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_AUTHENTICATING,
160 : NVME_RDMA_QPAIR_STATE_RUNNING,
161 : NVME_RDMA_QPAIR_STATE_EXITING,
162 : NVME_RDMA_QPAIR_STATE_LINGERING,
163 : NVME_RDMA_QPAIR_STATE_EXITED,
164 : };
165 :
166 : typedef int (*nvme_rdma_cm_event_cb)(struct nvme_rdma_qpair *rqpair, int ret);
167 :
168 : struct nvme_rdma_rsp_opts {
169 : uint16_t num_entries;
170 : struct nvme_rdma_qpair *rqpair;
171 : struct spdk_rdma_provider_srq *srq;
172 : struct spdk_rdma_utils_mem_map *mr_map;
173 : };
174 :
175 : struct nvme_rdma_rsps {
176 : /* Parallel arrays of response buffers + response SGLs of size num_entries */
177 : struct ibv_sge *rsp_sgls;
178 : struct spdk_nvme_rdma_rsp *rsps;
179 :
180 : struct ibv_recv_wr *rsp_recv_wrs;
181 :
182 : /* Count of outstanding recv objects */
183 : uint16_t current_num_recvs;
184 :
185 : uint16_t num_entries;
186 : };
187 :
188 : /* NVMe RDMA qpair extensions for spdk_nvme_qpair */
189 : struct nvme_rdma_qpair {
190 : struct spdk_nvme_qpair qpair;
191 :
192 : struct spdk_rdma_provider_qp *rdma_qp;
193 : struct rdma_cm_id *cm_id;
194 : struct ibv_cq *cq;
195 : struct spdk_rdma_provider_srq *srq;
196 :
197 : struct spdk_nvme_rdma_req *rdma_reqs;
198 :
199 : uint32_t max_send_sge;
200 :
201 : uint16_t num_entries;
202 :
203 : bool delay_cmd_submit;
204 :
205 : uint32_t num_completions;
206 : uint32_t num_outstanding_reqs;
207 :
208 : struct nvme_rdma_rsps *rsps;
209 :
210 : /*
211 : * Array of num_entries NVMe commands registered as RDMA message buffers.
212 : * Indexed by rdma_req->id.
213 : */
214 : struct spdk_nvmf_cmd *cmds;
215 :
216 : struct spdk_rdma_utils_mem_map *mr_map;
217 :
218 : TAILQ_HEAD(, spdk_nvme_rdma_req) free_reqs;
219 : TAILQ_HEAD(, spdk_nvme_rdma_req) outstanding_reqs;
220 :
221 : /* Count of outstanding send objects */
222 : uint16_t current_num_sends;
223 :
224 : TAILQ_ENTRY(nvme_rdma_qpair) link_active;
225 :
226 : /* Placed at the end of the struct since it is not used frequently */
227 : struct rdma_cm_event *evt;
228 : struct nvme_rdma_poller *poller;
229 :
230 : uint64_t evt_timeout_ticks;
231 : nvme_rdma_cm_event_cb evt_cb;
232 : enum rdma_cm_event_type expected_evt_type;
233 :
234 : enum nvme_rdma_qpair_state state;
235 :
236 : bool in_connect_poll;
237 :
238 : uint8_t stale_conn_retry_count;
239 : bool need_destroy;
240 :
241 : TAILQ_ENTRY(nvme_rdma_qpair) link_connecting;
242 : };
243 :
244 : enum NVME_RDMA_COMPLETION_FLAGS {
245 : NVME_RDMA_SEND_COMPLETED = 1u << 0,
246 : NVME_RDMA_RECV_COMPLETED = 1u << 1,
247 : };
248 :
249 : struct spdk_nvme_rdma_req {
250 : uint16_t id;
251 : uint16_t completion_flags: 2;
252 : uint16_t reserved: 14;
253 : /* if completion of RDMA_RECV received before RDMA_SEND, we will complete nvme request
254 : * during processing of RDMA_SEND. To complete the request we must know the response
255 : * received in RDMA_RECV, so store it in this field */
256 : struct spdk_nvme_rdma_rsp *rdma_rsp;
257 :
258 : struct nvme_rdma_wr rdma_wr;
259 :
260 : struct ibv_send_wr send_wr;
261 :
262 : struct nvme_request *req;
263 :
264 : struct ibv_sge send_sgl[NVME_RDMA_DEFAULT_TX_SGE];
265 :
266 : TAILQ_ENTRY(spdk_nvme_rdma_req) link;
267 : };
268 :
269 : struct spdk_nvme_rdma_rsp {
270 : struct spdk_nvme_cpl cpl;
271 : struct nvme_rdma_qpair *rqpair;
272 : struct ibv_recv_wr *recv_wr;
273 : struct nvme_rdma_wr rdma_wr;
274 : };
275 :
276 : struct nvme_rdma_memory_translation_ctx {
277 : void *addr;
278 : size_t length;
279 : uint32_t lkey;
280 : uint32_t rkey;
281 : };
282 :
283 : static const char *rdma_cm_event_str[] = {
284 : "RDMA_CM_EVENT_ADDR_RESOLVED",
285 : "RDMA_CM_EVENT_ADDR_ERROR",
286 : "RDMA_CM_EVENT_ROUTE_RESOLVED",
287 : "RDMA_CM_EVENT_ROUTE_ERROR",
288 : "RDMA_CM_EVENT_CONNECT_REQUEST",
289 : "RDMA_CM_EVENT_CONNECT_RESPONSE",
290 : "RDMA_CM_EVENT_CONNECT_ERROR",
291 : "RDMA_CM_EVENT_UNREACHABLE",
292 : "RDMA_CM_EVENT_REJECTED",
293 : "RDMA_CM_EVENT_ESTABLISHED",
294 : "RDMA_CM_EVENT_DISCONNECTED",
295 : "RDMA_CM_EVENT_DEVICE_REMOVAL",
296 : "RDMA_CM_EVENT_MULTICAST_JOIN",
297 : "RDMA_CM_EVENT_MULTICAST_ERROR",
298 : "RDMA_CM_EVENT_ADDR_CHANGE",
299 : "RDMA_CM_EVENT_TIMEWAIT_EXIT"
300 : };
301 :
302 : static struct nvme_rdma_poller *nvme_rdma_poll_group_get_poller(struct nvme_rdma_poll_group *group,
303 : struct ibv_context *device);
304 : static void nvme_rdma_poll_group_put_poller(struct nvme_rdma_poll_group *group,
305 : struct nvme_rdma_poller *poller);
306 :
307 : static int nvme_rdma_ctrlr_delete_io_qpair(struct spdk_nvme_ctrlr *ctrlr,
308 : struct spdk_nvme_qpair *qpair);
309 :
310 : static inline struct nvme_rdma_qpair *
311 18 : nvme_rdma_qpair(struct spdk_nvme_qpair *qpair)
312 : {
313 18 : assert(qpair->trtype == SPDK_NVME_TRANSPORT_RDMA);
314 18 : return SPDK_CONTAINEROF(qpair, struct nvme_rdma_qpair, qpair);
315 : }
316 :
317 : static inline struct nvme_rdma_poll_group *
318 8 : nvme_rdma_poll_group(struct spdk_nvme_transport_poll_group *group)
319 : {
320 8 : return (SPDK_CONTAINEROF(group, struct nvme_rdma_poll_group, group));
321 : }
322 :
323 : static inline struct nvme_rdma_ctrlr *
324 8 : nvme_rdma_ctrlr(struct spdk_nvme_ctrlr *ctrlr)
325 : {
326 8 : assert(ctrlr->trid.trtype == SPDK_NVME_TRANSPORT_RDMA);
327 8 : return SPDK_CONTAINEROF(ctrlr, struct nvme_rdma_ctrlr, ctrlr);
328 : }
329 :
330 : static inline struct spdk_nvme_rdma_req *
331 3 : nvme_rdma_req_get(struct nvme_rdma_qpair *rqpair)
332 : {
333 : struct spdk_nvme_rdma_req *rdma_req;
334 :
335 3 : rdma_req = TAILQ_FIRST(&rqpair->free_reqs);
336 3 : if (spdk_likely(rdma_req)) {
337 2 : TAILQ_REMOVE(&rqpair->free_reqs, rdma_req, link);
338 : }
339 :
340 3 : return rdma_req;
341 : }
342 :
343 : static inline void
344 1 : nvme_rdma_req_put(struct nvme_rdma_qpair *rqpair, struct spdk_nvme_rdma_req *rdma_req)
345 : {
346 1 : rdma_req->completion_flags = 0;
347 1 : rdma_req->req = NULL;
348 1 : rdma_req->rdma_rsp = NULL;
349 1 : TAILQ_INSERT_HEAD(&rqpair->free_reqs, rdma_req, link);
350 1 : }
351 :
352 : static void
353 0 : nvme_rdma_req_complete(struct spdk_nvme_rdma_req *rdma_req,
354 : struct spdk_nvme_cpl *rsp,
355 : bool print_on_error)
356 : {
357 0 : struct nvme_request *req = rdma_req->req;
358 : struct nvme_rdma_qpair *rqpair;
359 : struct spdk_nvme_qpair *qpair;
360 : bool error, print_error;
361 :
362 0 : assert(req != NULL);
363 :
364 0 : qpair = req->qpair;
365 0 : rqpair = nvme_rdma_qpair(qpair);
366 :
367 0 : error = spdk_nvme_cpl_is_error(rsp);
368 0 : print_error = error && print_on_error && !qpair->ctrlr->opts.disable_error_logging;
369 :
370 0 : if (print_error) {
371 0 : spdk_nvme_qpair_print_command(qpair, &req->cmd);
372 : }
373 :
374 0 : if (print_error || SPDK_DEBUGLOG_FLAG_ENABLED("nvme")) {
375 0 : spdk_nvme_qpair_print_completion(qpair, rsp);
376 : }
377 :
378 0 : assert(rqpair->num_outstanding_reqs > 0);
379 0 : rqpair->num_outstanding_reqs--;
380 :
381 0 : TAILQ_REMOVE(&rqpair->outstanding_reqs, rdma_req, link);
382 :
383 0 : nvme_complete_request(req->cb_fn, req->cb_arg, qpair, req, rsp);
384 0 : nvme_rdma_req_put(rqpair, rdma_req);
385 0 : }
386 :
387 : static const char *
388 4 : nvme_rdma_cm_event_str_get(uint32_t event)
389 : {
390 4 : if (event < SPDK_COUNTOF(rdma_cm_event_str)) {
391 4 : return rdma_cm_event_str[event];
392 : } else {
393 0 : return "Undefined";
394 : }
395 : }
396 :
397 :
398 : static int
399 12 : nvme_rdma_qpair_process_cm_event(struct nvme_rdma_qpair *rqpair)
400 : {
401 12 : struct rdma_cm_event *event = rqpair->evt;
402 : struct spdk_nvmf_rdma_accept_private_data *accept_data;
403 12 : int rc = 0;
404 :
405 12 : if (event) {
406 12 : switch (event->event) {
407 1 : case RDMA_CM_EVENT_ADDR_RESOLVED:
408 : case RDMA_CM_EVENT_ADDR_ERROR:
409 : case RDMA_CM_EVENT_ROUTE_RESOLVED:
410 : case RDMA_CM_EVENT_ROUTE_ERROR:
411 1 : break;
412 1 : case RDMA_CM_EVENT_CONNECT_REQUEST:
413 1 : break;
414 1 : case RDMA_CM_EVENT_CONNECT_ERROR:
415 1 : break;
416 1 : case RDMA_CM_EVENT_UNREACHABLE:
417 : case RDMA_CM_EVENT_REJECTED:
418 1 : break;
419 2 : case RDMA_CM_EVENT_CONNECT_RESPONSE:
420 2 : rc = spdk_rdma_provider_qp_complete_connect(rqpair->rdma_qp);
421 : /* fall through */
422 2 : case RDMA_CM_EVENT_ESTABLISHED:
423 2 : accept_data = (struct spdk_nvmf_rdma_accept_private_data *)event->param.conn.private_data;
424 2 : if (accept_data == NULL) {
425 1 : rc = -1;
426 : } else {
427 1 : SPDK_DEBUGLOG(nvme, "Requested queue depth %d. Target receive queue depth %d.\n",
428 : rqpair->num_entries + 1, accept_data->crqsize);
429 : }
430 2 : break;
431 1 : case RDMA_CM_EVENT_DISCONNECTED:
432 1 : rqpair->qpair.transport_failure_reason = SPDK_NVME_QPAIR_FAILURE_REMOTE;
433 1 : break;
434 1 : case RDMA_CM_EVENT_DEVICE_REMOVAL:
435 1 : rqpair->qpair.transport_failure_reason = SPDK_NVME_QPAIR_FAILURE_LOCAL;
436 1 : rqpair->need_destroy = true;
437 1 : break;
438 1 : case RDMA_CM_EVENT_MULTICAST_JOIN:
439 : case RDMA_CM_EVENT_MULTICAST_ERROR:
440 1 : break;
441 1 : case RDMA_CM_EVENT_ADDR_CHANGE:
442 1 : rqpair->qpair.transport_failure_reason = SPDK_NVME_QPAIR_FAILURE_LOCAL;
443 1 : break;
444 1 : case RDMA_CM_EVENT_TIMEWAIT_EXIT:
445 1 : break;
446 1 : default:
447 1 : SPDK_ERRLOG("Unexpected Acceptor Event [%d]\n", event->event);
448 1 : break;
449 : }
450 12 : rqpair->evt = NULL;
451 12 : rdma_ack_cm_event(event);
452 : }
453 :
454 12 : return rc;
455 : }
456 :
457 : /*
458 : * This function must be called under the nvme controller's lock
459 : * because it touches global controller variables. The lock is taken
460 : * by the generic transport code before invoking a few of the functions
461 : * in this file: nvme_rdma_ctrlr_connect_qpair, nvme_rdma_ctrlr_delete_io_qpair,
462 : * and conditionally nvme_rdma_qpair_process_completions when it is calling
463 : * completions on the admin qpair. When adding a new call to this function, please
464 : * verify that it is in a situation where it falls under the lock.
465 : */
466 : static int
467 0 : nvme_rdma_poll_events(struct nvme_rdma_ctrlr *rctrlr)
468 : {
469 : struct nvme_rdma_cm_event_entry *entry, *tmp;
470 : struct nvme_rdma_qpair *event_qpair;
471 : struct rdma_cm_event *event;
472 0 : struct rdma_event_channel *channel = rctrlr->cm_channel;
473 :
474 0 : STAILQ_FOREACH_SAFE(entry, &rctrlr->pending_cm_events, link, tmp) {
475 0 : event_qpair = entry->evt->id->context;
476 0 : if (event_qpair->evt == NULL) {
477 0 : event_qpair->evt = entry->evt;
478 0 : STAILQ_REMOVE(&rctrlr->pending_cm_events, entry, nvme_rdma_cm_event_entry, link);
479 0 : STAILQ_INSERT_HEAD(&rctrlr->free_cm_events, entry, link);
480 : }
481 : }
482 :
483 0 : while (rdma_get_cm_event(channel, &event) == 0) {
484 0 : event_qpair = event->id->context;
485 0 : if (event_qpair->evt == NULL) {
486 0 : event_qpair->evt = event;
487 : } else {
488 0 : assert(rctrlr == nvme_rdma_ctrlr(event_qpair->qpair.ctrlr));
489 0 : entry = STAILQ_FIRST(&rctrlr->free_cm_events);
490 0 : if (entry == NULL) {
491 0 : rdma_ack_cm_event(event);
492 0 : return -ENOMEM;
493 : }
494 0 : STAILQ_REMOVE(&rctrlr->free_cm_events, entry, nvme_rdma_cm_event_entry, link);
495 0 : entry->evt = event;
496 0 : STAILQ_INSERT_TAIL(&rctrlr->pending_cm_events, entry, link);
497 : }
498 : }
499 :
500 : /* rdma_get_cm_event() returns -1 on error. If an error occurs, errno
501 : * will be set to indicate the failure reason. So return negated errno here.
502 : */
503 0 : return -errno;
504 : }
505 :
506 : static int
507 4 : nvme_rdma_validate_cm_event(enum rdma_cm_event_type expected_evt_type,
508 : struct rdma_cm_event *reaped_evt)
509 : {
510 4 : int rc = -EBADMSG;
511 :
512 4 : if (expected_evt_type == reaped_evt->event) {
513 1 : return 0;
514 : }
515 :
516 3 : switch (expected_evt_type) {
517 2 : case RDMA_CM_EVENT_ESTABLISHED:
518 : /*
519 : * There is an enum ib_cm_rej_reason in the kernel headers that sets 10 as
520 : * IB_CM_REJ_STALE_CONN. I can't find the corresponding userspace but we get
521 : * the same values here.
522 : */
523 2 : if (reaped_evt->event == RDMA_CM_EVENT_REJECTED && reaped_evt->status == 10) {
524 1 : rc = -ESTALE;
525 1 : } else if (reaped_evt->event == RDMA_CM_EVENT_CONNECT_RESPONSE) {
526 : /*
527 : * If we are using a qpair which is not created using rdma cm API
528 : * then we will receive RDMA_CM_EVENT_CONNECT_RESPONSE instead of
529 : * RDMA_CM_EVENT_ESTABLISHED.
530 : */
531 1 : return 0;
532 : }
533 1 : break;
534 1 : default:
535 1 : break;
536 : }
537 :
538 2 : SPDK_ERRLOG("Expected %s but received %s (%d) from CM event channel (status = %d)\n",
539 : nvme_rdma_cm_event_str_get(expected_evt_type),
540 : nvme_rdma_cm_event_str_get(reaped_evt->event), reaped_evt->event,
541 : reaped_evt->status);
542 2 : return rc;
543 : }
544 :
545 : static int
546 0 : nvme_rdma_process_event_start(struct nvme_rdma_qpair *rqpair,
547 : enum rdma_cm_event_type evt,
548 : nvme_rdma_cm_event_cb evt_cb)
549 : {
550 : int rc;
551 :
552 0 : assert(evt_cb != NULL);
553 :
554 0 : if (rqpair->evt != NULL) {
555 0 : rc = nvme_rdma_qpair_process_cm_event(rqpair);
556 0 : if (rc) {
557 0 : return rc;
558 : }
559 : }
560 :
561 0 : rqpair->expected_evt_type = evt;
562 0 : rqpair->evt_cb = evt_cb;
563 0 : rqpair->evt_timeout_ticks = (g_spdk_nvme_transport_opts.rdma_cm_event_timeout_ms * 1000 *
564 0 : spdk_get_ticks_hz()) / SPDK_SEC_TO_USEC + spdk_get_ticks();
565 :
566 0 : return 0;
567 : }
568 :
569 : static int
570 0 : nvme_rdma_process_event_poll(struct nvme_rdma_qpair *rqpair)
571 : {
572 : struct nvme_rdma_ctrlr *rctrlr;
573 0 : int rc = 0, rc2;
574 :
575 0 : rctrlr = nvme_rdma_ctrlr(rqpair->qpair.ctrlr);
576 0 : assert(rctrlr != NULL);
577 :
578 0 : if (!rqpair->evt && spdk_get_ticks() < rqpair->evt_timeout_ticks) {
579 0 : rc = nvme_rdma_poll_events(rctrlr);
580 0 : if (rc == -EAGAIN || rc == -EWOULDBLOCK) {
581 0 : return rc;
582 : }
583 : }
584 :
585 0 : if (rqpair->evt == NULL) {
586 0 : rc = -EADDRNOTAVAIL;
587 0 : goto exit;
588 : }
589 :
590 0 : rc = nvme_rdma_validate_cm_event(rqpair->expected_evt_type, rqpair->evt);
591 :
592 0 : rc2 = nvme_rdma_qpair_process_cm_event(rqpair);
593 : /* bad message takes precedence over the other error codes from processing the event. */
594 0 : rc = rc == 0 ? rc2 : rc;
595 :
596 0 : exit:
597 0 : assert(rqpair->evt_cb != NULL);
598 0 : return rqpair->evt_cb(rqpair, rc);
599 : }
600 :
601 : static int
602 3 : nvme_rdma_resize_cq(struct nvme_rdma_qpair *rqpair, struct nvme_rdma_poller *poller)
603 : {
604 : int current_num_wc, required_num_wc;
605 : int max_cq_size;
606 :
607 3 : required_num_wc = poller->required_num_wc + WC_PER_QPAIR(rqpair->num_entries);
608 3 : current_num_wc = poller->current_num_wc;
609 3 : if (current_num_wc < required_num_wc) {
610 2 : current_num_wc = spdk_max(current_num_wc * 2, required_num_wc);
611 : }
612 :
613 3 : max_cq_size = g_spdk_nvme_transport_opts.rdma_max_cq_size;
614 3 : if (max_cq_size != 0 && current_num_wc > max_cq_size) {
615 0 : current_num_wc = max_cq_size;
616 : }
617 :
618 3 : if (poller->current_num_wc != current_num_wc) {
619 2 : SPDK_DEBUGLOG(nvme, "Resize RDMA CQ from %d to %d\n", poller->current_num_wc,
620 : current_num_wc);
621 2 : if (ibv_resize_cq(poller->cq, current_num_wc)) {
622 1 : SPDK_ERRLOG("RDMA CQ resize failed: errno %d: %s\n", errno, spdk_strerror(errno));
623 1 : return -1;
624 : }
625 :
626 1 : poller->current_num_wc = current_num_wc;
627 : }
628 :
629 2 : poller->required_num_wc = required_num_wc;
630 2 : return 0;
631 : }
632 :
633 : static int
634 5 : nvme_rdma_qpair_set_poller(struct spdk_nvme_qpair *qpair)
635 : {
636 5 : struct nvme_rdma_qpair *rqpair = nvme_rdma_qpair(qpair);
637 5 : struct nvme_rdma_poll_group *group = nvme_rdma_poll_group(qpair->poll_group);
638 : struct nvme_rdma_poller *poller;
639 :
640 5 : assert(rqpair->cq == NULL);
641 :
642 5 : poller = nvme_rdma_poll_group_get_poller(group, rqpair->cm_id->verbs);
643 5 : if (!poller) {
644 2 : SPDK_ERRLOG("Unable to find a cq for qpair %p on poll group %p\n", qpair, qpair->poll_group);
645 2 : return -EINVAL;
646 : }
647 :
648 3 : if (!poller->srq) {
649 3 : if (nvme_rdma_resize_cq(rqpair, poller)) {
650 1 : nvme_rdma_poll_group_put_poller(group, poller);
651 1 : return -EPROTO;
652 : }
653 : }
654 :
655 2 : rqpair->cq = poller->cq;
656 2 : rqpair->srq = poller->srq;
657 2 : if (rqpair->srq) {
658 0 : rqpair->rsps = poller->rsps;
659 : }
660 2 : rqpair->poller = poller;
661 2 : return 0;
662 : }
663 :
664 : static int
665 1 : nvme_rdma_qpair_init(struct nvme_rdma_qpair *rqpair)
666 : {
667 : int rc;
668 1 : struct spdk_rdma_provider_qp_init_attr attr = {};
669 : struct ibv_device_attr dev_attr;
670 : struct nvme_rdma_ctrlr *rctrlr;
671 : uint32_t num_cqe, max_num_cqe;
672 :
673 1 : rc = ibv_query_device(rqpair->cm_id->verbs, &dev_attr);
674 1 : if (rc != 0) {
675 0 : SPDK_ERRLOG("Failed to query RDMA device attributes.\n");
676 0 : return -1;
677 : }
678 :
679 1 : if (rqpair->qpair.poll_group) {
680 0 : assert(!rqpair->cq);
681 0 : rc = nvme_rdma_qpair_set_poller(&rqpair->qpair);
682 0 : if (rc) {
683 0 : SPDK_ERRLOG("Unable to activate the rdmaqpair.\n");
684 0 : return -1;
685 : }
686 0 : assert(rqpair->cq);
687 : } else {
688 1 : num_cqe = rqpair->num_entries * 2;
689 1 : max_num_cqe = g_spdk_nvme_transport_opts.rdma_max_cq_size;
690 1 : if (max_num_cqe != 0 && num_cqe > max_num_cqe) {
691 0 : num_cqe = max_num_cqe;
692 : }
693 1 : rqpair->cq = ibv_create_cq(rqpair->cm_id->verbs, num_cqe, rqpair, NULL, 0);
694 1 : if (!rqpair->cq) {
695 0 : SPDK_ERRLOG("Unable to create completion queue: errno %d: %s\n", errno, spdk_strerror(errno));
696 0 : return -1;
697 : }
698 : }
699 :
700 1 : rctrlr = nvme_rdma_ctrlr(rqpair->qpair.ctrlr);
701 1 : if (g_nvme_hooks.get_ibv_pd) {
702 0 : attr.pd = g_nvme_hooks.get_ibv_pd(&rctrlr->ctrlr.trid, rqpair->cm_id->verbs);
703 : } else {
704 1 : attr.pd = spdk_rdma_utils_get_pd(rqpair->cm_id->verbs);
705 : }
706 :
707 1 : attr.stats = rqpair->poller ? &rqpair->poller->stats.rdma_stats : NULL;
708 1 : attr.send_cq = rqpair->cq;
709 1 : attr.recv_cq = rqpair->cq;
710 1 : attr.cap.max_send_wr = rqpair->num_entries; /* SEND operations */
711 1 : if (rqpair->srq) {
712 0 : attr.srq = rqpair->srq->srq;
713 : } else {
714 1 : attr.cap.max_recv_wr = rqpair->num_entries; /* RECV operations */
715 : }
716 1 : attr.cap.max_send_sge = spdk_min(NVME_RDMA_DEFAULT_TX_SGE, dev_attr.max_sge);
717 1 : attr.cap.max_recv_sge = spdk_min(NVME_RDMA_DEFAULT_RX_SGE, dev_attr.max_sge);
718 :
719 1 : rqpair->rdma_qp = spdk_rdma_provider_qp_create(rqpair->cm_id, &attr);
720 :
721 1 : if (!rqpair->rdma_qp) {
722 0 : return -1;
723 : }
724 :
725 : /* ibv_create_qp will change the values in attr.cap. Make sure we store the proper value. */
726 1 : rqpair->max_send_sge = spdk_min(NVME_RDMA_DEFAULT_TX_SGE, attr.cap.max_send_sge);
727 1 : rqpair->current_num_sends = 0;
728 :
729 1 : rqpair->cm_id->context = rqpair;
730 :
731 1 : return 0;
732 : }
733 :
734 : static void
735 0 : nvme_rdma_reset_failed_sends(struct nvme_rdma_qpair *rqpair,
736 : struct ibv_send_wr *bad_send_wr, int rc)
737 : {
738 0 : SPDK_ERRLOG("Failed to post WRs on send queue, errno %d (%s), bad_wr %p\n",
739 : rc, spdk_strerror(rc), bad_send_wr);
740 0 : while (bad_send_wr != NULL) {
741 0 : assert(rqpair->current_num_sends > 0);
742 0 : rqpair->current_num_sends--;
743 0 : bad_send_wr = bad_send_wr->next;
744 : }
745 0 : }
746 :
747 : static void
748 0 : nvme_rdma_reset_failed_recvs(struct nvme_rdma_rsps *rsps,
749 : struct ibv_recv_wr *bad_recv_wr, int rc)
750 : {
751 0 : SPDK_ERRLOG("Failed to post WRs on receive queue, errno %d (%s), bad_wr %p\n",
752 : rc, spdk_strerror(rc), bad_recv_wr);
753 0 : while (bad_recv_wr != NULL) {
754 0 : assert(rsps->current_num_recvs > 0);
755 0 : rsps->current_num_recvs--;
756 0 : bad_recv_wr = bad_recv_wr->next;
757 : }
758 0 : }
759 :
760 : static inline int
761 1 : nvme_rdma_qpair_submit_sends(struct nvme_rdma_qpair *rqpair)
762 : {
763 1 : struct ibv_send_wr *bad_send_wr = NULL;
764 : int rc;
765 :
766 1 : rc = spdk_rdma_provider_qp_flush_send_wrs(rqpair->rdma_qp, &bad_send_wr);
767 :
768 1 : if (spdk_unlikely(rc)) {
769 0 : nvme_rdma_reset_failed_sends(rqpair, bad_send_wr, rc);
770 : }
771 :
772 1 : return rc;
773 : }
774 :
775 : static inline int
776 0 : nvme_rdma_qpair_submit_recvs(struct nvme_rdma_qpair *rqpair)
777 : {
778 : struct ibv_recv_wr *bad_recv_wr;
779 0 : int rc = 0;
780 :
781 0 : rc = spdk_rdma_provider_qp_flush_recv_wrs(rqpair->rdma_qp, &bad_recv_wr);
782 0 : if (spdk_unlikely(rc)) {
783 0 : nvme_rdma_reset_failed_recvs(rqpair->rsps, bad_recv_wr, rc);
784 : }
785 :
786 0 : return rc;
787 : }
788 :
789 : static inline int
790 0 : nvme_rdma_poller_submit_recvs(struct nvme_rdma_poller *poller)
791 : {
792 : struct ibv_recv_wr *bad_recv_wr;
793 : int rc;
794 :
795 0 : rc = spdk_rdma_provider_srq_flush_recv_wrs(poller->srq, &bad_recv_wr);
796 0 : if (spdk_unlikely(rc)) {
797 0 : nvme_rdma_reset_failed_recvs(poller->rsps, bad_recv_wr, rc);
798 : }
799 :
800 0 : return rc;
801 : }
802 :
803 : #define nvme_rdma_trace_ibv_sge(sg_list) \
804 : if (sg_list) { \
805 : SPDK_DEBUGLOG(nvme, "local addr %p length 0x%x lkey 0x%x\n", \
806 : (void *)(sg_list)->addr, (sg_list)->length, (sg_list)->lkey); \
807 : }
808 :
809 : static void
810 3 : nvme_rdma_free_rsps(struct nvme_rdma_rsps *rsps)
811 : {
812 3 : if (!rsps) {
813 1 : return;
814 : }
815 :
816 2 : spdk_free(rsps->rsps);
817 2 : spdk_free(rsps->rsp_sgls);
818 2 : spdk_free(rsps->rsp_recv_wrs);
819 2 : spdk_free(rsps);
820 : }
821 :
822 : static struct nvme_rdma_rsps *
823 2 : nvme_rdma_create_rsps(struct nvme_rdma_rsp_opts *opts)
824 : {
825 : struct nvme_rdma_rsps *rsps;
826 : struct spdk_rdma_utils_memory_translation translation;
827 : uint16_t i;
828 : int rc;
829 :
830 2 : rsps = spdk_zmalloc(sizeof(*rsps), 0, NULL, SPDK_ENV_NUMA_ID_ANY, SPDK_MALLOC_DMA);
831 2 : if (!rsps) {
832 0 : SPDK_ERRLOG("Failed to allocate rsps object\n");
833 0 : return NULL;
834 : }
835 :
836 2 : rsps->rsp_sgls = spdk_zmalloc(opts->num_entries * sizeof(*rsps->rsp_sgls), 0, NULL,
837 : SPDK_ENV_NUMA_ID_ANY, SPDK_MALLOC_DMA);
838 2 : if (!rsps->rsp_sgls) {
839 1 : SPDK_ERRLOG("Failed to allocate rsp_sgls\n");
840 1 : goto fail;
841 : }
842 :
843 1 : rsps->rsp_recv_wrs = spdk_zmalloc(opts->num_entries * sizeof(*rsps->rsp_recv_wrs), 0, NULL,
844 : SPDK_ENV_NUMA_ID_ANY, SPDK_MALLOC_DMA);
845 1 : if (!rsps->rsp_recv_wrs) {
846 0 : SPDK_ERRLOG("Failed to allocate rsp_recv_wrs\n");
847 0 : goto fail;
848 : }
849 :
850 1 : rsps->rsps = spdk_zmalloc(opts->num_entries * sizeof(*rsps->rsps), 0, NULL,
851 : SPDK_ENV_NUMA_ID_ANY, SPDK_MALLOC_DMA);
852 1 : if (!rsps->rsps) {
853 0 : SPDK_ERRLOG("can not allocate rdma rsps\n");
854 0 : goto fail;
855 : }
856 :
857 2 : for (i = 0; i < opts->num_entries; i++) {
858 1 : struct ibv_sge *rsp_sgl = &rsps->rsp_sgls[i];
859 1 : struct spdk_nvme_rdma_rsp *rsp = &rsps->rsps[i];
860 1 : struct ibv_recv_wr *recv_wr = &rsps->rsp_recv_wrs[i];
861 :
862 1 : rsp->rqpair = opts->rqpair;
863 1 : rsp->rdma_wr.type = RDMA_WR_TYPE_RECV;
864 1 : rsp->recv_wr = recv_wr;
865 1 : rsp_sgl->addr = (uint64_t)rsp;
866 1 : rsp_sgl->length = sizeof(struct spdk_nvme_cpl);
867 1 : rc = spdk_rdma_utils_get_translation(opts->mr_map, rsp, sizeof(*rsp), &translation);
868 1 : if (rc) {
869 0 : goto fail;
870 : }
871 1 : rsp_sgl->lkey = spdk_rdma_utils_memory_translation_get_lkey(&translation);
872 :
873 1 : recv_wr->wr_id = (uint64_t)&rsp->rdma_wr;
874 1 : recv_wr->next = NULL;
875 1 : recv_wr->sg_list = rsp_sgl;
876 1 : recv_wr->num_sge = 1;
877 :
878 1 : nvme_rdma_trace_ibv_sge(recv_wr->sg_list);
879 :
880 1 : if (opts->rqpair) {
881 1 : spdk_rdma_provider_qp_queue_recv_wrs(opts->rqpair->rdma_qp, recv_wr);
882 : } else {
883 0 : spdk_rdma_provider_srq_queue_recv_wrs(opts->srq, recv_wr);
884 : }
885 : }
886 :
887 1 : rsps->num_entries = opts->num_entries;
888 1 : rsps->current_num_recvs = opts->num_entries;
889 :
890 1 : return rsps;
891 1 : fail:
892 1 : nvme_rdma_free_rsps(rsps);
893 1 : return NULL;
894 : }
895 :
896 : static void
897 3 : nvme_rdma_free_reqs(struct nvme_rdma_qpair *rqpair)
898 : {
899 3 : if (!rqpair->rdma_reqs) {
900 2 : return;
901 : }
902 :
903 1 : spdk_free(rqpair->cmds);
904 1 : rqpair->cmds = NULL;
905 :
906 1 : spdk_free(rqpair->rdma_reqs);
907 1 : rqpair->rdma_reqs = NULL;
908 : }
909 :
910 : static int
911 4 : nvme_rdma_create_reqs(struct nvme_rdma_qpair *rqpair)
912 : {
913 : struct spdk_rdma_utils_memory_translation translation;
914 : uint16_t i;
915 : int rc;
916 :
917 4 : assert(!rqpair->rdma_reqs);
918 4 : rqpair->rdma_reqs = spdk_zmalloc(rqpair->num_entries * sizeof(struct spdk_nvme_rdma_req), 0, NULL,
919 : SPDK_ENV_NUMA_ID_ANY, SPDK_MALLOC_DMA);
920 4 : if (rqpair->rdma_reqs == NULL) {
921 1 : SPDK_ERRLOG("Failed to allocate rdma_reqs\n");
922 1 : goto fail;
923 : }
924 :
925 3 : assert(!rqpair->cmds);
926 3 : rqpair->cmds = spdk_zmalloc(rqpair->num_entries * sizeof(*rqpair->cmds), 0, NULL,
927 : SPDK_ENV_NUMA_ID_ANY, SPDK_MALLOC_DMA);
928 3 : if (!rqpair->cmds) {
929 0 : SPDK_ERRLOG("Failed to allocate RDMA cmds\n");
930 0 : goto fail;
931 : }
932 :
933 3 : TAILQ_INIT(&rqpair->free_reqs);
934 3 : TAILQ_INIT(&rqpair->outstanding_reqs);
935 10 : for (i = 0; i < rqpair->num_entries; i++) {
936 : struct spdk_nvme_rdma_req *rdma_req;
937 : struct spdk_nvmf_cmd *cmd;
938 :
939 7 : rdma_req = &rqpair->rdma_reqs[i];
940 7 : rdma_req->rdma_wr.type = RDMA_WR_TYPE_SEND;
941 7 : cmd = &rqpair->cmds[i];
942 :
943 7 : rdma_req->id = i;
944 :
945 7 : rc = spdk_rdma_utils_get_translation(rqpair->mr_map, cmd, sizeof(*cmd), &translation);
946 7 : if (rc) {
947 0 : goto fail;
948 : }
949 7 : rdma_req->send_sgl[0].lkey = spdk_rdma_utils_memory_translation_get_lkey(&translation);
950 :
951 : /* The first RDMA sgl element will always point
952 : * at this data structure. Depending on whether
953 : * an NVMe-oF SGL is required, the length of
954 : * this element may change. */
955 7 : rdma_req->send_sgl[0].addr = (uint64_t)cmd;
956 7 : rdma_req->send_wr.wr_id = (uint64_t)&rdma_req->rdma_wr;
957 7 : rdma_req->send_wr.next = NULL;
958 7 : rdma_req->send_wr.opcode = IBV_WR_SEND;
959 7 : rdma_req->send_wr.send_flags = IBV_SEND_SIGNALED;
960 7 : rdma_req->send_wr.sg_list = rdma_req->send_sgl;
961 7 : rdma_req->send_wr.imm_data = 0;
962 :
963 7 : TAILQ_INSERT_TAIL(&rqpair->free_reqs, rdma_req, link);
964 : }
965 :
966 3 : return 0;
967 1 : fail:
968 1 : nvme_rdma_free_reqs(rqpair);
969 1 : return -ENOMEM;
970 : }
971 :
972 : static int nvme_rdma_connect(struct nvme_rdma_qpair *rqpair);
973 :
974 : static int
975 0 : nvme_rdma_route_resolved(struct nvme_rdma_qpair *rqpair, int ret)
976 : {
977 0 : if (ret) {
978 0 : SPDK_ERRLOG("RDMA route resolution error\n");
979 0 : return -1;
980 : }
981 :
982 0 : ret = nvme_rdma_qpair_init(rqpair);
983 0 : if (ret < 0) {
984 0 : SPDK_ERRLOG("nvme_rdma_qpair_init() failed\n");
985 0 : return -1;
986 : }
987 :
988 0 : return nvme_rdma_connect(rqpair);
989 : }
990 :
991 : static int
992 0 : nvme_rdma_addr_resolved(struct nvme_rdma_qpair *rqpair, int ret)
993 : {
994 0 : if (ret) {
995 0 : SPDK_ERRLOG("RDMA address resolution error\n");
996 0 : return -1;
997 : }
998 :
999 0 : if (rqpair->qpair.ctrlr->opts.transport_ack_timeout != SPDK_NVME_TRANSPORT_ACK_TIMEOUT_DISABLED) {
1000 : #ifdef SPDK_CONFIG_RDMA_SET_ACK_TIMEOUT
1001 0 : uint8_t timeout = rqpair->qpair.ctrlr->opts.transport_ack_timeout;
1002 0 : ret = rdma_set_option(rqpair->cm_id, RDMA_OPTION_ID,
1003 : RDMA_OPTION_ID_ACK_TIMEOUT,
1004 : &timeout, sizeof(timeout));
1005 0 : if (ret) {
1006 0 : SPDK_NOTICELOG("Can't apply RDMA_OPTION_ID_ACK_TIMEOUT %d, ret %d\n", timeout, ret);
1007 : }
1008 : #else
1009 : SPDK_DEBUGLOG(nvme, "transport_ack_timeout is not supported\n");
1010 : #endif
1011 : }
1012 :
1013 0 : if (rqpair->qpair.ctrlr->opts.transport_tos != SPDK_NVME_TRANSPORT_TOS_DISABLED) {
1014 : #ifdef SPDK_CONFIG_RDMA_SET_TOS
1015 0 : uint8_t tos = rqpair->qpair.ctrlr->opts.transport_tos;
1016 0 : ret = rdma_set_option(rqpair->cm_id, RDMA_OPTION_ID, RDMA_OPTION_ID_TOS, &tos, sizeof(tos));
1017 0 : if (ret) {
1018 0 : SPDK_NOTICELOG("Can't apply RDMA_OPTION_ID_TOS %u, ret %d\n", tos, ret);
1019 : }
1020 : #else
1021 : SPDK_DEBUGLOG(nvme, "transport_tos is not supported\n");
1022 : #endif
1023 : }
1024 :
1025 0 : ret = rdma_resolve_route(rqpair->cm_id, NVME_RDMA_TIME_OUT_IN_MS);
1026 0 : if (ret) {
1027 0 : SPDK_ERRLOG("rdma_resolve_route\n");
1028 0 : return ret;
1029 : }
1030 :
1031 0 : return nvme_rdma_process_event_start(rqpair, RDMA_CM_EVENT_ROUTE_RESOLVED,
1032 : nvme_rdma_route_resolved);
1033 : }
1034 :
1035 : static int
1036 0 : nvme_rdma_resolve_addr(struct nvme_rdma_qpair *rqpair,
1037 : struct sockaddr *src_addr,
1038 : struct sockaddr *dst_addr)
1039 : {
1040 : int ret;
1041 :
1042 0 : if (src_addr) {
1043 0 : int reuse = 1;
1044 :
1045 0 : ret = rdma_set_option(rqpair->cm_id, RDMA_OPTION_ID, RDMA_OPTION_ID_REUSEADDR,
1046 : &reuse, sizeof(reuse));
1047 0 : if (ret) {
1048 0 : SPDK_NOTICELOG("Can't apply RDMA_OPTION_ID_REUSEADDR %d, ret %d\n",
1049 : reuse, ret);
1050 : /* It is likely that rdma_resolve_addr() returns -EADDRINUSE, but
1051 : * we may missing something. We rely on rdma_resolve_addr().
1052 : */
1053 : }
1054 : }
1055 :
1056 0 : ret = rdma_resolve_addr(rqpair->cm_id, src_addr, dst_addr,
1057 : NVME_RDMA_TIME_OUT_IN_MS);
1058 0 : if (ret) {
1059 0 : SPDK_ERRLOG("rdma_resolve_addr, %d\n", errno);
1060 0 : return ret;
1061 : }
1062 :
1063 0 : return nvme_rdma_process_event_start(rqpair, RDMA_CM_EVENT_ADDR_RESOLVED,
1064 : nvme_rdma_addr_resolved);
1065 : }
1066 :
1067 : static int nvme_rdma_stale_conn_retry(struct nvme_rdma_qpair *rqpair);
1068 :
1069 : static int
1070 0 : nvme_rdma_connect_established(struct nvme_rdma_qpair *rqpair, int ret)
1071 : {
1072 0 : struct nvme_rdma_rsp_opts opts = {};
1073 :
1074 0 : if (ret == -ESTALE) {
1075 0 : return nvme_rdma_stale_conn_retry(rqpair);
1076 0 : } else if (ret) {
1077 0 : SPDK_ERRLOG("RDMA connect error %d\n", ret);
1078 0 : return ret;
1079 : }
1080 :
1081 0 : assert(!rqpair->mr_map);
1082 0 : rqpair->mr_map = spdk_rdma_utils_create_mem_map(rqpair->rdma_qp->qp->pd, &g_nvme_hooks,
1083 : IBV_ACCESS_LOCAL_WRITE | IBV_ACCESS_REMOTE_READ | IBV_ACCESS_REMOTE_WRITE);
1084 0 : if (!rqpair->mr_map) {
1085 0 : SPDK_ERRLOG("Unable to register RDMA memory translation map\n");
1086 0 : return -1;
1087 : }
1088 :
1089 0 : ret = nvme_rdma_create_reqs(rqpair);
1090 0 : SPDK_DEBUGLOG(nvme, "rc =%d\n", ret);
1091 0 : if (ret) {
1092 0 : SPDK_ERRLOG("Unable to create rqpair RDMA requests\n");
1093 0 : return -1;
1094 : }
1095 0 : SPDK_DEBUGLOG(nvme, "RDMA requests created\n");
1096 :
1097 0 : if (!rqpair->srq) {
1098 0 : opts.num_entries = rqpair->num_entries;
1099 0 : opts.rqpair = rqpair;
1100 0 : opts.srq = NULL;
1101 0 : opts.mr_map = rqpair->mr_map;
1102 :
1103 0 : assert(!rqpair->rsps);
1104 0 : rqpair->rsps = nvme_rdma_create_rsps(&opts);
1105 0 : if (!rqpair->rsps) {
1106 0 : SPDK_ERRLOG("Unable to create rqpair RDMA responses\n");
1107 0 : return -1;
1108 : }
1109 0 : SPDK_DEBUGLOG(nvme, "RDMA responses created\n");
1110 :
1111 0 : ret = nvme_rdma_qpair_submit_recvs(rqpair);
1112 0 : SPDK_DEBUGLOG(nvme, "rc =%d\n", ret);
1113 0 : if (ret) {
1114 0 : SPDK_ERRLOG("Unable to submit rqpair RDMA responses\n");
1115 0 : return -1;
1116 : }
1117 0 : SPDK_DEBUGLOG(nvme, "RDMA responses submitted\n");
1118 : }
1119 :
1120 0 : rqpair->state = NVME_RDMA_QPAIR_STATE_FABRIC_CONNECT_SEND;
1121 :
1122 0 : return 0;
1123 : }
1124 :
1125 : static int
1126 0 : nvme_rdma_connect(struct nvme_rdma_qpair *rqpair)
1127 : {
1128 0 : struct rdma_conn_param param = {};
1129 0 : struct spdk_nvmf_rdma_request_private_data request_data = {};
1130 : struct ibv_device_attr attr;
1131 : int ret;
1132 : struct spdk_nvme_ctrlr *ctrlr;
1133 :
1134 0 : ret = ibv_query_device(rqpair->cm_id->verbs, &attr);
1135 0 : if (ret != 0) {
1136 0 : SPDK_ERRLOG("Failed to query RDMA device attributes.\n");
1137 0 : return ret;
1138 : }
1139 :
1140 0 : param.responder_resources = attr.max_qp_rd_atom;
1141 :
1142 0 : ctrlr = rqpair->qpair.ctrlr;
1143 0 : if (!ctrlr) {
1144 0 : return -1;
1145 : }
1146 :
1147 0 : request_data.qid = rqpair->qpair.id;
1148 0 : request_data.hrqsize = rqpair->num_entries + 1;
1149 0 : request_data.hsqsize = rqpair->num_entries;
1150 0 : request_data.cntlid = ctrlr->cntlid;
1151 :
1152 0 : param.private_data = &request_data;
1153 0 : param.private_data_len = sizeof(request_data);
1154 0 : param.retry_count = ctrlr->opts.transport_retry_count;
1155 0 : param.rnr_retry_count = 7;
1156 :
1157 : /* Fields below are ignored by rdma cm if qpair has been
1158 : * created using rdma cm API. */
1159 0 : param.srq = 0;
1160 0 : param.qp_num = rqpair->rdma_qp->qp->qp_num;
1161 :
1162 0 : ret = rdma_connect(rqpair->cm_id, ¶m);
1163 0 : if (ret) {
1164 0 : SPDK_ERRLOG("nvme rdma connect error\n");
1165 0 : return ret;
1166 : }
1167 :
1168 0 : ctrlr->numa.id_valid = 1;
1169 0 : ctrlr->numa.id = spdk_rdma_cm_id_get_numa_id(rqpair->cm_id);
1170 :
1171 0 : return nvme_rdma_process_event_start(rqpair, RDMA_CM_EVENT_ESTABLISHED,
1172 : nvme_rdma_connect_established);
1173 : }
1174 :
1175 : static int
1176 0 : nvme_rdma_ctrlr_connect_qpair(struct spdk_nvme_ctrlr *ctrlr, struct spdk_nvme_qpair *qpair)
1177 : {
1178 : struct sockaddr_storage dst_addr;
1179 : struct sockaddr_storage src_addr;
1180 : bool src_addr_specified;
1181 0 : long int port, src_port = 0;
1182 : int rc;
1183 : struct nvme_rdma_ctrlr *rctrlr;
1184 : struct nvme_rdma_qpair *rqpair;
1185 : struct nvme_rdma_poll_group *group;
1186 : int family;
1187 :
1188 0 : rqpair = nvme_rdma_qpair(qpair);
1189 0 : rctrlr = nvme_rdma_ctrlr(ctrlr);
1190 0 : assert(rctrlr != NULL);
1191 :
1192 0 : switch (ctrlr->trid.adrfam) {
1193 0 : case SPDK_NVMF_ADRFAM_IPV4:
1194 0 : family = AF_INET;
1195 0 : break;
1196 0 : case SPDK_NVMF_ADRFAM_IPV6:
1197 0 : family = AF_INET6;
1198 0 : break;
1199 0 : default:
1200 0 : SPDK_ERRLOG("Unhandled ADRFAM %d\n", ctrlr->trid.adrfam);
1201 0 : return -1;
1202 : }
1203 :
1204 0 : SPDK_DEBUGLOG(nvme, "adrfam %d ai_family %d\n", ctrlr->trid.adrfam, family);
1205 :
1206 0 : memset(&dst_addr, 0, sizeof(dst_addr));
1207 :
1208 0 : SPDK_DEBUGLOG(nvme, "trsvcid is %s\n", ctrlr->trid.trsvcid);
1209 0 : rc = nvme_parse_addr(&dst_addr, family, ctrlr->trid.traddr, ctrlr->trid.trsvcid, &port);
1210 0 : if (rc != 0) {
1211 0 : SPDK_ERRLOG("dst_addr nvme_parse_addr() failed\n");
1212 0 : return -1;
1213 : }
1214 :
1215 0 : if (ctrlr->opts.src_addr[0] || ctrlr->opts.src_svcid[0]) {
1216 0 : memset(&src_addr, 0, sizeof(src_addr));
1217 0 : rc = nvme_parse_addr(&src_addr, family,
1218 0 : ctrlr->opts.src_addr[0] ? ctrlr->opts.src_addr : NULL,
1219 0 : ctrlr->opts.src_svcid[0] ? ctrlr->opts.src_svcid : NULL,
1220 : &src_port);
1221 0 : if (rc != 0) {
1222 0 : SPDK_ERRLOG("src_addr nvme_parse_addr() failed\n");
1223 0 : return -1;
1224 : }
1225 0 : src_addr_specified = true;
1226 : } else {
1227 0 : src_addr_specified = false;
1228 : }
1229 :
1230 0 : rc = rdma_create_id(rctrlr->cm_channel, &rqpair->cm_id, rqpair, RDMA_PS_TCP);
1231 0 : if (rc < 0) {
1232 0 : SPDK_ERRLOG("rdma_create_id() failed\n");
1233 0 : return -1;
1234 : }
1235 :
1236 0 : rc = nvme_rdma_resolve_addr(rqpair,
1237 : src_addr_specified ? (struct sockaddr *)&src_addr : NULL,
1238 : (struct sockaddr *)&dst_addr);
1239 0 : if (rc < 0) {
1240 0 : SPDK_ERRLOG("nvme_rdma_resolve_addr() failed\n");
1241 0 : return -1;
1242 : }
1243 :
1244 0 : rqpair->state = NVME_RDMA_QPAIR_STATE_INITIALIZING;
1245 :
1246 0 : if (qpair->poll_group != NULL) {
1247 0 : group = nvme_rdma_poll_group(qpair->poll_group);
1248 0 : TAILQ_INSERT_TAIL(&group->connecting_qpairs, rqpair, link_connecting);
1249 : }
1250 :
1251 0 : return 0;
1252 : }
1253 :
1254 : static int
1255 0 : nvme_rdma_stale_conn_reconnect(struct nvme_rdma_qpair *rqpair)
1256 : {
1257 0 : struct spdk_nvme_qpair *qpair = &rqpair->qpair;
1258 :
1259 0 : if (spdk_get_ticks() < rqpair->evt_timeout_ticks) {
1260 0 : return -EAGAIN;
1261 : }
1262 :
1263 0 : return nvme_rdma_ctrlr_connect_qpair(qpair->ctrlr, qpair);
1264 : }
1265 :
1266 : static int
1267 0 : nvme_rdma_ctrlr_connect_qpair_poll(struct spdk_nvme_ctrlr *ctrlr,
1268 : struct spdk_nvme_qpair *qpair)
1269 : {
1270 0 : struct nvme_rdma_qpair *rqpair = nvme_rdma_qpair(qpair);
1271 : int rc;
1272 :
1273 0 : if (rqpair->in_connect_poll) {
1274 0 : return -EAGAIN;
1275 : }
1276 :
1277 0 : rqpair->in_connect_poll = true;
1278 :
1279 0 : switch (rqpair->state) {
1280 0 : case NVME_RDMA_QPAIR_STATE_INVALID:
1281 0 : rc = -EAGAIN;
1282 0 : break;
1283 :
1284 0 : case NVME_RDMA_QPAIR_STATE_INITIALIZING:
1285 : case NVME_RDMA_QPAIR_STATE_EXITING:
1286 0 : if (!nvme_qpair_is_admin_queue(qpair)) {
1287 0 : nvme_ctrlr_lock(ctrlr);
1288 : }
1289 :
1290 0 : rc = nvme_rdma_process_event_poll(rqpair);
1291 :
1292 0 : if (!nvme_qpair_is_admin_queue(qpair)) {
1293 0 : nvme_ctrlr_unlock(ctrlr);
1294 : }
1295 :
1296 0 : if (rc == 0) {
1297 0 : rc = -EAGAIN;
1298 : }
1299 0 : rqpair->in_connect_poll = false;
1300 :
1301 0 : return rc;
1302 :
1303 0 : case NVME_RDMA_QPAIR_STATE_STALE_CONN:
1304 0 : rc = nvme_rdma_stale_conn_reconnect(rqpair);
1305 0 : if (rc == 0) {
1306 0 : rc = -EAGAIN;
1307 : }
1308 0 : break;
1309 0 : case NVME_RDMA_QPAIR_STATE_FABRIC_CONNECT_SEND:
1310 0 : rc = nvme_fabric_qpair_connect_async(qpair, rqpair->num_entries + 1);
1311 0 : if (rc == 0) {
1312 0 : rqpair->state = NVME_RDMA_QPAIR_STATE_FABRIC_CONNECT_POLL;
1313 0 : rc = -EAGAIN;
1314 : } else {
1315 0 : SPDK_ERRLOG("Failed to send an NVMe-oF Fabric CONNECT command\n");
1316 : }
1317 0 : break;
1318 0 : case NVME_RDMA_QPAIR_STATE_FABRIC_CONNECT_POLL:
1319 0 : rc = nvme_fabric_qpair_connect_poll(qpair);
1320 0 : if (rc == 0) {
1321 0 : if (nvme_fabric_qpair_auth_required(qpair)) {
1322 0 : rc = nvme_fabric_qpair_authenticate_async(qpair);
1323 0 : if (rc == 0) {
1324 0 : rqpair->state = NVME_RDMA_QPAIR_STATE_AUTHENTICATING;
1325 0 : rc = -EAGAIN;
1326 : }
1327 : } else {
1328 0 : rqpair->state = NVME_RDMA_QPAIR_STATE_RUNNING;
1329 0 : nvme_qpair_set_state(qpair, NVME_QPAIR_CONNECTED);
1330 : }
1331 0 : } else if (rc != -EAGAIN) {
1332 0 : SPDK_ERRLOG("Failed to poll NVMe-oF Fabric CONNECT command\n");
1333 : }
1334 0 : break;
1335 0 : case NVME_RDMA_QPAIR_STATE_AUTHENTICATING:
1336 0 : rc = nvme_fabric_qpair_authenticate_poll(qpair);
1337 0 : if (rc == 0) {
1338 0 : rqpair->state = NVME_RDMA_QPAIR_STATE_RUNNING;
1339 0 : nvme_qpair_set_state(qpair, NVME_QPAIR_CONNECTED);
1340 : }
1341 0 : break;
1342 0 : case NVME_RDMA_QPAIR_STATE_RUNNING:
1343 0 : rc = 0;
1344 0 : break;
1345 0 : default:
1346 0 : assert(false);
1347 : rc = -EINVAL;
1348 : break;
1349 : }
1350 :
1351 0 : rqpair->in_connect_poll = false;
1352 :
1353 0 : return rc;
1354 : }
1355 :
1356 : static inline int
1357 28 : nvme_rdma_get_memory_translation(struct nvme_request *req, struct nvme_rdma_qpair *rqpair,
1358 : struct nvme_rdma_memory_translation_ctx *_ctx)
1359 : {
1360 : struct spdk_memory_domain_translation_ctx ctx;
1361 28 : struct spdk_memory_domain_translation_result dma_translation = {.iov_count = 0};
1362 : struct spdk_rdma_utils_memory_translation rdma_translation;
1363 : int rc;
1364 :
1365 28 : assert(req);
1366 28 : assert(rqpair);
1367 28 : assert(_ctx);
1368 :
1369 28 : if (req->payload.opts && req->payload.opts->memory_domain) {
1370 2 : ctx.size = sizeof(struct spdk_memory_domain_translation_ctx);
1371 2 : ctx.rdma.ibv_qp = rqpair->rdma_qp->qp;
1372 2 : dma_translation.size = sizeof(struct spdk_memory_domain_translation_result);
1373 :
1374 6 : rc = spdk_memory_domain_translate_data(req->payload.opts->memory_domain,
1375 2 : req->payload.opts->memory_domain_ctx,
1376 2 : rqpair->rdma_qp->domain, &ctx, _ctx->addr,
1377 : _ctx->length, &dma_translation);
1378 2 : if (spdk_unlikely(rc) || dma_translation.iov_count != 1) {
1379 1 : SPDK_ERRLOG("DMA memory translation failed, rc %d, iov count %u\n", rc, dma_translation.iov_count);
1380 1 : return rc;
1381 : }
1382 :
1383 1 : _ctx->lkey = dma_translation.rdma.lkey;
1384 1 : _ctx->rkey = dma_translation.rdma.rkey;
1385 1 : _ctx->addr = dma_translation.iov.iov_base;
1386 1 : _ctx->length = dma_translation.iov.iov_len;
1387 : } else {
1388 26 : rc = spdk_rdma_utils_get_translation(rqpair->mr_map, _ctx->addr, _ctx->length, &rdma_translation);
1389 26 : if (spdk_unlikely(rc)) {
1390 2 : SPDK_ERRLOG("RDMA memory translation failed, rc %d\n", rc);
1391 2 : return rc;
1392 : }
1393 24 : if (rdma_translation.translation_type == SPDK_RDMA_UTILS_TRANSLATION_MR) {
1394 24 : _ctx->lkey = rdma_translation.mr_or_key.mr->lkey;
1395 24 : _ctx->rkey = rdma_translation.mr_or_key.mr->rkey;
1396 : } else {
1397 0 : _ctx->lkey = _ctx->rkey = (uint32_t)rdma_translation.mr_or_key.key;
1398 : }
1399 : }
1400 :
1401 25 : return 0;
1402 : }
1403 :
1404 :
1405 : /*
1406 : * Build SGL describing empty payload.
1407 : */
1408 : static int
1409 2 : nvme_rdma_build_null_request(struct spdk_nvme_rdma_req *rdma_req)
1410 : {
1411 2 : struct nvme_request *req = rdma_req->req;
1412 :
1413 2 : req->cmd.psdt = SPDK_NVME_PSDT_SGL_MPTR_CONTIG;
1414 :
1415 : /* The first element of this SGL is pointing at an
1416 : * spdk_nvmf_cmd object. For this particular command,
1417 : * we only need the first 64 bytes corresponding to
1418 : * the NVMe command. */
1419 2 : rdma_req->send_sgl[0].length = sizeof(struct spdk_nvme_cmd);
1420 :
1421 : /* The RDMA SGL needs one element describing the NVMe command. */
1422 2 : rdma_req->send_wr.num_sge = 1;
1423 :
1424 2 : req->cmd.dptr.sgl1.keyed.type = SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK;
1425 2 : req->cmd.dptr.sgl1.keyed.subtype = SPDK_NVME_SGL_SUBTYPE_ADDRESS;
1426 2 : req->cmd.dptr.sgl1.keyed.length = 0;
1427 2 : req->cmd.dptr.sgl1.keyed.key = 0;
1428 2 : req->cmd.dptr.sgl1.address = 0;
1429 :
1430 2 : return 0;
1431 : }
1432 :
1433 : static inline void
1434 3 : nvme_rdma_configure_contig_inline_request(struct spdk_nvme_rdma_req *rdma_req,
1435 : struct nvme_request *req, struct nvme_rdma_memory_translation_ctx *ctx)
1436 : {
1437 3 : rdma_req->send_sgl[1].lkey = ctx->lkey;
1438 :
1439 : /* The first element of this SGL is pointing at an
1440 : * spdk_nvmf_cmd object. For this particular command,
1441 : * we only need the first 64 bytes corresponding to
1442 : * the NVMe command. */
1443 3 : rdma_req->send_sgl[0].length = sizeof(struct spdk_nvme_cmd);
1444 :
1445 3 : rdma_req->send_sgl[1].addr = (uint64_t)ctx->addr;
1446 3 : rdma_req->send_sgl[1].length = (uint32_t)ctx->length;
1447 :
1448 : /* The RDMA SGL contains two elements. The first describes
1449 : * the NVMe command and the second describes the data
1450 : * payload. */
1451 3 : rdma_req->send_wr.num_sge = 2;
1452 :
1453 3 : req->cmd.psdt = SPDK_NVME_PSDT_SGL_MPTR_CONTIG;
1454 3 : req->cmd.dptr.sgl1.unkeyed.type = SPDK_NVME_SGL_TYPE_DATA_BLOCK;
1455 3 : req->cmd.dptr.sgl1.unkeyed.subtype = SPDK_NVME_SGL_SUBTYPE_OFFSET;
1456 3 : req->cmd.dptr.sgl1.unkeyed.length = (uint32_t)ctx->length;
1457 : /* Inline only supported for icdoff == 0 currently. This function will
1458 : * not get called for controllers with other values. */
1459 3 : req->cmd.dptr.sgl1.address = (uint64_t)0;
1460 3 : }
1461 :
1462 : /*
1463 : * Build inline SGL describing contiguous payload buffer.
1464 : */
1465 : static inline int
1466 3 : nvme_rdma_build_contig_inline_request(struct nvme_rdma_qpair *rqpair,
1467 : struct spdk_nvme_rdma_req *rdma_req)
1468 : {
1469 3 : struct nvme_request *req = rdma_req->req;
1470 9 : struct nvme_rdma_memory_translation_ctx ctx = {
1471 3 : .addr = (uint8_t *)req->payload.contig_or_cb_arg + req->payload_offset,
1472 3 : .length = req->payload_size
1473 : };
1474 : int rc;
1475 :
1476 3 : assert(ctx.length != 0);
1477 3 : assert(nvme_payload_type(&req->payload) == NVME_PAYLOAD_TYPE_CONTIG);
1478 :
1479 3 : rc = nvme_rdma_get_memory_translation(req, rqpair, &ctx);
1480 3 : if (spdk_unlikely(rc)) {
1481 0 : return -1;
1482 : }
1483 :
1484 3 : nvme_rdma_configure_contig_inline_request(rdma_req, req, &ctx);
1485 :
1486 3 : return 0;
1487 : }
1488 :
1489 : static inline void
1490 2 : nvme_rdma_configure_contig_request(struct spdk_nvme_rdma_req *rdma_req, struct nvme_request *req,
1491 : struct nvme_rdma_memory_translation_ctx *ctx)
1492 : {
1493 2 : req->cmd.dptr.sgl1.keyed.key = ctx->rkey;
1494 :
1495 : /* The first element of this SGL is pointing at an
1496 : * spdk_nvmf_cmd object. For this particular command,
1497 : * we only need the first 64 bytes corresponding to
1498 : * the NVMe command. */
1499 2 : rdma_req->send_sgl[0].length = sizeof(struct spdk_nvme_cmd);
1500 :
1501 : /* The RDMA SGL needs one element describing the NVMe command. */
1502 2 : rdma_req->send_wr.num_sge = 1;
1503 :
1504 2 : req->cmd.psdt = SPDK_NVME_PSDT_SGL_MPTR_CONTIG;
1505 2 : req->cmd.dptr.sgl1.keyed.type = SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK;
1506 2 : req->cmd.dptr.sgl1.keyed.subtype = SPDK_NVME_SGL_SUBTYPE_ADDRESS;
1507 2 : req->cmd.dptr.sgl1.keyed.length = (uint32_t)ctx->length;
1508 2 : req->cmd.dptr.sgl1.address = (uint64_t)ctx->addr;
1509 2 : }
1510 :
1511 : /*
1512 : * Build SGL describing contiguous payload buffer.
1513 : */
1514 : static inline int
1515 3 : nvme_rdma_build_contig_request(struct nvme_rdma_qpair *rqpair,
1516 : struct spdk_nvme_rdma_req *rdma_req)
1517 : {
1518 3 : struct nvme_request *req = rdma_req->req;
1519 9 : struct nvme_rdma_memory_translation_ctx ctx = {
1520 3 : .addr = (uint8_t *)req->payload.contig_or_cb_arg + req->payload_offset,
1521 3 : .length = req->payload_size
1522 : };
1523 : int rc;
1524 :
1525 3 : assert(req->payload_size != 0);
1526 3 : assert(nvme_payload_type(&req->payload) == NVME_PAYLOAD_TYPE_CONTIG);
1527 :
1528 3 : if (spdk_unlikely(req->payload_size > NVME_RDMA_MAX_KEYED_SGL_LENGTH)) {
1529 1 : SPDK_ERRLOG("SGL length %u exceeds max keyed SGL block size %u\n",
1530 : req->payload_size, NVME_RDMA_MAX_KEYED_SGL_LENGTH);
1531 1 : return -1;
1532 : }
1533 :
1534 2 : rc = nvme_rdma_get_memory_translation(req, rqpair, &ctx);
1535 2 : if (spdk_unlikely(rc)) {
1536 0 : return -1;
1537 : }
1538 :
1539 2 : nvme_rdma_configure_contig_request(rdma_req, req, &ctx);
1540 :
1541 2 : return 0;
1542 : }
1543 :
1544 : /*
1545 : * Build SGL describing scattered payload buffer.
1546 : */
1547 : static inline int
1548 7 : nvme_rdma_build_sgl_request(struct nvme_rdma_qpair *rqpair,
1549 : struct spdk_nvme_rdma_req *rdma_req)
1550 : {
1551 7 : struct nvme_request *req = rdma_req->req;
1552 7 : struct spdk_nvmf_cmd *cmd = &rqpair->cmds[rdma_req->id];
1553 : struct nvme_rdma_memory_translation_ctx ctx;
1554 : uint32_t remaining_size;
1555 : uint32_t sge_length;
1556 : int rc, max_num_sgl, num_sgl_desc;
1557 :
1558 7 : assert(req->payload_size != 0);
1559 7 : assert(nvme_payload_type(&req->payload) == NVME_PAYLOAD_TYPE_SGL);
1560 7 : assert(req->payload.reset_sgl_fn != NULL);
1561 7 : assert(req->payload.next_sge_fn != NULL);
1562 7 : req->payload.reset_sgl_fn(req->payload.contig_or_cb_arg, req->payload_offset);
1563 :
1564 7 : max_num_sgl = req->qpair->ctrlr->max_sges;
1565 :
1566 7 : remaining_size = req->payload_size;
1567 7 : num_sgl_desc = 0;
1568 : do {
1569 18 : rc = req->payload.next_sge_fn(req->payload.contig_or_cb_arg, &ctx.addr, &sge_length);
1570 18 : if (spdk_unlikely(rc)) {
1571 1 : return -1;
1572 : }
1573 :
1574 17 : sge_length = spdk_min(remaining_size, sge_length);
1575 :
1576 17 : if (spdk_unlikely(sge_length > NVME_RDMA_MAX_KEYED_SGL_LENGTH)) {
1577 1 : SPDK_ERRLOG("SGL length %u exceeds max keyed SGL block size %u\n",
1578 : sge_length, NVME_RDMA_MAX_KEYED_SGL_LENGTH);
1579 1 : return -1;
1580 : }
1581 16 : ctx.length = sge_length;
1582 16 : rc = nvme_rdma_get_memory_translation(req, rqpair, &ctx);
1583 16 : if (spdk_unlikely(rc)) {
1584 1 : return -1;
1585 : }
1586 :
1587 15 : cmd->sgl[num_sgl_desc].keyed.key = ctx.rkey;
1588 15 : cmd->sgl[num_sgl_desc].keyed.type = SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK;
1589 15 : cmd->sgl[num_sgl_desc].keyed.subtype = SPDK_NVME_SGL_SUBTYPE_ADDRESS;
1590 15 : cmd->sgl[num_sgl_desc].keyed.length = (uint32_t)ctx.length;
1591 15 : cmd->sgl[num_sgl_desc].address = (uint64_t)ctx.addr;
1592 :
1593 15 : remaining_size -= ctx.length;
1594 15 : num_sgl_desc++;
1595 15 : } while (remaining_size > 0 && num_sgl_desc < max_num_sgl);
1596 :
1597 :
1598 : /* Should be impossible if we did our sgl checks properly up the stack, but do a sanity check here. */
1599 4 : if (spdk_unlikely(remaining_size > 0)) {
1600 0 : return -1;
1601 : }
1602 :
1603 4 : req->cmd.psdt = SPDK_NVME_PSDT_SGL_MPTR_CONTIG;
1604 :
1605 : /* The RDMA SGL needs one element describing some portion
1606 : * of the spdk_nvmf_cmd structure. */
1607 4 : rdma_req->send_wr.num_sge = 1;
1608 :
1609 : /*
1610 : * If only one SGL descriptor is required, it can be embedded directly in the command
1611 : * as a data block descriptor.
1612 : */
1613 4 : if (num_sgl_desc == 1) {
1614 : /* The first element of this SGL is pointing at an
1615 : * spdk_nvmf_cmd object. For this particular command,
1616 : * we only need the first 64 bytes corresponding to
1617 : * the NVMe command. */
1618 2 : rdma_req->send_sgl[0].length = sizeof(struct spdk_nvme_cmd);
1619 :
1620 2 : req->cmd.dptr.sgl1.keyed.type = cmd->sgl[0].keyed.type;
1621 2 : req->cmd.dptr.sgl1.keyed.subtype = cmd->sgl[0].keyed.subtype;
1622 2 : req->cmd.dptr.sgl1.keyed.length = cmd->sgl[0].keyed.length;
1623 2 : req->cmd.dptr.sgl1.keyed.key = cmd->sgl[0].keyed.key;
1624 2 : req->cmd.dptr.sgl1.address = cmd->sgl[0].address;
1625 : } else {
1626 : /*
1627 : * Otherwise, The SGL descriptor embedded in the command must point to the list of
1628 : * SGL descriptors used to describe the operation. In that case it is a last segment descriptor.
1629 : */
1630 2 : uint32_t descriptors_size = sizeof(struct spdk_nvme_sgl_descriptor) * num_sgl_desc;
1631 :
1632 2 : if (spdk_unlikely(descriptors_size > rqpair->qpair.ctrlr->ioccsz_bytes)) {
1633 1 : SPDK_ERRLOG("Size of SGL descriptors (%u) exceeds ICD (%u)\n",
1634 : descriptors_size, rqpair->qpair.ctrlr->ioccsz_bytes);
1635 1 : return -1;
1636 : }
1637 1 : rdma_req->send_sgl[0].length = sizeof(struct spdk_nvme_cmd) + descriptors_size;
1638 :
1639 1 : req->cmd.dptr.sgl1.unkeyed.type = SPDK_NVME_SGL_TYPE_LAST_SEGMENT;
1640 1 : req->cmd.dptr.sgl1.unkeyed.subtype = SPDK_NVME_SGL_SUBTYPE_OFFSET;
1641 1 : req->cmd.dptr.sgl1.unkeyed.length = descriptors_size;
1642 1 : req->cmd.dptr.sgl1.address = (uint64_t)0;
1643 : }
1644 :
1645 3 : return 0;
1646 : }
1647 :
1648 : /*
1649 : * Build inline SGL describing sgl payload buffer.
1650 : */
1651 : static inline int
1652 3 : nvme_rdma_build_sgl_inline_request(struct nvme_rdma_qpair *rqpair,
1653 : struct spdk_nvme_rdma_req *rdma_req)
1654 : {
1655 3 : struct nvme_request *req = rdma_req->req;
1656 : struct nvme_rdma_memory_translation_ctx ctx;
1657 : uint32_t length;
1658 : int rc;
1659 :
1660 3 : assert(req->payload_size != 0);
1661 3 : assert(nvme_payload_type(&req->payload) == NVME_PAYLOAD_TYPE_SGL);
1662 3 : assert(req->payload.reset_sgl_fn != NULL);
1663 3 : assert(req->payload.next_sge_fn != NULL);
1664 3 : req->payload.reset_sgl_fn(req->payload.contig_or_cb_arg, req->payload_offset);
1665 :
1666 3 : rc = req->payload.next_sge_fn(req->payload.contig_or_cb_arg, &ctx.addr, &length);
1667 3 : if (spdk_unlikely(rc)) {
1668 0 : return -1;
1669 : }
1670 :
1671 3 : if (length < req->payload_size) {
1672 0 : SPDK_DEBUGLOG(nvme, "Inline SGL request split so sending separately.\n");
1673 0 : return nvme_rdma_build_sgl_request(rqpair, rdma_req);
1674 : }
1675 :
1676 3 : if (length > req->payload_size) {
1677 0 : length = req->payload_size;
1678 : }
1679 :
1680 3 : ctx.length = length;
1681 3 : rc = nvme_rdma_get_memory_translation(req, rqpair, &ctx);
1682 3 : if (spdk_unlikely(rc)) {
1683 0 : return -1;
1684 : }
1685 :
1686 3 : rdma_req->send_sgl[1].addr = (uint64_t)ctx.addr;
1687 3 : rdma_req->send_sgl[1].length = (uint32_t)ctx.length;
1688 3 : rdma_req->send_sgl[1].lkey = ctx.lkey;
1689 :
1690 3 : rdma_req->send_wr.num_sge = 2;
1691 :
1692 : /* The first element of this SGL is pointing at an
1693 : * spdk_nvmf_cmd object. For this particular command,
1694 : * we only need the first 64 bytes corresponding to
1695 : * the NVMe command. */
1696 3 : rdma_req->send_sgl[0].length = sizeof(struct spdk_nvme_cmd);
1697 :
1698 3 : req->cmd.psdt = SPDK_NVME_PSDT_SGL_MPTR_CONTIG;
1699 3 : req->cmd.dptr.sgl1.unkeyed.type = SPDK_NVME_SGL_TYPE_DATA_BLOCK;
1700 3 : req->cmd.dptr.sgl1.unkeyed.subtype = SPDK_NVME_SGL_SUBTYPE_OFFSET;
1701 3 : req->cmd.dptr.sgl1.unkeyed.length = (uint32_t)ctx.length;
1702 : /* Inline only supported for icdoff == 0 currently. This function will
1703 : * not get called for controllers with other values. */
1704 3 : req->cmd.dptr.sgl1.address = (uint64_t)0;
1705 :
1706 3 : return 0;
1707 : }
1708 :
1709 : static inline int
1710 6 : nvme_rdma_req_init(struct nvme_rdma_qpair *rqpair, struct spdk_nvme_rdma_req *rdma_req)
1711 : {
1712 6 : struct nvme_request *req = rdma_req->req;
1713 6 : struct spdk_nvme_ctrlr *ctrlr = rqpair->qpair.ctrlr;
1714 : enum nvme_payload_type payload_type;
1715 : bool icd_supported;
1716 6 : int rc = -1;
1717 :
1718 6 : payload_type = nvme_payload_type(&req->payload);
1719 : /*
1720 : * Check if icdoff is non zero, to avoid interop conflicts with
1721 : * targets with non-zero icdoff. Both SPDK and the Linux kernel
1722 : * targets use icdoff = 0. For targets with non-zero icdoff, we
1723 : * will currently just not use inline data for now.
1724 : */
1725 6 : icd_supported = spdk_nvme_opc_get_data_transfer(req->cmd.opc) == SPDK_NVME_DATA_HOST_TO_CONTROLLER
1726 6 : && req->payload_size <= ctrlr->ioccsz_bytes && ctrlr->icdoff == 0;
1727 :
1728 6 : if (spdk_unlikely(req->payload_size == 0)) {
1729 2 : rc = nvme_rdma_build_null_request(rdma_req);
1730 4 : } else if (payload_type == NVME_PAYLOAD_TYPE_CONTIG) {
1731 2 : if (icd_supported) {
1732 1 : rc = nvme_rdma_build_contig_inline_request(rqpair, rdma_req);
1733 : } else {
1734 1 : rc = nvme_rdma_build_contig_request(rqpair, rdma_req);
1735 : }
1736 2 : } else if (payload_type == NVME_PAYLOAD_TYPE_SGL) {
1737 2 : if (icd_supported) {
1738 1 : rc = nvme_rdma_build_sgl_inline_request(rqpair, rdma_req);
1739 : } else {
1740 1 : rc = nvme_rdma_build_sgl_request(rqpair, rdma_req);
1741 : }
1742 : }
1743 :
1744 6 : if (spdk_unlikely(rc)) {
1745 0 : return rc;
1746 : }
1747 :
1748 6 : memcpy(&rqpair->cmds[rdma_req->id], &req->cmd, sizeof(req->cmd));
1749 6 : return 0;
1750 : }
1751 :
1752 : static struct spdk_nvme_qpair *
1753 5 : nvme_rdma_ctrlr_create_qpair(struct spdk_nvme_ctrlr *ctrlr,
1754 : uint16_t qid, uint32_t qsize,
1755 : enum spdk_nvme_qprio qprio,
1756 : uint32_t num_requests,
1757 : bool delay_cmd_submit,
1758 : bool async)
1759 : {
1760 : struct nvme_rdma_qpair *rqpair;
1761 : struct spdk_nvme_qpair *qpair;
1762 : int rc;
1763 :
1764 5 : if (qsize < SPDK_NVME_QUEUE_MIN_ENTRIES) {
1765 2 : SPDK_ERRLOG("Failed to create qpair with size %u. Minimum queue size is %d.\n",
1766 : qsize, SPDK_NVME_QUEUE_MIN_ENTRIES);
1767 2 : return NULL;
1768 : }
1769 :
1770 3 : rqpair = spdk_zmalloc(sizeof(struct nvme_rdma_qpair), 0, NULL, SPDK_ENV_NUMA_ID_ANY,
1771 : SPDK_MALLOC_DMA);
1772 3 : if (!rqpair) {
1773 0 : SPDK_ERRLOG("failed to get create rqpair\n");
1774 0 : return NULL;
1775 : }
1776 :
1777 : /* Set num_entries one less than queue size. According to NVMe
1778 : * and NVMe-oF specs we can not submit queue size requests,
1779 : * one slot shall always remain empty.
1780 : */
1781 3 : rqpair->num_entries = qsize - 1;
1782 3 : rqpair->delay_cmd_submit = delay_cmd_submit;
1783 3 : rqpair->state = NVME_RDMA_QPAIR_STATE_INVALID;
1784 3 : qpair = &rqpair->qpair;
1785 3 : rc = nvme_qpair_init(qpair, qid, ctrlr, qprio, num_requests, async);
1786 3 : if (rc != 0) {
1787 0 : spdk_free(rqpair);
1788 0 : return NULL;
1789 : }
1790 :
1791 3 : return qpair;
1792 : }
1793 :
1794 : static void
1795 1 : nvme_rdma_qpair_destroy(struct nvme_rdma_qpair *rqpair)
1796 : {
1797 1 : struct spdk_nvme_qpair *qpair = &rqpair->qpair;
1798 : struct nvme_rdma_ctrlr *rctrlr;
1799 : struct nvme_rdma_cm_event_entry *entry, *tmp;
1800 :
1801 1 : spdk_rdma_utils_free_mem_map(&rqpair->mr_map);
1802 :
1803 1 : if (rqpair->evt) {
1804 0 : rdma_ack_cm_event(rqpair->evt);
1805 0 : rqpair->evt = NULL;
1806 : }
1807 :
1808 : /*
1809 : * This works because we have the controller lock both in
1810 : * this function and in the function where we add new events.
1811 : */
1812 1 : if (qpair->ctrlr != NULL) {
1813 1 : rctrlr = nvme_rdma_ctrlr(qpair->ctrlr);
1814 1 : STAILQ_FOREACH_SAFE(entry, &rctrlr->pending_cm_events, link, tmp) {
1815 0 : if (entry->evt->id->context == rqpair) {
1816 0 : STAILQ_REMOVE(&rctrlr->pending_cm_events, entry, nvme_rdma_cm_event_entry, link);
1817 0 : rdma_ack_cm_event(entry->evt);
1818 0 : STAILQ_INSERT_HEAD(&rctrlr->free_cm_events, entry, link);
1819 : }
1820 : }
1821 : }
1822 :
1823 1 : if (rqpair->cm_id) {
1824 0 : if (rqpair->rdma_qp) {
1825 0 : spdk_rdma_utils_put_pd(rqpair->rdma_qp->qp->pd);
1826 0 : spdk_rdma_provider_qp_destroy(rqpair->rdma_qp);
1827 0 : rqpair->rdma_qp = NULL;
1828 : }
1829 : }
1830 :
1831 1 : if (rqpair->poller) {
1832 : struct nvme_rdma_poll_group *group;
1833 :
1834 0 : assert(qpair->poll_group);
1835 0 : group = nvme_rdma_poll_group(qpair->poll_group);
1836 :
1837 0 : nvme_rdma_poll_group_put_poller(group, rqpair->poller);
1838 :
1839 0 : rqpair->poller = NULL;
1840 0 : rqpair->cq = NULL;
1841 0 : if (rqpair->srq) {
1842 0 : rqpair->srq = NULL;
1843 0 : rqpair->rsps = NULL;
1844 : }
1845 1 : } else if (rqpair->cq) {
1846 0 : ibv_destroy_cq(rqpair->cq);
1847 0 : rqpair->cq = NULL;
1848 : }
1849 :
1850 1 : nvme_rdma_free_reqs(rqpair);
1851 1 : nvme_rdma_free_rsps(rqpair->rsps);
1852 1 : rqpair->rsps = NULL;
1853 :
1854 : /* destroy cm_id last so cma device will not be freed before we destroy the cq. */
1855 1 : if (rqpair->cm_id) {
1856 0 : rdma_destroy_id(rqpair->cm_id);
1857 0 : rqpair->cm_id = NULL;
1858 : }
1859 1 : }
1860 :
1861 : static void nvme_rdma_qpair_abort_reqs(struct spdk_nvme_qpair *qpair, uint32_t dnr);
1862 :
1863 : static int
1864 1 : nvme_rdma_qpair_disconnected(struct nvme_rdma_qpair *rqpair, int ret)
1865 : {
1866 1 : if (ret) {
1867 0 : SPDK_DEBUGLOG(nvme, "Target did not respond to qpair disconnect.\n");
1868 0 : goto quiet;
1869 : }
1870 :
1871 1 : if (rqpair->poller == NULL) {
1872 : /* If poller is not used, cq is not shared.
1873 : * So complete disconnecting qpair immediately.
1874 : */
1875 1 : goto quiet;
1876 : }
1877 :
1878 0 : if (rqpair->rsps == NULL) {
1879 0 : goto quiet;
1880 : }
1881 :
1882 0 : if (rqpair->need_destroy ||
1883 0 : (rqpair->current_num_sends != 0 ||
1884 0 : (!rqpair->srq && rqpair->rsps->current_num_recvs != 0))) {
1885 0 : rqpair->state = NVME_RDMA_QPAIR_STATE_LINGERING;
1886 0 : rqpair->evt_timeout_ticks = (NVME_RDMA_DISCONNECTED_QPAIR_TIMEOUT_US * spdk_get_ticks_hz()) /
1887 0 : SPDK_SEC_TO_USEC + spdk_get_ticks();
1888 :
1889 0 : return -EAGAIN;
1890 : }
1891 :
1892 0 : quiet:
1893 1 : rqpair->state = NVME_RDMA_QPAIR_STATE_EXITED;
1894 :
1895 1 : nvme_rdma_qpair_abort_reqs(&rqpair->qpair, rqpair->qpair.abort_dnr);
1896 1 : nvme_rdma_qpair_destroy(rqpair);
1897 1 : nvme_transport_ctrlr_disconnect_qpair_done(&rqpair->qpair);
1898 :
1899 1 : return 0;
1900 : }
1901 :
1902 : static int
1903 0 : nvme_rdma_qpair_wait_until_quiet(struct nvme_rdma_qpair *rqpair)
1904 : {
1905 0 : struct spdk_nvme_qpair *qpair = &rqpair->qpair;
1906 0 : struct spdk_nvme_ctrlr *ctrlr = qpair->ctrlr;
1907 :
1908 0 : if (spdk_get_ticks() < rqpair->evt_timeout_ticks &&
1909 0 : (rqpair->current_num_sends != 0 ||
1910 0 : (!rqpair->srq && rqpair->rsps->current_num_recvs != 0))) {
1911 0 : return -EAGAIN;
1912 : }
1913 :
1914 0 : rqpair->state = NVME_RDMA_QPAIR_STATE_EXITED;
1915 0 : nvme_rdma_qpair_abort_reqs(qpair, qpair->abort_dnr);
1916 0 : if (!nvme_qpair_is_admin_queue(qpair)) {
1917 0 : nvme_robust_mutex_lock(&ctrlr->ctrlr_lock);
1918 : }
1919 0 : nvme_rdma_qpair_destroy(rqpair);
1920 0 : if (!nvme_qpair_is_admin_queue(qpair)) {
1921 0 : nvme_robust_mutex_unlock(&ctrlr->ctrlr_lock);
1922 : }
1923 0 : nvme_transport_ctrlr_disconnect_qpair_done(&rqpair->qpair);
1924 :
1925 0 : return 0;
1926 : }
1927 :
1928 : static void
1929 0 : _nvme_rdma_ctrlr_disconnect_qpair(struct spdk_nvme_ctrlr *ctrlr, struct spdk_nvme_qpair *qpair,
1930 : nvme_rdma_cm_event_cb disconnected_qpair_cb)
1931 : {
1932 0 : struct nvme_rdma_qpair *rqpair = nvme_rdma_qpair(qpair);
1933 : int rc;
1934 :
1935 0 : assert(disconnected_qpair_cb != NULL);
1936 :
1937 0 : rqpair->state = NVME_RDMA_QPAIR_STATE_EXITING;
1938 :
1939 0 : if (rqpair->cm_id) {
1940 0 : if (rqpair->rdma_qp) {
1941 0 : rc = spdk_rdma_provider_qp_disconnect(rqpair->rdma_qp);
1942 0 : if ((qpair->ctrlr != NULL) && (rc == 0)) {
1943 0 : rc = nvme_rdma_process_event_start(rqpair, RDMA_CM_EVENT_DISCONNECTED,
1944 : disconnected_qpair_cb);
1945 0 : if (rc == 0) {
1946 0 : return;
1947 : }
1948 : }
1949 : }
1950 : }
1951 :
1952 0 : disconnected_qpair_cb(rqpair, 0);
1953 : }
1954 :
1955 : static int
1956 0 : nvme_rdma_ctrlr_disconnect_qpair_poll(struct spdk_nvme_ctrlr *ctrlr, struct spdk_nvme_qpair *qpair)
1957 : {
1958 0 : struct nvme_rdma_qpair *rqpair = nvme_rdma_qpair(qpair);
1959 : int rc;
1960 :
1961 0 : switch (rqpair->state) {
1962 0 : case NVME_RDMA_QPAIR_STATE_EXITING:
1963 0 : if (!nvme_qpair_is_admin_queue(qpair)) {
1964 0 : nvme_ctrlr_lock(ctrlr);
1965 : }
1966 :
1967 0 : rc = nvme_rdma_process_event_poll(rqpair);
1968 :
1969 0 : if (!nvme_qpair_is_admin_queue(qpair)) {
1970 0 : nvme_ctrlr_unlock(ctrlr);
1971 : }
1972 0 : break;
1973 :
1974 0 : case NVME_RDMA_QPAIR_STATE_LINGERING:
1975 0 : rc = nvme_rdma_qpair_wait_until_quiet(rqpair);
1976 0 : break;
1977 0 : case NVME_RDMA_QPAIR_STATE_EXITED:
1978 0 : rc = 0;
1979 0 : break;
1980 :
1981 0 : default:
1982 0 : assert(false);
1983 : rc = -EAGAIN;
1984 : break;
1985 : }
1986 :
1987 0 : return rc;
1988 : }
1989 :
1990 : static void
1991 0 : nvme_rdma_ctrlr_disconnect_qpair(struct spdk_nvme_ctrlr *ctrlr, struct spdk_nvme_qpair *qpair)
1992 : {
1993 : int rc;
1994 :
1995 0 : _nvme_rdma_ctrlr_disconnect_qpair(ctrlr, qpair, nvme_rdma_qpair_disconnected);
1996 :
1997 : /* If the async mode is disabled, poll the qpair until it is actually disconnected.
1998 : * It is ensured that poll_group_process_completions() calls disconnected_qpair_cb
1999 : * for any disconnected qpair. Hence, we do not have to check if the qpair is in
2000 : * a poll group or not.
2001 : * At the same time, if the qpair is being destroyed, i.e. this function is called by
2002 : * spdk_nvme_ctrlr_free_io_qpair then we need to wait until qpair is disconnected, otherwise
2003 : * we may leak some resources.
2004 : */
2005 0 : if (qpair->async && !qpair->destroy_in_progress) {
2006 0 : return;
2007 : }
2008 :
2009 : while (1) {
2010 0 : rc = nvme_rdma_ctrlr_disconnect_qpair_poll(ctrlr, qpair);
2011 0 : if (rc != -EAGAIN) {
2012 0 : break;
2013 : }
2014 : }
2015 : }
2016 :
2017 : static int
2018 0 : nvme_rdma_stale_conn_disconnected(struct nvme_rdma_qpair *rqpair, int ret)
2019 : {
2020 0 : struct spdk_nvme_qpair *qpair = &rqpair->qpair;
2021 :
2022 0 : if (ret) {
2023 0 : SPDK_DEBUGLOG(nvme, "Target did not respond to qpair disconnect.\n");
2024 : }
2025 :
2026 0 : nvme_rdma_qpair_destroy(rqpair);
2027 :
2028 0 : qpair->last_transport_failure_reason = qpair->transport_failure_reason;
2029 0 : qpair->transport_failure_reason = SPDK_NVME_QPAIR_FAILURE_NONE;
2030 :
2031 0 : rqpair->state = NVME_RDMA_QPAIR_STATE_STALE_CONN;
2032 0 : rqpair->evt_timeout_ticks = (NVME_RDMA_STALE_CONN_RETRY_DELAY_US * spdk_get_ticks_hz()) /
2033 0 : SPDK_SEC_TO_USEC + spdk_get_ticks();
2034 :
2035 0 : return 0;
2036 : }
2037 :
2038 : static int
2039 0 : nvme_rdma_stale_conn_retry(struct nvme_rdma_qpair *rqpair)
2040 : {
2041 0 : struct spdk_nvme_qpair *qpair = &rqpair->qpair;
2042 :
2043 0 : if (rqpair->stale_conn_retry_count >= NVME_RDMA_STALE_CONN_RETRY_MAX) {
2044 0 : SPDK_ERRLOG("Retry failed %d times, give up stale connection to qpair (cntlid:%u, qid:%u).\n",
2045 : NVME_RDMA_STALE_CONN_RETRY_MAX, qpair->ctrlr->cntlid, qpair->id);
2046 0 : return -ESTALE;
2047 : }
2048 :
2049 0 : rqpair->stale_conn_retry_count++;
2050 :
2051 0 : SPDK_NOTICELOG("%d times, retry stale connection to qpair (cntlid:%u, qid:%u).\n",
2052 : rqpair->stale_conn_retry_count, qpair->ctrlr->cntlid, qpair->id);
2053 :
2054 0 : _nvme_rdma_ctrlr_disconnect_qpair(qpair->ctrlr, qpair, nvme_rdma_stale_conn_disconnected);
2055 :
2056 0 : return 0;
2057 : }
2058 :
2059 : static int
2060 1 : nvme_rdma_ctrlr_delete_io_qpair(struct spdk_nvme_ctrlr *ctrlr, struct spdk_nvme_qpair *qpair)
2061 : {
2062 : struct nvme_rdma_qpair *rqpair;
2063 :
2064 1 : assert(qpair != NULL);
2065 1 : rqpair = nvme_rdma_qpair(qpair);
2066 :
2067 1 : if (rqpair->state != NVME_RDMA_QPAIR_STATE_EXITED) {
2068 : int rc __attribute__((unused));
2069 :
2070 : /* qpair was removed from the poll group while the disconnect is not finished.
2071 : * Destroy rdma resources forcefully. */
2072 1 : rc = nvme_rdma_qpair_disconnected(rqpair, 0);
2073 1 : assert(rc == 0);
2074 : }
2075 :
2076 1 : nvme_rdma_qpair_abort_reqs(qpair, qpair->abort_dnr);
2077 1 : nvme_qpair_deinit(qpair);
2078 :
2079 1 : spdk_free(rqpair);
2080 :
2081 1 : return 0;
2082 : }
2083 :
2084 : static struct spdk_nvme_qpair *
2085 0 : nvme_rdma_ctrlr_create_io_qpair(struct spdk_nvme_ctrlr *ctrlr, uint16_t qid,
2086 : const struct spdk_nvme_io_qpair_opts *opts)
2087 : {
2088 0 : return nvme_rdma_ctrlr_create_qpair(ctrlr, qid, opts->io_queue_size, opts->qprio,
2089 0 : opts->io_queue_requests,
2090 0 : opts->delay_cmd_submit,
2091 0 : opts->async_mode);
2092 : }
2093 :
2094 : static int
2095 0 : nvme_rdma_ctrlr_enable(struct spdk_nvme_ctrlr *ctrlr)
2096 : {
2097 : /* do nothing here */
2098 0 : return 0;
2099 : }
2100 :
2101 : static int nvme_rdma_ctrlr_destruct(struct spdk_nvme_ctrlr *ctrlr);
2102 :
2103 : /* We have to use the typedef in the function declaration to appease astyle. */
2104 : typedef struct spdk_nvme_ctrlr spdk_nvme_ctrlr_t;
2105 :
2106 : static spdk_nvme_ctrlr_t *
2107 1 : nvme_rdma_ctrlr_construct(const struct spdk_nvme_transport_id *trid,
2108 : const struct spdk_nvme_ctrlr_opts *opts,
2109 : void *devhandle)
2110 : {
2111 : struct nvme_rdma_ctrlr *rctrlr;
2112 : struct ibv_context **contexts;
2113 : struct ibv_device_attr dev_attr;
2114 : int i, flag, rc;
2115 :
2116 1 : rctrlr = spdk_zmalloc(sizeof(struct nvme_rdma_ctrlr), 0, NULL, SPDK_ENV_NUMA_ID_ANY,
2117 : SPDK_MALLOC_DMA);
2118 1 : if (rctrlr == NULL) {
2119 0 : SPDK_ERRLOG("could not allocate ctrlr\n");
2120 0 : return NULL;
2121 : }
2122 :
2123 1 : rctrlr->ctrlr.opts = *opts;
2124 1 : rctrlr->ctrlr.trid = *trid;
2125 :
2126 1 : if (opts->transport_retry_count > NVME_RDMA_CTRLR_MAX_TRANSPORT_RETRY_COUNT) {
2127 1 : SPDK_NOTICELOG("transport_retry_count exceeds max value %d, use max value\n",
2128 : NVME_RDMA_CTRLR_MAX_TRANSPORT_RETRY_COUNT);
2129 1 : rctrlr->ctrlr.opts.transport_retry_count = NVME_RDMA_CTRLR_MAX_TRANSPORT_RETRY_COUNT;
2130 : }
2131 :
2132 1 : if (opts->transport_ack_timeout > NVME_RDMA_CTRLR_MAX_TRANSPORT_ACK_TIMEOUT) {
2133 1 : SPDK_NOTICELOG("transport_ack_timeout exceeds max value %d, use max value\n",
2134 : NVME_RDMA_CTRLR_MAX_TRANSPORT_ACK_TIMEOUT);
2135 1 : rctrlr->ctrlr.opts.transport_ack_timeout = NVME_RDMA_CTRLR_MAX_TRANSPORT_ACK_TIMEOUT;
2136 : }
2137 :
2138 1 : contexts = rdma_get_devices(NULL);
2139 1 : if (contexts == NULL) {
2140 0 : SPDK_ERRLOG("rdma_get_devices() failed: %s (%d)\n", spdk_strerror(errno), errno);
2141 0 : spdk_free(rctrlr);
2142 0 : return NULL;
2143 : }
2144 :
2145 1 : i = 0;
2146 1 : rctrlr->max_sge = NVME_RDMA_MAX_SGL_DESCRIPTORS;
2147 :
2148 3 : while (contexts[i] != NULL) {
2149 2 : rc = ibv_query_device(contexts[i], &dev_attr);
2150 2 : if (rc < 0) {
2151 0 : SPDK_ERRLOG("Failed to query RDMA device attributes.\n");
2152 0 : rdma_free_devices(contexts);
2153 0 : spdk_free(rctrlr);
2154 0 : return NULL;
2155 : }
2156 2 : rctrlr->max_sge = spdk_min(rctrlr->max_sge, (uint16_t)dev_attr.max_sge);
2157 2 : i++;
2158 : }
2159 :
2160 1 : rdma_free_devices(contexts);
2161 :
2162 1 : rc = nvme_ctrlr_construct(&rctrlr->ctrlr);
2163 1 : if (rc != 0) {
2164 0 : spdk_free(rctrlr);
2165 0 : return NULL;
2166 : }
2167 :
2168 1 : STAILQ_INIT(&rctrlr->pending_cm_events);
2169 1 : STAILQ_INIT(&rctrlr->free_cm_events);
2170 1 : rctrlr->cm_events = spdk_zmalloc(NVME_RDMA_NUM_CM_EVENTS * sizeof(*rctrlr->cm_events), 0, NULL,
2171 : SPDK_ENV_NUMA_ID_ANY, SPDK_MALLOC_DMA);
2172 1 : if (rctrlr->cm_events == NULL) {
2173 0 : SPDK_ERRLOG("unable to allocate buffers to hold CM events.\n");
2174 0 : goto destruct_ctrlr;
2175 : }
2176 :
2177 257 : for (i = 0; i < NVME_RDMA_NUM_CM_EVENTS; i++) {
2178 256 : STAILQ_INSERT_TAIL(&rctrlr->free_cm_events, &rctrlr->cm_events[i], link);
2179 : }
2180 :
2181 1 : rctrlr->cm_channel = rdma_create_event_channel();
2182 1 : if (rctrlr->cm_channel == NULL) {
2183 0 : SPDK_ERRLOG("rdma_create_event_channel() failed\n");
2184 0 : goto destruct_ctrlr;
2185 : }
2186 :
2187 1 : flag = fcntl(rctrlr->cm_channel->fd, F_GETFL);
2188 1 : if (fcntl(rctrlr->cm_channel->fd, F_SETFL, flag | O_NONBLOCK) < 0) {
2189 0 : SPDK_ERRLOG("Cannot set event channel to non blocking\n");
2190 0 : goto destruct_ctrlr;
2191 : }
2192 :
2193 1 : rctrlr->ctrlr.adminq = nvme_rdma_ctrlr_create_qpair(&rctrlr->ctrlr, 0,
2194 1 : rctrlr->ctrlr.opts.admin_queue_size, 0,
2195 1 : rctrlr->ctrlr.opts.admin_queue_size, false, true);
2196 1 : if (!rctrlr->ctrlr.adminq) {
2197 0 : SPDK_ERRLOG("failed to create admin qpair\n");
2198 0 : goto destruct_ctrlr;
2199 : }
2200 :
2201 1 : if (nvme_ctrlr_add_process(&rctrlr->ctrlr, 0) != 0) {
2202 0 : SPDK_ERRLOG("nvme_ctrlr_add_process() failed\n");
2203 0 : goto destruct_ctrlr;
2204 : }
2205 :
2206 1 : SPDK_DEBUGLOG(nvme, "successfully initialized the nvmf ctrlr\n");
2207 1 : return &rctrlr->ctrlr;
2208 :
2209 0 : destruct_ctrlr:
2210 0 : nvme_ctrlr_destruct(&rctrlr->ctrlr);
2211 0 : return NULL;
2212 : }
2213 :
2214 : static int
2215 1 : nvme_rdma_ctrlr_destruct(struct spdk_nvme_ctrlr *ctrlr)
2216 : {
2217 1 : struct nvme_rdma_ctrlr *rctrlr = nvme_rdma_ctrlr(ctrlr);
2218 : struct nvme_rdma_cm_event_entry *entry;
2219 :
2220 1 : if (ctrlr->adminq) {
2221 1 : nvme_rdma_ctrlr_delete_io_qpair(ctrlr, ctrlr->adminq);
2222 : }
2223 :
2224 1 : STAILQ_FOREACH(entry, &rctrlr->pending_cm_events, link) {
2225 0 : rdma_ack_cm_event(entry->evt);
2226 : }
2227 :
2228 1 : STAILQ_INIT(&rctrlr->free_cm_events);
2229 1 : STAILQ_INIT(&rctrlr->pending_cm_events);
2230 1 : spdk_free(rctrlr->cm_events);
2231 :
2232 1 : if (rctrlr->cm_channel) {
2233 1 : rdma_destroy_event_channel(rctrlr->cm_channel);
2234 1 : rctrlr->cm_channel = NULL;
2235 : }
2236 :
2237 1 : nvme_ctrlr_destruct_finish(ctrlr);
2238 :
2239 1 : spdk_free(rctrlr);
2240 :
2241 1 : return 0;
2242 : }
2243 :
2244 : static inline int
2245 1 : _nvme_rdma_qpair_submit_request(struct nvme_rdma_qpair *rqpair,
2246 : struct spdk_nvme_rdma_req *rdma_req)
2247 : {
2248 1 : struct spdk_nvme_qpair *qpair = &rqpair->qpair;
2249 : struct ibv_send_wr *wr;
2250 : struct nvme_rdma_poll_group *group;
2251 :
2252 1 : if (!rqpair->link_active.tqe_prev && qpair->poll_group) {
2253 0 : group = nvme_rdma_poll_group(qpair->poll_group);
2254 0 : TAILQ_INSERT_TAIL(&group->active_qpairs, rqpair, link_active);
2255 : }
2256 1 : assert(rqpair->current_num_sends < rqpair->num_entries);
2257 1 : rqpair->current_num_sends++;
2258 :
2259 1 : wr = &rdma_req->send_wr;
2260 1 : wr->next = NULL;
2261 1 : nvme_rdma_trace_ibv_sge(wr->sg_list);
2262 :
2263 1 : spdk_rdma_provider_qp_queue_send_wrs(rqpair->rdma_qp, wr);
2264 :
2265 1 : if (!rqpair->delay_cmd_submit) {
2266 1 : return nvme_rdma_qpair_submit_sends(rqpair);
2267 : }
2268 :
2269 0 : return 0;
2270 : }
2271 :
2272 : static int
2273 2 : nvme_rdma_qpair_submit_request(struct spdk_nvme_qpair *qpair,
2274 : struct nvme_request *req)
2275 : {
2276 : struct nvme_rdma_qpair *rqpair;
2277 : struct spdk_nvme_rdma_req *rdma_req;
2278 : int rc;
2279 :
2280 2 : rqpair = nvme_rdma_qpair(qpair);
2281 2 : assert(rqpair != NULL);
2282 2 : assert(req != NULL);
2283 :
2284 2 : rdma_req = nvme_rdma_req_get(rqpair);
2285 2 : if (spdk_unlikely(!rdma_req)) {
2286 1 : if (rqpair->poller) {
2287 1 : rqpair->poller->stats.queued_requests++;
2288 : }
2289 : /* Inform the upper layer to try again later. */
2290 1 : return -EAGAIN;
2291 : }
2292 :
2293 1 : assert(rdma_req->req == NULL);
2294 1 : rdma_req->req = req;
2295 1 : req->cmd.cid = rdma_req->id;
2296 1 : rc = nvme_rdma_req_init(rqpair, rdma_req);
2297 1 : if (spdk_unlikely(rc)) {
2298 0 : SPDK_ERRLOG("nvme_rdma_req_init() failed\n");
2299 0 : nvme_rdma_req_put(rqpair, rdma_req);
2300 0 : return -1;
2301 : }
2302 :
2303 1 : TAILQ_INSERT_TAIL(&rqpair->outstanding_reqs, rdma_req, link);
2304 1 : rqpair->num_outstanding_reqs++;
2305 :
2306 1 : return _nvme_rdma_qpair_submit_request(rqpair, rdma_req);
2307 : }
2308 :
2309 : static int
2310 0 : nvme_rdma_qpair_reset(struct spdk_nvme_qpair *qpair)
2311 : {
2312 : /* Currently, doing nothing here */
2313 0 : return 0;
2314 : }
2315 :
2316 : static void
2317 2 : nvme_rdma_qpair_abort_reqs(struct spdk_nvme_qpair *qpair, uint32_t dnr)
2318 : {
2319 : struct spdk_nvme_rdma_req *rdma_req, *tmp;
2320 : struct spdk_nvme_cpl cpl;
2321 2 : struct nvme_rdma_qpair *rqpair = nvme_rdma_qpair(qpair);
2322 :
2323 2 : cpl.sqid = qpair->id;
2324 2 : cpl.status.sc = SPDK_NVME_SC_ABORTED_SQ_DELETION;
2325 2 : cpl.status.sct = SPDK_NVME_SCT_GENERIC;
2326 2 : cpl.status.dnr = dnr;
2327 :
2328 : /*
2329 : * We cannot abort requests at the RDMA layer without
2330 : * unregistering them. If we do, we can still get error
2331 : * free completions on the shared completion queue.
2332 : */
2333 2 : if (nvme_qpair_get_state(qpair) > NVME_QPAIR_DISCONNECTING &&
2334 0 : nvme_qpair_get_state(qpair) != NVME_QPAIR_DESTROYING) {
2335 0 : nvme_ctrlr_disconnect_qpair(qpair);
2336 : }
2337 :
2338 2 : TAILQ_FOREACH_SAFE(rdma_req, &rqpair->outstanding_reqs, link, tmp) {
2339 0 : nvme_rdma_req_complete(rdma_req, &cpl, true);
2340 : }
2341 2 : }
2342 :
2343 : static void
2344 0 : nvme_rdma_qpair_check_timeout(struct spdk_nvme_qpair *qpair)
2345 : {
2346 : uint64_t t02;
2347 : struct spdk_nvme_rdma_req *rdma_req, *tmp;
2348 0 : struct nvme_rdma_qpair *rqpair = nvme_rdma_qpair(qpair);
2349 0 : struct spdk_nvme_ctrlr *ctrlr = qpair->ctrlr;
2350 : struct spdk_nvme_ctrlr_process *active_proc;
2351 :
2352 : /* Don't check timeouts during controller initialization. */
2353 0 : if (ctrlr->state != NVME_CTRLR_STATE_READY) {
2354 0 : return;
2355 : }
2356 :
2357 0 : if (nvme_qpair_is_admin_queue(qpair)) {
2358 0 : active_proc = nvme_ctrlr_get_current_process(ctrlr);
2359 : } else {
2360 0 : active_proc = qpair->active_proc;
2361 : }
2362 :
2363 : /* Only check timeouts if the current process has a timeout callback. */
2364 0 : if (active_proc == NULL || active_proc->timeout_cb_fn == NULL) {
2365 0 : return;
2366 : }
2367 :
2368 0 : t02 = spdk_get_ticks();
2369 0 : TAILQ_FOREACH_SAFE(rdma_req, &rqpair->outstanding_reqs, link, tmp) {
2370 0 : assert(rdma_req->req != NULL);
2371 :
2372 0 : if (nvme_request_check_timeout(rdma_req->req, rdma_req->id, active_proc, t02)) {
2373 : /*
2374 : * The requests are in order, so as soon as one has not timed out,
2375 : * stop iterating.
2376 : */
2377 0 : break;
2378 : }
2379 : }
2380 : }
2381 :
2382 : static inline void
2383 0 : nvme_rdma_request_ready(struct nvme_rdma_qpair *rqpair, struct spdk_nvme_rdma_req *rdma_req)
2384 : {
2385 0 : struct spdk_nvme_rdma_rsp *rdma_rsp = rdma_req->rdma_rsp;
2386 0 : struct ibv_recv_wr *recv_wr = rdma_rsp->recv_wr;
2387 :
2388 0 : nvme_rdma_req_complete(rdma_req, &rdma_rsp->cpl, true);
2389 :
2390 0 : assert(rqpair->rsps->current_num_recvs < rqpair->rsps->num_entries);
2391 0 : rqpair->rsps->current_num_recvs++;
2392 :
2393 0 : recv_wr->next = NULL;
2394 0 : nvme_rdma_trace_ibv_sge(recv_wr->sg_list);
2395 :
2396 0 : if (!rqpair->srq) {
2397 0 : spdk_rdma_provider_qp_queue_recv_wrs(rqpair->rdma_qp, recv_wr);
2398 : } else {
2399 0 : spdk_rdma_provider_srq_queue_recv_wrs(rqpair->srq, recv_wr);
2400 : }
2401 0 : }
2402 :
2403 : #define MAX_COMPLETIONS_PER_POLL 128
2404 :
2405 : static void
2406 0 : nvme_rdma_fail_qpair(struct spdk_nvme_qpair *qpair, int failure_reason)
2407 : {
2408 0 : if (failure_reason == IBV_WC_RETRY_EXC_ERR) {
2409 0 : qpair->transport_failure_reason = SPDK_NVME_QPAIR_FAILURE_REMOTE;
2410 0 : } else if (qpair->transport_failure_reason == SPDK_NVME_QPAIR_FAILURE_NONE) {
2411 0 : qpair->transport_failure_reason = SPDK_NVME_QPAIR_FAILURE_UNKNOWN;
2412 : }
2413 :
2414 0 : nvme_ctrlr_disconnect_qpair(qpair);
2415 0 : }
2416 :
2417 : static struct nvme_rdma_qpair *
2418 4 : get_rdma_qpair_from_wc(struct nvme_rdma_poll_group *group, struct ibv_wc *wc)
2419 : {
2420 : struct spdk_nvme_qpair *qpair;
2421 : struct nvme_rdma_qpair *rqpair;
2422 :
2423 5 : STAILQ_FOREACH(qpair, &group->group.connected_qpairs, poll_group_stailq) {
2424 2 : rqpair = nvme_rdma_qpair(qpair);
2425 2 : if (NVME_RDMA_POLL_GROUP_CHECK_QPN(rqpair, wc->qp_num)) {
2426 1 : return rqpair;
2427 : }
2428 : }
2429 :
2430 4 : STAILQ_FOREACH(qpair, &group->group.disconnected_qpairs, poll_group_stailq) {
2431 2 : rqpair = nvme_rdma_qpair(qpair);
2432 2 : if (NVME_RDMA_POLL_GROUP_CHECK_QPN(rqpair, wc->qp_num)) {
2433 1 : return rqpair;
2434 : }
2435 : }
2436 :
2437 2 : return NULL;
2438 : }
2439 :
2440 : static inline void
2441 0 : nvme_rdma_log_wc_status(struct nvme_rdma_qpair *rqpair, struct ibv_wc *wc)
2442 : {
2443 0 : struct nvme_rdma_wr *rdma_wr = (struct nvme_rdma_wr *)wc->wr_id;
2444 :
2445 0 : if (wc->status == IBV_WC_WR_FLUSH_ERR) {
2446 : /* If qpair is in ERR state, we will receive completions for all posted and not completed
2447 : * Work Requests with IBV_WC_WR_FLUSH_ERR status. Don't log an error in that case */
2448 0 : SPDK_DEBUGLOG(nvme, "WC error, qid %u, qp state %d, request 0x%lu type %d, status: (%d): %s\n",
2449 : rqpair->qpair.id, rqpair->qpair.state, wc->wr_id, rdma_wr->type, wc->status,
2450 : ibv_wc_status_str(wc->status));
2451 : } else {
2452 0 : SPDK_ERRLOG("WC error, qid %u, qp state %d, request 0x%lu type %d, status: (%d): %s\n",
2453 : rqpair->qpair.id, rqpair->qpair.state, wc->wr_id, rdma_wr->type, wc->status,
2454 : ibv_wc_status_str(wc->status));
2455 : }
2456 0 : }
2457 :
2458 : static inline int
2459 0 : nvme_rdma_process_recv_completion(struct nvme_rdma_poller *poller, struct ibv_wc *wc,
2460 : struct nvme_rdma_wr *rdma_wr)
2461 : {
2462 : struct nvme_rdma_qpair *rqpair;
2463 : struct spdk_nvme_rdma_req *rdma_req;
2464 : struct spdk_nvme_rdma_rsp *rdma_rsp;
2465 :
2466 0 : rdma_rsp = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvme_rdma_rsp, rdma_wr);
2467 :
2468 0 : if (poller && poller->srq) {
2469 0 : rqpair = get_rdma_qpair_from_wc(poller->group, wc);
2470 0 : if (spdk_unlikely(!rqpair)) {
2471 : /* Since we do not handle the LAST_WQE_REACHED event, we do not know when
2472 : * a Receive Queue in a QP, that is associated with an SRQ, is flushed.
2473 : * We may get a WC for a already destroyed QP.
2474 : *
2475 : * However, for the SRQ, this is not any error. Hence, just re-post the
2476 : * receive request to the SRQ to reuse for other QPs, and return 0.
2477 : */
2478 0 : spdk_rdma_provider_srq_queue_recv_wrs(poller->srq, rdma_rsp->recv_wr);
2479 0 : return 0;
2480 : }
2481 : } else {
2482 0 : rqpair = rdma_rsp->rqpair;
2483 0 : if (spdk_unlikely(!rqpair)) {
2484 : /* TODO: Fix forceful QP destroy when it is not async mode.
2485 : * CQ itself did not cause any error. Hence, return 0 for now.
2486 : */
2487 0 : SPDK_WARNLOG("QP might be already destroyed.\n");
2488 0 : return 0;
2489 : }
2490 : }
2491 :
2492 :
2493 0 : assert(rqpair->rsps->current_num_recvs > 0);
2494 0 : rqpair->rsps->current_num_recvs--;
2495 :
2496 0 : if (spdk_unlikely(wc->status)) {
2497 0 : nvme_rdma_log_wc_status(rqpair, wc);
2498 0 : goto err_wc;
2499 : }
2500 :
2501 0 : SPDK_DEBUGLOG(nvme, "CQ recv completion\n");
2502 :
2503 0 : if (spdk_unlikely(wc->byte_len < sizeof(struct spdk_nvme_cpl))) {
2504 0 : SPDK_ERRLOG("recv length %u less than expected response size\n", wc->byte_len);
2505 0 : goto err_wc;
2506 : }
2507 0 : rdma_req = &rqpair->rdma_reqs[rdma_rsp->cpl.cid];
2508 0 : rdma_req->completion_flags |= NVME_RDMA_RECV_COMPLETED;
2509 0 : rdma_req->rdma_rsp = rdma_rsp;
2510 :
2511 0 : if ((rdma_req->completion_flags & NVME_RDMA_SEND_COMPLETED) == 0) {
2512 0 : return 0;
2513 : }
2514 :
2515 0 : rqpair->num_completions++;
2516 :
2517 0 : nvme_rdma_request_ready(rqpair, rdma_req);
2518 :
2519 0 : if (!rqpair->delay_cmd_submit) {
2520 0 : if (spdk_unlikely(nvme_rdma_qpair_submit_recvs(rqpair))) {
2521 0 : SPDK_ERRLOG("Unable to re-post rx descriptor\n");
2522 0 : nvme_rdma_fail_qpair(&rqpair->qpair, 0);
2523 0 : return -ENXIO;
2524 : }
2525 : }
2526 :
2527 0 : return 1;
2528 :
2529 0 : err_wc:
2530 0 : nvme_rdma_fail_qpair(&rqpair->qpair, 0);
2531 0 : if (poller && poller->srq) {
2532 0 : spdk_rdma_provider_srq_queue_recv_wrs(poller->srq, rdma_rsp->recv_wr);
2533 : }
2534 0 : return -ENXIO;
2535 : }
2536 :
2537 : static inline int
2538 0 : nvme_rdma_process_send_completion(struct nvme_rdma_poller *poller,
2539 : struct nvme_rdma_qpair *rdma_qpair,
2540 : struct ibv_wc *wc, struct nvme_rdma_wr *rdma_wr)
2541 : {
2542 : struct nvme_rdma_qpair *rqpair;
2543 : struct spdk_nvme_rdma_req *rdma_req;
2544 :
2545 0 : rdma_req = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvme_rdma_req, rdma_wr);
2546 0 : rqpair = rdma_req->req ? nvme_rdma_qpair(rdma_req->req->qpair) : NULL;
2547 0 : if (spdk_unlikely(!rqpair)) {
2548 0 : rqpair = rdma_qpair != NULL ? rdma_qpair : get_rdma_qpair_from_wc(poller->group, wc);
2549 : }
2550 :
2551 : /* If we are flushing I/O */
2552 0 : if (spdk_unlikely(wc->status)) {
2553 0 : if (!rqpair) {
2554 : /* When poll_group is used, several qpairs share the same CQ and it is possible to
2555 : * receive a completion with error (e.g. IBV_WC_WR_FLUSH_ERR) for already disconnected qpair
2556 : * That happens due to qpair is destroyed while there are submitted but not completed send/receive
2557 : * Work Requests */
2558 0 : assert(poller);
2559 0 : return 0;
2560 : }
2561 0 : assert(rqpair->current_num_sends > 0);
2562 0 : rqpair->current_num_sends--;
2563 0 : nvme_rdma_log_wc_status(rqpair, wc);
2564 0 : nvme_rdma_fail_qpair(&rqpair->qpair, 0);
2565 0 : if (rdma_req->rdma_rsp && poller && poller->srq) {
2566 0 : spdk_rdma_provider_srq_queue_recv_wrs(poller->srq, rdma_req->rdma_rsp->recv_wr);
2567 : }
2568 0 : return -ENXIO;
2569 : }
2570 :
2571 : /* We do not support Soft Roce anymore. Other than Soft Roce's bug, we should not
2572 : * receive a completion without error status after qpair is disconnected/destroyed.
2573 : */
2574 0 : if (spdk_unlikely(rdma_req->req == NULL)) {
2575 : /*
2576 : * Some infiniband drivers do not guarantee the previous assumption after we
2577 : * received a RDMA_CM_EVENT_DEVICE_REMOVAL event.
2578 : */
2579 0 : SPDK_ERRLOG("Received malformed completion: request 0x%"PRIx64" type %d\n", wc->wr_id,
2580 : rdma_wr->type);
2581 0 : if (!rqpair || !rqpair->need_destroy) {
2582 0 : assert(0);
2583 : }
2584 0 : return -ENXIO;
2585 : }
2586 :
2587 0 : rdma_req->completion_flags |= NVME_RDMA_SEND_COMPLETED;
2588 0 : assert(rqpair->current_num_sends > 0);
2589 0 : rqpair->current_num_sends--;
2590 :
2591 0 : if ((rdma_req->completion_flags & NVME_RDMA_RECV_COMPLETED) == 0) {
2592 0 : return 0;
2593 : }
2594 :
2595 0 : rqpair->num_completions++;
2596 :
2597 0 : nvme_rdma_request_ready(rqpair, rdma_req);
2598 :
2599 0 : if (!rqpair->delay_cmd_submit) {
2600 0 : if (spdk_unlikely(nvme_rdma_qpair_submit_recvs(rqpair))) {
2601 0 : SPDK_ERRLOG("Unable to re-post rx descriptor\n");
2602 0 : nvme_rdma_fail_qpair(&rqpair->qpair, 0);
2603 0 : return -ENXIO;
2604 : }
2605 : }
2606 :
2607 0 : return 1;
2608 : }
2609 :
2610 : static inline int
2611 0 : nvme_rdma_cq_process_completions(struct ibv_cq *cq, uint32_t batch_size,
2612 : struct nvme_rdma_poller *poller,
2613 : struct nvme_rdma_qpair *rdma_qpair,
2614 : uint64_t *rdma_completions)
2615 : {
2616 : struct ibv_wc wc[MAX_COMPLETIONS_PER_POLL];
2617 : struct nvme_rdma_wr *rdma_wr;
2618 0 : uint32_t reaped = 0;
2619 0 : int completion_rc = 0;
2620 : int rc, _rc, i;
2621 :
2622 0 : rc = ibv_poll_cq(cq, batch_size, wc);
2623 0 : if (spdk_unlikely(rc < 0)) {
2624 0 : SPDK_ERRLOG("Error polling CQ! (%d): %s\n",
2625 : errno, spdk_strerror(errno));
2626 0 : return -ECANCELED;
2627 0 : } else if (rc == 0) {
2628 0 : return 0;
2629 : }
2630 :
2631 0 : for (i = 0; i < rc; i++) {
2632 0 : rdma_wr = (struct nvme_rdma_wr *)wc[i].wr_id;
2633 0 : switch (rdma_wr->type) {
2634 0 : case RDMA_WR_TYPE_RECV:
2635 0 : _rc = nvme_rdma_process_recv_completion(poller, &wc[i], rdma_wr);
2636 0 : break;
2637 :
2638 0 : case RDMA_WR_TYPE_SEND:
2639 0 : _rc = nvme_rdma_process_send_completion(poller, rdma_qpair, &wc[i], rdma_wr);
2640 0 : break;
2641 :
2642 0 : default:
2643 0 : SPDK_ERRLOG("Received an unexpected opcode on the CQ: %d\n", rdma_wr->type);
2644 0 : return -ECANCELED;
2645 : }
2646 0 : if (spdk_likely(_rc >= 0)) {
2647 0 : reaped += _rc;
2648 : } else {
2649 0 : completion_rc = _rc;
2650 : }
2651 : }
2652 :
2653 0 : *rdma_completions += rc;
2654 :
2655 0 : if (spdk_unlikely(completion_rc)) {
2656 0 : return completion_rc;
2657 : }
2658 :
2659 0 : return reaped;
2660 : }
2661 :
2662 : static void
2663 0 : dummy_disconnected_qpair_cb(struct spdk_nvme_qpair *qpair, void *poll_group_ctx)
2664 : {
2665 :
2666 0 : }
2667 :
2668 : static int
2669 0 : nvme_rdma_qpair_process_completions(struct spdk_nvme_qpair *qpair,
2670 : uint32_t max_completions)
2671 : {
2672 0 : struct nvme_rdma_qpair *rqpair = nvme_rdma_qpair(qpair);
2673 0 : struct nvme_rdma_ctrlr *rctrlr = nvme_rdma_ctrlr(qpair->ctrlr);
2674 0 : int rc = 0, batch_size;
2675 : struct ibv_cq *cq;
2676 0 : uint64_t rdma_completions = 0;
2677 :
2678 : /*
2679 : * This is used during the connection phase. It's possible that we are still reaping error completions
2680 : * from other qpairs so we need to call the poll group function. Also, it's more correct since the cq
2681 : * is shared.
2682 : */
2683 0 : if (qpair->poll_group != NULL) {
2684 0 : return spdk_nvme_poll_group_process_completions(qpair->poll_group->group, max_completions,
2685 : dummy_disconnected_qpair_cb);
2686 : }
2687 :
2688 0 : if (max_completions == 0) {
2689 0 : max_completions = rqpair->num_entries;
2690 : } else {
2691 0 : max_completions = spdk_min(max_completions, rqpair->num_entries);
2692 : }
2693 :
2694 0 : switch (nvme_qpair_get_state(qpair)) {
2695 0 : case NVME_QPAIR_CONNECTING:
2696 0 : rc = nvme_rdma_ctrlr_connect_qpair_poll(qpair->ctrlr, qpair);
2697 0 : if (rc == 0) {
2698 : /* Once the connection is completed, we can submit queued requests */
2699 0 : nvme_qpair_resubmit_requests(qpair, rqpair->num_entries);
2700 0 : } else if (rc != -EAGAIN) {
2701 0 : SPDK_ERRLOG("Failed to connect rqpair=%p\n", rqpair);
2702 0 : goto failed;
2703 0 : } else if (rqpair->state <= NVME_RDMA_QPAIR_STATE_INITIALIZING) {
2704 0 : return 0;
2705 : }
2706 0 : break;
2707 :
2708 0 : case NVME_QPAIR_DISCONNECTING:
2709 0 : nvme_rdma_ctrlr_disconnect_qpair_poll(qpair->ctrlr, qpair);
2710 0 : return -ENXIO;
2711 :
2712 0 : default:
2713 0 : if (nvme_qpair_is_admin_queue(qpair)) {
2714 0 : nvme_rdma_poll_events(rctrlr);
2715 : }
2716 0 : nvme_rdma_qpair_process_cm_event(rqpair);
2717 0 : break;
2718 : }
2719 :
2720 0 : if (spdk_unlikely(qpair->transport_failure_reason != SPDK_NVME_QPAIR_FAILURE_NONE)) {
2721 0 : goto failed;
2722 : }
2723 :
2724 0 : cq = rqpair->cq;
2725 :
2726 0 : rqpair->num_completions = 0;
2727 : do {
2728 0 : batch_size = spdk_min((max_completions - rqpair->num_completions), MAX_COMPLETIONS_PER_POLL);
2729 0 : rc = nvme_rdma_cq_process_completions(cq, batch_size, NULL, rqpair, &rdma_completions);
2730 :
2731 0 : if (rc == 0) {
2732 0 : break;
2733 : /* Handle the case where we fail to poll the cq. */
2734 0 : } else if (rc == -ECANCELED) {
2735 0 : goto failed;
2736 0 : } else if (rc == -ENXIO) {
2737 0 : return rc;
2738 : }
2739 0 : } while (rqpair->num_completions < max_completions);
2740 :
2741 0 : if (spdk_unlikely(nvme_rdma_qpair_submit_sends(rqpair) ||
2742 : nvme_rdma_qpair_submit_recvs(rqpair))) {
2743 0 : goto failed;
2744 : }
2745 :
2746 0 : if (spdk_unlikely(qpair->ctrlr->timeout_enabled)) {
2747 0 : nvme_rdma_qpair_check_timeout(qpair);
2748 : }
2749 :
2750 0 : return rqpair->num_completions;
2751 :
2752 0 : failed:
2753 0 : nvme_rdma_fail_qpair(qpair, 0);
2754 0 : return -ENXIO;
2755 : }
2756 :
2757 : static uint32_t
2758 0 : nvme_rdma_ctrlr_get_max_xfer_size(struct spdk_nvme_ctrlr *ctrlr)
2759 : {
2760 : /* max_mr_size by ibv_query_device indicates the largest value that we can
2761 : * set for a registered memory region. It is independent from the actual
2762 : * I/O size and is very likely to be larger than 2 MiB which is the
2763 : * granularity we currently register memory regions. Hence return
2764 : * UINT32_MAX here and let the generic layer use the controller data to
2765 : * moderate this value.
2766 : */
2767 0 : return UINT32_MAX;
2768 : }
2769 :
2770 : static uint16_t
2771 5 : nvme_rdma_ctrlr_get_max_sges(struct spdk_nvme_ctrlr *ctrlr)
2772 : {
2773 5 : struct nvme_rdma_ctrlr *rctrlr = nvme_rdma_ctrlr(ctrlr);
2774 5 : uint32_t max_sge = rctrlr->max_sge;
2775 5 : uint32_t max_in_capsule_sge = (ctrlr->cdata.nvmf_specific.ioccsz * 16 -
2776 5 : sizeof(struct spdk_nvme_cmd)) /
2777 : sizeof(struct spdk_nvme_sgl_descriptor);
2778 :
2779 : /* Max SGE is limited by capsule size */
2780 5 : max_sge = spdk_min(max_sge, max_in_capsule_sge);
2781 : /* Max SGE may be limited by MSDBD */
2782 5 : if (ctrlr->cdata.nvmf_specific.msdbd != 0) {
2783 5 : max_sge = spdk_min(max_sge, ctrlr->cdata.nvmf_specific.msdbd);
2784 : }
2785 :
2786 : /* Max SGE can't be less than 1 */
2787 5 : max_sge = spdk_max(1, max_sge);
2788 5 : return max_sge;
2789 : }
2790 :
2791 : static int
2792 0 : nvme_rdma_qpair_iterate_requests(struct spdk_nvme_qpair *qpair,
2793 : int (*iter_fn)(struct nvme_request *req, void *arg),
2794 : void *arg)
2795 : {
2796 0 : struct nvme_rdma_qpair *rqpair = nvme_rdma_qpair(qpair);
2797 : struct spdk_nvme_rdma_req *rdma_req, *tmp;
2798 : int rc;
2799 :
2800 0 : assert(iter_fn != NULL);
2801 :
2802 0 : TAILQ_FOREACH_SAFE(rdma_req, &rqpair->outstanding_reqs, link, tmp) {
2803 0 : assert(rdma_req->req != NULL);
2804 :
2805 0 : rc = iter_fn(rdma_req->req, arg);
2806 0 : if (rc != 0) {
2807 0 : return rc;
2808 : }
2809 : }
2810 :
2811 0 : return 0;
2812 : }
2813 :
2814 : static int
2815 0 : nvme_rdma_qpair_authenticate(struct spdk_nvme_qpair *qpair)
2816 : {
2817 0 : struct nvme_rdma_qpair *rqpair = nvme_rdma_qpair(qpair);
2818 : int rc;
2819 :
2820 : /* If the qpair is still connecting, it'll be forced to authenticate later on */
2821 0 : if (rqpair->state < NVME_RDMA_QPAIR_STATE_RUNNING) {
2822 0 : return 0;
2823 0 : } else if (rqpair->state != NVME_RDMA_QPAIR_STATE_RUNNING) {
2824 0 : return -ENOTCONN;
2825 : }
2826 :
2827 0 : rc = nvme_fabric_qpair_authenticate_async(qpair);
2828 0 : if (rc == 0) {
2829 0 : nvme_qpair_set_state(qpair, NVME_QPAIR_CONNECTING);
2830 0 : rqpair->state = NVME_RDMA_QPAIR_STATE_AUTHENTICATING;
2831 : }
2832 :
2833 0 : return rc;
2834 : }
2835 :
2836 : static void
2837 0 : nvme_rdma_admin_qpair_abort_aers(struct spdk_nvme_qpair *qpair)
2838 : {
2839 : struct spdk_nvme_rdma_req *rdma_req, *tmp;
2840 : struct spdk_nvme_cpl cpl;
2841 0 : struct nvme_rdma_qpair *rqpair = nvme_rdma_qpair(qpair);
2842 :
2843 0 : cpl.status.sc = SPDK_NVME_SC_ABORTED_SQ_DELETION;
2844 0 : cpl.status.sct = SPDK_NVME_SCT_GENERIC;
2845 :
2846 0 : TAILQ_FOREACH_SAFE(rdma_req, &rqpair->outstanding_reqs, link, tmp) {
2847 0 : assert(rdma_req->req != NULL);
2848 :
2849 0 : if (rdma_req->req->cmd.opc != SPDK_NVME_OPC_ASYNC_EVENT_REQUEST) {
2850 0 : continue;
2851 : }
2852 :
2853 0 : nvme_rdma_req_complete(rdma_req, &cpl, false);
2854 : }
2855 0 : }
2856 :
2857 : static void
2858 9 : nvme_rdma_poller_destroy(struct nvme_rdma_poller *poller)
2859 : {
2860 9 : if (poller->cq) {
2861 7 : ibv_destroy_cq(poller->cq);
2862 : }
2863 9 : if (poller->rsps) {
2864 0 : nvme_rdma_free_rsps(poller->rsps);
2865 : }
2866 9 : if (poller->srq) {
2867 0 : spdk_rdma_provider_srq_destroy(poller->srq);
2868 : }
2869 9 : if (poller->mr_map) {
2870 0 : spdk_rdma_utils_free_mem_map(&poller->mr_map);
2871 : }
2872 9 : if (poller->pd) {
2873 0 : spdk_rdma_utils_put_pd(poller->pd);
2874 : }
2875 9 : free(poller);
2876 9 : }
2877 :
2878 : static struct nvme_rdma_poller *
2879 9 : nvme_rdma_poller_create(struct nvme_rdma_poll_group *group, struct ibv_context *ctx)
2880 : {
2881 : struct nvme_rdma_poller *poller;
2882 : struct ibv_device_attr dev_attr;
2883 9 : struct spdk_rdma_provider_srq_init_attr srq_init_attr = {};
2884 : struct nvme_rdma_rsp_opts opts;
2885 : int num_cqe, max_num_cqe;
2886 : int rc;
2887 :
2888 9 : poller = calloc(1, sizeof(*poller));
2889 9 : if (poller == NULL) {
2890 0 : SPDK_ERRLOG("Unable to allocate poller.\n");
2891 0 : return NULL;
2892 : }
2893 :
2894 9 : poller->group = group;
2895 9 : poller->device = ctx;
2896 :
2897 9 : if (g_spdk_nvme_transport_opts.rdma_srq_size != 0) {
2898 0 : rc = ibv_query_device(ctx, &dev_attr);
2899 0 : if (rc) {
2900 0 : SPDK_ERRLOG("Unable to query RDMA device.\n");
2901 0 : goto fail;
2902 : }
2903 :
2904 0 : poller->pd = spdk_rdma_utils_get_pd(ctx);
2905 0 : if (poller->pd == NULL) {
2906 0 : SPDK_ERRLOG("Unable to get PD.\n");
2907 0 : goto fail;
2908 : }
2909 :
2910 0 : poller->mr_map = spdk_rdma_utils_create_mem_map(poller->pd, &g_nvme_hooks,
2911 : IBV_ACCESS_LOCAL_WRITE | IBV_ACCESS_REMOTE_READ | IBV_ACCESS_REMOTE_WRITE);
2912 0 : if (poller->mr_map == NULL) {
2913 0 : SPDK_ERRLOG("Unable to create memory map.\n");
2914 0 : goto fail;
2915 : }
2916 :
2917 0 : srq_init_attr.stats = &poller->stats.rdma_stats.recv;
2918 0 : srq_init_attr.pd = poller->pd;
2919 0 : srq_init_attr.srq_init_attr.attr.max_wr = spdk_min((uint32_t)dev_attr.max_srq_wr,
2920 : g_spdk_nvme_transport_opts.rdma_srq_size);
2921 0 : srq_init_attr.srq_init_attr.attr.max_sge = spdk_min(dev_attr.max_sge,
2922 : NVME_RDMA_DEFAULT_RX_SGE);
2923 :
2924 0 : poller->srq = spdk_rdma_provider_srq_create(&srq_init_attr);
2925 0 : if (poller->srq == NULL) {
2926 0 : SPDK_ERRLOG("Unable to create SRQ.\n");
2927 0 : goto fail;
2928 : }
2929 :
2930 0 : opts.num_entries = g_spdk_nvme_transport_opts.rdma_srq_size;
2931 0 : opts.rqpair = NULL;
2932 0 : opts.srq = poller->srq;
2933 0 : opts.mr_map = poller->mr_map;
2934 :
2935 0 : poller->rsps = nvme_rdma_create_rsps(&opts);
2936 0 : if (poller->rsps == NULL) {
2937 0 : SPDK_ERRLOG("Unable to create poller RDMA responses.\n");
2938 0 : goto fail;
2939 : }
2940 :
2941 0 : rc = nvme_rdma_poller_submit_recvs(poller);
2942 0 : if (rc) {
2943 0 : SPDK_ERRLOG("Unable to submit poller RDMA responses.\n");
2944 0 : goto fail;
2945 : }
2946 :
2947 : /*
2948 : * When using an srq, fix the size of the completion queue at startup.
2949 : * The initiator sends only send and recv WRs. Hence, the multiplier is 2.
2950 : * (The target sends also data WRs. Hence, the multiplier is 3.)
2951 : */
2952 0 : num_cqe = g_spdk_nvme_transport_opts.rdma_srq_size * 2;
2953 : } else {
2954 9 : num_cqe = DEFAULT_NVME_RDMA_CQ_SIZE;
2955 : }
2956 :
2957 9 : max_num_cqe = g_spdk_nvme_transport_opts.rdma_max_cq_size;
2958 9 : if (max_num_cqe != 0 && num_cqe > max_num_cqe) {
2959 0 : num_cqe = max_num_cqe;
2960 : }
2961 :
2962 9 : poller->cq = ibv_create_cq(poller->device, num_cqe, group, NULL, 0);
2963 :
2964 9 : if (poller->cq == NULL) {
2965 2 : SPDK_ERRLOG("Unable to create CQ, errno %d.\n", errno);
2966 2 : goto fail;
2967 : }
2968 :
2969 7 : STAILQ_INSERT_HEAD(&group->pollers, poller, link);
2970 7 : group->num_pollers++;
2971 7 : poller->current_num_wc = num_cqe;
2972 7 : poller->required_num_wc = 0;
2973 7 : return poller;
2974 :
2975 2 : fail:
2976 2 : nvme_rdma_poller_destroy(poller);
2977 2 : return NULL;
2978 : }
2979 :
2980 : static void
2981 3 : nvme_rdma_poll_group_free_pollers(struct nvme_rdma_poll_group *group)
2982 : {
2983 : struct nvme_rdma_poller *poller, *tmp_poller;
2984 :
2985 5 : STAILQ_FOREACH_SAFE(poller, &group->pollers, link, tmp_poller) {
2986 2 : assert(poller->refcnt == 0);
2987 2 : if (poller->refcnt) {
2988 0 : SPDK_WARNLOG("Destroying poller with non-zero ref count: poller %p, refcnt %d\n",
2989 : poller, poller->refcnt);
2990 : }
2991 :
2992 2 : STAILQ_REMOVE(&group->pollers, poller, nvme_rdma_poller, link);
2993 2 : nvme_rdma_poller_destroy(poller);
2994 : }
2995 3 : }
2996 :
2997 : static struct nvme_rdma_poller *
2998 8 : nvme_rdma_poll_group_get_poller(struct nvme_rdma_poll_group *group, struct ibv_context *device)
2999 : {
3000 8 : struct nvme_rdma_poller *poller = NULL;
3001 :
3002 10 : STAILQ_FOREACH(poller, &group->pollers, link) {
3003 3 : if (poller->device == device) {
3004 1 : break;
3005 : }
3006 : }
3007 :
3008 8 : if (!poller) {
3009 7 : poller = nvme_rdma_poller_create(group, device);
3010 7 : if (!poller) {
3011 2 : SPDK_ERRLOG("Failed to create a poller for device %p\n", device);
3012 2 : return NULL;
3013 : }
3014 : }
3015 :
3016 6 : poller->refcnt++;
3017 6 : return poller;
3018 : }
3019 :
3020 : static void
3021 6 : nvme_rdma_poll_group_put_poller(struct nvme_rdma_poll_group *group, struct nvme_rdma_poller *poller)
3022 : {
3023 6 : assert(poller->refcnt > 0);
3024 6 : if (--poller->refcnt == 0) {
3025 5 : STAILQ_REMOVE(&group->pollers, poller, nvme_rdma_poller, link);
3026 5 : group->num_pollers--;
3027 5 : nvme_rdma_poller_destroy(poller);
3028 : }
3029 6 : }
3030 :
3031 : static struct spdk_nvme_transport_poll_group *
3032 1 : nvme_rdma_poll_group_create(void)
3033 : {
3034 : struct nvme_rdma_poll_group *group;
3035 :
3036 1 : group = calloc(1, sizeof(*group));
3037 1 : if (group == NULL) {
3038 0 : SPDK_ERRLOG("Unable to allocate poll group.\n");
3039 0 : return NULL;
3040 : }
3041 :
3042 1 : STAILQ_INIT(&group->pollers);
3043 1 : TAILQ_INIT(&group->connecting_qpairs);
3044 1 : TAILQ_INIT(&group->active_qpairs);
3045 1 : return &group->group;
3046 : }
3047 :
3048 : static int
3049 0 : nvme_rdma_poll_group_connect_qpair(struct spdk_nvme_qpair *qpair)
3050 : {
3051 0 : return 0;
3052 : }
3053 :
3054 : static int
3055 0 : nvme_rdma_poll_group_disconnect_qpair(struct spdk_nvme_qpair *qpair)
3056 : {
3057 0 : struct nvme_rdma_qpair *rqpair = nvme_rdma_qpair(qpair);
3058 0 : struct nvme_rdma_poll_group *group = nvme_rdma_poll_group(qpair->poll_group);
3059 :
3060 0 : if (rqpair->link_connecting.tqe_prev) {
3061 0 : TAILQ_REMOVE(&group->connecting_qpairs, rqpair, link_connecting);
3062 : /* We use prev pointer to check if qpair is in connecting list or not .
3063 : * TAILQ_REMOVE doesn't do it. So, we do it manually.
3064 : */
3065 0 : rqpair->link_connecting.tqe_prev = NULL;
3066 : }
3067 :
3068 0 : return 0;
3069 : }
3070 :
3071 : static int
3072 0 : nvme_rdma_poll_group_add(struct spdk_nvme_transport_poll_group *tgroup,
3073 : struct spdk_nvme_qpair *qpair)
3074 : {
3075 0 : return 0;
3076 : }
3077 :
3078 : static int
3079 0 : nvme_rdma_poll_group_remove(struct spdk_nvme_transport_poll_group *tgroup,
3080 : struct spdk_nvme_qpair *qpair)
3081 : {
3082 0 : struct nvme_rdma_qpair *rqpair = nvme_rdma_qpair(qpair);
3083 0 : struct nvme_rdma_poll_group *group = nvme_rdma_poll_group(qpair->poll_group);
3084 :
3085 0 : if (rqpair->link_active.tqe_prev) {
3086 0 : TAILQ_REMOVE(&group->active_qpairs, rqpair, link_active);
3087 0 : rqpair->link_active.tqe_prev = NULL;
3088 : }
3089 :
3090 0 : return 0;
3091 : }
3092 :
3093 : static inline void
3094 0 : nvme_rdma_qpair_process_submits(struct nvme_rdma_poll_group *group,
3095 : struct nvme_rdma_qpair *rqpair)
3096 : {
3097 0 : struct spdk_nvme_qpair *qpair = &rqpair->qpair;
3098 :
3099 0 : assert(rqpair->link_active.tqe_prev != NULL);
3100 :
3101 0 : if (spdk_unlikely(rqpair->state <= NVME_RDMA_QPAIR_STATE_INITIALIZING ||
3102 : rqpair->state >= NVME_RDMA_QPAIR_STATE_EXITING)) {
3103 0 : return;
3104 : }
3105 :
3106 0 : if (spdk_unlikely(qpair->ctrlr->timeout_enabled)) {
3107 0 : nvme_rdma_qpair_check_timeout(qpair);
3108 : }
3109 :
3110 0 : nvme_rdma_qpair_submit_sends(rqpair);
3111 0 : if (!rqpair->srq) {
3112 0 : nvme_rdma_qpair_submit_recvs(rqpair);
3113 : }
3114 0 : if (rqpair->num_completions > 0) {
3115 0 : nvme_qpair_resubmit_requests(qpair, rqpair->num_completions);
3116 0 : rqpair->num_completions = 0;
3117 : }
3118 :
3119 0 : if (rqpair->num_outstanding_reqs == 0 && STAILQ_EMPTY(&qpair->queued_req)) {
3120 0 : TAILQ_REMOVE(&group->active_qpairs, rqpair, link_active);
3121 : /* We use prev pointer to check if qpair is in active list or not.
3122 : * TAILQ_REMOVE doesn't do it. So, we do it manually.
3123 : */
3124 0 : rqpair->link_active.tqe_prev = NULL;
3125 : }
3126 : }
3127 :
3128 : static int64_t
3129 0 : nvme_rdma_poll_group_process_completions(struct spdk_nvme_transport_poll_group *tgroup,
3130 : uint32_t completions_per_qpair, spdk_nvme_disconnected_qpair_cb disconnected_qpair_cb)
3131 : {
3132 : struct spdk_nvme_qpair *qpair, *tmp_qpair;
3133 : struct nvme_rdma_qpair *rqpair, *tmp_rqpair;
3134 : struct nvme_rdma_poll_group *group;
3135 : struct nvme_rdma_poller *poller;
3136 0 : int batch_size, rc, rc2 = 0;
3137 0 : int64_t total_completions = 0;
3138 0 : uint64_t completions_allowed = 0;
3139 0 : uint64_t completions_per_poller = 0;
3140 0 : uint64_t poller_completions = 0;
3141 : uint64_t rdma_completions;
3142 :
3143 0 : if (completions_per_qpair == 0) {
3144 0 : completions_per_qpair = MAX_COMPLETIONS_PER_POLL;
3145 : }
3146 :
3147 0 : group = nvme_rdma_poll_group(tgroup);
3148 :
3149 0 : STAILQ_FOREACH_SAFE(qpair, &tgroup->disconnected_qpairs, poll_group_stailq, tmp_qpair) {
3150 0 : rc = nvme_rdma_ctrlr_disconnect_qpair_poll(qpair->ctrlr, qpair);
3151 0 : if (rc == 0) {
3152 0 : disconnected_qpair_cb(qpair, tgroup->group->ctx);
3153 : }
3154 : }
3155 :
3156 0 : TAILQ_FOREACH_SAFE(rqpair, &group->connecting_qpairs, link_connecting, tmp_rqpair) {
3157 0 : qpair = &rqpair->qpair;
3158 :
3159 0 : rc = nvme_rdma_ctrlr_connect_qpair_poll(qpair->ctrlr, qpair);
3160 0 : if (rc == 0 || rc != -EAGAIN) {
3161 0 : TAILQ_REMOVE(&group->connecting_qpairs, rqpair, link_connecting);
3162 : /* We use prev pointer to check if qpair is in connecting list or not.
3163 : * TAILQ_REMOVE does not do it. So, we do it manually.
3164 : */
3165 0 : rqpair->link_connecting.tqe_prev = NULL;
3166 :
3167 0 : if (rc == 0) {
3168 : /* Once the connection is completed, we can submit queued requests */
3169 0 : nvme_qpair_resubmit_requests(qpair, rqpair->num_entries);
3170 0 : } else if (rc != -EAGAIN) {
3171 0 : SPDK_ERRLOG("Failed to connect rqpair=%p\n", rqpair);
3172 0 : nvme_rdma_fail_qpair(qpair, 0);
3173 : }
3174 : }
3175 : }
3176 :
3177 0 : STAILQ_FOREACH_SAFE(qpair, &tgroup->connected_qpairs, poll_group_stailq, tmp_qpair) {
3178 0 : rqpair = nvme_rdma_qpair(qpair);
3179 :
3180 0 : if (spdk_likely(nvme_qpair_get_state(qpair) != NVME_QPAIR_CONNECTING)) {
3181 0 : nvme_rdma_qpair_process_cm_event(rqpair);
3182 : }
3183 :
3184 0 : if (spdk_unlikely(qpair->transport_failure_reason != SPDK_NVME_QPAIR_FAILURE_NONE)) {
3185 0 : rc2 = -ENXIO;
3186 0 : nvme_rdma_fail_qpair(qpair, 0);
3187 : }
3188 : }
3189 :
3190 0 : completions_allowed = completions_per_qpair * tgroup->num_connected_qpairs;
3191 0 : if (spdk_likely(group->num_pollers)) {
3192 0 : completions_per_poller = spdk_max(completions_allowed / group->num_pollers, 1);
3193 : }
3194 :
3195 0 : STAILQ_FOREACH(poller, &group->pollers, link) {
3196 0 : poller_completions = 0;
3197 0 : rdma_completions = 0;
3198 : do {
3199 0 : poller->stats.polls++;
3200 0 : batch_size = spdk_min((completions_per_poller - poller_completions), MAX_COMPLETIONS_PER_POLL);
3201 0 : rc = nvme_rdma_cq_process_completions(poller->cq, batch_size, poller, NULL, &rdma_completions);
3202 0 : if (rc <= 0) {
3203 0 : if (rc == -ECANCELED) {
3204 0 : return -EIO;
3205 0 : } else if (rc == 0) {
3206 0 : poller->stats.idle_polls++;
3207 : }
3208 0 : break;
3209 : }
3210 :
3211 0 : poller_completions += rc;
3212 0 : } while (poller_completions < completions_per_poller);
3213 0 : total_completions += poller_completions;
3214 0 : poller->stats.completions += rdma_completions;
3215 0 : if (poller->srq) {
3216 0 : nvme_rdma_poller_submit_recvs(poller);
3217 : }
3218 : }
3219 :
3220 0 : TAILQ_FOREACH_SAFE(rqpair, &group->active_qpairs, link_active, tmp_rqpair) {
3221 0 : nvme_rdma_qpair_process_submits(group, rqpair);
3222 : }
3223 :
3224 0 : return rc2 != 0 ? rc2 : total_completions;
3225 : }
3226 :
3227 : /*
3228 : * Handle disconnected qpairs when interrupt support gets added.
3229 : */
3230 : static void
3231 0 : nvme_rdma_poll_group_check_disconnected_qpairs(struct spdk_nvme_transport_poll_group *tgroup,
3232 : spdk_nvme_disconnected_qpair_cb disconnected_qpair_cb)
3233 : {
3234 0 : }
3235 :
3236 : static int
3237 1 : nvme_rdma_poll_group_destroy(struct spdk_nvme_transport_poll_group *tgroup)
3238 : {
3239 1 : struct nvme_rdma_poll_group *group = nvme_rdma_poll_group(tgroup);
3240 :
3241 1 : if (!STAILQ_EMPTY(&tgroup->connected_qpairs) || !STAILQ_EMPTY(&tgroup->disconnected_qpairs)) {
3242 0 : return -EBUSY;
3243 : }
3244 :
3245 1 : nvme_rdma_poll_group_free_pollers(group);
3246 1 : free(group);
3247 :
3248 1 : return 0;
3249 : }
3250 :
3251 : static int
3252 3 : nvme_rdma_poll_group_get_stats(struct spdk_nvme_transport_poll_group *tgroup,
3253 : struct spdk_nvme_transport_poll_group_stat **_stats)
3254 : {
3255 : struct nvme_rdma_poll_group *group;
3256 : struct spdk_nvme_transport_poll_group_stat *stats;
3257 : struct spdk_nvme_rdma_device_stat *device_stat;
3258 : struct nvme_rdma_poller *poller;
3259 3 : uint32_t i = 0;
3260 :
3261 3 : if (tgroup == NULL || _stats == NULL) {
3262 2 : SPDK_ERRLOG("Invalid stats or group pointer\n");
3263 2 : return -EINVAL;
3264 : }
3265 :
3266 1 : group = nvme_rdma_poll_group(tgroup);
3267 1 : stats = calloc(1, sizeof(*stats));
3268 1 : if (!stats) {
3269 0 : SPDK_ERRLOG("Can't allocate memory for RDMA stats\n");
3270 0 : return -ENOMEM;
3271 : }
3272 1 : stats->trtype = SPDK_NVME_TRANSPORT_RDMA;
3273 1 : stats->rdma.num_devices = group->num_pollers;
3274 :
3275 1 : if (stats->rdma.num_devices == 0) {
3276 0 : *_stats = stats;
3277 0 : return 0;
3278 : }
3279 :
3280 1 : stats->rdma.device_stats = calloc(stats->rdma.num_devices, sizeof(*stats->rdma.device_stats));
3281 1 : if (!stats->rdma.device_stats) {
3282 0 : SPDK_ERRLOG("Can't allocate memory for RDMA device stats\n");
3283 0 : free(stats);
3284 0 : return -ENOMEM;
3285 : }
3286 :
3287 3 : STAILQ_FOREACH(poller, &group->pollers, link) {
3288 2 : device_stat = &stats->rdma.device_stats[i];
3289 2 : device_stat->name = poller->device->device->name;
3290 2 : device_stat->polls = poller->stats.polls;
3291 2 : device_stat->idle_polls = poller->stats.idle_polls;
3292 2 : device_stat->completions = poller->stats.completions;
3293 2 : device_stat->queued_requests = poller->stats.queued_requests;
3294 2 : device_stat->total_send_wrs = poller->stats.rdma_stats.send.num_submitted_wrs;
3295 2 : device_stat->send_doorbell_updates = poller->stats.rdma_stats.send.doorbell_updates;
3296 2 : device_stat->total_recv_wrs = poller->stats.rdma_stats.recv.num_submitted_wrs;
3297 2 : device_stat->recv_doorbell_updates = poller->stats.rdma_stats.recv.doorbell_updates;
3298 2 : i++;
3299 : }
3300 :
3301 1 : *_stats = stats;
3302 :
3303 1 : return 0;
3304 : }
3305 :
3306 : static void
3307 1 : nvme_rdma_poll_group_free_stats(struct spdk_nvme_transport_poll_group *tgroup,
3308 : struct spdk_nvme_transport_poll_group_stat *stats)
3309 : {
3310 1 : if (stats) {
3311 1 : free(stats->rdma.device_stats);
3312 : }
3313 1 : free(stats);
3314 1 : }
3315 :
3316 : static int
3317 4 : nvme_rdma_ctrlr_get_memory_domains(const struct spdk_nvme_ctrlr *ctrlr,
3318 : struct spdk_memory_domain **domains, int array_size)
3319 : {
3320 4 : struct nvme_rdma_qpair *rqpair = nvme_rdma_qpair(ctrlr->adminq);
3321 :
3322 4 : if (domains && array_size > 0) {
3323 1 : domains[0] = rqpair->rdma_qp->domain;
3324 : }
3325 :
3326 4 : return 1;
3327 : }
3328 :
3329 : void
3330 0 : spdk_nvme_rdma_init_hooks(struct spdk_nvme_rdma_hooks *hooks)
3331 : {
3332 0 : g_nvme_hooks = *hooks;
3333 0 : }
3334 :
3335 : const struct spdk_nvme_transport_ops rdma_ops = {
3336 : .name = "RDMA",
3337 : .type = SPDK_NVME_TRANSPORT_RDMA,
3338 : .ctrlr_construct = nvme_rdma_ctrlr_construct,
3339 : .ctrlr_scan = nvme_fabric_ctrlr_scan,
3340 : .ctrlr_destruct = nvme_rdma_ctrlr_destruct,
3341 : .ctrlr_enable = nvme_rdma_ctrlr_enable,
3342 :
3343 : .ctrlr_set_reg_4 = nvme_fabric_ctrlr_set_reg_4,
3344 : .ctrlr_set_reg_8 = nvme_fabric_ctrlr_set_reg_8,
3345 : .ctrlr_get_reg_4 = nvme_fabric_ctrlr_get_reg_4,
3346 : .ctrlr_get_reg_8 = nvme_fabric_ctrlr_get_reg_8,
3347 : .ctrlr_set_reg_4_async = nvme_fabric_ctrlr_set_reg_4_async,
3348 : .ctrlr_set_reg_8_async = nvme_fabric_ctrlr_set_reg_8_async,
3349 : .ctrlr_get_reg_4_async = nvme_fabric_ctrlr_get_reg_4_async,
3350 : .ctrlr_get_reg_8_async = nvme_fabric_ctrlr_get_reg_8_async,
3351 :
3352 : .ctrlr_get_max_xfer_size = nvme_rdma_ctrlr_get_max_xfer_size,
3353 : .ctrlr_get_max_sges = nvme_rdma_ctrlr_get_max_sges,
3354 :
3355 : .ctrlr_create_io_qpair = nvme_rdma_ctrlr_create_io_qpair,
3356 : .ctrlr_delete_io_qpair = nvme_rdma_ctrlr_delete_io_qpair,
3357 : .ctrlr_connect_qpair = nvme_rdma_ctrlr_connect_qpair,
3358 : .ctrlr_disconnect_qpair = nvme_rdma_ctrlr_disconnect_qpair,
3359 :
3360 : .ctrlr_get_memory_domains = nvme_rdma_ctrlr_get_memory_domains,
3361 :
3362 : .qpair_abort_reqs = nvme_rdma_qpair_abort_reqs,
3363 : .qpair_reset = nvme_rdma_qpair_reset,
3364 : .qpair_submit_request = nvme_rdma_qpair_submit_request,
3365 : .qpair_process_completions = nvme_rdma_qpair_process_completions,
3366 : .qpair_iterate_requests = nvme_rdma_qpair_iterate_requests,
3367 : .qpair_authenticate = nvme_rdma_qpair_authenticate,
3368 : .admin_qpair_abort_aers = nvme_rdma_admin_qpair_abort_aers,
3369 :
3370 : .poll_group_create = nvme_rdma_poll_group_create,
3371 : .poll_group_connect_qpair = nvme_rdma_poll_group_connect_qpair,
3372 : .poll_group_disconnect_qpair = nvme_rdma_poll_group_disconnect_qpair,
3373 : .poll_group_add = nvme_rdma_poll_group_add,
3374 : .poll_group_remove = nvme_rdma_poll_group_remove,
3375 : .poll_group_process_completions = nvme_rdma_poll_group_process_completions,
3376 : .poll_group_check_disconnected_qpairs = nvme_rdma_poll_group_check_disconnected_qpairs,
3377 : .poll_group_destroy = nvme_rdma_poll_group_destroy,
3378 : .poll_group_get_stats = nvme_rdma_poll_group_get_stats,
3379 : .poll_group_free_stats = nvme_rdma_poll_group_free_stats,
3380 : };
3381 :
3382 1 : SPDK_NVME_TRANSPORT_REGISTER(rdma, &rdma_ops);
|
