->req.qpair = NULL;
832 : }
833 128 : rdma_req->req.cmd = NULL;
834 128 : rdma_req->req.iovcnt = 0;
835 128 : rdma_req->req.stripped_data = NULL;
836 :
837 : /* Set up memory to send responses */
838 128 : rdma_req->req.rsp = &resources->cpls[i];
839 :
840 128 : rdma_req->rsp.sgl[0].addr = (uintptr_t)&resources->cpls[i];
841 128 : rdma_req->rsp.sgl[0].length = sizeof(resources->cpls[i]);
842 128 : rc = spdk_rdma_get_translation(opts->map, &resources->cpls[i], sizeof(resources->cpls[i]),
843 : &translation);
844 128 : if (rc) {
845 0 : goto cleanup;
846 : }
847 128 : rdma_req->rsp.sgl[0].lkey = spdk_rdma_memory_translation_get_lkey(&translation);
848 :
849 128 : rdma_req->rsp_wr.type = RDMA_WR_TYPE_SEND;
850 128 : rdma_req->rsp.wr.wr_id = (uintptr_t)&rdma_req->rsp_wr;
851 128 : rdma_req->rsp.wr.next = NULL;
852 128 : rdma_req->rsp.wr.opcode = IBV_WR_SEND;
853 128 : rdma_req->rsp.wr.send_flags = IBV_SEND_SIGNALED;
854 128 : rdma_req->rsp.wr.sg_list = rdma_req->rsp.sgl;
855 128 : rdma_req->rsp.wr.num_sge = SPDK_COUNTOF(rdma_req->rsp.sgl);
856 :
857 : /* Set up memory for data buffers */
858 128 : rdma_req->data_wr.type = RDMA_WR_TYPE_DATA;
859 128 : rdma_req->data.wr.wr_id = (uintptr_t)&rdma_req->data_wr;
860 128 : rdma_req->data.wr.next = NULL;
861 128 : rdma_req->data.wr.send_flags = IBV_SEND_SIGNALED;
862 128 : rdma_req->data.wr.sg_list = rdma_req->data.sgl;
863 128 : rdma_req->data.wr.num_sge = SPDK_COUNTOF(rdma_req->data.sgl);
864 :
865 : /* Initialize request state to FREE */
866 128 : rdma_req->state = RDMA_REQUEST_STATE_FREE;
867 128 : STAILQ_INSERT_TAIL(&resources->free_queue, rdma_req, state_link);
868 : }
869 :
870 1 : if (srq) {
871 0 : rc = spdk_rdma_srq_flush_recv_wrs(srq, &bad_wr);
872 : } else {
873 1 : rc = spdk_rdma_qp_flush_recv_wrs(qp, &bad_wr);
874 : }
875 :
876 1 : if (rc) {
877 0 : goto cleanup;
878 : }
879 :
880 1 : return resources;
881 :
882 0 : cleanup:
883 0 : nvmf_rdma_resources_destroy(resources);
884 0 : return NULL;
885 : }
886 :
887 : static void
888 0 : nvmf_rdma_qpair_clean_ibv_events(struct spdk_nvmf_rdma_qpair *rqpair)
889 : {
890 : struct spdk_nvmf_rdma_ibv_event_ctx *ctx, *tctx;
891 0 : STAILQ_FOREACH_SAFE(ctx, &rqpair->ibv_events, link, tctx) {
892 0 : ctx->rqpair = NULL;
893 : /* Memory allocated for ctx is freed in nvmf_rdma_qpair_process_ibv_event */
894 0 : STAILQ_REMOVE(&rqpair->ibv_events, ctx, spdk_nvmf_rdma_ibv_event_ctx, link);
895 : }
896 0 : }
897 :
898 : static void nvmf_rdma_poller_destroy(struct spdk_nvmf_rdma_poller *poller);
899 :
900 : static void
901 0 : nvmf_rdma_qpair_destroy(struct spdk_nvmf_rdma_qpair *rqpair)
902 : {
903 : struct spdk_nvmf_rdma_recv *rdma_recv, *recv_tmp;
904 0 : struct ibv_recv_wr *bad_recv_wr = NULL;
905 : int rc;
906 :
907 0 : spdk_trace_record(TRACE_RDMA_QP_DESTROY, 0, 0, (uintptr_t)rqpair);
908 :
909 0 : if (rqpair->qd != 0) {
910 0 : struct spdk_nvmf_qpair *qpair = &rqpair->qpair;
911 0 : struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(qpair->transport,
912 : struct spdk_nvmf_rdma_transport, transport);
913 : struct spdk_nvmf_rdma_request *req;
914 0 : uint32_t i, max_req_count = 0;
915 :
916 0 : SPDK_WARNLOG("Destroying qpair when queue depth is %d\n", rqpair->qd);
917 :
918 0 : if (rqpair->srq == NULL) {
919 0 : nvmf_rdma_dump_qpair_contents(rqpair);
920 0 : max_req_count = rqpair->max_queue_depth;
921 0 : } else if (rqpair->poller && rqpair->resources) {
922 0 : max_req_count = rqpair->poller->max_srq_depth;
923 : }
924 :
925 0 : SPDK_DEBUGLOG(rdma, "Release incomplete requests\n");
926 0 : for (i = 0; i < max_req_count; i++) {
927 0 : req = &rqpair->resources->reqs[i];
928 0 : if (req->req.qpair == qpair && req->state != RDMA_REQUEST_STATE_FREE) {
929 : /* nvmf_rdma_request_process checks qpair ibv and internal state
930 : * and completes a request */
931 0 : nvmf_rdma_request_process(rtransport, req);
932 : }
933 : }
934 0 : assert(rqpair->qd == 0);
935 : }
936 :
937 0 : if (rqpair->poller) {
938 0 : RB_REMOVE(qpairs_tree, &rqpair->poller->qpairs, rqpair);
939 :
940 0 : if (rqpair->srq != NULL && rqpair->resources != NULL) {
941 : /* Drop all received but unprocessed commands for this queue and return them to SRQ */
942 0 : STAILQ_FOREACH_SAFE(rdma_recv, &rqpair->resources->incoming_queue, link, recv_tmp) {
943 0 : if (rqpair == rdma_recv->qpair) {
944 0 : STAILQ_REMOVE(&rqpair->resources->incoming_queue, rdma_recv, spdk_nvmf_rdma_recv, link);
945 0 : spdk_rdma_srq_queue_recv_wrs(rqpair->srq, &rdma_recv->wr);
946 0 : rc = spdk_rdma_srq_flush_recv_wrs(rqpair->srq, &bad_recv_wr);
947 0 : if (rc) {
948 0 : SPDK_ERRLOG("Unable to re-post rx descriptor\n");
949 : }
950 : }
951 : }
952 : }
953 : }
954 :
955 0 : if (rqpair->cm_id) {
956 0 : if (rqpair->rdma_qp != NULL) {
957 0 : spdk_rdma_qp_destroy(rqpair->rdma_qp);
958 0 : rqpair->rdma_qp = NULL;
959 : }
960 :
961 0 : if (rqpair->poller != NULL && rqpair->srq == NULL) {
962 0 : rqpair->poller->required_num_wr -= MAX_WR_PER_QP(rqpair->max_queue_depth);
963 : }
964 : }
965 :
966 0 : if (rqpair->srq == NULL && rqpair->resources != NULL) {
967 0 : nvmf_rdma_resources_destroy(rqpair->resources);
968 : }
969 :
970 0 : nvmf_rdma_qpair_clean_ibv_events(rqpair);
971 :
972 0 : if (rqpair->destruct_channel) {
973 0 : spdk_put_io_channel(rqpair->destruct_channel);
974 0 : rqpair->destruct_channel = NULL;
975 : }
976 :
977 0 : if (rqpair->poller && rqpair->poller->need_destroy && RB_EMPTY(&rqpair->poller->qpairs)) {
978 0 : nvmf_rdma_poller_destroy(rqpair->poller);
979 : }
980 :
981 : /* destroy cm_id last so cma device will not be freed before we destroy the cq. */
982 0 : if (rqpair->cm_id) {
983 0 : rdma_destroy_id(rqpair->cm_id);
984 : }
985 :
986 0 : free(rqpair);
987 0 : }
988 :
989 : static int
990 5 : nvmf_rdma_resize_cq(struct spdk_nvmf_rdma_qpair *rqpair, struct spdk_nvmf_rdma_device *device)
991 : {
992 : struct spdk_nvmf_rdma_poller *rpoller;
993 : int rc, num_cqe, required_num_wr;
994 :
995 : /* Enlarge CQ size dynamically */
996 5 : rpoller = rqpair->poller;
997 5 : required_num_wr = rpoller->required_num_wr + MAX_WR_PER_QP(rqpair->max_queue_depth);
998 5 : num_cqe = rpoller->num_cqe;
999 5 : if (num_cqe < required_num_wr) {
1000 4 : num_cqe = spdk_max(num_cqe * 2, required_num_wr);
1001 4 : num_cqe = spdk_min(num_cqe, device->attr.max_cqe);
1002 : }
1003 :
1004 5 : if (rpoller->num_cqe != num_cqe) {
1005 4 : if (device->context->device->transport_type == IBV_TRANSPORT_IWARP) {
1006 1 : SPDK_ERRLOG("iWARP doesn't support CQ resize. Current capacity %u, required %u\n"
1007 : "Using CQ of insufficient size may lead to CQ overrun\n", rpoller->num_cqe, num_cqe);
1008 1 : return -1;
1009 : }
1010 3 : if (required_num_wr > device->attr.max_cqe) {
1011 1 : SPDK_ERRLOG("RDMA CQE requirement (%d) exceeds device max_cqe limitation (%d)\n",
1012 : required_num_wr, device->attr.max_cqe);
1013 1 : return -1;
1014 : }
1015 :
1016 2 : SPDK_DEBUGLOG(rdma, "Resize RDMA CQ from %d to %d\n", rpoller->num_cqe, num_cqe);
1017 2 : rc = ibv_resize_cq(rpoller->cq, num_cqe);
1018 2 : if (rc) {
1019 1 : SPDK_ERRLOG("RDMA CQ resize failed: errno %d: %s\n", errno, spdk_strerror(errno));
1020 1 : return -1;
1021 : }
1022 :
1023 1 : rpoller->num_cqe = num_cqe;
1024 : }
1025 :
1026 2 : rpoller->required_num_wr = required_num_wr;
1027 2 : return 0;
1028 : }
1029 :
1030 : static int
1031 0 : nvmf_rdma_qpair_initialize(struct spdk_nvmf_qpair *qpair)
1032 : {
1033 : struct spdk_nvmf_rdma_qpair *rqpair;
1034 : struct spdk_nvmf_rdma_transport *rtransport;
1035 : struct spdk_nvmf_transport *transport;
1036 0 : struct spdk_nvmf_rdma_resource_opts opts;
1037 : struct spdk_nvmf_rdma_device *device;
1038 0 : struct spdk_rdma_qp_init_attr qp_init_attr = {};
1039 :
1040 0 : rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair);
1041 0 : device = rqpair->device;
1042 :
1043 0 : qp_init_attr.qp_context = rqpair;
1044 0 : qp_init_attr.pd = device->pd;
1045 0 : qp_init_attr.send_cq = rqpair->poller->cq;
1046 0 : qp_init_attr.recv_cq = rqpair->poller->cq;
1047 :
1048 0 : if (rqpair->srq) {
1049 0 : qp_init_attr.srq = rqpair->srq->srq;
1050 : } else {
1051 0 : qp_init_attr.cap.max_recv_wr = rqpair->max_queue_depth;
1052 : }
1053 :
1054 : /* SEND, READ, and WRITE operations */
1055 0 : qp_init_attr.cap.max_send_wr = (uint32_t)rqpair->max_queue_depth * 2;
1056 0 : qp_init_attr.cap.max_send_sge = spdk_min((uint32_t)device->attr.max_sge, NVMF_DEFAULT_TX_SGE);
1057 0 : qp_init_attr.cap.max_recv_sge = spdk_min((uint32_t)device->attr.max_sge, NVMF_DEFAULT_RX_SGE);
1058 0 : qp_init_attr.stats = &rqpair->poller->stat.qp_stats;
1059 :
1060 0 : if (rqpair->srq == NULL && nvmf_rdma_resize_cq(rqpair, device) < 0) {
1061 0 : SPDK_ERRLOG("Failed to resize the completion queue. Cannot initialize qpair.\n");
1062 0 : goto error;
1063 : }
1064 :
1065 0 : rqpair->rdma_qp = spdk_rdma_qp_create(rqpair->cm_id, &qp_init_attr);
1066 0 : if (!rqpair->rdma_qp) {
1067 0 : goto error;
1068 : }
1069 :
1070 0 : rqpair->qp_num = rqpair->rdma_qp->qp->qp_num;
1071 :
1072 0 : rqpair->max_send_depth = spdk_min((uint32_t)(rqpair->max_queue_depth * 2),
1073 : qp_init_attr.cap.max_send_wr);
1074 0 : rqpair->max_send_sge = spdk_min(NVMF_DEFAULT_TX_SGE, qp_init_attr.cap.max_send_sge);
1075 0 : rqpair->max_recv_sge = spdk_min(NVMF_DEFAULT_RX_SGE, qp_init_attr.cap.max_recv_sge);
1076 0 : spdk_trace_record(TRACE_RDMA_QP_CREATE, 0, 0, (uintptr_t)rqpair);
1077 0 : SPDK_DEBUGLOG(rdma, "New RDMA Connection: %p\n", qpair);
1078 :
1079 0 : if (rqpair->poller->srq == NULL) {
1080 0 : rtransport = SPDK_CONTAINEROF(qpair->transport, struct spdk_nvmf_rdma_transport, transport);
1081 0 : transport = &rtransport->transport;
1082 :
1083 0 : opts.qp = rqpair->rdma_qp;
1084 0 : opts.map = device->map;
1085 0 : opts.qpair = rqpair;
1086 0 : opts.shared = false;
1087 0 : opts.max_queue_depth = rqpair->max_queue_depth;
1088 0 : opts.in_capsule_data_size = transport->opts.in_capsule_data_size;
1089 :
1090 0 : rqpair->resources = nvmf_rdma_resources_create(&opts);
1091 :
1092 0 : if (!rqpair->resources) {
1093 0 : SPDK_ERRLOG("Unable to allocate resources for receive queue.\n");
1094 0 : rdma_destroy_qp(rqpair->cm_id);
1095 0 : goto error;
1096 : }
1097 : } else {
1098 0 : rqpair->resources = rqpair->poller->resources;
1099 : }
1100 :
1101 0 : rqpair->current_recv_depth = 0;
1102 0 : STAILQ_INIT(&rqpair->pending_rdma_read_queue);
1103 0 : STAILQ_INIT(&rqpair->pending_rdma_write_queue);
1104 0 : STAILQ_INIT(&rqpair->pending_rdma_send_queue);
1105 :
1106 0 : return 0;
1107 :
1108 0 : error:
1109 0 : rdma_destroy_id(rqpair->cm_id);
1110 0 : rqpair->cm_id = NULL;
1111 0 : return -1;
1112 : }
1113 :
1114 : /* Append the given recv wr structure to the resource structs outstanding recvs list. */
1115 : /* This function accepts either a single wr or the first wr in a linked list. */
1116 : static void
1117 6 : nvmf_rdma_qpair_queue_recv_wrs(struct spdk_nvmf_rdma_qpair *rqpair, struct ibv_recv_wr *first)
1118 : {
1119 6 : struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(rqpair->qpair.transport,
1120 : struct spdk_nvmf_rdma_transport, transport);
1121 :
1122 6 : if (rqpair->srq != NULL) {
1123 0 : spdk_rdma_srq_queue_recv_wrs(rqpair->srq, first);
1124 : } else {
1125 6 : if (spdk_rdma_qp_queue_recv_wrs(rqpair->rdma_qp, first)) {
1126 6 : STAILQ_INSERT_TAIL(&rqpair->poller->qpairs_pending_recv, rqpair, recv_link);
1127 : }
1128 : }
1129 :
1130 6 : if (rtransport->rdma_opts.no_wr_batching) {
1131 0 : _poller_submit_recvs(rtransport, rqpair->poller);
1132 : }
1133 6 : }
1134 :
1135 : static int
1136 4 : request_transfer_in(struct spdk_nvmf_request *req)
1137 : {
1138 : struct spdk_nvmf_rdma_request *rdma_req;
1139 : struct spdk_nvmf_qpair *qpair;
1140 : struct spdk_nvmf_rdma_qpair *rqpair;
1141 : struct spdk_nvmf_rdma_transport *rtransport;
1142 :
1143 4 : qpair = req->qpair;
1144 4 : rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req);
1145 4 : rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair);
1146 4 : rtransport = SPDK_CONTAINEROF(rqpair->qpair.transport,
1147 : struct spdk_nvmf_rdma_transport, transport);
1148 :
1149 4 : assert(req->xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER);
1150 4 : assert(rdma_req != NULL);
1151 :
1152 4 : if (spdk_rdma_qp_queue_send_wrs(rqpair->rdma_qp, rdma_req->transfer_wr)) {
1153 4 : STAILQ_INSERT_TAIL(&rqpair->poller->qpairs_pending_send, rqpair, send_link);
1154 : }
1155 4 : if (rtransport->rdma_opts.no_wr_batching) {
1156 0 : _poller_submit_sends(rtransport, rqpair->poller);
1157 : }
1158 :
1159 4 : assert(rqpair->current_read_depth + rdma_req->num_outstanding_data_wr <= rqpair->max_read_depth);
1160 4 : rqpair->current_read_depth += rdma_req->num_outstanding_data_wr;
1161 4 : assert(rqpair->current_send_depth + rdma_req->num_outstanding_data_wr <= rqpair->max_send_depth);
1162 4 : rqpair->current_send_depth += rdma_req->num_outstanding_data_wr;
1163 4 : return 0;
1164 : }
1165 :
1166 : static inline void
1167 0 : nvmf_rdma_request_reset_transfer_in(struct spdk_nvmf_rdma_request *rdma_req,
1168 : struct spdk_nvmf_rdma_transport *rtransport)
1169 : {
1170 : /* Put completed WRs back to pool and move transfer_wr pointer */
1171 0 : _nvmf_rdma_request_free_data(rdma_req, rdma_req->transfer_wr, rtransport->data_wr_pool);
1172 0 : rdma_req->transfer_wr = rdma_req->remaining_tranfer_in_wrs;
1173 0 : rdma_req->remaining_tranfer_in_wrs = NULL;
1174 0 : rdma_req->num_outstanding_data_wr = rdma_req->num_remaining_data_wr;
1175 0 : rdma_req->num_remaining_data_wr = 0;
1176 0 : }
1177 :
1178 : static inline int
1179 0 : request_prepare_transfer_in_part(struct spdk_nvmf_request *req, uint32_t num_reads_available)
1180 : {
1181 : struct spdk_nvmf_rdma_request *rdma_req;
1182 : struct ibv_send_wr *wr;
1183 : uint32_t i;
1184 :
1185 0 : rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req);
1186 :
1187 0 : assert(req->xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER);
1188 0 : assert(rdma_req != NULL);
1189 0 : assert(num_reads_available > 0);
1190 0 : assert(rdma_req->num_outstanding_data_wr > num_reads_available);
1191 0 : wr = rdma_req->transfer_wr;
1192 :
1193 0 : for (i = 0; i < num_reads_available - 1; i++) {
1194 0 : wr = wr->next;
1195 : }
1196 :
1197 0 : rdma_req->remaining_tranfer_in_wrs = wr->next;
1198 0 : rdma_req->num_remaining_data_wr = rdma_req->num_outstanding_data_wr - num_reads_available;
1199 0 : rdma_req->num_outstanding_data_wr = num_reads_available;
1200 : /* Break chain of WRs to send only part. Once this portion completes, we continue sending RDMA_READs */
1201 0 : wr->next = NULL;
1202 :
1203 0 : return 0;
1204 : }
1205 :
1206 : static int
1207 6 : request_transfer_out(struct spdk_nvmf_request *req, int *data_posted)
1208 : {
1209 6 : int num_outstanding_data_wr = 0;
1210 : struct spdk_nvmf_rdma_request *rdma_req;
1211 : struct spdk_nvmf_qpair *qpair;
1212 : struct spdk_nvmf_rdma_qpair *rqpair;
1213 : struct spdk_nvme_cpl *rsp;
1214 6 : struct ibv_send_wr *first = NULL;
1215 : struct spdk_nvmf_rdma_transport *rtransport;
1216 :
1217 6 : *data_posted = 0;
1218 6 : qpair = req->qpair;
1219 6 : rsp = &req->rsp->nvme_cpl;
1220 6 : rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req);
1221 6 : rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair);
1222 6 : rtransport = SPDK_CONTAINEROF(rqpair->qpair.transport,
1223 : struct spdk_nvmf_rdma_transport, transport);
1224 :
1225 : /* Advance our sq_head pointer */
1226 6 : if (qpair->sq_head == qpair->sq_head_max) {
1227 6 : qpair->sq_head = 0;
1228 : } else {
1229 0 : qpair->sq_head++;
1230 : }
1231 6 : rsp->sqhd = qpair->sq_head;
1232 :
1233 : /* queue the capsule for the recv buffer */
1234 6 : assert(rdma_req->recv != NULL);
1235 :
1236 6 : nvmf_rdma_qpair_queue_recv_wrs(rqpair, &rdma_req->recv->wr);
1237 :
1238 6 : rdma_req->recv = NULL;
1239 6 : assert(rqpair->current_recv_depth > 0);
1240 6 : rqpair->current_recv_depth--;
1241 :
1242 : /* Build the response which consists of optional
1243 : * RDMA WRITEs to transfer data, plus an RDMA SEND
1244 : * containing the response.
1245 : */
1246 6 : first = &rdma_req->rsp.wr;
1247 :
1248 6 : if (rsp->status.sc != SPDK_NVME_SC_SUCCESS) {
1249 : /* On failure, data was not read from the controller. So clear the
1250 : * number of outstanding data WRs to zero.
1251 : */
1252 1 : rdma_req->num_outstanding_data_wr = 0;
1253 5 : } else if (req->xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) {
1254 1 : first = rdma_req->transfer_wr;
1255 1 : *data_posted = 1;
1256 1 : num_outstanding_data_wr = rdma_req->num_outstanding_data_wr;
1257 : }
1258 6 : if (spdk_rdma_qp_queue_send_wrs(rqpair->rdma_qp, first)) {
1259 6 : STAILQ_INSERT_TAIL(&rqpair->poller->qpairs_pending_send, rqpair, send_link);
1260 : }
1261 6 : if (rtransport->rdma_opts.no_wr_batching) {
1262 0 : _poller_submit_sends(rtransport, rqpair->poller);
1263 : }
1264 :
1265 : /* +1 for the rsp wr */
1266 6 : assert(rqpair->current_send_depth + num_outstanding_data_wr + 1 <= rqpair->max_send_depth);
1267 6 : rqpair->current_send_depth += num_outstanding_data_wr + 1;
1268 :
1269 6 : return 0;
1270 : }
1271 :
1272 : static int
1273 0 : nvmf_rdma_event_accept(struct rdma_cm_id *id, struct spdk_nvmf_rdma_qpair *rqpair)
1274 : {
1275 0 : struct spdk_nvmf_rdma_accept_private_data accept_data;
1276 0 : struct rdma_conn_param ctrlr_event_data = {};
1277 : int rc;
1278 :
1279 0 : accept_data.recfmt = 0;
1280 0 : accept_data.crqsize = rqpair->max_queue_depth;
1281 :
1282 0 : ctrlr_event_data.private_data = &accept_data;
1283 0 : ctrlr_event_data.private_data_len = sizeof(accept_data);
1284 0 : if (id->ps == RDMA_PS_TCP) {
1285 0 : ctrlr_event_data.responder_resources = 0; /* We accept 0 reads from the host */
1286 0 : ctrlr_event_data.initiator_depth = rqpair->max_read_depth;
1287 : }
1288 :
1289 : /* Configure infinite retries for the initiator side qpair.
1290 : * We need to pass this value to the initiator to prevent the
1291 : * initiator side NIC from completing SEND requests back to the
1292 : * initiator with status rnr_retry_count_exceeded. */
1293 0 : ctrlr_event_data.rnr_retry_count = 0x7;
1294 :
1295 : /* When qpair is created without use of rdma cm API, an additional
1296 : * information must be provided to initiator in the connection response:
1297 : * whether qpair is using SRQ and its qp_num
1298 : * Fields below are ignored by rdma cm if qpair has been
1299 : * created using rdma cm API. */
1300 0 : ctrlr_event_data.srq = rqpair->srq ? 1 : 0;
1301 0 : ctrlr_event_data.qp_num = rqpair->qp_num;
1302 :
1303 0 : rc = spdk_rdma_qp_accept(rqpair->rdma_qp, &ctrlr_event_data);
1304 0 : if (rc) {
1305 0 : SPDK_ERRLOG("Error %d on spdk_rdma_qp_accept\n", errno);
1306 : } else {
1307 0 : SPDK_DEBUGLOG(rdma, "Sent back the accept\n");
1308 : }
1309 :
1310 0 : return rc;
1311 : }
1312 :
1313 : static void
1314 0 : nvmf_rdma_event_reject(struct rdma_cm_id *id, enum spdk_nvmf_rdma_transport_error error)
1315 : {
1316 0 : struct spdk_nvmf_rdma_reject_private_data rej_data;
1317 :
1318 0 : rej_data.recfmt = 0;
1319 0 : rej_data.sts = error;
1320 :
1321 0 : rdma_reject(id, &rej_data, sizeof(rej_data));
1322 0 : }
1323 :
1324 : static int
1325 0 : nvmf_rdma_connect(struct spdk_nvmf_transport *transport, struct rdma_cm_event *event)
1326 : {
1327 : struct spdk_nvmf_rdma_transport *rtransport;
1328 0 : struct spdk_nvmf_rdma_qpair *rqpair = NULL;
1329 : struct spdk_nvmf_rdma_port *port;
1330 0 : struct rdma_conn_param *rdma_param = NULL;
1331 0 : const struct spdk_nvmf_rdma_request_private_data *private_data = NULL;
1332 : uint16_t max_queue_depth;
1333 : uint16_t max_read_depth;
1334 :
1335 0 : rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport);
1336 :
1337 0 : assert(event->id != NULL); /* Impossible. Can't even reject the connection. */
1338 0 : assert(event->id->verbs != NULL); /* Impossible. No way to handle this. */
1339 :
1340 0 : rdma_param = &event->param.conn;
1341 0 : if (rdma_param->private_data == NULL ||
1342 0 : rdma_param->private_data_len < sizeof(struct spdk_nvmf_rdma_request_private_data)) {
1343 0 : SPDK_ERRLOG("connect request: no private data provided\n");
1344 0 : nvmf_rdma_event_reject(event->id, SPDK_NVMF_RDMA_ERROR_INVALID_PRIVATE_DATA_LENGTH);
1345 0 : return -1;
1346 : }
1347 :
1348 0 : private_data = rdma_param->private_data;
1349 0 : if (private_data->recfmt != 0) {
1350 0 : SPDK_ERRLOG("Received RDMA private data with RECFMT != 0\n");
1351 0 : nvmf_rdma_event_reject(event->id, SPDK_NVMF_RDMA_ERROR_INVALID_RECFMT);
1352 0 : return -1;
1353 : }
1354 :
1355 0 : SPDK_DEBUGLOG(rdma, "Connect Recv on fabric intf name %s, dev_name %s\n",
1356 : event->id->verbs->device->name, event->id->verbs->device->dev_name);
1357 :
1358 0 : port = event->listen_id->context;
1359 0 : SPDK_DEBUGLOG(rdma, "Listen Id was %p with verbs %p. ListenAddr: %p\n",
1360 : event->listen_id, event->listen_id->verbs, port);
1361 :
1362 : /* Figure out the supported queue depth. This is a multi-step process
1363 : * that takes into account hardware maximums, host provided values,
1364 : * and our target's internal memory limits */
1365 :
1366 0 : SPDK_DEBUGLOG(rdma, "Calculating Queue Depth\n");
1367 :
1368 : /* Start with the maximum queue depth allowed by the target */
1369 0 : max_queue_depth = rtransport->transport.opts.max_queue_depth;
1370 0 : max_read_depth = rtransport->transport.opts.max_queue_depth;
1371 0 : SPDK_DEBUGLOG(rdma, "Target Max Queue Depth: %d\n",
1372 : rtransport->transport.opts.max_queue_depth);
1373 :
1374 : /* Next check the local NIC's hardware limitations */
1375 0 : SPDK_DEBUGLOG(rdma,
1376 : "Local NIC Max Send/Recv Queue Depth: %d Max Read/Write Queue Depth: %d\n",
1377 : port->device->attr.max_qp_wr, port->device->attr.max_qp_rd_atom);
1378 0 : max_queue_depth = spdk_min(max_queue_depth, port->device->attr.max_qp_wr);
1379 0 : max_read_depth = spdk_min(max_read_depth, port->device->attr.max_qp_init_rd_atom);
1380 :
1381 : /* Next check the remote NIC's hardware limitations */
1382 0 : SPDK_DEBUGLOG(rdma,
1383 : "Host (Initiator) NIC Max Incoming RDMA R/W operations: %d Max Outgoing RDMA R/W operations: %d\n",
1384 : rdma_param->initiator_depth, rdma_param->responder_resources);
1385 : /* from man3 rdma_get_cm_event
1386 : * responder_resources - Specifies the number of responder resources that is requested by the recipient.
1387 : * The responder_resources field must match the initiator depth specified by the remote node when running
1388 : * the rdma_connect and rdma_accept functions. */
1389 0 : if (rdma_param->responder_resources != 0) {
1390 0 : if (private_data->qid) {
1391 0 : SPDK_DEBUGLOG(rdma, "Host (Initiator) is not allowed to use RDMA operations,"
1392 : " responder_resources must be 0 but set to %u\n",
1393 : rdma_param->responder_resources);
1394 : } else {
1395 0 : SPDK_WARNLOG("Host (Initiator) is not allowed to use RDMA operations,"
1396 : " responder_resources must be 0 but set to %u\n",
1397 : rdma_param->responder_resources);
1398 : }
1399 : }
1400 : /* from man3 rdma_get_cm_event
1401 : * initiator_depth - Specifies the maximum number of outstanding RDMA read operations that the recipient holds.
1402 : * The initiator_depth field must match the responder resources specified by the remote node when running
1403 : * the rdma_connect and rdma_accept functions. */
1404 0 : if (rdma_param->initiator_depth == 0) {
1405 0 : SPDK_ERRLOG("Host (Initiator) doesn't support RDMA_READ or atomic operations\n");
1406 0 : nvmf_rdma_event_reject(event->id, SPDK_NVMF_RDMA_ERROR_INVALID_IRD);
1407 0 : return -1;
1408 : }
1409 0 : max_read_depth = spdk_min(max_read_depth, rdma_param->initiator_depth);
1410 :
1411 0 : SPDK_DEBUGLOG(rdma, "Host Receive Queue Size: %d\n", private_data->hrqsize);
1412 0 : SPDK_DEBUGLOG(rdma, "Host Send Queue Size: %d\n", private_data->hsqsize);
1413 0 : max_queue_depth = spdk_min(max_queue_depth, private_data->hrqsize);
1414 0 : max_queue_depth = spdk_min(max_queue_depth, private_data->hsqsize + 1);
1415 :
1416 0 : SPDK_DEBUGLOG(rdma, "Final Negotiated Queue Depth: %d R/W Depth: %d\n",
1417 : max_queue_depth, max_read_depth);
1418 :
1419 0 : rqpair = calloc(1, sizeof(struct spdk_nvmf_rdma_qpair));
1420 0 : if (rqpair == NULL) {
1421 0 : SPDK_ERRLOG("Could not allocate new connection.\n");
1422 0 : nvmf_rdma_event_reject(event->id, SPDK_NVMF_RDMA_ERROR_NO_RESOURCES);
1423 0 : return -1;
1424 : }
1425 :
1426 0 : rqpair->device = port->device;
1427 0 : rqpair->max_queue_depth = max_queue_depth;
1428 0 : rqpair->max_read_depth = max_read_depth;
1429 0 : rqpair->cm_id = event->id;
1430 0 : rqpair->listen_id = event->listen_id;
1431 0 : rqpair->qpair.transport = transport;
1432 0 : STAILQ_INIT(&rqpair->ibv_events);
1433 : /* use qid from the private data to determine the qpair type
1434 : qid will be set to the appropriate value when the controller is created */
1435 0 : rqpair->qpair.qid = private_data->qid;
1436 :
1437 0 : event->id->context = &rqpair->qpair;
1438 :
1439 0 : spdk_nvmf_tgt_new_qpair(transport->tgt, &rqpair->qpair);
1440 :
1441 0 : return 0;
1442 : }
1443 :
1444 : static inline void
1445 28 : nvmf_rdma_setup_wr(struct ibv_send_wr *wr, struct ibv_send_wr *next,
1446 : enum spdk_nvme_data_transfer xfer)
1447 : {
1448 28 : if (xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) {
1449 24 : wr->opcode = IBV_WR_RDMA_WRITE;
1450 24 : wr->send_flags = 0;
1451 24 : wr->next = next;
1452 4 : } else if (xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER) {
1453 4 : wr->opcode = IBV_WR_RDMA_READ;
1454 4 : wr->send_flags = IBV_SEND_SIGNALED;
1455 4 : wr->next = NULL;
1456 : } else {
1457 0 : assert(0);
1458 : }
1459 28 : }
1460 :
1461 : static int
1462 6 : nvmf_request_alloc_wrs(struct spdk_nvmf_rdma_transport *rtransport,
1463 : struct spdk_nvmf_rdma_request *rdma_req,
1464 : uint32_t num_sgl_descriptors)
1465 : {
1466 6 : struct spdk_nvmf_rdma_request_data *work_requests[SPDK_NVMF_MAX_SGL_ENTRIES];
1467 : struct spdk_nvmf_rdma_request_data *current_data_wr;
1468 : uint32_t i;
1469 :
1470 6 : if (num_sgl_descriptors > SPDK_NVMF_MAX_SGL_ENTRIES) {
1471 0 : SPDK_ERRLOG("Requested too much entries (%u), the limit is %u\n",
1472 : num_sgl_descriptors, SPDK_NVMF_MAX_SGL_ENTRIES);
1473 0 : return -EINVAL;
1474 : }
1475 :
1476 6 : if (spdk_mempool_get_bulk(rtransport->data_wr_pool, (void **)work_requests, num_sgl_descriptors)) {
1477 0 : return -ENOMEM;
1478 : }
1479 :
1480 6 : current_data_wr = &rdma_req->data;
1481 :
1482 12 : for (i = 0; i < num_sgl_descriptors; i++) {
1483 6 : nvmf_rdma_setup_wr(¤t_data_wr->wr, &work_requests[i]->wr, rdma_req->req.xfer);
1484 6 : current_data_wr->wr.next = &work_requests[i]->wr;
1485 6 : current_data_wr = work_requests[i];
1486 6 : current_data_wr->wr.sg_list = current_data_wr->sgl;
1487 6 : current_data_wr->wr.wr_id = rdma_req->data.wr.wr_id;
1488 : }
1489 :
1490 6 : nvmf_rdma_setup_wr(¤t_data_wr->wr, &rdma_req->rsp.wr, rdma_req->req.xfer);
1491 :
1492 6 : return 0;
1493 : }
1494 :
1495 : static inline void
1496 16 : nvmf_rdma_setup_request(struct spdk_nvmf_rdma_request *rdma_req)
1497 : {
1498 16 : struct ibv_send_wr *wr = &rdma_req->data.wr;
1499 16 : struct spdk_nvme_sgl_descriptor *sgl = &rdma_req->req.cmd->nvme_cmd.dptr.sgl1;
1500 :
1501 16 : wr->wr.rdma.rkey = sgl->keyed.key;
1502 16 : wr->wr.rdma.remote_addr = sgl->address;
1503 16 : nvmf_rdma_setup_wr(wr, &rdma_req->rsp.wr, rdma_req->req.xfer);
1504 16 : }
1505 :
1506 : static inline void
1507 1 : nvmf_rdma_update_remote_addr(struct spdk_nvmf_rdma_request *rdma_req, uint32_t num_wrs)
1508 : {
1509 1 : struct ibv_send_wr *wr = &rdma_req->data.wr;
1510 1 : struct spdk_nvme_sgl_descriptor *sgl = &rdma_req->req.cmd->nvme_cmd.dptr.sgl1;
1511 : uint32_t i;
1512 : int j;
1513 1 : uint64_t remote_addr_offset = 0;
1514 :
1515 3 : for (i = 0; i < num_wrs; ++i) {
1516 2 : wr->wr.rdma.rkey = sgl->keyed.key;
1517 2 : wr->wr.rdma.remote_addr = sgl->address + remote_addr_offset;
1518 19 : for (j = 0; j < wr->num_sge; ++j) {
1519 17 : remote_addr_offset += wr->sg_list[j].length;
1520 : }
1521 2 : wr = wr->next;
1522 : }
1523 1 : }
1524 :
1525 : static int
1526 15 : nvmf_rdma_fill_wr_sgl(struct spdk_nvmf_rdma_poll_group *rgroup,
1527 : struct spdk_nvmf_rdma_device *device,
1528 : struct spdk_nvmf_rdma_request *rdma_req,
1529 : struct ibv_send_wr *wr,
1530 : uint32_t total_length)
1531 : {
1532 15 : struct spdk_rdma_memory_translation mem_translation;
1533 : struct ibv_sge *sg_ele;
1534 : struct iovec *iov;
1535 : uint32_t lkey, remaining;
1536 : int rc;
1537 :
1538 15 : wr->num_sge = 0;
1539 :
1540 74 : while (total_length && wr->num_sge < SPDK_NVMF_MAX_SGL_ENTRIES) {
1541 59 : iov = &rdma_req->req.iov[rdma_req->iovpos];
1542 59 : rc = spdk_rdma_get_translation(device->map, iov->iov_base, iov->iov_len, &mem_translation);
1543 59 : if (spdk_unlikely(rc)) {
1544 0 : return rc;
1545 : }
1546 :
1547 59 : lkey = spdk_rdma_memory_translation_get_lkey(&mem_translation);
1548 59 : sg_ele = &wr->sg_list[wr->num_sge];
1549 59 : remaining = spdk_min((uint32_t)iov->iov_len - rdma_req->offset, total_length);
1550 :
1551 59 : sg_ele->lkey = lkey;
1552 59 : sg_ele->addr = (uintptr_t)iov->iov_base + rdma_req->offset;
1553 59 : sg_ele->length = remaining;
1554 59 : SPDK_DEBUGLOG(rdma, "sge[%d] %p addr 0x%"PRIx64", len %u\n", wr->num_sge, sg_ele, sg_ele->addr,
1555 : sg_ele->length);
1556 59 : rdma_req->offset += sg_ele->length;
1557 59 : total_length -= sg_ele->length;
1558 59 : wr->num_sge++;
1559 :
1560 59 : if (rdma_req->offset == iov->iov_len) {
1561 57 : rdma_req->offset = 0;
1562 57 : rdma_req->iovpos++;
1563 : }
1564 : }
1565 :
1566 15 : if (total_length) {
1567 0 : SPDK_ERRLOG("Not enough SG entries to hold data buffer\n");
1568 0 : return -EINVAL;
1569 : }
1570 :
1571 15 : return 0;
1572 : }
1573 :
1574 : static int
1575 10 : nvmf_rdma_fill_wr_sgl_with_dif(struct spdk_nvmf_rdma_poll_group *rgroup,
1576 : struct spdk_nvmf_rdma_device *device,
1577 : struct spdk_nvmf_rdma_request *rdma_req,
1578 : struct ibv_send_wr *wr,
1579 : uint32_t total_length,
1580 : uint32_t num_extra_wrs)
1581 : {
1582 10 : struct spdk_rdma_memory_translation mem_translation;
1583 10 : struct spdk_dif_ctx *dif_ctx = &rdma_req->req.dif.dif_ctx;
1584 : struct ibv_sge *sg_ele;
1585 : struct iovec *iov;
1586 : struct iovec *rdma_iov;
1587 : uint32_t lkey, remaining;
1588 : uint32_t remaining_data_block, data_block_size, md_size;
1589 : uint32_t sge_len;
1590 : int rc;
1591 :
1592 10 : data_block_size = dif_ctx->block_size - dif_ctx->md_size;
1593 :
1594 10 : if (spdk_likely(!rdma_req->req.stripped_data)) {
1595 5 : rdma_iov = rdma_req->req.iov;
1596 5 : remaining_data_block = data_block_size;
1597 5 : md_size = dif_ctx->md_size;
1598 : } else {
1599 5 : rdma_iov = rdma_req->req.stripped_data->iov;
1600 5 : total_length = total_length / dif_ctx->block_size * data_block_size;
1601 5 : remaining_data_block = total_length;
1602 5 : md_size = 0;
1603 : }
1604 :
1605 10 : wr->num_sge = 0;
1606 :
1607 25 : while (total_length && (num_extra_wrs || wr->num_sge < SPDK_NVMF_MAX_SGL_ENTRIES)) {
1608 15 : iov = rdma_iov + rdma_req->iovpos;
1609 15 : rc = spdk_rdma_get_translation(device->map, iov->iov_base, iov->iov_len, &mem_translation);
1610 15 : if (spdk_unlikely(rc)) {
1611 0 : return rc;
1612 : }
1613 :
1614 15 : lkey = spdk_rdma_memory_translation_get_lkey(&mem_translation);
1615 15 : sg_ele = &wr->sg_list[wr->num_sge];
1616 15 : remaining = spdk_min((uint32_t)iov->iov_len - rdma_req->offset, total_length);
1617 :
1618 53 : while (remaining) {
1619 38 : if (wr->num_sge >= SPDK_NVMF_MAX_SGL_ENTRIES) {
1620 1 : if (num_extra_wrs > 0 && wr->next) {
1621 1 : wr = wr->next;
1622 1 : wr->num_sge = 0;
1623 1 : sg_ele = &wr->sg_list[wr->num_sge];
1624 1 : num_extra_wrs--;
1625 : } else {
1626 : break;
1627 : }
1628 : }
1629 38 : sg_ele->lkey = lkey;
1630 38 : sg_ele->addr = (uintptr_t)((char *)iov->iov_base + rdma_req->offset);
1631 38 : sge_len = spdk_min(remaining, remaining_data_block);
1632 38 : sg_ele->length = sge_len;
1633 38 : SPDK_DEBUGLOG(rdma, "sge[%d] %p addr 0x%"PRIx64", len %u\n", wr->num_sge, sg_ele,
1634 : sg_ele->addr, sg_ele->length);
1635 38 : remaining -= sge_len;
1636 38 : remaining_data_block -= sge_len;
1637 38 : rdma_req->offset += sge_len;
1638 38 : total_length -= sge_len;
1639 :
1640 38 : sg_ele++;
1641 38 : wr->num_sge++;
1642 :
1643 38 : if (remaining_data_block == 0) {
1644 : /* skip metadata */
1645 34 : rdma_req->offset += md_size;
1646 34 : total_length -= md_size;
1647 : /* Metadata that do not fit this IO buffer will be included in the next IO buffer */
1648 34 : remaining -= spdk_min(remaining, md_size);
1649 34 : remaining_data_block = data_block_size;
1650 : }
1651 :
1652 38 : if (remaining == 0) {
1653 : /* By subtracting the size of the last IOV from the offset, we ensure that we skip
1654 : the remaining metadata bits at the beginning of the next buffer */
1655 15 : rdma_req->offset -= spdk_min(iov->iov_len, rdma_req->offset);
1656 15 : rdma_req->iovpos++;
1657 : }
1658 : }
1659 : }
1660 :
1661 10 : if (total_length) {
1662 0 : SPDK_ERRLOG("Not enough SG entries to hold data buffer\n");
1663 0 : return -EINVAL;
1664 : }
1665 :
1666 10 : return 0;
1667 : }
1668 :
1669 : static inline uint32_t
1670 8 : nvmf_rdma_calc_num_wrs(uint32_t length, uint32_t io_unit_size, uint32_t block_size)
1671 : {
1672 : /* estimate the number of SG entries and WRs needed to process the request */
1673 8 : uint32_t num_sge = 0;
1674 : uint32_t i;
1675 8 : uint32_t num_buffers = SPDK_CEIL_DIV(length, io_unit_size);
1676 :
1677 23 : for (i = 0; i < num_buffers && length > 0; i++) {
1678 15 : uint32_t buffer_len = spdk_min(length, io_unit_size);
1679 15 : uint32_t num_sge_in_block = SPDK_CEIL_DIV(buffer_len, block_size);
1680 :
1681 15 : if (num_sge_in_block * block_size > buffer_len) {
1682 11 : ++num_sge_in_block;
1683 : }
1684 15 : num_sge += num_sge_in_block;
1685 15 : length -= buffer_len;
1686 : }
1687 8 : return SPDK_CEIL_DIV(num_sge, SPDK_NVMF_MAX_SGL_ENTRIES);
1688 : }
1689 :
1690 : static int
1691 16 : nvmf_rdma_request_fill_iovs(struct spdk_nvmf_rdma_transport *rtransport,
1692 : struct spdk_nvmf_rdma_device *device,
1693 : struct spdk_nvmf_rdma_request *rdma_req)
1694 : {
1695 : struct spdk_nvmf_rdma_qpair *rqpair;
1696 : struct spdk_nvmf_rdma_poll_group *rgroup;
1697 16 : struct spdk_nvmf_request *req = &rdma_req->req;
1698 16 : struct ibv_send_wr *wr = &rdma_req->data.wr;
1699 : int rc;
1700 16 : uint32_t num_wrs = 1;
1701 : uint32_t length;
1702 :
1703 16 : rqpair = SPDK_CONTAINEROF(req->qpair, struct spdk_nvmf_rdma_qpair, qpair);
1704 16 : rgroup = rqpair->poller->group;
1705 :
1706 : /* rdma wr specifics */
1707 16 : nvmf_rdma_setup_request(rdma_req);
1708 :
1709 16 : length = req->length;
1710 16 : if (spdk_unlikely(req->dif_enabled)) {
1711 8 : req->dif.orig_length = length;
1712 8 : length = spdk_dif_get_length_with_md(length, &req->dif.dif_ctx);
1713 8 : req->dif.elba_length = length;
1714 : }
1715 :
1716 16 : rc = spdk_nvmf_request_get_buffers(req, &rgroup->group, &rtransport->transport,
1717 : length);
1718 16 : if (rc != 0) {
1719 1 : return rc;
1720 : }
1721 :
1722 15 : assert(req->iovcnt <= rqpair->max_send_sge);
1723 :
1724 : /* When dif_insert_or_strip is true and the I/O data length is greater than one block,
1725 : * the stripped_buffers are got for DIF stripping. */
1726 15 : if (spdk_unlikely(req->dif_enabled && (req->xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST)
1727 : && (req->dif.elba_length > req->dif.dif_ctx.block_size))) {
1728 7 : rc = nvmf_request_get_stripped_buffers(req, &rgroup->group,
1729 : &rtransport->transport, req->dif.orig_length);
1730 7 : if (rc != 0) {
1731 4 : SPDK_INFOLOG(rdma, "Get stripped buffers fail %d, fallback to req.iov.\n", rc);
1732 : }
1733 : }
1734 :
1735 15 : rdma_req->iovpos = 0;
1736 :
1737 15 : if (spdk_unlikely(req->dif_enabled)) {
1738 8 : num_wrs = nvmf_rdma_calc_num_wrs(length, rtransport->transport.opts.io_unit_size,
1739 : req->dif.dif_ctx.block_size);
1740 8 : if (num_wrs > 1) {
1741 1 : rc = nvmf_request_alloc_wrs(rtransport, rdma_req, num_wrs - 1);
1742 1 : if (rc != 0) {
1743 0 : goto err_exit;
1744 : }
1745 : }
1746 :
1747 8 : rc = nvmf_rdma_fill_wr_sgl_with_dif(rgroup, device, rdma_req, wr, length, num_wrs - 1);
1748 8 : if (spdk_unlikely(rc != 0)) {
1749 0 : goto err_exit;
1750 : }
1751 :
1752 8 : if (num_wrs > 1) {
1753 1 : nvmf_rdma_update_remote_addr(rdma_req, num_wrs);
1754 : }
1755 : } else {
1756 7 : rc = nvmf_rdma_fill_wr_sgl(rgroup, device, rdma_req, wr, length);
1757 7 : if (spdk_unlikely(rc != 0)) {
1758 0 : goto err_exit;
1759 : }
1760 : }
1761 :
1762 : /* set the number of outstanding data WRs for this request. */
1763 15 : rdma_req->num_outstanding_data_wr = num_wrs;
1764 :
1765 15 : return rc;
1766 :
1767 0 : err_exit:
1768 0 : spdk_nvmf_request_free_buffers(req, &rgroup->group, &rtransport->transport);
1769 0 : nvmf_rdma_request_free_data(rdma_req, rtransport);
1770 0 : req->iovcnt = 0;
1771 0 : return rc;
1772 : }
1773 :
1774 : static int
1775 5 : nvmf_rdma_request_fill_iovs_multi_sgl(struct spdk_nvmf_rdma_transport *rtransport,
1776 : struct spdk_nvmf_rdma_device *device,
1777 : struct spdk_nvmf_rdma_request *rdma_req)
1778 : {
1779 : struct spdk_nvmf_rdma_qpair *rqpair;
1780 : struct spdk_nvmf_rdma_poll_group *rgroup;
1781 : struct ibv_send_wr *current_wr;
1782 5 : struct spdk_nvmf_request *req = &rdma_req->req;
1783 : struct spdk_nvme_sgl_descriptor *inline_segment, *desc;
1784 : uint32_t num_sgl_descriptors;
1785 5 : uint32_t lengths[SPDK_NVMF_MAX_SGL_ENTRIES], total_length = 0;
1786 : uint32_t i;
1787 : int rc;
1788 :
1789 5 : rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair);
1790 5 : rgroup = rqpair->poller->group;
1791 :
1792 5 : inline_segment = &req->cmd->nvme_cmd.dptr.sgl1;
1793 5 : assert(inline_segment->generic.type == SPDK_NVME_SGL_TYPE_LAST_SEGMENT);
1794 5 : assert(inline_segment->unkeyed.subtype == SPDK_NVME_SGL_SUBTYPE_OFFSET);
1795 :
1796 5 : num_sgl_descriptors = inline_segment->unkeyed.length / sizeof(struct spdk_nvme_sgl_descriptor);
1797 5 : assert(num_sgl_descriptors <= SPDK_NVMF_MAX_SGL_ENTRIES);
1798 :
1799 5 : desc = (struct spdk_nvme_sgl_descriptor *)rdma_req->recv->buf + inline_segment->address;
1800 15 : for (i = 0; i < num_sgl_descriptors; i++) {
1801 10 : if (spdk_likely(!req->dif_enabled)) {
1802 8 : lengths[i] = desc->keyed.length;
1803 : } else {
1804 2 : req->dif.orig_length += desc->keyed.length;
1805 2 : lengths[i] = spdk_dif_get_length_with_md(desc->keyed.length, &req->dif.dif_ctx);
1806 2 : req->dif.elba_length += lengths[i];
1807 : }
1808 10 : total_length += lengths[i];
1809 10 : desc++;
1810 : }
1811 :
1812 5 : if (total_length > rtransport->transport.opts.max_io_size) {
1813 0 : SPDK_ERRLOG("Multi SGL length 0x%x exceeds max io size 0x%x\n",
1814 : total_length, rtransport->transport.opts.max_io_size);
1815 0 : req->rsp->nvme_cpl.status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID;
1816 0 : return -EINVAL;
1817 : }
1818 :
1819 5 : if (nvmf_request_alloc_wrs(rtransport, rdma_req, num_sgl_descriptors - 1) != 0) {
1820 0 : return -ENOMEM;
1821 : }
1822 :
1823 5 : rc = spdk_nvmf_request_get_buffers(req, &rgroup->group, &rtransport->transport, total_length);
1824 5 : if (rc != 0) {
1825 0 : nvmf_rdma_request_free_data(rdma_req, rtransport);
1826 0 : return rc;
1827 : }
1828 :
1829 : /* When dif_insert_or_strip is true and the I/O data length is greater than one block,
1830 : * the stripped_buffers are got for DIF stripping. */
1831 5 : if (spdk_unlikely(req->dif_enabled && (req->xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST)
1832 : && (req->dif.elba_length > req->dif.dif_ctx.block_size))) {
1833 1 : rc = nvmf_request_get_stripped_buffers(req, &rgroup->group,
1834 : &rtransport->transport, req->dif.orig_length);
1835 1 : if (rc != 0) {
1836 0 : SPDK_INFOLOG(rdma, "Get stripped buffers fail %d, fallback to req.iov.\n", rc);
1837 : }
1838 : }
1839 :
1840 : /* The first WR must always be the embedded data WR. This is how we unwind them later. */
1841 5 : current_wr = &rdma_req->data.wr;
1842 5 : assert(current_wr != NULL);
1843 :
1844 5 : req->length = 0;
1845 5 : rdma_req->iovpos = 0;
1846 5 : desc = (struct spdk_nvme_sgl_descriptor *)rdma_req->recv->buf + inline_segment->address;
1847 15 : for (i = 0; i < num_sgl_descriptors; i++) {
1848 : /* The descriptors must be keyed data block descriptors with an address, not an offset. */
1849 10 : if (spdk_unlikely(desc->generic.type != SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK ||
1850 : desc->keyed.subtype != SPDK_NVME_SGL_SUBTYPE_ADDRESS)) {
1851 0 : rc = -EINVAL;
1852 0 : goto err_exit;
1853 : }
1854 :
1855 10 : if (spdk_likely(!req->dif_enabled)) {
1856 8 : rc = nvmf_rdma_fill_wr_sgl(rgroup, device, rdma_req, current_wr, lengths[i]);
1857 : } else {
1858 2 : rc = nvmf_rdma_fill_wr_sgl_with_dif(rgroup, device, rdma_req, current_wr,
1859 : lengths[i], 0);
1860 : }
1861 10 : if (rc != 0) {
1862 0 : rc = -ENOMEM;
1863 0 : goto err_exit;
1864 : }
1865 :
1866 10 : req->length += desc->keyed.length;
1867 10 : current_wr->wr.rdma.rkey = desc->keyed.key;
1868 10 : current_wr->wr.rdma.remote_addr = desc->address;
1869 10 : current_wr = current_wr->next;
1870 10 : desc++;
1871 : }
1872 :
1873 : #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL
1874 : /* Go back to the last descriptor in the list. */
1875 5 : desc--;
1876 5 : if ((device->attr.device_cap_flags & IBV_DEVICE_MEM_MGT_EXTENSIONS) != 0) {
1877 0 : if (desc->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_INVALIDATE_KEY) {
1878 0 : rdma_req->rsp.wr.opcode = IBV_WR_SEND_WITH_INV;
1879 0 : rdma_req->rsp.wr.imm_data = desc->keyed.key;
1880 : }
1881 : }
1882 : #endif
1883 :
1884 5 : rdma_req->num_outstanding_data_wr = num_sgl_descriptors;
1885 :
1886 5 : return 0;
1887 :
1888 0 : err_exit:
1889 0 : spdk_nvmf_request_free_buffers(req, &rgroup->group, &rtransport->transport);
1890 0 : nvmf_rdma_request_free_data(rdma_req, rtransport);
1891 0 : return rc;
1892 : }
1893 :
1894 : static int
1895 25 : nvmf_rdma_request_parse_sgl(struct spdk_nvmf_rdma_transport *rtransport,
1896 : struct spdk_nvmf_rdma_device *device,
1897 : struct spdk_nvmf_rdma_request *rdma_req)
1898 : {
1899 25 : struct spdk_nvmf_request *req = &rdma_req->req;
1900 : struct spdk_nvme_cpl *rsp;
1901 : struct spdk_nvme_sgl_descriptor *sgl;
1902 : int rc;
1903 : uint32_t length;
1904 :
1905 25 : rsp = &req->rsp->nvme_cpl;
1906 25 : sgl = &req->cmd->nvme_cmd.dptr.sgl1;
1907 :
1908 25 : if (sgl->generic.type == SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK &&
1909 17 : (sgl->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_ADDRESS ||
1910 0 : sgl->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_INVALIDATE_KEY)) {
1911 :
1912 17 : length = sgl->keyed.length;
1913 17 : if (length > rtransport->transport.opts.max_io_size) {
1914 1 : SPDK_ERRLOG("SGL length 0x%x exceeds max io size 0x%x\n",
1915 : length, rtransport->transport.opts.max_io_size);
1916 1 : rsp->status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID;
1917 1 : return -1;
1918 : }
1919 : #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL
1920 16 : if ((device->attr.device_cap_flags & IBV_DEVICE_MEM_MGT_EXTENSIONS) != 0) {
1921 0 : if (sgl->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_INVALIDATE_KEY) {
1922 0 : rdma_req->rsp.wr.opcode = IBV_WR_SEND_WITH_INV;
1923 0 : rdma_req->rsp.wr.imm_data = sgl->keyed.key;
1924 : }
1925 : }
1926 : #endif
1927 :
1928 : /* fill request length and populate iovs */
1929 16 : req->length = length;
1930 :
1931 16 : rc = nvmf_rdma_request_fill_iovs(rtransport, device, rdma_req);
1932 16 : if (spdk_unlikely(rc < 0)) {
1933 1 : if (rc == -EINVAL) {
1934 0 : SPDK_ERRLOG("SGL length exceeds the max I/O size\n");
1935 0 : rsp->status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID;
1936 0 : return -1;
1937 : }
1938 : /* No available buffers. Queue this request up. */
1939 1 : SPDK_DEBUGLOG(rdma, "No available large data buffers. Queueing request %p\n", rdma_req);
1940 1 : return 0;
1941 : }
1942 :
1943 15 : SPDK_DEBUGLOG(rdma, "Request %p took %d buffer/s from central pool\n", rdma_req,
1944 : req->iovcnt);
1945 :
1946 15 : return 0;
1947 8 : } else if (sgl->generic.type == SPDK_NVME_SGL_TYPE_DATA_BLOCK &&
1948 3 : sgl->unkeyed.subtype == SPDK_NVME_SGL_SUBTYPE_OFFSET) {
1949 3 : uint64_t offset = sgl->address;
1950 3 : uint32_t max_len = rtransport->transport.opts.in_capsule_data_size;
1951 :
1952 3 : SPDK_DEBUGLOG(nvmf, "In-capsule data: offset 0x%" PRIx64 ", length 0x%x\n",
1953 : offset, sgl->unkeyed.length);
1954 :
1955 3 : if (offset > max_len) {
1956 0 : SPDK_ERRLOG("In-capsule offset 0x%" PRIx64 " exceeds capsule length 0x%x\n",
1957 : offset, max_len);
1958 0 : rsp->status.sc = SPDK_NVME_SC_INVALID_SGL_OFFSET;
1959 0 : return -1;
1960 : }
1961 3 : max_len -= (uint32_t)offset;
1962 :
1963 3 : if (sgl->unkeyed.length > max_len) {
1964 2 : SPDK_ERRLOG("In-capsule data length 0x%x exceeds capsule length 0x%x\n",
1965 : sgl->unkeyed.length, max_len);
1966 2 : rsp->status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID;
1967 2 : return -1;
1968 : }
1969 :
1970 1 : rdma_req->num_outstanding_data_wr = 0;
1971 1 : req->data_from_pool = false;
1972 1 : req->length = sgl->unkeyed.length;
1973 :
1974 1 : req->iov[0].iov_base = rdma_req->recv->buf + offset;
1975 1 : req->iov[0].iov_len = req->length;
1976 1 : req->iovcnt = 1;
1977 :
1978 1 : return 0;
1979 5 : } else if (sgl->generic.type == SPDK_NVME_SGL_TYPE_LAST_SEGMENT &&
1980 5 : sgl->unkeyed.subtype == SPDK_NVME_SGL_SUBTYPE_OFFSET) {
1981 :
1982 5 : rc = nvmf_rdma_request_fill_iovs_multi_sgl(rtransport, device, rdma_req);
1983 5 : if (rc == -ENOMEM) {
1984 0 : SPDK_DEBUGLOG(rdma, "No available large data buffers. Queueing request %p\n", rdma_req);
1985 0 : return 0;
1986 5 : } else if (rc == -EINVAL) {
1987 0 : SPDK_ERRLOG("Multi SGL element request length exceeds the max I/O size\n");
1988 0 : rsp->status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID;
1989 0 : return -1;
1990 : }
1991 :
1992 5 : SPDK_DEBUGLOG(rdma, "Request %p took %d buffer/s from central pool\n", rdma_req,
1993 : req->iovcnt);
1994 :
1995 5 : return 0;
1996 : }
1997 :
1998 0 : SPDK_ERRLOG("Invalid NVMf I/O Command SGL: Type 0x%x, Subtype 0x%x\n",
1999 : sgl->generic.type, sgl->generic.subtype);
2000 0 : rsp->status.sc = SPDK_NVME_SC_SGL_DESCRIPTOR_TYPE_INVALID;
2001 0 : return -1;
2002 : }
2003 :
2004 : static void
2005 6 : _nvmf_rdma_request_free(struct spdk_nvmf_rdma_request *rdma_req,
2006 : struct spdk_nvmf_rdma_transport *rtransport)
2007 : {
2008 : struct spdk_nvmf_rdma_qpair *rqpair;
2009 : struct spdk_nvmf_rdma_poll_group *rgroup;
2010 :
2011 6 : rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair);
2012 6 : if (rdma_req->req.data_from_pool) {
2013 5 : rgroup = rqpair->poller->group;
2014 :
2015 5 : spdk_nvmf_request_free_buffers(&rdma_req->req, &rgroup->group, &rtransport->transport);
2016 : }
2017 6 : if (rdma_req->req.stripped_data) {
2018 0 : nvmf_request_free_stripped_buffers(&rdma_req->req,
2019 0 : &rqpair->poller->group->group,
2020 : &rtransport->transport);
2021 : }
2022 6 : nvmf_rdma_request_free_data(rdma_req, rtransport);
2023 6 : rdma_req->req.length = 0;
2024 6 : rdma_req->req.iovcnt = 0;
2025 6 : rdma_req->offset = 0;
2026 6 : rdma_req->req.dif_enabled = false;
2027 6 : rdma_req->fused_failed = false;
2028 6 : rdma_req->transfer_wr = NULL;
2029 6 : if (rdma_req->fused_pair) {
2030 : /* This req was part of a valid fused pair, but failed before it got to
2031 : * READ_TO_EXECUTE state. This means we need to fail the other request
2032 : * in the pair, because it is no longer part of a valid pair. If the pair
2033 : * already reached READY_TO_EXECUTE state, we need to kick it.
2034 : */
2035 0 : rdma_req->fused_pair->fused_failed = true;
2036 0 : if (rdma_req->fused_pair->state == RDMA_REQUEST_STATE_READY_TO_EXECUTE) {
2037 0 : nvmf_rdma_request_process(rtransport, rdma_req->fused_pair);
2038 : }
2039 0 : rdma_req->fused_pair = NULL;
2040 : }
2041 6 : memset(&rdma_req->req.dif, 0, sizeof(rdma_req->req.dif));
2042 6 : rqpair->qd--;
2043 :
2044 6 : STAILQ_INSERT_HEAD(&rqpair->resources->free_queue, rdma_req, state_link);
2045 6 : rdma_req->state = RDMA_REQUEST_STATE_FREE;
2046 6 : }
2047 :
2048 : static void
2049 6 : nvmf_rdma_check_fused_ordering(struct spdk_nvmf_rdma_transport *rtransport,
2050 : struct spdk_nvmf_rdma_qpair *rqpair,
2051 : struct spdk_nvmf_rdma_request *rdma_req)
2052 : {
2053 : enum spdk_nvme_cmd_fuse last, next;
2054 :
2055 6 : last = rqpair->fused_first ? rqpair->fused_first->req.cmd->nvme_cmd.fuse : SPDK_NVME_CMD_FUSE_NONE;
2056 6 : next = rdma_req->req.cmd->nvme_cmd.fuse;
2057 :
2058 6 : assert(last != SPDK_NVME_CMD_FUSE_SECOND);
2059 :
2060 6 : if (spdk_likely(last == SPDK_NVME_CMD_FUSE_NONE && next == SPDK_NVME_CMD_FUSE_NONE)) {
2061 6 : return;
2062 : }
2063 :
2064 0 : if (last == SPDK_NVME_CMD_FUSE_FIRST) {
2065 0 : if (next == SPDK_NVME_CMD_FUSE_SECOND) {
2066 : /* This is a valid pair of fused commands. Point them at each other
2067 : * so they can be submitted consecutively once ready to be executed.
2068 : */
2069 0 : rqpair->fused_first->fused_pair = rdma_req;
2070 0 : rdma_req->fused_pair = rqpair->fused_first;
2071 0 : rqpair->fused_first = NULL;
2072 0 : return;
2073 : } else {
2074 : /* Mark the last req as failed since it wasn't followed by a SECOND. */
2075 0 : rqpair->fused_first->fused_failed = true;
2076 :
2077 : /* If the last req is in READY_TO_EXECUTE state, then call
2078 : * nvmf_rdma_request_process(), otherwise nothing else will kick it.
2079 : */
2080 0 : if (rqpair->fused_first->state == RDMA_REQUEST_STATE_READY_TO_EXECUTE) {
2081 0 : nvmf_rdma_request_process(rtransport, rqpair->fused_first);
2082 : }
2083 :
2084 0 : rqpair->fused_first = NULL;
2085 : }
2086 : }
2087 :
2088 0 : if (next == SPDK_NVME_CMD_FUSE_FIRST) {
2089 : /* Set rqpair->fused_first here so that we know to check that the next request
2090 : * is a SECOND (and to fail this one if it isn't).
2091 : */
2092 0 : rqpair->fused_first = rdma_req;
2093 0 : } else if (next == SPDK_NVME_CMD_FUSE_SECOND) {
2094 : /* Mark this req failed since it ia SECOND and the last one was not a FIRST. */
2095 0 : rdma_req->fused_failed = true;
2096 : }
2097 : }
2098 :
2099 : bool
2100 23 : nvmf_rdma_request_process(struct spdk_nvmf_rdma_transport *rtransport,
2101 : struct spdk_nvmf_rdma_request *rdma_req)
2102 : {
2103 : struct spdk_nvmf_rdma_qpair *rqpair;
2104 : struct spdk_nvmf_rdma_device *device;
2105 : struct spdk_nvmf_rdma_poll_group *rgroup;
2106 23 : struct spdk_nvme_cpl *rsp = &rdma_req->req.rsp->nvme_cpl;
2107 : int rc;
2108 : struct spdk_nvmf_rdma_recv *rdma_recv;
2109 : enum spdk_nvmf_rdma_request_state prev_state;
2110 23 : bool progress = false;
2111 23 : int data_posted;
2112 : uint32_t num_blocks, num_rdma_reads_available, qdepth;
2113 :
2114 23 : rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair);
2115 23 : device = rqpair->device;
2116 23 : rgroup = rqpair->poller->group;
2117 :
2118 23 : assert(rdma_req->state != RDMA_REQUEST_STATE_FREE);
2119 :
2120 : /* If the queue pair is in an error state, force the request to the completed state
2121 : * to release resources. */
2122 23 : if (rqpair->ibv_state == IBV_QPS_ERR || rqpair->qpair.state != SPDK_NVMF_QPAIR_ACTIVE) {
2123 0 : switch (rdma_req->state) {
2124 0 : case RDMA_REQUEST_STATE_NEED_BUFFER:
2125 0 : STAILQ_REMOVE(&rgroup->group.pending_buf_queue, &rdma_req->req, spdk_nvmf_request, buf_link);
2126 0 : break;
2127 0 : case RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING:
2128 0 : STAILQ_REMOVE(&rqpair->pending_rdma_read_queue, rdma_req, spdk_nvmf_rdma_request, state_link);
2129 0 : break;
2130 0 : case RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING:
2131 0 : STAILQ_REMOVE(&rqpair->pending_rdma_write_queue, rdma_req, spdk_nvmf_rdma_request, state_link);
2132 0 : break;
2133 0 : case RDMA_REQUEST_STATE_READY_TO_COMPLETE_PENDING:
2134 0 : STAILQ_REMOVE(&rqpair->pending_rdma_send_queue, rdma_req, spdk_nvmf_rdma_request, state_link);
2135 0 : break;
2136 0 : default:
2137 0 : break;
2138 : }
2139 0 : rdma_req->state = RDMA_REQUEST_STATE_COMPLETED;
2140 : }
2141 :
2142 : /* The loop here is to allow for several back-to-back state changes. */
2143 : do {
2144 66 : prev_state = rdma_req->state;
2145 :
2146 66 : SPDK_DEBUGLOG(rdma, "Request %p entering state %d\n", rdma_req, prev_state);
2147 :
2148 66 : switch (rdma_req->state) {
2149 6 : case RDMA_REQUEST_STATE_FREE:
2150 : /* Some external code must kick a request into RDMA_REQUEST_STATE_NEW
2151 : * to escape this state. */
2152 6 : break;
2153 6 : case RDMA_REQUEST_STATE_NEW:
2154 6 : spdk_trace_record(TRACE_RDMA_REQUEST_STATE_NEW, 0, 0,
2155 : (uintptr_t)rdma_req, (uintptr_t)rqpair);
2156 6 : rdma_recv = rdma_req->recv;
2157 :
2158 : /* The first element of the SGL is the NVMe command */
2159 6 : rdma_req->req.cmd = (union nvmf_h2c_msg *)rdma_recv->sgl[0].addr;
2160 6 : memset(rdma_req->req.rsp, 0, sizeof(*rdma_req->req.rsp));
2161 6 : rdma_req->transfer_wr = &rdma_req->data.wr;
2162 :
2163 6 : if (rqpair->ibv_state == IBV_QPS_ERR || rqpair->qpair.state != SPDK_NVMF_QPAIR_ACTIVE) {
2164 0 : rdma_req->state = RDMA_REQUEST_STATE_COMPLETED;
2165 0 : break;
2166 : }
2167 :
2168 6 : if (spdk_unlikely(spdk_nvmf_request_get_dif_ctx(&rdma_req->req, &rdma_req->req.dif.dif_ctx))) {
2169 0 : rdma_req->req.dif_enabled = true;
2170 : }
2171 :
2172 6 : nvmf_rdma_check_fused_ordering(rtransport, rqpair, rdma_req);
2173 :
2174 : #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL
2175 6 : rdma_req->rsp.wr.opcode = IBV_WR_SEND;
2176 6 : rdma_req->rsp.wr.imm_data = 0;
2177 : #endif
2178 :
2179 : /* The next state transition depends on the data transfer needs of this request. */
2180 6 : rdma_req->req.xfer = spdk_nvmf_req_get_xfer(&rdma_req->req);
2181 :
2182 6 : if (spdk_unlikely(rdma_req->req.xfer == SPDK_NVME_DATA_BIDIRECTIONAL)) {
2183 1 : rsp->status.sct = SPDK_NVME_SCT_GENERIC;
2184 1 : rsp->status.sc = SPDK_NVME_SC_INVALID_OPCODE;
2185 1 : STAILQ_INSERT_TAIL(&rqpair->pending_rdma_send_queue, rdma_req, state_link);
2186 1 : rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE_PENDING;
2187 1 : SPDK_DEBUGLOG(rdma, "Request %p: invalid xfer type (BIDIRECTIONAL)\n", rdma_req);
2188 1 : break;
2189 : }
2190 :
2191 : /* If no data to transfer, ready to execute. */
2192 5 : if (rdma_req->req.xfer == SPDK_NVME_DATA_NONE) {
2193 0 : rdma_req->state = RDMA_REQUEST_STATE_READY_TO_EXECUTE;
2194 0 : break;
2195 : }
2196 :
2197 5 : rdma_req->state = RDMA_REQUEST_STATE_NEED_BUFFER;
2198 5 : STAILQ_INSERT_TAIL(&rgroup->group.pending_buf_queue, &rdma_req->req, buf_link);
2199 5 : break;
2200 5 : case RDMA_REQUEST_STATE_NEED_BUFFER:
2201 5 : spdk_trace_record(TRACE_RDMA_REQUEST_STATE_NEED_BUFFER, 0, 0,
2202 : (uintptr_t)rdma_req, (uintptr_t)rqpair);
2203 :
2204 5 : assert(rdma_req->req.xfer != SPDK_NVME_DATA_NONE);
2205 :
2206 5 : if (&rdma_req->req != STAILQ_FIRST(&rgroup->group.pending_buf_queue)) {
2207 : /* This request needs to wait in line to obtain a buffer */
2208 0 : break;
2209 : }
2210 :
2211 : /* Try to get a data buffer */
2212 5 : rc = nvmf_rdma_request_parse_sgl(rtransport, device, rdma_req);
2213 5 : if (rc < 0) {
2214 0 : STAILQ_REMOVE_HEAD(&rgroup->group.pending_buf_queue, buf_link);
2215 0 : STAILQ_INSERT_TAIL(&rqpair->pending_rdma_send_queue, rdma_req, state_link);
2216 0 : rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE_PENDING;
2217 0 : break;
2218 : }
2219 :
2220 5 : if (rdma_req->req.iovcnt == 0) {
2221 : /* No buffers available. */
2222 0 : rgroup->stat.pending_data_buffer++;
2223 0 : break;
2224 : }
2225 :
2226 5 : STAILQ_REMOVE_HEAD(&rgroup->group.pending_buf_queue, buf_link);
2227 :
2228 : /* If data is transferring from host to controller and the data didn't
2229 : * arrive using in capsule data, we need to do a transfer from the host.
2230 : */
2231 5 : if (rdma_req->req.xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER &&
2232 : rdma_req->req.data_from_pool) {
2233 4 : STAILQ_INSERT_TAIL(&rqpair->pending_rdma_read_queue, rdma_req, state_link);
2234 4 : rdma_req->state = RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING;
2235 4 : break;
2236 : }
2237 :
2238 1 : rdma_req->state = RDMA_REQUEST_STATE_READY_TO_EXECUTE;
2239 1 : break;
2240 4 : case RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING:
2241 4 : spdk_trace_record(TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING, 0, 0,
2242 : (uintptr_t)rdma_req, (uintptr_t)rqpair);
2243 :
2244 4 : if (rdma_req != STAILQ_FIRST(&rqpair->pending_rdma_read_queue)) {
2245 : /* This request needs to wait in line to perform RDMA */
2246 0 : break;
2247 : }
2248 4 : assert(rqpair->max_send_depth >= rqpair->current_send_depth);
2249 4 : qdepth = rqpair->max_send_depth - rqpair->current_send_depth;
2250 4 : assert(rqpair->max_read_depth >= rqpair->current_read_depth);
2251 4 : num_rdma_reads_available = rqpair->max_read_depth - rqpair->current_read_depth;
2252 4 : if (rdma_req->num_outstanding_data_wr > qdepth ||
2253 4 : rdma_req->num_outstanding_data_wr > num_rdma_reads_available) {
2254 0 : if (num_rdma_reads_available && qdepth) {
2255 : /* Send as much as we can */
2256 0 : request_prepare_transfer_in_part(&rdma_req->req, spdk_min(num_rdma_reads_available, qdepth));
2257 : } else {
2258 : /* We can only have so many WRs outstanding. we have to wait until some finish. */
2259 0 : rqpair->poller->stat.pending_rdma_read++;
2260 0 : break;
2261 : }
2262 : }
2263 :
2264 : /* We have already verified that this request is the head of the queue. */
2265 4 : if (rdma_req->num_remaining_data_wr == 0) {
2266 4 : STAILQ_REMOVE_HEAD(&rqpair->pending_rdma_read_queue, state_link);
2267 : }
2268 :
2269 4 : rc = request_transfer_in(&rdma_req->req);
2270 4 : if (!rc) {
2271 4 : rdma_req->state = RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER;
2272 : } else {
2273 0 : rsp->status.sc = SPDK_NVME_SC_INTERNAL_DEVICE_ERROR;
2274 0 : STAILQ_INSERT_TAIL(&rqpair->pending_rdma_send_queue, rdma_req, state_link);
2275 0 : rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE_PENDING;
2276 : }
2277 4 : break;
2278 4 : case RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER:
2279 4 : spdk_trace_record(TRACE_RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER, 0, 0,
2280 : (uintptr_t)rdma_req, (uintptr_t)rqpair);
2281 : /* Some external code must kick a request into RDMA_REQUEST_STATE_READY_TO_EXECUTE
2282 : * to escape this state. */
2283 4 : break;
2284 5 : case RDMA_REQUEST_STATE_READY_TO_EXECUTE:
2285 5 : spdk_trace_record(TRACE_RDMA_REQUEST_STATE_READY_TO_EXECUTE, 0, 0,
2286 : (uintptr_t)rdma_req, (uintptr_t)rqpair);
2287 :
2288 5 : if (spdk_unlikely(rdma_req->req.dif_enabled)) {
2289 0 : if (rdma_req->req.xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER) {
2290 : /* generate DIF for write operation */
2291 0 : num_blocks = SPDK_CEIL_DIV(rdma_req->req.dif.elba_length, rdma_req->req.dif.dif_ctx.block_size);
2292 0 : assert(num_blocks > 0);
2293 :
2294 0 : rc = spdk_dif_generate(rdma_req->req.iov, rdma_req->req.iovcnt,
2295 0 : num_blocks, &rdma_req->req.dif.dif_ctx);
2296 0 : if (rc != 0) {
2297 0 : SPDK_ERRLOG("DIF generation failed\n");
2298 0 : rdma_req->state = RDMA_REQUEST_STATE_COMPLETED;
2299 0 : spdk_nvmf_qpair_disconnect(&rqpair->qpair, NULL, NULL);
2300 0 : break;
2301 : }
2302 : }
2303 :
2304 0 : assert(rdma_req->req.dif.elba_length >= rdma_req->req.length);
2305 : /* set extended length before IO operation */
2306 0 : rdma_req->req.length = rdma_req->req.dif.elba_length;
2307 : }
2308 :
2309 5 : if (rdma_req->req.cmd->nvme_cmd.fuse != SPDK_NVME_CMD_FUSE_NONE) {
2310 0 : if (rdma_req->fused_failed) {
2311 : /* This request failed FUSED semantics. Fail it immediately, without
2312 : * even sending it to the target layer.
2313 : */
2314 0 : rsp->status.sct = SPDK_NVME_SCT_GENERIC;
2315 0 : rsp->status.sc = SPDK_NVME_SC_ABORTED_MISSING_FUSED;
2316 0 : STAILQ_INSERT_TAIL(&rqpair->pending_rdma_send_queue, rdma_req, state_link);
2317 0 : rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE_PENDING;
2318 0 : break;
2319 : }
2320 :
2321 0 : if (rdma_req->fused_pair == NULL ||
2322 0 : rdma_req->fused_pair->state != RDMA_REQUEST_STATE_READY_TO_EXECUTE) {
2323 : /* This request is ready to execute, but either we don't know yet if it's
2324 : * valid - i.e. this is a FIRST but we haven't received the next
2325 : * request yet or the other request of this fused pair isn't ready to
2326 : * execute. So break here and this request will get processed later either
2327 : * when the other request is ready or we find that this request isn't valid.
2328 : */
2329 : break;
2330 : }
2331 : }
2332 :
2333 : /* If we get to this point, and this request is a fused command, we know that
2334 : * it is part of valid sequence (FIRST followed by a SECOND) and that both
2335 : * requests are READY_TO_EXECUTE. So call spdk_nvmf_request_exec() both on this
2336 : * request, and the other request of the fused pair, in the correct order.
2337 : * Also clear the ->fused_pair pointers on both requests, since after this point
2338 : * we no longer need to maintain the relationship between these two requests.
2339 : */
2340 5 : if (rdma_req->req.cmd->nvme_cmd.fuse == SPDK_NVME_CMD_FUSE_SECOND) {
2341 0 : assert(rdma_req->fused_pair != NULL);
2342 0 : assert(rdma_req->fused_pair->fused_pair != NULL);
2343 0 : rdma_req->fused_pair->state = RDMA_REQUEST_STATE_EXECUTING;
2344 0 : spdk_nvmf_request_exec(&rdma_req->fused_pair->req);
2345 0 : rdma_req->fused_pair->fused_pair = NULL;
2346 0 : rdma_req->fused_pair = NULL;
2347 : }
2348 5 : rdma_req->state = RDMA_REQUEST_STATE_EXECUTING;
2349 5 : spdk_nvmf_request_exec(&rdma_req->req);
2350 5 : if (rdma_req->req.cmd->nvme_cmd.fuse == SPDK_NVME_CMD_FUSE_FIRST) {
2351 0 : assert(rdma_req->fused_pair != NULL);
2352 0 : assert(rdma_req->fused_pair->fused_pair != NULL);
2353 0 : rdma_req->fused_pair->state = RDMA_REQUEST_STATE_EXECUTING;
2354 0 : spdk_nvmf_request_exec(&rdma_req->fused_pair->req);
2355 0 : rdma_req->fused_pair->fused_pair = NULL;
2356 0 : rdma_req->fused_pair = NULL;
2357 : }
2358 5 : break;
2359 5 : case RDMA_REQUEST_STATE_EXECUTING:
2360 5 : spdk_trace_record(TRACE_RDMA_REQUEST_STATE_EXECUTING, 0, 0,
2361 : (uintptr_t)rdma_req, (uintptr_t)rqpair);
2362 : /* Some external code must kick a request into RDMA_REQUEST_STATE_EXECUTED
2363 : * to escape this state. */
2364 5 : break;
2365 5 : case RDMA_REQUEST_STATE_EXECUTED:
2366 5 : spdk_trace_record(TRACE_RDMA_REQUEST_STATE_EXECUTED, 0, 0,
2367 : (uintptr_t)rdma_req, (uintptr_t)rqpair);
2368 5 : if (rsp->status.sc == SPDK_NVME_SC_SUCCESS &&
2369 5 : rdma_req->req.xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) {
2370 1 : STAILQ_INSERT_TAIL(&rqpair->pending_rdma_write_queue, rdma_req, state_link);
2371 1 : rdma_req->state = RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING;
2372 : } else {
2373 4 : STAILQ_INSERT_TAIL(&rqpair->pending_rdma_send_queue, rdma_req, state_link);
2374 4 : rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE_PENDING;
2375 : }
2376 5 : if (spdk_unlikely(rdma_req->req.dif_enabled)) {
2377 : /* restore the original length */
2378 0 : rdma_req->req.length = rdma_req->req.dif.orig_length;
2379 :
2380 0 : if (rdma_req->req.xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) {
2381 0 : struct spdk_dif_error error_blk;
2382 :
2383 0 : num_blocks = SPDK_CEIL_DIV(rdma_req->req.dif.elba_length, rdma_req->req.dif.dif_ctx.block_size);
2384 0 : if (!rdma_req->req.stripped_data) {
2385 0 : rc = spdk_dif_verify(rdma_req->req.iov, rdma_req->req.iovcnt, num_blocks,
2386 0 : &rdma_req->req.dif.dif_ctx, &error_blk);
2387 : } else {
2388 0 : rc = spdk_dif_verify_copy(rdma_req->req.stripped_data->iov,
2389 0 : rdma_req->req.stripped_data->iovcnt,
2390 0 : rdma_req->req.iov, rdma_req->req.iovcnt, num_blocks,
2391 0 : &rdma_req->req.dif.dif_ctx, &error_blk);
2392 : }
2393 0 : if (rc) {
2394 0 : struct spdk_nvme_cpl *rsp = &rdma_req->req.rsp->nvme_cpl;
2395 :
2396 0 : SPDK_ERRLOG("DIF error detected. type=%d, offset=%" PRIu32 "\n", error_blk.err_type,
2397 : error_blk.err_offset);
2398 0 : rsp->status.sct = SPDK_NVME_SCT_MEDIA_ERROR;
2399 0 : rsp->status.sc = nvmf_rdma_dif_error_to_compl_status(error_blk.err_type);
2400 0 : STAILQ_REMOVE(&rqpair->pending_rdma_write_queue, rdma_req, spdk_nvmf_rdma_request, state_link);
2401 0 : STAILQ_INSERT_TAIL(&rqpair->pending_rdma_send_queue, rdma_req, state_link);
2402 0 : rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE_PENDING;
2403 : }
2404 : }
2405 : }
2406 5 : break;
2407 1 : case RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING:
2408 1 : spdk_trace_record(TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING, 0, 0,
2409 : (uintptr_t)rdma_req, (uintptr_t)rqpair);
2410 :
2411 1 : if (rdma_req != STAILQ_FIRST(&rqpair->pending_rdma_write_queue)) {
2412 : /* This request needs to wait in line to perform RDMA */
2413 0 : break;
2414 : }
2415 1 : if ((rqpair->current_send_depth + rdma_req->num_outstanding_data_wr + 1) >
2416 1 : rqpair->max_send_depth) {
2417 : /* We can only have so many WRs outstanding. we have to wait until some finish.
2418 : * +1 since each request has an additional wr in the resp. */
2419 0 : rqpair->poller->stat.pending_rdma_write++;
2420 0 : break;
2421 : }
2422 :
2423 : /* We have already verified that this request is the head of the queue. */
2424 1 : STAILQ_REMOVE_HEAD(&rqpair->pending_rdma_write_queue, state_link);
2425 :
2426 : /* The data transfer will be kicked off from
2427 : * RDMA_REQUEST_STATE_READY_TO_COMPLETE state.
2428 : * We verified that data + response fit into send queue, so we can go to the next state directly
2429 : */
2430 1 : rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE;
2431 1 : break;
2432 7 : case RDMA_REQUEST_STATE_READY_TO_COMPLETE_PENDING:
2433 7 : spdk_trace_record(TRACE_RDMA_REQUEST_STATE_READY_TO_COMPLETE_PENDING, 0, 0,
2434 : (uintptr_t)rdma_req, (uintptr_t)rqpair);
2435 :
2436 7 : if (rdma_req != STAILQ_FIRST(&rqpair->pending_rdma_send_queue)) {
2437 : /* This request needs to wait in line to send the completion */
2438 0 : break;
2439 : }
2440 :
2441 7 : assert(rqpair->current_send_depth <= rqpair->max_send_depth);
2442 7 : if (rqpair->current_send_depth == rqpair->max_send_depth) {
2443 : /* We can only have so many WRs outstanding. we have to wait until some finish */
2444 2 : rqpair->poller->stat.pending_rdma_send++;
2445 2 : break;
2446 : }
2447 :
2448 : /* We have already verified that this request is the head of the queue. */
2449 5 : STAILQ_REMOVE_HEAD(&rqpair->pending_rdma_send_queue, state_link);
2450 :
2451 : /* The response sending will be kicked off from
2452 : * RDMA_REQUEST_STATE_READY_TO_COMPLETE state.
2453 : */
2454 5 : rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE;
2455 5 : break;
2456 6 : case RDMA_REQUEST_STATE_READY_TO_COMPLETE:
2457 6 : spdk_trace_record(TRACE_RDMA_REQUEST_STATE_READY_TO_COMPLETE, 0, 0,
2458 : (uintptr_t)rdma_req, (uintptr_t)rqpair);
2459 6 : rc = request_transfer_out(&rdma_req->req, &data_posted);
2460 6 : assert(rc == 0); /* No good way to handle this currently */
2461 6 : if (rc) {
2462 0 : rdma_req->state = RDMA_REQUEST_STATE_COMPLETED;
2463 : } else {
2464 6 : rdma_req->state = data_posted ? RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST :
2465 : RDMA_REQUEST_STATE_COMPLETING;
2466 : }
2467 6 : break;
2468 1 : case RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST:
2469 1 : spdk_trace_record(TRACE_RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST, 0, 0,
2470 : (uintptr_t)rdma_req, (uintptr_t)rqpair);
2471 : /* Some external code must kick a request into RDMA_REQUEST_STATE_COMPLETED
2472 : * to escape this state. */
2473 1 : break;
2474 5 : case RDMA_REQUEST_STATE_COMPLETING:
2475 5 : spdk_trace_record(TRACE_RDMA_REQUEST_STATE_COMPLETING, 0, 0,
2476 : (uintptr_t)rdma_req, (uintptr_t)rqpair);
2477 : /* Some external code must kick a request into RDMA_REQUEST_STATE_COMPLETED
2478 : * to escape this state. */
2479 5 : break;
2480 6 : case RDMA_REQUEST_STATE_COMPLETED:
2481 6 : spdk_trace_record(TRACE_RDMA_REQUEST_STATE_COMPLETED, 0, 0,
2482 : (uintptr_t)rdma_req, (uintptr_t)rqpair);
2483 :
2484 6 : rqpair->poller->stat.request_latency += spdk_get_ticks() - rdma_req->receive_tsc;
2485 6 : _nvmf_rdma_request_free(rdma_req, rtransport);
2486 6 : break;
2487 0 : case RDMA_REQUEST_NUM_STATES:
2488 : default:
2489 0 : assert(0);
2490 : break;
2491 : }
2492 :
2493 66 : if (rdma_req->state != prev_state) {
2494 43 : progress = true;
2495 : }
2496 66 : } while (rdma_req->state != prev_state);
2497 :
2498 23 : return progress;
2499 : }
2500 :
2501 : /* Public API callbacks begin here */
2502 :
2503 : #define SPDK_NVMF_RDMA_DEFAULT_MAX_QUEUE_DEPTH 128
2504 : #define SPDK_NVMF_RDMA_DEFAULT_AQ_DEPTH 128
2505 : #define SPDK_NVMF_RDMA_DEFAULT_SRQ_DEPTH 4096
2506 : #define SPDK_NVMF_RDMA_DEFAULT_MAX_QPAIRS_PER_CTRLR 128
2507 : #define SPDK_NVMF_RDMA_DEFAULT_IN_CAPSULE_DATA_SIZE 4096
2508 : #define SPDK_NVMF_RDMA_DEFAULT_MAX_IO_SIZE 131072
2509 : #define SPDK_NVMF_RDMA_MIN_IO_BUFFER_SIZE (SPDK_NVMF_RDMA_DEFAULT_MAX_IO_SIZE / SPDK_NVMF_MAX_SGL_ENTRIES)
2510 : #define SPDK_NVMF_RDMA_DEFAULT_NUM_SHARED_BUFFERS 4095
2511 : #define SPDK_NVMF_RDMA_DEFAULT_BUFFER_CACHE_SIZE UINT32_MAX
2512 : #define SPDK_NVMF_RDMA_DEFAULT_NO_SRQ false
2513 : #define SPDK_NVMF_RDMA_DIF_INSERT_OR_STRIP false
2514 : #define SPDK_NVMF_RDMA_ACCEPTOR_BACKLOG 100
2515 : #define SPDK_NVMF_RDMA_DEFAULT_ABORT_TIMEOUT_SEC 1
2516 : #define SPDK_NVMF_RDMA_DEFAULT_NO_WR_BATCHING false
2517 :
2518 : static void
2519 1 : nvmf_rdma_opts_init(struct spdk_nvmf_transport_opts *opts)
2520 : {
2521 1 : opts->max_queue_depth = SPDK_NVMF_RDMA_DEFAULT_MAX_QUEUE_DEPTH;
2522 1 : opts->max_qpairs_per_ctrlr = SPDK_NVMF_RDMA_DEFAULT_MAX_QPAIRS_PER_CTRLR;
2523 1 : opts->in_capsule_data_size = SPDK_NVMF_RDMA_DEFAULT_IN_CAPSULE_DATA_SIZE;
2524 1 : opts->max_io_size = SPDK_NVMF_RDMA_DEFAULT_MAX_IO_SIZE;
2525 1 : opts->io_unit_size = SPDK_NVMF_RDMA_MIN_IO_BUFFER_SIZE;
2526 1 : opts->max_aq_depth = SPDK_NVMF_RDMA_DEFAULT_AQ_DEPTH;
2527 1 : opts->num_shared_buffers = SPDK_NVMF_RDMA_DEFAULT_NUM_SHARED_BUFFERS;
2528 1 : opts->buf_cache_size = SPDK_NVMF_RDMA_DEFAULT_BUFFER_CACHE_SIZE;
2529 1 : opts->dif_insert_or_strip = SPDK_NVMF_RDMA_DIF_INSERT_OR_STRIP;
2530 1 : opts->abort_timeout_sec = SPDK_NVMF_RDMA_DEFAULT_ABORT_TIMEOUT_SEC;
2531 1 : opts->transport_specific = NULL;
2532 1 : }
2533 :
2534 : static int nvmf_rdma_destroy(struct spdk_nvmf_transport *transport,
2535 : spdk_nvmf_transport_destroy_done_cb cb_fn, void *cb_arg);
2536 :
2537 : static inline bool
2538 0 : nvmf_rdma_is_rxe_device(struct spdk_nvmf_rdma_device *device)
2539 : {
2540 0 : return device->attr.vendor_id == SPDK_RDMA_RXE_VENDOR_ID_OLD ||
2541 0 : device->attr.vendor_id == SPDK_RDMA_RXE_VENDOR_ID_NEW;
2542 : }
2543 :
2544 : static int nvmf_rdma_accept(void *ctx);
2545 : static bool nvmf_rdma_retry_listen_port(struct spdk_nvmf_rdma_transport *rtransport);
2546 : static void destroy_ib_device(struct spdk_nvmf_rdma_transport *rtransport,
2547 : struct spdk_nvmf_rdma_device *device);
2548 :
2549 : static int
2550 0 : create_ib_device(struct spdk_nvmf_rdma_transport *rtransport, struct ibv_context *context,
2551 : struct spdk_nvmf_rdma_device **new_device)
2552 : {
2553 : struct spdk_nvmf_rdma_device *device;
2554 0 : int flag = 0;
2555 0 : int rc = 0;
2556 :
2557 0 : device = calloc(1, sizeof(*device));
2558 0 : if (!device) {
2559 0 : SPDK_ERRLOG("Unable to allocate memory for RDMA devices.\n");
2560 0 : return -ENOMEM;
2561 : }
2562 0 : device->context = context;
2563 0 : rc = ibv_query_device(device->context, &device->attr);
2564 0 : if (rc < 0) {
2565 0 : SPDK_ERRLOG("Failed to query RDMA device attributes.\n");
2566 0 : free(device);
2567 0 : return rc;
2568 : }
2569 :
2570 : #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL
2571 0 : if ((device->attr.device_cap_flags & IBV_DEVICE_MEM_MGT_EXTENSIONS) == 0) {
2572 0 : SPDK_WARNLOG("The libibverbs on this system supports SEND_WITH_INVALIDATE,");
2573 0 : SPDK_WARNLOG("but the device with vendor ID %u does not.\n", device->attr.vendor_id);
2574 : }
2575 :
2576 : /**
2577 : * The vendor ID is assigned by the IEEE and an ID of 0 implies Soft-RoCE.
2578 : * The Soft-RoCE RXE driver does not currently support send with invalidate,
2579 : * but incorrectly reports that it does. There are changes making their way
2580 : * through the kernel now that will enable this feature. When they are merged,
2581 : * we can conditionally enable this feature.
2582 : *
2583 : * TODO: enable this for versions of the kernel rxe driver that support it.
2584 : */
2585 0 : if (nvmf_rdma_is_rxe_device(device)) {
2586 0 : device->attr.device_cap_flags &= ~(IBV_DEVICE_MEM_MGT_EXTENSIONS);
2587 : }
2588 : #endif
2589 :
2590 : /* set up device context async ev fd as NON_BLOCKING */
2591 0 : flag = fcntl(device->context->async_fd, F_GETFL);
2592 0 : rc = fcntl(device->context->async_fd, F_SETFL, flag | O_NONBLOCK);
2593 0 : if (rc < 0) {
2594 0 : SPDK_ERRLOG("Failed to set context async fd to NONBLOCK.\n");
2595 0 : free(device);
2596 0 : return rc;
2597 : }
2598 :
2599 0 : TAILQ_INSERT_TAIL(&rtransport->devices, device, link);
2600 0 : SPDK_DEBUGLOG(rdma, "New device %p is added to RDMA trasport\n", device);
2601 :
2602 0 : if (g_nvmf_hooks.get_ibv_pd) {
2603 0 : device->pd = g_nvmf_hooks.get_ibv_pd(NULL, device->context);
2604 : } else {
2605 0 : device->pd = ibv_alloc_pd(device->context);
2606 : }
2607 :
2608 0 : if (!device->pd) {
2609 0 : SPDK_ERRLOG("Unable to allocate protection domain.\n");
2610 0 : destroy_ib_device(rtransport, device);
2611 0 : return -ENOMEM;
2612 : }
2613 :
2614 0 : assert(device->map == NULL);
2615 :
2616 0 : device->map = spdk_rdma_create_mem_map(device->pd, &g_nvmf_hooks, SPDK_RDMA_MEMORY_MAP_ROLE_TARGET);
2617 0 : if (!device->map) {
2618 0 : SPDK_ERRLOG("Unable to allocate memory map for listen address\n");
2619 0 : destroy_ib_device(rtransport, device);
2620 0 : return -ENOMEM;
2621 : }
2622 :
2623 0 : assert(device->map != NULL);
2624 0 : assert(device->pd != NULL);
2625 :
2626 0 : if (new_device) {
2627 0 : *new_device = device;
2628 : }
2629 0 : SPDK_NOTICELOG("Create IB device %s(%p/%p) succeed.\n", ibv_get_device_name(context->device),
2630 : device, context);
2631 :
2632 0 : return 0;
2633 : }
2634 :
2635 : static void
2636 0 : free_poll_fds(struct spdk_nvmf_rdma_transport *rtransport)
2637 : {
2638 0 : if (rtransport->poll_fds) {
2639 0 : free(rtransport->poll_fds);
2640 0 : rtransport->poll_fds = NULL;
2641 : }
2642 0 : rtransport->npoll_fds = 0;
2643 0 : }
2644 :
2645 : static int
2646 0 : generate_poll_fds(struct spdk_nvmf_rdma_transport *rtransport)
2647 : {
2648 : /* Set up poll descriptor array to monitor events from RDMA and IB
2649 : * in a single poll syscall
2650 : */
2651 0 : int device_count = 0;
2652 0 : int i = 0;
2653 : struct spdk_nvmf_rdma_device *device, *tmp;
2654 :
2655 0 : TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, tmp) {
2656 0 : device_count++;
2657 : }
2658 :
2659 0 : rtransport->npoll_fds = device_count + 1;
2660 :
2661 0 : rtransport->poll_fds = calloc(rtransport->npoll_fds, sizeof(struct pollfd));
2662 0 : if (rtransport->poll_fds == NULL) {
2663 0 : SPDK_ERRLOG("poll_fds allocation failed\n");
2664 0 : return -ENOMEM;
2665 : }
2666 :
2667 0 : rtransport->poll_fds[i].fd = rtransport->event_channel->fd;
2668 0 : rtransport->poll_fds[i++].events = POLLIN;
2669 :
2670 0 : TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, tmp) {
2671 0 : rtransport->poll_fds[i].fd = device->context->async_fd;
2672 0 : rtransport->poll_fds[i++].events = POLLIN;
2673 : }
2674 :
2675 0 : return 0;
2676 : }
2677 :
2678 : static struct spdk_nvmf_transport *
2679 0 : nvmf_rdma_create(struct spdk_nvmf_transport_opts *opts)
2680 : {
2681 : int rc;
2682 : struct spdk_nvmf_rdma_transport *rtransport;
2683 0 : struct spdk_nvmf_rdma_device *device;
2684 : struct ibv_context **contexts;
2685 : uint32_t i;
2686 : int flag;
2687 : uint32_t sge_count;
2688 : uint32_t min_shared_buffers;
2689 : uint32_t min_in_capsule_data_size;
2690 0 : int max_device_sge = SPDK_NVMF_MAX_SGL_ENTRIES;
2691 :
2692 0 : rtransport = calloc(1, sizeof(*rtransport));
2693 0 : if (!rtransport) {
2694 0 : return NULL;
2695 : }
2696 :
2697 0 : TAILQ_INIT(&rtransport->devices);
2698 0 : TAILQ_INIT(&rtransport->ports);
2699 0 : TAILQ_INIT(&rtransport->poll_groups);
2700 0 : TAILQ_INIT(&rtransport->retry_ports);
2701 :
2702 0 : rtransport->transport.ops = &spdk_nvmf_transport_rdma;
2703 0 : rtransport->rdma_opts.num_cqe = DEFAULT_NVMF_RDMA_CQ_SIZE;
2704 0 : rtransport->rdma_opts.max_srq_depth = SPDK_NVMF_RDMA_DEFAULT_SRQ_DEPTH;
2705 0 : rtransport->rdma_opts.no_srq = SPDK_NVMF_RDMA_DEFAULT_NO_SRQ;
2706 0 : rtransport->rdma_opts.acceptor_backlog = SPDK_NVMF_RDMA_ACCEPTOR_BACKLOG;
2707 0 : rtransport->rdma_opts.no_wr_batching = SPDK_NVMF_RDMA_DEFAULT_NO_WR_BATCHING;
2708 0 : if (opts->transport_specific != NULL &&
2709 0 : spdk_json_decode_object_relaxed(opts->transport_specific, rdma_transport_opts_decoder,
2710 : SPDK_COUNTOF(rdma_transport_opts_decoder),
2711 0 : &rtransport->rdma_opts)) {
2712 0 : SPDK_ERRLOG("spdk_json_decode_object_relaxed failed\n");
2713 0 : nvmf_rdma_destroy(&rtransport->transport, NULL, NULL);
2714 0 : return NULL;
2715 : }
2716 :
2717 0 : SPDK_INFOLOG(rdma, "*** RDMA Transport Init ***\n"
2718 : " Transport opts: max_ioq_depth=%d, max_io_size=%d,\n"
2719 : " max_io_qpairs_per_ctrlr=%d, io_unit_size=%d,\n"
2720 : " in_capsule_data_size=%d, max_aq_depth=%d,\n"
2721 : " num_shared_buffers=%d, num_cqe=%d, max_srq_depth=%d, no_srq=%d,"
2722 : " acceptor_backlog=%d, no_wr_batching=%d abort_timeout_sec=%d\n",
2723 : opts->max_queue_depth,
2724 : opts->max_io_size,
2725 : opts->max_qpairs_per_ctrlr - 1,
2726 : opts->io_unit_size,
2727 : opts->in_capsule_data_size,
2728 : opts->max_aq_depth,
2729 : opts->num_shared_buffers,
2730 : rtransport->rdma_opts.num_cqe,
2731 : rtransport->rdma_opts.max_srq_depth,
2732 : rtransport->rdma_opts.no_srq,
2733 : rtransport->rdma_opts.acceptor_backlog,
2734 : rtransport->rdma_opts.no_wr_batching,
2735 : opts->abort_timeout_sec);
2736 :
2737 : /* I/O unit size cannot be larger than max I/O size */
2738 0 : if (opts->io_unit_size > opts->max_io_size) {
2739 0 : opts->io_unit_size = opts->max_io_size;
2740 : }
2741 :
2742 0 : if (rtransport->rdma_opts.acceptor_backlog <= 0) {
2743 0 : SPDK_ERRLOG("The acceptor backlog cannot be less than 1, setting to the default value of (%d).\n",
2744 : SPDK_NVMF_RDMA_ACCEPTOR_BACKLOG);
2745 0 : rtransport->rdma_opts.acceptor_backlog = SPDK_NVMF_RDMA_ACCEPTOR_BACKLOG;
2746 : }
2747 :
2748 0 : if (opts->num_shared_buffers < (SPDK_NVMF_MAX_SGL_ENTRIES * 2)) {
2749 0 : SPDK_ERRLOG("The number of shared data buffers (%d) is less than"
2750 : "the minimum number required to guarantee that forward progress can be made (%d)\n",
2751 : opts->num_shared_buffers, (SPDK_NVMF_MAX_SGL_ENTRIES * 2));
2752 0 : nvmf_rdma_destroy(&rtransport->transport, NULL, NULL);
2753 0 : return NULL;
2754 : }
2755 :
2756 : /* If buf_cache_size == UINT32_MAX, we will dynamically pick a cache size later that we know will fit. */
2757 0 : if (opts->buf_cache_size < UINT32_MAX) {
2758 0 : min_shared_buffers = spdk_env_get_core_count() * opts->buf_cache_size;
2759 0 : if (min_shared_buffers > opts->num_shared_buffers) {
2760 0 : SPDK_ERRLOG("There are not enough buffers to satisfy"
2761 : "per-poll group caches for each thread. (%" PRIu32 ")"
2762 : "supplied. (%" PRIu32 ") required\n", opts->num_shared_buffers, min_shared_buffers);
2763 0 : SPDK_ERRLOG("Please specify a larger number of shared buffers\n");
2764 0 : nvmf_rdma_destroy(&rtransport->transport, NULL, NULL);
2765 0 : return NULL;
2766 : }
2767 : }
2768 :
2769 0 : sge_count = opts->max_io_size / opts->io_unit_size;
2770 0 : if (sge_count > NVMF_DEFAULT_TX_SGE) {
2771 0 : SPDK_ERRLOG("Unsupported IO Unit size specified, %d bytes\n", opts->io_unit_size);
2772 0 : nvmf_rdma_destroy(&rtransport->transport, NULL, NULL);
2773 0 : return NULL;
2774 : }
2775 :
2776 0 : min_in_capsule_data_size = sizeof(struct spdk_nvme_sgl_descriptor) * SPDK_NVMF_MAX_SGL_ENTRIES;
2777 0 : if (opts->in_capsule_data_size < min_in_capsule_data_size) {
2778 0 : SPDK_WARNLOG("In capsule data size is set to %u, this is minimum size required to support msdbd=16\n",
2779 : min_in_capsule_data_size);
2780 0 : opts->in_capsule_data_size = min_in_capsule_data_size;
2781 : }
2782 :
2783 0 : rtransport->event_channel = rdma_create_event_channel();
2784 0 : if (rtransport->event_channel == NULL) {
2785 0 : SPDK_ERRLOG("rdma_create_event_channel() failed, %s\n", spdk_strerror(errno));
2786 0 : nvmf_rdma_destroy(&rtransport->transport, NULL, NULL);
2787 0 : return NULL;
2788 : }
2789 :
2790 0 : flag = fcntl(rtransport->event_channel->fd, F_GETFL);
2791 0 : if (fcntl(rtransport->event_channel->fd, F_SETFL, flag | O_NONBLOCK) < 0) {
2792 0 : SPDK_ERRLOG("fcntl can't set nonblocking mode for socket, fd: %d (%s)\n",
2793 : rtransport->event_channel->fd, spdk_strerror(errno));
2794 0 : nvmf_rdma_destroy(&rtransport->transport, NULL, NULL);
2795 0 : return NULL;
2796 : }
2797 :
2798 0 : rtransport->data_wr_pool = spdk_mempool_create("spdk_nvmf_rdma_wr_data",
2799 0 : opts->max_queue_depth * SPDK_NVMF_MAX_SGL_ENTRIES,
2800 : sizeof(struct spdk_nvmf_rdma_request_data),
2801 : SPDK_MEMPOOL_DEFAULT_CACHE_SIZE,
2802 : SPDK_ENV_SOCKET_ID_ANY);
2803 0 : if (!rtransport->data_wr_pool) {
2804 0 : if (spdk_mempool_lookup("spdk_nvmf_rdma_wr_data") != NULL) {
2805 0 : SPDK_ERRLOG("Unable to allocate work request pool for poll group: already exists\n");
2806 0 : SPDK_ERRLOG("Probably running in multiprocess environment, which is "
2807 : "unsupported by the nvmf library\n");
2808 : } else {
2809 0 : SPDK_ERRLOG("Unable to allocate work request pool for poll group\n");
2810 : }
2811 0 : nvmf_rdma_destroy(&rtransport->transport, NULL, NULL);
2812 0 : return NULL;
2813 : }
2814 :
2815 0 : contexts = rdma_get_devices(NULL);
2816 0 : if (contexts == NULL) {
2817 0 : SPDK_ERRLOG("rdma_get_devices() failed: %s (%d)\n", spdk_strerror(errno), errno);
2818 0 : nvmf_rdma_destroy(&rtransport->transport, NULL, NULL);
2819 0 : return NULL;
2820 : }
2821 :
2822 0 : i = 0;
2823 0 : rc = 0;
2824 0 : while (contexts[i] != NULL) {
2825 0 : rc = create_ib_device(rtransport, contexts[i], &device);
2826 0 : if (rc < 0) {
2827 0 : break;
2828 : }
2829 0 : i++;
2830 0 : max_device_sge = spdk_min(max_device_sge, device->attr.max_sge);
2831 0 : device->is_ready = true;
2832 : }
2833 0 : rdma_free_devices(contexts);
2834 :
2835 0 : if (opts->io_unit_size * max_device_sge < opts->max_io_size) {
2836 : /* divide and round up. */
2837 0 : opts->io_unit_size = (opts->max_io_size + max_device_sge - 1) / max_device_sge;
2838 :
2839 : /* round up to the nearest 4k. */
2840 0 : opts->io_unit_size = (opts->io_unit_size + NVMF_DATA_BUFFER_ALIGNMENT - 1) & ~NVMF_DATA_BUFFER_MASK;
2841 :
2842 0 : opts->io_unit_size = spdk_max(opts->io_unit_size, SPDK_NVMF_RDMA_MIN_IO_BUFFER_SIZE);
2843 0 : SPDK_NOTICELOG("Adjusting the io unit size to fit the device's maximum I/O size. New I/O unit size %u\n",
2844 : opts->io_unit_size);
2845 : }
2846 :
2847 0 : if (rc < 0) {
2848 0 : nvmf_rdma_destroy(&rtransport->transport, NULL, NULL);
2849 0 : return NULL;
2850 : }
2851 :
2852 0 : rc = generate_poll_fds(rtransport);
2853 0 : if (rc < 0) {
2854 0 : nvmf_rdma_destroy(&rtransport->transport, NULL, NULL);
2855 0 : return NULL;
2856 : }
2857 :
2858 0 : rtransport->accept_poller = SPDK_POLLER_REGISTER(nvmf_rdma_accept, &rtransport->transport,
2859 : opts->acceptor_poll_rate);
2860 0 : if (!rtransport->accept_poller) {
2861 0 : nvmf_rdma_destroy(&rtransport->transport, NULL, NULL);
2862 0 : return NULL;
2863 : }
2864 :
2865 0 : return &rtransport->transport;
2866 : }
2867 :
2868 : static void
2869 0 : destroy_ib_device(struct spdk_nvmf_rdma_transport *rtransport,
2870 : struct spdk_nvmf_rdma_device *device)
2871 : {
2872 0 : TAILQ_REMOVE(&rtransport->devices, device, link);
2873 0 : spdk_rdma_free_mem_map(&device->map);
2874 0 : if (device->pd) {
2875 0 : if (!g_nvmf_hooks.get_ibv_pd) {
2876 0 : ibv_dealloc_pd(device->pd);
2877 : }
2878 : }
2879 0 : SPDK_DEBUGLOG(rdma, "IB device [%p] is destroyed.\n", device);
2880 0 : free(device);
2881 0 : }
2882 :
2883 : static void
2884 0 : nvmf_rdma_dump_opts(struct spdk_nvmf_transport *transport, struct spdk_json_write_ctx *w)
2885 : {
2886 : struct spdk_nvmf_rdma_transport *rtransport;
2887 0 : assert(w != NULL);
2888 :
2889 0 : rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport);
2890 0 : spdk_json_write_named_uint32(w, "max_srq_depth", rtransport->rdma_opts.max_srq_depth);
2891 0 : spdk_json_write_named_bool(w, "no_srq", rtransport->rdma_opts.no_srq);
2892 0 : if (rtransport->rdma_opts.no_srq == true) {
2893 0 : spdk_json_write_named_int32(w, "num_cqe", rtransport->rdma_opts.num_cqe);
2894 : }
2895 0 : spdk_json_write_named_int32(w, "acceptor_backlog", rtransport->rdma_opts.acceptor_backlog);
2896 0 : spdk_json_write_named_bool(w, "no_wr_batching", rtransport->rdma_opts.no_wr_batching);
2897 0 : }
2898 :
2899 : static int
2900 0 : nvmf_rdma_destroy(struct spdk_nvmf_transport *transport,
2901 : spdk_nvmf_transport_destroy_done_cb cb_fn, void *cb_arg)
2902 : {
2903 : struct spdk_nvmf_rdma_transport *rtransport;
2904 : struct spdk_nvmf_rdma_port *port, *port_tmp;
2905 : struct spdk_nvmf_rdma_device *device, *device_tmp;
2906 :
2907 0 : rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport);
2908 :
2909 0 : TAILQ_FOREACH_SAFE(port, &rtransport->retry_ports, link, port_tmp) {
2910 0 : TAILQ_REMOVE(&rtransport->retry_ports, port, link);
2911 0 : free(port);
2912 : }
2913 :
2914 0 : TAILQ_FOREACH_SAFE(port, &rtransport->ports, link, port_tmp) {
2915 0 : TAILQ_REMOVE(&rtransport->ports, port, link);
2916 0 : rdma_destroy_id(port->id);
2917 0 : free(port);
2918 : }
2919 :
2920 0 : free_poll_fds(rtransport);
2921 :
2922 0 : if (rtransport->event_channel != NULL) {
2923 0 : rdma_destroy_event_channel(rtransport->event_channel);
2924 : }
2925 :
2926 0 : TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, device_tmp) {
2927 0 : destroy_ib_device(rtransport, device);
2928 : }
2929 :
2930 0 : if (rtransport->data_wr_pool != NULL) {
2931 0 : if (spdk_mempool_count(rtransport->data_wr_pool) !=
2932 0 : (transport->opts.max_queue_depth * SPDK_NVMF_MAX_SGL_ENTRIES)) {
2933 0 : SPDK_ERRLOG("transport wr pool count is %zu but should be %u\n",
2934 : spdk_mempool_count(rtransport->data_wr_pool),
2935 : transport->opts.max_queue_depth * SPDK_NVMF_MAX_SGL_ENTRIES);
2936 : }
2937 : }
2938 :
2939 0 : spdk_mempool_free(rtransport->data_wr_pool);
2940 :
2941 0 : spdk_poller_unregister(&rtransport->accept_poller);
2942 0 : free(rtransport);
2943 :
2944 0 : if (cb_fn) {
2945 0 : cb_fn(cb_arg);
2946 : }
2947 0 : return 0;
2948 : }
2949 :
2950 : static int nvmf_rdma_trid_from_cm_id(struct rdma_cm_id *id,
2951 : struct spdk_nvme_transport_id *trid,
2952 : bool peer);
2953 :
2954 : static bool nvmf_rdma_rescan_devices(struct spdk_nvmf_rdma_transport *rtransport);
2955 :
2956 : static int
2957 0 : nvmf_rdma_listen(struct spdk_nvmf_transport *transport, const struct spdk_nvme_transport_id *trid,
2958 : struct spdk_nvmf_listen_opts *listen_opts)
2959 : {
2960 : struct spdk_nvmf_rdma_transport *rtransport;
2961 : struct spdk_nvmf_rdma_device *device;
2962 : struct spdk_nvmf_rdma_port *port, *tmp_port;
2963 0 : struct addrinfo *res;
2964 0 : struct addrinfo hints;
2965 : int family;
2966 : int rc;
2967 : long int port_val;
2968 0 : bool is_retry = false;
2969 :
2970 0 : if (!strlen(trid->trsvcid)) {
2971 0 : SPDK_ERRLOG("Service id is required\n");
2972 0 : return -EINVAL;
2973 : }
2974 :
2975 0 : rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport);
2976 0 : assert(rtransport->event_channel != NULL);
2977 :
2978 0 : port = calloc(1, sizeof(*port));
2979 0 : if (!port) {
2980 0 : SPDK_ERRLOG("Port allocation failed\n");
2981 0 : return -ENOMEM;
2982 : }
2983 :
2984 0 : port->trid = trid;
2985 :
2986 0 : switch (trid->adrfam) {
2987 0 : case SPDK_NVMF_ADRFAM_IPV4:
2988 0 : family = AF_INET;
2989 0 : break;
2990 0 : case SPDK_NVMF_ADRFAM_IPV6:
2991 0 : family = AF_INET6;
2992 0 : break;
2993 0 : default:
2994 0 : SPDK_ERRLOG("Unhandled ADRFAM %d\n", trid->adrfam);
2995 0 : free(port);
2996 0 : return -EINVAL;
2997 : }
2998 :
2999 0 : memset(&hints, 0, sizeof(hints));
3000 0 : hints.ai_family = family;
3001 0 : hints.ai_flags = AI_NUMERICSERV;
3002 0 : hints.ai_socktype = SOCK_STREAM;
3003 0 : hints.ai_protocol = 0;
3004 :
3005 : /* Range check the trsvcid. Fail in 3 cases:
3006 : * < 0: means that spdk_strtol hit an error
3007 : * 0: this results in ephemeral port which we don't want
3008 : * > 65535: port too high
3009 : */
3010 0 : port_val = spdk_strtol(trid->trsvcid, 10);
3011 0 : if (port_val <= 0 || port_val > 65535) {
3012 0 : SPDK_ERRLOG("invalid trsvcid %s\n", trid->trsvcid);
3013 0 : free(port);
3014 0 : return -EINVAL;
3015 : }
3016 :
3017 0 : rc = getaddrinfo(trid->traddr, trid->trsvcid, &hints, &res);
3018 0 : if (rc) {
3019 0 : SPDK_ERRLOG("getaddrinfo failed: %s (%d)\n", gai_strerror(rc), rc);
3020 0 : free(port);
3021 0 : return -(abs(rc));
3022 : }
3023 :
3024 0 : rc = rdma_create_id(rtransport->event_channel, &port->id, port, RDMA_PS_TCP);
3025 0 : if (rc < 0) {
3026 0 : SPDK_ERRLOG("rdma_create_id() failed\n");
3027 0 : freeaddrinfo(res);
3028 0 : free(port);
3029 0 : return rc;
3030 : }
3031 :
3032 0 : rc = rdma_bind_addr(port->id, res->ai_addr);
3033 0 : freeaddrinfo(res);
3034 :
3035 0 : if (rc < 0) {
3036 0 : TAILQ_FOREACH(tmp_port, &rtransport->retry_ports, link) {
3037 0 : if (spdk_nvme_transport_id_compare(tmp_port->trid, trid) == 0) {
3038 0 : is_retry = true;
3039 0 : break;
3040 : }
3041 : }
3042 0 : if (!is_retry) {
3043 0 : SPDK_ERRLOG("rdma_bind_addr() failed\n");
3044 : }
3045 0 : rdma_destroy_id(port->id);
3046 0 : free(port);
3047 0 : return rc;
3048 : }
3049 :
3050 0 : if (!port->id->verbs) {
3051 0 : SPDK_ERRLOG("ibv_context is null\n");
3052 0 : rdma_destroy_id(port->id);
3053 0 : free(port);
3054 0 : return -1;
3055 : }
3056 :
3057 0 : rc = rdma_listen(port->id, rtransport->rdma_opts.acceptor_backlog);
3058 0 : if (rc < 0) {
3059 0 : SPDK_ERRLOG("rdma_listen() failed\n");
3060 0 : rdma_destroy_id(port->id);
3061 0 : free(port);
3062 0 : return rc;
3063 : }
3064 :
3065 0 : TAILQ_FOREACH(device, &rtransport->devices, link) {
3066 0 : if (device->context == port->id->verbs && device->is_ready) {
3067 0 : port->device = device;
3068 0 : break;
3069 : }
3070 : }
3071 0 : if (!port->device) {
3072 0 : SPDK_ERRLOG("Accepted a connection with verbs %p, but unable to find a corresponding device.\n",
3073 : port->id->verbs);
3074 0 : rdma_destroy_id(port->id);
3075 0 : free(port);
3076 0 : nvmf_rdma_rescan_devices(rtransport);
3077 0 : return -EINVAL;
3078 : }
3079 :
3080 0 : SPDK_NOTICELOG("*** NVMe/RDMA Target Listening on %s port %s ***\n",
3081 : trid->traddr, trid->trsvcid);
3082 :
3083 0 : TAILQ_INSERT_TAIL(&rtransport->ports, port, link);
3084 0 : return 0;
3085 : }
3086 :
3087 : static void
3088 0 : nvmf_rdma_stop_listen_ex(struct spdk_nvmf_transport *transport,
3089 : const struct spdk_nvme_transport_id *trid, bool need_retry)
3090 : {
3091 : struct spdk_nvmf_rdma_transport *rtransport;
3092 : struct spdk_nvmf_rdma_port *port, *tmp;
3093 :
3094 0 : rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport);
3095 :
3096 0 : if (!need_retry) {
3097 0 : TAILQ_FOREACH_SAFE(port, &rtransport->retry_ports, link, tmp) {
3098 0 : if (spdk_nvme_transport_id_compare(port->trid, trid) == 0) {
3099 0 : TAILQ_REMOVE(&rtransport->retry_ports, port, link);
3100 0 : free(port);
3101 : }
3102 : }
3103 : }
3104 :
3105 0 : TAILQ_FOREACH_SAFE(port, &rtransport->ports, link, tmp) {
3106 0 : if (spdk_nvme_transport_id_compare(port->trid, trid) == 0) {
3107 0 : SPDK_DEBUGLOG(rdma, "Port %s:%s removed. need retry: %d\n",
3108 : port->trid->traddr, port->trid->trsvcid, need_retry);
3109 0 : TAILQ_REMOVE(&rtransport->ports, port, link);
3110 0 : rdma_destroy_id(port->id);
3111 0 : port->id = NULL;
3112 0 : port->device = NULL;
3113 0 : if (need_retry) {
3114 0 : TAILQ_INSERT_TAIL(&rtransport->retry_ports, port, link);
3115 : } else {
3116 0 : free(port);
3117 : }
3118 0 : break;
3119 : }
3120 : }
3121 0 : }
3122 :
3123 : static void
3124 0 : nvmf_rdma_stop_listen(struct spdk_nvmf_transport *transport,
3125 : const struct spdk_nvme_transport_id *trid)
3126 : {
3127 0 : nvmf_rdma_stop_listen_ex(transport, trid, false);
3128 0 : }
3129 :
3130 : static void _nvmf_rdma_register_poller_in_group(void *c);
3131 : static void _nvmf_rdma_remove_poller_in_group(void *c);
3132 :
3133 : static bool
3134 0 : nvmf_rdma_all_pollers_management_done(void *c)
3135 : {
3136 0 : struct poller_manage_ctx *ctx = c;
3137 : int counter;
3138 :
3139 0 : counter = __atomic_sub_fetch(ctx->inflight_op_counter, 1, __ATOMIC_SEQ_CST);
3140 0 : SPDK_DEBUGLOG(rdma, "nvmf_rdma_all_pollers_management_done called. counter: %d, poller: %p\n",
3141 : counter, ctx->rpoller);
3142 :
3143 0 : if (counter == 0) {
3144 0 : free((void *)ctx->inflight_op_counter);
3145 : }
3146 0 : free(ctx);
3147 :
3148 0 : return counter == 0;
3149 : }
3150 :
3151 : static int
3152 0 : nvmf_rdma_manage_poller(struct spdk_nvmf_rdma_transport *rtransport,
3153 : struct spdk_nvmf_rdma_device *device, bool *has_inflight, bool is_add)
3154 : {
3155 : struct spdk_nvmf_rdma_poll_group *rgroup;
3156 : struct spdk_nvmf_rdma_poller *rpoller;
3157 : struct spdk_nvmf_poll_group *poll_group;
3158 : struct poller_manage_ctx *ctx;
3159 : bool found;
3160 : int *inflight_counter;
3161 : spdk_msg_fn do_fn;
3162 :
3163 0 : *has_inflight = false;
3164 0 : do_fn = is_add ? _nvmf_rdma_register_poller_in_group : _nvmf_rdma_remove_poller_in_group;
3165 0 : inflight_counter = calloc(1, sizeof(int));
3166 0 : if (!inflight_counter) {
3167 0 : SPDK_ERRLOG("Failed to allocate inflight counter when removing pollers\n");
3168 0 : return -ENOMEM;
3169 : }
3170 :
3171 0 : TAILQ_FOREACH(rgroup, &rtransport->poll_groups, link) {
3172 0 : (*inflight_counter)++;
3173 : }
3174 :
3175 0 : TAILQ_FOREACH(rgroup, &rtransport->poll_groups, link) {
3176 0 : found = false;
3177 0 : TAILQ_FOREACH(rpoller, &rgroup->pollers, link) {
3178 0 : if (rpoller->device == device) {
3179 0 : found = true;
3180 0 : break;
3181 : }
3182 : }
3183 0 : if (found == is_add) {
3184 0 : __atomic_fetch_sub(inflight_counter, 1, __ATOMIC_SEQ_CST);
3185 0 : continue;
3186 : }
3187 :
3188 0 : ctx = calloc(1, sizeof(struct poller_manage_ctx));
3189 0 : if (!ctx) {
3190 0 : SPDK_ERRLOG("Failed to allocate poller_manage_ctx when removing pollers\n");
3191 0 : if (!*has_inflight) {
3192 0 : free(inflight_counter);
3193 : }
3194 0 : return -ENOMEM;
3195 : }
3196 :
3197 0 : ctx->rtransport = rtransport;
3198 0 : ctx->rgroup = rgroup;
3199 0 : ctx->rpoller = rpoller;
3200 0 : ctx->device = device;
3201 0 : ctx->thread = spdk_get_thread();
3202 0 : ctx->inflight_op_counter = inflight_counter;
3203 0 : *has_inflight = true;
3204 :
3205 0 : poll_group = rgroup->group.group;
3206 0 : if (poll_group->thread != spdk_get_thread()) {
3207 0 : spdk_thread_send_msg(poll_group->thread, do_fn, ctx);
3208 : } else {
3209 0 : do_fn(ctx);
3210 : }
3211 : }
3212 :
3213 0 : if (!*has_inflight) {
3214 0 : free(inflight_counter);
3215 : }
3216 :
3217 0 : return 0;
3218 : }
3219 :
3220 : static void nvmf_rdma_handle_device_removal(struct spdk_nvmf_rdma_transport *rtransport,
3221 : struct spdk_nvmf_rdma_device *device);
3222 :
3223 : static struct spdk_nvmf_rdma_device *
3224 0 : nvmf_rdma_find_ib_device(struct spdk_nvmf_rdma_transport *rtransport,
3225 : struct ibv_context *context)
3226 : {
3227 : struct spdk_nvmf_rdma_device *device, *tmp_device;
3228 :
3229 0 : TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, tmp_device) {
3230 0 : if (device->need_destroy) {
3231 0 : continue;
3232 : }
3233 :
3234 0 : if (strcmp(device->context->device->dev_name, context->device->dev_name) == 0) {
3235 0 : return device;
3236 : }
3237 : }
3238 :
3239 0 : return NULL;
3240 : }
3241 :
3242 : static bool
3243 0 : nvmf_rdma_check_devices_context(struct spdk_nvmf_rdma_transport *rtransport,
3244 : struct ibv_context *context)
3245 : {
3246 0 : struct spdk_nvmf_rdma_device *old_device, *new_device;
3247 0 : int rc = 0;
3248 0 : bool has_inflight;
3249 :
3250 0 : old_device = nvmf_rdma_find_ib_device(rtransport, context);
3251 :
3252 0 : if (old_device) {
3253 0 : if (old_device->context != context && !old_device->need_destroy && old_device->is_ready) {
3254 : /* context may not have time to be cleaned when rescan. exactly one context
3255 : * is valid for a device so this context must be invalid and just remove it. */
3256 0 : SPDK_WARNLOG("Device %p has a invalid context %p\n", old_device, old_device->context);
3257 0 : old_device->need_destroy = true;
3258 0 : nvmf_rdma_handle_device_removal(rtransport, old_device);
3259 : }
3260 0 : return false;
3261 : }
3262 :
3263 0 : rc = create_ib_device(rtransport, context, &new_device);
3264 : /* TODO: update transport opts. */
3265 0 : if (rc < 0) {
3266 0 : SPDK_ERRLOG("Failed to create ib device for context: %s(%p)\n",
3267 : ibv_get_device_name(context->device), context);
3268 0 : return false;
3269 : }
3270 :
3271 0 : rc = nvmf_rdma_manage_poller(rtransport, new_device, &has_inflight, true);
3272 0 : if (rc < 0) {
3273 0 : SPDK_ERRLOG("Failed to add poller for device context: %s(%p)\n",
3274 : ibv_get_device_name(context->device), context);
3275 0 : return false;
3276 : }
3277 :
3278 0 : if (has_inflight) {
3279 0 : new_device->is_ready = true;
3280 : }
3281 :
3282 0 : return true;
3283 : }
3284 :
3285 : static bool
3286 0 : nvmf_rdma_rescan_devices(struct spdk_nvmf_rdma_transport *rtransport)
3287 : {
3288 : struct spdk_nvmf_rdma_device *device;
3289 0 : struct ibv_device **ibv_device_list = NULL;
3290 0 : struct ibv_context **contexts = NULL;
3291 0 : int i = 0;
3292 0 : int num_dev = 0;
3293 0 : bool new_create = false, has_new_device = false;
3294 0 : struct ibv_context *tmp_verbs = NULL;
3295 :
3296 : /* do not rescan when any device is destroying, or context may be freed when
3297 : * regenerating the poll fds.
3298 : */
3299 0 : TAILQ_FOREACH(device, &rtransport->devices, link) {
3300 0 : if (device->need_destroy) {
3301 0 : return false;
3302 : }
3303 : }
3304 :
3305 0 : ibv_device_list = ibv_get_device_list(&num_dev);
3306 :
3307 : /* There is a bug in librdmacm. If verbs init failed in rdma_get_devices, it'll be
3308 : * marked as dead verbs and never be init again. So we need to make sure the
3309 : * verbs is available before we call rdma_get_devices. */
3310 0 : if (num_dev >= 0) {
3311 0 : for (i = 0; i < num_dev; i++) {
3312 0 : tmp_verbs = ibv_open_device(ibv_device_list[i]);
3313 0 : if (!tmp_verbs) {
3314 0 : SPDK_WARNLOG("Failed to init ibv device %p, err %d. Skip rescan.\n", ibv_device_list[i], errno);
3315 0 : break;
3316 : }
3317 0 : if (nvmf_rdma_find_ib_device(rtransport, tmp_verbs) == NULL) {
3318 0 : SPDK_DEBUGLOG(rdma, "Find new verbs init ibv device %p(%s).\n", ibv_device_list[i],
3319 : tmp_verbs->device->dev_name);
3320 0 : has_new_device = true;
3321 : }
3322 0 : ibv_close_device(tmp_verbs);
3323 : }
3324 0 : ibv_free_device_list(ibv_device_list);
3325 0 : if (!tmp_verbs || !has_new_device) {
3326 0 : return false;
3327 : }
3328 : }
3329 :
3330 0 : contexts = rdma_get_devices(NULL);
3331 :
3332 0 : for (i = 0; contexts && contexts[i] != NULL; i++) {
3333 0 : new_create |= nvmf_rdma_check_devices_context(rtransport, contexts[i]);
3334 : }
3335 :
3336 0 : if (new_create) {
3337 0 : free_poll_fds(rtransport);
3338 0 : generate_poll_fds(rtransport);
3339 : }
3340 :
3341 0 : if (contexts) {
3342 0 : rdma_free_devices(contexts);
3343 : }
3344 :
3345 0 : return new_create;
3346 : }
3347 :
3348 : static bool
3349 0 : nvmf_rdma_retry_listen_port(struct spdk_nvmf_rdma_transport *rtransport)
3350 : {
3351 : struct spdk_nvmf_rdma_port *port, *tmp_port;
3352 0 : int rc = 0;
3353 0 : bool new_create = false;
3354 :
3355 0 : if (TAILQ_EMPTY(&rtransport->retry_ports)) {
3356 0 : return false;
3357 : }
3358 :
3359 0 : new_create = nvmf_rdma_rescan_devices(rtransport);
3360 :
3361 0 : TAILQ_FOREACH_SAFE(port, &rtransport->retry_ports, link, tmp_port) {
3362 0 : rc = nvmf_rdma_listen(&rtransport->transport, port->trid, NULL);
3363 :
3364 0 : TAILQ_REMOVE(&rtransport->retry_ports, port, link);
3365 0 : if (rc) {
3366 0 : if (new_create) {
3367 0 : SPDK_ERRLOG("Found new IB device but port %s:%s is still failed(%d) to listen.\n",
3368 : port->trid->traddr, port->trid->trsvcid, rc);
3369 : }
3370 0 : TAILQ_INSERT_TAIL(&rtransport->retry_ports, port, link);
3371 0 : break;
3372 : } else {
3373 0 : SPDK_NOTICELOG("Port %s:%s come back\n", port->trid->traddr, port->trid->trsvcid);
3374 0 : free(port);
3375 : }
3376 : }
3377 :
3378 0 : return true;
3379 : }
3380 :
3381 : static void
3382 0 : nvmf_rdma_qpair_process_pending(struct spdk_nvmf_rdma_transport *rtransport,
3383 : struct spdk_nvmf_rdma_qpair *rqpair, bool drain)
3384 : {
3385 : struct spdk_nvmf_request *req, *tmp;
3386 : struct spdk_nvmf_rdma_request *rdma_req, *req_tmp;
3387 : struct spdk_nvmf_rdma_resources *resources;
3388 :
3389 : /* First process requests which are waiting for response to be sent */
3390 0 : STAILQ_FOREACH_SAFE(rdma_req, &rqpair->pending_rdma_send_queue, state_link, req_tmp) {
3391 0 : if (nvmf_rdma_request_process(rtransport, rdma_req) == false && drain == false) {
3392 0 : break;
3393 : }
3394 : }
3395 :
3396 : /* We process I/O in the data transfer pending queue at the highest priority. */
3397 0 : STAILQ_FOREACH_SAFE(rdma_req, &rqpair->pending_rdma_read_queue, state_link, req_tmp) {
3398 0 : if (nvmf_rdma_request_process(rtransport, rdma_req) == false && drain == false) {
3399 0 : break;
3400 : }
3401 : }
3402 :
3403 : /* Then RDMA writes since reads have stronger restrictions than writes */
3404 0 : STAILQ_FOREACH_SAFE(rdma_req, &rqpair->pending_rdma_write_queue, state_link, req_tmp) {
3405 0 : if (nvmf_rdma_request_process(rtransport, rdma_req) == false && drain == false) {
3406 0 : break;
3407 : }
3408 : }
3409 :
3410 : /* Then we handle request waiting on memory buffers. */
3411 0 : STAILQ_FOREACH_SAFE(req, &rqpair->poller->group->group.pending_buf_queue, buf_link, tmp) {
3412 0 : rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req);
3413 0 : if (nvmf_rdma_request_process(rtransport, rdma_req) == false && drain == false) {
3414 0 : break;
3415 : }
3416 : }
3417 :
3418 0 : resources = rqpair->resources;
3419 0 : while (!STAILQ_EMPTY(&resources->free_queue) && !STAILQ_EMPTY(&resources->incoming_queue)) {
3420 0 : rdma_req = STAILQ_FIRST(&resources->free_queue);
3421 0 : STAILQ_REMOVE_HEAD(&resources->free_queue, state_link);
3422 0 : rdma_req->recv = STAILQ_FIRST(&resources->incoming_queue);
3423 0 : STAILQ_REMOVE_HEAD(&resources->incoming_queue, link);
3424 :
3425 0 : if (rqpair->srq != NULL) {
3426 0 : rdma_req->req.qpair = &rdma_req->recv->qpair->qpair;
3427 0 : rdma_req->recv->qpair->qd++;
3428 : } else {
3429 0 : rqpair->qd++;
3430 : }
3431 :
3432 0 : rdma_req->receive_tsc = rdma_req->recv->receive_tsc;
3433 0 : rdma_req->state = RDMA_REQUEST_STATE_NEW;
3434 0 : if (nvmf_rdma_request_process(rtransport, rdma_req) == false) {
3435 0 : break;
3436 : }
3437 : }
3438 0 : if (!STAILQ_EMPTY(&resources->incoming_queue) && STAILQ_EMPTY(&resources->free_queue)) {
3439 0 : rqpair->poller->stat.pending_free_request++;
3440 : }
3441 0 : }
3442 :
3443 : static inline bool
3444 0 : nvmf_rdma_can_ignore_last_wqe_reached(struct spdk_nvmf_rdma_device *device)
3445 : {
3446 : /* iWARP transport and SoftRoCE driver don't support LAST_WQE_REACHED ibv async event */
3447 0 : return nvmf_rdma_is_rxe_device(device) ||
3448 0 : device->context->device->transport_type == IBV_TRANSPORT_IWARP;
3449 : }
3450 :
3451 : static void
3452 0 : nvmf_rdma_destroy_drained_qpair(struct spdk_nvmf_rdma_qpair *rqpair)
3453 : {
3454 0 : struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(rqpair->qpair.transport,
3455 : struct spdk_nvmf_rdma_transport, transport);
3456 :
3457 0 : nvmf_rdma_qpair_process_pending(rtransport, rqpair, true);
3458 :
3459 : /* nvmf_rdma_close_qpair is not called */
3460 0 : if (!rqpair->to_close) {
3461 0 : return;
3462 : }
3463 :
3464 : /* device is already destroyed and we should force destroy this qpair. */
3465 0 : if (rqpair->poller && rqpair->poller->need_destroy) {
3466 0 : nvmf_rdma_qpair_destroy(rqpair);
3467 0 : return;
3468 : }
3469 :
3470 : /* In non SRQ path, we will reach rqpair->max_queue_depth. In SRQ path, we will get the last_wqe event. */
3471 0 : if (rqpair->current_send_depth != 0) {
3472 0 : return;
3473 : }
3474 :
3475 0 : if (rqpair->srq == NULL && rqpair->current_recv_depth != rqpair->max_queue_depth) {
3476 0 : return;
3477 : }
3478 :
3479 0 : if (rqpair->srq != NULL && rqpair->last_wqe_reached == false &&
3480 0 : !nvmf_rdma_can_ignore_last_wqe_reached(rqpair->device)) {
3481 0 : return;
3482 : }
3483 :
3484 0 : assert(rqpair->qpair.state == SPDK_NVMF_QPAIR_ERROR);
3485 :
3486 0 : nvmf_rdma_qpair_destroy(rqpair);
3487 : }
3488 :
3489 : static int
3490 0 : nvmf_rdma_disconnect(struct rdma_cm_event *evt, bool *event_acked)
3491 : {
3492 : struct spdk_nvmf_qpair *qpair;
3493 : struct spdk_nvmf_rdma_qpair *rqpair;
3494 :
3495 0 : if (evt->id == NULL) {
3496 0 : SPDK_ERRLOG("disconnect request: missing cm_id\n");
3497 0 : return -1;
3498 : }
3499 :
3500 0 : qpair = evt->id->context;
3501 0 : if (qpair == NULL) {
3502 0 : SPDK_ERRLOG("disconnect request: no active connection\n");
3503 0 : return -1;
3504 : }
3505 :
3506 0 : rdma_ack_cm_event(evt);
3507 0 : *event_acked = true;
3508 :
3509 0 : rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair);
3510 :
3511 0 : spdk_trace_record(TRACE_RDMA_QP_DISCONNECT, 0, 0, (uintptr_t)rqpair);
3512 :
3513 0 : spdk_nvmf_qpair_disconnect(&rqpair->qpair, NULL, NULL);
3514 :
3515 0 : return 0;
3516 : }
3517 :
3518 : #ifdef DEBUG
3519 : static const char *CM_EVENT_STR[] = {
3520 : "RDMA_CM_EVENT_ADDR_RESOLVED",
3521 : "RDMA_CM_EVENT_ADDR_ERROR",
3522 : "RDMA_CM_EVENT_ROUTE_RESOLVED",
3523 : "RDMA_CM_EVENT_ROUTE_ERROR",
3524 : "RDMA_CM_EVENT_CONNECT_REQUEST",
3525 : "RDMA_CM_EVENT_CONNECT_RESPONSE",
3526 : "RDMA_CM_EVENT_CONNECT_ERROR",
3527 : "RDMA_CM_EVENT_UNREACHABLE",
3528 : "RDMA_CM_EVENT_REJECTED",
3529 : "RDMA_CM_EVENT_ESTABLISHED",
3530 : "RDMA_CM_EVENT_DISCONNECTED",
3531 : "RDMA_CM_EVENT_DEVICE_REMOVAL",
3532 : "RDMA_CM_EVENT_MULTICAST_JOIN",
3533 : "RDMA_CM_EVENT_MULTICAST_ERROR",
3534 : "RDMA_CM_EVENT_ADDR_CHANGE",
3535 : "RDMA_CM_EVENT_TIMEWAIT_EXIT"
3536 : };
3537 : #endif /* DEBUG */
3538 :
3539 : static void
3540 0 : nvmf_rdma_disconnect_qpairs_on_port(struct spdk_nvmf_rdma_transport *rtransport,
3541 : struct spdk_nvmf_rdma_port *port)
3542 : {
3543 : struct spdk_nvmf_rdma_poll_group *rgroup;
3544 : struct spdk_nvmf_rdma_poller *rpoller;
3545 : struct spdk_nvmf_rdma_qpair *rqpair;
3546 :
3547 0 : TAILQ_FOREACH(rgroup, &rtransport->poll_groups, link) {
3548 0 : TAILQ_FOREACH(rpoller, &rgroup->pollers, link) {
3549 0 : RB_FOREACH(rqpair, qpairs_tree, &rpoller->qpairs) {
3550 0 : if (rqpair->listen_id == port->id) {
3551 0 : spdk_nvmf_qpair_disconnect(&rqpair->qpair, NULL, NULL);
3552 : }
3553 : }
3554 : }
3555 : }
3556 0 : }
3557 :
3558 : static bool
3559 0 : nvmf_rdma_handle_cm_event_addr_change(struct spdk_nvmf_transport *transport,
3560 : struct rdma_cm_event *event)
3561 : {
3562 : const struct spdk_nvme_transport_id *trid;
3563 : struct spdk_nvmf_rdma_port *port;
3564 : struct spdk_nvmf_rdma_transport *rtransport;
3565 0 : bool event_acked = false;
3566 :
3567 0 : rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport);
3568 0 : TAILQ_FOREACH(port, &rtransport->ports, link) {
3569 0 : if (port->id == event->id) {
3570 0 : SPDK_ERRLOG("ADDR_CHANGE: IP %s:%s migrated\n", port->trid->traddr, port->trid->trsvcid);
3571 0 : rdma_ack_cm_event(event);
3572 0 : event_acked = true;
3573 0 : trid = port->trid;
3574 0 : break;
3575 : }
3576 : }
3577 :
3578 0 : if (event_acked) {
3579 0 : nvmf_rdma_disconnect_qpairs_on_port(rtransport, port);
3580 :
3581 0 : nvmf_rdma_stop_listen(transport, trid);
3582 0 : nvmf_rdma_listen(transport, trid, NULL);
3583 : }
3584 :
3585 0 : return event_acked;
3586 : }
3587 :
3588 : static void
3589 0 : nvmf_rdma_handle_device_removal(struct spdk_nvmf_rdma_transport *rtransport,
3590 : struct spdk_nvmf_rdma_device *device)
3591 : {
3592 : struct spdk_nvmf_rdma_port *port, *port_tmp;
3593 : int rc;
3594 0 : bool has_inflight;
3595 :
3596 0 : rc = nvmf_rdma_manage_poller(rtransport, device, &has_inflight, false);
3597 0 : if (rc) {
3598 0 : SPDK_ERRLOG("Failed to handle device removal, rc %d\n", rc);
3599 0 : return;
3600 : }
3601 :
3602 0 : if (!has_inflight) {
3603 : /* no pollers, destroy the device */
3604 0 : device->ready_to_destroy = true;
3605 0 : spdk_thread_send_msg(spdk_get_thread(), _nvmf_rdma_remove_destroyed_device, rtransport);
3606 : }
3607 :
3608 0 : TAILQ_FOREACH_SAFE(port, &rtransport->ports, link, port_tmp) {
3609 0 : if (port->device == device) {
3610 0 : SPDK_NOTICELOG("Port %s:%s on device %s is being removed.\n",
3611 : port->trid->traddr,
3612 : port->trid->trsvcid,
3613 : ibv_get_device_name(port->device->context->device));
3614 :
3615 : /* keep NVMF listener and only destroy structures of the
3616 : * RDMA transport. when the device comes back we can retry listening
3617 : * and the application's workflow will not be interrupted.
3618 : */
3619 0 : nvmf_rdma_stop_listen_ex(&rtransport->transport, port->trid, true);
3620 : }
3621 : }
3622 : }
3623 :
3624 : static void
3625 0 : nvmf_rdma_handle_cm_event_port_removal(struct spdk_nvmf_transport *transport,
3626 : struct rdma_cm_event *event)
3627 : {
3628 : struct spdk_nvmf_rdma_port *port, *tmp_port;
3629 : struct spdk_nvmf_rdma_transport *rtransport;
3630 :
3631 0 : port = event->id->context;
3632 0 : rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport);
3633 :
3634 0 : rdma_ack_cm_event(event);
3635 :
3636 : /* if device removal happens during ctrl qpair disconnecting, it's possible that we receive
3637 : * an DEVICE_REMOVAL event on qpair but the id->qp is just NULL. So we should make sure that
3638 : * we are handling a port event here.
3639 : */
3640 0 : TAILQ_FOREACH(tmp_port, &rtransport->ports, link) {
3641 0 : if (port == tmp_port && port->device && !port->device->need_destroy) {
3642 0 : port->device->need_destroy = true;
3643 0 : nvmf_rdma_handle_device_removal(rtransport, port->device);
3644 : }
3645 : }
3646 0 : }
3647 :
3648 : static void
3649 0 : nvmf_process_cm_event(struct spdk_nvmf_transport *transport)
3650 : {
3651 : struct spdk_nvmf_rdma_transport *rtransport;
3652 0 : struct rdma_cm_event *event;
3653 : int rc;
3654 0 : bool event_acked;
3655 :
3656 0 : rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport);
3657 :
3658 0 : if (rtransport->event_channel == NULL) {
3659 0 : return;
3660 : }
3661 :
3662 : while (1) {
3663 0 : event_acked = false;
3664 0 : rc = rdma_get_cm_event(rtransport->event_channel, &event);
3665 0 : if (rc) {
3666 0 : if (errno != EAGAIN && errno != EWOULDBLOCK) {
3667 0 : SPDK_ERRLOG("Acceptor Event Error: %s\n", spdk_strerror(errno));
3668 : }
3669 0 : break;
3670 : }
3671 :
3672 0 : SPDK_DEBUGLOG(rdma, "Acceptor Event: %s\n", CM_EVENT_STR[event->event]);
3673 :
3674 0 : spdk_trace_record(TRACE_RDMA_CM_ASYNC_EVENT, 0, 0, 0, event->event);
3675 :
3676 0 : switch (event->event) {
3677 0 : case RDMA_CM_EVENT_ADDR_RESOLVED:
3678 : case RDMA_CM_EVENT_ADDR_ERROR:
3679 : case RDMA_CM_EVENT_ROUTE_RESOLVED:
3680 : case RDMA_CM_EVENT_ROUTE_ERROR:
3681 : /* No action required. The target never attempts to resolve routes. */
3682 0 : break;
3683 0 : case RDMA_CM_EVENT_CONNECT_REQUEST:
3684 0 : rc = nvmf_rdma_connect(transport, event);
3685 0 : if (rc < 0) {
3686 0 : SPDK_ERRLOG("Unable to process connect event. rc: %d\n", rc);
3687 0 : break;
3688 : }
3689 0 : break;
3690 0 : case RDMA_CM_EVENT_CONNECT_RESPONSE:
3691 : /* The target never initiates a new connection. So this will not occur. */
3692 0 : break;
3693 0 : case RDMA_CM_EVENT_CONNECT_ERROR:
3694 : /* Can this happen? The docs say it can, but not sure what causes it. */
3695 0 : break;
3696 0 : case RDMA_CM_EVENT_UNREACHABLE:
3697 : case RDMA_CM_EVENT_REJECTED:
3698 : /* These only occur on the client side. */
3699 0 : break;
3700 0 : case RDMA_CM_EVENT_ESTABLISHED:
3701 : /* TODO: Should we be waiting for this event anywhere? */
3702 0 : break;
3703 0 : case RDMA_CM_EVENT_DISCONNECTED:
3704 0 : rc = nvmf_rdma_disconnect(event, &event_acked);
3705 0 : if (rc < 0) {
3706 0 : SPDK_ERRLOG("Unable to process disconnect event. rc: %d\n", rc);
3707 0 : break;
3708 : }
3709 0 : break;
3710 0 : case RDMA_CM_EVENT_DEVICE_REMOVAL:
3711 : /* In case of device removal, kernel IB part triggers IBV_EVENT_DEVICE_FATAL
3712 : * which triggers RDMA_CM_EVENT_DEVICE_REMOVAL on all cma_id’s.
3713 : * Once these events are sent to SPDK, we should release all IB resources and
3714 : * don't make attempts to call any ibv_query/modify/create functions. We can only call
3715 : * ibv_destroy* functions to release user space memory allocated by IB. All kernel
3716 : * resources are already cleaned. */
3717 0 : if (event->id->qp) {
3718 : /* If rdma_cm event has a valid `qp` pointer then the event refers to the
3719 : * corresponding qpair. Otherwise the event refers to a listening device. */
3720 0 : rc = nvmf_rdma_disconnect(event, &event_acked);
3721 0 : if (rc < 0) {
3722 0 : SPDK_ERRLOG("Unable to process disconnect event. rc: %d\n", rc);
3723 0 : break;
3724 : }
3725 : } else {
3726 0 : nvmf_rdma_handle_cm_event_port_removal(transport, event);
3727 0 : event_acked = true;
3728 : }
3729 0 : break;
3730 0 : case RDMA_CM_EVENT_MULTICAST_JOIN:
3731 : case RDMA_CM_EVENT_MULTICAST_ERROR:
3732 : /* Multicast is not used */
3733 0 : break;
3734 0 : case RDMA_CM_EVENT_ADDR_CHANGE:
3735 0 : event_acked = nvmf_rdma_handle_cm_event_addr_change(transport, event);
3736 0 : break;
3737 0 : case RDMA_CM_EVENT_TIMEWAIT_EXIT:
3738 : /* For now, do nothing. The target never re-uses queue pairs. */
3739 0 : break;
3740 0 : default:
3741 0 : SPDK_ERRLOG("Unexpected Acceptor Event [%d]\n", event->event);
3742 0 : break;
3743 : }
3744 0 : if (!event_acked) {
3745 0 : rdma_ack_cm_event(event);
3746 : }
3747 : }
3748 : }
3749 :
3750 : static void
3751 0 : nvmf_rdma_handle_last_wqe_reached(struct spdk_nvmf_rdma_qpair *rqpair)
3752 : {
3753 0 : rqpair->last_wqe_reached = true;
3754 0 : nvmf_rdma_destroy_drained_qpair(rqpair);
3755 0 : }
3756 :
3757 : static void
3758 0 : nvmf_rdma_qpair_process_ibv_event(void *ctx)
3759 : {
3760 0 : struct spdk_nvmf_rdma_ibv_event_ctx *event_ctx = ctx;
3761 :
3762 0 : if (event_ctx->rqpair) {
3763 0 : STAILQ_REMOVE(&event_ctx->rqpair->ibv_events, event_ctx, spdk_nvmf_rdma_ibv_event_ctx, link);
3764 0 : if (event_ctx->cb_fn) {
3765 0 : event_ctx->cb_fn(event_ctx->rqpair);
3766 : }
3767 : }
3768 0 : free(event_ctx);
3769 0 : }
3770 :
3771 : static int
3772 0 : nvmf_rdma_send_qpair_async_event(struct spdk_nvmf_rdma_qpair *rqpair,
3773 : spdk_nvmf_rdma_qpair_ibv_event fn)
3774 : {
3775 : struct spdk_nvmf_rdma_ibv_event_ctx *ctx;
3776 0 : struct spdk_thread *thr = NULL;
3777 : int rc;
3778 :
3779 0 : if (rqpair->qpair.group) {
3780 0 : thr = rqpair->qpair.group->thread;
3781 0 : } else if (rqpair->destruct_channel) {
3782 0 : thr = spdk_io_channel_get_thread(rqpair->destruct_channel);
3783 : }
3784 :
3785 0 : if (!thr) {
3786 0 : SPDK_DEBUGLOG(rdma, "rqpair %p has no thread\n", rqpair);
3787 0 : return -EINVAL;
3788 : }
3789 :
3790 0 : ctx = calloc(1, sizeof(*ctx));
3791 0 : if (!ctx) {
3792 0 : return -ENOMEM;
3793 : }
3794 :
3795 0 : ctx->rqpair = rqpair;
3796 0 : ctx->cb_fn = fn;
3797 0 : STAILQ_INSERT_TAIL(&rqpair->ibv_events, ctx, link);
3798 :
3799 0 : rc = spdk_thread_send_msg(thr, nvmf_rdma_qpair_process_ibv_event, ctx);
3800 0 : if (rc) {
3801 0 : STAILQ_REMOVE(&rqpair->ibv_events, ctx, spdk_nvmf_rdma_ibv_event_ctx, link);
3802 0 : free(ctx);
3803 : }
3804 :
3805 0 : return rc;
3806 : }
3807 :
3808 : static int
3809 0 : nvmf_process_ib_event(struct spdk_nvmf_rdma_device *device)
3810 : {
3811 : int rc;
3812 0 : struct spdk_nvmf_rdma_qpair *rqpair = NULL;
3813 0 : struct ibv_async_event event;
3814 :
3815 0 : rc = ibv_get_async_event(device->context, &event);
3816 :
3817 0 : if (rc) {
3818 : /* In non-blocking mode -1 means there are no events available */
3819 0 : return rc;
3820 : }
3821 :
3822 0 : switch (event.event_type) {
3823 0 : case IBV_EVENT_QP_FATAL:
3824 : case IBV_EVENT_QP_LAST_WQE_REACHED:
3825 : case IBV_EVENT_SQ_DRAINED:
3826 : case IBV_EVENT_QP_REQ_ERR:
3827 : case IBV_EVENT_QP_ACCESS_ERR:
3828 : case IBV_EVENT_COMM_EST:
3829 : case IBV_EVENT_PATH_MIG:
3830 : case IBV_EVENT_PATH_MIG_ERR:
3831 0 : rqpair = event.element.qp->qp_context;
3832 0 : if (!rqpair) {
3833 : /* Any QP event for NVMe-RDMA initiator may be returned. */
3834 0 : SPDK_NOTICELOG("Async QP event for unknown QP: %s\n",
3835 : ibv_event_type_str(event.event_type));
3836 0 : break;
3837 : }
3838 :
3839 0 : switch (event.event_type) {
3840 0 : case IBV_EVENT_QP_FATAL:
3841 0 : SPDK_ERRLOG("Fatal event received for rqpair %p\n", rqpair);
3842 0 : spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0,
3843 : (uintptr_t)rqpair, event.event_type);
3844 0 : nvmf_rdma_update_ibv_state(rqpair);
3845 0 : spdk_nvmf_qpair_disconnect(&rqpair->qpair, NULL, NULL);
3846 0 : break;
3847 0 : case IBV_EVENT_QP_LAST_WQE_REACHED:
3848 : /* This event only occurs for shared receive queues. */
3849 0 : SPDK_DEBUGLOG(rdma, "Last WQE reached event received for rqpair %p\n", rqpair);
3850 0 : rc = nvmf_rdma_send_qpair_async_event(rqpair, nvmf_rdma_handle_last_wqe_reached);
3851 0 : if (rc) {
3852 0 : SPDK_WARNLOG("Failed to send LAST_WQE_REACHED event. rqpair %p, err %d\n", rqpair, rc);
3853 0 : rqpair->last_wqe_reached = true;
3854 : }
3855 0 : break;
3856 0 : case IBV_EVENT_SQ_DRAINED:
3857 : /* This event occurs frequently in both error and non-error states.
3858 : * Check if the qpair is in an error state before sending a message. */
3859 0 : SPDK_DEBUGLOG(rdma, "Last sq drained event received for rqpair %p\n", rqpair);
3860 0 : spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0,
3861 : (uintptr_t)rqpair, event.event_type);
3862 0 : if (nvmf_rdma_update_ibv_state(rqpair) == IBV_QPS_ERR) {
3863 0 : spdk_nvmf_qpair_disconnect(&rqpair->qpair, NULL, NULL);
3864 : }
3865 0 : break;
3866 0 : case IBV_EVENT_QP_REQ_ERR:
3867 : case IBV_EVENT_QP_ACCESS_ERR:
3868 : case IBV_EVENT_COMM_EST:
3869 : case IBV_EVENT_PATH_MIG:
3870 : case IBV_EVENT_PATH_MIG_ERR:
3871 0 : SPDK_NOTICELOG("Async QP event: %s\n",
3872 : ibv_event_type_str(event.event_type));
3873 0 : spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0,
3874 : (uintptr_t)rqpair, event.event_type);
3875 0 : nvmf_rdma_update_ibv_state(rqpair);
3876 0 : break;
3877 0 : default:
3878 0 : break;
3879 : }
3880 0 : break;
3881 0 : case IBV_EVENT_DEVICE_FATAL:
3882 0 : SPDK_ERRLOG("Device Fatal event[%s] received on %s. device: %p\n",
3883 : ibv_event_type_str(event.event_type), ibv_get_device_name(device->context->device), device);
3884 0 : device->need_destroy = true;
3885 0 : break;
3886 0 : case IBV_EVENT_CQ_ERR:
3887 : case IBV_EVENT_PORT_ACTIVE:
3888 : case IBV_EVENT_PORT_ERR:
3889 : case IBV_EVENT_LID_CHANGE:
3890 : case IBV_EVENT_PKEY_CHANGE:
3891 : case IBV_EVENT_SM_CHANGE:
3892 : case IBV_EVENT_SRQ_ERR:
3893 : case IBV_EVENT_SRQ_LIMIT_REACHED:
3894 : case IBV_EVENT_CLIENT_REREGISTER:
3895 : case IBV_EVENT_GID_CHANGE:
3896 : default:
3897 0 : SPDK_NOTICELOG("Async event: %s\n",
3898 : ibv_event_type_str(event.event_type));
3899 0 : spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0, 0, event.event_type);
3900 0 : break;
3901 : }
3902 0 : ibv_ack_async_event(&event);
3903 :
3904 0 : return 0;
3905 : }
3906 :
3907 : static void
3908 0 : nvmf_process_ib_events(struct spdk_nvmf_rdma_device *device, uint32_t max_events)
3909 : {
3910 0 : int rc = 0;
3911 0 : uint32_t i = 0;
3912 :
3913 0 : for (i = 0; i < max_events; i++) {
3914 0 : rc = nvmf_process_ib_event(device);
3915 0 : if (rc) {
3916 0 : break;
3917 : }
3918 : }
3919 :
3920 0 : SPDK_DEBUGLOG(rdma, "Device %s: %u events processed\n", device->context->device->name, i);
3921 0 : }
3922 :
3923 : static int
3924 0 : nvmf_rdma_accept(void *ctx)
3925 : {
3926 0 : int nfds, i = 0;
3927 0 : struct spdk_nvmf_transport *transport = ctx;
3928 : struct spdk_nvmf_rdma_transport *rtransport;
3929 : struct spdk_nvmf_rdma_device *device, *tmp;
3930 : uint32_t count;
3931 : short revents;
3932 : bool do_retry;
3933 :
3934 0 : rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport);
3935 0 : do_retry = nvmf_rdma_retry_listen_port(rtransport);
3936 :
3937 0 : count = nfds = poll(rtransport->poll_fds, rtransport->npoll_fds, 0);
3938 :
3939 0 : if (nfds <= 0) {
3940 0 : return do_retry ? SPDK_POLLER_BUSY : SPDK_POLLER_IDLE;
3941 : }
3942 :
3943 : /* The first poll descriptor is RDMA CM event */
3944 0 : if (rtransport->poll_fds[i++].revents & POLLIN) {
3945 0 : nvmf_process_cm_event(transport);
3946 0 : nfds--;
3947 : }
3948 :
3949 0 : if (nfds == 0) {
3950 0 : return SPDK_POLLER_BUSY;
3951 : }
3952 :
3953 : /* Second and subsequent poll descriptors are IB async events */
3954 0 : TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, tmp) {
3955 0 : revents = rtransport->poll_fds[i++].revents;
3956 0 : if (revents & POLLIN) {
3957 0 : if (spdk_likely(!device->need_destroy)) {
3958 0 : nvmf_process_ib_events(device, 32);
3959 0 : if (spdk_unlikely(device->need_destroy)) {
3960 0 : nvmf_rdma_handle_device_removal(rtransport, device);
3961 : }
3962 : }
3963 0 : nfds--;
3964 0 : } else if (revents & POLLNVAL || revents & POLLHUP) {
3965 0 : SPDK_ERRLOG("Receive unknown revent %x on device %p\n", (int)revents, device);
3966 0 : nfds--;
3967 : }
3968 : }
3969 : /* check all flagged fd's have been served */
3970 0 : assert(nfds == 0);
3971 :
3972 0 : return count > 0 ? SPDK_POLLER_BUSY : SPDK_POLLER_IDLE;
3973 : }
3974 :
3975 : static void
3976 0 : nvmf_rdma_cdata_init(struct spdk_nvmf_transport *transport, struct spdk_nvmf_subsystem *subsystem,
3977 : struct spdk_nvmf_ctrlr_data *cdata)
3978 : {
3979 0 : cdata->nvmf_specific.msdbd = NVMF_DEFAULT_MSDBD;
3980 :
3981 : /* Disable in-capsule data transfer for RDMA controller when dif_insert_or_strip is enabled
3982 : since in-capsule data only works with NVME drives that support SGL memory layout */
3983 0 : if (transport->opts.dif_insert_or_strip) {
3984 0 : cdata->nvmf_specific.ioccsz = sizeof(struct spdk_nvme_cmd) / 16;
3985 : }
3986 :
3987 0 : if (cdata->nvmf_specific.ioccsz > ((sizeof(struct spdk_nvme_cmd) + 0x1000) / 16)) {
3988 0 : SPDK_WARNLOG("RDMA is configured to support up to 16 SGL entries while in capsule"
3989 : " data is greater than 4KiB.\n");
3990 0 : SPDK_WARNLOG("When used in conjunction with the NVMe-oF initiator from the Linux "
3991 : "kernel between versions 5.4 and 5.12 data corruption may occur for "
3992 : "writes that are not a multiple of 4KiB in size.\n");
3993 : }
3994 0 : }
3995 :
3996 : static void
3997 0 : nvmf_rdma_discover(struct spdk_nvmf_transport *transport,
3998 : struct spdk_nvme_transport_id *trid,
3999 : struct spdk_nvmf_discovery_log_page_entry *entry)
4000 : {
4001 0 : entry->trtype = SPDK_NVMF_TRTYPE_RDMA;
4002 0 : entry->adrfam = trid->adrfam;
4003 0 : entry->treq.secure_channel = SPDK_NVMF_TREQ_SECURE_CHANNEL_NOT_REQUIRED;
4004 :
4005 0 : spdk_strcpy_pad(entry->trsvcid, trid->trsvcid, sizeof(entry->trsvcid), ' ');
4006 0 : spdk_strcpy_pad(entry->traddr, trid->traddr, sizeof(entry->traddr), ' ');
4007 :
4008 0 : entry->tsas.rdma.rdma_qptype = SPDK_NVMF_RDMA_QPTYPE_RELIABLE_CONNECTED;
4009 0 : entry->tsas.rdma.rdma_prtype = SPDK_NVMF_RDMA_PRTYPE_NONE;
4010 0 : entry->tsas.rdma.rdma_cms = SPDK_NVMF_RDMA_CMS_RDMA_CM;
4011 0 : }
4012 :
4013 : static int
4014 0 : nvmf_rdma_poller_create(struct spdk_nvmf_rdma_transport *rtransport,
4015 : struct spdk_nvmf_rdma_poll_group *rgroup, struct spdk_nvmf_rdma_device *device,
4016 : struct spdk_nvmf_rdma_poller **out_poller)
4017 : {
4018 : struct spdk_nvmf_rdma_poller *poller;
4019 0 : struct spdk_rdma_srq_init_attr srq_init_attr;
4020 0 : struct spdk_nvmf_rdma_resource_opts opts;
4021 : int num_cqe;
4022 :
4023 0 : poller = calloc(1, sizeof(*poller));
4024 0 : if (!poller) {
4025 0 : SPDK_ERRLOG("Unable to allocate memory for new RDMA poller\n");
4026 0 : return -1;
4027 : }
4028 :
4029 0 : poller->device = device;
4030 0 : poller->group = rgroup;
4031 0 : *out_poller = poller;
4032 :
4033 0 : RB_INIT(&poller->qpairs);
4034 0 : STAILQ_INIT(&poller->qpairs_pending_send);
4035 0 : STAILQ_INIT(&poller->qpairs_pending_recv);
4036 :
4037 0 : TAILQ_INSERT_TAIL(&rgroup->pollers, poller, link);
4038 0 : SPDK_DEBUGLOG(rdma, "Create poller %p on device %p in poll group %p.\n", poller, device, rgroup);
4039 0 : if (rtransport->rdma_opts.no_srq == false && device->num_srq < device->attr.max_srq) {
4040 0 : if ((int)rtransport->rdma_opts.max_srq_depth > device->attr.max_srq_wr) {
4041 0 : SPDK_WARNLOG("Requested SRQ depth %u, max supported by dev %s is %d\n",
4042 : rtransport->rdma_opts.max_srq_depth, device->context->device->name, device->attr.max_srq_wr);
4043 : }
4044 0 : poller->max_srq_depth = spdk_min((int)rtransport->rdma_opts.max_srq_depth, device->attr.max_srq_wr);
4045 :
4046 0 : device->num_srq++;
4047 0 : memset(&srq_init_attr, 0, sizeof(srq_init_attr));
4048 0 : srq_init_attr.pd = device->pd;
4049 0 : srq_init_attr.stats = &poller->stat.qp_stats.recv;
4050 0 : srq_init_attr.srq_init_attr.attr.max_wr = poller->max_srq_depth;
4051 0 : srq_init_attr.srq_init_attr.attr.max_sge = spdk_min(device->attr.max_sge, NVMF_DEFAULT_RX_SGE);
4052 0 : poller->srq = spdk_rdma_srq_create(&srq_init_attr);
4053 0 : if (!poller->srq) {
4054 0 : SPDK_ERRLOG("Unable to create shared receive queue, errno %d\n", errno);
4055 0 : return -1;
4056 : }
4057 :
4058 0 : opts.qp = poller->srq;
4059 0 : opts.map = device->map;
4060 0 : opts.qpair = NULL;
4061 0 : opts.shared = true;
4062 0 : opts.max_queue_depth = poller->max_srq_depth;
4063 0 : opts.in_capsule_data_size = rtransport->transport.opts.in_capsule_data_size;
4064 :
4065 0 : poller->resources = nvmf_rdma_resources_create(&opts);
4066 0 : if (!poller->resources) {
4067 0 : SPDK_ERRLOG("Unable to allocate resources for shared receive queue.\n");
4068 0 : return -1;
4069 : }
4070 : }
4071 :
4072 : /*
4073 : * When using an srq, we can limit the completion queue at startup.
4074 : * The following formula represents the calculation:
4075 : * num_cqe = num_recv + num_data_wr + num_send_wr.
4076 : * where num_recv=num_data_wr=and num_send_wr=poller->max_srq_depth
4077 : */
4078 0 : if (poller->srq) {
4079 0 : num_cqe = poller->max_srq_depth * 3;
4080 : } else {
4081 0 : num_cqe = rtransport->rdma_opts.num_cqe;
4082 : }
4083 :
4084 0 : poller->cq = ibv_create_cq(device->context, num_cqe, poller, NULL, 0);
4085 0 : if (!poller->cq) {
4086 0 : SPDK_ERRLOG("Unable to create completion queue\n");
4087 0 : return -1;
4088 : }
4089 0 : poller->num_cqe = num_cqe;
4090 0 : return 0;
4091 : }
4092 :
4093 : static void
4094 0 : _nvmf_rdma_register_poller_in_group(void *c)
4095 : {
4096 0 : struct spdk_nvmf_rdma_poller *poller;
4097 0 : struct poller_manage_ctx *ctx = c;
4098 : struct spdk_nvmf_rdma_device *device;
4099 : int rc;
4100 :
4101 0 : rc = nvmf_rdma_poller_create(ctx->rtransport, ctx->rgroup, ctx->device, &poller);
4102 0 : if (rc < 0 && poller) {
4103 0 : nvmf_rdma_poller_destroy(poller);
4104 : }
4105 :
4106 0 : device = ctx->device;
4107 0 : if (nvmf_rdma_all_pollers_management_done(ctx)) {
4108 0 : device->is_ready = true;
4109 : }
4110 0 : }
4111 :
4112 : static void nvmf_rdma_poll_group_destroy(struct spdk_nvmf_transport_poll_group *group);
4113 :
4114 : static struct spdk_nvmf_transport_poll_group *
4115 5 : nvmf_rdma_poll_group_create(struct spdk_nvmf_transport *transport,
4116 : struct spdk_nvmf_poll_group *group)
4117 : {
4118 : struct spdk_nvmf_rdma_transport *rtransport;
4119 : struct spdk_nvmf_rdma_poll_group *rgroup;
4120 5 : struct spdk_nvmf_rdma_poller *poller;
4121 : struct spdk_nvmf_rdma_device *device;
4122 : int rc;
4123 :
4124 5 : rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport);
4125 :
4126 5 : rgroup = calloc(1, sizeof(*rgroup));
4127 5 : if (!rgroup) {
4128 0 : return NULL;
4129 : }
4130 :
4131 5 : TAILQ_INIT(&rgroup->pollers);
4132 :
4133 5 : TAILQ_FOREACH(device, &rtransport->devices, link) {
4134 0 : rc = nvmf_rdma_poller_create(rtransport, rgroup, device, &poller);
4135 0 : if (rc < 0) {
4136 0 : nvmf_rdma_poll_group_destroy(&rgroup->group);
4137 0 : return NULL;
4138 : }
4139 : }
4140 :
4141 5 : TAILQ_INSERT_TAIL(&rtransport->poll_groups, rgroup, link);
4142 5 : if (rtransport->conn_sched.next_admin_pg == NULL) {
4143 1 : rtransport->conn_sched.next_admin_pg = rgroup;
4144 1 : rtransport->conn_sched.next_io_pg = rgroup;
4145 : }
4146 :
4147 5 : return &rgroup->group;
4148 : }
4149 :
4150 : static uint32_t
4151 12 : nvmf_poll_group_get_io_qpair_count(struct spdk_nvmf_poll_group *pg)
4152 : {
4153 : uint32_t count;
4154 :
4155 : /* Just assume that unassociated qpairs will eventually be io
4156 : * qpairs. This is close enough for the use cases for this
4157 : * function.
4158 : */
4159 12 : pthread_mutex_lock(&pg->mutex);
4160 12 : count = pg->stat.current_io_qpairs + pg->current_unassociated_qpairs;
4161 12 : pthread_mutex_unlock(&pg->mutex);
4162 :
4163 12 : return count;
4164 : }
4165 :
4166 : static struct spdk_nvmf_transport_poll_group *
4167 14 : nvmf_rdma_get_optimal_poll_group(struct spdk_nvmf_qpair *qpair)
4168 : {
4169 : struct spdk_nvmf_rdma_transport *rtransport;
4170 : struct spdk_nvmf_rdma_poll_group **pg;
4171 : struct spdk_nvmf_transport_poll_group *result;
4172 : uint32_t count;
4173 :
4174 14 : rtransport = SPDK_CONTAINEROF(qpair->transport, struct spdk_nvmf_rdma_transport, transport);
4175 :
4176 14 : if (TAILQ_EMPTY(&rtransport->poll_groups)) {
4177 2 : return NULL;
4178 : }
4179 :
4180 12 : if (qpair->qid == 0) {
4181 6 : pg = &rtransport->conn_sched.next_admin_pg;
4182 : } else {
4183 : struct spdk_nvmf_rdma_poll_group *pg_min, *pg_start, *pg_current;
4184 : uint32_t min_value;
4185 :
4186 6 : pg = &rtransport->conn_sched.next_io_pg;
4187 6 : pg_min = *pg;
4188 6 : pg_start = *pg;
4189 6 : pg_current = *pg;
4190 6 : min_value = nvmf_poll_group_get_io_qpair_count(pg_current->group.group);
4191 :
4192 : while (1) {
4193 6 : count = nvmf_poll_group_get_io_qpair_count(pg_current->group.group);
4194 :
4195 6 : if (count < min_value) {
4196 0 : min_value = count;
4197 0 : pg_min = pg_current;
4198 : }
4199 :
4200 6 : pg_current = TAILQ_NEXT(pg_current, link);
4201 6 : if (pg_current == NULL) {
4202 2 : pg_current = TAILQ_FIRST(&rtransport->poll_groups);
4203 : }
4204 :
4205 6 : if (pg_current == pg_start || min_value == 0) {
4206 : break;
4207 : }
4208 : }
4209 6 : *pg = pg_min;
4210 : }
4211 :
4212 12 : assert(*pg != NULL);
4213 :
4214 12 : result = &(*pg)->group;
4215 :
4216 12 : *pg = TAILQ_NEXT(*pg, link);
4217 12 : if (*pg == NULL) {
4218 4 : *pg = TAILQ_FIRST(&rtransport->poll_groups);
4219 : }
4220 :
4221 12 : return result;
4222 : }
4223 :
4224 : static void
4225 0 : nvmf_rdma_poller_destroy(struct spdk_nvmf_rdma_poller *poller)
4226 : {
4227 : struct spdk_nvmf_rdma_qpair *qpair, *tmp_qpair;
4228 : int rc;
4229 :
4230 0 : TAILQ_REMOVE(&poller->group->pollers, poller, link);
4231 0 : RB_FOREACH_SAFE(qpair, qpairs_tree, &poller->qpairs, tmp_qpair) {
4232 0 : nvmf_rdma_qpair_destroy(qpair);
4233 : }
4234 :
4235 0 : if (poller->srq) {
4236 0 : if (poller->resources) {
4237 0 : nvmf_rdma_resources_destroy(poller->resources);
4238 : }
4239 0 : spdk_rdma_srq_destroy(poller->srq);
4240 0 : SPDK_DEBUGLOG(rdma, "Destroyed RDMA shared queue %p\n", poller->srq);
4241 : }
4242 :
4243 0 : if (poller->cq) {
4244 0 : rc = ibv_destroy_cq(poller->cq);
4245 0 : if (rc != 0) {
4246 0 : SPDK_ERRLOG("Destroy cq return %d, error: %s\n", rc, strerror(errno));
4247 : }
4248 : }
4249 :
4250 0 : if (poller->destroy_cb) {
4251 0 : poller->destroy_cb(poller->destroy_cb_ctx);
4252 0 : poller->destroy_cb = NULL;
4253 : }
4254 :
4255 0 : free(poller);
4256 0 : }
4257 :
4258 : static void
4259 5 : nvmf_rdma_poll_group_destroy(struct spdk_nvmf_transport_poll_group *group)
4260 : {
4261 : struct spdk_nvmf_rdma_poll_group *rgroup, *next_rgroup;
4262 : struct spdk_nvmf_rdma_poller *poller, *tmp;
4263 : struct spdk_nvmf_rdma_transport *rtransport;
4264 :
4265 5 : rgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_rdma_poll_group, group);
4266 5 : if (!rgroup) {
4267 0 : return;
4268 : }
4269 :
4270 5 : TAILQ_FOREACH_SAFE(poller, &rgroup->pollers, link, tmp) {
4271 0 : nvmf_rdma_poller_destroy(poller);
4272 : }
4273 :
4274 5 : if (rgroup->group.transport == NULL) {
4275 : /* Transport can be NULL when nvmf_rdma_poll_group_create()
4276 : * calls this function directly in a failure path. */
4277 0 : free(rgroup);
4278 0 : return;
4279 : }
4280 :
4281 5 : rtransport = SPDK_CONTAINEROF(rgroup->group.transport, struct spdk_nvmf_rdma_transport, transport);
4282 :
4283 5 : next_rgroup = TAILQ_NEXT(rgroup, link);
4284 5 : TAILQ_REMOVE(&rtransport->poll_groups, rgroup, link);
4285 5 : if (next_rgroup == NULL) {
4286 1 : next_rgroup = TAILQ_FIRST(&rtransport->poll_groups);
4287 : }
4288 5 : if (rtransport->conn_sched.next_admin_pg == rgroup) {
4289 5 : rtransport->conn_sched.next_admin_pg = next_rgroup;
4290 : }
4291 5 : if (rtransport->conn_sched.next_io_pg == rgroup) {
4292 5 : rtransport->conn_sched.next_io_pg = next_rgroup;
4293 : }
4294 :
4295 5 : free(rgroup);
4296 : }
4297 :
4298 : static void
4299 0 : nvmf_rdma_qpair_reject_connection(struct spdk_nvmf_rdma_qpair *rqpair)
4300 : {
4301 0 : if (rqpair->cm_id != NULL) {
4302 0 : nvmf_rdma_event_reject(rqpair->cm_id, SPDK_NVMF_RDMA_ERROR_NO_RESOURCES);
4303 : }
4304 0 : }
4305 :
4306 : static int
4307 0 : nvmf_rdma_poll_group_add(struct spdk_nvmf_transport_poll_group *group,
4308 : struct spdk_nvmf_qpair *qpair)
4309 : {
4310 : struct spdk_nvmf_rdma_poll_group *rgroup;
4311 : struct spdk_nvmf_rdma_qpair *rqpair;
4312 : struct spdk_nvmf_rdma_device *device;
4313 : struct spdk_nvmf_rdma_poller *poller;
4314 : int rc;
4315 :
4316 0 : rgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_rdma_poll_group, group);
4317 0 : rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair);
4318 :
4319 0 : device = rqpair->device;
4320 :
4321 0 : TAILQ_FOREACH(poller, &rgroup->pollers, link) {
4322 0 : if (poller->device == device) {
4323 0 : break;
4324 : }
4325 : }
4326 :
4327 0 : if (!poller) {
4328 0 : SPDK_ERRLOG("No poller found for device.\n");
4329 0 : return -1;
4330 : }
4331 :
4332 0 : if (poller->need_destroy) {
4333 0 : SPDK_ERRLOG("Poller is destroying.\n");
4334 0 : return -1;
4335 : }
4336 :
4337 0 : rqpair->poller = poller;
4338 0 : rqpair->srq = rqpair->poller->srq;
4339 :
4340 0 : rc = nvmf_rdma_qpair_initialize(qpair);
4341 0 : if (rc < 0) {
4342 0 : SPDK_ERRLOG("Failed to initialize nvmf_rdma_qpair with qpair=%p\n", qpair);
4343 0 : rqpair->poller = NULL;
4344 0 : rqpair->srq = NULL;
4345 0 : return -1;
4346 : }
4347 :
4348 0 : RB_INSERT(qpairs_tree, &poller->qpairs, rqpair);
4349 :
4350 0 : rc = nvmf_rdma_event_accept(rqpair->cm_id, rqpair);
4351 0 : if (rc) {
4352 : /* Try to reject, but we probably can't */
4353 0 : nvmf_rdma_qpair_reject_connection(rqpair);
4354 0 : return -1;
4355 : }
4356 :
4357 0 : nvmf_rdma_update_ibv_state(rqpair);
4358 :
4359 0 : return 0;
4360 : }
4361 :
4362 : static int
4363 0 : nvmf_rdma_poll_group_remove(struct spdk_nvmf_transport_poll_group *group,
4364 : struct spdk_nvmf_qpair *qpair)
4365 : {
4366 : struct spdk_nvmf_rdma_qpair *rqpair;
4367 :
4368 0 : rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair);
4369 0 : assert(group->transport->tgt != NULL);
4370 :
4371 0 : rqpair->destruct_channel = spdk_get_io_channel(group->transport->tgt);
4372 :
4373 0 : if (!rqpair->destruct_channel) {
4374 0 : SPDK_WARNLOG("failed to get io_channel, qpair %p\n", qpair);
4375 0 : return 0;
4376 : }
4377 :
4378 : /* Sanity check that we get io_channel on the correct thread */
4379 0 : if (qpair->group) {
4380 0 : assert(qpair->group->thread == spdk_io_channel_get_thread(rqpair->destruct_channel));
4381 : }
4382 :
4383 0 : return 0;
4384 : }
4385 :
4386 : static int
4387 0 : nvmf_rdma_request_free(struct spdk_nvmf_request *req)
4388 : {
4389 0 : struct spdk_nvmf_rdma_request *rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req);
4390 0 : struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(req->qpair->transport,
4391 : struct spdk_nvmf_rdma_transport, transport);
4392 0 : struct spdk_nvmf_rdma_qpair *rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair,
4393 : struct spdk_nvmf_rdma_qpair, qpair);
4394 :
4395 : /*
4396 : * AER requests are freed when a qpair is destroyed. The recv corresponding to that request
4397 : * needs to be returned to the shared receive queue or the poll group will eventually be
4398 : * starved of RECV structures.
4399 : */
4400 0 : if (rqpair->srq && rdma_req->recv) {
4401 : int rc;
4402 0 : struct ibv_recv_wr *bad_recv_wr;
4403 :
4404 0 : spdk_rdma_srq_queue_recv_wrs(rqpair->srq, &rdma_req->recv->wr);
4405 0 : rc = spdk_rdma_srq_flush_recv_wrs(rqpair->srq, &bad_recv_wr);
4406 0 : if (rc) {
4407 0 : SPDK_ERRLOG("Unable to re-post rx descriptor\n");
4408 : }
4409 : }
4410 :
4411 0 : _nvmf_rdma_request_free(rdma_req, rtransport);
4412 0 : return 0;
4413 : }
4414 :
4415 : static int
4416 0 : nvmf_rdma_request_complete(struct spdk_nvmf_request *req)
4417 : {
4418 0 : struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(req->qpair->transport,
4419 : struct spdk_nvmf_rdma_transport, transport);
4420 0 : struct spdk_nvmf_rdma_request *rdma_req = SPDK_CONTAINEROF(req,
4421 : struct spdk_nvmf_rdma_request, req);
4422 0 : struct spdk_nvmf_rdma_qpair *rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair,
4423 : struct spdk_nvmf_rdma_qpair, qpair);
4424 :
4425 0 : if (rqpair->ibv_state != IBV_QPS_ERR) {
4426 : /* The connection is alive, so process the request as normal */
4427 0 : rdma_req->state = RDMA_REQUEST_STATE_EXECUTED;
4428 : } else {
4429 : /* The connection is dead. Move the request directly to the completed state. */
4430 0 : rdma_req->state = RDMA_REQUEST_STATE_COMPLETED;
4431 : }
4432 :
4433 0 : nvmf_rdma_request_process(rtransport, rdma_req);
4434 :
4435 0 : return 0;
4436 : }
4437 :
4438 : static void
4439 0 : nvmf_rdma_close_qpair(struct spdk_nvmf_qpair *qpair,
4440 : spdk_nvmf_transport_qpair_fini_cb cb_fn, void *cb_arg)
4441 : {
4442 0 : struct spdk_nvmf_rdma_qpair *rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair);
4443 :
4444 0 : rqpair->to_close = true;
4445 :
4446 : /* This happens only when the qpair is disconnected before
4447 : * it is added to the poll group. Since there is no poll group,
4448 : * the RDMA qp has not been initialized yet and the RDMA CM
4449 : * event has not yet been acknowledged, so we need to reject it.
4450 : */
4451 0 : if (rqpair->qpair.state == SPDK_NVMF_QPAIR_UNINITIALIZED) {
4452 0 : nvmf_rdma_qpair_reject_connection(rqpair);
4453 0 : nvmf_rdma_qpair_destroy(rqpair);
4454 0 : return;
4455 : }
4456 :
4457 0 : if (rqpair->rdma_qp) {
4458 0 : spdk_rdma_qp_disconnect(rqpair->rdma_qp);
4459 : }
4460 :
4461 0 : nvmf_rdma_destroy_drained_qpair(rqpair);
4462 :
4463 0 : if (cb_fn) {
4464 0 : cb_fn(cb_arg);
4465 : }
4466 : }
4467 :
4468 : static struct spdk_nvmf_rdma_qpair *
4469 0 : get_rdma_qpair_from_wc(struct spdk_nvmf_rdma_poller *rpoller, struct ibv_wc *wc)
4470 : {
4471 0 : struct spdk_nvmf_rdma_qpair find;
4472 :
4473 0 : find.qp_num = wc->qp_num;
4474 :
4475 0 : return RB_FIND(qpairs_tree, &rpoller->qpairs, &find);
4476 : }
4477 :
4478 : #ifdef DEBUG
4479 : static int
4480 0 : nvmf_rdma_req_is_completing(struct spdk_nvmf_rdma_request *rdma_req)
4481 : {
4482 0 : return rdma_req->state == RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST ||
4483 0 : rdma_req->state == RDMA_REQUEST_STATE_COMPLETING;
4484 : }
4485 : #endif
4486 :
4487 : static void
4488 0 : _poller_reset_failed_recvs(struct spdk_nvmf_rdma_poller *rpoller, struct ibv_recv_wr *bad_recv_wr,
4489 : int rc)
4490 : {
4491 : struct spdk_nvmf_rdma_recv *rdma_recv;
4492 : struct spdk_nvmf_rdma_wr *bad_rdma_wr;
4493 :
4494 0 : SPDK_ERRLOG("Failed to post a recv for the poller %p with errno %d\n", rpoller, -rc);
4495 0 : while (bad_recv_wr != NULL) {
4496 0 : bad_rdma_wr = (struct spdk_nvmf_rdma_wr *)bad_recv_wr->wr_id;
4497 0 : rdma_recv = SPDK_CONTAINEROF(bad_rdma_wr, struct spdk_nvmf_rdma_recv, rdma_wr);
4498 :
4499 0 : rdma_recv->qpair->current_recv_depth++;
4500 0 : bad_recv_wr = bad_recv_wr->next;
4501 0 : SPDK_ERRLOG("Failed to post a recv for the qpair %p with errno %d\n", rdma_recv->qpair, -rc);
4502 0 : spdk_nvmf_qpair_disconnect(&rdma_recv->qpair->qpair, NULL, NULL);
4503 : }
4504 0 : }
4505 :
4506 : static void
4507 0 : _qp_reset_failed_recvs(struct spdk_nvmf_rdma_qpair *rqpair, struct ibv_recv_wr *bad_recv_wr, int rc)
4508 : {
4509 0 : SPDK_ERRLOG("Failed to post a recv for the qpair %p with errno %d\n", rqpair, -rc);
4510 0 : while (bad_recv_wr != NULL) {
4511 0 : bad_recv_wr = bad_recv_wr->next;
4512 0 : rqpair->current_recv_depth++;
4513 : }
4514 0 : spdk_nvmf_qpair_disconnect(&rqpair->qpair, NULL, NULL);
4515 0 : }
4516 :
4517 : static void
4518 0 : _poller_submit_recvs(struct spdk_nvmf_rdma_transport *rtransport,
4519 : struct spdk_nvmf_rdma_poller *rpoller)
4520 : {
4521 : struct spdk_nvmf_rdma_qpair *rqpair;
4522 0 : struct ibv_recv_wr *bad_recv_wr;
4523 : int rc;
4524 :
4525 0 : if (rpoller->srq) {
4526 0 : rc = spdk_rdma_srq_flush_recv_wrs(rpoller->srq, &bad_recv_wr);
4527 0 : if (rc) {
4528 0 : _poller_reset_failed_recvs(rpoller, bad_recv_wr, rc);
4529 : }
4530 : } else {
4531 0 : while (!STAILQ_EMPTY(&rpoller->qpairs_pending_recv)) {
4532 0 : rqpair = STAILQ_FIRST(&rpoller->qpairs_pending_recv);
4533 0 : rc = spdk_rdma_qp_flush_recv_wrs(rqpair->rdma_qp, &bad_recv_wr);
4534 0 : if (rc) {
4535 0 : _qp_reset_failed_recvs(rqpair, bad_recv_wr, rc);
4536 : }
4537 0 : STAILQ_REMOVE_HEAD(&rpoller->qpairs_pending_recv, recv_link);
4538 : }
4539 : }
4540 0 : }
4541 :
4542 : static void
4543 0 : _qp_reset_failed_sends(struct spdk_nvmf_rdma_transport *rtransport,
4544 : struct spdk_nvmf_rdma_qpair *rqpair, struct ibv_send_wr *bad_wr, int rc)
4545 : {
4546 : struct spdk_nvmf_rdma_wr *bad_rdma_wr;
4547 0 : struct spdk_nvmf_rdma_request *prev_rdma_req = NULL, *cur_rdma_req = NULL;
4548 :
4549 0 : SPDK_ERRLOG("Failed to post a send for the qpair %p with errno %d\n", rqpair, -rc);
4550 0 : for (; bad_wr != NULL; bad_wr = bad_wr->next) {
4551 0 : bad_rdma_wr = (struct spdk_nvmf_rdma_wr *)bad_wr->wr_id;
4552 0 : assert(rqpair->current_send_depth > 0);
4553 0 : rqpair->current_send_depth--;
4554 0 : switch (bad_rdma_wr->type) {
4555 0 : case RDMA_WR_TYPE_DATA:
4556 0 : cur_rdma_req = SPDK_CONTAINEROF(bad_rdma_wr, struct spdk_nvmf_rdma_request, data_wr);
4557 0 : if (bad_wr->opcode == IBV_WR_RDMA_READ) {
4558 0 : assert(rqpair->current_read_depth > 0);
4559 0 : rqpair->current_read_depth--;
4560 : }
4561 0 : break;
4562 0 : case RDMA_WR_TYPE_SEND:
4563 0 : cur_rdma_req = SPDK_CONTAINEROF(bad_rdma_wr, struct spdk_nvmf_rdma_request, rsp_wr);
4564 0 : break;
4565 0 : default:
4566 0 : SPDK_ERRLOG("Found a RECV in the list of pending SEND requests for qpair %p\n", rqpair);
4567 0 : prev_rdma_req = cur_rdma_req;
4568 0 : continue;
4569 : }
4570 :
4571 0 : if (prev_rdma_req == cur_rdma_req) {
4572 : /* this request was handled by an earlier wr. i.e. we were performing an nvme read. */
4573 : /* We only have to check against prev_wr since each requests wrs are contiguous in this list. */
4574 0 : continue;
4575 : }
4576 :
4577 0 : switch (cur_rdma_req->state) {
4578 0 : case RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER:
4579 0 : cur_rdma_req->req.rsp->nvme_cpl.status.sc = SPDK_NVME_SC_INTERNAL_DEVICE_ERROR;
4580 0 : STAILQ_INSERT_TAIL(&rqpair->pending_rdma_send_queue, cur_rdma_req, state_link);
4581 0 : cur_rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE_PENDING;
4582 0 : break;
4583 0 : case RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST:
4584 : case RDMA_REQUEST_STATE_COMPLETING:
4585 0 : cur_rdma_req->state = RDMA_REQUEST_STATE_COMPLETED;
4586 0 : break;
4587 0 : default:
4588 0 : SPDK_ERRLOG("Found a request in a bad state %d when draining pending SEND requests for qpair %p\n",
4589 : cur_rdma_req->state, rqpair);
4590 0 : continue;
4591 : }
4592 :
4593 0 : nvmf_rdma_request_process(rtransport, cur_rdma_req);
4594 0 : prev_rdma_req = cur_rdma_req;
4595 : }
4596 :
4597 0 : if (rqpair->qpair.state == SPDK_NVMF_QPAIR_ACTIVE) {
4598 : /* Disconnect the connection. */
4599 0 : spdk_nvmf_qpair_disconnect(&rqpair->qpair, NULL, NULL);
4600 : }
4601 :
4602 0 : }
4603 :
4604 : static void
4605 0 : _poller_submit_sends(struct spdk_nvmf_rdma_transport *rtransport,
4606 : struct spdk_nvmf_rdma_poller *rpoller)
4607 : {
4608 : struct spdk_nvmf_rdma_qpair *rqpair;
4609 0 : struct ibv_send_wr *bad_wr = NULL;
4610 : int rc;
4611 :
4612 0 : while (!STAILQ_EMPTY(&rpoller->qpairs_pending_send)) {
4613 0 : rqpair = STAILQ_FIRST(&rpoller->qpairs_pending_send);
4614 0 : rc = spdk_rdma_qp_flush_send_wrs(rqpair->rdma_qp, &bad_wr);
4615 :
4616 : /* bad wr always points to the first wr that failed. */
4617 0 : if (rc) {
4618 0 : _qp_reset_failed_sends(rtransport, rqpair, bad_wr, rc);
4619 : }
4620 0 : STAILQ_REMOVE_HEAD(&rpoller->qpairs_pending_send, send_link);
4621 : }
4622 0 : }
4623 :
4624 : static const char *
4625 0 : nvmf_rdma_wr_type_str(enum spdk_nvmf_rdma_wr_type wr_type)
4626 : {
4627 0 : switch (wr_type) {
4628 0 : case RDMA_WR_TYPE_RECV:
4629 0 : return "RECV";
4630 0 : case RDMA_WR_TYPE_SEND:
4631 0 : return "SEND";
4632 0 : case RDMA_WR_TYPE_DATA:
4633 0 : return "DATA";
4634 0 : default:
4635 0 : SPDK_ERRLOG("Unknown WR type %d\n", wr_type);
4636 0 : SPDK_UNREACHABLE();
4637 : }
4638 : }
4639 :
4640 : static inline void
4641 0 : nvmf_rdma_log_wc_status(struct spdk_nvmf_rdma_qpair *rqpair, struct ibv_wc *wc)
4642 : {
4643 0 : enum spdk_nvmf_rdma_wr_type wr_type = ((struct spdk_nvmf_rdma_wr *)wc->wr_id)->type;
4644 :
4645 0 : if (wc->status == IBV_WC_WR_FLUSH_ERR) {
4646 : /* If qpair is in ERR state, we will receive completions for all posted and not completed
4647 : * Work Requests with IBV_WC_WR_FLUSH_ERR status. Don't log an error in that case */
4648 0 : SPDK_DEBUGLOG(rdma,
4649 : "Error on CQ %p, (qp state %d ibv_state %d) request 0x%lu, type %s, status: (%d): %s\n",
4650 : rqpair->poller->cq, rqpair->qpair.state, rqpair->ibv_state, wc->wr_id,
4651 : nvmf_rdma_wr_type_str(wr_type), wc->status, ibv_wc_status_str(wc->status));
4652 : } else {
4653 0 : SPDK_ERRLOG("Error on CQ %p, (qp state %d ibv_state %d) request 0x%lu, type %s, status: (%d): %s\n",
4654 : rqpair->poller->cq, rqpair->qpair.state, rqpair->ibv_state, wc->wr_id,
4655 : nvmf_rdma_wr_type_str(wr_type), wc->status, ibv_wc_status_str(wc->status));
4656 : }
4657 0 : }
4658 :
4659 : static int
4660 0 : nvmf_rdma_poller_poll(struct spdk_nvmf_rdma_transport *rtransport,
4661 : struct spdk_nvmf_rdma_poller *rpoller)
4662 : {
4663 0 : struct ibv_wc wc[32];
4664 : struct spdk_nvmf_rdma_wr *rdma_wr;
4665 : struct spdk_nvmf_rdma_request *rdma_req;
4666 : struct spdk_nvmf_rdma_recv *rdma_recv;
4667 : struct spdk_nvmf_rdma_qpair *rqpair, *tmp_rqpair;
4668 : int reaped, i;
4669 0 : int count = 0;
4670 : int rc;
4671 0 : bool error = false;
4672 0 : uint64_t poll_tsc = spdk_get_ticks();
4673 :
4674 0 : if (spdk_unlikely(rpoller->need_destroy)) {
4675 : /* If qpair is closed before poller destroy, nvmf_rdma_destroy_drained_qpair may not
4676 : * be called because we cannot poll anything from cq. So we call that here to force
4677 : * destroy the qpair after to_close turning true.
4678 : */
4679 0 : RB_FOREACH_SAFE(rqpair, qpairs_tree, &rpoller->qpairs, tmp_rqpair) {
4680 0 : nvmf_rdma_destroy_drained_qpair(rqpair);
4681 : }
4682 0 : return 0;
4683 : }
4684 :
4685 : /* Poll for completing operations. */
4686 0 : reaped = ibv_poll_cq(rpoller->cq, 32, wc);
4687 0 : if (reaped < 0) {
4688 0 : SPDK_ERRLOG("Error polling CQ! (%d): %s\n",
4689 : errno, spdk_strerror(errno));
4690 0 : return -1;
4691 0 : } else if (reaped == 0) {
4692 0 : rpoller->stat.idle_polls++;
4693 : }
4694 :
4695 0 : rpoller->stat.polls++;
4696 0 : rpoller->stat.completions += reaped;
4697 :
4698 0 : for (i = 0; i < reaped; i++) {
4699 :
4700 0 : rdma_wr = (struct spdk_nvmf_rdma_wr *)wc[i].wr_id;
4701 :
4702 0 : switch (rdma_wr->type) {
4703 0 : case RDMA_WR_TYPE_SEND:
4704 0 : rdma_req = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvmf_rdma_request, rsp_wr);
4705 0 : rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair);
4706 :
4707 0 : if (!wc[i].status) {
4708 0 : count++;
4709 0 : assert(wc[i].opcode == IBV_WC_SEND);
4710 0 : assert(nvmf_rdma_req_is_completing(rdma_req));
4711 : }
4712 :
4713 0 : rdma_req->state = RDMA_REQUEST_STATE_COMPLETED;
4714 : /* RDMA_WRITE operation completed. +1 since it was chained with rsp WR */
4715 0 : assert(rqpair->current_send_depth >= (uint32_t)rdma_req->num_outstanding_data_wr + 1);
4716 0 : rqpair->current_send_depth -= rdma_req->num_outstanding_data_wr + 1;
4717 0 : rdma_req->num_outstanding_data_wr = 0;
4718 :
4719 0 : nvmf_rdma_request_process(rtransport, rdma_req);
4720 0 : break;
4721 0 : case RDMA_WR_TYPE_RECV:
4722 : /* rdma_recv->qpair will be invalid if using an SRQ. In that case we have to get the qpair from the wc. */
4723 0 : rdma_recv = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvmf_rdma_recv, rdma_wr);
4724 0 : if (rpoller->srq != NULL) {
4725 0 : rdma_recv->qpair = get_rdma_qpair_from_wc(rpoller, &wc[i]);
4726 : /* It is possible that there are still some completions for destroyed QP
4727 : * associated with SRQ. We just ignore these late completions and re-post
4728 : * receive WRs back to SRQ.
4729 : */
4730 0 : if (spdk_unlikely(NULL == rdma_recv->qpair)) {
4731 0 : struct ibv_recv_wr *bad_wr;
4732 :
4733 0 : rdma_recv->wr.next = NULL;
4734 0 : spdk_rdma_srq_queue_recv_wrs(rpoller->srq, &rdma_recv->wr);
4735 0 : rc = spdk_rdma_srq_flush_recv_wrs(rpoller->srq, &bad_wr);
4736 0 : if (rc) {
4737 0 : SPDK_ERRLOG("Failed to re-post recv WR to SRQ, err %d\n", rc);
4738 : }
4739 0 : continue;
4740 : }
4741 : }
4742 0 : rqpair = rdma_recv->qpair;
4743 :
4744 0 : assert(rqpair != NULL);
4745 0 : if (!wc[i].status) {
4746 0 : assert(wc[i].opcode == IBV_WC_RECV);
4747 0 : if (rqpair->current_recv_depth >= rqpair->max_queue_depth) {
4748 0 : spdk_nvmf_qpair_disconnect(&rqpair->qpair, NULL, NULL);
4749 0 : break;
4750 : }
4751 : }
4752 :
4753 0 : rdma_recv->wr.next = NULL;
4754 0 : rqpair->current_recv_depth++;
4755 0 : rdma_recv->receive_tsc = poll_tsc;
4756 0 : rpoller->stat.requests++;
4757 0 : STAILQ_INSERT_HEAD(&rqpair->resources->incoming_queue, rdma_recv, link);
4758 0 : break;
4759 0 : case RDMA_WR_TYPE_DATA:
4760 0 : rdma_req = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvmf_rdma_request, data_wr);
4761 0 : rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair);
4762 :
4763 0 : assert(rdma_req->num_outstanding_data_wr > 0);
4764 :
4765 0 : rqpair->current_send_depth--;
4766 0 : rdma_req->num_outstanding_data_wr--;
4767 0 : if (!wc[i].status) {
4768 0 : assert(wc[i].opcode == IBV_WC_RDMA_READ);
4769 0 : rqpair->current_read_depth--;
4770 : /* wait for all outstanding reads associated with the same rdma_req to complete before proceeding. */
4771 0 : if (rdma_req->num_outstanding_data_wr == 0) {
4772 0 : if (rdma_req->num_remaining_data_wr) {
4773 : /* Only part of RDMA_READ operations was submitted, process the rest */
4774 0 : nvmf_rdma_request_reset_transfer_in(rdma_req, rtransport);
4775 0 : rdma_req->state = RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING;
4776 0 : nvmf_rdma_request_process(rtransport, rdma_req);
4777 0 : break;
4778 : }
4779 0 : rdma_req->state = RDMA_REQUEST_STATE_READY_TO_EXECUTE;
4780 0 : nvmf_rdma_request_process(rtransport, rdma_req);
4781 : }
4782 : } else {
4783 : /* If the data transfer fails still force the queue into the error state,
4784 : * if we were performing an RDMA_READ, we need to force the request into a
4785 : * completed state since it wasn't linked to a send. However, in the RDMA_WRITE
4786 : * case, we should wait for the SEND to complete. */
4787 0 : if (rdma_req->data.wr.opcode == IBV_WR_RDMA_READ) {
4788 0 : rqpair->current_read_depth--;
4789 0 : if (rdma_req->num_outstanding_data_wr == 0) {
4790 0 : rdma_req->state = RDMA_REQUEST_STATE_COMPLETED;
4791 : }
4792 : }
4793 : }
4794 0 : break;
4795 0 : default:
4796 0 : SPDK_ERRLOG("Received an unknown opcode on the CQ: %d\n", wc[i].opcode);
4797 0 : continue;
4798 : }
4799 :
4800 : /* Handle error conditions */
4801 0 : if (wc[i].status) {
4802 0 : nvmf_rdma_update_ibv_state(rqpair);
4803 0 : nvmf_rdma_log_wc_status(rqpair, &wc[i]);
4804 :
4805 0 : error = true;
4806 :
4807 0 : if (rqpair->qpair.state == SPDK_NVMF_QPAIR_ACTIVE) {
4808 : /* Disconnect the connection. */
4809 0 : spdk_nvmf_qpair_disconnect(&rqpair->qpair, NULL, NULL);
4810 : } else {
4811 0 : nvmf_rdma_destroy_drained_qpair(rqpair);
4812 : }
4813 0 : continue;
4814 : }
4815 :
4816 0 : nvmf_rdma_qpair_process_pending(rtransport, rqpair, false);
4817 :
4818 0 : if (rqpair->qpair.state != SPDK_NVMF_QPAIR_ACTIVE) {
4819 0 : nvmf_rdma_destroy_drained_qpair(rqpair);
4820 : }
4821 : }
4822 :
4823 0 : if (error == true) {
4824 0 : return -1;
4825 : }
4826 :
4827 : /* submit outstanding work requests. */
4828 0 : _poller_submit_recvs(rtransport, rpoller);
4829 0 : _poller_submit_sends(rtransport, rpoller);
4830 :
4831 0 : return count;
4832 : }
4833 :
4834 : static void
4835 0 : _nvmf_rdma_remove_destroyed_device(void *c)
4836 : {
4837 0 : struct spdk_nvmf_rdma_transport *rtransport = c;
4838 : struct spdk_nvmf_rdma_device *device, *device_tmp;
4839 : int rc;
4840 :
4841 0 : TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, device_tmp) {
4842 0 : if (device->ready_to_destroy) {
4843 0 : destroy_ib_device(rtransport, device);
4844 : }
4845 : }
4846 :
4847 0 : free_poll_fds(rtransport);
4848 0 : rc = generate_poll_fds(rtransport);
4849 : /* cannot handle fd allocation error here */
4850 0 : if (rc != 0) {
4851 0 : SPDK_ERRLOG("Failed to generate poll fds after remove ib device.\n");
4852 : }
4853 0 : }
4854 :
4855 : static void
4856 0 : _nvmf_rdma_remove_poller_in_group_cb(void *c)
4857 : {
4858 0 : struct poller_manage_ctx *ctx = c;
4859 0 : struct spdk_nvmf_rdma_transport *rtransport = ctx->rtransport;
4860 0 : struct spdk_nvmf_rdma_device *device = ctx->device;
4861 0 : struct spdk_thread *thread = ctx->thread;
4862 :
4863 0 : if (nvmf_rdma_all_pollers_management_done(c)) {
4864 : /* destroy device when last poller is destroyed */
4865 0 : device->ready_to_destroy = true;
4866 0 : spdk_thread_send_msg(thread, _nvmf_rdma_remove_destroyed_device, rtransport);
4867 : }
4868 0 : }
4869 :
4870 : static void
4871 0 : _nvmf_rdma_remove_poller_in_group(void *c)
4872 : {
4873 0 : struct poller_manage_ctx *ctx = c;
4874 :
4875 0 : ctx->rpoller->need_destroy = true;
4876 0 : ctx->rpoller->destroy_cb_ctx = ctx;
4877 0 : ctx->rpoller->destroy_cb = _nvmf_rdma_remove_poller_in_group_cb;
4878 :
4879 : /* qp will be disconnected after receiving a RDMA_CM_EVENT_DEVICE_REMOVAL event. */
4880 0 : if (RB_EMPTY(&ctx->rpoller->qpairs)) {
4881 0 : nvmf_rdma_poller_destroy(ctx->rpoller);
4882 : }
4883 0 : }
4884 :
4885 : static int
4886 0 : nvmf_rdma_poll_group_poll(struct spdk_nvmf_transport_poll_group *group)
4887 : {
4888 : struct spdk_nvmf_rdma_transport *rtransport;
4889 : struct spdk_nvmf_rdma_poll_group *rgroup;
4890 : struct spdk_nvmf_rdma_poller *rpoller, *tmp;
4891 : int count, rc;
4892 :
4893 0 : rtransport = SPDK_CONTAINEROF(group->transport, struct spdk_nvmf_rdma_transport, transport);
4894 0 : rgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_rdma_poll_group, group);
4895 :
4896 0 : count = 0;
4897 0 : TAILQ_FOREACH_SAFE(rpoller, &rgroup->pollers, link, tmp) {
4898 0 : rc = nvmf_rdma_poller_poll(rtransport, rpoller);
4899 0 : if (rc < 0) {
4900 0 : return rc;
4901 : }
4902 0 : count += rc;
4903 : }
4904 :
4905 0 : return count;
4906 : }
4907 :
4908 : static int
4909 0 : nvmf_rdma_trid_from_cm_id(struct rdma_cm_id *id,
4910 : struct spdk_nvme_transport_id *trid,
4911 : bool peer)
4912 : {
4913 : struct sockaddr *saddr;
4914 : uint16_t port;
4915 :
4916 0 : spdk_nvme_trid_populate_transport(trid, SPDK_NVME_TRANSPORT_RDMA);
4917 :
4918 0 : if (peer) {
4919 0 : saddr = rdma_get_peer_addr(id);
4920 : } else {
4921 0 : saddr = rdma_get_local_addr(id);
4922 : }
4923 0 : switch (saddr->sa_family) {
4924 0 : case AF_INET: {
4925 0 : struct sockaddr_in *saddr_in = (struct sockaddr_in *)saddr;
4926 :
4927 0 : trid->adrfam = SPDK_NVMF_ADRFAM_IPV4;
4928 0 : inet_ntop(AF_INET, &saddr_in->sin_addr,
4929 0 : trid->traddr, sizeof(trid->traddr));
4930 0 : if (peer) {
4931 0 : port = ntohs(rdma_get_dst_port(id));
4932 : } else {
4933 0 : port = ntohs(rdma_get_src_port(id));
4934 : }
4935 0 : snprintf(trid->trsvcid, sizeof(trid->trsvcid), "%u", port);
4936 0 : break;
4937 : }
4938 0 : case AF_INET6: {
4939 0 : struct sockaddr_in6 *saddr_in = (struct sockaddr_in6 *)saddr;
4940 0 : trid->adrfam = SPDK_NVMF_ADRFAM_IPV6;
4941 0 : inet_ntop(AF_INET6, &saddr_in->sin6_addr,
4942 0 : trid->traddr, sizeof(trid->traddr));
4943 0 : if (peer) {
4944 0 : port = ntohs(rdma_get_dst_port(id));
4945 : } else {
4946 0 : port = ntohs(rdma_get_src_port(id));
4947 : }
4948 0 : snprintf(trid->trsvcid, sizeof(trid->trsvcid), "%u", port);
4949 0 : break;
4950 : }
4951 0 : default:
4952 0 : return -1;
4953 :
4954 : }
4955 :
4956 0 : return 0;
4957 : }
4958 :
4959 : static int
4960 0 : nvmf_rdma_qpair_get_peer_trid(struct spdk_nvmf_qpair *qpair,
4961 : struct spdk_nvme_transport_id *trid)
4962 : {
4963 : struct spdk_nvmf_rdma_qpair *rqpair;
4964 :
4965 0 : rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair);
4966 :
4967 0 : return nvmf_rdma_trid_from_cm_id(rqpair->cm_id, trid, true);
4968 : }
4969 :
4970 : static int
4971 0 : nvmf_rdma_qpair_get_local_trid(struct spdk_nvmf_qpair *qpair,
4972 : struct spdk_nvme_transport_id *trid)
4973 : {
4974 : struct spdk_nvmf_rdma_qpair *rqpair;
4975 :
4976 0 : rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair);
4977 :
4978 0 : return nvmf_rdma_trid_from_cm_id(rqpair->cm_id, trid, false);
4979 : }
4980 :
4981 : static int
4982 0 : nvmf_rdma_qpair_get_listen_trid(struct spdk_nvmf_qpair *qpair,
4983 : struct spdk_nvme_transport_id *trid)
4984 : {
4985 : struct spdk_nvmf_rdma_qpair *rqpair;
4986 :
4987 0 : rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair);
4988 :
4989 0 : return nvmf_rdma_trid_from_cm_id(rqpair->listen_id, trid, false);
4990 : }
4991 :
4992 : void
4993 0 : spdk_nvmf_rdma_init_hooks(struct spdk_nvme_rdma_hooks *hooks)
4994 : {
4995 0 : g_nvmf_hooks = *hooks;
4996 0 : }
4997 :
4998 : static void
4999 0 : nvmf_rdma_request_set_abort_status(struct spdk_nvmf_request *req,
5000 : struct spdk_nvmf_rdma_request *rdma_req_to_abort,
5001 : struct spdk_nvmf_rdma_qpair *rqpair)
5002 : {
5003 0 : rdma_req_to_abort->req.rsp->nvme_cpl.status.sct = SPDK_NVME_SCT_GENERIC;
5004 0 : rdma_req_to_abort->req.rsp->nvme_cpl.status.sc = SPDK_NVME_SC_ABORTED_BY_REQUEST;
5005 :
5006 0 : STAILQ_INSERT_TAIL(&rqpair->pending_rdma_send_queue, rdma_req_to_abort, state_link);
5007 0 : rdma_req_to_abort->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE_PENDING;
5008 :
5009 0 : req->rsp->nvme_cpl.cdw0 &= ~1U; /* Command was successfully aborted. */
5010 0 : }
5011 :
5012 : static int
5013 0 : _nvmf_rdma_qpair_abort_request(void *ctx)
5014 : {
5015 0 : struct spdk_nvmf_request *req = ctx;
5016 0 : struct spdk_nvmf_rdma_request *rdma_req_to_abort = SPDK_CONTAINEROF(
5017 : req->req_to_abort, struct spdk_nvmf_rdma_request, req);
5018 0 : struct spdk_nvmf_rdma_qpair *rqpair = SPDK_CONTAINEROF(req->req_to_abort->qpair,
5019 : struct spdk_nvmf_rdma_qpair, qpair);
5020 : int rc;
5021 :
5022 0 : spdk_poller_unregister(&req->poller);
5023 :
5024 0 : switch (rdma_req_to_abort->state) {
5025 0 : case RDMA_REQUEST_STATE_EXECUTING:
5026 0 : rc = nvmf_ctrlr_abort_request(req);
5027 0 : if (rc == SPDK_NVMF_REQUEST_EXEC_STATUS_ASYNCHRONOUS) {
5028 0 : return SPDK_POLLER_BUSY;
5029 : }
5030 0 : break;
5031 :
5032 0 : case RDMA_REQUEST_STATE_NEED_BUFFER:
5033 0 : STAILQ_REMOVE(&rqpair->poller->group->group.pending_buf_queue,
5034 : &rdma_req_to_abort->req, spdk_nvmf_request, buf_link);
5035 :
5036 0 : nvmf_rdma_request_set_abort_status(req, rdma_req_to_abort, rqpair);
5037 0 : break;
5038 :
5039 0 : case RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING:
5040 0 : STAILQ_REMOVE(&rqpair->pending_rdma_read_queue, rdma_req_to_abort,
5041 : spdk_nvmf_rdma_request, state_link);
5042 :
5043 0 : nvmf_rdma_request_set_abort_status(req, rdma_req_to_abort, rqpair);
5044 0 : break;
5045 :
5046 0 : case RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING:
5047 0 : STAILQ_REMOVE(&rqpair->pending_rdma_write_queue, rdma_req_to_abort,
5048 : spdk_nvmf_rdma_request, state_link);
5049 :
5050 0 : nvmf_rdma_request_set_abort_status(req, rdma_req_to_abort, rqpair);
5051 0 : break;
5052 :
5053 0 : case RDMA_REQUEST_STATE_READY_TO_COMPLETE_PENDING:
5054 : /* Remove req from the list here to re-use common function */
5055 0 : STAILQ_REMOVE(&rqpair->pending_rdma_send_queue, rdma_req_to_abort,
5056 : spdk_nvmf_rdma_request, state_link);
5057 :
5058 0 : nvmf_rdma_request_set_abort_status(req, rdma_req_to_abort, rqpair);
5059 0 : break;
5060 :
5061 0 : case RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER:
5062 0 : if (spdk_get_ticks() < req->timeout_tsc) {
5063 0 : req->poller = SPDK_POLLER_REGISTER(_nvmf_rdma_qpair_abort_request, req, 0);
5064 0 : return SPDK_POLLER_BUSY;
5065 : }
5066 0 : break;
5067 :
5068 0 : default:
5069 0 : break;
5070 : }
5071 :
5072 0 : spdk_nvmf_request_complete(req);
5073 0 : return SPDK_POLLER_BUSY;
5074 : }
5075 :
5076 : static void
5077 0 : nvmf_rdma_qpair_abort_request(struct spdk_nvmf_qpair *qpair,
5078 : struct spdk_nvmf_request *req)
5079 : {
5080 : struct spdk_nvmf_rdma_qpair *rqpair;
5081 : struct spdk_nvmf_rdma_transport *rtransport;
5082 : struct spdk_nvmf_transport *transport;
5083 : uint16_t cid;
5084 : uint32_t i, max_req_count;
5085 0 : struct spdk_nvmf_rdma_request *rdma_req_to_abort = NULL, *rdma_req;
5086 :
5087 0 : rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair);
5088 0 : rtransport = SPDK_CONTAINEROF(qpair->transport, struct spdk_nvmf_rdma_transport, transport);
5089 0 : transport = &rtransport->transport;
5090 :
5091 0 : cid = req->cmd->nvme_cmd.cdw10_bits.abort.cid;
5092 0 : max_req_count = rqpair->srq == NULL ? rqpair->max_queue_depth : rqpair->poller->max_srq_depth;
5093 :
5094 0 : for (i = 0; i < max_req_count; i++) {
5095 0 : rdma_req = &rqpair->resources->reqs[i];
5096 : /* When SRQ == NULL, rqpair has its own requests and req.qpair pointer always points to the qpair
5097 : * When SRQ != NULL all rqpairs share common requests and qpair pointer is assigned when we start to
5098 : * process a request. So in both cases all requests which are not in FREE state have valid qpair ptr */
5099 0 : if (rdma_req->state != RDMA_REQUEST_STATE_FREE && rdma_req->req.cmd->nvme_cmd.cid == cid &&
5100 0 : rdma_req->req.qpair == qpair) {
5101 0 : rdma_req_to_abort = rdma_req;
5102 0 : break;
5103 : }
5104 : }
5105 :
5106 0 : if (rdma_req_to_abort == NULL) {
5107 0 : spdk_nvmf_request_complete(req);
5108 0 : return;
5109 : }
5110 :
5111 0 : req->req_to_abort = &rdma_req_to_abort->req;
5112 0 : req->timeout_tsc = spdk_get_ticks() +
5113 0 : transport->opts.abort_timeout_sec * spdk_get_ticks_hz();
5114 0 : req->poller = NULL;
5115 :
5116 0 : _nvmf_rdma_qpair_abort_request(req);
5117 : }
5118 :
5119 : static void
5120 0 : nvmf_rdma_poll_group_dump_stat(struct spdk_nvmf_transport_poll_group *group,
5121 : struct spdk_json_write_ctx *w)
5122 : {
5123 : struct spdk_nvmf_rdma_poll_group *rgroup;
5124 : struct spdk_nvmf_rdma_poller *rpoller;
5125 :
5126 0 : assert(w != NULL);
5127 :
5128 0 : rgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_rdma_poll_group, group);
5129 :
5130 0 : spdk_json_write_named_uint64(w, "pending_data_buffer", rgroup->stat.pending_data_buffer);
5131 :
5132 0 : spdk_json_write_named_array_begin(w, "devices");
5133 :
5134 0 : TAILQ_FOREACH(rpoller, &rgroup->pollers, link) {
5135 0 : spdk_json_write_object_begin(w);
5136 0 : spdk_json_write_named_string(w, "name",
5137 0 : ibv_get_device_name(rpoller->device->context->device));
5138 0 : spdk_json_write_named_uint64(w, "polls",
5139 : rpoller->stat.polls);
5140 0 : spdk_json_write_named_uint64(w, "idle_polls",
5141 : rpoller->stat.idle_polls);
5142 0 : spdk_json_write_named_uint64(w, "completions",
5143 : rpoller->stat.completions);
5144 0 : spdk_json_write_named_uint64(w, "requests",
5145 : rpoller->stat.requests);
5146 0 : spdk_json_write_named_uint64(w, "request_latency",
5147 : rpoller->stat.request_latency);
5148 0 : spdk_json_write_named_uint64(w, "pending_free_request",
5149 : rpoller->stat.pending_free_request);
5150 0 : spdk_json_write_named_uint64(w, "pending_rdma_read",
5151 : rpoller->stat.pending_rdma_read);
5152 0 : spdk_json_write_named_uint64(w, "pending_rdma_write",
5153 : rpoller->stat.pending_rdma_write);
5154 0 : spdk_json_write_named_uint64(w, "pending_rdma_send",
5155 : rpoller->stat.pending_rdma_send);
5156 0 : spdk_json_write_named_uint64(w, "total_send_wrs",
5157 : rpoller->stat.qp_stats.send.num_submitted_wrs);
5158 0 : spdk_json_write_named_uint64(w, "send_doorbell_updates",
5159 : rpoller->stat.qp_stats.send.doorbell_updates);
5160 0 : spdk_json_write_named_uint64(w, "total_recv_wrs",
5161 : rpoller->stat.qp_stats.recv.num_submitted_wrs);
5162 0 : spdk_json_write_named_uint64(w, "recv_doorbell_updates",
5163 : rpoller->stat.qp_stats.recv.doorbell_updates);
5164 0 : spdk_json_write_object_end(w);
5165 : }
5166 :
5167 0 : spdk_json_write_array_end(w);
5168 0 : }
5169 :
5170 : const struct spdk_nvmf_transport_ops spdk_nvmf_transport_rdma = {
5171 : .name = "RDMA",
5172 : .type = SPDK_NVME_TRANSPORT_RDMA,
5173 : .opts_init = nvmf_rdma_opts_init,
5174 : .create = nvmf_rdma_create,
5175 : .dump_opts = nvmf_rdma_dump_opts,
5176 : .destroy = nvmf_rdma_destroy,
5177 :
5178 : .listen = nvmf_rdma_listen,
5179 : .stop_listen = nvmf_rdma_stop_listen,
5180 : .cdata_init = nvmf_rdma_cdata_init,
5181 :
5182 : .listener_discover = nvmf_rdma_discover,
5183 :
5184 : .poll_group_create = nvmf_rdma_poll_group_create,
5185 : .get_optimal_poll_group = nvmf_rdma_get_optimal_poll_group,
5186 : .poll_group_destroy = nvmf_rdma_poll_group_destroy,
5187 : .poll_group_add = nvmf_rdma_poll_group_add,
5188 : .poll_group_remove = nvmf_rdma_poll_group_remove,
5189 : .poll_group_poll = nvmf_rdma_poll_group_poll,
5190 :
5191 : .req_free = nvmf_rdma_request_free,
5192 : .req_complete = nvmf_rdma_request_complete,
5193 :
5194 : .qpair_fini = nvmf_rdma_close_qpair,
5195 : .qpair_get_peer_trid = nvmf_rdma_qpair_get_peer_trid,
5196 : .qpair_get_local_trid = nvmf_rdma_qpair_get_local_trid,
5197 : .qpair_get_listen_trid = nvmf_rdma_qpair_get_listen_trid,
5198 : .qpair_abort_request = nvmf_rdma_qpair_abort_request,
5199 :
5200 : .poll_group_dump_stat = nvmf_rdma_poll_group_dump_stat,
5201 : };
5202 :
5203 2 : SPDK_NVMF_TRANSPORT_REGISTER(rdma, &spdk_nvmf_transport_rdma);
5204 2 : SPDK_LOG_REGISTER_COMPONENT(rdma)
|