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 : /*
8 : * NVMe over RDMA transport
9 : */
10 :
11 : #include "spdk/stdinc.h"
12 :
13 : #include "spdk/assert.h"
14 : #include "spdk/dma.h"
15 : #include "spdk/log.h"
16 : #include "spdk/trace.h"
17 : #include "spdk/queue.h"
18 : #include "spdk/nvme.h"
19 : #include "spdk/nvmf_spec.h"
20 : #include "spdk/string.h"
21 : #include "spdk/endian.h"
22 : #include "spdk/likely.h"
23 : #include "spdk/config.h"
24 : #include "spdk/net.h"
25 : #include "spdk/file.h"
26 :
27 : #include "nvme_internal.h"
28 : #include "spdk_internal/rdma_provider.h"
29 : #include "spdk_internal/rdma_utils.h"
30 :
31 : #define NVME_RDMA_TIME_OUT_IN_MS 2000
32 : #define NVME_RDMA_RW_BUFFER_SIZE 131072
33 :
34 : /*
35 : * NVME RDMA qpair Resource Defaults
36 : */
37 : #define NVME_RDMA_DEFAULT_TX_SGE 2
38 : #define NVME_RDMA_DEFAULT_RX_SGE 1
39 :
40 : /* Max number of NVMe-oF SGL descriptors supported by the host */
41 : #define NVME_RDMA_MAX_SGL_DESCRIPTORS 16
42 :
43 : /* number of STAILQ entries for holding pending RDMA CM events. */
44 : #define NVME_RDMA_NUM_CM_EVENTS 256
45 :
46 : /* The default size for a shared rdma completion queue. */
47 : #define DEFAULT_NVME_RDMA_CQ_SIZE 4096
48 :
49 : /*
50 : * In the special case of a stale connection we don't expose a mechanism
51 : * for the user to retry the connection so we need to handle it internally.
52 : */
53 : #define NVME_RDMA_STALE_CONN_RETRY_MAX 5
54 : #define NVME_RDMA_STALE_CONN_RETRY_DELAY_US 10000
55 :
56 : /*
57 : * Maximum value of transport_retry_count used by RDMA controller
58 : */
59 : #define NVME_RDMA_CTRLR_MAX_TRANSPORT_RETRY_COUNT 7
60 :
61 : /*
62 : * Maximum value of transport_ack_timeout used by RDMA controller
63 : */
64 : #define NVME_RDMA_CTRLR_MAX_TRANSPORT_ACK_TIMEOUT 31
65 :
66 : /*
67 : * Number of microseconds to wait until the lingering qpair becomes quiet.
68 : */
69 : #define NVME_RDMA_DISCONNECTED_QPAIR_TIMEOUT_US 1000000ull
70 :
71 : /*
72 : * The max length of keyed SGL data block (3 bytes)
73 : */
74 : #define NVME_RDMA_MAX_KEYED_SGL_LENGTH ((1u << 24u) - 1)
75 :
76 : #define WC_PER_QPAIR(queue_depth) (queue_depth * 2)
77 :
78 : #define NVME_RDMA_POLL_GROUP_CHECK_QPN(_rqpair, qpn) \
79 : ((_rqpair)->rdma_qp && (_rqpair)->rdma_qp->qp->qp_num == (qpn)) \
80 :
81 : enum nvme_rdma_wr_type {
82 : RDMA_WR_TYPE_RECV,
83 : RDMA_WR_TYPE_SEND,
84 : };
85 :
86 : struct nvme_rdma_wr {
87 : /* Using this instead of the enum allows this struct to only occupy one byte. */
88 : uint8_t type;
89 : };
90 :
91 : struct spdk_nvmf_cmd {
92 : struct spdk_nvme_cmd cmd;
93 : struct spdk_nvme_sgl_descriptor sgl[NVME_RDMA_MAX_SGL_DESCRIPTORS];
94 : };
95 :
96 : struct spdk_nvme_rdma_hooks g_nvme_hooks = {};
97 :
98 : /* STAILQ wrapper for cm events. */
99 : struct nvme_rdma_cm_event_entry {
100 : struct rdma_cm_event *evt;
101 : STAILQ_ENTRY(nvme_rdma_cm_event_entry) link;
102 : };
103 :
104 : /* NVMe RDMA transport extensions for spdk_nvme_ctrlr */
105 : struct nvme_rdma_ctrlr {
106 : struct spdk_nvme_ctrlr ctrlr;
107 :
108 : uint16_t max_sge;
109 :
110 : struct rdma_event_channel *cm_channel;
111 :
112 : STAILQ_HEAD(, nvme_rdma_cm_event_entry) pending_cm_events;
113 :
114 : STAILQ_HEAD(, nvme_rdma_cm_event_entry) free_cm_events;
115 :
116 : struct nvme_rdma_cm_event_entry *cm_events;
117 : };
118 :
119 : struct nvme_rdma_poller_stats {
120 : uint64_t polls;
121 : uint64_t idle_polls;
122 : uint64_t queued_requests;
123 : uint64_t completions;
124 : struct spdk_rdma_provider_qp_stats rdma_stats;
125 : };
126 :
127 : struct nvme_rdma_poll_group;
128 : struct nvme_rdma_rsps;
129 :
130 : struct nvme_rdma_poller {
131 : struct ibv_context *device;
132 : struct ibv_cq *cq;
133 : struct spdk_rdma_provider_srq *srq;
134 : struct nvme_rdma_rsps *rsps;
135 : struct ibv_pd *pd;
136 : struct spdk_rdma_utils_mem_map *mr_map;
137 : uint32_t refcnt;
138 : int required_num_wc;
139 : int current_num_wc;
140 : struct nvme_rdma_poller_stats stats;
141 : struct nvme_rdma_poll_group *group;
142 : STAILQ_ENTRY(nvme_rdma_poller) link;
143 : };
144 :
145 : struct nvme_rdma_qpair;
146 :
147 : struct nvme_rdma_poll_group {
148 : struct spdk_nvme_transport_poll_group group;
149 : STAILQ_HEAD(, nvme_rdma_poller) pollers;
150 : uint32_t num_pollers;
151 : TAILQ_HEAD(, nvme_rdma_qpair) connecting_qpairs;
152 : TAILQ_HEAD(, nvme_rdma_qpair) active_qpairs;
153 : };
154 :
155 : enum nvme_rdma_qpair_state {
156 : NVME_RDMA_QPAIR_STATE_INVALID = 0,
157 : NVME_RDMA_QPAIR_STATE_STALE_CONN,
158 : NVME_RDMA_QPAIR_STATE_INITIALIZING,
159 : NVME_RDMA_QPAIR_STATE_FABRIC_CONNECT_SEND,
160 : NVME_RDMA_QPAIR_STATE_FABRIC_CONNECT_POLL,
161 : NVME_RDMA_QPAIR_STATE_RUNNING,
162 : NVME_RDMA_QPAIR_STATE_EXITING,
163 : NVME_RDMA_QPAIR_STATE_LINGERING,
164 : NVME_RDMA_QPAIR_STATE_EXITED,
165 : };
166 :
167 : typedef int (*nvme_rdma_cm_event_cb)(struct nvme_rdma_qpair *rqpair, int ret);
168 :
169 : struct nvme_rdma_rsp_opts {
170 : uint16_t num_entries;
171 : struct nvme_rdma_qpair *rqpair;
172 : struct spdk_rdma_provider_srq *srq;
173 : struct spdk_rdma_utils_mem_map *mr_map;
174 : };
175 :
176 : struct nvme_rdma_rsps {
177 : /* Parallel arrays of response buffers + response SGLs of size num_entries */
178 : struct ibv_sge *rsp_sgls;
179 : struct spdk_nvme_rdma_rsp *rsps;
180 :
181 : struct ibv_recv_wr *rsp_recv_wrs;
182 :
183 : /* Count of outstanding recv objects */
184 : uint16_t current_num_recvs;
185 :
186 : uint16_t num_entries;
187 : };
188 :
189 : /* NVMe RDMA qpair extensions for spdk_nvme_qpair */
190 : struct nvme_rdma_qpair {
191 : struct spdk_nvme_qpair qpair;
192 :
193 : struct spdk_rdma_provider_qp *rdma_qp;
194 : struct rdma_cm_id *cm_id;
195 : struct ibv_cq *cq;
196 : struct spdk_rdma_provider_srq *srq;
197 :
198 : struct spdk_nvme_rdma_req *rdma_reqs;
199 :
200 : uint32_t max_send_sge;
201 :
202 : uint32_t max_recv_sge;
203 :
204 : uint16_t num_entries;
205 :
206 : bool delay_cmd_submit;
207 :
208 : uint32_t num_completions;
209 : uint32_t num_outstanding_reqs;
210 :
211 : struct nvme_rdma_rsps *rsps;
212 :
213 : /*
214 : * Array of num_entries NVMe commands registered as RDMA message buffers.
215 : * Indexed by rdma_req->id.
216 : */
217 : struct spdk_nvmf_cmd *cmds;
218 :
219 : struct spdk_rdma_utils_mem_map *mr_map;
220 :
221 : TAILQ_HEAD(, spdk_nvme_rdma_req) free_reqs;
222 : TAILQ_HEAD(, spdk_nvme_rdma_req) outstanding_reqs;
223 :
224 : struct spdk_memory_domain *memory_domain;
225 :
226 : /* Count of outstanding send objects */
227 : uint16_t current_num_sends;
228 :
229 : TAILQ_ENTRY(nvme_rdma_qpair) link_active;
230 :
231 : /* Placed at the end of the struct since it is not used frequently */
232 : struct rdma_cm_event *evt;
233 : struct nvme_rdma_poller *poller;
234 :
235 : uint64_t evt_timeout_ticks;
236 : nvme_rdma_cm_event_cb evt_cb;
237 : enum rdma_cm_event_type expected_evt_type;
238 :
239 : enum nvme_rdma_qpair_state state;
240 :
241 : bool in_connect_poll;
242 :
243 : uint8_t stale_conn_retry_count;
244 : bool need_destroy;
245 :
246 : TAILQ_ENTRY(nvme_rdma_qpair) link_connecting;
247 : };
248 :
249 : enum NVME_RDMA_COMPLETION_FLAGS {
250 : NVME_RDMA_SEND_COMPLETED = 1u << 0,
251 : NVME_RDMA_RECV_COMPLETED = 1u << 1,
252 : };
253 :
254 : struct spdk_nvme_rdma_req {
255 : uint16_t id;
256 : uint16_t completion_flags: 2;
257 : uint16_t reserved: 14;
258 : /* if completion of RDMA_RECV received before RDMA_SEND, we will complete nvme request
259 : * during processing of RDMA_SEND. To complete the request we must know the response
260 : * received in RDMA_RECV, so store it in this field */
261 : struct spdk_nvme_rdma_rsp *rdma_rsp;
262 :
263 : struct nvme_rdma_wr rdma_wr;
264 :
265 : struct ibv_send_wr send_wr;
266 :
267 : struct nvme_request *req;
268 :
269 : struct ibv_sge send_sgl[NVME_RDMA_DEFAULT_TX_SGE];
270 :
271 : TAILQ_ENTRY(spdk_nvme_rdma_req) link;
272 : };
273 :
274 : struct spdk_nvme_rdma_rsp {
275 : struct spdk_nvme_cpl cpl;
276 : struct nvme_rdma_qpair *rqpair;
277 : struct ibv_recv_wr *recv_wr;
278 : struct nvme_rdma_wr rdma_wr;
279 : };
280 :
281 : struct nvme_rdma_memory_translation_ctx {
282 : void *addr;
283 : size_t length;
284 : uint32_t lkey;
285 : uint32_t rkey;
286 : };
287 :
288 : static const char *rdma_cm_event_str[] = {
289 : "RDMA_CM_EVENT_ADDR_RESOLVED",
290 : "RDMA_CM_EVENT_ADDR_ERROR",
291 : "RDMA_CM_EVENT_ROUTE_RESOLVED",
292 : "RDMA_CM_EVENT_ROUTE_ERROR",
293 : "RDMA_CM_EVENT_CONNECT_REQUEST",
294 : "RDMA_CM_EVENT_CONNECT_RESPONSE",
295 : "RDMA_CM_EVENT_CONNECT_ERROR",
296 : "RDMA_CM_EVENT_UNREACHABLE",
297 : "RDMA_CM_EVENT_REJECTED",
298 : "RDMA_CM_EVENT_ESTABLISHED",
299 : "RDMA_CM_EVENT_DISCONNECTED",
300 : "RDMA_CM_EVENT_DEVICE_REMOVAL",
301 : "RDMA_CM_EVENT_MULTICAST_JOIN",
302 : "RDMA_CM_EVENT_MULTICAST_ERROR",
303 : "RDMA_CM_EVENT_ADDR_CHANGE",
304 : "RDMA_CM_EVENT_TIMEWAIT_EXIT"
305 : };
306 :
307 : static struct nvme_rdma_poller *nvme_rdma_poll_group_get_poller(struct nvme_rdma_poll_group *group,
308 : struct ibv_context *device);
309 : static void nvme_rdma_poll_group_put_poller(struct nvme_rdma_poll_group *group,
310 : struct nvme_rdma_poller *poller);
311 :
312 : static int nvme_rdma_ctrlr_delete_io_qpair(struct spdk_nvme_ctrlr *ctrlr,
313 : struct spdk_nvme_qpair *qpair);
314 :
315 : static inline struct nvme_rdma_qpair *
316 18 : nvme_rdma_qpair(struct spdk_nvme_qpair *qpair)
317 : {
318 18 : assert(qpair->trtype == SPDK_NVME_TRANSPORT_RDMA);
319 18 : return SPDK_CONTAINEROF(qpair, struct nvme_rdma_qpair, qpair);
320 : }
321 :
322 : static inline struct nvme_rdma_poll_group *
323 8 : nvme_rdma_poll_group(struct spdk_nvme_transport_poll_group *group)
324 : {
325 8 : return (SPDK_CONTAINEROF(group, struct nvme_rdma_poll_group, group));
326 : }
327 :
328 : static inline struct nvme_rdma_ctrlr *
329 8 : nvme_rdma_ctrlr(struct spdk_nvme_ctrlr *ctrlr)
330 : {
331 8 : assert(ctrlr->trid.trtype == SPDK_NVME_TRANSPORT_RDMA);
332 8 : return SPDK_CONTAINEROF(ctrlr, struct nvme_rdma_ctrlr, ctrlr);
333 : }
334 :
335 : static struct spdk_nvme_rdma_req *
336 3 : nvme_rdma_req_get(struct nvme_rdma_qpair *rqpair)
337 : {
338 : struct spdk_nvme_rdma_req *rdma_req;
339 :
340 3 : rdma_req = TAILQ_FIRST(&rqpair->free_reqs);
341 3 : if (rdma_req) {
342 2 : TAILQ_REMOVE(&rqpair->free_reqs, rdma_req, link);
343 : }
344 :
345 3 : return rdma_req;
346 : }
347 :
348 : static void
349 1 : nvme_rdma_req_put(struct nvme_rdma_qpair *rqpair, struct spdk_nvme_rdma_req *rdma_req)
350 : {
351 1 : rdma_req->completion_flags = 0;
352 1 : rdma_req->req = NULL;
353 1 : rdma_req->rdma_rsp = NULL;
354 1 : TAILQ_INSERT_HEAD(&rqpair->free_reqs, rdma_req, link);
355 1 : }
356 :
357 : static void
358 0 : nvme_rdma_req_complete(struct spdk_nvme_rdma_req *rdma_req,
359 : struct spdk_nvme_cpl *rsp,
360 : bool print_on_error)
361 : {
362 0 : struct nvme_request *req = rdma_req->req;
363 : struct nvme_rdma_qpair *rqpair;
364 : struct spdk_nvme_qpair *qpair;
365 : bool error, print_error;
366 :
367 0 : assert(req != NULL);
368 :
369 0 : qpair = req->qpair;
370 0 : rqpair = nvme_rdma_qpair(qpair);
371 :
372 0 : error = spdk_nvme_cpl_is_error(rsp);
373 0 : print_error = error && print_on_error && !qpair->ctrlr->opts.disable_error_logging;
374 :
375 0 : if (print_error) {
376 0 : spdk_nvme_qpair_print_command(qpair, &req->cmd);
377 : }
378 :
379 0 : if (print_error || SPDK_DEBUGLOG_FLAG_ENABLED("nvme")) {
380 0 : spdk_nvme_qpair_print_completion(qpair, rsp);
381 : }
382 :
383 0 : assert(rqpair->num_outstanding_reqs > 0);
384 0 : rqpair->num_outstanding_reqs--;
385 :
386 0 : TAILQ_REMOVE(&rqpair->outstanding_reqs, rdma_req, link);
387 :
388 0 : nvme_complete_request(req->cb_fn, req->cb_arg, qpair, req, rsp);
389 0 : nvme_rdma_req_put(rqpair, rdma_req);
390 0 : }
391 :
392 : static const char *
393 4 : nvme_rdma_cm_event_str_get(uint32_t event)
394 : {
395 4 : if (event < SPDK_COUNTOF(rdma_cm_event_str)) {
396 4 : return rdma_cm_event_str[event];
397 : } else {
398 0 : return "Undefined";
399 : }
400 : }
401 :
402 :
403 : static int
404 12 : nvme_rdma_qpair_process_cm_event(struct nvme_rdma_qpair *rqpair)
405 : {
406 12 : struct rdma_cm_event *event = rqpair->evt;
407 : struct spdk_nvmf_rdma_accept_private_data *accept_data;
408 12 : int rc = 0;
409 :
410 12 : if (event) {
411 12 : switch (event->event) {
412 1 : case RDMA_CM_EVENT_ADDR_RESOLVED:
413 : case RDMA_CM_EVENT_ADDR_ERROR:
414 : case RDMA_CM_EVENT_ROUTE_RESOLVED:
415 : case RDMA_CM_EVENT_ROUTE_ERROR:
416 1 : break;
417 1 : case RDMA_CM_EVENT_CONNECT_REQUEST:
418 1 : break;
419 1 : case RDMA_CM_EVENT_CONNECT_ERROR:
420 1 : break;
421 1 : case RDMA_CM_EVENT_UNREACHABLE:
422 : case RDMA_CM_EVENT_REJECTED:
423 1 : break;
424 2 : case RDMA_CM_EVENT_CONNECT_RESPONSE:
425 2 : rc = spdk_rdma_provider_qp_complete_connect(rqpair->rdma_qp);
426 : /* fall through */
427 2 : case RDMA_CM_EVENT_ESTABLISHED:
428 2 : accept_data = (struct spdk_nvmf_rdma_accept_private_data *)event->param.conn.private_data;
429 2 : if (accept_data == NULL) {
430 1 : rc = -1;
431 : } else {
432 1 : SPDK_DEBUGLOG(nvme, "Requested queue depth %d. Target receive queue depth %d.\n",
433 : rqpair->num_entries + 1, accept_data->crqsize);
434 : }
435 2 : break;
436 1 : case RDMA_CM_EVENT_DISCONNECTED:
437 1 : rqpair->qpair.transport_failure_reason = SPDK_NVME_QPAIR_FAILURE_REMOTE;
438 1 : break;
439 1 : case RDMA_CM_EVENT_DEVICE_REMOVAL:
440 1 : rqpair->qpair.transport_failure_reason = SPDK_NVME_QPAIR_FAILURE_LOCAL;
441 1 : rqpair->need_destroy = true;
442 1 : break;
443 1 : case RDMA_CM_EVENT_MULTICAST_JOIN:
444 : case RDMA_CM_EVENT_MULTICAST_ERROR:
445 1 : break;
446 1 : case RDMA_CM_EVENT_ADDR_CHANGE:
447 1 : rqpair->qpair.transport_failure_reason = SPDK_NVME_QPAIR_FAILURE_LOCAL;
448 1 : break;
449 1 : case RDMA_CM_EVENT_TIMEWAIT_EXIT:
450 1 : break;
451 1 : default:
452 1 : SPDK_ERRLOG("Unexpected Acceptor Event [%d]\n", event->event);
453 1 : break;
454 : }
455 12 : rqpair->evt = NULL;
456 12 : rdma_ack_cm_event(event);
457 : }
458 :
459 12 : return rc;
460 : }
461 :
462 : /*
463 : * This function must be called under the nvme controller's lock
464 : * because it touches global controller variables. The lock is taken
465 : * by the generic transport code before invoking a few of the functions
466 : * in this file: nvme_rdma_ctrlr_connect_qpair, nvme_rdma_ctrlr_delete_io_qpair,
467 : * and conditionally nvme_rdma_qpair_process_completions when it is calling
468 : * completions on the admin qpair. When adding a new call to this function, please
469 : * verify that it is in a situation where it falls under the lock.
470 : */
471 : static int
472 0 : nvme_rdma_poll_events(struct nvme_rdma_ctrlr *rctrlr)
473 : {
474 : struct nvme_rdma_cm_event_entry *entry, *tmp;
475 : struct nvme_rdma_qpair *event_qpair;
476 0 : struct rdma_cm_event *event;
477 0 : struct rdma_event_channel *channel = rctrlr->cm_channel;
478 :
479 0 : STAILQ_FOREACH_SAFE(entry, &rctrlr->pending_cm_events, link, tmp) {
480 0 : event_qpair = entry->evt->id->context;
481 0 : if (event_qpair->evt == NULL) {
482 0 : event_qpair->evt = entry->evt;
483 0 : STAILQ_REMOVE(&rctrlr->pending_cm_events, entry, nvme_rdma_cm_event_entry, link);
484 0 : STAILQ_INSERT_HEAD(&rctrlr->free_cm_events, entry, link);
485 : }
486 : }
487 :
488 0 : while (rdma_get_cm_event(channel, &event) == 0) {
489 0 : event_qpair = event->id->context;
490 0 : if (event_qpair->evt == NULL) {
491 0 : event_qpair->evt = event;
492 : } else {
493 0 : assert(rctrlr == nvme_rdma_ctrlr(event_qpair->qpair.ctrlr));
494 0 : entry = STAILQ_FIRST(&rctrlr->free_cm_events);
495 0 : if (entry == NULL) {
496 0 : rdma_ack_cm_event(event);
497 0 : return -ENOMEM;
498 : }
499 0 : STAILQ_REMOVE(&rctrlr->free_cm_events, entry, nvme_rdma_cm_event_entry, link);
500 0 : entry->evt = event;
501 0 : STAILQ_INSERT_TAIL(&rctrlr->pending_cm_events, entry, link);
502 : }
503 : }
504 :
505 : /* rdma_get_cm_event() returns -1 on error. If an error occurs, errno
506 : * will be set to indicate the failure reason. So return negated errno here.
507 : */
508 0 : return -errno;
509 : }
510 :
511 : static int
512 4 : nvme_rdma_validate_cm_event(enum rdma_cm_event_type expected_evt_type,
513 : struct rdma_cm_event *reaped_evt)
514 : {
515 4 : int rc = -EBADMSG;
516 :
517 4 : if (expected_evt_type == reaped_evt->event) {
518 1 : return 0;
519 : }
520 :
521 3 : switch (expected_evt_type) {
522 2 : case RDMA_CM_EVENT_ESTABLISHED:
523 : /*
524 : * There is an enum ib_cm_rej_reason in the kernel headers that sets 10 as
525 : * IB_CM_REJ_STALE_CONN. I can't find the corresponding userspace but we get
526 : * the same values here.
527 : */
528 2 : if (reaped_evt->event == RDMA_CM_EVENT_REJECTED && reaped_evt->status == 10) {
529 1 : rc = -ESTALE;
530 1 : } else if (reaped_evt->event == RDMA_CM_EVENT_CONNECT_RESPONSE) {
531 : /*
532 : * If we are using a qpair which is not created using rdma cm API
533 : * then we will receive RDMA_CM_EVENT_CONNECT_RESPONSE instead of
534 : * RDMA_CM_EVENT_ESTABLISHED.
535 : */
536 1 : return 0;
537 : }
538 1 : break;
539 1 : default:
540 1 : break;
541 : }
542 :
543 2 : SPDK_ERRLOG("Expected %s but received %s (%d) from CM event channel (status = %d)\n",
544 : nvme_rdma_cm_event_str_get(expected_evt_type),
545 : nvme_rdma_cm_event_str_get(reaped_evt->event), reaped_evt->event,
546 : reaped_evt->status);
547 2 : return rc;
548 : }
549 :
550 : static int
551 0 : nvme_rdma_process_event_start(struct nvme_rdma_qpair *rqpair,
552 : enum rdma_cm_event_type evt,
553 : nvme_rdma_cm_event_cb evt_cb)
554 : {
555 : int rc;
556 :
557 0 : assert(evt_cb != NULL);
558 :
559 0 : if (rqpair->evt != NULL) {
560 0 : rc = nvme_rdma_qpair_process_cm_event(rqpair);
561 0 : if (rc) {
562 0 : return rc;
563 : }
564 : }
565 :
566 0 : rqpair->expected_evt_type = evt;
567 0 : rqpair->evt_cb = evt_cb;
568 0 : rqpair->evt_timeout_ticks = (g_spdk_nvme_transport_opts.rdma_cm_event_timeout_ms * 1000 *
569 0 : spdk_get_ticks_hz()) / SPDK_SEC_TO_USEC + spdk_get_ticks();
570 :
571 0 : return 0;
572 : }
573 :
574 : static int
575 0 : nvme_rdma_process_event_poll(struct nvme_rdma_qpair *rqpair)
576 : {
577 : struct nvme_rdma_ctrlr *rctrlr;
578 0 : int rc = 0, rc2;
579 :
580 0 : rctrlr = nvme_rdma_ctrlr(rqpair->qpair.ctrlr);
581 0 : assert(rctrlr != NULL);
582 :
583 0 : if (!rqpair->evt && spdk_get_ticks() < rqpair->evt_timeout_ticks) {
584 0 : rc = nvme_rdma_poll_events(rctrlr);
585 0 : if (rc == -EAGAIN || rc == -EWOULDBLOCK) {
586 0 : return rc;
587 : }
588 : }
589 :
590 0 : if (rqpair->evt == NULL) {
591 0 : rc = -EADDRNOTAVAIL;
592 0 : goto exit;
593 : }
594 :
595 0 : rc = nvme_rdma_validate_cm_event(rqpair->expected_evt_type, rqpair->evt);
596 :
597 0 : rc2 = nvme_rdma_qpair_process_cm_event(rqpair);
598 : /* bad message takes precedence over the other error codes from processing the event. */
599 0 : rc = rc == 0 ? rc2 : rc;
600 :
601 0 : exit:
602 0 : assert(rqpair->evt_cb != NULL);
603 0 : return rqpair->evt_cb(rqpair, rc);
604 : }
605 :
606 : static int
607 3 : nvme_rdma_resize_cq(struct nvme_rdma_qpair *rqpair, struct nvme_rdma_poller *poller)
608 : {
609 : int current_num_wc, required_num_wc;
610 : int max_cq_size;
611 :
612 3 : required_num_wc = poller->required_num_wc + WC_PER_QPAIR(rqpair->num_entries);
613 3 : current_num_wc = poller->current_num_wc;
614 3 : if (current_num_wc < required_num_wc) {
615 2 : current_num_wc = spdk_max(current_num_wc * 2, required_num_wc);
616 : }
617 :
618 3 : max_cq_size = g_spdk_nvme_transport_opts.rdma_max_cq_size;
619 3 : if (max_cq_size != 0 && current_num_wc > max_cq_size) {
620 0 : current_num_wc = max_cq_size;
621 : }
622 :
623 3 : if (poller->current_num_wc != current_num_wc) {
624 2 : SPDK_DEBUGLOG(nvme, "Resize RDMA CQ from %d to %d\n", poller->current_num_wc,
625 : current_num_wc);
626 2 : if (ibv_resize_cq(poller->cq, current_num_wc)) {
627 1 : SPDK_ERRLOG("RDMA CQ resize failed: errno %d: %s\n", errno, spdk_strerror(errno));
628 1 : return -1;
629 : }
630 :
631 1 : poller->current_num_wc = current_num_wc;
632 : }
633 :
634 2 : poller->required_num_wc = required_num_wc;
635 2 : return 0;
636 : }
637 :
638 : static int
639 5 : nvme_rdma_qpair_set_poller(struct spdk_nvme_qpair *qpair)
640 : {
641 5 : struct nvme_rdma_qpair *rqpair = nvme_rdma_qpair(qpair);
642 5 : struct nvme_rdma_poll_group *group = nvme_rdma_poll_group(qpair->poll_group);
643 : struct nvme_rdma_poller *poller;
644 :
645 5 : assert(rqpair->cq == NULL);
646 :
647 5 : poller = nvme_rdma_poll_group_get_poller(group, rqpair->cm_id->verbs);
648 5 : if (!poller) {
649 2 : SPDK_ERRLOG("Unable to find a cq for qpair %p on poll group %p\n", qpair, qpair->poll_group);
650 2 : return -EINVAL;
651 : }
652 :
653 3 : if (!poller->srq) {
654 3 : if (nvme_rdma_resize_cq(rqpair, poller)) {
655 1 : nvme_rdma_poll_group_put_poller(group, poller);
656 1 : return -EPROTO;
657 : }
658 : }
659 :
660 2 : rqpair->cq = poller->cq;
661 2 : rqpair->srq = poller->srq;
662 2 : if (rqpair->srq) {
663 0 : rqpair->rsps = poller->rsps;
664 : }
665 2 : rqpair->poller = poller;
666 2 : return 0;
667 : }
668 :
669 : static int
670 1 : nvme_rdma_qpair_init(struct nvme_rdma_qpair *rqpair)
671 : {
672 : int rc;
673 1 : struct spdk_rdma_provider_qp_init_attr attr = {};
674 1 : struct ibv_device_attr dev_attr;
675 : struct nvme_rdma_ctrlr *rctrlr;
676 : uint32_t num_cqe, max_num_cqe;
677 :
678 1 : rc = ibv_query_device(rqpair->cm_id->verbs, &dev_attr);
679 1 : if (rc != 0) {
680 0 : SPDK_ERRLOG("Failed to query RDMA device attributes.\n");
681 0 : return -1;
682 : }
683 :
684 1 : if (rqpair->qpair.poll_group) {
685 0 : assert(!rqpair->cq);
686 0 : rc = nvme_rdma_qpair_set_poller(&rqpair->qpair);
687 0 : if (rc) {
688 0 : SPDK_ERRLOG("Unable to activate the rdmaqpair.\n");
689 0 : return -1;
690 : }
691 0 : assert(rqpair->cq);
692 : } else {
693 1 : num_cqe = rqpair->num_entries * 2;
694 1 : max_num_cqe = g_spdk_nvme_transport_opts.rdma_max_cq_size;
695 1 : if (max_num_cqe != 0 && num_cqe > max_num_cqe) {
696 0 : num_cqe = max_num_cqe;
697 : }
698 1 : rqpair->cq = ibv_create_cq(rqpair->cm_id->verbs, num_cqe, rqpair, NULL, 0);
699 1 : if (!rqpair->cq) {
700 0 : SPDK_ERRLOG("Unable to create completion queue: errno %d: %s\n", errno, spdk_strerror(errno));
701 0 : return -1;
702 : }
703 : }
704 :
705 1 : rctrlr = nvme_rdma_ctrlr(rqpair->qpair.ctrlr);
706 1 : if (g_nvme_hooks.get_ibv_pd) {
707 0 : attr.pd = g_nvme_hooks.get_ibv_pd(&rctrlr->ctrlr.trid, rqpair->cm_id->verbs);
708 : } else {
709 1 : attr.pd = spdk_rdma_utils_get_pd(rqpair->cm_id->verbs);
710 : }
711 :
712 1 : attr.stats = rqpair->poller ? &rqpair->poller->stats.rdma_stats : NULL;
713 1 : attr.send_cq = rqpair->cq;
714 1 : attr.recv_cq = rqpair->cq;
715 1 : attr.cap.max_send_wr = rqpair->num_entries; /* SEND operations */
716 1 : if (rqpair->srq) {
717 0 : attr.srq = rqpair->srq->srq;
718 : } else {
719 1 : attr.cap.max_recv_wr = rqpair->num_entries; /* RECV operations */
720 : }
721 1 : attr.cap.max_send_sge = spdk_min(NVME_RDMA_DEFAULT_TX_SGE, dev_attr.max_sge);
722 1 : attr.cap.max_recv_sge = spdk_min(NVME_RDMA_DEFAULT_RX_SGE, dev_attr.max_sge);
723 :
724 1 : rqpair->rdma_qp = spdk_rdma_provider_qp_create(rqpair->cm_id, &attr);
725 :
726 1 : if (!rqpair->rdma_qp) {
727 0 : return -1;
728 : }
729 :
730 1 : rqpair->memory_domain = spdk_rdma_utils_get_memory_domain(rqpair->rdma_qp->qp->pd);
731 1 : if (!rqpair->memory_domain) {
732 0 : SPDK_ERRLOG("Failed to get memory domain\n");
733 0 : return -1;
734 : }
735 :
736 : /* ibv_create_qp will change the values in attr.cap. Make sure we store the proper value. */
737 1 : rqpair->max_send_sge = spdk_min(NVME_RDMA_DEFAULT_TX_SGE, attr.cap.max_send_sge);
738 1 : rqpair->max_recv_sge = spdk_min(NVME_RDMA_DEFAULT_RX_SGE, attr.cap.max_recv_sge);
739 1 : rqpair->current_num_sends = 0;
740 :
741 1 : rqpair->cm_id->context = rqpair;
742 :
743 1 : return 0;
744 : }
745 :
746 : static void
747 0 : nvme_rdma_reset_failed_sends(struct nvme_rdma_qpair *rqpair,
748 : struct ibv_send_wr *bad_send_wr, int rc)
749 : {
750 0 : SPDK_ERRLOG("Failed to post WRs on send queue, errno %d (%s), bad_wr %p\n",
751 : rc, spdk_strerror(rc), bad_send_wr);
752 0 : while (bad_send_wr != NULL) {
753 0 : assert(rqpair->current_num_sends > 0);
754 0 : rqpair->current_num_sends--;
755 0 : bad_send_wr = bad_send_wr->next;
756 : }
757 0 : }
758 :
759 : static void
760 0 : nvme_rdma_reset_failed_recvs(struct nvme_rdma_rsps *rsps,
761 : struct ibv_recv_wr *bad_recv_wr, int rc)
762 : {
763 0 : SPDK_ERRLOG("Failed to post WRs on receive queue, errno %d (%s), bad_wr %p\n",
764 : rc, spdk_strerror(rc), bad_recv_wr);
765 0 : while (bad_recv_wr != NULL) {
766 0 : assert(rsps->current_num_recvs > 0);
767 0 : rsps->current_num_recvs--;
768 0 : bad_recv_wr = bad_recv_wr->next;
769 : }
770 0 : }
771 :
772 : static inline int
773 1 : nvme_rdma_qpair_submit_sends(struct nvme_rdma_qpair *rqpair)
774 : {
775 1 : struct ibv_send_wr *bad_send_wr = NULL;
776 : int rc;
777 :
778 1 : rc = spdk_rdma_provider_qp_flush_send_wrs(rqpair->rdma_qp, &bad_send_wr);
779 :
780 1 : if (spdk_unlikely(rc)) {
781 0 : nvme_rdma_reset_failed_sends(rqpair, bad_send_wr, rc);
782 : }
783 :
784 1 : return rc;
785 : }
786 :
787 : static inline int
788 0 : nvme_rdma_qpair_submit_recvs(struct nvme_rdma_qpair *rqpair)
789 : {
790 0 : struct ibv_recv_wr *bad_recv_wr;
791 0 : int rc = 0;
792 :
793 0 : rc = spdk_rdma_provider_qp_flush_recv_wrs(rqpair->rdma_qp, &bad_recv_wr);
794 0 : if (spdk_unlikely(rc)) {
795 0 : nvme_rdma_reset_failed_recvs(rqpair->rsps, bad_recv_wr, rc);
796 : }
797 :
798 0 : return rc;
799 : }
800 :
801 : static inline int
802 0 : nvme_rdma_poller_submit_recvs(struct nvme_rdma_poller *poller)
803 : {
804 0 : struct ibv_recv_wr *bad_recv_wr;
805 : int rc;
806 :
807 0 : rc = spdk_rdma_provider_srq_flush_recv_wrs(poller->srq, &bad_recv_wr);
808 0 : if (spdk_unlikely(rc)) {
809 0 : nvme_rdma_reset_failed_recvs(poller->rsps, bad_recv_wr, rc);
810 : }
811 :
812 0 : return rc;
813 : }
814 :
815 : #define nvme_rdma_trace_ibv_sge(sg_list) \
816 : if (sg_list) { \
817 : SPDK_DEBUGLOG(nvme, "local addr %p length 0x%x lkey 0x%x\n", \
818 : (void *)(sg_list)->addr, (sg_list)->length, (sg_list)->lkey); \
819 : }
820 :
821 : static void
822 3 : nvme_rdma_free_rsps(struct nvme_rdma_rsps *rsps)
823 : {
824 3 : if (!rsps) {
825 1 : return;
826 : }
827 :
828 2 : spdk_free(rsps->rsps);
829 2 : spdk_free(rsps->rsp_sgls);
830 2 : spdk_free(rsps->rsp_recv_wrs);
831 2 : spdk_free(rsps);
832 : }
833 :
834 : static struct nvme_rdma_rsps *
835 2 : nvme_rdma_create_rsps(struct nvme_rdma_rsp_opts *opts)
836 : {
837 : struct nvme_rdma_rsps *rsps;
838 2 : struct spdk_rdma_utils_memory_translation translation;
839 : uint16_t i;
840 : int rc;
841 :
842 2 : rsps = spdk_zmalloc(sizeof(*rsps), 0, NULL, SPDK_ENV_SOCKET_ID_ANY, SPDK_MALLOC_DMA);
843 2 : if (!rsps) {
844 0 : SPDK_ERRLOG("Failed to allocate rsps object\n");
845 0 : return NULL;
846 : }
847 :
848 2 : rsps->rsp_sgls = spdk_zmalloc(opts->num_entries * sizeof(*rsps->rsp_sgls), 0, NULL,
849 : SPDK_ENV_SOCKET_ID_ANY, SPDK_MALLOC_DMA);
850 2 : if (!rsps->rsp_sgls) {
851 1 : SPDK_ERRLOG("Failed to allocate rsp_sgls\n");
852 1 : goto fail;
853 : }
854 :
855 1 : rsps->rsp_recv_wrs = spdk_zmalloc(opts->num_entries * sizeof(*rsps->rsp_recv_wrs), 0, NULL,
856 : SPDK_ENV_SOCKET_ID_ANY, SPDK_MALLOC_DMA);
857 1 : if (!rsps->rsp_recv_wrs) {
858 0 : SPDK_ERRLOG("Failed to allocate rsp_recv_wrs\n");
859 0 : goto fail;
860 : }
861 :
862 1 : rsps->rsps = spdk_zmalloc(opts->num_entries * sizeof(*rsps->rsps), 0, NULL,
863 : SPDK_ENV_SOCKET_ID_ANY, SPDK_MALLOC_DMA);
864 1 : if (!rsps->rsps) {
865 0 : SPDK_ERRLOG("can not allocate rdma rsps\n");
866 0 : goto fail;
867 : }
868 :
869 2 : for (i = 0; i < opts->num_entries; i++) {
870 1 : struct ibv_sge *rsp_sgl = &rsps->rsp_sgls[i];
871 1 : struct spdk_nvme_rdma_rsp *rsp = &rsps->rsps[i];
872 1 : struct ibv_recv_wr *recv_wr = &rsps->rsp_recv_wrs[i];
873 :
874 1 : rsp->rqpair = opts->rqpair;
875 1 : rsp->rdma_wr.type = RDMA_WR_TYPE_RECV;
876 1 : rsp->recv_wr = recv_wr;
877 1 : rsp_sgl->addr = (uint64_t)rsp;
878 1 : rsp_sgl->length = sizeof(struct spdk_nvme_cpl);
879 1 : rc = spdk_rdma_utils_get_translation(opts->mr_map, rsp, sizeof(*rsp), &translation);
880 1 : if (rc) {
881 0 : goto fail;
882 : }
883 1 : rsp_sgl->lkey = spdk_rdma_utils_memory_translation_get_lkey(&translation);
884 :
885 1 : recv_wr->wr_id = (uint64_t)&rsp->rdma_wr;
886 1 : recv_wr->next = NULL;
887 1 : recv_wr->sg_list = rsp_sgl;
888 1 : recv_wr->num_sge = 1;
889 :
890 1 : nvme_rdma_trace_ibv_sge(recv_wr->sg_list);
891 :
892 1 : if (opts->rqpair) {
893 1 : spdk_rdma_provider_qp_queue_recv_wrs(opts->rqpair->rdma_qp, recv_wr);
894 : } else {
895 0 : spdk_rdma_provider_srq_queue_recv_wrs(opts->srq, recv_wr);
896 : }
897 : }
898 :
899 1 : rsps->num_entries = opts->num_entries;
900 1 : rsps->current_num_recvs = opts->num_entries;
901 :
902 1 : return rsps;
903 1 : fail:
904 1 : nvme_rdma_free_rsps(rsps);
905 1 : return NULL;
906 : }
907 :
908 : static void
909 3 : nvme_rdma_free_reqs(struct nvme_rdma_qpair *rqpair)
910 : {
911 3 : if (!rqpair->rdma_reqs) {
912 2 : return;
913 : }
914 :
915 1 : spdk_free(rqpair->cmds);
916 1 : rqpair->cmds = NULL;
917 :
918 1 : spdk_free(rqpair->rdma_reqs);
919 1 : rqpair->rdma_reqs = NULL;
920 : }
921 :
922 : static int
923 4 : nvme_rdma_create_reqs(struct nvme_rdma_qpair *rqpair)
924 : {
925 4 : struct spdk_rdma_utils_memory_translation translation;
926 : uint16_t i;
927 : int rc;
928 :
929 4 : assert(!rqpair->rdma_reqs);
930 4 : rqpair->rdma_reqs = spdk_zmalloc(rqpair->num_entries * sizeof(struct spdk_nvme_rdma_req), 0, NULL,
931 : SPDK_ENV_SOCKET_ID_ANY, SPDK_MALLOC_DMA);
932 4 : if (rqpair->rdma_reqs == NULL) {
933 1 : SPDK_ERRLOG("Failed to allocate rdma_reqs\n");
934 1 : goto fail;
935 : }
936 :
937 3 : assert(!rqpair->cmds);
938 3 : rqpair->cmds = spdk_zmalloc(rqpair->num_entries * sizeof(*rqpair->cmds), 0, NULL,
939 : SPDK_ENV_SOCKET_ID_ANY, SPDK_MALLOC_DMA);
940 3 : if (!rqpair->cmds) {
941 0 : SPDK_ERRLOG("Failed to allocate RDMA cmds\n");
942 0 : goto fail;
943 : }
944 :
945 3 : TAILQ_INIT(&rqpair->free_reqs);
946 3 : TAILQ_INIT(&rqpair->outstanding_reqs);
947 10 : for (i = 0; i < rqpair->num_entries; i++) {
948 : struct spdk_nvme_rdma_req *rdma_req;
949 : struct spdk_nvmf_cmd *cmd;
950 :
951 7 : rdma_req = &rqpair->rdma_reqs[i];
952 7 : rdma_req->rdma_wr.type = RDMA_WR_TYPE_SEND;
953 7 : cmd = &rqpair->cmds[i];
954 :
955 7 : rdma_req->id = i;
956 :
957 7 : rc = spdk_rdma_utils_get_translation(rqpair->mr_map, cmd, sizeof(*cmd), &translation);
958 7 : if (rc) {
959 0 : goto fail;
960 : }
961 7 : rdma_req->send_sgl[0].lkey = spdk_rdma_utils_memory_translation_get_lkey(&translation);
962 :
963 : /* The first RDMA sgl element will always point
964 : * at this data structure. Depending on whether
965 : * an NVMe-oF SGL is required, the length of
966 : * this element may change. */
967 7 : rdma_req->send_sgl[0].addr = (uint64_t)cmd;
968 7 : rdma_req->send_wr.wr_id = (uint64_t)&rdma_req->rdma_wr;
969 7 : rdma_req->send_wr.next = NULL;
970 7 : rdma_req->send_wr.opcode = IBV_WR_SEND;
971 7 : rdma_req->send_wr.send_flags = IBV_SEND_SIGNALED;
972 7 : rdma_req->send_wr.sg_list = rdma_req->send_sgl;
973 7 : rdma_req->send_wr.imm_data = 0;
974 :
975 7 : TAILQ_INSERT_TAIL(&rqpair->free_reqs, rdma_req, link);
976 : }
977 :
978 3 : return 0;
979 1 : fail:
980 1 : nvme_rdma_free_reqs(rqpair);
981 1 : return -ENOMEM;
982 : }
983 :
984 : static int nvme_rdma_connect(struct nvme_rdma_qpair *rqpair);
985 :
986 : static int
987 0 : nvme_rdma_route_resolved(struct nvme_rdma_qpair *rqpair, int ret)
988 : {
989 0 : if (ret) {
990 0 : SPDK_ERRLOG("RDMA route resolution error\n");
991 0 : return -1;
992 : }
993 :
994 0 : ret = nvme_rdma_qpair_init(rqpair);
995 0 : if (ret < 0) {
996 0 : SPDK_ERRLOG("nvme_rdma_qpair_init() failed\n");
997 0 : return -1;
998 : }
999 :
1000 0 : return nvme_rdma_connect(rqpair);
1001 : }
1002 :
1003 : static int
1004 0 : nvme_rdma_addr_resolved(struct nvme_rdma_qpair *rqpair, int ret)
1005 : {
1006 0 : if (ret) {
1007 0 : SPDK_ERRLOG("RDMA address resolution error\n");
1008 0 : return -1;
1009 : }
1010 :
1011 0 : if (rqpair->qpair.ctrlr->opts.transport_ack_timeout != SPDK_NVME_TRANSPORT_ACK_TIMEOUT_DISABLED) {
1012 : #ifdef SPDK_CONFIG_RDMA_SET_ACK_TIMEOUT
1013 0 : uint8_t timeout = rqpair->qpair.ctrlr->opts.transport_ack_timeout;
1014 0 : ret = rdma_set_option(rqpair->cm_id, RDMA_OPTION_ID,
1015 : RDMA_OPTION_ID_ACK_TIMEOUT,
1016 : &timeout, sizeof(timeout));
1017 0 : if (ret) {
1018 0 : SPDK_NOTICELOG("Can't apply RDMA_OPTION_ID_ACK_TIMEOUT %d, ret %d\n", timeout, ret);
1019 : }
1020 : #else
1021 : SPDK_DEBUGLOG(nvme, "transport_ack_timeout is not supported\n");
1022 : #endif
1023 : }
1024 :
1025 0 : if (rqpair->qpair.ctrlr->opts.transport_tos != SPDK_NVME_TRANSPORT_TOS_DISABLED) {
1026 : #ifdef SPDK_CONFIG_RDMA_SET_TOS
1027 0 : uint8_t tos = rqpair->qpair.ctrlr->opts.transport_tos;
1028 0 : ret = rdma_set_option(rqpair->cm_id, RDMA_OPTION_ID, RDMA_OPTION_ID_TOS, &tos, sizeof(tos));
1029 0 : if (ret) {
1030 0 : SPDK_NOTICELOG("Can't apply RDMA_OPTION_ID_TOS %u, ret %d\n", tos, ret);
1031 : }
1032 : #else
1033 : SPDK_DEBUGLOG(nvme, "transport_tos is not supported\n");
1034 : #endif
1035 : }
1036 :
1037 0 : ret = rdma_resolve_route(rqpair->cm_id, NVME_RDMA_TIME_OUT_IN_MS);
1038 0 : if (ret) {
1039 0 : SPDK_ERRLOG("rdma_resolve_route\n");
1040 0 : return ret;
1041 : }
1042 :
1043 0 : return nvme_rdma_process_event_start(rqpair, RDMA_CM_EVENT_ROUTE_RESOLVED,
1044 : nvme_rdma_route_resolved);
1045 : }
1046 :
1047 : static int
1048 0 : nvme_rdma_resolve_addr(struct nvme_rdma_qpair *rqpair,
1049 : struct sockaddr *src_addr,
1050 : struct sockaddr *dst_addr)
1051 : {
1052 : int ret;
1053 :
1054 0 : if (src_addr) {
1055 0 : int reuse = 1;
1056 :
1057 0 : ret = rdma_set_option(rqpair->cm_id, RDMA_OPTION_ID, RDMA_OPTION_ID_REUSEADDR,
1058 : &reuse, sizeof(reuse));
1059 0 : if (ret) {
1060 0 : SPDK_NOTICELOG("Can't apply RDMA_OPTION_ID_REUSEADDR %d, ret %d\n",
1061 : reuse, ret);
1062 : /* It is likely that rdma_resolve_addr() returns -EADDRINUSE, but
1063 : * we may missing something. We rely on rdma_resolve_addr().
1064 : */
1065 : }
1066 : }
1067 :
1068 0 : ret = rdma_resolve_addr(rqpair->cm_id, src_addr, dst_addr,
1069 : NVME_RDMA_TIME_OUT_IN_MS);
1070 0 : if (ret) {
1071 0 : SPDK_ERRLOG("rdma_resolve_addr, %d\n", errno);
1072 0 : return ret;
1073 : }
1074 :
1075 0 : return nvme_rdma_process_event_start(rqpair, RDMA_CM_EVENT_ADDR_RESOLVED,
1076 : nvme_rdma_addr_resolved);
1077 : }
1078 :
1079 : static int nvme_rdma_stale_conn_retry(struct nvme_rdma_qpair *rqpair);
1080 :
1081 : static int
1082 0 : nvme_rdma_connect_established(struct nvme_rdma_qpair *rqpair, int ret)
1083 : {
1084 0 : struct nvme_rdma_rsp_opts opts = {};
1085 :
1086 0 : if (ret == -ESTALE) {
1087 0 : return nvme_rdma_stale_conn_retry(rqpair);
1088 0 : } else if (ret) {
1089 0 : SPDK_ERRLOG("RDMA connect error %d\n", ret);
1090 0 : return ret;
1091 : }
1092 :
1093 0 : assert(!rqpair->mr_map);
1094 0 : rqpair->mr_map = spdk_rdma_utils_create_mem_map(rqpair->rdma_qp->qp->pd, &g_nvme_hooks,
1095 : IBV_ACCESS_LOCAL_WRITE | IBV_ACCESS_REMOTE_READ | IBV_ACCESS_REMOTE_WRITE);
1096 0 : if (!rqpair->mr_map) {
1097 0 : SPDK_ERRLOG("Unable to register RDMA memory translation map\n");
1098 0 : return -1;
1099 : }
1100 :
1101 0 : ret = nvme_rdma_create_reqs(rqpair);
1102 0 : SPDK_DEBUGLOG(nvme, "rc =%d\n", ret);
1103 0 : if (ret) {
1104 0 : SPDK_ERRLOG("Unable to create rqpair RDMA requests\n");
1105 0 : return -1;
1106 : }
1107 0 : SPDK_DEBUGLOG(nvme, "RDMA requests created\n");
1108 :
1109 0 : if (!rqpair->srq) {
1110 0 : opts.num_entries = rqpair->num_entries;
1111 0 : opts.rqpair = rqpair;
1112 0 : opts.srq = NULL;
1113 0 : opts.mr_map = rqpair->mr_map;
1114 :
1115 0 : assert(!rqpair->rsps);
1116 0 : rqpair->rsps = nvme_rdma_create_rsps(&opts);
1117 0 : if (!rqpair->rsps) {
1118 0 : SPDK_ERRLOG("Unable to create rqpair RDMA responses\n");
1119 0 : return -1;
1120 : }
1121 0 : SPDK_DEBUGLOG(nvme, "RDMA responses created\n");
1122 :
1123 0 : ret = nvme_rdma_qpair_submit_recvs(rqpair);
1124 0 : SPDK_DEBUGLOG(nvme, "rc =%d\n", ret);
1125 0 : if (ret) {
1126 0 : SPDK_ERRLOG("Unable to submit rqpair RDMA responses\n");
1127 0 : return -1;
1128 : }
1129 0 : SPDK_DEBUGLOG(nvme, "RDMA responses submitted\n");
1130 : }
1131 :
1132 0 : rqpair->state = NVME_RDMA_QPAIR_STATE_FABRIC_CONNECT_SEND;
1133 :
1134 0 : return 0;
1135 : }
1136 :
1137 : static int
1138 0 : nvme_rdma_connect(struct nvme_rdma_qpair *rqpair)
1139 : {
1140 0 : struct rdma_conn_param param = {};
1141 0 : struct spdk_nvmf_rdma_request_private_data request_data = {};
1142 0 : struct ibv_device_attr attr;
1143 : int ret;
1144 : struct spdk_nvme_ctrlr *ctrlr;
1145 : struct sockaddr *sa;
1146 0 : uint32_t numa_socket_id;
1147 0 : char host[64];
1148 0 : char ifc[64];
1149 :
1150 0 : ret = ibv_query_device(rqpair->cm_id->verbs, &attr);
1151 0 : if (ret != 0) {
1152 0 : SPDK_ERRLOG("Failed to query RDMA device attributes.\n");
1153 0 : return ret;
1154 : }
1155 :
1156 0 : param.responder_resources = attr.max_qp_rd_atom;
1157 :
1158 0 : ctrlr = rqpair->qpair.ctrlr;
1159 0 : if (!ctrlr) {
1160 0 : return -1;
1161 : }
1162 :
1163 0 : request_data.qid = rqpair->qpair.id;
1164 0 : request_data.hrqsize = rqpair->num_entries + 1;
1165 0 : request_data.hsqsize = rqpair->num_entries;
1166 0 : request_data.cntlid = ctrlr->cntlid;
1167 :
1168 0 : param.private_data = &request_data;
1169 0 : param.private_data_len = sizeof(request_data);
1170 0 : param.retry_count = ctrlr->opts.transport_retry_count;
1171 0 : param.rnr_retry_count = 7;
1172 :
1173 : /* Fields below are ignored by rdma cm if qpair has been
1174 : * created using rdma cm API. */
1175 0 : param.srq = 0;
1176 0 : param.qp_num = rqpair->rdma_qp->qp->qp_num;
1177 :
1178 0 : ret = rdma_connect(rqpair->cm_id, ¶m);
1179 0 : if (ret) {
1180 0 : SPDK_ERRLOG("nvme rdma connect error\n");
1181 0 : return ret;
1182 : }
1183 :
1184 0 : sa = rdma_get_local_addr(rqpair->cm_id);
1185 0 : if (sa == NULL) {
1186 0 : goto out;
1187 : }
1188 0 : ret = spdk_net_get_address_string(sa, host, sizeof(host));
1189 0 : if (ret != 0) {
1190 0 : goto out;
1191 : }
1192 0 : ret = spdk_net_get_interface_name(host, ifc, sizeof(ifc));
1193 0 : if (ret != 0) {
1194 0 : goto out;
1195 : }
1196 0 : ret = spdk_read_sysfs_attribute_uint32(&numa_socket_id,
1197 : "/sys/class/net/%s/device/numa_node", ifc);
1198 0 : if (ret == 0) {
1199 0 : ctrlr->socket_id_valid = true;
1200 0 : ctrlr->socket_id = numa_socket_id;
1201 : }
1202 :
1203 0 : out:
1204 0 : return nvme_rdma_process_event_start(rqpair, RDMA_CM_EVENT_ESTABLISHED,
1205 : nvme_rdma_connect_established);
1206 : }
1207 :
1208 : static int
1209 0 : nvme_rdma_ctrlr_connect_qpair(struct spdk_nvme_ctrlr *ctrlr, struct spdk_nvme_qpair *qpair)
1210 : {
1211 0 : struct sockaddr_storage dst_addr;
1212 0 : struct sockaddr_storage src_addr;
1213 : bool src_addr_specified;
1214 0 : long int port, src_port;
1215 : int rc;
1216 : struct nvme_rdma_ctrlr *rctrlr;
1217 : struct nvme_rdma_qpair *rqpair;
1218 : struct nvme_rdma_poll_group *group;
1219 : int family;
1220 :
1221 0 : rqpair = nvme_rdma_qpair(qpair);
1222 0 : rctrlr = nvme_rdma_ctrlr(ctrlr);
1223 0 : assert(rctrlr != NULL);
1224 :
1225 0 : switch (ctrlr->trid.adrfam) {
1226 0 : case SPDK_NVMF_ADRFAM_IPV4:
1227 0 : family = AF_INET;
1228 0 : break;
1229 0 : case SPDK_NVMF_ADRFAM_IPV6:
1230 0 : family = AF_INET6;
1231 0 : break;
1232 0 : default:
1233 0 : SPDK_ERRLOG("Unhandled ADRFAM %d\n", ctrlr->trid.adrfam);
1234 0 : return -1;
1235 : }
1236 :
1237 0 : SPDK_DEBUGLOG(nvme, "adrfam %d ai_family %d\n", ctrlr->trid.adrfam, family);
1238 :
1239 0 : memset(&dst_addr, 0, sizeof(dst_addr));
1240 :
1241 0 : SPDK_DEBUGLOG(nvme, "trsvcid is %s\n", ctrlr->trid.trsvcid);
1242 0 : rc = nvme_parse_addr(&dst_addr, family, ctrlr->trid.traddr, ctrlr->trid.trsvcid, &port);
1243 0 : if (rc != 0) {
1244 0 : SPDK_ERRLOG("dst_addr nvme_parse_addr() failed\n");
1245 0 : return -1;
1246 : }
1247 :
1248 0 : if (ctrlr->opts.src_addr[0] || ctrlr->opts.src_svcid[0]) {
1249 0 : memset(&src_addr, 0, sizeof(src_addr));
1250 0 : rc = nvme_parse_addr(&src_addr, family, ctrlr->opts.src_addr, ctrlr->opts.src_svcid, &src_port);
1251 0 : if (rc != 0) {
1252 0 : SPDK_ERRLOG("src_addr nvme_parse_addr() failed\n");
1253 0 : return -1;
1254 : }
1255 0 : src_addr_specified = true;
1256 : } else {
1257 0 : src_addr_specified = false;
1258 : }
1259 :
1260 0 : rc = rdma_create_id(rctrlr->cm_channel, &rqpair->cm_id, rqpair, RDMA_PS_TCP);
1261 0 : if (rc < 0) {
1262 0 : SPDK_ERRLOG("rdma_create_id() failed\n");
1263 0 : return -1;
1264 : }
1265 :
1266 0 : rc = nvme_rdma_resolve_addr(rqpair,
1267 : src_addr_specified ? (struct sockaddr *)&src_addr : NULL,
1268 : (struct sockaddr *)&dst_addr);
1269 0 : if (rc < 0) {
1270 0 : SPDK_ERRLOG("nvme_rdma_resolve_addr() failed\n");
1271 0 : return -1;
1272 : }
1273 :
1274 0 : rqpair->state = NVME_RDMA_QPAIR_STATE_INITIALIZING;
1275 :
1276 0 : if (qpair->poll_group != NULL) {
1277 0 : group = nvme_rdma_poll_group(qpair->poll_group);
1278 0 : TAILQ_INSERT_TAIL(&group->connecting_qpairs, rqpair, link_connecting);
1279 : }
1280 :
1281 0 : return 0;
1282 : }
1283 :
1284 : static int
1285 0 : nvme_rdma_stale_conn_reconnect(struct nvme_rdma_qpair *rqpair)
1286 : {
1287 0 : struct spdk_nvme_qpair *qpair = &rqpair->qpair;
1288 :
1289 0 : if (spdk_get_ticks() < rqpair->evt_timeout_ticks) {
1290 0 : return -EAGAIN;
1291 : }
1292 :
1293 0 : return nvme_rdma_ctrlr_connect_qpair(qpair->ctrlr, qpair);
1294 : }
1295 :
1296 : static int
1297 0 : nvme_rdma_ctrlr_connect_qpair_poll(struct spdk_nvme_ctrlr *ctrlr,
1298 : struct spdk_nvme_qpair *qpair)
1299 : {
1300 0 : struct nvme_rdma_qpair *rqpair = nvme_rdma_qpair(qpair);
1301 : int rc;
1302 :
1303 0 : if (rqpair->in_connect_poll) {
1304 0 : return -EAGAIN;
1305 : }
1306 :
1307 0 : rqpair->in_connect_poll = true;
1308 :
1309 0 : switch (rqpair->state) {
1310 0 : case NVME_RDMA_QPAIR_STATE_INVALID:
1311 0 : rc = -EAGAIN;
1312 0 : break;
1313 :
1314 0 : case NVME_RDMA_QPAIR_STATE_INITIALIZING:
1315 : case NVME_RDMA_QPAIR_STATE_EXITING:
1316 0 : if (!nvme_qpair_is_admin_queue(qpair)) {
1317 0 : nvme_ctrlr_lock(ctrlr);
1318 : }
1319 :
1320 0 : rc = nvme_rdma_process_event_poll(rqpair);
1321 :
1322 0 : if (!nvme_qpair_is_admin_queue(qpair)) {
1323 0 : nvme_ctrlr_unlock(ctrlr);
1324 : }
1325 :
1326 0 : if (rc == 0) {
1327 0 : rc = -EAGAIN;
1328 : }
1329 0 : rqpair->in_connect_poll = false;
1330 :
1331 0 : return rc;
1332 :
1333 0 : case NVME_RDMA_QPAIR_STATE_STALE_CONN:
1334 0 : rc = nvme_rdma_stale_conn_reconnect(rqpair);
1335 0 : if (rc == 0) {
1336 0 : rc = -EAGAIN;
1337 : }
1338 0 : break;
1339 0 : case NVME_RDMA_QPAIR_STATE_FABRIC_CONNECT_SEND:
1340 0 : rc = nvme_fabric_qpair_connect_async(qpair, rqpair->num_entries + 1);
1341 0 : if (rc == 0) {
1342 0 : rqpair->state = NVME_RDMA_QPAIR_STATE_FABRIC_CONNECT_POLL;
1343 0 : rc = -EAGAIN;
1344 : } else {
1345 0 : SPDK_ERRLOG("Failed to send an NVMe-oF Fabric CONNECT command\n");
1346 : }
1347 0 : break;
1348 0 : case NVME_RDMA_QPAIR_STATE_FABRIC_CONNECT_POLL:
1349 0 : rc = nvme_fabric_qpair_connect_poll(qpair);
1350 0 : if (rc == 0) {
1351 0 : rqpair->state = NVME_RDMA_QPAIR_STATE_RUNNING;
1352 0 : nvme_qpair_set_state(qpair, NVME_QPAIR_CONNECTED);
1353 0 : } else if (rc != -EAGAIN) {
1354 0 : SPDK_ERRLOG("Failed to poll NVMe-oF Fabric CONNECT command\n");
1355 : }
1356 0 : break;
1357 0 : case NVME_RDMA_QPAIR_STATE_RUNNING:
1358 0 : rc = 0;
1359 0 : break;
1360 0 : default:
1361 0 : assert(false);
1362 : rc = -EINVAL;
1363 : break;
1364 : }
1365 :
1366 0 : rqpair->in_connect_poll = false;
1367 :
1368 0 : return rc;
1369 : }
1370 :
1371 : static inline int
1372 28 : nvme_rdma_get_memory_translation(struct nvme_request *req, struct nvme_rdma_qpair *rqpair,
1373 : struct nvme_rdma_memory_translation_ctx *_ctx)
1374 : {
1375 28 : struct spdk_memory_domain_translation_ctx ctx;
1376 28 : struct spdk_memory_domain_translation_result dma_translation = {.iov_count = 0};
1377 28 : struct spdk_rdma_utils_memory_translation rdma_translation;
1378 : int rc;
1379 :
1380 28 : assert(req);
1381 28 : assert(rqpair);
1382 28 : assert(_ctx);
1383 :
1384 28 : if (req->payload.opts && req->payload.opts->memory_domain) {
1385 2 : ctx.size = sizeof(struct spdk_memory_domain_translation_ctx);
1386 2 : ctx.rdma.ibv_qp = rqpair->rdma_qp->qp;
1387 2 : dma_translation.size = sizeof(struct spdk_memory_domain_translation_result);
1388 :
1389 2 : rc = spdk_memory_domain_translate_data(req->payload.opts->memory_domain,
1390 2 : req->payload.opts->memory_domain_ctx,
1391 : rqpair->memory_domain, &ctx, _ctx->addr,
1392 : _ctx->length, &dma_translation);
1393 2 : if (spdk_unlikely(rc) || dma_translation.iov_count != 1) {
1394 1 : SPDK_ERRLOG("DMA memory translation failed, rc %d, iov count %u\n", rc, dma_translation.iov_count);
1395 1 : return rc;
1396 : }
1397 :
1398 1 : _ctx->lkey = dma_translation.rdma.lkey;
1399 1 : _ctx->rkey = dma_translation.rdma.rkey;
1400 1 : _ctx->addr = dma_translation.iov.iov_base;
1401 1 : _ctx->length = dma_translation.iov.iov_len;
1402 : } else {
1403 26 : rc = spdk_rdma_utils_get_translation(rqpair->mr_map, _ctx->addr, _ctx->length, &rdma_translation);
1404 26 : if (spdk_unlikely(rc)) {
1405 2 : SPDK_ERRLOG("RDMA memory translation failed, rc %d\n", rc);
1406 2 : return rc;
1407 : }
1408 24 : if (rdma_translation.translation_type == SPDK_RDMA_UTILS_TRANSLATION_MR) {
1409 24 : _ctx->lkey = rdma_translation.mr_or_key.mr->lkey;
1410 24 : _ctx->rkey = rdma_translation.mr_or_key.mr->rkey;
1411 : } else {
1412 0 : _ctx->lkey = _ctx->rkey = (uint32_t)rdma_translation.mr_or_key.key;
1413 : }
1414 : }
1415 :
1416 25 : return 0;
1417 : }
1418 :
1419 :
1420 : /*
1421 : * Build SGL describing empty payload.
1422 : */
1423 : static int
1424 2 : nvme_rdma_build_null_request(struct spdk_nvme_rdma_req *rdma_req)
1425 : {
1426 2 : struct nvme_request *req = rdma_req->req;
1427 :
1428 2 : req->cmd.psdt = SPDK_NVME_PSDT_SGL_MPTR_CONTIG;
1429 :
1430 : /* The first element of this SGL is pointing at an
1431 : * spdk_nvmf_cmd object. For this particular command,
1432 : * we only need the first 64 bytes corresponding to
1433 : * the NVMe command. */
1434 2 : rdma_req->send_sgl[0].length = sizeof(struct spdk_nvme_cmd);
1435 :
1436 : /* The RDMA SGL needs one element describing the NVMe command. */
1437 2 : rdma_req->send_wr.num_sge = 1;
1438 :
1439 2 : req->cmd.dptr.sgl1.keyed.type = SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK;
1440 2 : req->cmd.dptr.sgl1.keyed.subtype = SPDK_NVME_SGL_SUBTYPE_ADDRESS;
1441 2 : req->cmd.dptr.sgl1.keyed.length = 0;
1442 2 : req->cmd.dptr.sgl1.keyed.key = 0;
1443 2 : req->cmd.dptr.sgl1.address = 0;
1444 :
1445 2 : return 0;
1446 : }
1447 :
1448 : /*
1449 : * Build inline SGL describing contiguous payload buffer.
1450 : */
1451 : static int
1452 3 : nvme_rdma_build_contig_inline_request(struct nvme_rdma_qpair *rqpair,
1453 : struct spdk_nvme_rdma_req *rdma_req)
1454 : {
1455 3 : struct nvme_request *req = rdma_req->req;
1456 3 : struct nvme_rdma_memory_translation_ctx ctx = {
1457 3 : .addr = (uint8_t *)req->payload.contig_or_cb_arg + req->payload_offset,
1458 3 : .length = req->payload_size
1459 : };
1460 : int rc;
1461 :
1462 3 : assert(ctx.length != 0);
1463 3 : assert(nvme_payload_type(&req->payload) == NVME_PAYLOAD_TYPE_CONTIG);
1464 :
1465 3 : rc = nvme_rdma_get_memory_translation(req, rqpair, &ctx);
1466 3 : if (spdk_unlikely(rc)) {
1467 0 : return -1;
1468 : }
1469 :
1470 3 : rdma_req->send_sgl[1].lkey = ctx.lkey;
1471 :
1472 : /* The first element of this SGL is pointing at an
1473 : * spdk_nvmf_cmd object. For this particular command,
1474 : * we only need the first 64 bytes corresponding to
1475 : * the NVMe command. */
1476 3 : rdma_req->send_sgl[0].length = sizeof(struct spdk_nvme_cmd);
1477 :
1478 3 : rdma_req->send_sgl[1].addr = (uint64_t)ctx.addr;
1479 3 : rdma_req->send_sgl[1].length = (uint32_t)ctx.length;
1480 :
1481 : /* The RDMA SGL contains two elements. The first describes
1482 : * the NVMe command and the second describes the data
1483 : * payload. */
1484 3 : rdma_req->send_wr.num_sge = 2;
1485 :
1486 3 : req->cmd.psdt = SPDK_NVME_PSDT_SGL_MPTR_CONTIG;
1487 3 : req->cmd.dptr.sgl1.unkeyed.type = SPDK_NVME_SGL_TYPE_DATA_BLOCK;
1488 3 : req->cmd.dptr.sgl1.unkeyed.subtype = SPDK_NVME_SGL_SUBTYPE_OFFSET;
1489 3 : req->cmd.dptr.sgl1.unkeyed.length = (uint32_t)ctx.length;
1490 : /* Inline only supported for icdoff == 0 currently. This function will
1491 : * not get called for controllers with other values. */
1492 3 : req->cmd.dptr.sgl1.address = (uint64_t)0;
1493 :
1494 3 : return 0;
1495 : }
1496 :
1497 : /*
1498 : * Build SGL describing contiguous payload buffer.
1499 : */
1500 : static int
1501 3 : nvme_rdma_build_contig_request(struct nvme_rdma_qpair *rqpair,
1502 : struct spdk_nvme_rdma_req *rdma_req)
1503 : {
1504 3 : struct nvme_request *req = rdma_req->req;
1505 3 : struct nvme_rdma_memory_translation_ctx ctx = {
1506 3 : .addr = (uint8_t *)req->payload.contig_or_cb_arg + req->payload_offset,
1507 3 : .length = req->payload_size
1508 : };
1509 : int rc;
1510 :
1511 3 : assert(req->payload_size != 0);
1512 3 : assert(nvme_payload_type(&req->payload) == NVME_PAYLOAD_TYPE_CONTIG);
1513 :
1514 3 : if (spdk_unlikely(req->payload_size > NVME_RDMA_MAX_KEYED_SGL_LENGTH)) {
1515 1 : SPDK_ERRLOG("SGL length %u exceeds max keyed SGL block size %u\n",
1516 : req->payload_size, NVME_RDMA_MAX_KEYED_SGL_LENGTH);
1517 1 : return -1;
1518 : }
1519 :
1520 2 : rc = nvme_rdma_get_memory_translation(req, rqpair, &ctx);
1521 2 : if (spdk_unlikely(rc)) {
1522 0 : return -1;
1523 : }
1524 :
1525 2 : req->cmd.dptr.sgl1.keyed.key = ctx.rkey;
1526 :
1527 : /* The first element of this SGL is pointing at an
1528 : * spdk_nvmf_cmd object. For this particular command,
1529 : * we only need the first 64 bytes corresponding to
1530 : * the NVMe command. */
1531 2 : rdma_req->send_sgl[0].length = sizeof(struct spdk_nvme_cmd);
1532 :
1533 : /* The RDMA SGL needs one element describing the NVMe command. */
1534 2 : rdma_req->send_wr.num_sge = 1;
1535 :
1536 2 : req->cmd.psdt = SPDK_NVME_PSDT_SGL_MPTR_CONTIG;
1537 2 : req->cmd.dptr.sgl1.keyed.type = SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK;
1538 2 : req->cmd.dptr.sgl1.keyed.subtype = SPDK_NVME_SGL_SUBTYPE_ADDRESS;
1539 2 : req->cmd.dptr.sgl1.keyed.length = (uint32_t)ctx.length;
1540 2 : req->cmd.dptr.sgl1.address = (uint64_t)ctx.addr;
1541 :
1542 2 : return 0;
1543 : }
1544 :
1545 : /*
1546 : * Build SGL describing scattered payload buffer.
1547 : */
1548 : static int
1549 7 : nvme_rdma_build_sgl_request(struct nvme_rdma_qpair *rqpair,
1550 : struct spdk_nvme_rdma_req *rdma_req)
1551 : {
1552 7 : struct nvme_request *req = rdma_req->req;
1553 7 : struct spdk_nvmf_cmd *cmd = &rqpair->cmds[rdma_req->id];
1554 7 : struct nvme_rdma_memory_translation_ctx ctx;
1555 : uint32_t remaining_size;
1556 7 : uint32_t sge_length;
1557 : int rc, max_num_sgl, num_sgl_desc;
1558 :
1559 7 : assert(req->payload_size != 0);
1560 7 : assert(nvme_payload_type(&req->payload) == NVME_PAYLOAD_TYPE_SGL);
1561 7 : assert(req->payload.reset_sgl_fn != NULL);
1562 7 : assert(req->payload.next_sge_fn != NULL);
1563 7 : req->payload.reset_sgl_fn(req->payload.contig_or_cb_arg, req->payload_offset);
1564 :
1565 7 : max_num_sgl = req->qpair->ctrlr->max_sges;
1566 :
1567 7 : remaining_size = req->payload_size;
1568 7 : num_sgl_desc = 0;
1569 : do {
1570 18 : rc = req->payload.next_sge_fn(req->payload.contig_or_cb_arg, &ctx.addr, &sge_length);
1571 18 : if (rc) {
1572 1 : return -1;
1573 : }
1574 :
1575 17 : sge_length = spdk_min(remaining_size, sge_length);
1576 :
1577 17 : if (spdk_unlikely(sge_length > NVME_RDMA_MAX_KEYED_SGL_LENGTH)) {
1578 1 : SPDK_ERRLOG("SGL length %u exceeds max keyed SGL block size %u\n",
1579 : sge_length, NVME_RDMA_MAX_KEYED_SGL_LENGTH);
1580 1 : return -1;
1581 : }
1582 16 : ctx.length = sge_length;
1583 16 : rc = nvme_rdma_get_memory_translation(req, rqpair, &ctx);
1584 16 : if (spdk_unlikely(rc)) {
1585 1 : return -1;
1586 : }
1587 :
1588 15 : cmd->sgl[num_sgl_desc].keyed.key = ctx.rkey;
1589 15 : cmd->sgl[num_sgl_desc].keyed.type = SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK;
1590 15 : cmd->sgl[num_sgl_desc].keyed.subtype = SPDK_NVME_SGL_SUBTYPE_ADDRESS;
1591 15 : cmd->sgl[num_sgl_desc].keyed.length = (uint32_t)ctx.length;
1592 15 : cmd->sgl[num_sgl_desc].address = (uint64_t)ctx.addr;
1593 :
1594 15 : remaining_size -= ctx.length;
1595 15 : num_sgl_desc++;
1596 15 : } while (remaining_size > 0 && num_sgl_desc < max_num_sgl);
1597 :
1598 :
1599 : /* Should be impossible if we did our sgl checks properly up the stack, but do a sanity check here. */
1600 4 : if (remaining_size > 0) {
1601 0 : return -1;
1602 : }
1603 :
1604 4 : req->cmd.psdt = SPDK_NVME_PSDT_SGL_MPTR_CONTIG;
1605 :
1606 : /* The RDMA SGL needs one element describing some portion
1607 : * of the spdk_nvmf_cmd structure. */
1608 4 : rdma_req->send_wr.num_sge = 1;
1609 :
1610 : /*
1611 : * If only one SGL descriptor is required, it can be embedded directly in the command
1612 : * as a data block descriptor.
1613 : */
1614 4 : if (num_sgl_desc == 1) {
1615 : /* The first element of this SGL is pointing at an
1616 : * spdk_nvmf_cmd object. For this particular command,
1617 : * we only need the first 64 bytes corresponding to
1618 : * the NVMe command. */
1619 2 : rdma_req->send_sgl[0].length = sizeof(struct spdk_nvme_cmd);
1620 :
1621 2 : req->cmd.dptr.sgl1.keyed.type = cmd->sgl[0].keyed.type;
1622 2 : req->cmd.dptr.sgl1.keyed.subtype = cmd->sgl[0].keyed.subtype;
1623 2 : req->cmd.dptr.sgl1.keyed.length = cmd->sgl[0].keyed.length;
1624 2 : req->cmd.dptr.sgl1.keyed.key = cmd->sgl[0].keyed.key;
1625 2 : req->cmd.dptr.sgl1.address = cmd->sgl[0].address;
1626 : } else {
1627 : /*
1628 : * Otherwise, The SGL descriptor embedded in the command must point to the list of
1629 : * SGL descriptors used to describe the operation. In that case it is a last segment descriptor.
1630 : */
1631 2 : uint32_t descriptors_size = sizeof(struct spdk_nvme_sgl_descriptor) * num_sgl_desc;
1632 :
1633 2 : if (spdk_unlikely(descriptors_size > rqpair->qpair.ctrlr->ioccsz_bytes)) {
1634 1 : SPDK_ERRLOG("Size of SGL descriptors (%u) exceeds ICD (%u)\n",
1635 : descriptors_size, rqpair->qpair.ctrlr->ioccsz_bytes);
1636 1 : return -1;
1637 : }
1638 1 : rdma_req->send_sgl[0].length = sizeof(struct spdk_nvme_cmd) + descriptors_size;
1639 :
1640 1 : req->cmd.dptr.sgl1.unkeyed.type = SPDK_NVME_SGL_TYPE_LAST_SEGMENT;
1641 1 : req->cmd.dptr.sgl1.unkeyed.subtype = SPDK_NVME_SGL_SUBTYPE_OFFSET;
1642 1 : req->cmd.dptr.sgl1.unkeyed.length = descriptors_size;
1643 1 : req->cmd.dptr.sgl1.address = (uint64_t)0;
1644 : }
1645 :
1646 3 : return 0;
1647 : }
1648 :
1649 : /*
1650 : * Build inline SGL describing sgl payload buffer.
1651 : */
1652 : static int
1653 3 : nvme_rdma_build_sgl_inline_request(struct nvme_rdma_qpair *rqpair,
1654 : struct spdk_nvme_rdma_req *rdma_req)
1655 : {
1656 3 : struct nvme_request *req = rdma_req->req;
1657 3 : struct nvme_rdma_memory_translation_ctx ctx;
1658 3 : uint32_t length;
1659 : int rc;
1660 :
1661 3 : assert(req->payload_size != 0);
1662 3 : assert(nvme_payload_type(&req->payload) == NVME_PAYLOAD_TYPE_SGL);
1663 3 : assert(req->payload.reset_sgl_fn != NULL);
1664 3 : assert(req->payload.next_sge_fn != NULL);
1665 3 : req->payload.reset_sgl_fn(req->payload.contig_or_cb_arg, req->payload_offset);
1666 :
1667 3 : rc = req->payload.next_sge_fn(req->payload.contig_or_cb_arg, &ctx.addr, &length);
1668 3 : if (rc) {
1669 0 : return -1;
1670 : }
1671 :
1672 3 : if (length < req->payload_size) {
1673 0 : SPDK_DEBUGLOG(nvme, "Inline SGL request split so sending separately.\n");
1674 0 : return nvme_rdma_build_sgl_request(rqpair, rdma_req);
1675 : }
1676 :
1677 3 : if (length > req->payload_size) {
1678 0 : length = req->payload_size;
1679 : }
1680 :
1681 3 : ctx.length = length;
1682 3 : rc = nvme_rdma_get_memory_translation(req, rqpair, &ctx);
1683 3 : if (spdk_unlikely(rc)) {
1684 0 : return -1;
1685 : }
1686 :
1687 3 : rdma_req->send_sgl[1].addr = (uint64_t)ctx.addr;
1688 3 : rdma_req->send_sgl[1].length = (uint32_t)ctx.length;
1689 3 : rdma_req->send_sgl[1].lkey = ctx.lkey;
1690 :
1691 3 : rdma_req->send_wr.num_sge = 2;
1692 :
1693 : /* The first element of this SGL is pointing at an
1694 : * spdk_nvmf_cmd object. For this particular command,
1695 : * we only need the first 64 bytes corresponding to
1696 : * the NVMe command. */
1697 3 : rdma_req->send_sgl[0].length = sizeof(struct spdk_nvme_cmd);
1698 :
1699 3 : req->cmd.psdt = SPDK_NVME_PSDT_SGL_MPTR_CONTIG;
1700 3 : req->cmd.dptr.sgl1.unkeyed.type = SPDK_NVME_SGL_TYPE_DATA_BLOCK;
1701 3 : req->cmd.dptr.sgl1.unkeyed.subtype = SPDK_NVME_SGL_SUBTYPE_OFFSET;
1702 3 : req->cmd.dptr.sgl1.unkeyed.length = (uint32_t)ctx.length;
1703 : /* Inline only supported for icdoff == 0 currently. This function will
1704 : * not get called for controllers with other values. */
1705 3 : req->cmd.dptr.sgl1.address = (uint64_t)0;
1706 :
1707 3 : return 0;
1708 : }
1709 :
1710 : static int
1711 6 : nvme_rdma_req_init(struct nvme_rdma_qpair *rqpair, struct nvme_request *req,
1712 : struct spdk_nvme_rdma_req *rdma_req)
1713 : {
1714 6 : struct spdk_nvme_ctrlr *ctrlr = rqpair->qpair.ctrlr;
1715 : enum nvme_payload_type payload_type;
1716 : bool icd_supported;
1717 : int rc;
1718 :
1719 6 : assert(rdma_req->req == NULL);
1720 6 : rdma_req->req = req;
1721 6 : req->cmd.cid = rdma_req->id;
1722 6 : payload_type = nvme_payload_type(&req->payload);
1723 : /*
1724 : * Check if icdoff is non zero, to avoid interop conflicts with
1725 : * targets with non-zero icdoff. Both SPDK and the Linux kernel
1726 : * targets use icdoff = 0. For targets with non-zero icdoff, we
1727 : * will currently just not use inline data for now.
1728 : */
1729 6 : icd_supported = spdk_nvme_opc_get_data_transfer(req->cmd.opc) == SPDK_NVME_DATA_HOST_TO_CONTROLLER
1730 6 : && req->payload_size <= ctrlr->ioccsz_bytes && ctrlr->icdoff == 0;
1731 :
1732 6 : if (req->payload_size == 0) {
1733 2 : rc = nvme_rdma_build_null_request(rdma_req);
1734 4 : } else if (payload_type == NVME_PAYLOAD_TYPE_CONTIG) {
1735 2 : if (icd_supported) {
1736 1 : rc = nvme_rdma_build_contig_inline_request(rqpair, rdma_req);
1737 : } else {
1738 1 : rc = nvme_rdma_build_contig_request(rqpair, rdma_req);
1739 : }
1740 2 : } else if (payload_type == NVME_PAYLOAD_TYPE_SGL) {
1741 2 : if (icd_supported) {
1742 1 : rc = nvme_rdma_build_sgl_inline_request(rqpair, rdma_req);
1743 : } else {
1744 1 : rc = nvme_rdma_build_sgl_request(rqpair, rdma_req);
1745 : }
1746 : } else {
1747 0 : rc = -1;
1748 : }
1749 :
1750 6 : if (rc) {
1751 0 : rdma_req->req = NULL;
1752 0 : return rc;
1753 : }
1754 :
1755 6 : memcpy(&rqpair->cmds[rdma_req->id], &req->cmd, sizeof(req->cmd));
1756 6 : return 0;
1757 : }
1758 :
1759 : static struct spdk_nvme_qpair *
1760 5 : nvme_rdma_ctrlr_create_qpair(struct spdk_nvme_ctrlr *ctrlr,
1761 : uint16_t qid, uint32_t qsize,
1762 : enum spdk_nvme_qprio qprio,
1763 : uint32_t num_requests,
1764 : bool delay_cmd_submit,
1765 : bool async)
1766 : {
1767 : struct nvme_rdma_qpair *rqpair;
1768 : struct spdk_nvme_qpair *qpair;
1769 : int rc;
1770 :
1771 5 : if (qsize < SPDK_NVME_QUEUE_MIN_ENTRIES) {
1772 2 : SPDK_ERRLOG("Failed to create qpair with size %u. Minimum queue size is %d.\n",
1773 : qsize, SPDK_NVME_QUEUE_MIN_ENTRIES);
1774 2 : return NULL;
1775 : }
1776 :
1777 3 : rqpair = spdk_zmalloc(sizeof(struct nvme_rdma_qpair), 0, NULL, SPDK_ENV_SOCKET_ID_ANY,
1778 : SPDK_MALLOC_DMA);
1779 3 : if (!rqpair) {
1780 0 : SPDK_ERRLOG("failed to get create rqpair\n");
1781 0 : return NULL;
1782 : }
1783 :
1784 : /* Set num_entries one less than queue size. According to NVMe
1785 : * and NVMe-oF specs we can not submit queue size requests,
1786 : * one slot shall always remain empty.
1787 : */
1788 3 : rqpair->num_entries = qsize - 1;
1789 3 : rqpair->delay_cmd_submit = delay_cmd_submit;
1790 3 : rqpair->state = NVME_RDMA_QPAIR_STATE_INVALID;
1791 3 : qpair = &rqpair->qpair;
1792 3 : rc = nvme_qpair_init(qpair, qid, ctrlr, qprio, num_requests, async);
1793 3 : if (rc != 0) {
1794 0 : spdk_free(rqpair);
1795 0 : return NULL;
1796 : }
1797 :
1798 3 : return qpair;
1799 : }
1800 :
1801 : static void
1802 1 : nvme_rdma_qpair_destroy(struct nvme_rdma_qpair *rqpair)
1803 : {
1804 1 : struct spdk_nvme_qpair *qpair = &rqpair->qpair;
1805 : struct nvme_rdma_ctrlr *rctrlr;
1806 : struct nvme_rdma_cm_event_entry *entry, *tmp;
1807 :
1808 1 : spdk_rdma_utils_free_mem_map(&rqpair->mr_map);
1809 :
1810 1 : if (rqpair->evt) {
1811 0 : rdma_ack_cm_event(rqpair->evt);
1812 0 : rqpair->evt = NULL;
1813 : }
1814 :
1815 : /*
1816 : * This works because we have the controller lock both in
1817 : * this function and in the function where we add new events.
1818 : */
1819 1 : if (qpair->ctrlr != NULL) {
1820 1 : rctrlr = nvme_rdma_ctrlr(qpair->ctrlr);
1821 1 : STAILQ_FOREACH_SAFE(entry, &rctrlr->pending_cm_events, link, tmp) {
1822 0 : if (entry->evt->id->context == rqpair) {
1823 0 : STAILQ_REMOVE(&rctrlr->pending_cm_events, entry, nvme_rdma_cm_event_entry, link);
1824 0 : rdma_ack_cm_event(entry->evt);
1825 0 : STAILQ_INSERT_HEAD(&rctrlr->free_cm_events, entry, link);
1826 : }
1827 : }
1828 : }
1829 :
1830 1 : if (rqpair->cm_id) {
1831 0 : if (rqpair->rdma_qp) {
1832 0 : spdk_rdma_utils_put_pd(rqpair->rdma_qp->qp->pd);
1833 0 : spdk_rdma_provider_qp_destroy(rqpair->rdma_qp);
1834 0 : rqpair->rdma_qp = NULL;
1835 : }
1836 : }
1837 :
1838 1 : if (rqpair->poller) {
1839 : struct nvme_rdma_poll_group *group;
1840 :
1841 0 : assert(qpair->poll_group);
1842 0 : group = nvme_rdma_poll_group(qpair->poll_group);
1843 :
1844 0 : nvme_rdma_poll_group_put_poller(group, rqpair->poller);
1845 :
1846 0 : rqpair->poller = NULL;
1847 0 : rqpair->cq = NULL;
1848 0 : if (rqpair->srq) {
1849 0 : rqpair->srq = NULL;
1850 0 : rqpair->rsps = NULL;
1851 : }
1852 1 : } else if (rqpair->cq) {
1853 0 : ibv_destroy_cq(rqpair->cq);
1854 0 : rqpair->cq = NULL;
1855 : }
1856 :
1857 1 : nvme_rdma_free_reqs(rqpair);
1858 1 : nvme_rdma_free_rsps(rqpair->rsps);
1859 1 : rqpair->rsps = NULL;
1860 :
1861 : /* destroy cm_id last so cma device will not be freed before we destroy the cq. */
1862 1 : if (rqpair->cm_id) {
1863 0 : rdma_destroy_id(rqpair->cm_id);
1864 0 : rqpair->cm_id = NULL;
1865 : }
1866 1 : }
1867 :
1868 : static void nvme_rdma_qpair_abort_reqs(struct spdk_nvme_qpair *qpair, uint32_t dnr);
1869 :
1870 : static int
1871 1 : nvme_rdma_qpair_disconnected(struct nvme_rdma_qpair *rqpair, int ret)
1872 : {
1873 1 : if (ret) {
1874 0 : SPDK_DEBUGLOG(nvme, "Target did not respond to qpair disconnect.\n");
1875 0 : goto quiet;
1876 : }
1877 :
1878 1 : if (rqpair->poller == NULL) {
1879 : /* If poller is not used, cq is not shared.
1880 : * So complete disconnecting qpair immediately.
1881 : */
1882 1 : goto quiet;
1883 : }
1884 :
1885 0 : if (rqpair->rsps == NULL) {
1886 0 : goto quiet;
1887 : }
1888 :
1889 0 : if (rqpair->need_destroy ||
1890 0 : (rqpair->current_num_sends != 0 ||
1891 0 : (!rqpair->srq && rqpair->rsps->current_num_recvs != 0))) {
1892 0 : rqpair->state = NVME_RDMA_QPAIR_STATE_LINGERING;
1893 0 : rqpair->evt_timeout_ticks = (NVME_RDMA_DISCONNECTED_QPAIR_TIMEOUT_US * spdk_get_ticks_hz()) /
1894 0 : SPDK_SEC_TO_USEC + spdk_get_ticks();
1895 :
1896 0 : return -EAGAIN;
1897 : }
1898 :
1899 0 : quiet:
1900 1 : rqpair->state = NVME_RDMA_QPAIR_STATE_EXITED;
1901 :
1902 1 : nvme_rdma_qpair_abort_reqs(&rqpair->qpair, 0);
1903 1 : nvme_rdma_qpair_destroy(rqpair);
1904 1 : nvme_transport_ctrlr_disconnect_qpair_done(&rqpair->qpair);
1905 :
1906 1 : return 0;
1907 : }
1908 :
1909 : static int
1910 0 : nvme_rdma_qpair_wait_until_quiet(struct nvme_rdma_qpair *rqpair)
1911 : {
1912 0 : struct spdk_nvme_qpair *qpair = &rqpair->qpair;
1913 0 : struct spdk_nvme_ctrlr *ctrlr = qpair->ctrlr;
1914 :
1915 0 : if (spdk_get_ticks() < rqpair->evt_timeout_ticks &&
1916 0 : (rqpair->current_num_sends != 0 ||
1917 0 : (!rqpair->srq && rqpair->rsps->current_num_recvs != 0))) {
1918 0 : return -EAGAIN;
1919 : }
1920 :
1921 0 : rqpair->state = NVME_RDMA_QPAIR_STATE_EXITED;
1922 0 : nvme_rdma_qpair_abort_reqs(&rqpair->qpair, 0);
1923 0 : if (!nvme_qpair_is_admin_queue(qpair)) {
1924 0 : nvme_robust_mutex_lock(&ctrlr->ctrlr_lock);
1925 : }
1926 0 : nvme_rdma_qpair_destroy(rqpair);
1927 0 : if (!nvme_qpair_is_admin_queue(qpair)) {
1928 0 : nvme_robust_mutex_unlock(&ctrlr->ctrlr_lock);
1929 : }
1930 0 : nvme_transport_ctrlr_disconnect_qpair_done(&rqpair->qpair);
1931 :
1932 0 : return 0;
1933 : }
1934 :
1935 : static void
1936 0 : _nvme_rdma_ctrlr_disconnect_qpair(struct spdk_nvme_ctrlr *ctrlr, struct spdk_nvme_qpair *qpair,
1937 : nvme_rdma_cm_event_cb disconnected_qpair_cb)
1938 : {
1939 0 : struct nvme_rdma_qpair *rqpair = nvme_rdma_qpair(qpair);
1940 : int rc;
1941 :
1942 0 : assert(disconnected_qpair_cb != NULL);
1943 :
1944 0 : rqpair->state = NVME_RDMA_QPAIR_STATE_EXITING;
1945 :
1946 0 : if (rqpair->cm_id) {
1947 0 : if (rqpair->rdma_qp) {
1948 0 : rc = spdk_rdma_provider_qp_disconnect(rqpair->rdma_qp);
1949 0 : if ((qpair->ctrlr != NULL) && (rc == 0)) {
1950 0 : rc = nvme_rdma_process_event_start(rqpair, RDMA_CM_EVENT_DISCONNECTED,
1951 : disconnected_qpair_cb);
1952 0 : if (rc == 0) {
1953 0 : return;
1954 : }
1955 : }
1956 : }
1957 : }
1958 :
1959 0 : disconnected_qpair_cb(rqpair, 0);
1960 : }
1961 :
1962 : static int
1963 0 : nvme_rdma_ctrlr_disconnect_qpair_poll(struct spdk_nvme_ctrlr *ctrlr, struct spdk_nvme_qpair *qpair)
1964 : {
1965 0 : struct nvme_rdma_qpair *rqpair = nvme_rdma_qpair(qpair);
1966 : int rc;
1967 :
1968 0 : switch (rqpair->state) {
1969 0 : case NVME_RDMA_QPAIR_STATE_EXITING:
1970 0 : if (!nvme_qpair_is_admin_queue(qpair)) {
1971 0 : nvme_ctrlr_lock(ctrlr);
1972 : }
1973 :
1974 0 : rc = nvme_rdma_process_event_poll(rqpair);
1975 :
1976 0 : if (!nvme_qpair_is_admin_queue(qpair)) {
1977 0 : nvme_ctrlr_unlock(ctrlr);
1978 : }
1979 0 : break;
1980 :
1981 0 : case NVME_RDMA_QPAIR_STATE_LINGERING:
1982 0 : rc = nvme_rdma_qpair_wait_until_quiet(rqpair);
1983 0 : break;
1984 0 : case NVME_RDMA_QPAIR_STATE_EXITED:
1985 0 : rc = 0;
1986 0 : break;
1987 :
1988 0 : default:
1989 0 : assert(false);
1990 : rc = -EAGAIN;
1991 : break;
1992 : }
1993 :
1994 0 : return rc;
1995 : }
1996 :
1997 : static void
1998 0 : nvme_rdma_ctrlr_disconnect_qpair(struct spdk_nvme_ctrlr *ctrlr, struct spdk_nvme_qpair *qpair)
1999 : {
2000 : int rc;
2001 :
2002 0 : _nvme_rdma_ctrlr_disconnect_qpair(ctrlr, qpair, nvme_rdma_qpair_disconnected);
2003 :
2004 : /* If the async mode is disabled, poll the qpair until it is actually disconnected.
2005 : * It is ensured that poll_group_process_completions() calls disconnected_qpair_cb
2006 : * for any disconnected qpair. Hence, we do not have to check if the qpair is in
2007 : * a poll group or not.
2008 : * At the same time, if the qpair is being destroyed, i.e. this function is called by
2009 : * spdk_nvme_ctrlr_free_io_qpair then we need to wait until qpair is disconnected, otherwise
2010 : * we may leak some resources.
2011 : */
2012 0 : if (qpair->async && !qpair->destroy_in_progress) {
2013 0 : return;
2014 : }
2015 :
2016 : while (1) {
2017 0 : rc = nvme_rdma_ctrlr_disconnect_qpair_poll(ctrlr, qpair);
2018 0 : if (rc != -EAGAIN) {
2019 0 : break;
2020 : }
2021 : }
2022 : }
2023 :
2024 : static int
2025 0 : nvme_rdma_stale_conn_disconnected(struct nvme_rdma_qpair *rqpair, int ret)
2026 : {
2027 0 : struct spdk_nvme_qpair *qpair = &rqpair->qpair;
2028 :
2029 0 : if (ret) {
2030 0 : SPDK_DEBUGLOG(nvme, "Target did not respond to qpair disconnect.\n");
2031 : }
2032 :
2033 0 : nvme_rdma_qpair_destroy(rqpair);
2034 :
2035 0 : qpair->last_transport_failure_reason = qpair->transport_failure_reason;
2036 0 : qpair->transport_failure_reason = SPDK_NVME_QPAIR_FAILURE_NONE;
2037 :
2038 0 : rqpair->state = NVME_RDMA_QPAIR_STATE_STALE_CONN;
2039 0 : rqpair->evt_timeout_ticks = (NVME_RDMA_STALE_CONN_RETRY_DELAY_US * spdk_get_ticks_hz()) /
2040 0 : SPDK_SEC_TO_USEC + spdk_get_ticks();
2041 :
2042 0 : return 0;
2043 : }
2044 :
2045 : static int
2046 0 : nvme_rdma_stale_conn_retry(struct nvme_rdma_qpair *rqpair)
2047 : {
2048 0 : struct spdk_nvme_qpair *qpair = &rqpair->qpair;
2049 :
2050 0 : if (rqpair->stale_conn_retry_count >= NVME_RDMA_STALE_CONN_RETRY_MAX) {
2051 0 : SPDK_ERRLOG("Retry failed %d times, give up stale connection to qpair (cntlid:%u, qid:%u).\n",
2052 : NVME_RDMA_STALE_CONN_RETRY_MAX, qpair->ctrlr->cntlid, qpair->id);
2053 0 : return -ESTALE;
2054 : }
2055 :
2056 0 : rqpair->stale_conn_retry_count++;
2057 :
2058 0 : SPDK_NOTICELOG("%d times, retry stale connection to qpair (cntlid:%u, qid:%u).\n",
2059 : rqpair->stale_conn_retry_count, qpair->ctrlr->cntlid, qpair->id);
2060 :
2061 0 : _nvme_rdma_ctrlr_disconnect_qpair(qpair->ctrlr, qpair, nvme_rdma_stale_conn_disconnected);
2062 :
2063 0 : return 0;
2064 : }
2065 :
2066 : static int
2067 1 : nvme_rdma_ctrlr_delete_io_qpair(struct spdk_nvme_ctrlr *ctrlr, struct spdk_nvme_qpair *qpair)
2068 : {
2069 : struct nvme_rdma_qpair *rqpair;
2070 :
2071 1 : assert(qpair != NULL);
2072 1 : rqpair = nvme_rdma_qpair(qpair);
2073 :
2074 1 : if (rqpair->state != NVME_RDMA_QPAIR_STATE_EXITED) {
2075 : int rc __attribute__((unused));
2076 :
2077 : /* qpair was removed from the poll group while the disconnect is not finished.
2078 : * Destroy rdma resources forcefully. */
2079 1 : rc = nvme_rdma_qpair_disconnected(rqpair, 0);
2080 1 : assert(rc == 0);
2081 : }
2082 :
2083 1 : nvme_rdma_qpair_abort_reqs(qpair, 0);
2084 1 : nvme_qpair_deinit(qpair);
2085 :
2086 1 : if (spdk_rdma_utils_put_memory_domain(rqpair->memory_domain) != 0) {
2087 0 : SPDK_ERRLOG("Failed to release memory domain\n");
2088 0 : assert(0);
2089 : }
2090 :
2091 1 : spdk_free(rqpair);
2092 :
2093 1 : return 0;
2094 : }
2095 :
2096 : static struct spdk_nvme_qpair *
2097 0 : nvme_rdma_ctrlr_create_io_qpair(struct spdk_nvme_ctrlr *ctrlr, uint16_t qid,
2098 : const struct spdk_nvme_io_qpair_opts *opts)
2099 : {
2100 0 : return nvme_rdma_ctrlr_create_qpair(ctrlr, qid, opts->io_queue_size, opts->qprio,
2101 0 : opts->io_queue_requests,
2102 0 : opts->delay_cmd_submit,
2103 0 : opts->async_mode);
2104 : }
2105 :
2106 : static int
2107 0 : nvme_rdma_ctrlr_enable(struct spdk_nvme_ctrlr *ctrlr)
2108 : {
2109 : /* do nothing here */
2110 0 : return 0;
2111 : }
2112 :
2113 : static int nvme_rdma_ctrlr_destruct(struct spdk_nvme_ctrlr *ctrlr);
2114 :
2115 : /* We have to use the typedef in the function declaration to appease astyle. */
2116 : typedef struct spdk_nvme_ctrlr spdk_nvme_ctrlr_t;
2117 :
2118 : static spdk_nvme_ctrlr_t *
2119 1 : nvme_rdma_ctrlr_construct(const struct spdk_nvme_transport_id *trid,
2120 : const struct spdk_nvme_ctrlr_opts *opts,
2121 : void *devhandle)
2122 : {
2123 : struct nvme_rdma_ctrlr *rctrlr;
2124 : struct ibv_context **contexts;
2125 1 : struct ibv_device_attr dev_attr;
2126 : int i, flag, rc;
2127 :
2128 1 : rctrlr = spdk_zmalloc(sizeof(struct nvme_rdma_ctrlr), 0, NULL, SPDK_ENV_SOCKET_ID_ANY,
2129 : SPDK_MALLOC_DMA);
2130 1 : if (rctrlr == NULL) {
2131 0 : SPDK_ERRLOG("could not allocate ctrlr\n");
2132 0 : return NULL;
2133 : }
2134 :
2135 1 : rctrlr->ctrlr.opts = *opts;
2136 1 : rctrlr->ctrlr.trid = *trid;
2137 :
2138 1 : if (opts->transport_retry_count > NVME_RDMA_CTRLR_MAX_TRANSPORT_RETRY_COUNT) {
2139 1 : SPDK_NOTICELOG("transport_retry_count exceeds max value %d, use max value\n",
2140 : NVME_RDMA_CTRLR_MAX_TRANSPORT_RETRY_COUNT);
2141 1 : rctrlr->ctrlr.opts.transport_retry_count = NVME_RDMA_CTRLR_MAX_TRANSPORT_RETRY_COUNT;
2142 : }
2143 :
2144 1 : if (opts->transport_ack_timeout > NVME_RDMA_CTRLR_MAX_TRANSPORT_ACK_TIMEOUT) {
2145 1 : SPDK_NOTICELOG("transport_ack_timeout exceeds max value %d, use max value\n",
2146 : NVME_RDMA_CTRLR_MAX_TRANSPORT_ACK_TIMEOUT);
2147 1 : rctrlr->ctrlr.opts.transport_ack_timeout = NVME_RDMA_CTRLR_MAX_TRANSPORT_ACK_TIMEOUT;
2148 : }
2149 :
2150 1 : contexts = rdma_get_devices(NULL);
2151 1 : if (contexts == NULL) {
2152 0 : SPDK_ERRLOG("rdma_get_devices() failed: %s (%d)\n", spdk_strerror(errno), errno);
2153 0 : spdk_free(rctrlr);
2154 0 : return NULL;
2155 : }
2156 :
2157 1 : i = 0;
2158 1 : rctrlr->max_sge = NVME_RDMA_MAX_SGL_DESCRIPTORS;
2159 :
2160 3 : while (contexts[i] != NULL) {
2161 2 : rc = ibv_query_device(contexts[i], &dev_attr);
2162 2 : if (rc < 0) {
2163 0 : SPDK_ERRLOG("Failed to query RDMA device attributes.\n");
2164 0 : rdma_free_devices(contexts);
2165 0 : spdk_free(rctrlr);
2166 0 : return NULL;
2167 : }
2168 2 : rctrlr->max_sge = spdk_min(rctrlr->max_sge, (uint16_t)dev_attr.max_sge);
2169 2 : i++;
2170 : }
2171 :
2172 1 : rdma_free_devices(contexts);
2173 :
2174 1 : rc = nvme_ctrlr_construct(&rctrlr->ctrlr);
2175 1 : if (rc != 0) {
2176 0 : spdk_free(rctrlr);
2177 0 : return NULL;
2178 : }
2179 :
2180 1 : STAILQ_INIT(&rctrlr->pending_cm_events);
2181 1 : STAILQ_INIT(&rctrlr->free_cm_events);
2182 1 : rctrlr->cm_events = spdk_zmalloc(NVME_RDMA_NUM_CM_EVENTS * sizeof(*rctrlr->cm_events), 0, NULL,
2183 : SPDK_ENV_SOCKET_ID_ANY, SPDK_MALLOC_DMA);
2184 1 : if (rctrlr->cm_events == NULL) {
2185 0 : SPDK_ERRLOG("unable to allocate buffers to hold CM events.\n");
2186 0 : goto destruct_ctrlr;
2187 : }
2188 :
2189 257 : for (i = 0; i < NVME_RDMA_NUM_CM_EVENTS; i++) {
2190 256 : STAILQ_INSERT_TAIL(&rctrlr->free_cm_events, &rctrlr->cm_events[i], link);
2191 : }
2192 :
2193 1 : rctrlr->cm_channel = rdma_create_event_channel();
2194 1 : if (rctrlr->cm_channel == NULL) {
2195 0 : SPDK_ERRLOG("rdma_create_event_channel() failed\n");
2196 0 : goto destruct_ctrlr;
2197 : }
2198 :
2199 1 : flag = fcntl(rctrlr->cm_channel->fd, F_GETFL);
2200 1 : if (fcntl(rctrlr->cm_channel->fd, F_SETFL, flag | O_NONBLOCK) < 0) {
2201 0 : SPDK_ERRLOG("Cannot set event channel to non blocking\n");
2202 0 : goto destruct_ctrlr;
2203 : }
2204 :
2205 2 : rctrlr->ctrlr.adminq = nvme_rdma_ctrlr_create_qpair(&rctrlr->ctrlr, 0,
2206 1 : rctrlr->ctrlr.opts.admin_queue_size, 0,
2207 1 : rctrlr->ctrlr.opts.admin_queue_size, false, true);
2208 1 : if (!rctrlr->ctrlr.adminq) {
2209 0 : SPDK_ERRLOG("failed to create admin qpair\n");
2210 0 : goto destruct_ctrlr;
2211 : }
2212 :
2213 1 : if (nvme_ctrlr_add_process(&rctrlr->ctrlr, 0) != 0) {
2214 0 : SPDK_ERRLOG("nvme_ctrlr_add_process() failed\n");
2215 0 : goto destruct_ctrlr;
2216 : }
2217 :
2218 1 : SPDK_DEBUGLOG(nvme, "successfully initialized the nvmf ctrlr\n");
2219 1 : return &rctrlr->ctrlr;
2220 :
2221 0 : destruct_ctrlr:
2222 0 : nvme_ctrlr_destruct(&rctrlr->ctrlr);
2223 0 : return NULL;
2224 : }
2225 :
2226 : static int
2227 1 : nvme_rdma_ctrlr_destruct(struct spdk_nvme_ctrlr *ctrlr)
2228 : {
2229 1 : struct nvme_rdma_ctrlr *rctrlr = nvme_rdma_ctrlr(ctrlr);
2230 : struct nvme_rdma_cm_event_entry *entry;
2231 :
2232 1 : if (ctrlr->adminq) {
2233 1 : nvme_rdma_ctrlr_delete_io_qpair(ctrlr, ctrlr->adminq);
2234 : }
2235 :
2236 1 : STAILQ_FOREACH(entry, &rctrlr->pending_cm_events, link) {
2237 0 : rdma_ack_cm_event(entry->evt);
2238 : }
2239 :
2240 1 : STAILQ_INIT(&rctrlr->free_cm_events);
2241 1 : STAILQ_INIT(&rctrlr->pending_cm_events);
2242 1 : spdk_free(rctrlr->cm_events);
2243 :
2244 1 : if (rctrlr->cm_channel) {
2245 1 : rdma_destroy_event_channel(rctrlr->cm_channel);
2246 1 : rctrlr->cm_channel = NULL;
2247 : }
2248 :
2249 1 : nvme_ctrlr_destruct_finish(ctrlr);
2250 :
2251 1 : spdk_free(rctrlr);
2252 :
2253 1 : return 0;
2254 : }
2255 :
2256 : static int
2257 2 : nvme_rdma_qpair_submit_request(struct spdk_nvme_qpair *qpair,
2258 : struct nvme_request *req)
2259 : {
2260 : struct nvme_rdma_qpair *rqpair;
2261 : struct spdk_nvme_rdma_req *rdma_req;
2262 : struct ibv_send_wr *wr;
2263 : struct nvme_rdma_poll_group *group;
2264 :
2265 2 : rqpair = nvme_rdma_qpair(qpair);
2266 2 : assert(rqpair != NULL);
2267 2 : assert(req != NULL);
2268 :
2269 2 : rdma_req = nvme_rdma_req_get(rqpair);
2270 2 : if (spdk_unlikely(!rdma_req)) {
2271 1 : if (rqpair->poller) {
2272 1 : rqpair->poller->stats.queued_requests++;
2273 : }
2274 : /* Inform the upper layer to try again later. */
2275 1 : return -EAGAIN;
2276 : }
2277 :
2278 1 : if (nvme_rdma_req_init(rqpair, req, rdma_req)) {
2279 0 : SPDK_ERRLOG("nvme_rdma_req_init() failed\n");
2280 0 : nvme_rdma_req_put(rqpair, rdma_req);
2281 0 : return -1;
2282 : }
2283 :
2284 1 : TAILQ_INSERT_TAIL(&rqpair->outstanding_reqs, rdma_req, link);
2285 :
2286 1 : if (!rqpair->link_active.tqe_prev && qpair->poll_group) {
2287 0 : group = nvme_rdma_poll_group(qpair->poll_group);
2288 0 : TAILQ_INSERT_TAIL(&group->active_qpairs, rqpair, link_active);
2289 : }
2290 1 : rqpair->num_outstanding_reqs++;
2291 :
2292 1 : assert(rqpair->current_num_sends < rqpair->num_entries);
2293 1 : rqpair->current_num_sends++;
2294 :
2295 1 : wr = &rdma_req->send_wr;
2296 1 : wr->next = NULL;
2297 1 : nvme_rdma_trace_ibv_sge(wr->sg_list);
2298 :
2299 1 : spdk_rdma_provider_qp_queue_send_wrs(rqpair->rdma_qp, wr);
2300 :
2301 1 : if (!rqpair->delay_cmd_submit) {
2302 1 : return nvme_rdma_qpair_submit_sends(rqpair);
2303 : }
2304 :
2305 0 : return 0;
2306 : }
2307 :
2308 : static int
2309 0 : nvme_rdma_qpair_reset(struct spdk_nvme_qpair *qpair)
2310 : {
2311 : /* Currently, doing nothing here */
2312 0 : return 0;
2313 : }
2314 :
2315 : static void
2316 2 : nvme_rdma_qpair_abort_reqs(struct spdk_nvme_qpair *qpair, uint32_t dnr)
2317 : {
2318 : struct spdk_nvme_rdma_req *rdma_req, *tmp;
2319 2 : struct spdk_nvme_cpl cpl;
2320 2 : struct nvme_rdma_qpair *rqpair = nvme_rdma_qpair(qpair);
2321 :
2322 2 : cpl.sqid = qpair->id;
2323 2 : cpl.status.sc = SPDK_NVME_SC_ABORTED_SQ_DELETION;
2324 2 : cpl.status.sct = SPDK_NVME_SCT_GENERIC;
2325 2 : cpl.status.dnr = dnr;
2326 :
2327 : /*
2328 : * We cannot abort requests at the RDMA layer without
2329 : * unregistering them. If we do, we can still get error
2330 : * free completions on the shared completion queue.
2331 : */
2332 2 : if (nvme_qpair_get_state(qpair) > NVME_QPAIR_DISCONNECTING &&
2333 0 : nvme_qpair_get_state(qpair) != NVME_QPAIR_DESTROYING) {
2334 0 : nvme_ctrlr_disconnect_qpair(qpair);
2335 : }
2336 :
2337 2 : TAILQ_FOREACH_SAFE(rdma_req, &rqpair->outstanding_reqs, link, tmp) {
2338 0 : nvme_rdma_req_complete(rdma_req, &cpl, true);
2339 : }
2340 2 : }
2341 :
2342 : static void
2343 0 : nvme_rdma_qpair_check_timeout(struct spdk_nvme_qpair *qpair)
2344 : {
2345 : uint64_t t02;
2346 : struct spdk_nvme_rdma_req *rdma_req, *tmp;
2347 0 : struct nvme_rdma_qpair *rqpair = nvme_rdma_qpair(qpair);
2348 0 : struct spdk_nvme_ctrlr *ctrlr = qpair->ctrlr;
2349 : struct spdk_nvme_ctrlr_process *active_proc;
2350 :
2351 : /* Don't check timeouts during controller initialization. */
2352 0 : if (ctrlr->state != NVME_CTRLR_STATE_READY) {
2353 0 : return;
2354 : }
2355 :
2356 0 : if (nvme_qpair_is_admin_queue(qpair)) {
2357 0 : active_proc = nvme_ctrlr_get_current_process(ctrlr);
2358 : } else {
2359 0 : active_proc = qpair->active_proc;
2360 : }
2361 :
2362 : /* Only check timeouts if the current process has a timeout callback. */
2363 0 : if (active_proc == NULL || active_proc->timeout_cb_fn == NULL) {
2364 0 : return;
2365 : }
2366 :
2367 0 : t02 = spdk_get_ticks();
2368 0 : TAILQ_FOREACH_SAFE(rdma_req, &rqpair->outstanding_reqs, link, tmp) {
2369 0 : assert(rdma_req->req != NULL);
2370 :
2371 0 : if (nvme_request_check_timeout(rdma_req->req, rdma_req->id, active_proc, t02)) {
2372 : /*
2373 : * The requests are in order, so as soon as one has not timed out,
2374 : * stop iterating.
2375 : */
2376 0 : break;
2377 : }
2378 : }
2379 : }
2380 :
2381 : static inline void
2382 0 : nvme_rdma_request_ready(struct nvme_rdma_qpair *rqpair, struct spdk_nvme_rdma_req *rdma_req)
2383 : {
2384 0 : struct spdk_nvme_rdma_rsp *rdma_rsp = rdma_req->rdma_rsp;
2385 0 : struct ibv_recv_wr *recv_wr = rdma_rsp->recv_wr;
2386 :
2387 0 : nvme_rdma_req_complete(rdma_req, &rdma_rsp->cpl, true);
2388 :
2389 0 : assert(rqpair->rsps->current_num_recvs < rqpair->rsps->num_entries);
2390 0 : rqpair->rsps->current_num_recvs++;
2391 :
2392 0 : recv_wr->next = NULL;
2393 0 : nvme_rdma_trace_ibv_sge(recv_wr->sg_list);
2394 :
2395 0 : if (!rqpair->srq) {
2396 0 : spdk_rdma_provider_qp_queue_recv_wrs(rqpair->rdma_qp, recv_wr);
2397 : } else {
2398 0 : spdk_rdma_provider_srq_queue_recv_wrs(rqpair->srq, recv_wr);
2399 : }
2400 0 : }
2401 :
2402 : #define MAX_COMPLETIONS_PER_POLL 128
2403 :
2404 : static void
2405 0 : nvme_rdma_fail_qpair(struct spdk_nvme_qpair *qpair, int failure_reason)
2406 : {
2407 0 : if (failure_reason == IBV_WC_RETRY_EXC_ERR) {
2408 0 : qpair->transport_failure_reason = SPDK_NVME_QPAIR_FAILURE_REMOTE;
2409 0 : } else if (qpair->transport_failure_reason == SPDK_NVME_QPAIR_FAILURE_NONE) {
2410 0 : qpair->transport_failure_reason = SPDK_NVME_QPAIR_FAILURE_UNKNOWN;
2411 : }
2412 :
2413 0 : nvme_ctrlr_disconnect_qpair(qpair);
2414 0 : }
2415 :
2416 : static struct nvme_rdma_qpair *
2417 4 : get_rdma_qpair_from_wc(struct nvme_rdma_poll_group *group, struct ibv_wc *wc)
2418 : {
2419 : struct spdk_nvme_qpair *qpair;
2420 : struct nvme_rdma_qpair *rqpair;
2421 :
2422 5 : STAILQ_FOREACH(qpair, &group->group.connected_qpairs, poll_group_stailq) {
2423 2 : rqpair = nvme_rdma_qpair(qpair);
2424 2 : if (NVME_RDMA_POLL_GROUP_CHECK_QPN(rqpair, wc->qp_num)) {
2425 1 : return rqpair;
2426 : }
2427 : }
2428 :
2429 4 : STAILQ_FOREACH(qpair, &group->group.disconnected_qpairs, poll_group_stailq) {
2430 2 : rqpair = nvme_rdma_qpair(qpair);
2431 2 : if (NVME_RDMA_POLL_GROUP_CHECK_QPN(rqpair, wc->qp_num)) {
2432 1 : return rqpair;
2433 : }
2434 : }
2435 :
2436 2 : return NULL;
2437 : }
2438 :
2439 : static inline void
2440 0 : nvme_rdma_log_wc_status(struct nvme_rdma_qpair *rqpair, struct ibv_wc *wc)
2441 : {
2442 0 : struct nvme_rdma_wr *rdma_wr = (struct nvme_rdma_wr *)wc->wr_id;
2443 :
2444 0 : if (wc->status == IBV_WC_WR_FLUSH_ERR) {
2445 : /* If qpair is in ERR state, we will receive completions for all posted and not completed
2446 : * Work Requests with IBV_WC_WR_FLUSH_ERR status. Don't log an error in that case */
2447 0 : SPDK_DEBUGLOG(nvme, "WC error, qid %u, qp state %d, request 0x%lu type %d, status: (%d): %s\n",
2448 : rqpair->qpair.id, rqpair->qpair.state, wc->wr_id, rdma_wr->type, wc->status,
2449 : ibv_wc_status_str(wc->status));
2450 : } else {
2451 0 : SPDK_ERRLOG("WC error, qid %u, qp state %d, request 0x%lu type %d, status: (%d): %s\n",
2452 : rqpair->qpair.id, rqpair->qpair.state, wc->wr_id, rdma_wr->type, wc->status,
2453 : ibv_wc_status_str(wc->status));
2454 : }
2455 0 : }
2456 :
2457 : static inline int
2458 0 : nvme_rdma_process_recv_completion(struct nvme_rdma_poller *poller, struct ibv_wc *wc,
2459 : struct nvme_rdma_wr *rdma_wr)
2460 : {
2461 : struct nvme_rdma_qpair *rqpair;
2462 : struct spdk_nvme_rdma_req *rdma_req;
2463 : struct spdk_nvme_rdma_rsp *rdma_rsp;
2464 :
2465 0 : rdma_rsp = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvme_rdma_rsp, rdma_wr);
2466 :
2467 0 : if (poller && poller->srq) {
2468 0 : rqpair = get_rdma_qpair_from_wc(poller->group, wc);
2469 0 : if (spdk_unlikely(!rqpair)) {
2470 : /* Since we do not handle the LAST_WQE_REACHED event, we do not know when
2471 : * a Receive Queue in a QP, that is associated with an SRQ, is flushed.
2472 : * We may get a WC for a already destroyed QP.
2473 : *
2474 : * However, for the SRQ, this is not any error. Hence, just re-post the
2475 : * receive request to the SRQ to reuse for other QPs, and return 0.
2476 : */
2477 0 : spdk_rdma_provider_srq_queue_recv_wrs(poller->srq, rdma_rsp->recv_wr);
2478 0 : return 0;
2479 : }
2480 : } else {
2481 0 : rqpair = rdma_rsp->rqpair;
2482 0 : if (spdk_unlikely(!rqpair)) {
2483 : /* TODO: Fix forceful QP destroy when it is not async mode.
2484 : * CQ itself did not cause any error. Hence, return 0 for now.
2485 : */
2486 0 : SPDK_WARNLOG("QP might be already destroyed.\n");
2487 0 : return 0;
2488 : }
2489 : }
2490 :
2491 :
2492 0 : assert(rqpair->rsps->current_num_recvs > 0);
2493 0 : rqpair->rsps->current_num_recvs--;
2494 :
2495 0 : if (wc->status) {
2496 0 : nvme_rdma_log_wc_status(rqpair, wc);
2497 0 : goto err_wc;
2498 : }
2499 :
2500 0 : SPDK_DEBUGLOG(nvme, "CQ recv completion\n");
2501 :
2502 0 : if (wc->byte_len < sizeof(struct spdk_nvme_cpl)) {
2503 0 : SPDK_ERRLOG("recv length %u less than expected response size\n", wc->byte_len);
2504 0 : goto err_wc;
2505 : }
2506 0 : rdma_req = &rqpair->rdma_reqs[rdma_rsp->cpl.cid];
2507 0 : rdma_req->completion_flags |= NVME_RDMA_RECV_COMPLETED;
2508 0 : rdma_req->rdma_rsp = rdma_rsp;
2509 :
2510 0 : if ((rdma_req->completion_flags & NVME_RDMA_SEND_COMPLETED) == 0) {
2511 0 : return 0;
2512 : }
2513 :
2514 0 : rqpair->num_completions++;
2515 :
2516 0 : nvme_rdma_request_ready(rqpair, rdma_req);
2517 :
2518 0 : if (!rqpair->delay_cmd_submit) {
2519 0 : if (spdk_unlikely(nvme_rdma_qpair_submit_recvs(rqpair))) {
2520 0 : SPDK_ERRLOG("Unable to re-post rx descriptor\n");
2521 0 : nvme_rdma_fail_qpair(&rqpair->qpair, 0);
2522 0 : return -ENXIO;
2523 : }
2524 : }
2525 :
2526 0 : return 1;
2527 :
2528 0 : err_wc:
2529 0 : nvme_rdma_fail_qpair(&rqpair->qpair, 0);
2530 0 : if (poller && poller->srq) {
2531 0 : spdk_rdma_provider_srq_queue_recv_wrs(poller->srq, rdma_rsp->recv_wr);
2532 : }
2533 0 : return -ENXIO;
2534 : }
2535 :
2536 : static inline int
2537 0 : nvme_rdma_process_send_completion(struct nvme_rdma_poller *poller,
2538 : struct nvme_rdma_qpair *rdma_qpair,
2539 : struct ibv_wc *wc, struct nvme_rdma_wr *rdma_wr)
2540 : {
2541 : struct nvme_rdma_qpair *rqpair;
2542 : struct spdk_nvme_rdma_req *rdma_req;
2543 :
2544 0 : rdma_req = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvme_rdma_req, rdma_wr);
2545 0 : rqpair = rdma_req->req ? nvme_rdma_qpair(rdma_req->req->qpair) : NULL;
2546 0 : if (!rqpair) {
2547 0 : rqpair = rdma_qpair != NULL ? rdma_qpair : get_rdma_qpair_from_wc(poller->group, wc);
2548 : }
2549 :
2550 : /* If we are flushing I/O */
2551 0 : if (wc->status) {
2552 0 : if (!rqpair) {
2553 : /* When poll_group is used, several qpairs share the same CQ and it is possible to
2554 : * receive a completion with error (e.g. IBV_WC_WR_FLUSH_ERR) for already disconnected qpair
2555 : * That happens due to qpair is destroyed while there are submitted but not completed send/receive
2556 : * Work Requests */
2557 0 : assert(poller);
2558 0 : return 0;
2559 : }
2560 0 : assert(rqpair->current_num_sends > 0);
2561 0 : rqpair->current_num_sends--;
2562 0 : nvme_rdma_log_wc_status(rqpair, wc);
2563 0 : nvme_rdma_fail_qpair(&rqpair->qpair, 0);
2564 0 : if (rdma_req->rdma_rsp && poller && poller->srq) {
2565 0 : spdk_rdma_provider_srq_queue_recv_wrs(poller->srq, rdma_req->rdma_rsp->recv_wr);
2566 : }
2567 0 : return -ENXIO;
2568 : }
2569 :
2570 : /* We do not support Soft Roce anymore. Other than Soft Roce's bug, we should not
2571 : * receive a completion without error status after qpair is disconnected/destroyed.
2572 : */
2573 0 : if (spdk_unlikely(rdma_req->req == NULL)) {
2574 : /*
2575 : * Some infiniband drivers do not guarantee the previous assumption after we
2576 : * received a RDMA_CM_EVENT_DEVICE_REMOVAL event.
2577 : */
2578 0 : SPDK_ERRLOG("Received malformed completion: request 0x%"PRIx64" type %d\n", wc->wr_id,
2579 : rdma_wr->type);
2580 0 : if (!rqpair || !rqpair->need_destroy) {
2581 0 : assert(0);
2582 : }
2583 0 : return -ENXIO;
2584 : }
2585 :
2586 0 : rdma_req->completion_flags |= NVME_RDMA_SEND_COMPLETED;
2587 0 : assert(rqpair->current_num_sends > 0);
2588 0 : rqpair->current_num_sends--;
2589 :
2590 0 : if ((rdma_req->completion_flags & NVME_RDMA_RECV_COMPLETED) == 0) {
2591 0 : return 0;
2592 : }
2593 :
2594 0 : rqpair->num_completions++;
2595 :
2596 0 : nvme_rdma_request_ready(rqpair, rdma_req);
2597 :
2598 0 : if (!rqpair->delay_cmd_submit) {
2599 0 : if (spdk_unlikely(nvme_rdma_qpair_submit_recvs(rqpair))) {
2600 0 : SPDK_ERRLOG("Unable to re-post rx descriptor\n");
2601 0 : nvme_rdma_fail_qpair(&rqpair->qpair, 0);
2602 0 : return -ENXIO;
2603 : }
2604 : }
2605 :
2606 0 : return 1;
2607 : }
2608 :
2609 : static int
2610 0 : nvme_rdma_cq_process_completions(struct ibv_cq *cq, uint32_t batch_size,
2611 : struct nvme_rdma_poller *poller,
2612 : struct nvme_rdma_qpair *rdma_qpair,
2613 : uint64_t *rdma_completions)
2614 : {
2615 0 : struct ibv_wc wc[MAX_COMPLETIONS_PER_POLL];
2616 : struct nvme_rdma_wr *rdma_wr;
2617 0 : uint32_t reaped = 0;
2618 0 : int completion_rc = 0;
2619 : int rc, _rc, i;
2620 :
2621 0 : rc = ibv_poll_cq(cq, batch_size, wc);
2622 0 : if (rc < 0) {
2623 0 : SPDK_ERRLOG("Error polling CQ! (%d): %s\n",
2624 : errno, spdk_strerror(errno));
2625 0 : return -ECANCELED;
2626 0 : } else if (rc == 0) {
2627 0 : return 0;
2628 : }
2629 :
2630 0 : for (i = 0; i < rc; i++) {
2631 0 : rdma_wr = (struct nvme_rdma_wr *)wc[i].wr_id;
2632 0 : switch (rdma_wr->type) {
2633 0 : case RDMA_WR_TYPE_RECV:
2634 0 : _rc = nvme_rdma_process_recv_completion(poller, &wc[i], rdma_wr);
2635 0 : break;
2636 :
2637 0 : case RDMA_WR_TYPE_SEND:
2638 0 : _rc = nvme_rdma_process_send_completion(poller, rdma_qpair, &wc[i], rdma_wr);
2639 0 : break;
2640 :
2641 0 : default:
2642 0 : SPDK_ERRLOG("Received an unexpected opcode on the CQ: %d\n", rdma_wr->type);
2643 0 : return -ECANCELED;
2644 : }
2645 0 : if (spdk_likely(_rc >= 0)) {
2646 0 : reaped += _rc;
2647 : } else {
2648 0 : completion_rc = _rc;
2649 : }
2650 : }
2651 :
2652 0 : *rdma_completions += rc;
2653 :
2654 0 : if (completion_rc) {
2655 0 : return completion_rc;
2656 : }
2657 :
2658 0 : return reaped;
2659 : }
2660 :
2661 : static void
2662 0 : dummy_disconnected_qpair_cb(struct spdk_nvme_qpair *qpair, void *poll_group_ctx)
2663 : {
2664 :
2665 0 : }
2666 :
2667 : static int
2668 0 : nvme_rdma_qpair_process_completions(struct spdk_nvme_qpair *qpair,
2669 : uint32_t max_completions)
2670 : {
2671 0 : struct nvme_rdma_qpair *rqpair = nvme_rdma_qpair(qpair);
2672 0 : struct nvme_rdma_ctrlr *rctrlr = nvme_rdma_ctrlr(qpair->ctrlr);
2673 0 : int rc = 0, batch_size;
2674 : struct ibv_cq *cq;
2675 0 : uint64_t rdma_completions = 0;
2676 :
2677 : /*
2678 : * This is used during the connection phase. It's possible that we are still reaping error completions
2679 : * from other qpairs so we need to call the poll group function. Also, it's more correct since the cq
2680 : * is shared.
2681 : */
2682 0 : if (qpair->poll_group != NULL) {
2683 0 : return spdk_nvme_poll_group_process_completions(qpair->poll_group->group, max_completions,
2684 : dummy_disconnected_qpair_cb);
2685 : }
2686 :
2687 0 : if (max_completions == 0) {
2688 0 : max_completions = rqpair->num_entries;
2689 : } else {
2690 0 : max_completions = spdk_min(max_completions, rqpair->num_entries);
2691 : }
2692 :
2693 0 : switch (nvme_qpair_get_state(qpair)) {
2694 0 : case NVME_QPAIR_CONNECTING:
2695 0 : rc = nvme_rdma_ctrlr_connect_qpair_poll(qpair->ctrlr, qpair);
2696 0 : if (rc == 0) {
2697 : /* Once the connection is completed, we can submit queued requests */
2698 0 : nvme_qpair_resubmit_requests(qpair, rqpair->num_entries);
2699 0 : } else if (rc != -EAGAIN) {
2700 0 : SPDK_ERRLOG("Failed to connect rqpair=%p\n", rqpair);
2701 0 : goto failed;
2702 0 : } else if (rqpair->state <= NVME_RDMA_QPAIR_STATE_INITIALIZING) {
2703 0 : return 0;
2704 : }
2705 0 : break;
2706 :
2707 0 : case NVME_QPAIR_DISCONNECTING:
2708 0 : nvme_rdma_ctrlr_disconnect_qpair_poll(qpair->ctrlr, qpair);
2709 0 : return -ENXIO;
2710 :
2711 0 : default:
2712 0 : if (nvme_qpair_is_admin_queue(qpair)) {
2713 0 : nvme_rdma_poll_events(rctrlr);
2714 : }
2715 0 : nvme_rdma_qpair_process_cm_event(rqpair);
2716 0 : break;
2717 : }
2718 :
2719 0 : if (spdk_unlikely(qpair->transport_failure_reason != SPDK_NVME_QPAIR_FAILURE_NONE)) {
2720 0 : goto failed;
2721 : }
2722 :
2723 0 : cq = rqpair->cq;
2724 :
2725 0 : rqpair->num_completions = 0;
2726 : do {
2727 0 : batch_size = spdk_min((max_completions - rqpair->num_completions), MAX_COMPLETIONS_PER_POLL);
2728 0 : rc = nvme_rdma_cq_process_completions(cq, batch_size, NULL, rqpair, &rdma_completions);
2729 :
2730 0 : if (rc == 0) {
2731 0 : break;
2732 : /* Handle the case where we fail to poll the cq. */
2733 0 : } else if (rc == -ECANCELED) {
2734 0 : goto failed;
2735 0 : } else if (rc == -ENXIO) {
2736 0 : return rc;
2737 : }
2738 0 : } while (rqpair->num_completions < max_completions);
2739 :
2740 0 : if (spdk_unlikely(nvme_rdma_qpair_submit_sends(rqpair) ||
2741 : nvme_rdma_qpair_submit_recvs(rqpair))) {
2742 0 : goto failed;
2743 : }
2744 :
2745 0 : if (spdk_unlikely(qpair->ctrlr->timeout_enabled)) {
2746 0 : nvme_rdma_qpair_check_timeout(qpair);
2747 : }
2748 :
2749 0 : return rqpair->num_completions;
2750 :
2751 0 : failed:
2752 0 : nvme_rdma_fail_qpair(qpair, 0);
2753 0 : return -ENXIO;
2754 : }
2755 :
2756 : static uint32_t
2757 0 : nvme_rdma_ctrlr_get_max_xfer_size(struct spdk_nvme_ctrlr *ctrlr)
2758 : {
2759 : /* max_mr_size by ibv_query_device indicates the largest value that we can
2760 : * set for a registered memory region. It is independent from the actual
2761 : * I/O size and is very likely to be larger than 2 MiB which is the
2762 : * granularity we currently register memory regions. Hence return
2763 : * UINT32_MAX here and let the generic layer use the controller data to
2764 : * moderate this value.
2765 : */
2766 0 : return UINT32_MAX;
2767 : }
2768 :
2769 : static uint16_t
2770 5 : nvme_rdma_ctrlr_get_max_sges(struct spdk_nvme_ctrlr *ctrlr)
2771 : {
2772 5 : struct nvme_rdma_ctrlr *rctrlr = nvme_rdma_ctrlr(ctrlr);
2773 5 : uint32_t max_sge = rctrlr->max_sge;
2774 5 : uint32_t max_in_capsule_sge = (ctrlr->cdata.nvmf_specific.ioccsz * 16 -
2775 5 : sizeof(struct spdk_nvme_cmd)) /
2776 : sizeof(struct spdk_nvme_sgl_descriptor);
2777 :
2778 : /* Max SGE is limited by capsule size */
2779 5 : max_sge = spdk_min(max_sge, max_in_capsule_sge);
2780 : /* Max SGE may be limited by MSDBD */
2781 5 : if (ctrlr->cdata.nvmf_specific.msdbd != 0) {
2782 5 : max_sge = spdk_min(max_sge, ctrlr->cdata.nvmf_specific.msdbd);
2783 : }
2784 :
2785 : /* Max SGE can't be less than 1 */
2786 5 : max_sge = spdk_max(1, max_sge);
2787 5 : return max_sge;
2788 : }
2789 :
2790 : static int
2791 0 : nvme_rdma_qpair_iterate_requests(struct spdk_nvme_qpair *qpair,
2792 : int (*iter_fn)(struct nvme_request *req, void *arg),
2793 : void *arg)
2794 : {
2795 0 : struct nvme_rdma_qpair *rqpair = nvme_rdma_qpair(qpair);
2796 : struct spdk_nvme_rdma_req *rdma_req, *tmp;
2797 : int rc;
2798 :
2799 0 : assert(iter_fn != NULL);
2800 :
2801 0 : TAILQ_FOREACH_SAFE(rdma_req, &rqpair->outstanding_reqs, link, tmp) {
2802 0 : assert(rdma_req->req != NULL);
2803 :
2804 0 : rc = iter_fn(rdma_req->req, arg);
2805 0 : if (rc != 0) {
2806 0 : return rc;
2807 : }
2808 : }
2809 :
2810 0 : return 0;
2811 : }
2812 :
2813 : static void
2814 0 : nvme_rdma_admin_qpair_abort_aers(struct spdk_nvme_qpair *qpair)
2815 : {
2816 : struct spdk_nvme_rdma_req *rdma_req, *tmp;
2817 0 : struct spdk_nvme_cpl cpl;
2818 0 : struct nvme_rdma_qpair *rqpair = nvme_rdma_qpair(qpair);
2819 :
2820 0 : cpl.status.sc = SPDK_NVME_SC_ABORTED_SQ_DELETION;
2821 0 : cpl.status.sct = SPDK_NVME_SCT_GENERIC;
2822 :
2823 0 : TAILQ_FOREACH_SAFE(rdma_req, &rqpair->outstanding_reqs, link, tmp) {
2824 0 : assert(rdma_req->req != NULL);
2825 :
2826 0 : if (rdma_req->req->cmd.opc != SPDK_NVME_OPC_ASYNC_EVENT_REQUEST) {
2827 0 : continue;
2828 : }
2829 :
2830 0 : nvme_rdma_req_complete(rdma_req, &cpl, false);
2831 : }
2832 0 : }
2833 :
2834 : static void
2835 9 : nvme_rdma_poller_destroy(struct nvme_rdma_poller *poller)
2836 : {
2837 9 : if (poller->cq) {
2838 7 : ibv_destroy_cq(poller->cq);
2839 : }
2840 9 : if (poller->rsps) {
2841 0 : nvme_rdma_free_rsps(poller->rsps);
2842 : }
2843 9 : if (poller->srq) {
2844 0 : spdk_rdma_provider_srq_destroy(poller->srq);
2845 : }
2846 9 : if (poller->mr_map) {
2847 0 : spdk_rdma_utils_free_mem_map(&poller->mr_map);
2848 : }
2849 9 : if (poller->pd) {
2850 0 : spdk_rdma_utils_put_pd(poller->pd);
2851 : }
2852 9 : free(poller);
2853 9 : }
2854 :
2855 : static struct nvme_rdma_poller *
2856 9 : nvme_rdma_poller_create(struct nvme_rdma_poll_group *group, struct ibv_context *ctx)
2857 : {
2858 : struct nvme_rdma_poller *poller;
2859 9 : struct ibv_device_attr dev_attr;
2860 9 : struct spdk_rdma_provider_srq_init_attr srq_init_attr = {};
2861 9 : struct nvme_rdma_rsp_opts opts;
2862 : int num_cqe, max_num_cqe;
2863 : int rc;
2864 :
2865 9 : poller = calloc(1, sizeof(*poller));
2866 9 : if (poller == NULL) {
2867 0 : SPDK_ERRLOG("Unable to allocate poller.\n");
2868 0 : return NULL;
2869 : }
2870 :
2871 9 : poller->group = group;
2872 9 : poller->device = ctx;
2873 :
2874 9 : if (g_spdk_nvme_transport_opts.rdma_srq_size != 0) {
2875 0 : rc = ibv_query_device(ctx, &dev_attr);
2876 0 : if (rc) {
2877 0 : SPDK_ERRLOG("Unable to query RDMA device.\n");
2878 0 : goto fail;
2879 : }
2880 :
2881 0 : poller->pd = spdk_rdma_utils_get_pd(ctx);
2882 0 : if (poller->pd == NULL) {
2883 0 : SPDK_ERRLOG("Unable to get PD.\n");
2884 0 : goto fail;
2885 : }
2886 :
2887 0 : poller->mr_map = spdk_rdma_utils_create_mem_map(poller->pd, &g_nvme_hooks,
2888 : IBV_ACCESS_LOCAL_WRITE | IBV_ACCESS_REMOTE_READ | IBV_ACCESS_REMOTE_WRITE);
2889 0 : if (poller->mr_map == NULL) {
2890 0 : SPDK_ERRLOG("Unable to create memory map.\n");
2891 0 : goto fail;
2892 : }
2893 :
2894 0 : srq_init_attr.stats = &poller->stats.rdma_stats.recv;
2895 0 : srq_init_attr.pd = poller->pd;
2896 0 : srq_init_attr.srq_init_attr.attr.max_wr = spdk_min((uint32_t)dev_attr.max_srq_wr,
2897 : g_spdk_nvme_transport_opts.rdma_srq_size);
2898 0 : srq_init_attr.srq_init_attr.attr.max_sge = spdk_min(dev_attr.max_sge,
2899 : NVME_RDMA_DEFAULT_RX_SGE);
2900 :
2901 0 : poller->srq = spdk_rdma_provider_srq_create(&srq_init_attr);
2902 0 : if (poller->srq == NULL) {
2903 0 : SPDK_ERRLOG("Unable to create SRQ.\n");
2904 0 : goto fail;
2905 : }
2906 :
2907 0 : opts.num_entries = g_spdk_nvme_transport_opts.rdma_srq_size;
2908 0 : opts.rqpair = NULL;
2909 0 : opts.srq = poller->srq;
2910 0 : opts.mr_map = poller->mr_map;
2911 :
2912 0 : poller->rsps = nvme_rdma_create_rsps(&opts);
2913 0 : if (poller->rsps == NULL) {
2914 0 : SPDK_ERRLOG("Unable to create poller RDMA responses.\n");
2915 0 : goto fail;
2916 : }
2917 :
2918 0 : rc = nvme_rdma_poller_submit_recvs(poller);
2919 0 : if (rc) {
2920 0 : SPDK_ERRLOG("Unable to submit poller RDMA responses.\n");
2921 0 : goto fail;
2922 : }
2923 :
2924 : /*
2925 : * When using an srq, fix the size of the completion queue at startup.
2926 : * The initiator sends only send and recv WRs. Hence, the multiplier is 2.
2927 : * (The target sends also data WRs. Hence, the multiplier is 3.)
2928 : */
2929 0 : num_cqe = g_spdk_nvme_transport_opts.rdma_srq_size * 2;
2930 : } else {
2931 9 : num_cqe = DEFAULT_NVME_RDMA_CQ_SIZE;
2932 : }
2933 :
2934 9 : max_num_cqe = g_spdk_nvme_transport_opts.rdma_max_cq_size;
2935 9 : if (max_num_cqe != 0 && num_cqe > max_num_cqe) {
2936 0 : num_cqe = max_num_cqe;
2937 : }
2938 :
2939 9 : poller->cq = ibv_create_cq(poller->device, num_cqe, group, NULL, 0);
2940 :
2941 9 : if (poller->cq == NULL) {
2942 2 : SPDK_ERRLOG("Unable to create CQ, errno %d.\n", errno);
2943 2 : goto fail;
2944 : }
2945 :
2946 7 : STAILQ_INSERT_HEAD(&group->pollers, poller, link);
2947 7 : group->num_pollers++;
2948 7 : poller->current_num_wc = num_cqe;
2949 7 : poller->required_num_wc = 0;
2950 7 : return poller;
2951 :
2952 2 : fail:
2953 2 : nvme_rdma_poller_destroy(poller);
2954 2 : return NULL;
2955 : }
2956 :
2957 : static void
2958 3 : nvme_rdma_poll_group_free_pollers(struct nvme_rdma_poll_group *group)
2959 : {
2960 : struct nvme_rdma_poller *poller, *tmp_poller;
2961 :
2962 5 : STAILQ_FOREACH_SAFE(poller, &group->pollers, link, tmp_poller) {
2963 2 : assert(poller->refcnt == 0);
2964 2 : if (poller->refcnt) {
2965 0 : SPDK_WARNLOG("Destroying poller with non-zero ref count: poller %p, refcnt %d\n",
2966 : poller, poller->refcnt);
2967 : }
2968 :
2969 2 : STAILQ_REMOVE(&group->pollers, poller, nvme_rdma_poller, link);
2970 2 : nvme_rdma_poller_destroy(poller);
2971 : }
2972 3 : }
2973 :
2974 : static struct nvme_rdma_poller *
2975 8 : nvme_rdma_poll_group_get_poller(struct nvme_rdma_poll_group *group, struct ibv_context *device)
2976 : {
2977 8 : struct nvme_rdma_poller *poller = NULL;
2978 :
2979 10 : STAILQ_FOREACH(poller, &group->pollers, link) {
2980 3 : if (poller->device == device) {
2981 1 : break;
2982 : }
2983 : }
2984 :
2985 8 : if (!poller) {
2986 7 : poller = nvme_rdma_poller_create(group, device);
2987 7 : if (!poller) {
2988 2 : SPDK_ERRLOG("Failed to create a poller for device %p\n", device);
2989 2 : return NULL;
2990 : }
2991 : }
2992 :
2993 6 : poller->refcnt++;
2994 6 : return poller;
2995 : }
2996 :
2997 : static void
2998 6 : nvme_rdma_poll_group_put_poller(struct nvme_rdma_poll_group *group, struct nvme_rdma_poller *poller)
2999 : {
3000 6 : assert(poller->refcnt > 0);
3001 6 : if (--poller->refcnt == 0) {
3002 5 : STAILQ_REMOVE(&group->pollers, poller, nvme_rdma_poller, link);
3003 5 : group->num_pollers--;
3004 5 : nvme_rdma_poller_destroy(poller);
3005 : }
3006 6 : }
3007 :
3008 : static struct spdk_nvme_transport_poll_group *
3009 1 : nvme_rdma_poll_group_create(void)
3010 : {
3011 : struct nvme_rdma_poll_group *group;
3012 :
3013 1 : group = calloc(1, sizeof(*group));
3014 1 : if (group == NULL) {
3015 0 : SPDK_ERRLOG("Unable to allocate poll group.\n");
3016 0 : return NULL;
3017 : }
3018 :
3019 1 : STAILQ_INIT(&group->pollers);
3020 1 : TAILQ_INIT(&group->connecting_qpairs);
3021 1 : TAILQ_INIT(&group->active_qpairs);
3022 1 : return &group->group;
3023 : }
3024 :
3025 : static int
3026 0 : nvme_rdma_poll_group_connect_qpair(struct spdk_nvme_qpair *qpair)
3027 : {
3028 0 : return 0;
3029 : }
3030 :
3031 : static int
3032 0 : nvme_rdma_poll_group_disconnect_qpair(struct spdk_nvme_qpair *qpair)
3033 : {
3034 0 : struct nvme_rdma_qpair *rqpair = nvme_rdma_qpair(qpair);
3035 0 : struct nvme_rdma_poll_group *group = nvme_rdma_poll_group(qpair->poll_group);
3036 :
3037 0 : if (rqpair->link_connecting.tqe_prev) {
3038 0 : TAILQ_REMOVE(&group->connecting_qpairs, rqpair, link_connecting);
3039 : /* We use prev pointer to check if qpair is in connecting list or not .
3040 : * TAILQ_REMOVE doesn't do it. So, we do it manually.
3041 : */
3042 0 : rqpair->link_connecting.tqe_prev = NULL;
3043 : }
3044 :
3045 0 : return 0;
3046 : }
3047 :
3048 : static int
3049 0 : nvme_rdma_poll_group_add(struct spdk_nvme_transport_poll_group *tgroup,
3050 : struct spdk_nvme_qpair *qpair)
3051 : {
3052 0 : return 0;
3053 : }
3054 :
3055 : static int
3056 0 : nvme_rdma_poll_group_remove(struct spdk_nvme_transport_poll_group *tgroup,
3057 : struct spdk_nvme_qpair *qpair)
3058 : {
3059 0 : struct nvme_rdma_qpair *rqpair = nvme_rdma_qpair(qpair);
3060 0 : struct nvme_rdma_poll_group *group = nvme_rdma_poll_group(qpair->poll_group);
3061 :
3062 0 : if (rqpair->link_active.tqe_prev) {
3063 0 : TAILQ_REMOVE(&group->active_qpairs, rqpair, link_active);
3064 0 : rqpair->link_active.tqe_prev = NULL;
3065 : }
3066 :
3067 0 : return 0;
3068 : }
3069 :
3070 : static inline void
3071 0 : nvme_rdma_qpair_process_submits(struct nvme_rdma_poll_group *group,
3072 : struct nvme_rdma_qpair *rqpair)
3073 : {
3074 0 : struct spdk_nvme_qpair *qpair = &rqpair->qpair;
3075 :
3076 0 : assert(rqpair->link_active.tqe_prev != NULL);
3077 :
3078 0 : if (spdk_unlikely(rqpair->state <= NVME_RDMA_QPAIR_STATE_INITIALIZING ||
3079 : rqpair->state >= NVME_RDMA_QPAIR_STATE_EXITING)) {
3080 0 : return;
3081 : }
3082 :
3083 0 : if (spdk_unlikely(qpair->ctrlr->timeout_enabled)) {
3084 0 : nvme_rdma_qpair_check_timeout(qpair);
3085 : }
3086 :
3087 0 : nvme_rdma_qpair_submit_sends(rqpair);
3088 0 : if (!rqpair->srq) {
3089 0 : nvme_rdma_qpair_submit_recvs(rqpair);
3090 : }
3091 0 : if (rqpair->num_completions > 0) {
3092 0 : nvme_qpair_resubmit_requests(qpair, rqpair->num_completions);
3093 0 : rqpair->num_completions = 0;
3094 : }
3095 :
3096 0 : if (rqpair->num_outstanding_reqs == 0 && STAILQ_EMPTY(&qpair->queued_req)) {
3097 0 : TAILQ_REMOVE(&group->active_qpairs, rqpair, link_active);
3098 : /* We use prev pointer to check if qpair is in active list or not.
3099 : * TAILQ_REMOVE doesn't do it. So, we do it manually.
3100 : */
3101 0 : rqpair->link_active.tqe_prev = NULL;
3102 : }
3103 : }
3104 :
3105 : static int64_t
3106 0 : nvme_rdma_poll_group_process_completions(struct spdk_nvme_transport_poll_group *tgroup,
3107 : uint32_t completions_per_qpair, spdk_nvme_disconnected_qpair_cb disconnected_qpair_cb)
3108 : {
3109 : struct spdk_nvme_qpair *qpair, *tmp_qpair;
3110 : struct nvme_rdma_qpair *rqpair, *tmp_rqpair;
3111 : struct nvme_rdma_poll_group *group;
3112 : struct nvme_rdma_poller *poller;
3113 0 : int batch_size, rc, rc2 = 0;
3114 0 : int64_t total_completions = 0;
3115 0 : uint64_t completions_allowed = 0;
3116 0 : uint64_t completions_per_poller = 0;
3117 0 : uint64_t poller_completions = 0;
3118 0 : uint64_t rdma_completions;
3119 :
3120 0 : if (completions_per_qpair == 0) {
3121 0 : completions_per_qpair = MAX_COMPLETIONS_PER_POLL;
3122 : }
3123 :
3124 0 : group = nvme_rdma_poll_group(tgroup);
3125 :
3126 0 : STAILQ_FOREACH_SAFE(qpair, &tgroup->disconnected_qpairs, poll_group_stailq, tmp_qpair) {
3127 0 : rc = nvme_rdma_ctrlr_disconnect_qpair_poll(qpair->ctrlr, qpair);
3128 0 : if (rc == 0) {
3129 0 : disconnected_qpair_cb(qpair, tgroup->group->ctx);
3130 : }
3131 : }
3132 :
3133 0 : TAILQ_FOREACH_SAFE(rqpair, &group->connecting_qpairs, link_connecting, tmp_rqpair) {
3134 0 : qpair = &rqpair->qpair;
3135 :
3136 0 : rc = nvme_rdma_ctrlr_connect_qpair_poll(qpair->ctrlr, qpair);
3137 0 : if (rc == 0 || rc != -EAGAIN) {
3138 0 : TAILQ_REMOVE(&group->connecting_qpairs, rqpair, link_connecting);
3139 : /* We use prev pointer to check if qpair is in connecting list or not.
3140 : * TAILQ_REMOVE does not do it. So, we do it manually.
3141 : */
3142 0 : rqpair->link_connecting.tqe_prev = NULL;
3143 :
3144 0 : if (rc == 0) {
3145 : /* Once the connection is completed, we can submit queued requests */
3146 0 : nvme_qpair_resubmit_requests(qpair, rqpair->num_entries);
3147 0 : } else if (rc != -EAGAIN) {
3148 0 : SPDK_ERRLOG("Failed to connect rqpair=%p\n", rqpair);
3149 0 : nvme_rdma_fail_qpair(qpair, 0);
3150 : }
3151 : }
3152 : }
3153 :
3154 0 : STAILQ_FOREACH_SAFE(qpair, &tgroup->connected_qpairs, poll_group_stailq, tmp_qpair) {
3155 0 : rqpair = nvme_rdma_qpair(qpair);
3156 :
3157 0 : if (spdk_likely(nvme_qpair_get_state(qpair) != NVME_QPAIR_CONNECTING)) {
3158 0 : nvme_rdma_qpair_process_cm_event(rqpair);
3159 : }
3160 :
3161 0 : if (spdk_unlikely(qpair->transport_failure_reason != SPDK_NVME_QPAIR_FAILURE_NONE)) {
3162 0 : rc2 = -ENXIO;
3163 0 : nvme_rdma_fail_qpair(qpair, 0);
3164 : }
3165 : }
3166 :
3167 0 : completions_allowed = completions_per_qpair * tgroup->num_connected_qpairs;
3168 0 : if (group->num_pollers) {
3169 0 : completions_per_poller = spdk_max(completions_allowed / group->num_pollers, 1);
3170 : }
3171 :
3172 0 : STAILQ_FOREACH(poller, &group->pollers, link) {
3173 0 : poller_completions = 0;
3174 0 : rdma_completions = 0;
3175 : do {
3176 0 : poller->stats.polls++;
3177 0 : batch_size = spdk_min((completions_per_poller - poller_completions), MAX_COMPLETIONS_PER_POLL);
3178 0 : rc = nvme_rdma_cq_process_completions(poller->cq, batch_size, poller, NULL, &rdma_completions);
3179 0 : if (rc <= 0) {
3180 0 : if (rc == -ECANCELED) {
3181 0 : return -EIO;
3182 0 : } else if (rc == 0) {
3183 0 : poller->stats.idle_polls++;
3184 : }
3185 0 : break;
3186 : }
3187 :
3188 0 : poller_completions += rc;
3189 0 : } while (poller_completions < completions_per_poller);
3190 0 : total_completions += poller_completions;
3191 0 : poller->stats.completions += rdma_completions;
3192 0 : if (poller->srq) {
3193 0 : nvme_rdma_poller_submit_recvs(poller);
3194 : }
3195 : }
3196 :
3197 0 : TAILQ_FOREACH_SAFE(rqpair, &group->active_qpairs, link_active, tmp_rqpair) {
3198 0 : nvme_rdma_qpair_process_submits(group, rqpair);
3199 : }
3200 :
3201 0 : return rc2 != 0 ? rc2 : total_completions;
3202 : }
3203 :
3204 : static int
3205 1 : nvme_rdma_poll_group_destroy(struct spdk_nvme_transport_poll_group *tgroup)
3206 : {
3207 1 : struct nvme_rdma_poll_group *group = nvme_rdma_poll_group(tgroup);
3208 :
3209 1 : if (!STAILQ_EMPTY(&tgroup->connected_qpairs) || !STAILQ_EMPTY(&tgroup->disconnected_qpairs)) {
3210 0 : return -EBUSY;
3211 : }
3212 :
3213 1 : nvme_rdma_poll_group_free_pollers(group);
3214 1 : free(group);
3215 :
3216 1 : return 0;
3217 : }
3218 :
3219 : static int
3220 3 : nvme_rdma_poll_group_get_stats(struct spdk_nvme_transport_poll_group *tgroup,
3221 : struct spdk_nvme_transport_poll_group_stat **_stats)
3222 : {
3223 : struct nvme_rdma_poll_group *group;
3224 : struct spdk_nvme_transport_poll_group_stat *stats;
3225 : struct spdk_nvme_rdma_device_stat *device_stat;
3226 : struct nvme_rdma_poller *poller;
3227 3 : uint32_t i = 0;
3228 :
3229 3 : if (tgroup == NULL || _stats == NULL) {
3230 2 : SPDK_ERRLOG("Invalid stats or group pointer\n");
3231 2 : return -EINVAL;
3232 : }
3233 :
3234 1 : group = nvme_rdma_poll_group(tgroup);
3235 1 : stats = calloc(1, sizeof(*stats));
3236 1 : if (!stats) {
3237 0 : SPDK_ERRLOG("Can't allocate memory for RDMA stats\n");
3238 0 : return -ENOMEM;
3239 : }
3240 1 : stats->trtype = SPDK_NVME_TRANSPORT_RDMA;
3241 1 : stats->rdma.num_devices = group->num_pollers;
3242 :
3243 1 : if (stats->rdma.num_devices == 0) {
3244 0 : *_stats = stats;
3245 0 : return 0;
3246 : }
3247 :
3248 1 : stats->rdma.device_stats = calloc(stats->rdma.num_devices, sizeof(*stats->rdma.device_stats));
3249 1 : if (!stats->rdma.device_stats) {
3250 0 : SPDK_ERRLOG("Can't allocate memory for RDMA device stats\n");
3251 0 : free(stats);
3252 0 : return -ENOMEM;
3253 : }
3254 :
3255 3 : STAILQ_FOREACH(poller, &group->pollers, link) {
3256 2 : device_stat = &stats->rdma.device_stats[i];
3257 2 : device_stat->name = poller->device->device->name;
3258 2 : device_stat->polls = poller->stats.polls;
3259 2 : device_stat->idle_polls = poller->stats.idle_polls;
3260 2 : device_stat->completions = poller->stats.completions;
3261 2 : device_stat->queued_requests = poller->stats.queued_requests;
3262 2 : device_stat->total_send_wrs = poller->stats.rdma_stats.send.num_submitted_wrs;
3263 2 : device_stat->send_doorbell_updates = poller->stats.rdma_stats.send.doorbell_updates;
3264 2 : device_stat->total_recv_wrs = poller->stats.rdma_stats.recv.num_submitted_wrs;
3265 2 : device_stat->recv_doorbell_updates = poller->stats.rdma_stats.recv.doorbell_updates;
3266 2 : i++;
3267 : }
3268 :
3269 1 : *_stats = stats;
3270 :
3271 1 : return 0;
3272 : }
3273 :
3274 : static void
3275 1 : nvme_rdma_poll_group_free_stats(struct spdk_nvme_transport_poll_group *tgroup,
3276 : struct spdk_nvme_transport_poll_group_stat *stats)
3277 : {
3278 1 : if (stats) {
3279 1 : free(stats->rdma.device_stats);
3280 : }
3281 1 : free(stats);
3282 1 : }
3283 :
3284 : static int
3285 4 : nvme_rdma_ctrlr_get_memory_domains(const struct spdk_nvme_ctrlr *ctrlr,
3286 : struct spdk_memory_domain **domains, int array_size)
3287 : {
3288 4 : struct nvme_rdma_qpair *rqpair = nvme_rdma_qpair(ctrlr->adminq);
3289 :
3290 4 : if (domains && array_size > 0) {
3291 1 : domains[0] = rqpair->memory_domain;
3292 : }
3293 :
3294 4 : return 1;
3295 : }
3296 :
3297 : void
3298 0 : spdk_nvme_rdma_init_hooks(struct spdk_nvme_rdma_hooks *hooks)
3299 : {
3300 0 : g_nvme_hooks = *hooks;
3301 0 : }
3302 :
3303 : const struct spdk_nvme_transport_ops rdma_ops = {
3304 : .name = "RDMA",
3305 : .type = SPDK_NVME_TRANSPORT_RDMA,
3306 : .ctrlr_construct = nvme_rdma_ctrlr_construct,
3307 : .ctrlr_scan = nvme_fabric_ctrlr_scan,
3308 : .ctrlr_destruct = nvme_rdma_ctrlr_destruct,
3309 : .ctrlr_enable = nvme_rdma_ctrlr_enable,
3310 :
3311 : .ctrlr_set_reg_4 = nvme_fabric_ctrlr_set_reg_4,
3312 : .ctrlr_set_reg_8 = nvme_fabric_ctrlr_set_reg_8,
3313 : .ctrlr_get_reg_4 = nvme_fabric_ctrlr_get_reg_4,
3314 : .ctrlr_get_reg_8 = nvme_fabric_ctrlr_get_reg_8,
3315 : .ctrlr_set_reg_4_async = nvme_fabric_ctrlr_set_reg_4_async,
3316 : .ctrlr_set_reg_8_async = nvme_fabric_ctrlr_set_reg_8_async,
3317 : .ctrlr_get_reg_4_async = nvme_fabric_ctrlr_get_reg_4_async,
3318 : .ctrlr_get_reg_8_async = nvme_fabric_ctrlr_get_reg_8_async,
3319 :
3320 : .ctrlr_get_max_xfer_size = nvme_rdma_ctrlr_get_max_xfer_size,
3321 : .ctrlr_get_max_sges = nvme_rdma_ctrlr_get_max_sges,
3322 :
3323 : .ctrlr_create_io_qpair = nvme_rdma_ctrlr_create_io_qpair,
3324 : .ctrlr_delete_io_qpair = nvme_rdma_ctrlr_delete_io_qpair,
3325 : .ctrlr_connect_qpair = nvme_rdma_ctrlr_connect_qpair,
3326 : .ctrlr_disconnect_qpair = nvme_rdma_ctrlr_disconnect_qpair,
3327 :
3328 : .ctrlr_get_memory_domains = nvme_rdma_ctrlr_get_memory_domains,
3329 :
3330 : .qpair_abort_reqs = nvme_rdma_qpair_abort_reqs,
3331 : .qpair_reset = nvme_rdma_qpair_reset,
3332 : .qpair_submit_request = nvme_rdma_qpair_submit_request,
3333 : .qpair_process_completions = nvme_rdma_qpair_process_completions,
3334 : .qpair_iterate_requests = nvme_rdma_qpair_iterate_requests,
3335 : .admin_qpair_abort_aers = nvme_rdma_admin_qpair_abort_aers,
3336 :
3337 : .poll_group_create = nvme_rdma_poll_group_create,
3338 : .poll_group_connect_qpair = nvme_rdma_poll_group_connect_qpair,
3339 : .poll_group_disconnect_qpair = nvme_rdma_poll_group_disconnect_qpair,
3340 : .poll_group_add = nvme_rdma_poll_group_add,
3341 : .poll_group_remove = nvme_rdma_poll_group_remove,
3342 : .poll_group_process_completions = nvme_rdma_poll_group_process_completions,
3343 : .poll_group_destroy = nvme_rdma_poll_group_destroy,
3344 : .poll_group_get_stats = nvme_rdma_poll_group_get_stats,
3345 : .poll_group_free_stats = nvme_rdma_poll_group_free_stats,
3346 : };
3347 :
3348 1 : SPDK_NVME_TRANSPORT_REGISTER(rdma, &rdma_ops);
|