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