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) 2021-2023 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 :
25 : #include "nvme_internal.h"
26 : #include "spdk_internal/rdma.h"
27 :
28 : #define NVME_RDMA_TIME_OUT_IN_MS 2000
29 : #define NVME_RDMA_RW_BUFFER_SIZE 131072
30 :
31 : /*
32 : * NVME RDMA qpair Resource Defaults
33 : */
34 : #define NVME_RDMA_DEFAULT_TX_SGE 2
35 : #define NVME_RDMA_DEFAULT_RX_SGE 1
36 :
37 : /* Max number of NVMe-oF SGL descriptors supported by the host */
38 : #define NVME_RDMA_MAX_SGL_DESCRIPTORS 16
39 :
40 : /* number of STAILQ entries for holding pending RDMA CM events. */
41 : #define NVME_RDMA_NUM_CM_EVENTS 256
42 :
43 : /* CM event processing timeout */
44 : #define NVME_RDMA_QPAIR_CM_EVENT_TIMEOUT_US 1000000
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 : struct nvme_rdma_memory_domain {
82 : TAILQ_ENTRY(nvme_rdma_memory_domain) link;
83 : uint32_t ref;
84 : struct ibv_pd *pd;
85 : struct spdk_memory_domain *domain;
86 : struct spdk_memory_domain_rdma_ctx rdma_ctx;
87 : };
88 :
89 : enum nvme_rdma_wr_type {
90 : RDMA_WR_TYPE_RECV,
91 : RDMA_WR_TYPE_SEND,
92 : };
93 :
94 : struct nvme_rdma_wr {
95 : /* Using this instead of the enum allows this struct to only occupy one byte. */
96 : uint8_t type;
97 : };
98 :
99 : struct spdk_nvmf_cmd {
100 : struct spdk_nvme_cmd cmd;
101 : struct spdk_nvme_sgl_descriptor sgl[NVME_RDMA_MAX_SGL_DESCRIPTORS];
102 : };
103 :
104 : struct spdk_nvme_rdma_hooks g_nvme_hooks = {};
105 :
106 : /* STAILQ wrapper for cm events. */
107 : struct nvme_rdma_cm_event_entry {
108 : struct rdma_cm_event *evt;
109 : STAILQ_ENTRY(nvme_rdma_cm_event_entry) link;
110 : };
111 :
112 : /* NVMe RDMA transport extensions for spdk_nvme_ctrlr */
113 : struct nvme_rdma_ctrlr {
114 : struct spdk_nvme_ctrlr ctrlr;
115 :
116 : uint16_t max_sge;
117 :
118 : struct rdma_event_channel *cm_channel;
119 :
120 : STAILQ_HEAD(, nvme_rdma_cm_event_entry) pending_cm_events;
121 :
122 : STAILQ_HEAD(, nvme_rdma_cm_event_entry) free_cm_events;
123 :
124 : struct nvme_rdma_cm_event_entry *cm_events;
125 : };
126 :
127 : struct nvme_rdma_poller_stats {
128 : uint64_t polls;
129 : uint64_t idle_polls;
130 : uint64_t queued_requests;
131 : uint64_t completions;
132 : struct spdk_rdma_qp_stats rdma_stats;
133 : };
134 :
135 : struct nvme_rdma_poll_group;
136 : struct nvme_rdma_rsps;
137 :
138 : struct nvme_rdma_poller {
139 : struct ibv_context *device;
140 : struct ibv_cq *cq;
141 : struct spdk_rdma_srq *srq;
142 : struct nvme_rdma_rsps *rsps;
143 : struct ibv_pd *pd;
144 : struct spdk_rdma_mem_map *mr_map;
145 : uint32_t refcnt;
146 : int required_num_wc;
147 : int current_num_wc;
148 : struct nvme_rdma_poller_stats stats;
149 : struct nvme_rdma_poll_group *group;
150 : STAILQ_ENTRY(nvme_rdma_poller) link;
151 : };
152 :
153 : struct nvme_rdma_qpair;
154 :
155 : struct nvme_rdma_poll_group {
156 : struct spdk_nvme_transport_poll_group group;
157 : STAILQ_HEAD(, nvme_rdma_poller) pollers;
158 : uint32_t num_pollers;
159 : TAILQ_HEAD(, nvme_rdma_qpair) connecting_qpairs;
160 : };
161 :
162 : enum nvme_rdma_qpair_state {
163 : NVME_RDMA_QPAIR_STATE_INVALID = 0,
164 : NVME_RDMA_QPAIR_STATE_STALE_CONN,
165 : NVME_RDMA_QPAIR_STATE_INITIALIZING,
166 : NVME_RDMA_QPAIR_STATE_FABRIC_CONNECT_SEND,
167 : NVME_RDMA_QPAIR_STATE_FABRIC_CONNECT_POLL,
168 : NVME_RDMA_QPAIR_STATE_RUNNING,
169 : NVME_RDMA_QPAIR_STATE_EXITING,
170 : NVME_RDMA_QPAIR_STATE_LINGERING,
171 : NVME_RDMA_QPAIR_STATE_EXITED,
172 : };
173 :
174 : typedef int (*nvme_rdma_cm_event_cb)(struct nvme_rdma_qpair *rqpair, int ret);
175 :
176 : struct nvme_rdma_rsp_opts {
177 : uint16_t num_entries;
178 : struct nvme_rdma_qpair *rqpair;
179 : struct spdk_rdma_srq *srq;
180 : struct spdk_rdma_mem_map *mr_map;
181 : };
182 :
183 : struct nvme_rdma_rsps {
184 : /* Parallel arrays of response buffers + response SGLs of size num_entries */
185 : struct ibv_sge *rsp_sgls;
186 : struct spdk_nvme_rdma_rsp *rsps;
187 :
188 : struct ibv_recv_wr *rsp_recv_wrs;
189 :
190 : /* Count of outstanding recv objects */
191 : uint16_t current_num_recvs;
192 :
193 : uint16_t num_entries;
194 : };
195 :
196 : /* NVMe RDMA qpair extensions for spdk_nvme_qpair */
197 : struct nvme_rdma_qpair {
198 : struct spdk_nvme_qpair qpair;
199 :
200 : struct spdk_rdma_qp *rdma_qp;
201 : struct rdma_cm_id *cm_id;
202 : struct ibv_cq *cq;
203 : struct spdk_rdma_srq *srq;
204 :
205 : struct spdk_nvme_rdma_req *rdma_reqs;
206 :
207 : uint32_t max_send_sge;
208 :
209 : uint32_t max_recv_sge;
210 :
211 : uint16_t num_entries;
212 :
213 : bool delay_cmd_submit;
214 :
215 : uint32_t num_completions;
216 :
217 : struct nvme_rdma_rsps *rsps;
218 :
219 : /*
220 : * Array of num_entries NVMe commands registered as RDMA message buffers.
221 : * Indexed by rdma_req->id.
222 : */
223 : struct spdk_nvmf_cmd *cmds;
224 :
225 : struct spdk_rdma_mem_map *mr_map;
226 :
227 : TAILQ_HEAD(, spdk_nvme_rdma_req) free_reqs;
228 : TAILQ_HEAD(, spdk_nvme_rdma_req) outstanding_reqs;
229 :
230 : struct nvme_rdma_memory_domain *memory_domain;
231 :
232 : /* Count of outstanding send objects */
233 : uint16_t current_num_sends;
234 :
235 : /* Placed at the end of the struct since it is not used frequently */
236 : struct rdma_cm_event *evt;
237 : struct nvme_rdma_poller *poller;
238 :
239 : uint64_t evt_timeout_ticks;
240 : nvme_rdma_cm_event_cb evt_cb;
241 : enum rdma_cm_event_type expected_evt_type;
242 :
243 : enum nvme_rdma_qpair_state state;
244 :
245 : bool in_connect_poll;
246 :
247 : uint8_t stale_conn_retry_count;
248 : bool need_destroy;
249 :
250 : TAILQ_ENTRY(nvme_rdma_qpair) link_connecting;
251 : };
252 :
253 : enum NVME_RDMA_COMPLETION_FLAGS {
254 : NVME_RDMA_SEND_COMPLETED = 1u << 0,
255 : NVME_RDMA_RECV_COMPLETED = 1u << 1,
256 : };
257 :
258 : struct spdk_nvme_rdma_req {
259 : uint16_t id;
260 : uint16_t completion_flags: 2;
261 : uint16_t reserved: 14;
262 : /* if completion of RDMA_RECV received before RDMA_SEND, we will complete nvme request
263 : * during processing of RDMA_SEND. To complete the request we must know the response
264 : * received in RDMA_RECV, so store it in this field */
265 : struct spdk_nvme_rdma_rsp *rdma_rsp;
266 :
267 : struct nvme_rdma_wr rdma_wr;
268 :
269 : struct ibv_send_wr send_wr;
270 :
271 : struct nvme_request *req;
272 :
273 : struct ibv_sge send_sgl[NVME_RDMA_DEFAULT_TX_SGE];
274 :
275 : TAILQ_ENTRY(spdk_nvme_rdma_req) link;
276 : };
277 :
278 : struct spdk_nvme_rdma_rsp {
279 : struct spdk_nvme_cpl cpl;
280 : struct nvme_rdma_qpair *rqpair;
281 : struct ibv_recv_wr *recv_wr;
282 : struct nvme_rdma_wr rdma_wr;
283 : };
284 :
285 : struct nvme_rdma_memory_translation_ctx {
286 : void *addr;
287 : size_t length;
288 : uint32_t lkey;
289 : uint32_t rkey;
290 : };
291 :
292 : static const char *rdma_cm_event_str[] = {
293 : "RDMA_CM_EVENT_ADDR_RESOLVED",
294 : "RDMA_CM_EVENT_ADDR_ERROR",
295 : "RDMA_CM_EVENT_ROUTE_RESOLVED",
296 : "RDMA_CM_EVENT_ROUTE_ERROR",
297 : "RDMA_CM_EVENT_CONNECT_REQUEST",
298 : "RDMA_CM_EVENT_CONNECT_RESPONSE",
299 : "RDMA_CM_EVENT_CONNECT_ERROR",
300 : "RDMA_CM_EVENT_UNREACHABLE",
301 : "RDMA_CM_EVENT_REJECTED",
302 : "RDMA_CM_EVENT_ESTABLISHED",
303 : "RDMA_CM_EVENT_DISCONNECTED",
304 : "RDMA_CM_EVENT_DEVICE_REMOVAL",
305 : "RDMA_CM_EVENT_MULTICAST_JOIN",
306 : "RDMA_CM_EVENT_MULTICAST_ERROR",
307 : "RDMA_CM_EVENT_ADDR_CHANGE",
308 : "RDMA_CM_EVENT_TIMEWAIT_EXIT"
309 : };
310 :
311 : static struct nvme_rdma_poller *nvme_rdma_poll_group_get_poller(struct nvme_rdma_poll_group *group,
312 : struct ibv_context *device);
313 : static void nvme_rdma_poll_group_put_poller(struct nvme_rdma_poll_group *group,
314 : struct nvme_rdma_poller *poller);
315 :
316 : static TAILQ_HEAD(, nvme_rdma_memory_domain) g_memory_domains = TAILQ_HEAD_INITIALIZER(
317 : g_memory_domains);
318 : static pthread_mutex_t g_memory_domains_lock = PTHREAD_MUTEX_INITIALIZER;
319 :
320 : static struct nvme_rdma_memory_domain *
321 6 : nvme_rdma_get_memory_domain(struct ibv_pd *pd)
322 : {
323 6 : struct nvme_rdma_memory_domain *domain = NULL;
324 6 : struct spdk_memory_domain_ctx ctx;
325 : int rc;
326 :
327 6 : pthread_mutex_lock(&g_memory_domains_lock);
328 :
329 11 : TAILQ_FOREACH(domain, &g_memory_domains, link) {
330 7 : if (domain->pd == pd) {
331 2 : domain->ref++;
332 2 : pthread_mutex_unlock(&g_memory_domains_lock);
333 2 : return domain;
334 : }
335 : }
336 :
337 4 : domain = calloc(1, sizeof(*domain));
338 4 : if (!domain) {
339 0 : SPDK_ERRLOG("Memory allocation failed\n");
340 0 : pthread_mutex_unlock(&g_memory_domains_lock);
341 0 : return NULL;
342 : }
343 :
344 4 : domain->rdma_ctx.size = sizeof(domain->rdma_ctx);
345 4 : domain->rdma_ctx.ibv_pd = pd;
346 4 : ctx.size = sizeof(ctx);
347 4 : ctx.user_ctx = &domain->rdma_ctx;
348 :
349 4 : rc = spdk_memory_domain_create(&domain->domain, SPDK_DMA_DEVICE_TYPE_RDMA, &ctx,
350 : SPDK_RDMA_DMA_DEVICE);
351 4 : if (rc) {
352 1 : SPDK_ERRLOG("Failed to create memory domain\n");
353 1 : free(domain);
354 1 : pthread_mutex_unlock(&g_memory_domains_lock);
355 1 : return NULL;
356 : }
357 :
358 3 : domain->pd = pd;
359 3 : domain->ref = 1;
360 3 : TAILQ_INSERT_TAIL(&g_memory_domains, domain, link);
361 :
362 3 : pthread_mutex_unlock(&g_memory_domains_lock);
363 :
364 3 : return domain;
365 : }
366 :
367 : static void
368 5 : nvme_rdma_put_memory_domain(struct nvme_rdma_memory_domain *device)
369 : {
370 5 : if (!device) {
371 1 : return;
372 : }
373 :
374 4 : pthread_mutex_lock(&g_memory_domains_lock);
375 :
376 4 : assert(device->ref > 0);
377 :
378 4 : device->ref--;
379 :
380 4 : if (device->ref == 0) {
381 2 : spdk_memory_domain_destroy(device->domain);
382 2 : TAILQ_REMOVE(&g_memory_domains, device, link);
383 2 : free(device);
384 : }
385 :
386 4 : pthread_mutex_unlock(&g_memory_domains_lock);
387 : }
388 :
389 : static int nvme_rdma_ctrlr_delete_io_qpair(struct spdk_nvme_ctrlr *ctrlr,
390 : struct spdk_nvme_qpair *qpair);
391 :
392 : static inline struct nvme_rdma_qpair *
393 18 : nvme_rdma_qpair(struct spdk_nvme_qpair *qpair)
394 : {
395 18 : assert(qpair->trtype == SPDK_NVME_TRANSPORT_RDMA);
396 18 : return SPDK_CONTAINEROF(qpair, struct nvme_rdma_qpair, qpair);
397 : }
398 :
399 : static inline struct nvme_rdma_poll_group *
400 8 : nvme_rdma_poll_group(struct spdk_nvme_transport_poll_group *group)
401 : {
402 8 : return (SPDK_CONTAINEROF(group, struct nvme_rdma_poll_group, group));
403 : }
404 :
405 : static inline struct nvme_rdma_ctrlr *
406 8 : nvme_rdma_ctrlr(struct spdk_nvme_ctrlr *ctrlr)
407 : {
408 8 : assert(ctrlr->trid.trtype == SPDK_NVME_TRANSPORT_RDMA);
409 8 : return SPDK_CONTAINEROF(ctrlr, struct nvme_rdma_ctrlr, ctrlr);
410 : }
411 :
412 : static struct spdk_nvme_rdma_req *
413 3 : nvme_rdma_req_get(struct nvme_rdma_qpair *rqpair)
414 : {
415 : struct spdk_nvme_rdma_req *rdma_req;
416 :
417 3 : rdma_req = TAILQ_FIRST(&rqpair->free_reqs);
418 3 : if (rdma_req) {
419 2 : TAILQ_REMOVE(&rqpair->free_reqs, rdma_req, link);
420 : }
421 :
422 3 : return rdma_req;
423 : }
424 :
425 : static void
426 1 : nvme_rdma_req_put(struct nvme_rdma_qpair *rqpair, struct spdk_nvme_rdma_req *rdma_req)
427 : {
428 1 : rdma_req->completion_flags = 0;
429 1 : rdma_req->req = NULL;
430 1 : rdma_req->rdma_rsp = NULL;
431 1 : TAILQ_INSERT_HEAD(&rqpair->free_reqs, rdma_req, link);
432 1 : }
433 :
434 : static void
435 0 : nvme_rdma_req_complete(struct spdk_nvme_rdma_req *rdma_req,
436 : struct spdk_nvme_cpl *rsp,
437 : bool print_on_error)
438 : {
439 0 : struct nvme_request *req = rdma_req->req;
440 : struct nvme_rdma_qpair *rqpair;
441 : struct spdk_nvme_qpair *qpair;
442 : bool error, print_error;
443 :
444 0 : assert(req != NULL);
445 :
446 0 : qpair = req->qpair;
447 0 : rqpair = nvme_rdma_qpair(qpair);
448 :
449 0 : error = spdk_nvme_cpl_is_error(rsp);
450 0 : print_error = error && print_on_error && !qpair->ctrlr->opts.disable_error_logging;
451 :
452 0 : if (print_error) {
453 0 : spdk_nvme_qpair_print_command(qpair, &req->cmd);
454 : }
455 :
456 0 : if (print_error || SPDK_DEBUGLOG_FLAG_ENABLED("nvme")) {
457 0 : spdk_nvme_qpair_print_completion(qpair, rsp);
458 : }
459 :
460 0 : TAILQ_REMOVE(&rqpair->outstanding_reqs, rdma_req, link);
461 :
462 0 : nvme_complete_request(req->cb_fn, req->cb_arg, qpair, req, rsp);
463 0 : nvme_rdma_req_put(rqpair, rdma_req);
464 0 : }
465 :
466 : static const char *
467 4 : nvme_rdma_cm_event_str_get(uint32_t event)
468 : {
469 4 : if (event < SPDK_COUNTOF(rdma_cm_event_str)) {
470 4 : return rdma_cm_event_str[event];
471 : } else {
472 0 : return "Undefined";
473 : }
474 : }
475 :
476 :
477 : static int
478 12 : nvme_rdma_qpair_process_cm_event(struct nvme_rdma_qpair *rqpair)
479 : {
480 12 : struct rdma_cm_event *event = rqpair->evt;
481 : struct spdk_nvmf_rdma_accept_private_data *accept_data;
482 12 : int rc = 0;
483 :
484 12 : if (event) {
485 12 : switch (event->event) {
486 1 : case RDMA_CM_EVENT_ADDR_RESOLVED:
487 : case RDMA_CM_EVENT_ADDR_ERROR:
488 : case RDMA_CM_EVENT_ROUTE_RESOLVED:
489 : case RDMA_CM_EVENT_ROUTE_ERROR:
490 1 : break;
491 1 : case RDMA_CM_EVENT_CONNECT_REQUEST:
492 1 : break;
493 1 : case RDMA_CM_EVENT_CONNECT_ERROR:
494 1 : break;
495 1 : case RDMA_CM_EVENT_UNREACHABLE:
496 : case RDMA_CM_EVENT_REJECTED:
497 1 : break;
498 2 : case RDMA_CM_EVENT_CONNECT_RESPONSE:
499 2 : rc = spdk_rdma_qp_complete_connect(rqpair->rdma_qp);
500 : /* fall through */
501 2 : case RDMA_CM_EVENT_ESTABLISHED:
502 2 : accept_data = (struct spdk_nvmf_rdma_accept_private_data *)event->param.conn.private_data;
503 2 : if (accept_data == NULL) {
504 1 : rc = -1;
505 : } else {
506 1 : SPDK_DEBUGLOG(nvme, "Requested queue depth %d. Target receive queue depth %d.\n",
507 : rqpair->num_entries + 1, accept_data->crqsize);
508 : }
509 2 : break;
510 1 : case RDMA_CM_EVENT_DISCONNECTED:
511 1 : rqpair->qpair.transport_failure_reason = SPDK_NVME_QPAIR_FAILURE_REMOTE;
512 1 : break;
513 1 : case RDMA_CM_EVENT_DEVICE_REMOVAL:
514 1 : rqpair->qpair.transport_failure_reason = SPDK_NVME_QPAIR_FAILURE_LOCAL;
515 1 : rqpair->need_destroy = true;
516 1 : break;
517 1 : case RDMA_CM_EVENT_MULTICAST_JOIN:
518 : case RDMA_CM_EVENT_MULTICAST_ERROR:
519 1 : break;
520 1 : case RDMA_CM_EVENT_ADDR_CHANGE:
521 1 : rqpair->qpair.transport_failure_reason = SPDK_NVME_QPAIR_FAILURE_LOCAL;
522 1 : break;
523 1 : case RDMA_CM_EVENT_TIMEWAIT_EXIT:
524 1 : break;
525 1 : default:
526 1 : SPDK_ERRLOG("Unexpected Acceptor Event [%d]\n", event->event);
527 1 : break;
528 : }
529 12 : rqpair->evt = NULL;
530 12 : rdma_ack_cm_event(event);
531 : }
532 :
533 12 : return rc;
534 : }
535 :
536 : /*
537 : * This function must be called under the nvme controller's lock
538 : * because it touches global controller variables. The lock is taken
539 : * by the generic transport code before invoking a few of the functions
540 : * in this file: nvme_rdma_ctrlr_connect_qpair, nvme_rdma_ctrlr_delete_io_qpair,
541 : * and conditionally nvme_rdma_qpair_process_completions when it is calling
542 : * completions on the admin qpair. When adding a new call to this function, please
543 : * verify that it is in a situation where it falls under the lock.
544 : */
545 : static int
546 0 : nvme_rdma_poll_events(struct nvme_rdma_ctrlr *rctrlr)
547 : {
548 : struct nvme_rdma_cm_event_entry *entry, *tmp;
549 : struct nvme_rdma_qpair *event_qpair;
550 0 : struct rdma_cm_event *event;
551 0 : struct rdma_event_channel *channel = rctrlr->cm_channel;
552 :
553 0 : STAILQ_FOREACH_SAFE(entry, &rctrlr->pending_cm_events, link, tmp) {
554 0 : event_qpair = entry->evt->id->context;
555 0 : if (event_qpair->evt == NULL) {
556 0 : event_qpair->evt = entry->evt;
557 0 : STAILQ_REMOVE(&rctrlr->pending_cm_events, entry, nvme_rdma_cm_event_entry, link);
558 0 : STAILQ_INSERT_HEAD(&rctrlr->free_cm_events, entry, link);
559 : }
560 : }
561 :
562 0 : while (rdma_get_cm_event(channel, &event) == 0) {
563 0 : event_qpair = event->id->context;
564 0 : if (event_qpair->evt == NULL) {
565 0 : event_qpair->evt = event;
566 : } else {
567 0 : assert(rctrlr == nvme_rdma_ctrlr(event_qpair->qpair.ctrlr));
568 0 : entry = STAILQ_FIRST(&rctrlr->free_cm_events);
569 0 : if (entry == NULL) {
570 0 : rdma_ack_cm_event(event);
571 0 : return -ENOMEM;
572 : }
573 0 : STAILQ_REMOVE(&rctrlr->free_cm_events, entry, nvme_rdma_cm_event_entry, link);
574 0 : entry->evt = event;
575 0 : STAILQ_INSERT_TAIL(&rctrlr->pending_cm_events, entry, link);
576 : }
577 : }
578 :
579 : /* rdma_get_cm_event() returns -1 on error. If an error occurs, errno
580 : * will be set to indicate the failure reason. So return negated errno here.
581 : */
582 0 : return -errno;
583 : }
584 :
585 : static int
586 4 : nvme_rdma_validate_cm_event(enum rdma_cm_event_type expected_evt_type,
587 : struct rdma_cm_event *reaped_evt)
588 : {
589 4 : int rc = -EBADMSG;
590 :
591 4 : if (expected_evt_type == reaped_evt->event) {
592 1 : return 0;
593 : }
594 :
595 3 : switch (expected_evt_type) {
596 2 : case RDMA_CM_EVENT_ESTABLISHED:
597 : /*
598 : * There is an enum ib_cm_rej_reason in the kernel headers that sets 10 as
599 : * IB_CM_REJ_STALE_CONN. I can't find the corresponding userspace but we get
600 : * the same values here.
601 : */
602 2 : if (reaped_evt->event == RDMA_CM_EVENT_REJECTED && reaped_evt->status == 10) {
603 1 : rc = -ESTALE;
604 1 : } else if (reaped_evt->event == RDMA_CM_EVENT_CONNECT_RESPONSE) {
605 : /*
606 : * If we are using a qpair which is not created using rdma cm API
607 : * then we will receive RDMA_CM_EVENT_CONNECT_RESPONSE instead of
608 : * RDMA_CM_EVENT_ESTABLISHED.
609 : */
610 1 : return 0;
611 : }
612 1 : break;
613 1 : default:
614 1 : break;
615 : }
616 :
617 2 : SPDK_ERRLOG("Expected %s but received %s (%d) from CM event channel (status = %d)\n",
618 : nvme_rdma_cm_event_str_get(expected_evt_type),
619 : nvme_rdma_cm_event_str_get(reaped_evt->event), reaped_evt->event,
620 : reaped_evt->status);
621 2 : return rc;
622 : }
623 :
624 : static int
625 0 : nvme_rdma_process_event_start(struct nvme_rdma_qpair *rqpair,
626 : enum rdma_cm_event_type evt,
627 : nvme_rdma_cm_event_cb evt_cb)
628 : {
629 : int rc;
630 :
631 0 : assert(evt_cb != NULL);
632 :
633 0 : if (rqpair->evt != NULL) {
634 0 : rc = nvme_rdma_qpair_process_cm_event(rqpair);
635 0 : if (rc) {
636 0 : return rc;
637 : }
638 : }
639 :
640 0 : rqpair->expected_evt_type = evt;
641 0 : rqpair->evt_cb = evt_cb;
642 0 : rqpair->evt_timeout_ticks = (NVME_RDMA_QPAIR_CM_EVENT_TIMEOUT_US * spdk_get_ticks_hz()) /
643 0 : SPDK_SEC_TO_USEC + spdk_get_ticks();
644 :
645 0 : return 0;
646 : }
647 :
648 : static int
649 0 : nvme_rdma_process_event_poll(struct nvme_rdma_qpair *rqpair)
650 : {
651 : struct nvme_rdma_ctrlr *rctrlr;
652 0 : int rc = 0, rc2;
653 :
654 0 : rctrlr = nvme_rdma_ctrlr(rqpair->qpair.ctrlr);
655 0 : assert(rctrlr != NULL);
656 :
657 0 : if (!rqpair->evt && spdk_get_ticks() < rqpair->evt_timeout_ticks) {
658 0 : rc = nvme_rdma_poll_events(rctrlr);
659 0 : if (rc == -EAGAIN || rc == -EWOULDBLOCK) {
660 0 : return rc;
661 : }
662 : }
663 :
664 0 : if (rqpair->evt == NULL) {
665 0 : rc = -EADDRNOTAVAIL;
666 0 : goto exit;
667 : }
668 :
669 0 : rc = nvme_rdma_validate_cm_event(rqpair->expected_evt_type, rqpair->evt);
670 :
671 0 : rc2 = nvme_rdma_qpair_process_cm_event(rqpair);
672 : /* bad message takes precedence over the other error codes from processing the event. */
673 0 : rc = rc == 0 ? rc2 : rc;
674 :
675 0 : exit:
676 0 : assert(rqpair->evt_cb != NULL);
677 0 : return rqpair->evt_cb(rqpair, rc);
678 : }
679 :
680 : static int
681 3 : nvme_rdma_resize_cq(struct nvme_rdma_qpair *rqpair, struct nvme_rdma_poller *poller)
682 : {
683 : int current_num_wc, required_num_wc;
684 : int max_cq_size;
685 :
686 3 : required_num_wc = poller->required_num_wc + WC_PER_QPAIR(rqpair->num_entries);
687 3 : current_num_wc = poller->current_num_wc;
688 3 : if (current_num_wc < required_num_wc) {
689 2 : current_num_wc = spdk_max(current_num_wc * 2, required_num_wc);
690 : }
691 :
692 3 : max_cq_size = g_spdk_nvme_transport_opts.rdma_max_cq_size;
693 3 : if (max_cq_size != 0 && current_num_wc > max_cq_size) {
694 0 : current_num_wc = max_cq_size;
695 : }
696 :
697 3 : if (poller->current_num_wc != current_num_wc) {
698 2 : SPDK_DEBUGLOG(nvme, "Resize RDMA CQ from %d to %d\n", poller->current_num_wc,
699 : current_num_wc);
700 2 : if (ibv_resize_cq(poller->cq, current_num_wc)) {
701 1 : SPDK_ERRLOG("RDMA CQ resize failed: errno %d: %s\n", errno, spdk_strerror(errno));
702 1 : return -1;
703 : }
704 :
705 1 : poller->current_num_wc = current_num_wc;
706 : }
707 :
708 2 : poller->required_num_wc = required_num_wc;
709 2 : return 0;
710 : }
711 :
712 : static int
713 5 : nvme_rdma_qpair_set_poller(struct spdk_nvme_qpair *qpair)
714 : {
715 5 : struct nvme_rdma_qpair *rqpair = nvme_rdma_qpair(qpair);
716 5 : struct nvme_rdma_poll_group *group = nvme_rdma_poll_group(qpair->poll_group);
717 : struct nvme_rdma_poller *poller;
718 :
719 5 : assert(rqpair->cq == NULL);
720 :
721 5 : poller = nvme_rdma_poll_group_get_poller(group, rqpair->cm_id->verbs);
722 5 : if (!poller) {
723 2 : SPDK_ERRLOG("Unable to find a cq for qpair %p on poll group %p\n", qpair, qpair->poll_group);
724 2 : return -EINVAL;
725 : }
726 :
727 3 : if (!poller->srq) {
728 3 : if (nvme_rdma_resize_cq(rqpair, poller)) {
729 1 : nvme_rdma_poll_group_put_poller(group, poller);
730 1 : return -EPROTO;
731 : }
732 : }
733 :
734 2 : rqpair->cq = poller->cq;
735 2 : rqpair->srq = poller->srq;
736 2 : if (rqpair->srq) {
737 0 : rqpair->rsps = poller->rsps;
738 : }
739 2 : rqpair->poller = poller;
740 2 : return 0;
741 : }
742 :
743 : static int
744 1 : nvme_rdma_qpair_init(struct nvme_rdma_qpair *rqpair)
745 : {
746 : int rc;
747 1 : struct spdk_rdma_qp_init_attr attr = {};
748 1 : struct ibv_device_attr dev_attr;
749 : struct nvme_rdma_ctrlr *rctrlr;
750 : uint32_t num_cqe, max_num_cqe;
751 :
752 1 : rc = ibv_query_device(rqpair->cm_id->verbs, &dev_attr);
753 1 : if (rc != 0) {
754 0 : SPDK_ERRLOG("Failed to query RDMA device attributes.\n");
755 0 : return -1;
756 : }
757 :
758 1 : if (rqpair->qpair.poll_group) {
759 0 : assert(!rqpair->cq);
760 0 : rc = nvme_rdma_qpair_set_poller(&rqpair->qpair);
761 0 : if (rc) {
762 0 : SPDK_ERRLOG("Unable to activate the rdmaqpair.\n");
763 0 : return -1;
764 : }
765 0 : assert(rqpair->cq);
766 : } else {
767 1 : num_cqe = rqpair->num_entries * 2;
768 1 : max_num_cqe = g_spdk_nvme_transport_opts.rdma_max_cq_size;
769 1 : if (max_num_cqe != 0 && num_cqe > max_num_cqe) {
770 0 : num_cqe = max_num_cqe;
771 : }
772 1 : rqpair->cq = ibv_create_cq(rqpair->cm_id->verbs, num_cqe, rqpair, NULL, 0);
773 1 : if (!rqpair->cq) {
774 0 : SPDK_ERRLOG("Unable to create completion queue: errno %d: %s\n", errno, spdk_strerror(errno));
775 0 : return -1;
776 : }
777 : }
778 :
779 1 : rctrlr = nvme_rdma_ctrlr(rqpair->qpair.ctrlr);
780 1 : if (g_nvme_hooks.get_ibv_pd) {
781 0 : attr.pd = g_nvme_hooks.get_ibv_pd(&rctrlr->ctrlr.trid, rqpair->cm_id->verbs);
782 : } else {
783 1 : attr.pd = spdk_rdma_get_pd(rqpair->cm_id->verbs);
784 : }
785 :
786 1 : attr.stats = rqpair->poller ? &rqpair->poller->stats.rdma_stats : NULL;
787 1 : attr.send_cq = rqpair->cq;
788 1 : attr.recv_cq = rqpair->cq;
789 1 : attr.cap.max_send_wr = rqpair->num_entries; /* SEND operations */
790 1 : if (rqpair->srq) {
791 0 : attr.srq = rqpair->srq->srq;
792 : } else {
793 1 : attr.cap.max_recv_wr = rqpair->num_entries; /* RECV operations */
794 : }
795 1 : attr.cap.max_send_sge = spdk_min(NVME_RDMA_DEFAULT_TX_SGE, dev_attr.max_sge);
796 1 : attr.cap.max_recv_sge = spdk_min(NVME_RDMA_DEFAULT_RX_SGE, dev_attr.max_sge);
797 :
798 1 : rqpair->rdma_qp = spdk_rdma_qp_create(rqpair->cm_id, &attr);
799 :
800 1 : if (!rqpair->rdma_qp) {
801 0 : return -1;
802 : }
803 :
804 1 : rqpair->memory_domain = nvme_rdma_get_memory_domain(rqpair->rdma_qp->qp->pd);
805 1 : if (!rqpair->memory_domain) {
806 0 : SPDK_ERRLOG("Failed to get memory domain\n");
807 0 : return -1;
808 : }
809 :
810 : /* ibv_create_qp will change the values in attr.cap. Make sure we store the proper value. */
811 1 : rqpair->max_send_sge = spdk_min(NVME_RDMA_DEFAULT_TX_SGE, attr.cap.max_send_sge);
812 1 : rqpair->max_recv_sge = spdk_min(NVME_RDMA_DEFAULT_RX_SGE, attr.cap.max_recv_sge);
813 1 : rqpair->current_num_sends = 0;
814 :
815 1 : rqpair->cm_id->context = rqpair;
816 :
817 1 : return 0;
818 : }
819 :
820 : static void
821 0 : nvme_rdma_reset_failed_sends(struct nvme_rdma_qpair *rqpair,
822 : struct ibv_send_wr *bad_send_wr, int rc)
823 : {
824 0 : SPDK_ERRLOG("Failed to post WRs on send queue, errno %d (%s), bad_wr %p\n",
825 : rc, spdk_strerror(rc), bad_send_wr);
826 0 : while (bad_send_wr != NULL) {
827 0 : assert(rqpair->current_num_sends > 0);
828 0 : rqpair->current_num_sends--;
829 0 : bad_send_wr = bad_send_wr->next;
830 : }
831 0 : }
832 :
833 : static void
834 0 : nvme_rdma_reset_failed_recvs(struct nvme_rdma_rsps *rsps,
835 : struct ibv_recv_wr *bad_recv_wr, int rc)
836 : {
837 0 : SPDK_ERRLOG("Failed to post WRs on receive queue, errno %d (%s), bad_wr %p\n",
838 : rc, spdk_strerror(rc), bad_recv_wr);
839 0 : while (bad_recv_wr != NULL) {
840 0 : assert(rsps->current_num_recvs > 0);
841 0 : rsps->current_num_recvs--;
842 0 : bad_recv_wr = bad_recv_wr->next;
843 : }
844 0 : }
845 :
846 : static inline int
847 1 : nvme_rdma_qpair_submit_sends(struct nvme_rdma_qpair *rqpair)
848 : {
849 1 : struct ibv_send_wr *bad_send_wr = NULL;
850 : int rc;
851 :
852 1 : rc = spdk_rdma_qp_flush_send_wrs(rqpair->rdma_qp, &bad_send_wr);
853 :
854 1 : if (spdk_unlikely(rc)) {
855 0 : nvme_rdma_reset_failed_sends(rqpair, bad_send_wr, rc);
856 : }
857 :
858 1 : return rc;
859 : }
860 :
861 : static inline int
862 0 : nvme_rdma_qpair_submit_recvs(struct nvme_rdma_qpair *rqpair)
863 : {
864 0 : struct ibv_recv_wr *bad_recv_wr;
865 0 : int rc = 0;
866 :
867 0 : rc = spdk_rdma_qp_flush_recv_wrs(rqpair->rdma_qp, &bad_recv_wr);
868 0 : if (spdk_unlikely(rc)) {
869 0 : nvme_rdma_reset_failed_recvs(rqpair->rsps, bad_recv_wr, rc);
870 : }
871 :
872 0 : return rc;
873 : }
874 :
875 : static inline int
876 0 : nvme_rdma_poller_submit_recvs(struct nvme_rdma_poller *poller)
877 : {
878 0 : struct ibv_recv_wr *bad_recv_wr;
879 : int rc;
880 :
881 0 : rc = spdk_rdma_srq_flush_recv_wrs(poller->srq, &bad_recv_wr);
882 0 : if (spdk_unlikely(rc)) {
883 0 : nvme_rdma_reset_failed_recvs(poller->rsps, bad_recv_wr, rc);
884 : }
885 :
886 0 : return rc;
887 : }
888 :
889 : #define nvme_rdma_trace_ibv_sge(sg_list) \
890 : if (sg_list) { \
891 : SPDK_DEBUGLOG(nvme, "local addr %p length 0x%x lkey 0x%x\n", \
892 : (void *)(sg_list)->addr, (sg_list)->length, (sg_list)->lkey); \
893 : }
894 :
895 : static void
896 3 : nvme_rdma_free_rsps(struct nvme_rdma_rsps *rsps)
897 : {
898 3 : if (!rsps) {
899 1 : return;
900 : }
901 :
902 2 : spdk_free(rsps->rsps);
903 2 : spdk_free(rsps->rsp_sgls);
904 2 : spdk_free(rsps->rsp_recv_wrs);
905 2 : spdk_free(rsps);
906 : }
907 :
908 : static struct nvme_rdma_rsps *
909 2 : nvme_rdma_create_rsps(struct nvme_rdma_rsp_opts *opts)
910 : {
911 : struct nvme_rdma_rsps *rsps;
912 2 : struct spdk_rdma_memory_translation translation;
913 : uint16_t i;
914 : int rc;
915 :
916 2 : rsps = spdk_zmalloc(sizeof(*rsps), 0, NULL, SPDK_ENV_SOCKET_ID_ANY, SPDK_MALLOC_DMA);
917 2 : if (!rsps) {
918 0 : SPDK_ERRLOG("Failed to allocate rsps object\n");
919 0 : return NULL;
920 : }
921 :
922 2 : rsps->rsp_sgls = spdk_zmalloc(opts->num_entries * sizeof(*rsps->rsp_sgls), 0, NULL,
923 : SPDK_ENV_SOCKET_ID_ANY, SPDK_MALLOC_DMA);
924 2 : if (!rsps->rsp_sgls) {
925 1 : SPDK_ERRLOG("Failed to allocate rsp_sgls\n");
926 1 : goto fail;
927 : }
928 :
929 1 : rsps->rsp_recv_wrs = spdk_zmalloc(opts->num_entries * sizeof(*rsps->rsp_recv_wrs), 0, NULL,
930 : SPDK_ENV_SOCKET_ID_ANY, SPDK_MALLOC_DMA);
931 1 : if (!rsps->rsp_recv_wrs) {
932 0 : SPDK_ERRLOG("Failed to allocate rsp_recv_wrs\n");
933 0 : goto fail;
934 : }
935 :
936 1 : rsps->rsps = spdk_zmalloc(opts->num_entries * sizeof(*rsps->rsps), 0, NULL,
937 : SPDK_ENV_SOCKET_ID_ANY, SPDK_MALLOC_DMA);
938 1 : if (!rsps->rsps) {
939 0 : SPDK_ERRLOG("can not allocate rdma rsps\n");
940 0 : goto fail;
941 : }
942 :
943 2 : for (i = 0; i < opts->num_entries; i++) {
944 1 : struct ibv_sge *rsp_sgl = &rsps->rsp_sgls[i];
945 1 : struct spdk_nvme_rdma_rsp *rsp = &rsps->rsps[i];
946 1 : struct ibv_recv_wr *recv_wr = &rsps->rsp_recv_wrs[i];
947 :
948 1 : rsp->rqpair = opts->rqpair;
949 1 : rsp->rdma_wr.type = RDMA_WR_TYPE_RECV;
950 1 : rsp->recv_wr = recv_wr;
951 1 : rsp_sgl->addr = (uint64_t)rsp;
952 1 : rsp_sgl->length = sizeof(struct spdk_nvme_cpl);
953 1 : rc = spdk_rdma_get_translation(opts->mr_map, rsp, sizeof(*rsp), &translation);
954 1 : if (rc) {
955 0 : goto fail;
956 : }
957 1 : rsp_sgl->lkey = spdk_rdma_memory_translation_get_lkey(&translation);
958 :
959 1 : recv_wr->wr_id = (uint64_t)&rsp->rdma_wr;
960 1 : recv_wr->next = NULL;
961 1 : recv_wr->sg_list = rsp_sgl;
962 1 : recv_wr->num_sge = 1;
963 :
964 1 : nvme_rdma_trace_ibv_sge(recv_wr->sg_list);
965 :
966 1 : if (opts->rqpair) {
967 1 : spdk_rdma_qp_queue_recv_wrs(opts->rqpair->rdma_qp, recv_wr);
968 : } else {
969 0 : spdk_rdma_srq_queue_recv_wrs(opts->srq, recv_wr);
970 : }
971 : }
972 :
973 1 : rsps->num_entries = opts->num_entries;
974 1 : rsps->current_num_recvs = opts->num_entries;
975 :
976 1 : return rsps;
977 1 : fail:
978 1 : nvme_rdma_free_rsps(rsps);
979 1 : return NULL;
980 : }
981 :
982 : static void
983 3 : nvme_rdma_free_reqs(struct nvme_rdma_qpair *rqpair)
984 : {
985 3 : if (!rqpair->rdma_reqs) {
986 2 : return;
987 : }
988 :
989 1 : spdk_free(rqpair->cmds);
990 1 : rqpair->cmds = NULL;
991 :
992 1 : spdk_free(rqpair->rdma_reqs);
993 1 : rqpair->rdma_reqs = NULL;
994 : }
995 :
996 : static int
997 4 : nvme_rdma_create_reqs(struct nvme_rdma_qpair *rqpair)
998 : {
999 4 : struct spdk_rdma_memory_translation translation;
1000 : uint16_t i;
1001 : int rc;
1002 :
1003 4 : assert(!rqpair->rdma_reqs);
1004 4 : rqpair->rdma_reqs = spdk_zmalloc(rqpair->num_entries * sizeof(struct spdk_nvme_rdma_req), 0, NULL,
1005 : SPDK_ENV_SOCKET_ID_ANY, SPDK_MALLOC_DMA);
1006 4 : if (rqpair->rdma_reqs == NULL) {
1007 1 : SPDK_ERRLOG("Failed to allocate rdma_reqs\n");
1008 1 : goto fail;
1009 : }
1010 :
1011 3 : assert(!rqpair->cmds);
1012 3 : rqpair->cmds = spdk_zmalloc(rqpair->num_entries * sizeof(*rqpair->cmds), 0, NULL,
1013 : SPDK_ENV_SOCKET_ID_ANY, SPDK_MALLOC_DMA);
1014 3 : if (!rqpair->cmds) {
1015 0 : SPDK_ERRLOG("Failed to allocate RDMA cmds\n");
1016 0 : goto fail;
1017 : }
1018 :
1019 3 : TAILQ_INIT(&rqpair->free_reqs);
1020 3 : TAILQ_INIT(&rqpair->outstanding_reqs);
1021 10 : for (i = 0; i < rqpair->num_entries; i++) {
1022 : struct spdk_nvme_rdma_req *rdma_req;
1023 : struct spdk_nvmf_cmd *cmd;
1024 :
1025 7 : rdma_req = &rqpair->rdma_reqs[i];
1026 7 : rdma_req->rdma_wr.type = RDMA_WR_TYPE_SEND;
1027 7 : cmd = &rqpair->cmds[i];
1028 :
1029 7 : rdma_req->id = i;
1030 :
1031 7 : rc = spdk_rdma_get_translation(rqpair->mr_map, cmd, sizeof(*cmd), &translation);
1032 7 : if (rc) {
1033 0 : goto fail;
1034 : }
1035 7 : rdma_req->send_sgl[0].lkey = spdk_rdma_memory_translation_get_lkey(&translation);
1036 :
1037 : /* The first RDMA sgl element will always point
1038 : * at this data structure. Depending on whether
1039 : * an NVMe-oF SGL is required, the length of
1040 : * this element may change. */
1041 7 : rdma_req->send_sgl[0].addr = (uint64_t)cmd;
1042 7 : rdma_req->send_wr.wr_id = (uint64_t)&rdma_req->rdma_wr;
1043 7 : rdma_req->send_wr.next = NULL;
1044 7 : rdma_req->send_wr.opcode = IBV_WR_SEND;
1045 7 : rdma_req->send_wr.send_flags = IBV_SEND_SIGNALED;
1046 7 : rdma_req->send_wr.sg_list = rdma_req->send_sgl;
1047 7 : rdma_req->send_wr.imm_data = 0;
1048 :
1049 7 : TAILQ_INSERT_TAIL(&rqpair->free_reqs, rdma_req, link);
1050 : }
1051 :
1052 3 : return 0;
1053 1 : fail:
1054 1 : nvme_rdma_free_reqs(rqpair);
1055 1 : return -ENOMEM;
1056 : }
1057 :
1058 : static int nvme_rdma_connect(struct nvme_rdma_qpair *rqpair);
1059 :
1060 : static int
1061 0 : nvme_rdma_route_resolved(struct nvme_rdma_qpair *rqpair, int ret)
1062 : {
1063 0 : if (ret) {
1064 0 : SPDK_ERRLOG("RDMA route resolution error\n");
1065 0 : return -1;
1066 : }
1067 :
1068 0 : ret = nvme_rdma_qpair_init(rqpair);
1069 0 : if (ret < 0) {
1070 0 : SPDK_ERRLOG("nvme_rdma_qpair_init() failed\n");
1071 0 : return -1;
1072 : }
1073 :
1074 0 : return nvme_rdma_connect(rqpair);
1075 : }
1076 :
1077 : static int
1078 0 : nvme_rdma_addr_resolved(struct nvme_rdma_qpair *rqpair, int ret)
1079 : {
1080 0 : if (ret) {
1081 0 : SPDK_ERRLOG("RDMA address resolution error\n");
1082 0 : return -1;
1083 : }
1084 :
1085 0 : if (rqpair->qpair.ctrlr->opts.transport_ack_timeout != SPDK_NVME_TRANSPORT_ACK_TIMEOUT_DISABLED) {
1086 : #ifdef SPDK_CONFIG_RDMA_SET_ACK_TIMEOUT
1087 0 : uint8_t timeout = rqpair->qpair.ctrlr->opts.transport_ack_timeout;
1088 0 : ret = rdma_set_option(rqpair->cm_id, RDMA_OPTION_ID,
1089 : RDMA_OPTION_ID_ACK_TIMEOUT,
1090 : &timeout, sizeof(timeout));
1091 0 : if (ret) {
1092 0 : SPDK_NOTICELOG("Can't apply RDMA_OPTION_ID_ACK_TIMEOUT %d, ret %d\n", timeout, ret);
1093 : }
1094 : #else
1095 : SPDK_DEBUGLOG(nvme, "transport_ack_timeout is not supported\n");
1096 : #endif
1097 : }
1098 :
1099 0 : if (rqpair->qpair.ctrlr->opts.transport_tos != SPDK_NVME_TRANSPORT_TOS_DISABLED) {
1100 : #ifdef SPDK_CONFIG_RDMA_SET_TOS
1101 0 : uint8_t tos = rqpair->qpair.ctrlr->opts.transport_tos;
1102 0 : ret = rdma_set_option(rqpair->cm_id, RDMA_OPTION_ID, RDMA_OPTION_ID_TOS, &tos, sizeof(tos));
1103 0 : if (ret) {
1104 0 : SPDK_NOTICELOG("Can't apply RDMA_OPTION_ID_TOS %u, ret %d\n", tos, ret);
1105 : }
1106 : #else
1107 : SPDK_DEBUGLOG(nvme, "transport_tos is not supported\n");
1108 : #endif
1109 : }
1110 :
1111 0 : ret = rdma_resolve_route(rqpair->cm_id, NVME_RDMA_TIME_OUT_IN_MS);
1112 0 : if (ret) {
1113 0 : SPDK_ERRLOG("rdma_resolve_route\n");
1114 0 : return ret;
1115 : }
1116 :
1117 0 : return nvme_rdma_process_event_start(rqpair, RDMA_CM_EVENT_ROUTE_RESOLVED,
1118 : nvme_rdma_route_resolved);
1119 : }
1120 :
1121 : static int
1122 0 : nvme_rdma_resolve_addr(struct nvme_rdma_qpair *rqpair,
1123 : struct sockaddr *src_addr,
1124 : struct sockaddr *dst_addr)
1125 : {
1126 : int ret;
1127 :
1128 0 : if (src_addr) {
1129 0 : int reuse = 1;
1130 :
1131 0 : ret = rdma_set_option(rqpair->cm_id, RDMA_OPTION_ID, RDMA_OPTION_ID_REUSEADDR,
1132 : &reuse, sizeof(reuse));
1133 0 : if (ret) {
1134 0 : SPDK_NOTICELOG("Can't apply RDMA_OPTION_ID_REUSEADDR %d, ret %d\n",
1135 : reuse, ret);
1136 : /* It is likely that rdma_resolve_addr() returns -EADDRINUSE, but
1137 : * we may missing something. We rely on rdma_resolve_addr().
1138 : */
1139 : }
1140 : }
1141 :
1142 0 : ret = rdma_resolve_addr(rqpair->cm_id, src_addr, dst_addr,
1143 : NVME_RDMA_TIME_OUT_IN_MS);
1144 0 : if (ret) {
1145 0 : SPDK_ERRLOG("rdma_resolve_addr, %d\n", errno);
1146 0 : return ret;
1147 : }
1148 :
1149 0 : return nvme_rdma_process_event_start(rqpair, RDMA_CM_EVENT_ADDR_RESOLVED,
1150 : nvme_rdma_addr_resolved);
1151 : }
1152 :
1153 : static int nvme_rdma_stale_conn_retry(struct nvme_rdma_qpair *rqpair);
1154 :
1155 : static int
1156 0 : nvme_rdma_connect_established(struct nvme_rdma_qpair *rqpair, int ret)
1157 : {
1158 0 : struct nvme_rdma_rsp_opts opts = {};
1159 :
1160 0 : if (ret == -ESTALE) {
1161 0 : return nvme_rdma_stale_conn_retry(rqpair);
1162 0 : } else if (ret) {
1163 0 : SPDK_ERRLOG("RDMA connect error %d\n", ret);
1164 0 : return ret;
1165 : }
1166 :
1167 0 : assert(!rqpair->mr_map);
1168 0 : rqpair->mr_map = spdk_rdma_create_mem_map(rqpair->rdma_qp->qp->pd, &g_nvme_hooks,
1169 : SPDK_RDMA_MEMORY_MAP_ROLE_INITIATOR);
1170 0 : if (!rqpair->mr_map) {
1171 0 : SPDK_ERRLOG("Unable to register RDMA memory translation map\n");
1172 0 : return -1;
1173 : }
1174 :
1175 0 : ret = nvme_rdma_create_reqs(rqpair);
1176 0 : SPDK_DEBUGLOG(nvme, "rc =%d\n", ret);
1177 0 : if (ret) {
1178 0 : SPDK_ERRLOG("Unable to create rqpair RDMA requests\n");
1179 0 : return -1;
1180 : }
1181 0 : SPDK_DEBUGLOG(nvme, "RDMA requests created\n");
1182 :
1183 0 : if (!rqpair->srq) {
1184 0 : opts.num_entries = rqpair->num_entries;
1185 0 : opts.rqpair = rqpair;
1186 0 : opts.srq = NULL;
1187 0 : opts.mr_map = rqpair->mr_map;
1188 :
1189 0 : assert(!rqpair->rsps);
1190 0 : rqpair->rsps = nvme_rdma_create_rsps(&opts);
1191 0 : if (!rqpair->rsps) {
1192 0 : SPDK_ERRLOG("Unable to create rqpair RDMA responses\n");
1193 0 : return -1;
1194 : }
1195 0 : SPDK_DEBUGLOG(nvme, "RDMA responses created\n");
1196 :
1197 0 : ret = nvme_rdma_qpair_submit_recvs(rqpair);
1198 0 : SPDK_DEBUGLOG(nvme, "rc =%d\n", ret);
1199 0 : if (ret) {
1200 0 : SPDK_ERRLOG("Unable to submit rqpair RDMA responses\n");
1201 0 : return -1;
1202 : }
1203 0 : SPDK_DEBUGLOG(nvme, "RDMA responses submitted\n");
1204 : }
1205 :
1206 0 : rqpair->state = NVME_RDMA_QPAIR_STATE_FABRIC_CONNECT_SEND;
1207 :
1208 0 : return 0;
1209 : }
1210 :
1211 : static int
1212 0 : nvme_rdma_connect(struct nvme_rdma_qpair *rqpair)
1213 : {
1214 0 : struct rdma_conn_param param = {};
1215 0 : struct spdk_nvmf_rdma_request_private_data request_data = {};
1216 0 : struct ibv_device_attr attr;
1217 : int ret;
1218 : struct spdk_nvme_ctrlr *ctrlr;
1219 :
1220 0 : ret = ibv_query_device(rqpair->cm_id->verbs, &attr);
1221 0 : if (ret != 0) {
1222 0 : SPDK_ERRLOG("Failed to query RDMA device attributes.\n");
1223 0 : return ret;
1224 : }
1225 :
1226 0 : param.responder_resources = attr.max_qp_rd_atom;
1227 :
1228 0 : ctrlr = rqpair->qpair.ctrlr;
1229 0 : if (!ctrlr) {
1230 0 : return -1;
1231 : }
1232 :
1233 0 : request_data.qid = rqpair->qpair.id;
1234 0 : request_data.hrqsize = rqpair->num_entries + 1;
1235 0 : request_data.hsqsize = rqpair->num_entries;
1236 0 : request_data.cntlid = ctrlr->cntlid;
1237 :
1238 0 : param.private_data = &request_data;
1239 0 : param.private_data_len = sizeof(request_data);
1240 0 : param.retry_count = ctrlr->opts.transport_retry_count;
1241 0 : param.rnr_retry_count = 7;
1242 :
1243 : /* Fields below are ignored by rdma cm if qpair has been
1244 : * created using rdma cm API. */
1245 0 : param.srq = 0;
1246 0 : param.qp_num = rqpair->rdma_qp->qp->qp_num;
1247 :
1248 0 : ret = rdma_connect(rqpair->cm_id, ¶m);
1249 0 : if (ret) {
1250 0 : SPDK_ERRLOG("nvme rdma connect error\n");
1251 0 : return ret;
1252 : }
1253 :
1254 0 : return nvme_rdma_process_event_start(rqpair, RDMA_CM_EVENT_ESTABLISHED,
1255 : nvme_rdma_connect_established);
1256 : }
1257 :
1258 : static int
1259 0 : nvme_rdma_ctrlr_connect_qpair(struct spdk_nvme_ctrlr *ctrlr, struct spdk_nvme_qpair *qpair)
1260 : {
1261 0 : struct sockaddr_storage dst_addr;
1262 0 : struct sockaddr_storage src_addr;
1263 : bool src_addr_specified;
1264 0 : long int port, src_port;
1265 : int rc;
1266 : struct nvme_rdma_ctrlr *rctrlr;
1267 : struct nvme_rdma_qpair *rqpair;
1268 : struct nvme_rdma_poll_group *group;
1269 : int family;
1270 :
1271 0 : rqpair = nvme_rdma_qpair(qpair);
1272 0 : rctrlr = nvme_rdma_ctrlr(ctrlr);
1273 0 : assert(rctrlr != NULL);
1274 :
1275 0 : switch (ctrlr->trid.adrfam) {
1276 0 : case SPDK_NVMF_ADRFAM_IPV4:
1277 0 : family = AF_INET;
1278 0 : break;
1279 0 : case SPDK_NVMF_ADRFAM_IPV6:
1280 0 : family = AF_INET6;
1281 0 : break;
1282 0 : default:
1283 0 : SPDK_ERRLOG("Unhandled ADRFAM %d\n", ctrlr->trid.adrfam);
1284 0 : return -1;
1285 : }
1286 :
1287 0 : SPDK_DEBUGLOG(nvme, "adrfam %d ai_family %d\n", ctrlr->trid.adrfam, family);
1288 :
1289 0 : memset(&dst_addr, 0, sizeof(dst_addr));
1290 :
1291 0 : SPDK_DEBUGLOG(nvme, "trsvcid is %s\n", ctrlr->trid.trsvcid);
1292 0 : rc = nvme_parse_addr(&dst_addr, family, ctrlr->trid.traddr, ctrlr->trid.trsvcid, &port);
1293 0 : if (rc != 0) {
1294 0 : SPDK_ERRLOG("dst_addr nvme_parse_addr() failed\n");
1295 0 : return -1;
1296 : }
1297 :
1298 0 : if (ctrlr->opts.src_addr[0] || ctrlr->opts.src_svcid[0]) {
1299 0 : memset(&src_addr, 0, sizeof(src_addr));
1300 0 : rc = nvme_parse_addr(&src_addr, family, ctrlr->opts.src_addr, ctrlr->opts.src_svcid, &src_port);
1301 0 : if (rc != 0) {
1302 0 : SPDK_ERRLOG("src_addr nvme_parse_addr() failed\n");
1303 0 : return -1;
1304 : }
1305 0 : src_addr_specified = true;
1306 : } else {
1307 0 : src_addr_specified = false;
1308 : }
1309 :
1310 0 : rc = rdma_create_id(rctrlr->cm_channel, &rqpair->cm_id, rqpair, RDMA_PS_TCP);
1311 0 : if (rc < 0) {
1312 0 : SPDK_ERRLOG("rdma_create_id() failed\n");
1313 0 : return -1;
1314 : }
1315 :
1316 0 : rc = nvme_rdma_resolve_addr(rqpair,
1317 : src_addr_specified ? (struct sockaddr *)&src_addr : NULL,
1318 : (struct sockaddr *)&dst_addr);
1319 0 : if (rc < 0) {
1320 0 : SPDK_ERRLOG("nvme_rdma_resolve_addr() failed\n");
1321 0 : return -1;
1322 : }
1323 :
1324 0 : rqpair->state = NVME_RDMA_QPAIR_STATE_INITIALIZING;
1325 :
1326 0 : if (qpair->poll_group != NULL) {
1327 0 : group = nvme_rdma_poll_group(qpair->poll_group);
1328 0 : TAILQ_INSERT_TAIL(&group->connecting_qpairs, rqpair, link_connecting);
1329 : }
1330 :
1331 0 : return 0;
1332 : }
1333 :
1334 : static int
1335 0 : nvme_rdma_stale_conn_reconnect(struct nvme_rdma_qpair *rqpair)
1336 : {
1337 0 : struct spdk_nvme_qpair *qpair = &rqpair->qpair;
1338 :
1339 0 : if (spdk_get_ticks() < rqpair->evt_timeout_ticks) {
1340 0 : return -EAGAIN;
1341 : }
1342 :
1343 0 : return nvme_rdma_ctrlr_connect_qpair(qpair->ctrlr, qpair);
1344 : }
1345 :
1346 : static int
1347 0 : nvme_rdma_ctrlr_connect_qpair_poll(struct spdk_nvme_ctrlr *ctrlr,
1348 : struct spdk_nvme_qpair *qpair)
1349 : {
1350 0 : struct nvme_rdma_qpair *rqpair = nvme_rdma_qpair(qpair);
1351 : int rc;
1352 :
1353 0 : if (rqpair->in_connect_poll) {
1354 0 : return -EAGAIN;
1355 : }
1356 :
1357 0 : rqpair->in_connect_poll = true;
1358 :
1359 0 : switch (rqpair->state) {
1360 0 : case NVME_RDMA_QPAIR_STATE_INVALID:
1361 0 : rc = -EAGAIN;
1362 0 : break;
1363 :
1364 0 : case NVME_RDMA_QPAIR_STATE_INITIALIZING:
1365 : case NVME_RDMA_QPAIR_STATE_EXITING:
1366 0 : if (!nvme_qpair_is_admin_queue(qpair)) {
1367 0 : nvme_ctrlr_lock(ctrlr);
1368 : }
1369 :
1370 0 : rc = nvme_rdma_process_event_poll(rqpair);
1371 :
1372 0 : if (!nvme_qpair_is_admin_queue(qpair)) {
1373 0 : nvme_ctrlr_unlock(ctrlr);
1374 : }
1375 :
1376 0 : if (rc == 0) {
1377 0 : rc = -EAGAIN;
1378 : }
1379 0 : rqpair->in_connect_poll = false;
1380 :
1381 0 : return rc;
1382 :
1383 0 : case NVME_RDMA_QPAIR_STATE_STALE_CONN:
1384 0 : rc = nvme_rdma_stale_conn_reconnect(rqpair);
1385 0 : if (rc == 0) {
1386 0 : rc = -EAGAIN;
1387 : }
1388 0 : break;
1389 0 : case NVME_RDMA_QPAIR_STATE_FABRIC_CONNECT_SEND:
1390 0 : rc = nvme_fabric_qpair_connect_async(qpair, rqpair->num_entries + 1);
1391 0 : if (rc == 0) {
1392 0 : rqpair->state = NVME_RDMA_QPAIR_STATE_FABRIC_CONNECT_POLL;
1393 0 : rc = -EAGAIN;
1394 : } else {
1395 0 : SPDK_ERRLOG("Failed to send an NVMe-oF Fabric CONNECT command\n");
1396 : }
1397 0 : break;
1398 0 : case NVME_RDMA_QPAIR_STATE_FABRIC_CONNECT_POLL:
1399 0 : rc = nvme_fabric_qpair_connect_poll(qpair);
1400 0 : if (rc == 0) {
1401 0 : rqpair->state = NVME_RDMA_QPAIR_STATE_RUNNING;
1402 0 : nvme_qpair_set_state(qpair, NVME_QPAIR_CONNECTED);
1403 0 : } else if (rc != -EAGAIN) {
1404 0 : SPDK_ERRLOG("Failed to poll NVMe-oF Fabric CONNECT command\n");
1405 : }
1406 0 : break;
1407 0 : case NVME_RDMA_QPAIR_STATE_RUNNING:
1408 0 : rc = 0;
1409 0 : break;
1410 0 : default:
1411 0 : assert(false);
1412 : rc = -EINVAL;
1413 : break;
1414 : }
1415 :
1416 0 : rqpair->in_connect_poll = false;
1417 :
1418 0 : return rc;
1419 : }
1420 :
1421 : static inline int
1422 28 : nvme_rdma_get_memory_translation(struct nvme_request *req, struct nvme_rdma_qpair *rqpair,
1423 : struct nvme_rdma_memory_translation_ctx *_ctx)
1424 : {
1425 28 : struct spdk_memory_domain_translation_ctx ctx;
1426 28 : struct spdk_memory_domain_translation_result dma_translation = {.iov_count = 0};
1427 28 : struct spdk_rdma_memory_translation rdma_translation;
1428 : int rc;
1429 :
1430 28 : assert(req);
1431 28 : assert(rqpair);
1432 28 : assert(_ctx);
1433 :
1434 28 : if (req->payload.opts && req->payload.opts->memory_domain) {
1435 2 : ctx.size = sizeof(struct spdk_memory_domain_translation_ctx);
1436 2 : ctx.rdma.ibv_qp = rqpair->rdma_qp->qp;
1437 2 : dma_translation.size = sizeof(struct spdk_memory_domain_translation_result);
1438 :
1439 2 : rc = spdk_memory_domain_translate_data(req->payload.opts->memory_domain,
1440 2 : req->payload.opts->memory_domain_ctx,
1441 2 : rqpair->memory_domain->domain, &ctx, _ctx->addr,
1442 : _ctx->length, &dma_translation);
1443 2 : if (spdk_unlikely(rc) || dma_translation.iov_count != 1) {
1444 1 : SPDK_ERRLOG("DMA memory translation failed, rc %d, iov count %u\n", rc, dma_translation.iov_count);
1445 1 : return rc;
1446 : }
1447 :
1448 1 : _ctx->lkey = dma_translation.rdma.lkey;
1449 1 : _ctx->rkey = dma_translation.rdma.rkey;
1450 1 : _ctx->addr = dma_translation.iov.iov_base;
1451 1 : _ctx->length = dma_translation.iov.iov_len;
1452 : } else {
1453 26 : rc = spdk_rdma_get_translation(rqpair->mr_map, _ctx->addr, _ctx->length, &rdma_translation);
1454 26 : if (spdk_unlikely(rc)) {
1455 2 : SPDK_ERRLOG("RDMA memory translation failed, rc %d\n", rc);
1456 2 : return rc;
1457 : }
1458 24 : if (rdma_translation.translation_type == SPDK_RDMA_TRANSLATION_MR) {
1459 24 : _ctx->lkey = rdma_translation.mr_or_key.mr->lkey;
1460 24 : _ctx->rkey = rdma_translation.mr_or_key.mr->rkey;
1461 : } else {
1462 0 : _ctx->lkey = _ctx->rkey = (uint32_t)rdma_translation.mr_or_key.key;
1463 : }
1464 : }
1465 :
1466 25 : return 0;
1467 : }
1468 :
1469 :
1470 : /*
1471 : * Build SGL describing empty payload.
1472 : */
1473 : static int
1474 2 : nvme_rdma_build_null_request(struct spdk_nvme_rdma_req *rdma_req)
1475 : {
1476 2 : struct nvme_request *req = rdma_req->req;
1477 :
1478 2 : req->cmd.psdt = SPDK_NVME_PSDT_SGL_MPTR_CONTIG;
1479 :
1480 : /* The first element of this SGL is pointing at an
1481 : * spdk_nvmf_cmd object. For this particular command,
1482 : * we only need the first 64 bytes corresponding to
1483 : * the NVMe command. */
1484 2 : rdma_req->send_sgl[0].length = sizeof(struct spdk_nvme_cmd);
1485 :
1486 : /* The RDMA SGL needs one element describing the NVMe command. */
1487 2 : rdma_req->send_wr.num_sge = 1;
1488 :
1489 2 : req->cmd.dptr.sgl1.keyed.type = SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK;
1490 2 : req->cmd.dptr.sgl1.keyed.subtype = SPDK_NVME_SGL_SUBTYPE_ADDRESS;
1491 2 : req->cmd.dptr.sgl1.keyed.length = 0;
1492 2 : req->cmd.dptr.sgl1.keyed.key = 0;
1493 2 : req->cmd.dptr.sgl1.address = 0;
1494 :
1495 2 : return 0;
1496 : }
1497 :
1498 : /*
1499 : * Build inline SGL describing contiguous payload buffer.
1500 : */
1501 : static int
1502 3 : nvme_rdma_build_contig_inline_request(struct nvme_rdma_qpair *rqpair,
1503 : struct spdk_nvme_rdma_req *rdma_req)
1504 : {
1505 3 : struct nvme_request *req = rdma_req->req;
1506 3 : struct nvme_rdma_memory_translation_ctx ctx = {
1507 3 : .addr = (uint8_t *)req->payload.contig_or_cb_arg + req->payload_offset,
1508 3 : .length = req->payload_size
1509 : };
1510 : int rc;
1511 :
1512 3 : assert(ctx.length != 0);
1513 3 : assert(nvme_payload_type(&req->payload) == NVME_PAYLOAD_TYPE_CONTIG);
1514 :
1515 3 : rc = nvme_rdma_get_memory_translation(req, rqpair, &ctx);
1516 3 : if (spdk_unlikely(rc)) {
1517 0 : return -1;
1518 : }
1519 :
1520 3 : rdma_req->send_sgl[1].lkey = ctx.lkey;
1521 :
1522 : /* The first element of this SGL is pointing at an
1523 : * spdk_nvmf_cmd object. For this particular command,
1524 : * we only need the first 64 bytes corresponding to
1525 : * the NVMe command. */
1526 3 : rdma_req->send_sgl[0].length = sizeof(struct spdk_nvme_cmd);
1527 :
1528 3 : rdma_req->send_sgl[1].addr = (uint64_t)ctx.addr;
1529 3 : rdma_req->send_sgl[1].length = (uint32_t)ctx.length;
1530 :
1531 : /* The RDMA SGL contains two elements. The first describes
1532 : * the NVMe command and the second describes the data
1533 : * payload. */
1534 3 : rdma_req->send_wr.num_sge = 2;
1535 :
1536 3 : req->cmd.psdt = SPDK_NVME_PSDT_SGL_MPTR_CONTIG;
1537 3 : req->cmd.dptr.sgl1.unkeyed.type = SPDK_NVME_SGL_TYPE_DATA_BLOCK;
1538 3 : req->cmd.dptr.sgl1.unkeyed.subtype = SPDK_NVME_SGL_SUBTYPE_OFFSET;
1539 3 : req->cmd.dptr.sgl1.unkeyed.length = (uint32_t)ctx.length;
1540 : /* Inline only supported for icdoff == 0 currently. This function will
1541 : * not get called for controllers with other values. */
1542 3 : req->cmd.dptr.sgl1.address = (uint64_t)0;
1543 :
1544 3 : return 0;
1545 : }
1546 :
1547 : /*
1548 : * Build SGL describing contiguous payload buffer.
1549 : */
1550 : static int
1551 3 : nvme_rdma_build_contig_request(struct nvme_rdma_qpair *rqpair,
1552 : struct spdk_nvme_rdma_req *rdma_req)
1553 : {
1554 3 : struct nvme_request *req = rdma_req->req;
1555 3 : struct nvme_rdma_memory_translation_ctx ctx = {
1556 3 : .addr = (uint8_t *)req->payload.contig_or_cb_arg + req->payload_offset,
1557 3 : .length = req->payload_size
1558 : };
1559 : int rc;
1560 :
1561 3 : assert(req->payload_size != 0);
1562 3 : assert(nvme_payload_type(&req->payload) == NVME_PAYLOAD_TYPE_CONTIG);
1563 :
1564 3 : if (spdk_unlikely(req->payload_size > NVME_RDMA_MAX_KEYED_SGL_LENGTH)) {
1565 1 : SPDK_ERRLOG("SGL length %u exceeds max keyed SGL block size %u\n",
1566 : req->payload_size, NVME_RDMA_MAX_KEYED_SGL_LENGTH);
1567 1 : return -1;
1568 : }
1569 :
1570 2 : rc = nvme_rdma_get_memory_translation(req, rqpair, &ctx);
1571 2 : if (spdk_unlikely(rc)) {
1572 0 : return -1;
1573 : }
1574 :
1575 2 : req->cmd.dptr.sgl1.keyed.key = ctx.rkey;
1576 :
1577 : /* The first element of this SGL is pointing at an
1578 : * spdk_nvmf_cmd object. For this particular command,
1579 : * we only need the first 64 bytes corresponding to
1580 : * the NVMe command. */
1581 2 : rdma_req->send_sgl[0].length = sizeof(struct spdk_nvme_cmd);
1582 :
1583 : /* The RDMA SGL needs one element describing the NVMe command. */
1584 2 : rdma_req->send_wr.num_sge = 1;
1585 :
1586 2 : req->cmd.psdt = SPDK_NVME_PSDT_SGL_MPTR_CONTIG;
1587 2 : req->cmd.dptr.sgl1.keyed.type = SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK;
1588 2 : req->cmd.dptr.sgl1.keyed.subtype = SPDK_NVME_SGL_SUBTYPE_ADDRESS;
1589 2 : req->cmd.dptr.sgl1.keyed.length = (uint32_t)ctx.length;
1590 2 : req->cmd.dptr.sgl1.address = (uint64_t)ctx.addr;
1591 :
1592 2 : return 0;
1593 : }
1594 :
1595 : /*
1596 : * Build SGL describing scattered payload buffer.
1597 : */
1598 : static int
1599 7 : nvme_rdma_build_sgl_request(struct nvme_rdma_qpair *rqpair,
1600 : struct spdk_nvme_rdma_req *rdma_req)
1601 : {
1602 7 : struct nvme_request *req = rdma_req->req;
1603 7 : struct spdk_nvmf_cmd *cmd = &rqpair->cmds[rdma_req->id];
1604 7 : struct nvme_rdma_memory_translation_ctx ctx;
1605 : uint32_t remaining_size;
1606 7 : uint32_t sge_length;
1607 : int rc, max_num_sgl, num_sgl_desc;
1608 :
1609 7 : assert(req->payload_size != 0);
1610 7 : assert(nvme_payload_type(&req->payload) == NVME_PAYLOAD_TYPE_SGL);
1611 7 : assert(req->payload.reset_sgl_fn != NULL);
1612 7 : assert(req->payload.next_sge_fn != NULL);
1613 7 : req->payload.reset_sgl_fn(req->payload.contig_or_cb_arg, req->payload_offset);
1614 :
1615 7 : max_num_sgl = req->qpair->ctrlr->max_sges;
1616 :
1617 7 : remaining_size = req->payload_size;
1618 7 : num_sgl_desc = 0;
1619 : do {
1620 18 : rc = req->payload.next_sge_fn(req->payload.contig_or_cb_arg, &ctx.addr, &sge_length);
1621 18 : if (rc) {
1622 1 : return -1;
1623 : }
1624 :
1625 17 : sge_length = spdk_min(remaining_size, sge_length);
1626 :
1627 17 : if (spdk_unlikely(sge_length > NVME_RDMA_MAX_KEYED_SGL_LENGTH)) {
1628 1 : SPDK_ERRLOG("SGL length %u exceeds max keyed SGL block size %u\n",
1629 : sge_length, NVME_RDMA_MAX_KEYED_SGL_LENGTH);
1630 1 : return -1;
1631 : }
1632 16 : ctx.length = sge_length;
1633 16 : rc = nvme_rdma_get_memory_translation(req, rqpair, &ctx);
1634 16 : if (spdk_unlikely(rc)) {
1635 1 : return -1;
1636 : }
1637 :
1638 15 : cmd->sgl[num_sgl_desc].keyed.key = ctx.rkey;
1639 15 : cmd->sgl[num_sgl_desc].keyed.type = SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK;
1640 15 : cmd->sgl[num_sgl_desc].keyed.subtype = SPDK_NVME_SGL_SUBTYPE_ADDRESS;
1641 15 : cmd->sgl[num_sgl_desc].keyed.length = (uint32_t)ctx.length;
1642 15 : cmd->sgl[num_sgl_desc].address = (uint64_t)ctx.addr;
1643 :
1644 15 : remaining_size -= ctx.length;
1645 15 : num_sgl_desc++;
1646 15 : } while (remaining_size > 0 && num_sgl_desc < max_num_sgl);
1647 :
1648 :
1649 : /* Should be impossible if we did our sgl checks properly up the stack, but do a sanity check here. */
1650 4 : if (remaining_size > 0) {
1651 0 : return -1;
1652 : }
1653 :
1654 4 : req->cmd.psdt = SPDK_NVME_PSDT_SGL_MPTR_CONTIG;
1655 :
1656 : /* The RDMA SGL needs one element describing some portion
1657 : * of the spdk_nvmf_cmd structure. */
1658 4 : rdma_req->send_wr.num_sge = 1;
1659 :
1660 : /*
1661 : * If only one SGL descriptor is required, it can be embedded directly in the command
1662 : * as a data block descriptor.
1663 : */
1664 4 : if (num_sgl_desc == 1) {
1665 : /* The first element of this SGL is pointing at an
1666 : * spdk_nvmf_cmd object. For this particular command,
1667 : * we only need the first 64 bytes corresponding to
1668 : * the NVMe command. */
1669 2 : rdma_req->send_sgl[0].length = sizeof(struct spdk_nvme_cmd);
1670 :
1671 2 : req->cmd.dptr.sgl1.keyed.type = cmd->sgl[0].keyed.type;
1672 2 : req->cmd.dptr.sgl1.keyed.subtype = cmd->sgl[0].keyed.subtype;
1673 2 : req->cmd.dptr.sgl1.keyed.length = cmd->sgl[0].keyed.length;
1674 2 : req->cmd.dptr.sgl1.keyed.key = cmd->sgl[0].keyed.key;
1675 2 : req->cmd.dptr.sgl1.address = cmd->sgl[0].address;
1676 : } else {
1677 : /*
1678 : * Otherwise, The SGL descriptor embedded in the command must point to the list of
1679 : * SGL descriptors used to describe the operation. In that case it is a last segment descriptor.
1680 : */
1681 2 : uint32_t descriptors_size = sizeof(struct spdk_nvme_sgl_descriptor) * num_sgl_desc;
1682 :
1683 2 : if (spdk_unlikely(descriptors_size > rqpair->qpair.ctrlr->ioccsz_bytes)) {
1684 1 : SPDK_ERRLOG("Size of SGL descriptors (%u) exceeds ICD (%u)\n",
1685 : descriptors_size, rqpair->qpair.ctrlr->ioccsz_bytes);
1686 1 : return -1;
1687 : }
1688 1 : rdma_req->send_sgl[0].length = sizeof(struct spdk_nvme_cmd) + descriptors_size;
1689 :
1690 1 : req->cmd.dptr.sgl1.unkeyed.type = SPDK_NVME_SGL_TYPE_LAST_SEGMENT;
1691 1 : req->cmd.dptr.sgl1.unkeyed.subtype = SPDK_NVME_SGL_SUBTYPE_OFFSET;
1692 1 : req->cmd.dptr.sgl1.unkeyed.length = descriptors_size;
1693 1 : req->cmd.dptr.sgl1.address = (uint64_t)0;
1694 : }
1695 :
1696 3 : return 0;
1697 : }
1698 :
1699 : /*
1700 : * Build inline SGL describing sgl payload buffer.
1701 : */
1702 : static int
1703 3 : nvme_rdma_build_sgl_inline_request(struct nvme_rdma_qpair *rqpair,
1704 : struct spdk_nvme_rdma_req *rdma_req)
1705 : {
1706 3 : struct nvme_request *req = rdma_req->req;
1707 3 : struct nvme_rdma_memory_translation_ctx ctx;
1708 3 : uint32_t length;
1709 : int rc;
1710 :
1711 3 : assert(req->payload_size != 0);
1712 3 : assert(nvme_payload_type(&req->payload) == NVME_PAYLOAD_TYPE_SGL);
1713 3 : assert(req->payload.reset_sgl_fn != NULL);
1714 3 : assert(req->payload.next_sge_fn != NULL);
1715 3 : req->payload.reset_sgl_fn(req->payload.contig_or_cb_arg, req->payload_offset);
1716 :
1717 3 : rc = req->payload.next_sge_fn(req->payload.contig_or_cb_arg, &ctx.addr, &length);
1718 3 : if (rc) {
1719 0 : return -1;
1720 : }
1721 :
1722 3 : if (length < req->payload_size) {
1723 0 : SPDK_DEBUGLOG(nvme, "Inline SGL request split so sending separately.\n");
1724 0 : return nvme_rdma_build_sgl_request(rqpair, rdma_req);
1725 : }
1726 :
1727 3 : if (length > req->payload_size) {
1728 0 : length = req->payload_size;
1729 : }
1730 :
1731 3 : ctx.length = length;
1732 3 : rc = nvme_rdma_get_memory_translation(req, rqpair, &ctx);
1733 3 : if (spdk_unlikely(rc)) {
1734 0 : return -1;
1735 : }
1736 :
1737 3 : rdma_req->send_sgl[1].addr = (uint64_t)ctx.addr;
1738 3 : rdma_req->send_sgl[1].length = (uint32_t)ctx.length;
1739 3 : rdma_req->send_sgl[1].lkey = ctx.lkey;
1740 :
1741 3 : rdma_req->send_wr.num_sge = 2;
1742 :
1743 : /* The first element of this SGL is pointing at an
1744 : * spdk_nvmf_cmd object. For this particular command,
1745 : * we only need the first 64 bytes corresponding to
1746 : * the NVMe command. */
1747 3 : rdma_req->send_sgl[0].length = sizeof(struct spdk_nvme_cmd);
1748 :
1749 3 : req->cmd.psdt = SPDK_NVME_PSDT_SGL_MPTR_CONTIG;
1750 3 : req->cmd.dptr.sgl1.unkeyed.type = SPDK_NVME_SGL_TYPE_DATA_BLOCK;
1751 3 : req->cmd.dptr.sgl1.unkeyed.subtype = SPDK_NVME_SGL_SUBTYPE_OFFSET;
1752 3 : req->cmd.dptr.sgl1.unkeyed.length = (uint32_t)ctx.length;
1753 : /* Inline only supported for icdoff == 0 currently. This function will
1754 : * not get called for controllers with other values. */
1755 3 : req->cmd.dptr.sgl1.address = (uint64_t)0;
1756 :
1757 3 : return 0;
1758 : }
1759 :
1760 : static int
1761 6 : nvme_rdma_req_init(struct nvme_rdma_qpair *rqpair, struct nvme_request *req,
1762 : struct spdk_nvme_rdma_req *rdma_req)
1763 : {
1764 6 : struct spdk_nvme_ctrlr *ctrlr = rqpair->qpair.ctrlr;
1765 : enum nvme_payload_type payload_type;
1766 : bool icd_supported;
1767 : int rc;
1768 :
1769 6 : assert(rdma_req->req == NULL);
1770 6 : rdma_req->req = req;
1771 6 : req->cmd.cid = rdma_req->id;
1772 6 : payload_type = nvme_payload_type(&req->payload);
1773 : /*
1774 : * Check if icdoff is non zero, to avoid interop conflicts with
1775 : * targets with non-zero icdoff. Both SPDK and the Linux kernel
1776 : * targets use icdoff = 0. For targets with non-zero icdoff, we
1777 : * will currently just not use inline data for now.
1778 : */
1779 6 : icd_supported = spdk_nvme_opc_get_data_transfer(req->cmd.opc) == SPDK_NVME_DATA_HOST_TO_CONTROLLER
1780 6 : && req->payload_size <= ctrlr->ioccsz_bytes && ctrlr->icdoff == 0;
1781 :
1782 6 : if (req->payload_size == 0) {
1783 2 : rc = nvme_rdma_build_null_request(rdma_req);
1784 4 : } else if (payload_type == NVME_PAYLOAD_TYPE_CONTIG) {
1785 2 : if (icd_supported) {
1786 1 : rc = nvme_rdma_build_contig_inline_request(rqpair, rdma_req);
1787 : } else {
1788 1 : rc = nvme_rdma_build_contig_request(rqpair, rdma_req);
1789 : }
1790 2 : } else if (payload_type == NVME_PAYLOAD_TYPE_SGL) {
1791 2 : if (icd_supported) {
1792 1 : rc = nvme_rdma_build_sgl_inline_request(rqpair, rdma_req);
1793 : } else {
1794 1 : rc = nvme_rdma_build_sgl_request(rqpair, rdma_req);
1795 : }
1796 : } else {
1797 0 : rc = -1;
1798 : }
1799 :
1800 6 : if (rc) {
1801 0 : rdma_req->req = NULL;
1802 0 : return rc;
1803 : }
1804 :
1805 6 : memcpy(&rqpair->cmds[rdma_req->id], &req->cmd, sizeof(req->cmd));
1806 6 : return 0;
1807 : }
1808 :
1809 : static struct spdk_nvme_qpair *
1810 5 : nvme_rdma_ctrlr_create_qpair(struct spdk_nvme_ctrlr *ctrlr,
1811 : uint16_t qid, uint32_t qsize,
1812 : enum spdk_nvme_qprio qprio,
1813 : uint32_t num_requests,
1814 : bool delay_cmd_submit,
1815 : bool async)
1816 : {
1817 : struct nvme_rdma_qpair *rqpair;
1818 : struct spdk_nvme_qpair *qpair;
1819 : int rc;
1820 :
1821 5 : if (qsize < SPDK_NVME_QUEUE_MIN_ENTRIES) {
1822 2 : SPDK_ERRLOG("Failed to create qpair with size %u. Minimum queue size is %d.\n",
1823 : qsize, SPDK_NVME_QUEUE_MIN_ENTRIES);
1824 2 : return NULL;
1825 : }
1826 :
1827 3 : rqpair = spdk_zmalloc(sizeof(struct nvme_rdma_qpair), 0, NULL, SPDK_ENV_SOCKET_ID_ANY,
1828 : SPDK_MALLOC_DMA);
1829 3 : if (!rqpair) {
1830 0 : SPDK_ERRLOG("failed to get create rqpair\n");
1831 0 : return NULL;
1832 : }
1833 :
1834 : /* Set num_entries one less than queue size. According to NVMe
1835 : * and NVMe-oF specs we can not submit queue size requests,
1836 : * one slot shall always remain empty.
1837 : */
1838 3 : rqpair->num_entries = qsize - 1;
1839 3 : rqpair->delay_cmd_submit = delay_cmd_submit;
1840 3 : rqpair->state = NVME_RDMA_QPAIR_STATE_INVALID;
1841 3 : qpair = &rqpair->qpair;
1842 3 : rc = nvme_qpair_init(qpair, qid, ctrlr, qprio, num_requests, async);
1843 3 : if (rc != 0) {
1844 0 : spdk_free(rqpair);
1845 0 : return NULL;
1846 : }
1847 :
1848 3 : return qpair;
1849 : }
1850 :
1851 : static void
1852 1 : nvme_rdma_qpair_destroy(struct nvme_rdma_qpair *rqpair)
1853 : {
1854 1 : struct spdk_nvme_qpair *qpair = &rqpair->qpair;
1855 : struct nvme_rdma_ctrlr *rctrlr;
1856 : struct nvme_rdma_cm_event_entry *entry, *tmp;
1857 :
1858 1 : spdk_rdma_free_mem_map(&rqpair->mr_map);
1859 :
1860 1 : if (rqpair->evt) {
1861 0 : rdma_ack_cm_event(rqpair->evt);
1862 0 : rqpair->evt = NULL;
1863 : }
1864 :
1865 : /*
1866 : * This works because we have the controller lock both in
1867 : * this function and in the function where we add new events.
1868 : */
1869 1 : if (qpair->ctrlr != NULL) {
1870 1 : rctrlr = nvme_rdma_ctrlr(qpair->ctrlr);
1871 1 : STAILQ_FOREACH_SAFE(entry, &rctrlr->pending_cm_events, link, tmp) {
1872 0 : if (entry->evt->id->context == rqpair) {
1873 0 : STAILQ_REMOVE(&rctrlr->pending_cm_events, entry, nvme_rdma_cm_event_entry, link);
1874 0 : rdma_ack_cm_event(entry->evt);
1875 0 : STAILQ_INSERT_HEAD(&rctrlr->free_cm_events, entry, link);
1876 : }
1877 : }
1878 : }
1879 :
1880 1 : if (rqpair->cm_id) {
1881 0 : if (rqpair->rdma_qp) {
1882 0 : spdk_rdma_put_pd(rqpair->rdma_qp->qp->pd);
1883 0 : spdk_rdma_qp_destroy(rqpair->rdma_qp);
1884 0 : rqpair->rdma_qp = NULL;
1885 : }
1886 : }
1887 :
1888 1 : if (rqpair->poller) {
1889 : struct nvme_rdma_poll_group *group;
1890 :
1891 0 : assert(qpair->poll_group);
1892 0 : group = nvme_rdma_poll_group(qpair->poll_group);
1893 :
1894 0 : nvme_rdma_poll_group_put_poller(group, rqpair->poller);
1895 :
1896 0 : rqpair->poller = NULL;
1897 0 : rqpair->cq = NULL;
1898 0 : if (rqpair->srq) {
1899 0 : rqpair->srq = NULL;
1900 0 : rqpair->rsps = NULL;
1901 : }
1902 1 : } else if (rqpair->cq) {
1903 0 : ibv_destroy_cq(rqpair->cq);
1904 0 : rqpair->cq = NULL;
1905 : }
1906 :
1907 1 : nvme_rdma_free_reqs(rqpair);
1908 1 : nvme_rdma_free_rsps(rqpair->rsps);
1909 1 : rqpair->rsps = NULL;
1910 :
1911 : /* destroy cm_id last so cma device will not be freed before we destroy the cq. */
1912 1 : if (rqpair->cm_id) {
1913 0 : rdma_destroy_id(rqpair->cm_id);
1914 0 : rqpair->cm_id = NULL;
1915 : }
1916 1 : }
1917 :
1918 : static void nvme_rdma_qpair_abort_reqs(struct spdk_nvme_qpair *qpair, uint32_t dnr);
1919 :
1920 : static int
1921 1 : nvme_rdma_qpair_disconnected(struct nvme_rdma_qpair *rqpair, int ret)
1922 : {
1923 1 : if (ret) {
1924 0 : SPDK_DEBUGLOG(nvme, "Target did not respond to qpair disconnect.\n");
1925 0 : goto quiet;
1926 : }
1927 :
1928 1 : if (rqpair->poller == NULL) {
1929 : /* If poller is not used, cq is not shared.
1930 : * So complete disconnecting qpair immediately.
1931 : */
1932 1 : goto quiet;
1933 : }
1934 :
1935 0 : if (rqpair->rsps == NULL) {
1936 0 : goto quiet;
1937 : }
1938 :
1939 0 : if (rqpair->need_destroy ||
1940 0 : (rqpair->current_num_sends != 0 ||
1941 0 : (!rqpair->srq && rqpair->rsps->current_num_recvs != 0))) {
1942 0 : rqpair->state = NVME_RDMA_QPAIR_STATE_LINGERING;
1943 0 : rqpair->evt_timeout_ticks = (NVME_RDMA_DISCONNECTED_QPAIR_TIMEOUT_US * spdk_get_ticks_hz()) /
1944 0 : SPDK_SEC_TO_USEC + spdk_get_ticks();
1945 :
1946 0 : return -EAGAIN;
1947 : }
1948 :
1949 0 : quiet:
1950 1 : rqpair->state = NVME_RDMA_QPAIR_STATE_EXITED;
1951 :
1952 1 : nvme_rdma_qpair_abort_reqs(&rqpair->qpair, 0);
1953 1 : nvme_rdma_qpair_destroy(rqpair);
1954 1 : nvme_transport_ctrlr_disconnect_qpair_done(&rqpair->qpair);
1955 :
1956 1 : return 0;
1957 : }
1958 :
1959 : static int
1960 0 : nvme_rdma_qpair_wait_until_quiet(struct nvme_rdma_qpair *rqpair)
1961 : {
1962 0 : if (spdk_get_ticks() < rqpair->evt_timeout_ticks &&
1963 0 : (rqpair->current_num_sends != 0 ||
1964 0 : (!rqpair->srq && rqpair->rsps->current_num_recvs != 0))) {
1965 0 : return -EAGAIN;
1966 : }
1967 :
1968 0 : rqpair->state = NVME_RDMA_QPAIR_STATE_EXITED;
1969 :
1970 0 : nvme_rdma_qpair_abort_reqs(&rqpair->qpair, 0);
1971 0 : nvme_rdma_qpair_destroy(rqpair);
1972 0 : nvme_transport_ctrlr_disconnect_qpair_done(&rqpair->qpair);
1973 :
1974 0 : return 0;
1975 : }
1976 :
1977 : static void
1978 0 : _nvme_rdma_ctrlr_disconnect_qpair(struct spdk_nvme_ctrlr *ctrlr, struct spdk_nvme_qpair *qpair,
1979 : nvme_rdma_cm_event_cb disconnected_qpair_cb)
1980 : {
1981 0 : struct nvme_rdma_qpair *rqpair = nvme_rdma_qpair(qpair);
1982 : int rc;
1983 :
1984 0 : assert(disconnected_qpair_cb != NULL);
1985 :
1986 0 : rqpair->state = NVME_RDMA_QPAIR_STATE_EXITING;
1987 :
1988 0 : if (rqpair->cm_id) {
1989 0 : if (rqpair->rdma_qp) {
1990 0 : rc = spdk_rdma_qp_disconnect(rqpair->rdma_qp);
1991 0 : if ((qpair->ctrlr != NULL) && (rc == 0)) {
1992 0 : rc = nvme_rdma_process_event_start(rqpair, RDMA_CM_EVENT_DISCONNECTED,
1993 : disconnected_qpair_cb);
1994 0 : if (rc == 0) {
1995 0 : return;
1996 : }
1997 : }
1998 : }
1999 : }
2000 :
2001 0 : disconnected_qpair_cb(rqpair, 0);
2002 : }
2003 :
2004 : static int
2005 0 : nvme_rdma_ctrlr_disconnect_qpair_poll(struct spdk_nvme_ctrlr *ctrlr, struct spdk_nvme_qpair *qpair)
2006 : {
2007 0 : struct nvme_rdma_qpair *rqpair = nvme_rdma_qpair(qpair);
2008 : int rc;
2009 :
2010 0 : switch (rqpair->state) {
2011 0 : case NVME_RDMA_QPAIR_STATE_EXITING:
2012 0 : if (!nvme_qpair_is_admin_queue(qpair)) {
2013 0 : nvme_ctrlr_lock(ctrlr);
2014 : }
2015 :
2016 0 : rc = nvme_rdma_process_event_poll(rqpair);
2017 :
2018 0 : if (!nvme_qpair_is_admin_queue(qpair)) {
2019 0 : nvme_ctrlr_unlock(ctrlr);
2020 : }
2021 0 : break;
2022 :
2023 0 : case NVME_RDMA_QPAIR_STATE_LINGERING:
2024 0 : rc = nvme_rdma_qpair_wait_until_quiet(rqpair);
2025 0 : break;
2026 0 : case NVME_RDMA_QPAIR_STATE_EXITED:
2027 0 : rc = 0;
2028 0 : break;
2029 :
2030 0 : default:
2031 0 : assert(false);
2032 : rc = -EAGAIN;
2033 : break;
2034 : }
2035 :
2036 0 : return rc;
2037 : }
2038 :
2039 : static void
2040 0 : nvme_rdma_ctrlr_disconnect_qpair(struct spdk_nvme_ctrlr *ctrlr, struct spdk_nvme_qpair *qpair)
2041 : {
2042 : int rc;
2043 :
2044 0 : _nvme_rdma_ctrlr_disconnect_qpair(ctrlr, qpair, nvme_rdma_qpair_disconnected);
2045 :
2046 : /* If the async mode is disabled, poll the qpair until it is actually disconnected.
2047 : * It is ensured that poll_group_process_completions() calls disconnected_qpair_cb
2048 : * for any disconnected qpair. Hence, we do not have to check if the qpair is in
2049 : * a poll group or not.
2050 : * At the same time, if the qpair is being destroyed, i.e. this function is called by
2051 : * spdk_nvme_ctrlr_free_io_qpair then we need to wait until qpair is disconnected, otherwise
2052 : * we may leak some resources.
2053 : */
2054 0 : if (qpair->async && !qpair->destroy_in_progress) {
2055 0 : return;
2056 : }
2057 :
2058 : while (1) {
2059 0 : rc = nvme_rdma_ctrlr_disconnect_qpair_poll(ctrlr, qpair);
2060 0 : if (rc != -EAGAIN) {
2061 0 : break;
2062 : }
2063 : }
2064 : }
2065 :
2066 : static int
2067 0 : nvme_rdma_stale_conn_disconnected(struct nvme_rdma_qpair *rqpair, int ret)
2068 : {
2069 0 : struct spdk_nvme_qpair *qpair = &rqpair->qpair;
2070 :
2071 0 : if (ret) {
2072 0 : SPDK_DEBUGLOG(nvme, "Target did not respond to qpair disconnect.\n");
2073 : }
2074 :
2075 0 : nvme_rdma_qpair_destroy(rqpair);
2076 :
2077 0 : qpair->last_transport_failure_reason = qpair->transport_failure_reason;
2078 0 : qpair->transport_failure_reason = SPDK_NVME_QPAIR_FAILURE_NONE;
2079 :
2080 0 : rqpair->state = NVME_RDMA_QPAIR_STATE_STALE_CONN;
2081 0 : rqpair->evt_timeout_ticks = (NVME_RDMA_STALE_CONN_RETRY_DELAY_US * spdk_get_ticks_hz()) /
2082 0 : SPDK_SEC_TO_USEC + spdk_get_ticks();
2083 :
2084 0 : return 0;
2085 : }
2086 :
2087 : static int
2088 0 : nvme_rdma_stale_conn_retry(struct nvme_rdma_qpair *rqpair)
2089 : {
2090 0 : struct spdk_nvme_qpair *qpair = &rqpair->qpair;
2091 :
2092 0 : if (rqpair->stale_conn_retry_count >= NVME_RDMA_STALE_CONN_RETRY_MAX) {
2093 0 : SPDK_ERRLOG("Retry failed %d times, give up stale connection to qpair (cntlid:%u, qid:%u).\n",
2094 : NVME_RDMA_STALE_CONN_RETRY_MAX, qpair->ctrlr->cntlid, qpair->id);
2095 0 : return -ESTALE;
2096 : }
2097 :
2098 0 : rqpair->stale_conn_retry_count++;
2099 :
2100 0 : SPDK_NOTICELOG("%d times, retry stale connection to qpair (cntlid:%u, qid:%u).\n",
2101 : rqpair->stale_conn_retry_count, qpair->ctrlr->cntlid, qpair->id);
2102 :
2103 0 : _nvme_rdma_ctrlr_disconnect_qpair(qpair->ctrlr, qpair, nvme_rdma_stale_conn_disconnected);
2104 :
2105 0 : return 0;
2106 : }
2107 :
2108 : static int
2109 1 : nvme_rdma_ctrlr_delete_io_qpair(struct spdk_nvme_ctrlr *ctrlr, struct spdk_nvme_qpair *qpair)
2110 : {
2111 : struct nvme_rdma_qpair *rqpair;
2112 :
2113 1 : assert(qpair != NULL);
2114 1 : rqpair = nvme_rdma_qpair(qpair);
2115 :
2116 1 : if (rqpair->state != NVME_RDMA_QPAIR_STATE_EXITED) {
2117 : int rc __attribute__((unused));
2118 :
2119 : /* qpair was removed from the poll group while the disconnect is not finished.
2120 : * Destroy rdma resources forcefully. */
2121 1 : rc = nvme_rdma_qpair_disconnected(rqpair, 0);
2122 1 : assert(rc == 0);
2123 : }
2124 :
2125 1 : nvme_rdma_qpair_abort_reqs(qpair, 0);
2126 1 : nvme_qpair_deinit(qpair);
2127 :
2128 1 : nvme_rdma_put_memory_domain(rqpair->memory_domain);
2129 :
2130 1 : spdk_free(rqpair);
2131 :
2132 1 : return 0;
2133 : }
2134 :
2135 : static struct spdk_nvme_qpair *
2136 0 : nvme_rdma_ctrlr_create_io_qpair(struct spdk_nvme_ctrlr *ctrlr, uint16_t qid,
2137 : const struct spdk_nvme_io_qpair_opts *opts)
2138 : {
2139 0 : return nvme_rdma_ctrlr_create_qpair(ctrlr, qid, opts->io_queue_size, opts->qprio,
2140 0 : opts->io_queue_requests,
2141 0 : opts->delay_cmd_submit,
2142 0 : opts->async_mode);
2143 : }
2144 :
2145 : static int
2146 0 : nvme_rdma_ctrlr_enable(struct spdk_nvme_ctrlr *ctrlr)
2147 : {
2148 : /* do nothing here */
2149 0 : return 0;
2150 : }
2151 :
2152 : static int nvme_rdma_ctrlr_destruct(struct spdk_nvme_ctrlr *ctrlr);
2153 :
2154 : /* We have to use the typedef in the function declaration to appease astyle. */
2155 : typedef struct spdk_nvme_ctrlr spdk_nvme_ctrlr_t;
2156 :
2157 : static spdk_nvme_ctrlr_t *
2158 1 : nvme_rdma_ctrlr_construct(const struct spdk_nvme_transport_id *trid,
2159 : const struct spdk_nvme_ctrlr_opts *opts,
2160 : void *devhandle)
2161 : {
2162 : struct nvme_rdma_ctrlr *rctrlr;
2163 : struct ibv_context **contexts;
2164 1 : struct ibv_device_attr dev_attr;
2165 : int i, flag, rc;
2166 :
2167 1 : rctrlr = spdk_zmalloc(sizeof(struct nvme_rdma_ctrlr), 0, NULL, SPDK_ENV_SOCKET_ID_ANY,
2168 : SPDK_MALLOC_DMA);
2169 1 : if (rctrlr == NULL) {
2170 0 : SPDK_ERRLOG("could not allocate ctrlr\n");
2171 0 : return NULL;
2172 : }
2173 :
2174 1 : rctrlr->ctrlr.opts = *opts;
2175 1 : rctrlr->ctrlr.trid = *trid;
2176 :
2177 1 : if (opts->transport_retry_count > NVME_RDMA_CTRLR_MAX_TRANSPORT_RETRY_COUNT) {
2178 1 : SPDK_NOTICELOG("transport_retry_count exceeds max value %d, use max value\n",
2179 : NVME_RDMA_CTRLR_MAX_TRANSPORT_RETRY_COUNT);
2180 1 : rctrlr->ctrlr.opts.transport_retry_count = NVME_RDMA_CTRLR_MAX_TRANSPORT_RETRY_COUNT;
2181 : }
2182 :
2183 1 : if (opts->transport_ack_timeout > NVME_RDMA_CTRLR_MAX_TRANSPORT_ACK_TIMEOUT) {
2184 1 : SPDK_NOTICELOG("transport_ack_timeout exceeds max value %d, use max value\n",
2185 : NVME_RDMA_CTRLR_MAX_TRANSPORT_ACK_TIMEOUT);
2186 1 : rctrlr->ctrlr.opts.transport_ack_timeout = NVME_RDMA_CTRLR_MAX_TRANSPORT_ACK_TIMEOUT;
2187 : }
2188 :
2189 1 : contexts = rdma_get_devices(NULL);
2190 1 : if (contexts == NULL) {
2191 0 : SPDK_ERRLOG("rdma_get_devices() failed: %s (%d)\n", spdk_strerror(errno), errno);
2192 0 : spdk_free(rctrlr);
2193 0 : return NULL;
2194 : }
2195 :
2196 1 : i = 0;
2197 1 : rctrlr->max_sge = NVME_RDMA_MAX_SGL_DESCRIPTORS;
2198 :
2199 3 : while (contexts[i] != NULL) {
2200 2 : rc = ibv_query_device(contexts[i], &dev_attr);
2201 2 : if (rc < 0) {
2202 0 : SPDK_ERRLOG("Failed to query RDMA device attributes.\n");
2203 0 : rdma_free_devices(contexts);
2204 0 : spdk_free(rctrlr);
2205 0 : return NULL;
2206 : }
2207 2 : rctrlr->max_sge = spdk_min(rctrlr->max_sge, (uint16_t)dev_attr.max_sge);
2208 2 : i++;
2209 : }
2210 :
2211 1 : rdma_free_devices(contexts);
2212 :
2213 1 : rc = nvme_ctrlr_construct(&rctrlr->ctrlr);
2214 1 : if (rc != 0) {
2215 0 : spdk_free(rctrlr);
2216 0 : return NULL;
2217 : }
2218 :
2219 1 : STAILQ_INIT(&rctrlr->pending_cm_events);
2220 1 : STAILQ_INIT(&rctrlr->free_cm_events);
2221 1 : rctrlr->cm_events = spdk_zmalloc(NVME_RDMA_NUM_CM_EVENTS * sizeof(*rctrlr->cm_events), 0, NULL,
2222 : SPDK_ENV_SOCKET_ID_ANY, SPDK_MALLOC_DMA);
2223 1 : if (rctrlr->cm_events == NULL) {
2224 0 : SPDK_ERRLOG("unable to allocate buffers to hold CM events.\n");
2225 0 : goto destruct_ctrlr;
2226 : }
2227 :
2228 257 : for (i = 0; i < NVME_RDMA_NUM_CM_EVENTS; i++) {
2229 256 : STAILQ_INSERT_TAIL(&rctrlr->free_cm_events, &rctrlr->cm_events[i], link);
2230 : }
2231 :
2232 1 : rctrlr->cm_channel = rdma_create_event_channel();
2233 1 : if (rctrlr->cm_channel == NULL) {
2234 0 : SPDK_ERRLOG("rdma_create_event_channel() failed\n");
2235 0 : goto destruct_ctrlr;
2236 : }
2237 :
2238 1 : flag = fcntl(rctrlr->cm_channel->fd, F_GETFL);
2239 1 : if (fcntl(rctrlr->cm_channel->fd, F_SETFL, flag | O_NONBLOCK) < 0) {
2240 0 : SPDK_ERRLOG("Cannot set event channel to non blocking\n");
2241 0 : goto destruct_ctrlr;
2242 : }
2243 :
2244 2 : rctrlr->ctrlr.adminq = nvme_rdma_ctrlr_create_qpair(&rctrlr->ctrlr, 0,
2245 1 : rctrlr->ctrlr.opts.admin_queue_size, 0,
2246 1 : rctrlr->ctrlr.opts.admin_queue_size, false, true);
2247 1 : if (!rctrlr->ctrlr.adminq) {
2248 0 : SPDK_ERRLOG("failed to create admin qpair\n");
2249 0 : goto destruct_ctrlr;
2250 : }
2251 :
2252 1 : if (nvme_ctrlr_add_process(&rctrlr->ctrlr, 0) != 0) {
2253 0 : SPDK_ERRLOG("nvme_ctrlr_add_process() failed\n");
2254 0 : goto destruct_ctrlr;
2255 : }
2256 :
2257 1 : SPDK_DEBUGLOG(nvme, "successfully initialized the nvmf ctrlr\n");
2258 1 : return &rctrlr->ctrlr;
2259 :
2260 0 : destruct_ctrlr:
2261 0 : nvme_ctrlr_destruct(&rctrlr->ctrlr);
2262 0 : return NULL;
2263 : }
2264 :
2265 : static int
2266 1 : nvme_rdma_ctrlr_destruct(struct spdk_nvme_ctrlr *ctrlr)
2267 : {
2268 1 : struct nvme_rdma_ctrlr *rctrlr = nvme_rdma_ctrlr(ctrlr);
2269 : struct nvme_rdma_cm_event_entry *entry;
2270 :
2271 1 : if (ctrlr->adminq) {
2272 1 : nvme_rdma_ctrlr_delete_io_qpair(ctrlr, ctrlr->adminq);
2273 : }
2274 :
2275 1 : STAILQ_FOREACH(entry, &rctrlr->pending_cm_events, link) {
2276 0 : rdma_ack_cm_event(entry->evt);
2277 : }
2278 :
2279 1 : STAILQ_INIT(&rctrlr->free_cm_events);
2280 1 : STAILQ_INIT(&rctrlr->pending_cm_events);
2281 1 : spdk_free(rctrlr->cm_events);
2282 :
2283 1 : if (rctrlr->cm_channel) {
2284 1 : rdma_destroy_event_channel(rctrlr->cm_channel);
2285 1 : rctrlr->cm_channel = NULL;
2286 : }
2287 :
2288 1 : nvme_ctrlr_destruct_finish(ctrlr);
2289 :
2290 1 : spdk_free(rctrlr);
2291 :
2292 1 : return 0;
2293 : }
2294 :
2295 : static int
2296 2 : nvme_rdma_qpair_submit_request(struct spdk_nvme_qpair *qpair,
2297 : struct nvme_request *req)
2298 : {
2299 : struct nvme_rdma_qpair *rqpair;
2300 : struct spdk_nvme_rdma_req *rdma_req;
2301 : struct ibv_send_wr *wr;
2302 :
2303 2 : rqpair = nvme_rdma_qpair(qpair);
2304 2 : assert(rqpair != NULL);
2305 2 : assert(req != NULL);
2306 :
2307 2 : rdma_req = nvme_rdma_req_get(rqpair);
2308 2 : if (spdk_unlikely(!rdma_req)) {
2309 1 : if (rqpair->poller) {
2310 1 : rqpair->poller->stats.queued_requests++;
2311 : }
2312 : /* Inform the upper layer to try again later. */
2313 1 : return -EAGAIN;
2314 : }
2315 :
2316 1 : if (nvme_rdma_req_init(rqpair, req, rdma_req)) {
2317 0 : SPDK_ERRLOG("nvme_rdma_req_init() failed\n");
2318 0 : nvme_rdma_req_put(rqpair, rdma_req);
2319 0 : return -1;
2320 : }
2321 :
2322 1 : TAILQ_INSERT_TAIL(&rqpair->outstanding_reqs, rdma_req, link);
2323 :
2324 1 : assert(rqpair->current_num_sends < rqpair->num_entries);
2325 1 : rqpair->current_num_sends++;
2326 :
2327 1 : wr = &rdma_req->send_wr;
2328 1 : wr->next = NULL;
2329 1 : nvme_rdma_trace_ibv_sge(wr->sg_list);
2330 :
2331 1 : spdk_rdma_qp_queue_send_wrs(rqpair->rdma_qp, wr);
2332 :
2333 1 : if (!rqpair->delay_cmd_submit) {
2334 1 : return nvme_rdma_qpair_submit_sends(rqpair);
2335 : }
2336 :
2337 0 : return 0;
2338 : }
2339 :
2340 : static int
2341 0 : nvme_rdma_qpair_reset(struct spdk_nvme_qpair *qpair)
2342 : {
2343 : /* Currently, doing nothing here */
2344 0 : return 0;
2345 : }
2346 :
2347 : static void
2348 2 : nvme_rdma_qpair_abort_reqs(struct spdk_nvme_qpair *qpair, uint32_t dnr)
2349 : {
2350 : struct spdk_nvme_rdma_req *rdma_req, *tmp;
2351 2 : struct spdk_nvme_cpl cpl;
2352 2 : struct nvme_rdma_qpair *rqpair = nvme_rdma_qpair(qpair);
2353 :
2354 2 : cpl.sqid = qpair->id;
2355 2 : cpl.status.sc = SPDK_NVME_SC_ABORTED_SQ_DELETION;
2356 2 : cpl.status.sct = SPDK_NVME_SCT_GENERIC;
2357 2 : cpl.status.dnr = dnr;
2358 :
2359 : /*
2360 : * We cannot abort requests at the RDMA layer without
2361 : * unregistering them. If we do, we can still get error
2362 : * free completions on the shared completion queue.
2363 : */
2364 2 : if (nvme_qpair_get_state(qpair) > NVME_QPAIR_DISCONNECTING &&
2365 0 : nvme_qpair_get_state(qpair) != NVME_QPAIR_DESTROYING) {
2366 0 : nvme_ctrlr_disconnect_qpair(qpair);
2367 : }
2368 :
2369 2 : TAILQ_FOREACH_SAFE(rdma_req, &rqpair->outstanding_reqs, link, tmp) {
2370 0 : nvme_rdma_req_complete(rdma_req, &cpl, true);
2371 : }
2372 2 : }
2373 :
2374 : static void
2375 0 : nvme_rdma_qpair_check_timeout(struct spdk_nvme_qpair *qpair)
2376 : {
2377 : uint64_t t02;
2378 : struct spdk_nvme_rdma_req *rdma_req, *tmp;
2379 0 : struct nvme_rdma_qpair *rqpair = nvme_rdma_qpair(qpair);
2380 0 : struct spdk_nvme_ctrlr *ctrlr = qpair->ctrlr;
2381 : struct spdk_nvme_ctrlr_process *active_proc;
2382 :
2383 : /* Don't check timeouts during controller initialization. */
2384 0 : if (ctrlr->state != NVME_CTRLR_STATE_READY) {
2385 0 : return;
2386 : }
2387 :
2388 0 : if (nvme_qpair_is_admin_queue(qpair)) {
2389 0 : active_proc = nvme_ctrlr_get_current_process(ctrlr);
2390 : } else {
2391 0 : active_proc = qpair->active_proc;
2392 : }
2393 :
2394 : /* Only check timeouts if the current process has a timeout callback. */
2395 0 : if (active_proc == NULL || active_proc->timeout_cb_fn == NULL) {
2396 0 : return;
2397 : }
2398 :
2399 0 : t02 = spdk_get_ticks();
2400 0 : TAILQ_FOREACH_SAFE(rdma_req, &rqpair->outstanding_reqs, link, tmp) {
2401 0 : assert(rdma_req->req != NULL);
2402 :
2403 0 : if (nvme_request_check_timeout(rdma_req->req, rdma_req->id, active_proc, t02)) {
2404 : /*
2405 : * The requests are in order, so as soon as one has not timed out,
2406 : * stop iterating.
2407 : */
2408 0 : break;
2409 : }
2410 : }
2411 : }
2412 :
2413 : static inline void
2414 0 : nvme_rdma_request_ready(struct nvme_rdma_qpair *rqpair, struct spdk_nvme_rdma_req *rdma_req)
2415 : {
2416 0 : struct spdk_nvme_rdma_rsp *rdma_rsp = rdma_req->rdma_rsp;
2417 0 : struct ibv_recv_wr *recv_wr = rdma_rsp->recv_wr;
2418 :
2419 0 : nvme_rdma_req_complete(rdma_req, &rdma_rsp->cpl, true);
2420 :
2421 0 : assert(rqpair->rsps->current_num_recvs < rqpair->rsps->num_entries);
2422 0 : rqpair->rsps->current_num_recvs++;
2423 :
2424 0 : recv_wr->next = NULL;
2425 0 : nvme_rdma_trace_ibv_sge(recv_wr->sg_list);
2426 :
2427 0 : if (!rqpair->srq) {
2428 0 : spdk_rdma_qp_queue_recv_wrs(rqpair->rdma_qp, recv_wr);
2429 : } else {
2430 0 : spdk_rdma_srq_queue_recv_wrs(rqpair->srq, recv_wr);
2431 : }
2432 0 : }
2433 :
2434 : #define MAX_COMPLETIONS_PER_POLL 128
2435 :
2436 : static void
2437 0 : nvme_rdma_fail_qpair(struct spdk_nvme_qpair *qpair, int failure_reason)
2438 : {
2439 0 : if (failure_reason == IBV_WC_RETRY_EXC_ERR) {
2440 0 : qpair->transport_failure_reason = SPDK_NVME_QPAIR_FAILURE_REMOTE;
2441 0 : } else if (qpair->transport_failure_reason == SPDK_NVME_QPAIR_FAILURE_NONE) {
2442 0 : qpair->transport_failure_reason = SPDK_NVME_QPAIR_FAILURE_UNKNOWN;
2443 : }
2444 :
2445 0 : nvme_ctrlr_disconnect_qpair(qpair);
2446 0 : }
2447 :
2448 : static struct nvme_rdma_qpair *
2449 4 : get_rdma_qpair_from_wc(struct nvme_rdma_poll_group *group, struct ibv_wc *wc)
2450 : {
2451 : struct spdk_nvme_qpair *qpair;
2452 : struct nvme_rdma_qpair *rqpair;
2453 :
2454 5 : STAILQ_FOREACH(qpair, &group->group.connected_qpairs, poll_group_stailq) {
2455 2 : rqpair = nvme_rdma_qpair(qpair);
2456 2 : if (NVME_RDMA_POLL_GROUP_CHECK_QPN(rqpair, wc->qp_num)) {
2457 1 : return rqpair;
2458 : }
2459 : }
2460 :
2461 4 : STAILQ_FOREACH(qpair, &group->group.disconnected_qpairs, poll_group_stailq) {
2462 2 : rqpair = nvme_rdma_qpair(qpair);
2463 2 : if (NVME_RDMA_POLL_GROUP_CHECK_QPN(rqpair, wc->qp_num)) {
2464 1 : return rqpair;
2465 : }
2466 : }
2467 :
2468 2 : return NULL;
2469 : }
2470 :
2471 : static inline void
2472 0 : nvme_rdma_log_wc_status(struct nvme_rdma_qpair *rqpair, struct ibv_wc *wc)
2473 : {
2474 0 : struct nvme_rdma_wr *rdma_wr = (struct nvme_rdma_wr *)wc->wr_id;
2475 :
2476 0 : if (wc->status == IBV_WC_WR_FLUSH_ERR) {
2477 : /* If qpair is in ERR state, we will receive completions for all posted and not completed
2478 : * Work Requests with IBV_WC_WR_FLUSH_ERR status. Don't log an error in that case */
2479 0 : SPDK_DEBUGLOG(nvme, "WC error, qid %u, qp state %d, request 0x%lu type %d, status: (%d): %s\n",
2480 : rqpair->qpair.id, rqpair->qpair.state, wc->wr_id, rdma_wr->type, wc->status,
2481 : ibv_wc_status_str(wc->status));
2482 : } else {
2483 0 : SPDK_ERRLOG("WC error, qid %u, qp state %d, request 0x%lu type %d, status: (%d): %s\n",
2484 : rqpair->qpair.id, rqpair->qpair.state, wc->wr_id, rdma_wr->type, wc->status,
2485 : ibv_wc_status_str(wc->status));
2486 : }
2487 0 : }
2488 :
2489 : static inline int
2490 0 : nvme_rdma_process_recv_completion(struct nvme_rdma_poller *poller, struct ibv_wc *wc,
2491 : struct nvme_rdma_wr *rdma_wr)
2492 : {
2493 : struct nvme_rdma_qpair *rqpair;
2494 : struct spdk_nvme_rdma_req *rdma_req;
2495 : struct spdk_nvme_rdma_rsp *rdma_rsp;
2496 :
2497 0 : rdma_rsp = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvme_rdma_rsp, rdma_wr);
2498 :
2499 0 : if (poller && poller->srq) {
2500 0 : rqpair = get_rdma_qpair_from_wc(poller->group, wc);
2501 0 : if (spdk_unlikely(!rqpair)) {
2502 : /* Since we do not handle the LAST_WQE_REACHED event, we do not know when
2503 : * a Receive Queue in a QP, that is associated with an SRQ, is flushed.
2504 : * We may get a WC for a already destroyed QP.
2505 : *
2506 : * However, for the SRQ, this is not any error. Hence, just re-post the
2507 : * receive request to the SRQ to reuse for other QPs, and return 0.
2508 : */
2509 0 : spdk_rdma_srq_queue_recv_wrs(poller->srq, rdma_rsp->recv_wr);
2510 0 : return 0;
2511 : }
2512 : } else {
2513 0 : rqpair = rdma_rsp->rqpair;
2514 0 : if (spdk_unlikely(!rqpair)) {
2515 : /* TODO: Fix forceful QP destroy when it is not async mode.
2516 : * CQ itself did not cause any error. Hence, return 0 for now.
2517 : */
2518 0 : SPDK_WARNLOG("QP might be already destroyed.\n");
2519 0 : return 0;
2520 : }
2521 : }
2522 :
2523 :
2524 0 : assert(rqpair->rsps->current_num_recvs > 0);
2525 0 : rqpair->rsps->current_num_recvs--;
2526 :
2527 0 : if (wc->status) {
2528 0 : nvme_rdma_log_wc_status(rqpair, wc);
2529 0 : goto err_wc;
2530 : }
2531 :
2532 0 : SPDK_DEBUGLOG(nvme, "CQ recv completion\n");
2533 :
2534 0 : if (wc->byte_len < sizeof(struct spdk_nvme_cpl)) {
2535 0 : SPDK_ERRLOG("recv length %u less than expected response size\n", wc->byte_len);
2536 0 : goto err_wc;
2537 : }
2538 0 : rdma_req = &rqpair->rdma_reqs[rdma_rsp->cpl.cid];
2539 0 : rdma_req->completion_flags |= NVME_RDMA_RECV_COMPLETED;
2540 0 : rdma_req->rdma_rsp = rdma_rsp;
2541 :
2542 0 : if ((rdma_req->completion_flags & NVME_RDMA_SEND_COMPLETED) == 0) {
2543 0 : return 0;
2544 : }
2545 :
2546 0 : nvme_rdma_request_ready(rqpair, rdma_req);
2547 :
2548 0 : if (!rqpair->delay_cmd_submit) {
2549 0 : if (spdk_unlikely(nvme_rdma_qpair_submit_recvs(rqpair))) {
2550 0 : SPDK_ERRLOG("Unable to re-post rx descriptor\n");
2551 0 : nvme_rdma_fail_qpair(&rqpair->qpair, 0);
2552 0 : return -ENXIO;
2553 : }
2554 : }
2555 :
2556 0 : rqpair->num_completions++;
2557 0 : return 1;
2558 :
2559 0 : err_wc:
2560 0 : nvme_rdma_fail_qpair(&rqpair->qpair, 0);
2561 0 : if (poller && poller->srq) {
2562 0 : spdk_rdma_srq_queue_recv_wrs(poller->srq, rdma_rsp->recv_wr);
2563 : }
2564 0 : return -ENXIO;
2565 : }
2566 :
2567 : static inline int
2568 0 : nvme_rdma_process_send_completion(struct nvme_rdma_poller *poller,
2569 : struct nvme_rdma_qpair *rdma_qpair,
2570 : struct ibv_wc *wc, struct nvme_rdma_wr *rdma_wr)
2571 : {
2572 : struct nvme_rdma_qpair *rqpair;
2573 : struct spdk_nvme_rdma_req *rdma_req;
2574 :
2575 0 : rdma_req = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvme_rdma_req, rdma_wr);
2576 0 : rqpair = rdma_req->req ? nvme_rdma_qpair(rdma_req->req->qpair) : NULL;
2577 0 : if (!rqpair) {
2578 0 : rqpair = rdma_qpair != NULL ? rdma_qpair : get_rdma_qpair_from_wc(poller->group, wc);
2579 : }
2580 :
2581 : /* If we are flushing I/O */
2582 0 : if (wc->status) {
2583 0 : if (!rqpair) {
2584 : /* When poll_group is used, several qpairs share the same CQ and it is possible to
2585 : * receive a completion with error (e.g. IBV_WC_WR_FLUSH_ERR) for already disconnected qpair
2586 : * That happens due to qpair is destroyed while there are submitted but not completed send/receive
2587 : * Work Requests */
2588 0 : assert(poller);
2589 0 : return 0;
2590 : }
2591 0 : assert(rqpair->current_num_sends > 0);
2592 0 : rqpair->current_num_sends--;
2593 0 : nvme_rdma_log_wc_status(rqpair, wc);
2594 0 : nvme_rdma_fail_qpair(&rqpair->qpair, 0);
2595 0 : if (rdma_req->rdma_rsp && poller && poller->srq) {
2596 0 : spdk_rdma_srq_queue_recv_wrs(poller->srq, rdma_req->rdma_rsp->recv_wr);
2597 : }
2598 0 : return -ENXIO;
2599 : }
2600 :
2601 : /* We do not support Soft Roce anymore. Other than Soft Roce's bug, we should not
2602 : * receive a completion without error status after qpair is disconnected/destroyed.
2603 : */
2604 0 : if (spdk_unlikely(rdma_req->req == NULL)) {
2605 : /*
2606 : * Some infiniband drivers do not guarantee the previous assumption after we
2607 : * received a RDMA_CM_EVENT_DEVICE_REMOVAL event.
2608 : */
2609 0 : SPDK_ERRLOG("Received malformed completion: request 0x%"PRIx64" type %d\n", wc->wr_id,
2610 : rdma_wr->type);
2611 0 : if (!rqpair || !rqpair->need_destroy) {
2612 0 : assert(0);
2613 : }
2614 0 : return -ENXIO;
2615 : }
2616 :
2617 0 : rdma_req->completion_flags |= NVME_RDMA_SEND_COMPLETED;
2618 0 : assert(rqpair->current_num_sends > 0);
2619 0 : rqpair->current_num_sends--;
2620 :
2621 0 : if ((rdma_req->completion_flags & NVME_RDMA_RECV_COMPLETED) == 0) {
2622 0 : return 0;
2623 : }
2624 :
2625 0 : nvme_rdma_request_ready(rqpair, rdma_req);
2626 :
2627 0 : if (!rqpair->delay_cmd_submit) {
2628 0 : if (spdk_unlikely(nvme_rdma_qpair_submit_recvs(rqpair))) {
2629 0 : SPDK_ERRLOG("Unable to re-post rx descriptor\n");
2630 0 : nvme_rdma_fail_qpair(&rqpair->qpair, 0);
2631 0 : return -ENXIO;
2632 : }
2633 : }
2634 :
2635 0 : rqpair->num_completions++;
2636 0 : return 1;
2637 : }
2638 :
2639 : static int
2640 0 : nvme_rdma_cq_process_completions(struct ibv_cq *cq, uint32_t batch_size,
2641 : struct nvme_rdma_poller *poller,
2642 : struct nvme_rdma_qpair *rdma_qpair,
2643 : uint64_t *rdma_completions)
2644 : {
2645 0 : struct ibv_wc wc[MAX_COMPLETIONS_PER_POLL];
2646 : struct nvme_rdma_wr *rdma_wr;
2647 0 : uint32_t reaped = 0;
2648 0 : int completion_rc = 0;
2649 : int rc, _rc, i;
2650 :
2651 0 : rc = ibv_poll_cq(cq, batch_size, wc);
2652 0 : if (rc < 0) {
2653 0 : SPDK_ERRLOG("Error polling CQ! (%d): %s\n",
2654 : errno, spdk_strerror(errno));
2655 0 : return -ECANCELED;
2656 0 : } else if (rc == 0) {
2657 0 : return 0;
2658 : }
2659 :
2660 0 : for (i = 0; i < rc; i++) {
2661 0 : rdma_wr = (struct nvme_rdma_wr *)wc[i].wr_id;
2662 0 : switch (rdma_wr->type) {
2663 0 : case RDMA_WR_TYPE_RECV:
2664 0 : _rc = nvme_rdma_process_recv_completion(poller, &wc[i], rdma_wr);
2665 0 : break;
2666 :
2667 0 : case RDMA_WR_TYPE_SEND:
2668 0 : _rc = nvme_rdma_process_send_completion(poller, rdma_qpair, &wc[i], rdma_wr);
2669 0 : break;
2670 :
2671 0 : default:
2672 0 : SPDK_ERRLOG("Received an unexpected opcode on the CQ: %d\n", rdma_wr->type);
2673 0 : return -ECANCELED;
2674 : }
2675 0 : if (spdk_likely(_rc >= 0)) {
2676 0 : reaped += _rc;
2677 : } else {
2678 0 : completion_rc = _rc;
2679 : }
2680 : }
2681 :
2682 0 : *rdma_completions += rc;
2683 :
2684 0 : if (completion_rc) {
2685 0 : return completion_rc;
2686 : }
2687 :
2688 0 : return reaped;
2689 : }
2690 :
2691 : static void
2692 0 : dummy_disconnected_qpair_cb(struct spdk_nvme_qpair *qpair, void *poll_group_ctx)
2693 : {
2694 :
2695 0 : }
2696 :
2697 : static int
2698 0 : nvme_rdma_qpair_process_completions(struct spdk_nvme_qpair *qpair,
2699 : uint32_t max_completions)
2700 : {
2701 0 : struct nvme_rdma_qpair *rqpair = nvme_rdma_qpair(qpair);
2702 0 : struct nvme_rdma_ctrlr *rctrlr = nvme_rdma_ctrlr(qpair->ctrlr);
2703 0 : int rc = 0, batch_size;
2704 : struct ibv_cq *cq;
2705 0 : uint64_t rdma_completions = 0;
2706 :
2707 : /*
2708 : * This is used during the connection phase. It's possible that we are still reaping error completions
2709 : * from other qpairs so we need to call the poll group function. Also, it's more correct since the cq
2710 : * is shared.
2711 : */
2712 0 : if (qpair->poll_group != NULL) {
2713 0 : return spdk_nvme_poll_group_process_completions(qpair->poll_group->group, max_completions,
2714 : dummy_disconnected_qpair_cb);
2715 : }
2716 :
2717 0 : if (max_completions == 0) {
2718 0 : max_completions = rqpair->num_entries;
2719 : } else {
2720 0 : max_completions = spdk_min(max_completions, rqpair->num_entries);
2721 : }
2722 :
2723 0 : switch (nvme_qpair_get_state(qpair)) {
2724 0 : case NVME_QPAIR_CONNECTING:
2725 0 : rc = nvme_rdma_ctrlr_connect_qpair_poll(qpair->ctrlr, qpair);
2726 0 : if (rc == 0) {
2727 : /* Once the connection is completed, we can submit queued requests */
2728 0 : nvme_qpair_resubmit_requests(qpair, rqpair->num_entries);
2729 0 : } else if (rc != -EAGAIN) {
2730 0 : SPDK_ERRLOG("Failed to connect rqpair=%p\n", rqpair);
2731 0 : goto failed;
2732 0 : } else if (rqpair->state <= NVME_RDMA_QPAIR_STATE_INITIALIZING) {
2733 0 : return 0;
2734 : }
2735 0 : break;
2736 :
2737 0 : case NVME_QPAIR_DISCONNECTING:
2738 0 : nvme_rdma_ctrlr_disconnect_qpair_poll(qpair->ctrlr, qpair);
2739 0 : return -ENXIO;
2740 :
2741 0 : default:
2742 0 : if (nvme_qpair_is_admin_queue(qpair)) {
2743 0 : nvme_rdma_poll_events(rctrlr);
2744 : }
2745 0 : nvme_rdma_qpair_process_cm_event(rqpair);
2746 0 : break;
2747 : }
2748 :
2749 0 : if (spdk_unlikely(qpair->transport_failure_reason != SPDK_NVME_QPAIR_FAILURE_NONE)) {
2750 0 : goto failed;
2751 : }
2752 :
2753 0 : cq = rqpair->cq;
2754 :
2755 0 : rqpair->num_completions = 0;
2756 : do {
2757 0 : batch_size = spdk_min((max_completions - rqpair->num_completions), MAX_COMPLETIONS_PER_POLL);
2758 0 : rc = nvme_rdma_cq_process_completions(cq, batch_size, NULL, rqpair, &rdma_completions);
2759 :
2760 0 : if (rc == 0) {
2761 0 : break;
2762 : /* Handle the case where we fail to poll the cq. */
2763 0 : } else if (rc == -ECANCELED) {
2764 0 : goto failed;
2765 0 : } else if (rc == -ENXIO) {
2766 0 : return rc;
2767 : }
2768 0 : } while (rqpair->num_completions < max_completions);
2769 :
2770 0 : if (spdk_unlikely(nvme_rdma_qpair_submit_sends(rqpair) ||
2771 : nvme_rdma_qpair_submit_recvs(rqpair))) {
2772 0 : goto failed;
2773 : }
2774 :
2775 0 : if (spdk_unlikely(qpair->ctrlr->timeout_enabled)) {
2776 0 : nvme_rdma_qpair_check_timeout(qpair);
2777 : }
2778 :
2779 0 : return rqpair->num_completions;
2780 :
2781 0 : failed:
2782 0 : nvme_rdma_fail_qpair(qpair, 0);
2783 0 : return -ENXIO;
2784 : }
2785 :
2786 : static uint32_t
2787 0 : nvme_rdma_ctrlr_get_max_xfer_size(struct spdk_nvme_ctrlr *ctrlr)
2788 : {
2789 : /* max_mr_size by ibv_query_device indicates the largest value that we can
2790 : * set for a registered memory region. It is independent from the actual
2791 : * I/O size and is very likely to be larger than 2 MiB which is the
2792 : * granularity we currently register memory regions. Hence return
2793 : * UINT32_MAX here and let the generic layer use the controller data to
2794 : * moderate this value.
2795 : */
2796 0 : return UINT32_MAX;
2797 : }
2798 :
2799 : static uint16_t
2800 5 : nvme_rdma_ctrlr_get_max_sges(struct spdk_nvme_ctrlr *ctrlr)
2801 : {
2802 5 : struct nvme_rdma_ctrlr *rctrlr = nvme_rdma_ctrlr(ctrlr);
2803 5 : uint32_t max_sge = rctrlr->max_sge;
2804 5 : uint32_t max_in_capsule_sge = (ctrlr->cdata.nvmf_specific.ioccsz * 16 -
2805 5 : sizeof(struct spdk_nvme_cmd)) /
2806 : sizeof(struct spdk_nvme_sgl_descriptor);
2807 :
2808 : /* Max SGE is limited by capsule size */
2809 5 : max_sge = spdk_min(max_sge, max_in_capsule_sge);
2810 : /* Max SGE may be limited by MSDBD */
2811 5 : if (ctrlr->cdata.nvmf_specific.msdbd != 0) {
2812 5 : max_sge = spdk_min(max_sge, ctrlr->cdata.nvmf_specific.msdbd);
2813 : }
2814 :
2815 : /* Max SGE can't be less than 1 */
2816 5 : max_sge = spdk_max(1, max_sge);
2817 5 : return max_sge;
2818 : }
2819 :
2820 : static int
2821 0 : nvme_rdma_qpair_iterate_requests(struct spdk_nvme_qpair *qpair,
2822 : int (*iter_fn)(struct nvme_request *req, void *arg),
2823 : void *arg)
2824 : {
2825 0 : struct nvme_rdma_qpair *rqpair = nvme_rdma_qpair(qpair);
2826 : struct spdk_nvme_rdma_req *rdma_req, *tmp;
2827 : int rc;
2828 :
2829 0 : assert(iter_fn != NULL);
2830 :
2831 0 : TAILQ_FOREACH_SAFE(rdma_req, &rqpair->outstanding_reqs, link, tmp) {
2832 0 : assert(rdma_req->req != NULL);
2833 :
2834 0 : rc = iter_fn(rdma_req->req, arg);
2835 0 : if (rc != 0) {
2836 0 : return rc;
2837 : }
2838 : }
2839 :
2840 0 : return 0;
2841 : }
2842 :
2843 : static void
2844 0 : nvme_rdma_admin_qpair_abort_aers(struct spdk_nvme_qpair *qpair)
2845 : {
2846 : struct spdk_nvme_rdma_req *rdma_req, *tmp;
2847 0 : struct spdk_nvme_cpl cpl;
2848 0 : struct nvme_rdma_qpair *rqpair = nvme_rdma_qpair(qpair);
2849 :
2850 0 : cpl.status.sc = SPDK_NVME_SC_ABORTED_SQ_DELETION;
2851 0 : cpl.status.sct = SPDK_NVME_SCT_GENERIC;
2852 :
2853 0 : TAILQ_FOREACH_SAFE(rdma_req, &rqpair->outstanding_reqs, link, tmp) {
2854 0 : assert(rdma_req->req != NULL);
2855 :
2856 0 : if (rdma_req->req->cmd.opc != SPDK_NVME_OPC_ASYNC_EVENT_REQUEST) {
2857 0 : continue;
2858 : }
2859 :
2860 0 : nvme_rdma_req_complete(rdma_req, &cpl, false);
2861 : }
2862 0 : }
2863 :
2864 : static void
2865 9 : nvme_rdma_poller_destroy(struct nvme_rdma_poller *poller)
2866 : {
2867 9 : if (poller->cq) {
2868 7 : ibv_destroy_cq(poller->cq);
2869 : }
2870 9 : if (poller->rsps) {
2871 0 : nvme_rdma_free_rsps(poller->rsps);
2872 : }
2873 9 : if (poller->srq) {
2874 0 : spdk_rdma_srq_destroy(poller->srq);
2875 : }
2876 9 : if (poller->mr_map) {
2877 0 : spdk_rdma_free_mem_map(&poller->mr_map);
2878 : }
2879 9 : if (poller->pd) {
2880 0 : spdk_rdma_put_pd(poller->pd);
2881 : }
2882 9 : free(poller);
2883 9 : }
2884 :
2885 : static struct nvme_rdma_poller *
2886 9 : nvme_rdma_poller_create(struct nvme_rdma_poll_group *group, struct ibv_context *ctx)
2887 : {
2888 : struct nvme_rdma_poller *poller;
2889 9 : struct ibv_device_attr dev_attr;
2890 9 : struct spdk_rdma_srq_init_attr srq_init_attr = {};
2891 9 : struct nvme_rdma_rsp_opts opts;
2892 : int num_cqe, max_num_cqe;
2893 : int rc;
2894 :
2895 9 : poller = calloc(1, sizeof(*poller));
2896 9 : if (poller == NULL) {
2897 0 : SPDK_ERRLOG("Unable to allocate poller.\n");
2898 0 : return NULL;
2899 : }
2900 :
2901 9 : poller->group = group;
2902 9 : poller->device = ctx;
2903 :
2904 9 : if (g_spdk_nvme_transport_opts.rdma_srq_size != 0) {
2905 0 : rc = ibv_query_device(ctx, &dev_attr);
2906 0 : if (rc) {
2907 0 : SPDK_ERRLOG("Unable to query RDMA device.\n");
2908 0 : goto fail;
2909 : }
2910 :
2911 0 : poller->pd = spdk_rdma_get_pd(ctx);
2912 0 : if (poller->pd == NULL) {
2913 0 : SPDK_ERRLOG("Unable to get PD.\n");
2914 0 : goto fail;
2915 : }
2916 :
2917 0 : poller->mr_map = spdk_rdma_create_mem_map(poller->pd, &g_nvme_hooks,
2918 : SPDK_RDMA_MEMORY_MAP_ROLE_INITIATOR);
2919 0 : if (poller->mr_map == NULL) {
2920 0 : SPDK_ERRLOG("Unable to create memory map.\n");
2921 0 : goto fail;
2922 : }
2923 :
2924 0 : srq_init_attr.stats = &poller->stats.rdma_stats.recv;
2925 0 : srq_init_attr.pd = poller->pd;
2926 0 : srq_init_attr.srq_init_attr.attr.max_wr = spdk_min((uint32_t)dev_attr.max_srq_wr,
2927 : g_spdk_nvme_transport_opts.rdma_srq_size);
2928 0 : srq_init_attr.srq_init_attr.attr.max_sge = spdk_min(dev_attr.max_sge,
2929 : NVME_RDMA_DEFAULT_RX_SGE);
2930 :
2931 0 : poller->srq = spdk_rdma_srq_create(&srq_init_attr);
2932 0 : if (poller->srq == NULL) {
2933 0 : SPDK_ERRLOG("Unable to create SRQ.\n");
2934 0 : goto fail;
2935 : }
2936 :
2937 0 : opts.num_entries = g_spdk_nvme_transport_opts.rdma_srq_size;
2938 0 : opts.rqpair = NULL;
2939 0 : opts.srq = poller->srq;
2940 0 : opts.mr_map = poller->mr_map;
2941 :
2942 0 : poller->rsps = nvme_rdma_create_rsps(&opts);
2943 0 : if (poller->rsps == NULL) {
2944 0 : SPDK_ERRLOG("Unable to create poller RDMA responses.\n");
2945 0 : goto fail;
2946 : }
2947 :
2948 0 : rc = nvme_rdma_poller_submit_recvs(poller);
2949 0 : if (rc) {
2950 0 : SPDK_ERRLOG("Unable to submit poller RDMA responses.\n");
2951 0 : goto fail;
2952 : }
2953 :
2954 : /*
2955 : * When using an srq, fix the size of the completion queue at startup.
2956 : * The initiator sends only send and recv WRs. Hence, the multiplier is 2.
2957 : * (The target sends also data WRs. Hence, the multiplier is 3.)
2958 : */
2959 0 : num_cqe = g_spdk_nvme_transport_opts.rdma_srq_size * 2;
2960 : } else {
2961 9 : num_cqe = DEFAULT_NVME_RDMA_CQ_SIZE;
2962 : }
2963 :
2964 9 : max_num_cqe = g_spdk_nvme_transport_opts.rdma_max_cq_size;
2965 9 : if (max_num_cqe != 0 && num_cqe > max_num_cqe) {
2966 0 : num_cqe = max_num_cqe;
2967 : }
2968 :
2969 9 : poller->cq = ibv_create_cq(poller->device, num_cqe, group, NULL, 0);
2970 :
2971 9 : if (poller->cq == NULL) {
2972 2 : SPDK_ERRLOG("Unable to create CQ, errno %d.\n", errno);
2973 2 : goto fail;
2974 : }
2975 :
2976 7 : STAILQ_INSERT_HEAD(&group->pollers, poller, link);
2977 7 : group->num_pollers++;
2978 7 : poller->current_num_wc = num_cqe;
2979 7 : poller->required_num_wc = 0;
2980 7 : return poller;
2981 :
2982 2 : fail:
2983 2 : nvme_rdma_poller_destroy(poller);
2984 2 : return NULL;
2985 : }
2986 :
2987 : static void
2988 3 : nvme_rdma_poll_group_free_pollers(struct nvme_rdma_poll_group *group)
2989 : {
2990 : struct nvme_rdma_poller *poller, *tmp_poller;
2991 :
2992 5 : STAILQ_FOREACH_SAFE(poller, &group->pollers, link, tmp_poller) {
2993 2 : assert(poller->refcnt == 0);
2994 2 : if (poller->refcnt) {
2995 0 : SPDK_WARNLOG("Destroying poller with non-zero ref count: poller %p, refcnt %d\n",
2996 : poller, poller->refcnt);
2997 : }
2998 :
2999 2 : STAILQ_REMOVE(&group->pollers, poller, nvme_rdma_poller, link);
3000 2 : nvme_rdma_poller_destroy(poller);
3001 : }
3002 3 : }
3003 :
3004 : static struct nvme_rdma_poller *
3005 8 : nvme_rdma_poll_group_get_poller(struct nvme_rdma_poll_group *group, struct ibv_context *device)
3006 : {
3007 8 : struct nvme_rdma_poller *poller = NULL;
3008 :
3009 10 : STAILQ_FOREACH(poller, &group->pollers, link) {
3010 3 : if (poller->device == device) {
3011 1 : break;
3012 : }
3013 : }
3014 :
3015 8 : if (!poller) {
3016 7 : poller = nvme_rdma_poller_create(group, device);
3017 7 : if (!poller) {
3018 2 : SPDK_ERRLOG("Failed to create a poller for device %p\n", device);
3019 2 : return NULL;
3020 : }
3021 : }
3022 :
3023 6 : poller->refcnt++;
3024 6 : return poller;
3025 : }
3026 :
3027 : static void
3028 6 : nvme_rdma_poll_group_put_poller(struct nvme_rdma_poll_group *group, struct nvme_rdma_poller *poller)
3029 : {
3030 6 : assert(poller->refcnt > 0);
3031 6 : if (--poller->refcnt == 0) {
3032 5 : STAILQ_REMOVE(&group->pollers, poller, nvme_rdma_poller, link);
3033 5 : group->num_pollers--;
3034 5 : nvme_rdma_poller_destroy(poller);
3035 : }
3036 6 : }
3037 :
3038 : static struct spdk_nvme_transport_poll_group *
3039 1 : nvme_rdma_poll_group_create(void)
3040 : {
3041 : struct nvme_rdma_poll_group *group;
3042 :
3043 1 : group = calloc(1, sizeof(*group));
3044 1 : if (group == NULL) {
3045 0 : SPDK_ERRLOG("Unable to allocate poll group.\n");
3046 0 : return NULL;
3047 : }
3048 :
3049 1 : STAILQ_INIT(&group->pollers);
3050 1 : TAILQ_INIT(&group->connecting_qpairs);
3051 1 : return &group->group;
3052 : }
3053 :
3054 : static int
3055 0 : nvme_rdma_poll_group_connect_qpair(struct spdk_nvme_qpair *qpair)
3056 : {
3057 0 : return 0;
3058 : }
3059 :
3060 : static int
3061 0 : nvme_rdma_poll_group_disconnect_qpair(struct spdk_nvme_qpair *qpair)
3062 : {
3063 0 : struct nvme_rdma_qpair *rqpair = nvme_rdma_qpair(qpair);
3064 0 : struct nvme_rdma_poll_group *group = nvme_rdma_poll_group(qpair->poll_group);;
3065 :
3066 0 : if (rqpair->link_connecting.tqe_prev) {
3067 0 : TAILQ_REMOVE(&group->connecting_qpairs, rqpair, link_connecting);
3068 : /* We use prev pointer to check if qpair is in connecting list or not .
3069 : * TAILQ_REMOVE doesn't do it. So, we do it manually.
3070 : */
3071 0 : rqpair->link_connecting.tqe_prev = NULL;
3072 : }
3073 :
3074 0 : return 0;
3075 : }
3076 :
3077 : static int
3078 0 : nvme_rdma_poll_group_add(struct spdk_nvme_transport_poll_group *tgroup,
3079 : struct spdk_nvme_qpair *qpair)
3080 : {
3081 0 : return 0;
3082 : }
3083 :
3084 : static int
3085 0 : nvme_rdma_poll_group_remove(struct spdk_nvme_transport_poll_group *tgroup,
3086 : struct spdk_nvme_qpair *qpair)
3087 : {
3088 0 : return 0;
3089 : }
3090 :
3091 : static inline void
3092 0 : nvme_rdma_qpair_process_submits(struct spdk_nvme_qpair *qpair)
3093 : {
3094 0 : struct nvme_rdma_qpair *rqpair = nvme_rdma_qpair(qpair);
3095 :
3096 0 : if (spdk_unlikely(rqpair->state <= NVME_RDMA_QPAIR_STATE_INITIALIZING)) {
3097 0 : return;
3098 : }
3099 :
3100 0 : if (spdk_unlikely(qpair->ctrlr->timeout_enabled)) {
3101 0 : nvme_rdma_qpair_check_timeout(qpair);
3102 : }
3103 :
3104 0 : nvme_rdma_qpair_submit_sends(rqpair);
3105 0 : if (!rqpair->srq) {
3106 0 : nvme_rdma_qpair_submit_recvs(rqpair);
3107 : }
3108 0 : if (rqpair->num_completions > 0) {
3109 0 : nvme_qpair_resubmit_requests(qpair, rqpair->num_completions);
3110 0 : rqpair->num_completions = 0;
3111 : }
3112 : }
3113 :
3114 : static int64_t
3115 0 : nvme_rdma_poll_group_process_completions(struct spdk_nvme_transport_poll_group *tgroup,
3116 : uint32_t completions_per_qpair, spdk_nvme_disconnected_qpair_cb disconnected_qpair_cb)
3117 : {
3118 : struct spdk_nvme_qpair *qpair, *tmp_qpair;
3119 : struct nvme_rdma_qpair *rqpair, *tmp_rqpair;
3120 : struct nvme_rdma_poll_group *group;
3121 : struct nvme_rdma_poller *poller;
3122 0 : int batch_size, rc, rc2 = 0;
3123 0 : int64_t total_completions = 0;
3124 0 : uint64_t completions_allowed = 0;
3125 0 : uint64_t completions_per_poller = 0;
3126 0 : uint64_t poller_completions = 0;
3127 0 : uint64_t rdma_completions;
3128 :
3129 0 : if (completions_per_qpair == 0) {
3130 0 : completions_per_qpair = MAX_COMPLETIONS_PER_POLL;
3131 : }
3132 :
3133 0 : group = nvme_rdma_poll_group(tgroup);
3134 :
3135 0 : STAILQ_FOREACH_SAFE(qpair, &tgroup->disconnected_qpairs, poll_group_stailq, tmp_qpair) {
3136 0 : rc = nvme_rdma_ctrlr_disconnect_qpair_poll(qpair->ctrlr, qpair);
3137 0 : if (rc == 0) {
3138 0 : disconnected_qpair_cb(qpair, tgroup->group->ctx);
3139 : }
3140 : }
3141 :
3142 0 : TAILQ_FOREACH_SAFE(rqpair, &group->connecting_qpairs, link_connecting, tmp_rqpair) {
3143 0 : qpair = &rqpair->qpair;
3144 :
3145 0 : rc = nvme_rdma_ctrlr_connect_qpair_poll(qpair->ctrlr, qpair);
3146 0 : if (rc == 0 || rc != -EAGAIN) {
3147 0 : TAILQ_REMOVE(&group->connecting_qpairs, rqpair, link_connecting);
3148 : /* We use prev pointer to check if qpair is in connecting list or not.
3149 : * TAILQ_REMOVE does not do it. So, we do it manually.
3150 : */
3151 0 : rqpair->link_connecting.tqe_prev = NULL;
3152 :
3153 0 : if (rc == 0) {
3154 : /* Once the connection is completed, we can submit queued requests */
3155 0 : nvme_qpair_resubmit_requests(qpair, rqpair->num_entries);
3156 0 : } else if (rc != -EAGAIN) {
3157 0 : SPDK_ERRLOG("Failed to connect rqpair=%p\n", rqpair);
3158 0 : nvme_rdma_fail_qpair(qpair, 0);
3159 : }
3160 : }
3161 : }
3162 :
3163 0 : STAILQ_FOREACH_SAFE(qpair, &tgroup->connected_qpairs, poll_group_stailq, tmp_qpair) {
3164 0 : rqpair = nvme_rdma_qpair(qpair);
3165 :
3166 0 : if (spdk_likely(nvme_qpair_get_state(qpair) != NVME_QPAIR_CONNECTING)) {
3167 0 : nvme_rdma_qpair_process_cm_event(rqpair);
3168 : }
3169 :
3170 0 : if (spdk_unlikely(qpair->transport_failure_reason != SPDK_NVME_QPAIR_FAILURE_NONE)) {
3171 0 : rc2 = -ENXIO;
3172 0 : nvme_rdma_fail_qpair(qpair, 0);
3173 : }
3174 : }
3175 :
3176 0 : completions_allowed = completions_per_qpair * tgroup->num_connected_qpairs;
3177 0 : if (group->num_pollers) {
3178 0 : completions_per_poller = spdk_max(completions_allowed / group->num_pollers, 1);
3179 : }
3180 :
3181 0 : STAILQ_FOREACH(poller, &group->pollers, link) {
3182 0 : poller_completions = 0;
3183 0 : rdma_completions = 0;
3184 : do {
3185 0 : poller->stats.polls++;
3186 0 : batch_size = spdk_min((completions_per_poller - poller_completions), MAX_COMPLETIONS_PER_POLL);
3187 0 : rc = nvme_rdma_cq_process_completions(poller->cq, batch_size, poller, NULL, &rdma_completions);
3188 0 : if (rc <= 0) {
3189 0 : if (rc == -ECANCELED) {
3190 0 : return -EIO;
3191 0 : } else if (rc == 0) {
3192 0 : poller->stats.idle_polls++;
3193 : }
3194 0 : break;
3195 : }
3196 :
3197 0 : poller_completions += rc;
3198 0 : } while (poller_completions < completions_per_poller);
3199 0 : total_completions += poller_completions;
3200 0 : poller->stats.completions += rdma_completions;
3201 0 : if (poller->srq) {
3202 0 : nvme_rdma_poller_submit_recvs(poller);
3203 : }
3204 : }
3205 :
3206 0 : STAILQ_FOREACH_SAFE(qpair, &tgroup->connected_qpairs, poll_group_stailq, tmp_qpair) {
3207 0 : nvme_rdma_qpair_process_submits(qpair);
3208 : }
3209 :
3210 0 : return rc2 != 0 ? rc2 : total_completions;
3211 : }
3212 :
3213 : static int
3214 1 : nvme_rdma_poll_group_destroy(struct spdk_nvme_transport_poll_group *tgroup)
3215 : {
3216 1 : struct nvme_rdma_poll_group *group = nvme_rdma_poll_group(tgroup);
3217 :
3218 1 : if (!STAILQ_EMPTY(&tgroup->connected_qpairs) || !STAILQ_EMPTY(&tgroup->disconnected_qpairs)) {
3219 0 : return -EBUSY;
3220 : }
3221 :
3222 1 : nvme_rdma_poll_group_free_pollers(group);
3223 1 : free(group);
3224 :
3225 1 : return 0;
3226 : }
3227 :
3228 : static int
3229 3 : nvme_rdma_poll_group_get_stats(struct spdk_nvme_transport_poll_group *tgroup,
3230 : struct spdk_nvme_transport_poll_group_stat **_stats)
3231 : {
3232 : struct nvme_rdma_poll_group *group;
3233 : struct spdk_nvme_transport_poll_group_stat *stats;
3234 : struct spdk_nvme_rdma_device_stat *device_stat;
3235 : struct nvme_rdma_poller *poller;
3236 3 : uint32_t i = 0;
3237 :
3238 3 : if (tgroup == NULL || _stats == NULL) {
3239 2 : SPDK_ERRLOG("Invalid stats or group pointer\n");
3240 2 : return -EINVAL;
3241 : }
3242 :
3243 1 : group = nvme_rdma_poll_group(tgroup);
3244 1 : stats = calloc(1, sizeof(*stats));
3245 1 : if (!stats) {
3246 0 : SPDK_ERRLOG("Can't allocate memory for RDMA stats\n");
3247 0 : return -ENOMEM;
3248 : }
3249 1 : stats->trtype = SPDK_NVME_TRANSPORT_RDMA;
3250 1 : stats->rdma.num_devices = group->num_pollers;
3251 :
3252 1 : if (stats->rdma.num_devices == 0) {
3253 0 : *_stats = stats;
3254 0 : return 0;
3255 : }
3256 :
3257 1 : stats->rdma.device_stats = calloc(stats->rdma.num_devices, sizeof(*stats->rdma.device_stats));
3258 1 : if (!stats->rdma.device_stats) {
3259 0 : SPDK_ERRLOG("Can't allocate memory for RDMA device stats\n");
3260 0 : free(stats);
3261 0 : return -ENOMEM;
3262 : }
3263 :
3264 3 : STAILQ_FOREACH(poller, &group->pollers, link) {
3265 2 : device_stat = &stats->rdma.device_stats[i];
3266 2 : device_stat->name = poller->device->device->name;
3267 2 : device_stat->polls = poller->stats.polls;
3268 2 : device_stat->idle_polls = poller->stats.idle_polls;
3269 2 : device_stat->completions = poller->stats.completions;
3270 2 : device_stat->queued_requests = poller->stats.queued_requests;
3271 2 : device_stat->total_send_wrs = poller->stats.rdma_stats.send.num_submitted_wrs;
3272 2 : device_stat->send_doorbell_updates = poller->stats.rdma_stats.send.doorbell_updates;
3273 2 : device_stat->total_recv_wrs = poller->stats.rdma_stats.recv.num_submitted_wrs;
3274 2 : device_stat->recv_doorbell_updates = poller->stats.rdma_stats.recv.doorbell_updates;
3275 2 : i++;
3276 : }
3277 :
3278 1 : *_stats = stats;
3279 :
3280 1 : return 0;
3281 : }
3282 :
3283 : static void
3284 1 : nvme_rdma_poll_group_free_stats(struct spdk_nvme_transport_poll_group *tgroup,
3285 : struct spdk_nvme_transport_poll_group_stat *stats)
3286 : {
3287 1 : if (stats) {
3288 1 : free(stats->rdma.device_stats);
3289 : }
3290 1 : free(stats);
3291 1 : }
3292 :
3293 : static int
3294 4 : nvme_rdma_ctrlr_get_memory_domains(const struct spdk_nvme_ctrlr *ctrlr,
3295 : struct spdk_memory_domain **domains, int array_size)
3296 : {
3297 4 : struct nvme_rdma_qpair *rqpair = nvme_rdma_qpair(ctrlr->adminq);
3298 :
3299 4 : if (domains && array_size > 0) {
3300 1 : domains[0] = rqpair->memory_domain->domain;
3301 : }
3302 :
3303 4 : return 1;
3304 : }
3305 :
3306 : void
3307 0 : spdk_nvme_rdma_init_hooks(struct spdk_nvme_rdma_hooks *hooks)
3308 : {
3309 0 : g_nvme_hooks = *hooks;
3310 0 : }
3311 :
3312 : const struct spdk_nvme_transport_ops rdma_ops = {
3313 : .name = "RDMA",
3314 : .type = SPDK_NVME_TRANSPORT_RDMA,
3315 : .ctrlr_construct = nvme_rdma_ctrlr_construct,
3316 : .ctrlr_scan = nvme_fabric_ctrlr_scan,
3317 : .ctrlr_destruct = nvme_rdma_ctrlr_destruct,
3318 : .ctrlr_enable = nvme_rdma_ctrlr_enable,
3319 :
3320 : .ctrlr_set_reg_4 = nvme_fabric_ctrlr_set_reg_4,
3321 : .ctrlr_set_reg_8 = nvme_fabric_ctrlr_set_reg_8,
3322 : .ctrlr_get_reg_4 = nvme_fabric_ctrlr_get_reg_4,
3323 : .ctrlr_get_reg_8 = nvme_fabric_ctrlr_get_reg_8,
3324 : .ctrlr_set_reg_4_async = nvme_fabric_ctrlr_set_reg_4_async,
3325 : .ctrlr_set_reg_8_async = nvme_fabric_ctrlr_set_reg_8_async,
3326 : .ctrlr_get_reg_4_async = nvme_fabric_ctrlr_get_reg_4_async,
3327 : .ctrlr_get_reg_8_async = nvme_fabric_ctrlr_get_reg_8_async,
3328 :
3329 : .ctrlr_get_max_xfer_size = nvme_rdma_ctrlr_get_max_xfer_size,
3330 : .ctrlr_get_max_sges = nvme_rdma_ctrlr_get_max_sges,
3331 :
3332 : .ctrlr_create_io_qpair = nvme_rdma_ctrlr_create_io_qpair,
3333 : .ctrlr_delete_io_qpair = nvme_rdma_ctrlr_delete_io_qpair,
3334 : .ctrlr_connect_qpair = nvme_rdma_ctrlr_connect_qpair,
3335 : .ctrlr_disconnect_qpair = nvme_rdma_ctrlr_disconnect_qpair,
3336 :
3337 : .ctrlr_get_memory_domains = nvme_rdma_ctrlr_get_memory_domains,
3338 :
3339 : .qpair_abort_reqs = nvme_rdma_qpair_abort_reqs,
3340 : .qpair_reset = nvme_rdma_qpair_reset,
3341 : .qpair_submit_request = nvme_rdma_qpair_submit_request,
3342 : .qpair_process_completions = nvme_rdma_qpair_process_completions,
3343 : .qpair_iterate_requests = nvme_rdma_qpair_iterate_requests,
3344 : .admin_qpair_abort_aers = nvme_rdma_admin_qpair_abort_aers,
3345 :
3346 : .poll_group_create = nvme_rdma_poll_group_create,
3347 : .poll_group_connect_qpair = nvme_rdma_poll_group_connect_qpair,
3348 : .poll_group_disconnect_qpair = nvme_rdma_poll_group_disconnect_qpair,
3349 : .poll_group_add = nvme_rdma_poll_group_add,
3350 : .poll_group_remove = nvme_rdma_poll_group_remove,
3351 : .poll_group_process_completions = nvme_rdma_poll_group_process_completions,
3352 : .poll_group_destroy = nvme_rdma_poll_group_destroy,
3353 : .poll_group_get_stats = nvme_rdma_poll_group_get_stats,
3354 : .poll_group_free_stats = nvme_rdma_poll_group_free_stats,
3355 : };
3356 :
3357 1 : SPDK_NVME_TRANSPORT_REGISTER(rdma, &rdma_ops);
|