1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /*
3 * RDMA Transport Layer
4 *
5 * Copyright (c) 2014 - 2018 ProfitBricks GmbH. All rights reserved.
6 * Copyright (c) 2018 - 2019 1&1 IONOS Cloud GmbH. All rights reserved.
7 * Copyright (c) 2019 - 2020 1&1 IONOS SE. All rights reserved.
8 */
9
10 #undef pr_fmt
11 #define pr_fmt(fmt) KBUILD_MODNAME " L" __stringify(__LINE__) ": " fmt
12
13 #include <linux/module.h>
14 #include <linux/rculist.h>
15 #include <linux/random.h>
16
17 #include "rtrs-clt.h"
18 #include "rtrs-log.h"
19
20 #define RTRS_CONNECT_TIMEOUT_MS 30000
21 /*
22 * Wait a bit before trying to reconnect after a failure
23 * in order to give server time to finish clean up which
24 * leads to "false positives" failed reconnect attempts
25 */
26 #define RTRS_RECONNECT_BACKOFF 1000
27 /*
28 * Wait for additional random time between 0 and 8 seconds
29 * before starting to reconnect to avoid clients reconnecting
30 * all at once in case of a major network outage
31 */
32 #define RTRS_RECONNECT_SEED 8
33
34 #define FIRST_CONN 0x01
35 /* limit to 128 * 4k = 512k max IO */
36 #define RTRS_MAX_SEGMENTS 128
37
38 MODULE_DESCRIPTION("RDMA Transport Client");
39 MODULE_LICENSE("GPL");
40
41 static const struct rtrs_rdma_dev_pd_ops dev_pd_ops;
42 static struct rtrs_rdma_dev_pd dev_pd = {
43 .ops = &dev_pd_ops
44 };
45
46 static struct workqueue_struct *rtrs_wq;
47 static struct class *rtrs_clt_dev_class;
48
rtrs_clt_is_connected(const struct rtrs_clt * clt)49 static inline bool rtrs_clt_is_connected(const struct rtrs_clt *clt)
50 {
51 struct rtrs_clt_path *clt_path;
52 bool connected = false;
53
54 rcu_read_lock();
55 list_for_each_entry_rcu(clt_path, &clt->paths_list, s.entry)
56 connected |= READ_ONCE(clt_path->state) == RTRS_CLT_CONNECTED;
57 rcu_read_unlock();
58
59 return connected;
60 }
61
62 static struct rtrs_permit *
__rtrs_get_permit(struct rtrs_clt * clt,enum rtrs_clt_con_type con_type)63 __rtrs_get_permit(struct rtrs_clt *clt, enum rtrs_clt_con_type con_type)
64 {
65 size_t max_depth = clt->queue_depth;
66 struct rtrs_permit *permit;
67 int bit;
68
69 /*
70 * Adapted from null_blk get_tag(). Callers from different cpus may
71 * grab the same bit, since find_first_zero_bit is not atomic.
72 * But then the test_and_set_bit_lock will fail for all the
73 * callers but one, so that they will loop again.
74 * This way an explicit spinlock is not required.
75 */
76 do {
77 bit = find_first_zero_bit(clt->permits_map, max_depth);
78 if (bit >= max_depth)
79 return NULL;
80 } while (test_and_set_bit_lock(bit, clt->permits_map));
81
82 permit = get_permit(clt, bit);
83 WARN_ON(permit->mem_id != bit);
84 permit->cpu_id = raw_smp_processor_id();
85 permit->con_type = con_type;
86
87 return permit;
88 }
89
__rtrs_put_permit(struct rtrs_clt * clt,struct rtrs_permit * permit)90 static inline void __rtrs_put_permit(struct rtrs_clt *clt,
91 struct rtrs_permit *permit)
92 {
93 clear_bit_unlock(permit->mem_id, clt->permits_map);
94 }
95
96 /**
97 * rtrs_clt_get_permit() - allocates permit for future RDMA operation
98 * @clt: Current session
99 * @con_type: Type of connection to use with the permit
100 * @can_wait: Wait type
101 *
102 * Description:
103 * Allocates permit for the following RDMA operation. Permit is used
104 * to preallocate all resources and to propagate memory pressure
105 * up earlier.
106 *
107 * Context:
108 * Can sleep if @wait == RTRS_PERMIT_WAIT
109 */
rtrs_clt_get_permit(struct rtrs_clt * clt,enum rtrs_clt_con_type con_type,enum wait_type can_wait)110 struct rtrs_permit *rtrs_clt_get_permit(struct rtrs_clt *clt,
111 enum rtrs_clt_con_type con_type,
112 enum wait_type can_wait)
113 {
114 struct rtrs_permit *permit;
115 DEFINE_WAIT(wait);
116
117 permit = __rtrs_get_permit(clt, con_type);
118 if (permit || !can_wait)
119 return permit;
120
121 do {
122 prepare_to_wait(&clt->permits_wait, &wait,
123 TASK_UNINTERRUPTIBLE);
124 permit = __rtrs_get_permit(clt, con_type);
125 if (permit)
126 break;
127
128 io_schedule();
129 } while (1);
130
131 finish_wait(&clt->permits_wait, &wait);
132
133 return permit;
134 }
135 EXPORT_SYMBOL(rtrs_clt_get_permit);
136
137 /**
138 * rtrs_clt_put_permit() - puts allocated permit
139 * @clt: Current session
140 * @permit: Permit to be freed
141 *
142 * Context:
143 * Does not matter
144 */
rtrs_clt_put_permit(struct rtrs_clt * clt,struct rtrs_permit * permit)145 void rtrs_clt_put_permit(struct rtrs_clt *clt, struct rtrs_permit *permit)
146 {
147 if (WARN_ON(!test_bit(permit->mem_id, clt->permits_map)))
148 return;
149
150 __rtrs_put_permit(clt, permit);
151
152 /*
153 * rtrs_clt_get_permit() adds itself to the &clt->permits_wait list
154 * before calling schedule(). So if rtrs_clt_get_permit() is sleeping
155 * it must have added itself to &clt->permits_wait before
156 * __rtrs_put_permit() finished.
157 * Hence it is safe to guard wake_up() with a waitqueue_active() test.
158 */
159 if (waitqueue_active(&clt->permits_wait))
160 wake_up(&clt->permits_wait);
161 }
162 EXPORT_SYMBOL(rtrs_clt_put_permit);
163
164 /**
165 * rtrs_permit_to_clt_con() - returns RDMA connection pointer by the permit
166 * @clt_path: client path pointer
167 * @permit: permit for the allocation of the RDMA buffer
168 * Note:
169 * IO connection starts from 1.
170 * 0 connection is for user messages.
171 */
172 static
rtrs_permit_to_clt_con(struct rtrs_clt_path * clt_path,struct rtrs_permit * permit)173 struct rtrs_clt_con *rtrs_permit_to_clt_con(struct rtrs_clt_path *clt_path,
174 struct rtrs_permit *permit)
175 {
176 int id = 0;
177
178 if (permit->con_type == RTRS_IO_CON)
179 id = (permit->cpu_id % (clt_path->s.irq_con_num - 1)) + 1;
180
181 return to_clt_con(clt_path->s.con[id]);
182 }
183
184 /**
185 * rtrs_clt_change_state() - change the session state through session state
186 * machine.
187 *
188 * @clt_path: client path to change the state of.
189 * @new_state: state to change to.
190 *
191 * returns true if sess's state is changed to new state, otherwise return false.
192 *
193 * Locks:
194 * state_wq lock must be hold.
195 */
rtrs_clt_change_state(struct rtrs_clt_path * clt_path,enum rtrs_clt_state new_state)196 static bool rtrs_clt_change_state(struct rtrs_clt_path *clt_path,
197 enum rtrs_clt_state new_state)
198 {
199 enum rtrs_clt_state old_state;
200 bool changed = false;
201
202 lockdep_assert_held(&clt_path->state_wq.lock);
203
204 old_state = clt_path->state;
205 switch (new_state) {
206 case RTRS_CLT_CONNECTING:
207 switch (old_state) {
208 case RTRS_CLT_RECONNECTING:
209 changed = true;
210 fallthrough;
211 default:
212 break;
213 }
214 break;
215 case RTRS_CLT_RECONNECTING:
216 switch (old_state) {
217 case RTRS_CLT_CONNECTED:
218 case RTRS_CLT_CONNECTING_ERR:
219 case RTRS_CLT_CLOSED:
220 changed = true;
221 fallthrough;
222 default:
223 break;
224 }
225 break;
226 case RTRS_CLT_CONNECTED:
227 switch (old_state) {
228 case RTRS_CLT_CONNECTING:
229 changed = true;
230 fallthrough;
231 default:
232 break;
233 }
234 break;
235 case RTRS_CLT_CONNECTING_ERR:
236 switch (old_state) {
237 case RTRS_CLT_CONNECTING:
238 changed = true;
239 fallthrough;
240 default:
241 break;
242 }
243 break;
244 case RTRS_CLT_CLOSING:
245 switch (old_state) {
246 case RTRS_CLT_CONNECTING:
247 case RTRS_CLT_CONNECTING_ERR:
248 case RTRS_CLT_RECONNECTING:
249 case RTRS_CLT_CONNECTED:
250 changed = true;
251 fallthrough;
252 default:
253 break;
254 }
255 break;
256 case RTRS_CLT_CLOSED:
257 switch (old_state) {
258 case RTRS_CLT_CLOSING:
259 changed = true;
260 fallthrough;
261 default:
262 break;
263 }
264 break;
265 case RTRS_CLT_DEAD:
266 switch (old_state) {
267 case RTRS_CLT_CLOSED:
268 changed = true;
269 fallthrough;
270 default:
271 break;
272 }
273 break;
274 default:
275 break;
276 }
277 if (changed) {
278 clt_path->state = new_state;
279 wake_up_locked(&clt_path->state_wq);
280 }
281
282 return changed;
283 }
284
rtrs_clt_change_state_from_to(struct rtrs_clt_path * clt_path,enum rtrs_clt_state old_state,enum rtrs_clt_state new_state)285 static bool rtrs_clt_change_state_from_to(struct rtrs_clt_path *clt_path,
286 enum rtrs_clt_state old_state,
287 enum rtrs_clt_state new_state)
288 {
289 bool changed = false;
290
291 spin_lock_irq(&clt_path->state_wq.lock);
292 if (clt_path->state == old_state)
293 changed = rtrs_clt_change_state(clt_path, new_state);
294 spin_unlock_irq(&clt_path->state_wq.lock);
295
296 return changed;
297 }
298
rtrs_rdma_error_recovery(struct rtrs_clt_con * con)299 static void rtrs_rdma_error_recovery(struct rtrs_clt_con *con)
300 {
301 struct rtrs_clt_path *clt_path = to_clt_path(con->c.path);
302
303 if (rtrs_clt_change_state_from_to(clt_path,
304 RTRS_CLT_CONNECTED,
305 RTRS_CLT_RECONNECTING)) {
306 struct rtrs_clt *clt = clt_path->clt;
307 unsigned int delay_ms;
308
309 /*
310 * Normal scenario, reconnect if we were successfully connected
311 */
312 delay_ms = clt->reconnect_delay_sec * 1000;
313 queue_delayed_work(rtrs_wq, &clt_path->reconnect_dwork,
314 msecs_to_jiffies(delay_ms +
315 prandom_u32() % RTRS_RECONNECT_SEED));
316 } else {
317 /*
318 * Error can happen just on establishing new connection,
319 * so notify waiter with error state, waiter is responsible
320 * for cleaning the rest and reconnect if needed.
321 */
322 rtrs_clt_change_state_from_to(clt_path,
323 RTRS_CLT_CONNECTING,
324 RTRS_CLT_CONNECTING_ERR);
325 }
326 }
327
rtrs_clt_fast_reg_done(struct ib_cq * cq,struct ib_wc * wc)328 static void rtrs_clt_fast_reg_done(struct ib_cq *cq, struct ib_wc *wc)
329 {
330 struct rtrs_clt_con *con = to_clt_con(wc->qp->qp_context);
331
332 if (wc->status != IB_WC_SUCCESS) {
333 rtrs_err(con->c.path, "Failed IB_WR_REG_MR: %s\n",
334 ib_wc_status_msg(wc->status));
335 rtrs_rdma_error_recovery(con);
336 }
337 }
338
339 static struct ib_cqe fast_reg_cqe = {
340 .done = rtrs_clt_fast_reg_done
341 };
342
343 static void complete_rdma_req(struct rtrs_clt_io_req *req, int errno,
344 bool notify, bool can_wait);
345
rtrs_clt_inv_rkey_done(struct ib_cq * cq,struct ib_wc * wc)346 static void rtrs_clt_inv_rkey_done(struct ib_cq *cq, struct ib_wc *wc)
347 {
348 struct rtrs_clt_io_req *req =
349 container_of(wc->wr_cqe, typeof(*req), inv_cqe);
350 struct rtrs_clt_con *con = to_clt_con(wc->qp->qp_context);
351
352 if (wc->status != IB_WC_SUCCESS) {
353 rtrs_err(con->c.path, "Failed IB_WR_LOCAL_INV: %s\n",
354 ib_wc_status_msg(wc->status));
355 rtrs_rdma_error_recovery(con);
356 }
357 req->need_inv = false;
358 if (req->need_inv_comp)
359 complete(&req->inv_comp);
360 else
361 /* Complete request from INV callback */
362 complete_rdma_req(req, req->inv_errno, true, false);
363 }
364
rtrs_inv_rkey(struct rtrs_clt_io_req * req)365 static int rtrs_inv_rkey(struct rtrs_clt_io_req *req)
366 {
367 struct rtrs_clt_con *con = req->con;
368 struct ib_send_wr wr = {
369 .opcode = IB_WR_LOCAL_INV,
370 .wr_cqe = &req->inv_cqe,
371 .send_flags = IB_SEND_SIGNALED,
372 .ex.invalidate_rkey = req->mr->rkey,
373 };
374 req->inv_cqe.done = rtrs_clt_inv_rkey_done;
375
376 return ib_post_send(con->c.qp, &wr, NULL);
377 }
378
complete_rdma_req(struct rtrs_clt_io_req * req,int errno,bool notify,bool can_wait)379 static void complete_rdma_req(struct rtrs_clt_io_req *req, int errno,
380 bool notify, bool can_wait)
381 {
382 struct rtrs_clt_con *con = req->con;
383 struct rtrs_clt_path *clt_path;
384 int err;
385
386 if (!req->in_use)
387 return;
388 if (WARN_ON(!req->con))
389 return;
390 clt_path = to_clt_path(con->c.path);
391
392 if (req->sg_cnt) {
393 if (req->dir == DMA_FROM_DEVICE && req->need_inv) {
394 /*
395 * We are here to invalidate read requests
396 * ourselves. In normal scenario server should
397 * send INV for all read requests, but
398 * we are here, thus two things could happen:
399 *
400 * 1. this is failover, when errno != 0
401 * and can_wait == 1,
402 *
403 * 2. something totally bad happened and
404 * server forgot to send INV, so we
405 * should do that ourselves.
406 */
407
408 if (can_wait) {
409 req->need_inv_comp = true;
410 } else {
411 /* This should be IO path, so always notify */
412 WARN_ON(!notify);
413 /* Save errno for INV callback */
414 req->inv_errno = errno;
415 }
416
417 refcount_inc(&req->ref);
418 err = rtrs_inv_rkey(req);
419 if (err) {
420 rtrs_err(con->c.path, "Send INV WR key=%#x: %d\n",
421 req->mr->rkey, err);
422 } else if (can_wait) {
423 wait_for_completion(&req->inv_comp);
424 } else {
425 /*
426 * Something went wrong, so request will be
427 * completed from INV callback.
428 */
429 WARN_ON_ONCE(1);
430
431 return;
432 }
433 if (!refcount_dec_and_test(&req->ref))
434 return;
435 }
436 ib_dma_unmap_sg(clt_path->s.dev->ib_dev, req->sglist,
437 req->sg_cnt, req->dir);
438 }
439 if (!refcount_dec_and_test(&req->ref))
440 return;
441 if (req->mp_policy == MP_POLICY_MIN_INFLIGHT)
442 atomic_dec(&clt_path->stats->inflight);
443
444 req->in_use = false;
445 req->con = NULL;
446
447 if (errno) {
448 rtrs_err_rl(con->c.path, "IO request failed: error=%d path=%s [%s:%u] notify=%d\n",
449 errno, kobject_name(&clt_path->kobj), clt_path->hca_name,
450 clt_path->hca_port, notify);
451 }
452
453 if (notify)
454 req->conf(req->priv, errno);
455 }
456
rtrs_post_send_rdma(struct rtrs_clt_con * con,struct rtrs_clt_io_req * req,struct rtrs_rbuf * rbuf,u32 off,u32 imm,struct ib_send_wr * wr)457 static int rtrs_post_send_rdma(struct rtrs_clt_con *con,
458 struct rtrs_clt_io_req *req,
459 struct rtrs_rbuf *rbuf, u32 off,
460 u32 imm, struct ib_send_wr *wr)
461 {
462 struct rtrs_clt_path *clt_path = to_clt_path(con->c.path);
463 enum ib_send_flags flags;
464 struct ib_sge sge;
465
466 if (!req->sg_size) {
467 rtrs_wrn(con->c.path,
468 "Doing RDMA Write failed, no data supplied\n");
469 return -EINVAL;
470 }
471
472 /* user data and user message in the first list element */
473 sge.addr = req->iu->dma_addr;
474 sge.length = req->sg_size;
475 sge.lkey = clt_path->s.dev->ib_pd->local_dma_lkey;
476
477 /*
478 * From time to time we have to post signalled sends,
479 * or send queue will fill up and only QP reset can help.
480 */
481 flags = atomic_inc_return(&con->c.wr_cnt) % clt_path->s.signal_interval ?
482 0 : IB_SEND_SIGNALED;
483
484 ib_dma_sync_single_for_device(clt_path->s.dev->ib_dev,
485 req->iu->dma_addr,
486 req->sg_size, DMA_TO_DEVICE);
487
488 return rtrs_iu_post_rdma_write_imm(&con->c, req->iu, &sge, 1,
489 rbuf->rkey, rbuf->addr + off,
490 imm, flags, wr, NULL);
491 }
492
process_io_rsp(struct rtrs_clt_path * clt_path,u32 msg_id,s16 errno,bool w_inval)493 static void process_io_rsp(struct rtrs_clt_path *clt_path, u32 msg_id,
494 s16 errno, bool w_inval)
495 {
496 struct rtrs_clt_io_req *req;
497
498 if (WARN_ON(msg_id >= clt_path->queue_depth))
499 return;
500
501 req = &clt_path->reqs[msg_id];
502 /* Drop need_inv if server responded with send with invalidation */
503 req->need_inv &= !w_inval;
504 complete_rdma_req(req, errno, true, false);
505 }
506
rtrs_clt_recv_done(struct rtrs_clt_con * con,struct ib_wc * wc)507 static void rtrs_clt_recv_done(struct rtrs_clt_con *con, struct ib_wc *wc)
508 {
509 struct rtrs_iu *iu;
510 int err;
511 struct rtrs_clt_path *clt_path = to_clt_path(con->c.path);
512
513 WARN_ON((clt_path->flags & RTRS_MSG_NEW_RKEY_F) == 0);
514 iu = container_of(wc->wr_cqe, struct rtrs_iu,
515 cqe);
516 err = rtrs_iu_post_recv(&con->c, iu);
517 if (err) {
518 rtrs_err(con->c.path, "post iu failed %d\n", err);
519 rtrs_rdma_error_recovery(con);
520 }
521 }
522
rtrs_clt_rkey_rsp_done(struct rtrs_clt_con * con,struct ib_wc * wc)523 static void rtrs_clt_rkey_rsp_done(struct rtrs_clt_con *con, struct ib_wc *wc)
524 {
525 struct rtrs_clt_path *clt_path = to_clt_path(con->c.path);
526 struct rtrs_msg_rkey_rsp *msg;
527 u32 imm_type, imm_payload;
528 bool w_inval = false;
529 struct rtrs_iu *iu;
530 u32 buf_id;
531 int err;
532
533 WARN_ON((clt_path->flags & RTRS_MSG_NEW_RKEY_F) == 0);
534
535 iu = container_of(wc->wr_cqe, struct rtrs_iu, cqe);
536
537 if (wc->byte_len < sizeof(*msg)) {
538 rtrs_err(con->c.path, "rkey response is malformed: size %d\n",
539 wc->byte_len);
540 goto out;
541 }
542 ib_dma_sync_single_for_cpu(clt_path->s.dev->ib_dev, iu->dma_addr,
543 iu->size, DMA_FROM_DEVICE);
544 msg = iu->buf;
545 if (le16_to_cpu(msg->type) != RTRS_MSG_RKEY_RSP) {
546 rtrs_err(clt_path->clt,
547 "rkey response is malformed: type %d\n",
548 le16_to_cpu(msg->type));
549 goto out;
550 }
551 buf_id = le16_to_cpu(msg->buf_id);
552 if (WARN_ON(buf_id >= clt_path->queue_depth))
553 goto out;
554
555 rtrs_from_imm(be32_to_cpu(wc->ex.imm_data), &imm_type, &imm_payload);
556 if (imm_type == RTRS_IO_RSP_IMM ||
557 imm_type == RTRS_IO_RSP_W_INV_IMM) {
558 u32 msg_id;
559
560 w_inval = (imm_type == RTRS_IO_RSP_W_INV_IMM);
561 rtrs_from_io_rsp_imm(imm_payload, &msg_id, &err);
562
563 if (WARN_ON(buf_id != msg_id))
564 goto out;
565 clt_path->rbufs[buf_id].rkey = le32_to_cpu(msg->rkey);
566 process_io_rsp(clt_path, msg_id, err, w_inval);
567 }
568 ib_dma_sync_single_for_device(clt_path->s.dev->ib_dev, iu->dma_addr,
569 iu->size, DMA_FROM_DEVICE);
570 return rtrs_clt_recv_done(con, wc);
571 out:
572 rtrs_rdma_error_recovery(con);
573 }
574
575 static void rtrs_clt_rdma_done(struct ib_cq *cq, struct ib_wc *wc);
576
577 static struct ib_cqe io_comp_cqe = {
578 .done = rtrs_clt_rdma_done
579 };
580
581 /*
582 * Post x2 empty WRs: first is for this RDMA with IMM,
583 * second is for RECV with INV, which happened earlier.
584 */
rtrs_post_recv_empty_x2(struct rtrs_con * con,struct ib_cqe * cqe)585 static int rtrs_post_recv_empty_x2(struct rtrs_con *con, struct ib_cqe *cqe)
586 {
587 struct ib_recv_wr wr_arr[2], *wr;
588 int i;
589
590 memset(wr_arr, 0, sizeof(wr_arr));
591 for (i = 0; i < ARRAY_SIZE(wr_arr); i++) {
592 wr = &wr_arr[i];
593 wr->wr_cqe = cqe;
594 if (i)
595 /* Chain backwards */
596 wr->next = &wr_arr[i - 1];
597 }
598
599 return ib_post_recv(con->qp, wr, NULL);
600 }
601
rtrs_clt_rdma_done(struct ib_cq * cq,struct ib_wc * wc)602 static void rtrs_clt_rdma_done(struct ib_cq *cq, struct ib_wc *wc)
603 {
604 struct rtrs_clt_con *con = to_clt_con(wc->qp->qp_context);
605 struct rtrs_clt_path *clt_path = to_clt_path(con->c.path);
606 u32 imm_type, imm_payload;
607 bool w_inval = false;
608 int err;
609
610 if (wc->status != IB_WC_SUCCESS) {
611 if (wc->status != IB_WC_WR_FLUSH_ERR) {
612 rtrs_err(clt_path->clt, "RDMA failed: %s\n",
613 ib_wc_status_msg(wc->status));
614 rtrs_rdma_error_recovery(con);
615 }
616 return;
617 }
618 rtrs_clt_update_wc_stats(con);
619
620 switch (wc->opcode) {
621 case IB_WC_RECV_RDMA_WITH_IMM:
622 /*
623 * post_recv() RDMA write completions of IO reqs (read/write)
624 * and hb
625 */
626 if (WARN_ON(wc->wr_cqe->done != rtrs_clt_rdma_done))
627 return;
628 rtrs_from_imm(be32_to_cpu(wc->ex.imm_data),
629 &imm_type, &imm_payload);
630 if (imm_type == RTRS_IO_RSP_IMM ||
631 imm_type == RTRS_IO_RSP_W_INV_IMM) {
632 u32 msg_id;
633
634 w_inval = (imm_type == RTRS_IO_RSP_W_INV_IMM);
635 rtrs_from_io_rsp_imm(imm_payload, &msg_id, &err);
636
637 process_io_rsp(clt_path, msg_id, err, w_inval);
638 } else if (imm_type == RTRS_HB_MSG_IMM) {
639 WARN_ON(con->c.cid);
640 rtrs_send_hb_ack(&clt_path->s);
641 if (clt_path->flags & RTRS_MSG_NEW_RKEY_F)
642 return rtrs_clt_recv_done(con, wc);
643 } else if (imm_type == RTRS_HB_ACK_IMM) {
644 WARN_ON(con->c.cid);
645 clt_path->s.hb_missed_cnt = 0;
646 clt_path->s.hb_cur_latency =
647 ktime_sub(ktime_get(), clt_path->s.hb_last_sent);
648 if (clt_path->flags & RTRS_MSG_NEW_RKEY_F)
649 return rtrs_clt_recv_done(con, wc);
650 } else {
651 rtrs_wrn(con->c.path, "Unknown IMM type %u\n",
652 imm_type);
653 }
654 if (w_inval)
655 /*
656 * Post x2 empty WRs: first is for this RDMA with IMM,
657 * second is for RECV with INV, which happened earlier.
658 */
659 err = rtrs_post_recv_empty_x2(&con->c, &io_comp_cqe);
660 else
661 err = rtrs_post_recv_empty(&con->c, &io_comp_cqe);
662 if (err) {
663 rtrs_err(con->c.path, "rtrs_post_recv_empty(): %d\n",
664 err);
665 rtrs_rdma_error_recovery(con);
666 }
667 break;
668 case IB_WC_RECV:
669 /*
670 * Key invalidations from server side
671 */
672 WARN_ON(!(wc->wc_flags & IB_WC_WITH_INVALIDATE ||
673 wc->wc_flags & IB_WC_WITH_IMM));
674 WARN_ON(wc->wr_cqe->done != rtrs_clt_rdma_done);
675 if (clt_path->flags & RTRS_MSG_NEW_RKEY_F) {
676 if (wc->wc_flags & IB_WC_WITH_INVALIDATE)
677 return rtrs_clt_recv_done(con, wc);
678
679 return rtrs_clt_rkey_rsp_done(con, wc);
680 }
681 break;
682 case IB_WC_RDMA_WRITE:
683 /*
684 * post_send() RDMA write completions of IO reqs (read/write)
685 * and hb.
686 */
687 break;
688
689 default:
690 rtrs_wrn(clt_path->clt, "Unexpected WC type: %d\n", wc->opcode);
691 return;
692 }
693 }
694
post_recv_io(struct rtrs_clt_con * con,size_t q_size)695 static int post_recv_io(struct rtrs_clt_con *con, size_t q_size)
696 {
697 int err, i;
698 struct rtrs_clt_path *clt_path = to_clt_path(con->c.path);
699
700 for (i = 0; i < q_size; i++) {
701 if (clt_path->flags & RTRS_MSG_NEW_RKEY_F) {
702 struct rtrs_iu *iu = &con->rsp_ius[i];
703
704 err = rtrs_iu_post_recv(&con->c, iu);
705 } else {
706 err = rtrs_post_recv_empty(&con->c, &io_comp_cqe);
707 }
708 if (err)
709 return err;
710 }
711
712 return 0;
713 }
714
post_recv_path(struct rtrs_clt_path * clt_path)715 static int post_recv_path(struct rtrs_clt_path *clt_path)
716 {
717 size_t q_size = 0;
718 int err, cid;
719
720 for (cid = 0; cid < clt_path->s.con_num; cid++) {
721 if (cid == 0)
722 q_size = SERVICE_CON_QUEUE_DEPTH;
723 else
724 q_size = clt_path->queue_depth;
725
726 /*
727 * x2 for RDMA read responses + FR key invalidations,
728 * RDMA writes do not require any FR registrations.
729 */
730 q_size *= 2;
731
732 err = post_recv_io(to_clt_con(clt_path->s.con[cid]), q_size);
733 if (err) {
734 rtrs_err(clt_path->clt, "post_recv_io(), err: %d\n",
735 err);
736 return err;
737 }
738 }
739
740 return 0;
741 }
742
743 struct path_it {
744 int i;
745 struct list_head skip_list;
746 struct rtrs_clt *clt;
747 struct rtrs_clt_path *(*next_path)(struct path_it *it);
748 };
749
750 /*
751 * rtrs_clt_get_next_path_or_null - get clt path from the list or return NULL
752 * @head: the head for the list.
753 * @clt_path: The element to take the next clt_path from.
754 *
755 * Next clt path returned in round-robin fashion, i.e. head will be skipped,
756 * but if list is observed as empty, NULL will be returned.
757 *
758 * This function may safely run concurrently with the _rcu list-mutation
759 * primitives such as list_add_rcu() as long as it's guarded by rcu_read_lock().
760 */
761 static inline struct rtrs_clt_path *
rtrs_clt_get_next_path_or_null(struct list_head * head,struct rtrs_clt_path * clt_path)762 rtrs_clt_get_next_path_or_null(struct list_head *head, struct rtrs_clt_path *clt_path)
763 {
764 return list_next_or_null_rcu(head, &clt_path->s.entry, typeof(*clt_path), s.entry) ?:
765 list_next_or_null_rcu(head,
766 READ_ONCE((&clt_path->s.entry)->next),
767 typeof(*clt_path), s.entry);
768 }
769
770 /**
771 * get_next_path_rr() - Returns path in round-robin fashion.
772 * @it: the path pointer
773 *
774 * Related to @MP_POLICY_RR
775 *
776 * Locks:
777 * rcu_read_lock() must be hold.
778 */
get_next_path_rr(struct path_it * it)779 static struct rtrs_clt_path *get_next_path_rr(struct path_it *it)
780 {
781 struct rtrs_clt_path __rcu **ppcpu_path;
782 struct rtrs_clt_path *path;
783 struct rtrs_clt *clt;
784
785 clt = it->clt;
786
787 /*
788 * Here we use two RCU objects: @paths_list and @pcpu_path
789 * pointer. See rtrs_clt_remove_path_from_arr() for details
790 * how that is handled.
791 */
792
793 ppcpu_path = this_cpu_ptr(clt->pcpu_path);
794 path = rcu_dereference(*ppcpu_path);
795 if (!path)
796 path = list_first_or_null_rcu(&clt->paths_list,
797 typeof(*path), s.entry);
798 else
799 path = rtrs_clt_get_next_path_or_null(&clt->paths_list, path);
800
801 rcu_assign_pointer(*ppcpu_path, path);
802
803 return path;
804 }
805
806 /**
807 * get_next_path_min_inflight() - Returns path with minimal inflight count.
808 * @it: the path pointer
809 *
810 * Related to @MP_POLICY_MIN_INFLIGHT
811 *
812 * Locks:
813 * rcu_read_lock() must be hold.
814 */
get_next_path_min_inflight(struct path_it * it)815 static struct rtrs_clt_path *get_next_path_min_inflight(struct path_it *it)
816 {
817 struct rtrs_clt_path *min_path = NULL;
818 struct rtrs_clt *clt = it->clt;
819 struct rtrs_clt_path *clt_path;
820 int min_inflight = INT_MAX;
821 int inflight;
822
823 list_for_each_entry_rcu(clt_path, &clt->paths_list, s.entry) {
824 if (READ_ONCE(clt_path->state) != RTRS_CLT_CONNECTED)
825 continue;
826
827 if (!list_empty(raw_cpu_ptr(clt_path->mp_skip_entry)))
828 continue;
829
830 inflight = atomic_read(&clt_path->stats->inflight);
831
832 if (inflight < min_inflight) {
833 min_inflight = inflight;
834 min_path = clt_path;
835 }
836 }
837
838 /*
839 * add the path to the skip list, so that next time we can get
840 * a different one
841 */
842 if (min_path)
843 list_add(raw_cpu_ptr(min_path->mp_skip_entry), &it->skip_list);
844
845 return min_path;
846 }
847
848 /**
849 * get_next_path_min_latency() - Returns path with minimal latency.
850 * @it: the path pointer
851 *
852 * Return: a path with the lowest latency or NULL if all paths are tried
853 *
854 * Locks:
855 * rcu_read_lock() must be hold.
856 *
857 * Related to @MP_POLICY_MIN_LATENCY
858 *
859 * This DOES skip an already-tried path.
860 * There is a skip-list to skip a path if the path has tried but failed.
861 * It will try the minimum latency path and then the second minimum latency
862 * path and so on. Finally it will return NULL if all paths are tried.
863 * Therefore the caller MUST check the returned
864 * path is NULL and trigger the IO error.
865 */
get_next_path_min_latency(struct path_it * it)866 static struct rtrs_clt_path *get_next_path_min_latency(struct path_it *it)
867 {
868 struct rtrs_clt_path *min_path = NULL;
869 struct rtrs_clt *clt = it->clt;
870 struct rtrs_clt_path *clt_path;
871 ktime_t min_latency = KTIME_MAX;
872 ktime_t latency;
873
874 list_for_each_entry_rcu(clt_path, &clt->paths_list, s.entry) {
875 if (READ_ONCE(clt_path->state) != RTRS_CLT_CONNECTED)
876 continue;
877
878 if (!list_empty(raw_cpu_ptr(clt_path->mp_skip_entry)))
879 continue;
880
881 latency = clt_path->s.hb_cur_latency;
882
883 if (latency < min_latency) {
884 min_latency = latency;
885 min_path = clt_path;
886 }
887 }
888
889 /*
890 * add the path to the skip list, so that next time we can get
891 * a different one
892 */
893 if (min_path)
894 list_add(raw_cpu_ptr(min_path->mp_skip_entry), &it->skip_list);
895
896 return min_path;
897 }
898
path_it_init(struct path_it * it,struct rtrs_clt * clt)899 static inline void path_it_init(struct path_it *it, struct rtrs_clt *clt)
900 {
901 INIT_LIST_HEAD(&it->skip_list);
902 it->clt = clt;
903 it->i = 0;
904
905 if (clt->mp_policy == MP_POLICY_RR)
906 it->next_path = get_next_path_rr;
907 else if (clt->mp_policy == MP_POLICY_MIN_INFLIGHT)
908 it->next_path = get_next_path_min_inflight;
909 else
910 it->next_path = get_next_path_min_latency;
911 }
912
path_it_deinit(struct path_it * it)913 static inline void path_it_deinit(struct path_it *it)
914 {
915 struct list_head *skip, *tmp;
916 /*
917 * The skip_list is used only for the MIN_INFLIGHT policy.
918 * We need to remove paths from it, so that next IO can insert
919 * paths (->mp_skip_entry) into a skip_list again.
920 */
921 list_for_each_safe(skip, tmp, &it->skip_list)
922 list_del_init(skip);
923 }
924
925 /**
926 * rtrs_clt_init_req() - Initialize an rtrs_clt_io_req holding information
927 * about an inflight IO.
928 * The user buffer holding user control message (not data) is copied into
929 * the corresponding buffer of rtrs_iu (req->iu->buf), which later on will
930 * also hold the control message of rtrs.
931 * @req: an io request holding information about IO.
932 * @clt_path: client path
933 * @conf: conformation callback function to notify upper layer.
934 * @permit: permit for allocation of RDMA remote buffer
935 * @priv: private pointer
936 * @vec: kernel vector containing control message
937 * @usr_len: length of the user message
938 * @sg: scater list for IO data
939 * @sg_cnt: number of scater list entries
940 * @data_len: length of the IO data
941 * @dir: direction of the IO.
942 */
rtrs_clt_init_req(struct rtrs_clt_io_req * req,struct rtrs_clt_path * clt_path,void (* conf)(void * priv,int errno),struct rtrs_permit * permit,void * priv,const struct kvec * vec,size_t usr_len,struct scatterlist * sg,size_t sg_cnt,size_t data_len,int dir)943 static void rtrs_clt_init_req(struct rtrs_clt_io_req *req,
944 struct rtrs_clt_path *clt_path,
945 void (*conf)(void *priv, int errno),
946 struct rtrs_permit *permit, void *priv,
947 const struct kvec *vec, size_t usr_len,
948 struct scatterlist *sg, size_t sg_cnt,
949 size_t data_len, int dir)
950 {
951 struct iov_iter iter;
952 size_t len;
953
954 req->permit = permit;
955 req->in_use = true;
956 req->usr_len = usr_len;
957 req->data_len = data_len;
958 req->sglist = sg;
959 req->sg_cnt = sg_cnt;
960 req->priv = priv;
961 req->dir = dir;
962 req->con = rtrs_permit_to_clt_con(clt_path, permit);
963 req->conf = conf;
964 req->need_inv = false;
965 req->need_inv_comp = false;
966 req->inv_errno = 0;
967 refcount_set(&req->ref, 1);
968 req->mp_policy = clt_path->clt->mp_policy;
969
970 iov_iter_kvec(&iter, WRITE, vec, 1, usr_len);
971 len = _copy_from_iter(req->iu->buf, usr_len, &iter);
972 WARN_ON(len != usr_len);
973
974 reinit_completion(&req->inv_comp);
975 }
976
977 static struct rtrs_clt_io_req *
rtrs_clt_get_req(struct rtrs_clt_path * clt_path,void (* conf)(void * priv,int errno),struct rtrs_permit * permit,void * priv,const struct kvec * vec,size_t usr_len,struct scatterlist * sg,size_t sg_cnt,size_t data_len,int dir)978 rtrs_clt_get_req(struct rtrs_clt_path *clt_path,
979 void (*conf)(void *priv, int errno),
980 struct rtrs_permit *permit, void *priv,
981 const struct kvec *vec, size_t usr_len,
982 struct scatterlist *sg, size_t sg_cnt,
983 size_t data_len, int dir)
984 {
985 struct rtrs_clt_io_req *req;
986
987 req = &clt_path->reqs[permit->mem_id];
988 rtrs_clt_init_req(req, clt_path, conf, permit, priv, vec, usr_len,
989 sg, sg_cnt, data_len, dir);
990 return req;
991 }
992
993 static struct rtrs_clt_io_req *
rtrs_clt_get_copy_req(struct rtrs_clt_path * alive_path,struct rtrs_clt_io_req * fail_req)994 rtrs_clt_get_copy_req(struct rtrs_clt_path *alive_path,
995 struct rtrs_clt_io_req *fail_req)
996 {
997 struct rtrs_clt_io_req *req;
998 struct kvec vec = {
999 .iov_base = fail_req->iu->buf,
1000 .iov_len = fail_req->usr_len
1001 };
1002
1003 req = &alive_path->reqs[fail_req->permit->mem_id];
1004 rtrs_clt_init_req(req, alive_path, fail_req->conf, fail_req->permit,
1005 fail_req->priv, &vec, fail_req->usr_len,
1006 fail_req->sglist, fail_req->sg_cnt,
1007 fail_req->data_len, fail_req->dir);
1008 return req;
1009 }
1010
rtrs_post_rdma_write_sg(struct rtrs_clt_con * con,struct rtrs_clt_io_req * req,struct rtrs_rbuf * rbuf,bool fr_en,u32 count,u32 size,u32 imm,struct ib_send_wr * wr,struct ib_send_wr * tail)1011 static int rtrs_post_rdma_write_sg(struct rtrs_clt_con *con,
1012 struct rtrs_clt_io_req *req,
1013 struct rtrs_rbuf *rbuf, bool fr_en,
1014 u32 count, u32 size, u32 imm,
1015 struct ib_send_wr *wr,
1016 struct ib_send_wr *tail)
1017 {
1018 struct rtrs_clt_path *clt_path = to_clt_path(con->c.path);
1019 struct ib_sge *sge = req->sge;
1020 enum ib_send_flags flags;
1021 struct scatterlist *sg;
1022 size_t num_sge;
1023 int i;
1024 struct ib_send_wr *ptail = NULL;
1025
1026 if (fr_en) {
1027 i = 0;
1028 sge[i].addr = req->mr->iova;
1029 sge[i].length = req->mr->length;
1030 sge[i].lkey = req->mr->lkey;
1031 i++;
1032 num_sge = 2;
1033 ptail = tail;
1034 } else {
1035 for_each_sg(req->sglist, sg, count, i) {
1036 sge[i].addr = sg_dma_address(sg);
1037 sge[i].length = sg_dma_len(sg);
1038 sge[i].lkey = clt_path->s.dev->ib_pd->local_dma_lkey;
1039 }
1040 num_sge = 1 + count;
1041 }
1042 sge[i].addr = req->iu->dma_addr;
1043 sge[i].length = size;
1044 sge[i].lkey = clt_path->s.dev->ib_pd->local_dma_lkey;
1045
1046 /*
1047 * From time to time we have to post signalled sends,
1048 * or send queue will fill up and only QP reset can help.
1049 */
1050 flags = atomic_inc_return(&con->c.wr_cnt) % clt_path->s.signal_interval ?
1051 0 : IB_SEND_SIGNALED;
1052
1053 ib_dma_sync_single_for_device(clt_path->s.dev->ib_dev,
1054 req->iu->dma_addr,
1055 size, DMA_TO_DEVICE);
1056
1057 return rtrs_iu_post_rdma_write_imm(&con->c, req->iu, sge, num_sge,
1058 rbuf->rkey, rbuf->addr, imm,
1059 flags, wr, ptail);
1060 }
1061
rtrs_map_sg_fr(struct rtrs_clt_io_req * req,size_t count)1062 static int rtrs_map_sg_fr(struct rtrs_clt_io_req *req, size_t count)
1063 {
1064 int nr;
1065
1066 /* Align the MR to a 4K page size to match the block virt boundary */
1067 nr = ib_map_mr_sg(req->mr, req->sglist, count, NULL, SZ_4K);
1068 if (nr < 0)
1069 return nr;
1070 if (nr < req->sg_cnt)
1071 return -EINVAL;
1072 ib_update_fast_reg_key(req->mr, ib_inc_rkey(req->mr->rkey));
1073
1074 return nr;
1075 }
1076
rtrs_clt_write_req(struct rtrs_clt_io_req * req)1077 static int rtrs_clt_write_req(struct rtrs_clt_io_req *req)
1078 {
1079 struct rtrs_clt_con *con = req->con;
1080 struct rtrs_path *s = con->c.path;
1081 struct rtrs_clt_path *clt_path = to_clt_path(s);
1082 struct rtrs_msg_rdma_write *msg;
1083
1084 struct rtrs_rbuf *rbuf;
1085 int ret, count = 0;
1086 u32 imm, buf_id;
1087 struct ib_reg_wr rwr;
1088 struct ib_send_wr inv_wr;
1089 struct ib_send_wr *wr = NULL;
1090 bool fr_en = false;
1091
1092 const size_t tsize = sizeof(*msg) + req->data_len + req->usr_len;
1093
1094 if (tsize > clt_path->chunk_size) {
1095 rtrs_wrn(s, "Write request failed, size too big %zu > %d\n",
1096 tsize, clt_path->chunk_size);
1097 return -EMSGSIZE;
1098 }
1099 if (req->sg_cnt) {
1100 count = ib_dma_map_sg(clt_path->s.dev->ib_dev, req->sglist,
1101 req->sg_cnt, req->dir);
1102 if (!count) {
1103 rtrs_wrn(s, "Write request failed, map failed\n");
1104 return -EINVAL;
1105 }
1106 }
1107 /* put rtrs msg after sg and user message */
1108 msg = req->iu->buf + req->usr_len;
1109 msg->type = cpu_to_le16(RTRS_MSG_WRITE);
1110 msg->usr_len = cpu_to_le16(req->usr_len);
1111
1112 /* rtrs message on server side will be after user data and message */
1113 imm = req->permit->mem_off + req->data_len + req->usr_len;
1114 imm = rtrs_to_io_req_imm(imm);
1115 buf_id = req->permit->mem_id;
1116 req->sg_size = tsize;
1117 rbuf = &clt_path->rbufs[buf_id];
1118
1119 if (count) {
1120 ret = rtrs_map_sg_fr(req, count);
1121 if (ret < 0) {
1122 rtrs_err_rl(s,
1123 "Write request failed, failed to map fast reg. data, err: %d\n",
1124 ret);
1125 ib_dma_unmap_sg(clt_path->s.dev->ib_dev, req->sglist,
1126 req->sg_cnt, req->dir);
1127 return ret;
1128 }
1129 inv_wr = (struct ib_send_wr) {
1130 .opcode = IB_WR_LOCAL_INV,
1131 .wr_cqe = &req->inv_cqe,
1132 .send_flags = IB_SEND_SIGNALED,
1133 .ex.invalidate_rkey = req->mr->rkey,
1134 };
1135 req->inv_cqe.done = rtrs_clt_inv_rkey_done;
1136 rwr = (struct ib_reg_wr) {
1137 .wr.opcode = IB_WR_REG_MR,
1138 .wr.wr_cqe = &fast_reg_cqe,
1139 .mr = req->mr,
1140 .key = req->mr->rkey,
1141 .access = (IB_ACCESS_LOCAL_WRITE),
1142 };
1143 wr = &rwr.wr;
1144 fr_en = true;
1145 refcount_inc(&req->ref);
1146 }
1147 /*
1148 * Update stats now, after request is successfully sent it is not
1149 * safe anymore to touch it.
1150 */
1151 rtrs_clt_update_all_stats(req, WRITE);
1152
1153 ret = rtrs_post_rdma_write_sg(req->con, req, rbuf, fr_en, count,
1154 req->usr_len + sizeof(*msg),
1155 imm, wr, &inv_wr);
1156 if (ret) {
1157 rtrs_err_rl(s,
1158 "Write request failed: error=%d path=%s [%s:%u]\n",
1159 ret, kobject_name(&clt_path->kobj), clt_path->hca_name,
1160 clt_path->hca_port);
1161 if (req->mp_policy == MP_POLICY_MIN_INFLIGHT)
1162 atomic_dec(&clt_path->stats->inflight);
1163 if (req->sg_cnt)
1164 ib_dma_unmap_sg(clt_path->s.dev->ib_dev, req->sglist,
1165 req->sg_cnt, req->dir);
1166 }
1167
1168 return ret;
1169 }
1170
rtrs_clt_read_req(struct rtrs_clt_io_req * req)1171 static int rtrs_clt_read_req(struct rtrs_clt_io_req *req)
1172 {
1173 struct rtrs_clt_con *con = req->con;
1174 struct rtrs_path *s = con->c.path;
1175 struct rtrs_clt_path *clt_path = to_clt_path(s);
1176 struct rtrs_msg_rdma_read *msg;
1177 struct rtrs_ib_dev *dev = clt_path->s.dev;
1178
1179 struct ib_reg_wr rwr;
1180 struct ib_send_wr *wr = NULL;
1181
1182 int ret, count = 0;
1183 u32 imm, buf_id;
1184
1185 const size_t tsize = sizeof(*msg) + req->data_len + req->usr_len;
1186
1187 if (tsize > clt_path->chunk_size) {
1188 rtrs_wrn(s,
1189 "Read request failed, message size is %zu, bigger than CHUNK_SIZE %d\n",
1190 tsize, clt_path->chunk_size);
1191 return -EMSGSIZE;
1192 }
1193
1194 if (req->sg_cnt) {
1195 count = ib_dma_map_sg(dev->ib_dev, req->sglist, req->sg_cnt,
1196 req->dir);
1197 if (!count) {
1198 rtrs_wrn(s,
1199 "Read request failed, dma map failed\n");
1200 return -EINVAL;
1201 }
1202 }
1203 /* put our message into req->buf after user message*/
1204 msg = req->iu->buf + req->usr_len;
1205 msg->type = cpu_to_le16(RTRS_MSG_READ);
1206 msg->usr_len = cpu_to_le16(req->usr_len);
1207
1208 if (count) {
1209 ret = rtrs_map_sg_fr(req, count);
1210 if (ret < 0) {
1211 rtrs_err_rl(s,
1212 "Read request failed, failed to map fast reg. data, err: %d\n",
1213 ret);
1214 ib_dma_unmap_sg(dev->ib_dev, req->sglist, req->sg_cnt,
1215 req->dir);
1216 return ret;
1217 }
1218 rwr = (struct ib_reg_wr) {
1219 .wr.opcode = IB_WR_REG_MR,
1220 .wr.wr_cqe = &fast_reg_cqe,
1221 .mr = req->mr,
1222 .key = req->mr->rkey,
1223 .access = (IB_ACCESS_LOCAL_WRITE |
1224 IB_ACCESS_REMOTE_WRITE),
1225 };
1226 wr = &rwr.wr;
1227
1228 msg->sg_cnt = cpu_to_le16(1);
1229 msg->flags = cpu_to_le16(RTRS_MSG_NEED_INVAL_F);
1230
1231 msg->desc[0].addr = cpu_to_le64(req->mr->iova);
1232 msg->desc[0].key = cpu_to_le32(req->mr->rkey);
1233 msg->desc[0].len = cpu_to_le32(req->mr->length);
1234
1235 /* Further invalidation is required */
1236 req->need_inv = !!RTRS_MSG_NEED_INVAL_F;
1237
1238 } else {
1239 msg->sg_cnt = 0;
1240 msg->flags = 0;
1241 }
1242 /*
1243 * rtrs message will be after the space reserved for disk data and
1244 * user message
1245 */
1246 imm = req->permit->mem_off + req->data_len + req->usr_len;
1247 imm = rtrs_to_io_req_imm(imm);
1248 buf_id = req->permit->mem_id;
1249
1250 req->sg_size = sizeof(*msg);
1251 req->sg_size += le16_to_cpu(msg->sg_cnt) * sizeof(struct rtrs_sg_desc);
1252 req->sg_size += req->usr_len;
1253
1254 /*
1255 * Update stats now, after request is successfully sent it is not
1256 * safe anymore to touch it.
1257 */
1258 rtrs_clt_update_all_stats(req, READ);
1259
1260 ret = rtrs_post_send_rdma(req->con, req, &clt_path->rbufs[buf_id],
1261 req->data_len, imm, wr);
1262 if (ret) {
1263 rtrs_err_rl(s,
1264 "Read request failed: error=%d path=%s [%s:%u]\n",
1265 ret, kobject_name(&clt_path->kobj), clt_path->hca_name,
1266 clt_path->hca_port);
1267 if (req->mp_policy == MP_POLICY_MIN_INFLIGHT)
1268 atomic_dec(&clt_path->stats->inflight);
1269 req->need_inv = false;
1270 if (req->sg_cnt)
1271 ib_dma_unmap_sg(dev->ib_dev, req->sglist,
1272 req->sg_cnt, req->dir);
1273 }
1274
1275 return ret;
1276 }
1277
1278 /**
1279 * rtrs_clt_failover_req() - Try to find an active path for a failed request
1280 * @clt: clt context
1281 * @fail_req: a failed io request.
1282 */
rtrs_clt_failover_req(struct rtrs_clt * clt,struct rtrs_clt_io_req * fail_req)1283 static int rtrs_clt_failover_req(struct rtrs_clt *clt,
1284 struct rtrs_clt_io_req *fail_req)
1285 {
1286 struct rtrs_clt_path *alive_path;
1287 struct rtrs_clt_io_req *req;
1288 int err = -ECONNABORTED;
1289 struct path_it it;
1290
1291 rcu_read_lock();
1292 for (path_it_init(&it, clt);
1293 (alive_path = it.next_path(&it)) && it.i < it.clt->paths_num;
1294 it.i++) {
1295 if (READ_ONCE(alive_path->state) != RTRS_CLT_CONNECTED)
1296 continue;
1297 req = rtrs_clt_get_copy_req(alive_path, fail_req);
1298 if (req->dir == DMA_TO_DEVICE)
1299 err = rtrs_clt_write_req(req);
1300 else
1301 err = rtrs_clt_read_req(req);
1302 if (err) {
1303 req->in_use = false;
1304 continue;
1305 }
1306 /* Success path */
1307 rtrs_clt_inc_failover_cnt(alive_path->stats);
1308 break;
1309 }
1310 path_it_deinit(&it);
1311 rcu_read_unlock();
1312
1313 return err;
1314 }
1315
fail_all_outstanding_reqs(struct rtrs_clt_path * clt_path)1316 static void fail_all_outstanding_reqs(struct rtrs_clt_path *clt_path)
1317 {
1318 struct rtrs_clt *clt = clt_path->clt;
1319 struct rtrs_clt_io_req *req;
1320 int i, err;
1321
1322 if (!clt_path->reqs)
1323 return;
1324 for (i = 0; i < clt_path->queue_depth; ++i) {
1325 req = &clt_path->reqs[i];
1326 if (!req->in_use)
1327 continue;
1328
1329 /*
1330 * Safely (without notification) complete failed request.
1331 * After completion this request is still useble and can
1332 * be failovered to another path.
1333 */
1334 complete_rdma_req(req, -ECONNABORTED, false, true);
1335
1336 err = rtrs_clt_failover_req(clt, req);
1337 if (err)
1338 /* Failover failed, notify anyway */
1339 req->conf(req->priv, err);
1340 }
1341 }
1342
free_path_reqs(struct rtrs_clt_path * clt_path)1343 static void free_path_reqs(struct rtrs_clt_path *clt_path)
1344 {
1345 struct rtrs_clt_io_req *req;
1346 int i;
1347
1348 if (!clt_path->reqs)
1349 return;
1350 for (i = 0; i < clt_path->queue_depth; ++i) {
1351 req = &clt_path->reqs[i];
1352 if (req->mr)
1353 ib_dereg_mr(req->mr);
1354 kfree(req->sge);
1355 rtrs_iu_free(req->iu, clt_path->s.dev->ib_dev, 1);
1356 }
1357 kfree(clt_path->reqs);
1358 clt_path->reqs = NULL;
1359 }
1360
alloc_path_reqs(struct rtrs_clt_path * clt_path)1361 static int alloc_path_reqs(struct rtrs_clt_path *clt_path)
1362 {
1363 struct rtrs_clt_io_req *req;
1364 int i, err = -ENOMEM;
1365
1366 clt_path->reqs = kcalloc(clt_path->queue_depth,
1367 sizeof(*clt_path->reqs),
1368 GFP_KERNEL);
1369 if (!clt_path->reqs)
1370 return -ENOMEM;
1371
1372 for (i = 0; i < clt_path->queue_depth; ++i) {
1373 req = &clt_path->reqs[i];
1374 req->iu = rtrs_iu_alloc(1, clt_path->max_hdr_size, GFP_KERNEL,
1375 clt_path->s.dev->ib_dev,
1376 DMA_TO_DEVICE,
1377 rtrs_clt_rdma_done);
1378 if (!req->iu)
1379 goto out;
1380
1381 req->sge = kcalloc(2, sizeof(*req->sge), GFP_KERNEL);
1382 if (!req->sge)
1383 goto out;
1384
1385 req->mr = ib_alloc_mr(clt_path->s.dev->ib_pd,
1386 IB_MR_TYPE_MEM_REG,
1387 clt_path->max_pages_per_mr);
1388 if (IS_ERR(req->mr)) {
1389 err = PTR_ERR(req->mr);
1390 req->mr = NULL;
1391 pr_err("Failed to alloc clt_path->max_pages_per_mr %d\n",
1392 clt_path->max_pages_per_mr);
1393 goto out;
1394 }
1395
1396 init_completion(&req->inv_comp);
1397 }
1398
1399 return 0;
1400
1401 out:
1402 free_path_reqs(clt_path);
1403
1404 return err;
1405 }
1406
alloc_permits(struct rtrs_clt * clt)1407 static int alloc_permits(struct rtrs_clt *clt)
1408 {
1409 unsigned int chunk_bits;
1410 int err, i;
1411
1412 clt->permits_map = kcalloc(BITS_TO_LONGS(clt->queue_depth),
1413 sizeof(long), GFP_KERNEL);
1414 if (!clt->permits_map) {
1415 err = -ENOMEM;
1416 goto out_err;
1417 }
1418 clt->permits = kcalloc(clt->queue_depth, permit_size(clt), GFP_KERNEL);
1419 if (!clt->permits) {
1420 err = -ENOMEM;
1421 goto err_map;
1422 }
1423 chunk_bits = ilog2(clt->queue_depth - 1) + 1;
1424 for (i = 0; i < clt->queue_depth; i++) {
1425 struct rtrs_permit *permit;
1426
1427 permit = get_permit(clt, i);
1428 permit->mem_id = i;
1429 permit->mem_off = i << (MAX_IMM_PAYL_BITS - chunk_bits);
1430 }
1431
1432 return 0;
1433
1434 err_map:
1435 kfree(clt->permits_map);
1436 clt->permits_map = NULL;
1437 out_err:
1438 return err;
1439 }
1440
free_permits(struct rtrs_clt * clt)1441 static void free_permits(struct rtrs_clt *clt)
1442 {
1443 if (clt->permits_map) {
1444 size_t sz = clt->queue_depth;
1445
1446 wait_event(clt->permits_wait,
1447 find_first_bit(clt->permits_map, sz) >= sz);
1448 }
1449 kfree(clt->permits_map);
1450 clt->permits_map = NULL;
1451 kfree(clt->permits);
1452 clt->permits = NULL;
1453 }
1454
query_fast_reg_mode(struct rtrs_clt_path * clt_path)1455 static void query_fast_reg_mode(struct rtrs_clt_path *clt_path)
1456 {
1457 struct ib_device *ib_dev;
1458 u64 max_pages_per_mr;
1459 int mr_page_shift;
1460
1461 ib_dev = clt_path->s.dev->ib_dev;
1462
1463 /*
1464 * Use the smallest page size supported by the HCA, down to a
1465 * minimum of 4096 bytes. We're unlikely to build large sglists
1466 * out of smaller entries.
1467 */
1468 mr_page_shift = max(12, ffs(ib_dev->attrs.page_size_cap) - 1);
1469 max_pages_per_mr = ib_dev->attrs.max_mr_size;
1470 do_div(max_pages_per_mr, (1ull << mr_page_shift));
1471 clt_path->max_pages_per_mr =
1472 min3(clt_path->max_pages_per_mr, (u32)max_pages_per_mr,
1473 ib_dev->attrs.max_fast_reg_page_list_len);
1474 clt_path->clt->max_segments =
1475 min(clt_path->max_pages_per_mr, clt_path->clt->max_segments);
1476 }
1477
rtrs_clt_change_state_get_old(struct rtrs_clt_path * clt_path,enum rtrs_clt_state new_state,enum rtrs_clt_state * old_state)1478 static bool rtrs_clt_change_state_get_old(struct rtrs_clt_path *clt_path,
1479 enum rtrs_clt_state new_state,
1480 enum rtrs_clt_state *old_state)
1481 {
1482 bool changed;
1483
1484 spin_lock_irq(&clt_path->state_wq.lock);
1485 if (old_state)
1486 *old_state = clt_path->state;
1487 changed = rtrs_clt_change_state(clt_path, new_state);
1488 spin_unlock_irq(&clt_path->state_wq.lock);
1489
1490 return changed;
1491 }
1492
rtrs_clt_hb_err_handler(struct rtrs_con * c)1493 static void rtrs_clt_hb_err_handler(struct rtrs_con *c)
1494 {
1495 struct rtrs_clt_con *con = container_of(c, typeof(*con), c);
1496
1497 rtrs_rdma_error_recovery(con);
1498 }
1499
rtrs_clt_init_hb(struct rtrs_clt_path * clt_path)1500 static void rtrs_clt_init_hb(struct rtrs_clt_path *clt_path)
1501 {
1502 rtrs_init_hb(&clt_path->s, &io_comp_cqe,
1503 RTRS_HB_INTERVAL_MS,
1504 RTRS_HB_MISSED_MAX,
1505 rtrs_clt_hb_err_handler,
1506 rtrs_wq);
1507 }
1508
1509 static void rtrs_clt_reconnect_work(struct work_struct *work);
1510 static void rtrs_clt_close_work(struct work_struct *work);
1511
alloc_path(struct rtrs_clt * clt,const struct rtrs_addr * path,size_t con_num,u32 nr_poll_queues)1512 static struct rtrs_clt_path *alloc_path(struct rtrs_clt *clt,
1513 const struct rtrs_addr *path,
1514 size_t con_num, u32 nr_poll_queues)
1515 {
1516 struct rtrs_clt_path *clt_path;
1517 int err = -ENOMEM;
1518 int cpu;
1519 size_t total_con;
1520
1521 clt_path = kzalloc(sizeof(*clt_path), GFP_KERNEL);
1522 if (!clt_path)
1523 goto err;
1524
1525 /*
1526 * irqmode and poll
1527 * +1: Extra connection for user messages
1528 */
1529 total_con = con_num + nr_poll_queues + 1;
1530 clt_path->s.con = kcalloc(total_con, sizeof(*clt_path->s.con),
1531 GFP_KERNEL);
1532 if (!clt_path->s.con)
1533 goto err_free_path;
1534
1535 clt_path->s.con_num = total_con;
1536 clt_path->s.irq_con_num = con_num + 1;
1537
1538 clt_path->stats = kzalloc(sizeof(*clt_path->stats), GFP_KERNEL);
1539 if (!clt_path->stats)
1540 goto err_free_con;
1541
1542 mutex_init(&clt_path->init_mutex);
1543 uuid_gen(&clt_path->s.uuid);
1544 memcpy(&clt_path->s.dst_addr, path->dst,
1545 rdma_addr_size((struct sockaddr *)path->dst));
1546
1547 /*
1548 * rdma_resolve_addr() passes src_addr to cma_bind_addr, which
1549 * checks the sa_family to be non-zero. If user passed src_addr=NULL
1550 * the sess->src_addr will contain only zeros, which is then fine.
1551 */
1552 if (path->src)
1553 memcpy(&clt_path->s.src_addr, path->src,
1554 rdma_addr_size((struct sockaddr *)path->src));
1555 strscpy(clt_path->s.sessname, clt->sessname,
1556 sizeof(clt_path->s.sessname));
1557 clt_path->clt = clt;
1558 clt_path->max_pages_per_mr = RTRS_MAX_SEGMENTS;
1559 init_waitqueue_head(&clt_path->state_wq);
1560 clt_path->state = RTRS_CLT_CONNECTING;
1561 atomic_set(&clt_path->connected_cnt, 0);
1562 INIT_WORK(&clt_path->close_work, rtrs_clt_close_work);
1563 INIT_DELAYED_WORK(&clt_path->reconnect_dwork, rtrs_clt_reconnect_work);
1564 rtrs_clt_init_hb(clt_path);
1565
1566 clt_path->mp_skip_entry = alloc_percpu(typeof(*clt_path->mp_skip_entry));
1567 if (!clt_path->mp_skip_entry)
1568 goto err_free_stats;
1569
1570 for_each_possible_cpu(cpu)
1571 INIT_LIST_HEAD(per_cpu_ptr(clt_path->mp_skip_entry, cpu));
1572
1573 err = rtrs_clt_init_stats(clt_path->stats);
1574 if (err)
1575 goto err_free_percpu;
1576
1577 return clt_path;
1578
1579 err_free_percpu:
1580 free_percpu(clt_path->mp_skip_entry);
1581 err_free_stats:
1582 kfree(clt_path->stats);
1583 err_free_con:
1584 kfree(clt_path->s.con);
1585 err_free_path:
1586 kfree(clt_path);
1587 err:
1588 return ERR_PTR(err);
1589 }
1590
free_path(struct rtrs_clt_path * clt_path)1591 void free_path(struct rtrs_clt_path *clt_path)
1592 {
1593 free_percpu(clt_path->mp_skip_entry);
1594 mutex_destroy(&clt_path->init_mutex);
1595 kfree(clt_path->s.con);
1596 kfree(clt_path->rbufs);
1597 kfree(clt_path);
1598 }
1599
create_con(struct rtrs_clt_path * clt_path,unsigned int cid)1600 static int create_con(struct rtrs_clt_path *clt_path, unsigned int cid)
1601 {
1602 struct rtrs_clt_con *con;
1603
1604 con = kzalloc(sizeof(*con), GFP_KERNEL);
1605 if (!con)
1606 return -ENOMEM;
1607
1608 /* Map first two connections to the first CPU */
1609 con->cpu = (cid ? cid - 1 : 0) % nr_cpu_ids;
1610 con->c.cid = cid;
1611 con->c.path = &clt_path->s;
1612 /* Align with srv, init as 1 */
1613 atomic_set(&con->c.wr_cnt, 1);
1614 mutex_init(&con->con_mutex);
1615
1616 clt_path->s.con[cid] = &con->c;
1617
1618 return 0;
1619 }
1620
destroy_con(struct rtrs_clt_con * con)1621 static void destroy_con(struct rtrs_clt_con *con)
1622 {
1623 struct rtrs_clt_path *clt_path = to_clt_path(con->c.path);
1624
1625 clt_path->s.con[con->c.cid] = NULL;
1626 mutex_destroy(&con->con_mutex);
1627 kfree(con);
1628 }
1629
create_con_cq_qp(struct rtrs_clt_con * con)1630 static int create_con_cq_qp(struct rtrs_clt_con *con)
1631 {
1632 struct rtrs_clt_path *clt_path = to_clt_path(con->c.path);
1633 u32 max_send_wr, max_recv_wr, cq_num, max_send_sge, wr_limit;
1634 int err, cq_vector;
1635 struct rtrs_msg_rkey_rsp *rsp;
1636
1637 lockdep_assert_held(&con->con_mutex);
1638 if (con->c.cid == 0) {
1639 max_send_sge = 1;
1640 /* We must be the first here */
1641 if (WARN_ON(clt_path->s.dev))
1642 return -EINVAL;
1643
1644 /*
1645 * The whole session uses device from user connection.
1646 * Be careful not to close user connection before ib dev
1647 * is gracefully put.
1648 */
1649 clt_path->s.dev = rtrs_ib_dev_find_or_add(con->c.cm_id->device,
1650 &dev_pd);
1651 if (!clt_path->s.dev) {
1652 rtrs_wrn(clt_path->clt,
1653 "rtrs_ib_dev_find_get_or_add(): no memory\n");
1654 return -ENOMEM;
1655 }
1656 clt_path->s.dev_ref = 1;
1657 query_fast_reg_mode(clt_path);
1658 wr_limit = clt_path->s.dev->ib_dev->attrs.max_qp_wr;
1659 /*
1660 * Two (request + registration) completion for send
1661 * Two for recv if always_invalidate is set on server
1662 * or one for recv.
1663 * + 2 for drain and heartbeat
1664 * in case qp gets into error state.
1665 */
1666 max_send_wr =
1667 min_t(int, wr_limit, SERVICE_CON_QUEUE_DEPTH * 2 + 2);
1668 max_recv_wr = max_send_wr;
1669 } else {
1670 /*
1671 * Here we assume that session members are correctly set.
1672 * This is always true if user connection (cid == 0) is
1673 * established first.
1674 */
1675 if (WARN_ON(!clt_path->s.dev))
1676 return -EINVAL;
1677 if (WARN_ON(!clt_path->queue_depth))
1678 return -EINVAL;
1679
1680 wr_limit = clt_path->s.dev->ib_dev->attrs.max_qp_wr;
1681 /* Shared between connections */
1682 clt_path->s.dev_ref++;
1683 max_send_wr = min_t(int, wr_limit,
1684 /* QD * (REQ + RSP + FR REGS or INVS) + drain */
1685 clt_path->queue_depth * 4 + 1);
1686 max_recv_wr = min_t(int, wr_limit,
1687 clt_path->queue_depth * 3 + 1);
1688 max_send_sge = 2;
1689 }
1690 atomic_set(&con->c.sq_wr_avail, max_send_wr);
1691 cq_num = max_send_wr + max_recv_wr;
1692 /* alloc iu to recv new rkey reply when server reports flags set */
1693 if (clt_path->flags & RTRS_MSG_NEW_RKEY_F || con->c.cid == 0) {
1694 con->rsp_ius = rtrs_iu_alloc(cq_num, sizeof(*rsp),
1695 GFP_KERNEL,
1696 clt_path->s.dev->ib_dev,
1697 DMA_FROM_DEVICE,
1698 rtrs_clt_rdma_done);
1699 if (!con->rsp_ius)
1700 return -ENOMEM;
1701 con->queue_num = cq_num;
1702 }
1703 cq_num = max_send_wr + max_recv_wr;
1704 cq_vector = con->cpu % clt_path->s.dev->ib_dev->num_comp_vectors;
1705 if (con->c.cid >= clt_path->s.irq_con_num)
1706 err = rtrs_cq_qp_create(&clt_path->s, &con->c, max_send_sge,
1707 cq_vector, cq_num, max_send_wr,
1708 max_recv_wr, IB_POLL_DIRECT);
1709 else
1710 err = rtrs_cq_qp_create(&clt_path->s, &con->c, max_send_sge,
1711 cq_vector, cq_num, max_send_wr,
1712 max_recv_wr, IB_POLL_SOFTIRQ);
1713 /*
1714 * In case of error we do not bother to clean previous allocations,
1715 * since destroy_con_cq_qp() must be called.
1716 */
1717 return err;
1718 }
1719
destroy_con_cq_qp(struct rtrs_clt_con * con)1720 static void destroy_con_cq_qp(struct rtrs_clt_con *con)
1721 {
1722 struct rtrs_clt_path *clt_path = to_clt_path(con->c.path);
1723
1724 /*
1725 * Be careful here: destroy_con_cq_qp() can be called even
1726 * create_con_cq_qp() failed, see comments there.
1727 */
1728 lockdep_assert_held(&con->con_mutex);
1729 rtrs_cq_qp_destroy(&con->c);
1730 if (con->rsp_ius) {
1731 rtrs_iu_free(con->rsp_ius, clt_path->s.dev->ib_dev,
1732 con->queue_num);
1733 con->rsp_ius = NULL;
1734 con->queue_num = 0;
1735 }
1736 if (clt_path->s.dev_ref && !--clt_path->s.dev_ref) {
1737 rtrs_ib_dev_put(clt_path->s.dev);
1738 clt_path->s.dev = NULL;
1739 }
1740 }
1741
stop_cm(struct rtrs_clt_con * con)1742 static void stop_cm(struct rtrs_clt_con *con)
1743 {
1744 rdma_disconnect(con->c.cm_id);
1745 if (con->c.qp)
1746 ib_drain_qp(con->c.qp);
1747 }
1748
destroy_cm(struct rtrs_clt_con * con)1749 static void destroy_cm(struct rtrs_clt_con *con)
1750 {
1751 rdma_destroy_id(con->c.cm_id);
1752 con->c.cm_id = NULL;
1753 }
1754
rtrs_rdma_addr_resolved(struct rtrs_clt_con * con)1755 static int rtrs_rdma_addr_resolved(struct rtrs_clt_con *con)
1756 {
1757 struct rtrs_path *s = con->c.path;
1758 int err;
1759
1760 mutex_lock(&con->con_mutex);
1761 err = create_con_cq_qp(con);
1762 mutex_unlock(&con->con_mutex);
1763 if (err) {
1764 rtrs_err(s, "create_con_cq_qp(), err: %d\n", err);
1765 return err;
1766 }
1767 err = rdma_resolve_route(con->c.cm_id, RTRS_CONNECT_TIMEOUT_MS);
1768 if (err)
1769 rtrs_err(s, "Resolving route failed, err: %d\n", err);
1770
1771 return err;
1772 }
1773
rtrs_rdma_route_resolved(struct rtrs_clt_con * con)1774 static int rtrs_rdma_route_resolved(struct rtrs_clt_con *con)
1775 {
1776 struct rtrs_clt_path *clt_path = to_clt_path(con->c.path);
1777 struct rtrs_clt *clt = clt_path->clt;
1778 struct rtrs_msg_conn_req msg;
1779 struct rdma_conn_param param;
1780
1781 int err;
1782
1783 param = (struct rdma_conn_param) {
1784 .retry_count = 7,
1785 .rnr_retry_count = 7,
1786 .private_data = &msg,
1787 .private_data_len = sizeof(msg),
1788 };
1789
1790 msg = (struct rtrs_msg_conn_req) {
1791 .magic = cpu_to_le16(RTRS_MAGIC),
1792 .version = cpu_to_le16(RTRS_PROTO_VER),
1793 .cid = cpu_to_le16(con->c.cid),
1794 .cid_num = cpu_to_le16(clt_path->s.con_num),
1795 .recon_cnt = cpu_to_le16(clt_path->s.recon_cnt),
1796 };
1797 msg.first_conn = clt_path->for_new_clt ? FIRST_CONN : 0;
1798 uuid_copy(&msg.sess_uuid, &clt_path->s.uuid);
1799 uuid_copy(&msg.paths_uuid, &clt->paths_uuid);
1800
1801 err = rdma_connect_locked(con->c.cm_id, ¶m);
1802 if (err)
1803 rtrs_err(clt, "rdma_connect_locked(): %d\n", err);
1804
1805 return err;
1806 }
1807
rtrs_rdma_conn_established(struct rtrs_clt_con * con,struct rdma_cm_event * ev)1808 static int rtrs_rdma_conn_established(struct rtrs_clt_con *con,
1809 struct rdma_cm_event *ev)
1810 {
1811 struct rtrs_clt_path *clt_path = to_clt_path(con->c.path);
1812 struct rtrs_clt *clt = clt_path->clt;
1813 const struct rtrs_msg_conn_rsp *msg;
1814 u16 version, queue_depth;
1815 int errno;
1816 u8 len;
1817
1818 msg = ev->param.conn.private_data;
1819 len = ev->param.conn.private_data_len;
1820 if (len < sizeof(*msg)) {
1821 rtrs_err(clt, "Invalid RTRS connection response\n");
1822 return -ECONNRESET;
1823 }
1824 if (le16_to_cpu(msg->magic) != RTRS_MAGIC) {
1825 rtrs_err(clt, "Invalid RTRS magic\n");
1826 return -ECONNRESET;
1827 }
1828 version = le16_to_cpu(msg->version);
1829 if (version >> 8 != RTRS_PROTO_VER_MAJOR) {
1830 rtrs_err(clt, "Unsupported major RTRS version: %d, expected %d\n",
1831 version >> 8, RTRS_PROTO_VER_MAJOR);
1832 return -ECONNRESET;
1833 }
1834 errno = le16_to_cpu(msg->errno);
1835 if (errno) {
1836 rtrs_err(clt, "Invalid RTRS message: errno %d\n",
1837 errno);
1838 return -ECONNRESET;
1839 }
1840 if (con->c.cid == 0) {
1841 queue_depth = le16_to_cpu(msg->queue_depth);
1842
1843 if (clt_path->queue_depth > 0 && queue_depth != clt_path->queue_depth) {
1844 rtrs_err(clt, "Error: queue depth changed\n");
1845
1846 /*
1847 * Stop any more reconnection attempts
1848 */
1849 clt_path->reconnect_attempts = -1;
1850 rtrs_err(clt,
1851 "Disabling auto-reconnect. Trigger a manual reconnect after issue is resolved\n");
1852 return -ECONNRESET;
1853 }
1854
1855 if (!clt_path->rbufs) {
1856 clt_path->rbufs = kcalloc(queue_depth,
1857 sizeof(*clt_path->rbufs),
1858 GFP_KERNEL);
1859 if (!clt_path->rbufs)
1860 return -ENOMEM;
1861 }
1862 clt_path->queue_depth = queue_depth;
1863 clt_path->s.signal_interval = min_not_zero(queue_depth,
1864 (unsigned short) SERVICE_CON_QUEUE_DEPTH);
1865 clt_path->max_hdr_size = le32_to_cpu(msg->max_hdr_size);
1866 clt_path->max_io_size = le32_to_cpu(msg->max_io_size);
1867 clt_path->flags = le32_to_cpu(msg->flags);
1868 clt_path->chunk_size = clt_path->max_io_size + clt_path->max_hdr_size;
1869
1870 /*
1871 * Global IO size is always a minimum.
1872 * If while a reconnection server sends us a value a bit
1873 * higher - client does not care and uses cached minimum.
1874 *
1875 * Since we can have several sessions (paths) restablishing
1876 * connections in parallel, use lock.
1877 */
1878 mutex_lock(&clt->paths_mutex);
1879 clt->queue_depth = clt_path->queue_depth;
1880 clt->max_io_size = min_not_zero(clt_path->max_io_size,
1881 clt->max_io_size);
1882 mutex_unlock(&clt->paths_mutex);
1883
1884 /*
1885 * Cache the hca_port and hca_name for sysfs
1886 */
1887 clt_path->hca_port = con->c.cm_id->port_num;
1888 scnprintf(clt_path->hca_name, sizeof(clt_path->hca_name),
1889 clt_path->s.dev->ib_dev->name);
1890 clt_path->s.src_addr = con->c.cm_id->route.addr.src_addr;
1891 /* set for_new_clt, to allow future reconnect on any path */
1892 clt_path->for_new_clt = 1;
1893 }
1894
1895 return 0;
1896 }
1897
flag_success_on_conn(struct rtrs_clt_con * con)1898 static inline void flag_success_on_conn(struct rtrs_clt_con *con)
1899 {
1900 struct rtrs_clt_path *clt_path = to_clt_path(con->c.path);
1901
1902 atomic_inc(&clt_path->connected_cnt);
1903 con->cm_err = 1;
1904 }
1905
rtrs_rdma_conn_rejected(struct rtrs_clt_con * con,struct rdma_cm_event * ev)1906 static int rtrs_rdma_conn_rejected(struct rtrs_clt_con *con,
1907 struct rdma_cm_event *ev)
1908 {
1909 struct rtrs_path *s = con->c.path;
1910 const struct rtrs_msg_conn_rsp *msg;
1911 const char *rej_msg;
1912 int status, errno;
1913 u8 data_len;
1914
1915 status = ev->status;
1916 rej_msg = rdma_reject_msg(con->c.cm_id, status);
1917 msg = rdma_consumer_reject_data(con->c.cm_id, ev, &data_len);
1918
1919 if (msg && data_len >= sizeof(*msg)) {
1920 errno = (int16_t)le16_to_cpu(msg->errno);
1921 if (errno == -EBUSY)
1922 rtrs_err(s,
1923 "Previous session is still exists on the server, please reconnect later\n");
1924 else
1925 rtrs_err(s,
1926 "Connect rejected: status %d (%s), rtrs errno %d\n",
1927 status, rej_msg, errno);
1928 } else {
1929 rtrs_err(s,
1930 "Connect rejected but with malformed message: status %d (%s)\n",
1931 status, rej_msg);
1932 }
1933
1934 return -ECONNRESET;
1935 }
1936
rtrs_clt_close_conns(struct rtrs_clt_path * clt_path,bool wait)1937 void rtrs_clt_close_conns(struct rtrs_clt_path *clt_path, bool wait)
1938 {
1939 if (rtrs_clt_change_state_get_old(clt_path, RTRS_CLT_CLOSING, NULL))
1940 queue_work(rtrs_wq, &clt_path->close_work);
1941 if (wait)
1942 flush_work(&clt_path->close_work);
1943 }
1944
flag_error_on_conn(struct rtrs_clt_con * con,int cm_err)1945 static inline void flag_error_on_conn(struct rtrs_clt_con *con, int cm_err)
1946 {
1947 if (con->cm_err == 1) {
1948 struct rtrs_clt_path *clt_path;
1949
1950 clt_path = to_clt_path(con->c.path);
1951 if (atomic_dec_and_test(&clt_path->connected_cnt))
1952
1953 wake_up(&clt_path->state_wq);
1954 }
1955 con->cm_err = cm_err;
1956 }
1957
rtrs_clt_rdma_cm_handler(struct rdma_cm_id * cm_id,struct rdma_cm_event * ev)1958 static int rtrs_clt_rdma_cm_handler(struct rdma_cm_id *cm_id,
1959 struct rdma_cm_event *ev)
1960 {
1961 struct rtrs_clt_con *con = cm_id->context;
1962 struct rtrs_path *s = con->c.path;
1963 struct rtrs_clt_path *clt_path = to_clt_path(s);
1964 int cm_err = 0;
1965
1966 switch (ev->event) {
1967 case RDMA_CM_EVENT_ADDR_RESOLVED:
1968 cm_err = rtrs_rdma_addr_resolved(con);
1969 break;
1970 case RDMA_CM_EVENT_ROUTE_RESOLVED:
1971 cm_err = rtrs_rdma_route_resolved(con);
1972 break;
1973 case RDMA_CM_EVENT_ESTABLISHED:
1974 cm_err = rtrs_rdma_conn_established(con, ev);
1975 if (!cm_err) {
1976 /*
1977 * Report success and wake up. Here we abuse state_wq,
1978 * i.e. wake up without state change, but we set cm_err.
1979 */
1980 flag_success_on_conn(con);
1981 wake_up(&clt_path->state_wq);
1982 return 0;
1983 }
1984 break;
1985 case RDMA_CM_EVENT_REJECTED:
1986 cm_err = rtrs_rdma_conn_rejected(con, ev);
1987 break;
1988 case RDMA_CM_EVENT_DISCONNECTED:
1989 /* No message for disconnecting */
1990 cm_err = -ECONNRESET;
1991 break;
1992 case RDMA_CM_EVENT_CONNECT_ERROR:
1993 case RDMA_CM_EVENT_UNREACHABLE:
1994 case RDMA_CM_EVENT_ADDR_CHANGE:
1995 case RDMA_CM_EVENT_TIMEWAIT_EXIT:
1996 rtrs_wrn(s, "CM error (CM event: %s, err: %d)\n",
1997 rdma_event_msg(ev->event), ev->status);
1998 cm_err = -ECONNRESET;
1999 break;
2000 case RDMA_CM_EVENT_ADDR_ERROR:
2001 case RDMA_CM_EVENT_ROUTE_ERROR:
2002 rtrs_wrn(s, "CM error (CM event: %s, err: %d)\n",
2003 rdma_event_msg(ev->event), ev->status);
2004 cm_err = -EHOSTUNREACH;
2005 break;
2006 case RDMA_CM_EVENT_DEVICE_REMOVAL:
2007 /*
2008 * Device removal is a special case. Queue close and return 0.
2009 */
2010 rtrs_clt_close_conns(clt_path, false);
2011 return 0;
2012 default:
2013 rtrs_err(s, "Unexpected RDMA CM error (CM event: %s, err: %d)\n",
2014 rdma_event_msg(ev->event), ev->status);
2015 cm_err = -ECONNRESET;
2016 break;
2017 }
2018
2019 if (cm_err) {
2020 /*
2021 * cm error makes sense only on connection establishing,
2022 * in other cases we rely on normal procedure of reconnecting.
2023 */
2024 flag_error_on_conn(con, cm_err);
2025 rtrs_rdma_error_recovery(con);
2026 }
2027
2028 return 0;
2029 }
2030
2031 /* The caller should do the cleanup in case of error */
create_cm(struct rtrs_clt_con * con)2032 static int create_cm(struct rtrs_clt_con *con)
2033 {
2034 struct rtrs_path *s = con->c.path;
2035 struct rtrs_clt_path *clt_path = to_clt_path(s);
2036 struct rdma_cm_id *cm_id;
2037 int err;
2038
2039 cm_id = rdma_create_id(&init_net, rtrs_clt_rdma_cm_handler, con,
2040 clt_path->s.dst_addr.ss_family == AF_IB ?
2041 RDMA_PS_IB : RDMA_PS_TCP, IB_QPT_RC);
2042 if (IS_ERR(cm_id)) {
2043 err = PTR_ERR(cm_id);
2044 rtrs_err(s, "Failed to create CM ID, err: %d\n", err);
2045
2046 return err;
2047 }
2048 con->c.cm_id = cm_id;
2049 con->cm_err = 0;
2050 /* allow the port to be reused */
2051 err = rdma_set_reuseaddr(cm_id, 1);
2052 if (err != 0) {
2053 rtrs_err(s, "Set address reuse failed, err: %d\n", err);
2054 return err;
2055 }
2056 err = rdma_resolve_addr(cm_id, (struct sockaddr *)&clt_path->s.src_addr,
2057 (struct sockaddr *)&clt_path->s.dst_addr,
2058 RTRS_CONNECT_TIMEOUT_MS);
2059 if (err) {
2060 rtrs_err(s, "Failed to resolve address, err: %d\n", err);
2061 return err;
2062 }
2063 /*
2064 * Combine connection status and session events. This is needed
2065 * for waiting two possible cases: cm_err has something meaningful
2066 * or session state was really changed to error by device removal.
2067 */
2068 err = wait_event_interruptible_timeout(
2069 clt_path->state_wq,
2070 con->cm_err || clt_path->state != RTRS_CLT_CONNECTING,
2071 msecs_to_jiffies(RTRS_CONNECT_TIMEOUT_MS));
2072 if (err == 0 || err == -ERESTARTSYS) {
2073 if (err == 0)
2074 err = -ETIMEDOUT;
2075 /* Timedout or interrupted */
2076 return err;
2077 }
2078 if (con->cm_err < 0)
2079 return con->cm_err;
2080 if (READ_ONCE(clt_path->state) != RTRS_CLT_CONNECTING)
2081 /* Device removal */
2082 return -ECONNABORTED;
2083
2084 return 0;
2085 }
2086
rtrs_clt_path_up(struct rtrs_clt_path * clt_path)2087 static void rtrs_clt_path_up(struct rtrs_clt_path *clt_path)
2088 {
2089 struct rtrs_clt *clt = clt_path->clt;
2090 int up;
2091
2092 /*
2093 * We can fire RECONNECTED event only when all paths were
2094 * connected on rtrs_clt_open(), then each was disconnected
2095 * and the first one connected again. That's why this nasty
2096 * game with counter value.
2097 */
2098
2099 mutex_lock(&clt->paths_ev_mutex);
2100 up = ++clt->paths_up;
2101 /*
2102 * Here it is safe to access paths num directly since up counter
2103 * is greater than MAX_PATHS_NUM only while rtrs_clt_open() is
2104 * in progress, thus paths removals are impossible.
2105 */
2106 if (up > MAX_PATHS_NUM && up == MAX_PATHS_NUM + clt->paths_num)
2107 clt->paths_up = clt->paths_num;
2108 else if (up == 1)
2109 clt->link_ev(clt->priv, RTRS_CLT_LINK_EV_RECONNECTED);
2110 mutex_unlock(&clt->paths_ev_mutex);
2111
2112 /* Mark session as established */
2113 clt_path->established = true;
2114 clt_path->reconnect_attempts = 0;
2115 clt_path->stats->reconnects.successful_cnt++;
2116 }
2117
rtrs_clt_path_down(struct rtrs_clt_path * clt_path)2118 static void rtrs_clt_path_down(struct rtrs_clt_path *clt_path)
2119 {
2120 struct rtrs_clt *clt = clt_path->clt;
2121
2122 if (!clt_path->established)
2123 return;
2124
2125 clt_path->established = false;
2126 mutex_lock(&clt->paths_ev_mutex);
2127 WARN_ON(!clt->paths_up);
2128 if (--clt->paths_up == 0)
2129 clt->link_ev(clt->priv, RTRS_CLT_LINK_EV_DISCONNECTED);
2130 mutex_unlock(&clt->paths_ev_mutex);
2131 }
2132
rtrs_clt_stop_and_destroy_conns(struct rtrs_clt_path * clt_path)2133 static void rtrs_clt_stop_and_destroy_conns(struct rtrs_clt_path *clt_path)
2134 {
2135 struct rtrs_clt_con *con;
2136 unsigned int cid;
2137
2138 WARN_ON(READ_ONCE(clt_path->state) == RTRS_CLT_CONNECTED);
2139
2140 /*
2141 * Possible race with rtrs_clt_open(), when DEVICE_REMOVAL comes
2142 * exactly in between. Start destroying after it finishes.
2143 */
2144 mutex_lock(&clt_path->init_mutex);
2145 mutex_unlock(&clt_path->init_mutex);
2146
2147 /*
2148 * All IO paths must observe !CONNECTED state before we
2149 * free everything.
2150 */
2151 synchronize_rcu();
2152
2153 rtrs_stop_hb(&clt_path->s);
2154
2155 /*
2156 * The order it utterly crucial: firstly disconnect and complete all
2157 * rdma requests with error (thus set in_use=false for requests),
2158 * then fail outstanding requests checking in_use for each, and
2159 * eventually notify upper layer about session disconnection.
2160 */
2161
2162 for (cid = 0; cid < clt_path->s.con_num; cid++) {
2163 if (!clt_path->s.con[cid])
2164 break;
2165 con = to_clt_con(clt_path->s.con[cid]);
2166 stop_cm(con);
2167 }
2168 fail_all_outstanding_reqs(clt_path);
2169 free_path_reqs(clt_path);
2170 rtrs_clt_path_down(clt_path);
2171
2172 /*
2173 * Wait for graceful shutdown, namely when peer side invokes
2174 * rdma_disconnect(). 'connected_cnt' is decremented only on
2175 * CM events, thus if other side had crashed and hb has detected
2176 * something is wrong, here we will stuck for exactly timeout ms,
2177 * since CM does not fire anything. That is fine, we are not in
2178 * hurry.
2179 */
2180 wait_event_timeout(clt_path->state_wq,
2181 !atomic_read(&clt_path->connected_cnt),
2182 msecs_to_jiffies(RTRS_CONNECT_TIMEOUT_MS));
2183
2184 for (cid = 0; cid < clt_path->s.con_num; cid++) {
2185 if (!clt_path->s.con[cid])
2186 break;
2187 con = to_clt_con(clt_path->s.con[cid]);
2188 mutex_lock(&con->con_mutex);
2189 destroy_con_cq_qp(con);
2190 mutex_unlock(&con->con_mutex);
2191 destroy_cm(con);
2192 destroy_con(con);
2193 }
2194 }
2195
xchg_paths(struct rtrs_clt_path __rcu ** rcu_ppcpu_path,struct rtrs_clt_path * clt_path,struct rtrs_clt_path * next)2196 static inline bool xchg_paths(struct rtrs_clt_path __rcu **rcu_ppcpu_path,
2197 struct rtrs_clt_path *clt_path,
2198 struct rtrs_clt_path *next)
2199 {
2200 struct rtrs_clt_path **ppcpu_path;
2201
2202 /* Call cmpxchg() without sparse warnings */
2203 ppcpu_path = (typeof(ppcpu_path))rcu_ppcpu_path;
2204 return clt_path == cmpxchg(ppcpu_path, clt_path, next);
2205 }
2206
rtrs_clt_remove_path_from_arr(struct rtrs_clt_path * clt_path)2207 static void rtrs_clt_remove_path_from_arr(struct rtrs_clt_path *clt_path)
2208 {
2209 struct rtrs_clt *clt = clt_path->clt;
2210 struct rtrs_clt_path *next;
2211 bool wait_for_grace = false;
2212 int cpu;
2213
2214 mutex_lock(&clt->paths_mutex);
2215 list_del_rcu(&clt_path->s.entry);
2216
2217 /* Make sure everybody observes path removal. */
2218 synchronize_rcu();
2219
2220 /*
2221 * At this point nobody sees @sess in the list, but still we have
2222 * dangling pointer @pcpu_path which _can_ point to @sess. Since
2223 * nobody can observe @sess in the list, we guarantee that IO path
2224 * will not assign @sess to @pcpu_path, i.e. @pcpu_path can be equal
2225 * to @sess, but can never again become @sess.
2226 */
2227
2228 /*
2229 * Decrement paths number only after grace period, because
2230 * caller of do_each_path() must firstly observe list without
2231 * path and only then decremented paths number.
2232 *
2233 * Otherwise there can be the following situation:
2234 * o Two paths exist and IO is coming.
2235 * o One path is removed:
2236 * CPU#0 CPU#1
2237 * do_each_path(): rtrs_clt_remove_path_from_arr():
2238 * path = get_next_path()
2239 * ^^^ list_del_rcu(path)
2240 * [!CONNECTED path] clt->paths_num--
2241 * ^^^^^^^^^
2242 * load clt->paths_num from 2 to 1
2243 * ^^^^^^^^^
2244 * sees 1
2245 *
2246 * path is observed as !CONNECTED, but do_each_path() loop
2247 * ends, because expression i < clt->paths_num is false.
2248 */
2249 clt->paths_num--;
2250
2251 /*
2252 * Get @next connection from current @sess which is going to be
2253 * removed. If @sess is the last element, then @next is NULL.
2254 */
2255 rcu_read_lock();
2256 next = rtrs_clt_get_next_path_or_null(&clt->paths_list, clt_path);
2257 rcu_read_unlock();
2258
2259 /*
2260 * @pcpu paths can still point to the path which is going to be
2261 * removed, so change the pointer manually.
2262 */
2263 for_each_possible_cpu(cpu) {
2264 struct rtrs_clt_path __rcu **ppcpu_path;
2265
2266 ppcpu_path = per_cpu_ptr(clt->pcpu_path, cpu);
2267 if (rcu_dereference_protected(*ppcpu_path,
2268 lockdep_is_held(&clt->paths_mutex)) != clt_path)
2269 /*
2270 * synchronize_rcu() was called just after deleting
2271 * entry from the list, thus IO code path cannot
2272 * change pointer back to the pointer which is going
2273 * to be removed, we are safe here.
2274 */
2275 continue;
2276
2277 /*
2278 * We race with IO code path, which also changes pointer,
2279 * thus we have to be careful not to overwrite it.
2280 */
2281 if (xchg_paths(ppcpu_path, clt_path, next))
2282 /*
2283 * @ppcpu_path was successfully replaced with @next,
2284 * that means that someone could also pick up the
2285 * @sess and dereferencing it right now, so wait for
2286 * a grace period is required.
2287 */
2288 wait_for_grace = true;
2289 }
2290 if (wait_for_grace)
2291 synchronize_rcu();
2292
2293 mutex_unlock(&clt->paths_mutex);
2294 }
2295
rtrs_clt_add_path_to_arr(struct rtrs_clt_path * clt_path)2296 static void rtrs_clt_add_path_to_arr(struct rtrs_clt_path *clt_path)
2297 {
2298 struct rtrs_clt *clt = clt_path->clt;
2299
2300 mutex_lock(&clt->paths_mutex);
2301 clt->paths_num++;
2302
2303 list_add_tail_rcu(&clt_path->s.entry, &clt->paths_list);
2304 mutex_unlock(&clt->paths_mutex);
2305 }
2306
rtrs_clt_close_work(struct work_struct * work)2307 static void rtrs_clt_close_work(struct work_struct *work)
2308 {
2309 struct rtrs_clt_path *clt_path;
2310
2311 clt_path = container_of(work, struct rtrs_clt_path, close_work);
2312
2313 cancel_delayed_work_sync(&clt_path->reconnect_dwork);
2314 rtrs_clt_stop_and_destroy_conns(clt_path);
2315 rtrs_clt_change_state_get_old(clt_path, RTRS_CLT_CLOSED, NULL);
2316 }
2317
init_conns(struct rtrs_clt_path * clt_path)2318 static int init_conns(struct rtrs_clt_path *clt_path)
2319 {
2320 unsigned int cid;
2321 int err, i;
2322
2323 /*
2324 * On every new session connections increase reconnect counter
2325 * to avoid clashes with previous sessions not yet closed
2326 * sessions on a server side.
2327 */
2328 clt_path->s.recon_cnt++;
2329
2330 /* Establish all RDMA connections */
2331 for (cid = 0; cid < clt_path->s.con_num; cid++) {
2332 err = create_con(clt_path, cid);
2333 if (err)
2334 goto destroy;
2335
2336 err = create_cm(to_clt_con(clt_path->s.con[cid]));
2337 if (err)
2338 goto destroy;
2339 }
2340 err = alloc_path_reqs(clt_path);
2341 if (err)
2342 goto destroy;
2343
2344 return 0;
2345
2346 destroy:
2347 /* Make sure we do the cleanup in the order they are created */
2348 for (i = 0; i <= cid; i++) {
2349 struct rtrs_clt_con *con;
2350
2351 if (!clt_path->s.con[i])
2352 break;
2353
2354 con = to_clt_con(clt_path->s.con[i]);
2355 if (con->c.cm_id) {
2356 stop_cm(con);
2357 mutex_lock(&con->con_mutex);
2358 destroy_con_cq_qp(con);
2359 mutex_unlock(&con->con_mutex);
2360 destroy_cm(con);
2361 }
2362 destroy_con(con);
2363 }
2364 /*
2365 * If we've never taken async path and got an error, say,
2366 * doing rdma_resolve_addr(), switch to CONNECTION_ERR state
2367 * manually to keep reconnecting.
2368 */
2369 rtrs_clt_change_state_get_old(clt_path, RTRS_CLT_CONNECTING_ERR, NULL);
2370
2371 return err;
2372 }
2373
rtrs_clt_info_req_done(struct ib_cq * cq,struct ib_wc * wc)2374 static void rtrs_clt_info_req_done(struct ib_cq *cq, struct ib_wc *wc)
2375 {
2376 struct rtrs_clt_con *con = to_clt_con(wc->qp->qp_context);
2377 struct rtrs_clt_path *clt_path = to_clt_path(con->c.path);
2378 struct rtrs_iu *iu;
2379
2380 iu = container_of(wc->wr_cqe, struct rtrs_iu, cqe);
2381 rtrs_iu_free(iu, clt_path->s.dev->ib_dev, 1);
2382
2383 if (wc->status != IB_WC_SUCCESS) {
2384 rtrs_err(clt_path->clt, "Path info request send failed: %s\n",
2385 ib_wc_status_msg(wc->status));
2386 rtrs_clt_change_state_get_old(clt_path, RTRS_CLT_CONNECTING_ERR, NULL);
2387 return;
2388 }
2389
2390 rtrs_clt_update_wc_stats(con);
2391 }
2392
process_info_rsp(struct rtrs_clt_path * clt_path,const struct rtrs_msg_info_rsp * msg)2393 static int process_info_rsp(struct rtrs_clt_path *clt_path,
2394 const struct rtrs_msg_info_rsp *msg)
2395 {
2396 unsigned int sg_cnt, total_len;
2397 int i, sgi;
2398
2399 sg_cnt = le16_to_cpu(msg->sg_cnt);
2400 if (!sg_cnt || (clt_path->queue_depth % sg_cnt)) {
2401 rtrs_err(clt_path->clt,
2402 "Incorrect sg_cnt %d, is not multiple\n",
2403 sg_cnt);
2404 return -EINVAL;
2405 }
2406
2407 /*
2408 * Check if IB immediate data size is enough to hold the mem_id and
2409 * the offset inside the memory chunk.
2410 */
2411 if ((ilog2(sg_cnt - 1) + 1) + (ilog2(clt_path->chunk_size - 1) + 1) >
2412 MAX_IMM_PAYL_BITS) {
2413 rtrs_err(clt_path->clt,
2414 "RDMA immediate size (%db) not enough to encode %d buffers of size %dB\n",
2415 MAX_IMM_PAYL_BITS, sg_cnt, clt_path->chunk_size);
2416 return -EINVAL;
2417 }
2418 total_len = 0;
2419 for (sgi = 0, i = 0; sgi < sg_cnt && i < clt_path->queue_depth; sgi++) {
2420 const struct rtrs_sg_desc *desc = &msg->desc[sgi];
2421 u32 len, rkey;
2422 u64 addr;
2423
2424 addr = le64_to_cpu(desc->addr);
2425 rkey = le32_to_cpu(desc->key);
2426 len = le32_to_cpu(desc->len);
2427
2428 total_len += len;
2429
2430 if (!len || (len % clt_path->chunk_size)) {
2431 rtrs_err(clt_path->clt, "Incorrect [%d].len %d\n",
2432 sgi,
2433 len);
2434 return -EINVAL;
2435 }
2436 for ( ; len && i < clt_path->queue_depth; i++) {
2437 clt_path->rbufs[i].addr = addr;
2438 clt_path->rbufs[i].rkey = rkey;
2439
2440 len -= clt_path->chunk_size;
2441 addr += clt_path->chunk_size;
2442 }
2443 }
2444 /* Sanity check */
2445 if (sgi != sg_cnt || i != clt_path->queue_depth) {
2446 rtrs_err(clt_path->clt,
2447 "Incorrect sg vector, not fully mapped\n");
2448 return -EINVAL;
2449 }
2450 if (total_len != clt_path->chunk_size * clt_path->queue_depth) {
2451 rtrs_err(clt_path->clt, "Incorrect total_len %d\n", total_len);
2452 return -EINVAL;
2453 }
2454
2455 return 0;
2456 }
2457
rtrs_clt_info_rsp_done(struct ib_cq * cq,struct ib_wc * wc)2458 static void rtrs_clt_info_rsp_done(struct ib_cq *cq, struct ib_wc *wc)
2459 {
2460 struct rtrs_clt_con *con = to_clt_con(wc->qp->qp_context);
2461 struct rtrs_clt_path *clt_path = to_clt_path(con->c.path);
2462 struct rtrs_msg_info_rsp *msg;
2463 enum rtrs_clt_state state;
2464 struct rtrs_iu *iu;
2465 size_t rx_sz;
2466 int err;
2467
2468 state = RTRS_CLT_CONNECTING_ERR;
2469
2470 WARN_ON(con->c.cid);
2471 iu = container_of(wc->wr_cqe, struct rtrs_iu, cqe);
2472 if (wc->status != IB_WC_SUCCESS) {
2473 rtrs_err(clt_path->clt, "Path info response recv failed: %s\n",
2474 ib_wc_status_msg(wc->status));
2475 goto out;
2476 }
2477 WARN_ON(wc->opcode != IB_WC_RECV);
2478
2479 if (wc->byte_len < sizeof(*msg)) {
2480 rtrs_err(clt_path->clt, "Path info response is malformed: size %d\n",
2481 wc->byte_len);
2482 goto out;
2483 }
2484 ib_dma_sync_single_for_cpu(clt_path->s.dev->ib_dev, iu->dma_addr,
2485 iu->size, DMA_FROM_DEVICE);
2486 msg = iu->buf;
2487 if (le16_to_cpu(msg->type) != RTRS_MSG_INFO_RSP) {
2488 rtrs_err(clt_path->clt, "Path info response is malformed: type %d\n",
2489 le16_to_cpu(msg->type));
2490 goto out;
2491 }
2492 rx_sz = sizeof(*msg);
2493 rx_sz += sizeof(msg->desc[0]) * le16_to_cpu(msg->sg_cnt);
2494 if (wc->byte_len < rx_sz) {
2495 rtrs_err(clt_path->clt, "Path info response is malformed: size %d\n",
2496 wc->byte_len);
2497 goto out;
2498 }
2499 err = process_info_rsp(clt_path, msg);
2500 if (err)
2501 goto out;
2502
2503 err = post_recv_path(clt_path);
2504 if (err)
2505 goto out;
2506
2507 state = RTRS_CLT_CONNECTED;
2508
2509 out:
2510 rtrs_clt_update_wc_stats(con);
2511 rtrs_iu_free(iu, clt_path->s.dev->ib_dev, 1);
2512 rtrs_clt_change_state_get_old(clt_path, state, NULL);
2513 }
2514
rtrs_send_path_info(struct rtrs_clt_path * clt_path)2515 static int rtrs_send_path_info(struct rtrs_clt_path *clt_path)
2516 {
2517 struct rtrs_clt_con *usr_con = to_clt_con(clt_path->s.con[0]);
2518 struct rtrs_msg_info_req *msg;
2519 struct rtrs_iu *tx_iu, *rx_iu;
2520 size_t rx_sz;
2521 int err;
2522
2523 rx_sz = sizeof(struct rtrs_msg_info_rsp);
2524 rx_sz += sizeof(struct rtrs_sg_desc) * clt_path->queue_depth;
2525
2526 tx_iu = rtrs_iu_alloc(1, sizeof(struct rtrs_msg_info_req), GFP_KERNEL,
2527 clt_path->s.dev->ib_dev, DMA_TO_DEVICE,
2528 rtrs_clt_info_req_done);
2529 rx_iu = rtrs_iu_alloc(1, rx_sz, GFP_KERNEL, clt_path->s.dev->ib_dev,
2530 DMA_FROM_DEVICE, rtrs_clt_info_rsp_done);
2531 if (!tx_iu || !rx_iu) {
2532 err = -ENOMEM;
2533 goto out;
2534 }
2535 /* Prepare for getting info response */
2536 err = rtrs_iu_post_recv(&usr_con->c, rx_iu);
2537 if (err) {
2538 rtrs_err(clt_path->clt, "rtrs_iu_post_recv(), err: %d\n", err);
2539 goto out;
2540 }
2541 rx_iu = NULL;
2542
2543 msg = tx_iu->buf;
2544 msg->type = cpu_to_le16(RTRS_MSG_INFO_REQ);
2545 memcpy(msg->pathname, clt_path->s.sessname, sizeof(msg->pathname));
2546
2547 ib_dma_sync_single_for_device(clt_path->s.dev->ib_dev,
2548 tx_iu->dma_addr,
2549 tx_iu->size, DMA_TO_DEVICE);
2550
2551 /* Send info request */
2552 err = rtrs_iu_post_send(&usr_con->c, tx_iu, sizeof(*msg), NULL);
2553 if (err) {
2554 rtrs_err(clt_path->clt, "rtrs_iu_post_send(), err: %d\n", err);
2555 goto out;
2556 }
2557 tx_iu = NULL;
2558
2559 /* Wait for state change */
2560 wait_event_interruptible_timeout(clt_path->state_wq,
2561 clt_path->state != RTRS_CLT_CONNECTING,
2562 msecs_to_jiffies(
2563 RTRS_CONNECT_TIMEOUT_MS));
2564 if (READ_ONCE(clt_path->state) != RTRS_CLT_CONNECTED) {
2565 if (READ_ONCE(clt_path->state) == RTRS_CLT_CONNECTING_ERR)
2566 err = -ECONNRESET;
2567 else
2568 err = -ETIMEDOUT;
2569 }
2570
2571 out:
2572 if (tx_iu)
2573 rtrs_iu_free(tx_iu, clt_path->s.dev->ib_dev, 1);
2574 if (rx_iu)
2575 rtrs_iu_free(rx_iu, clt_path->s.dev->ib_dev, 1);
2576 if (err)
2577 /* If we've never taken async path because of malloc problems */
2578 rtrs_clt_change_state_get_old(clt_path,
2579 RTRS_CLT_CONNECTING_ERR, NULL);
2580
2581 return err;
2582 }
2583
2584 /**
2585 * init_path() - establishes all path connections and does handshake
2586 * @clt_path: client path.
2587 * In case of error full close or reconnect procedure should be taken,
2588 * because reconnect or close async works can be started.
2589 */
init_path(struct rtrs_clt_path * clt_path)2590 static int init_path(struct rtrs_clt_path *clt_path)
2591 {
2592 int err;
2593 char str[NAME_MAX];
2594 struct rtrs_addr path = {
2595 .src = &clt_path->s.src_addr,
2596 .dst = &clt_path->s.dst_addr,
2597 };
2598
2599 rtrs_addr_to_str(&path, str, sizeof(str));
2600
2601 mutex_lock(&clt_path->init_mutex);
2602 err = init_conns(clt_path);
2603 if (err) {
2604 rtrs_err(clt_path->clt,
2605 "init_conns() failed: err=%d path=%s [%s:%u]\n", err,
2606 str, clt_path->hca_name, clt_path->hca_port);
2607 goto out;
2608 }
2609 err = rtrs_send_path_info(clt_path);
2610 if (err) {
2611 rtrs_err(clt_path->clt,
2612 "rtrs_send_path_info() failed: err=%d path=%s [%s:%u]\n",
2613 err, str, clt_path->hca_name, clt_path->hca_port);
2614 goto out;
2615 }
2616 rtrs_clt_path_up(clt_path);
2617 rtrs_start_hb(&clt_path->s);
2618 out:
2619 mutex_unlock(&clt_path->init_mutex);
2620
2621 return err;
2622 }
2623
rtrs_clt_reconnect_work(struct work_struct * work)2624 static void rtrs_clt_reconnect_work(struct work_struct *work)
2625 {
2626 struct rtrs_clt_path *clt_path;
2627 struct rtrs_clt *clt;
2628 unsigned int delay_ms;
2629 int err;
2630
2631 clt_path = container_of(to_delayed_work(work), struct rtrs_clt_path,
2632 reconnect_dwork);
2633 clt = clt_path->clt;
2634
2635 if (READ_ONCE(clt_path->state) != RTRS_CLT_RECONNECTING)
2636 return;
2637
2638 if (clt_path->reconnect_attempts >= clt->max_reconnect_attempts) {
2639 /* Close a path completely if max attempts is reached */
2640 rtrs_clt_close_conns(clt_path, false);
2641 return;
2642 }
2643 clt_path->reconnect_attempts++;
2644
2645 /* Stop everything */
2646 rtrs_clt_stop_and_destroy_conns(clt_path);
2647 msleep(RTRS_RECONNECT_BACKOFF);
2648 if (rtrs_clt_change_state_get_old(clt_path, RTRS_CLT_CONNECTING, NULL)) {
2649 err = init_path(clt_path);
2650 if (err)
2651 goto reconnect_again;
2652 }
2653
2654 return;
2655
2656 reconnect_again:
2657 if (rtrs_clt_change_state_get_old(clt_path, RTRS_CLT_RECONNECTING, NULL)) {
2658 clt_path->stats->reconnects.fail_cnt++;
2659 delay_ms = clt->reconnect_delay_sec * 1000;
2660 queue_delayed_work(rtrs_wq, &clt_path->reconnect_dwork,
2661 msecs_to_jiffies(delay_ms +
2662 prandom_u32() %
2663 RTRS_RECONNECT_SEED));
2664 }
2665 }
2666
rtrs_clt_dev_release(struct device * dev)2667 static void rtrs_clt_dev_release(struct device *dev)
2668 {
2669 struct rtrs_clt *clt = container_of(dev, struct rtrs_clt, dev);
2670
2671 mutex_destroy(&clt->paths_ev_mutex);
2672 mutex_destroy(&clt->paths_mutex);
2673 kfree(clt);
2674 }
2675
alloc_clt(const char * sessname,size_t paths_num,u16 port,size_t pdu_sz,void * priv,void (* link_ev)(void * priv,enum rtrs_clt_link_ev ev),unsigned int reconnect_delay_sec,unsigned int max_reconnect_attempts)2676 static struct rtrs_clt *alloc_clt(const char *sessname, size_t paths_num,
2677 u16 port, size_t pdu_sz, void *priv,
2678 void (*link_ev)(void *priv,
2679 enum rtrs_clt_link_ev ev),
2680 unsigned int reconnect_delay_sec,
2681 unsigned int max_reconnect_attempts)
2682 {
2683 struct rtrs_clt *clt;
2684 int err;
2685
2686 if (!paths_num || paths_num > MAX_PATHS_NUM)
2687 return ERR_PTR(-EINVAL);
2688
2689 if (strlen(sessname) >= sizeof(clt->sessname))
2690 return ERR_PTR(-EINVAL);
2691
2692 clt = kzalloc(sizeof(*clt), GFP_KERNEL);
2693 if (!clt)
2694 return ERR_PTR(-ENOMEM);
2695
2696 clt->pcpu_path = alloc_percpu(typeof(*clt->pcpu_path));
2697 if (!clt->pcpu_path) {
2698 kfree(clt);
2699 return ERR_PTR(-ENOMEM);
2700 }
2701
2702 clt->dev.class = rtrs_clt_dev_class;
2703 clt->dev.release = rtrs_clt_dev_release;
2704 uuid_gen(&clt->paths_uuid);
2705 INIT_LIST_HEAD_RCU(&clt->paths_list);
2706 clt->paths_num = paths_num;
2707 clt->paths_up = MAX_PATHS_NUM;
2708 clt->port = port;
2709 clt->pdu_sz = pdu_sz;
2710 clt->max_segments = RTRS_MAX_SEGMENTS;
2711 clt->reconnect_delay_sec = reconnect_delay_sec;
2712 clt->max_reconnect_attempts = max_reconnect_attempts;
2713 clt->priv = priv;
2714 clt->link_ev = link_ev;
2715 clt->mp_policy = MP_POLICY_MIN_INFLIGHT;
2716 strscpy(clt->sessname, sessname, sizeof(clt->sessname));
2717 init_waitqueue_head(&clt->permits_wait);
2718 mutex_init(&clt->paths_ev_mutex);
2719 mutex_init(&clt->paths_mutex);
2720 device_initialize(&clt->dev);
2721
2722 err = dev_set_name(&clt->dev, "%s", sessname);
2723 if (err)
2724 goto err_put;
2725
2726 /*
2727 * Suppress user space notification until
2728 * sysfs files are created
2729 */
2730 dev_set_uevent_suppress(&clt->dev, true);
2731 err = device_add(&clt->dev);
2732 if (err)
2733 goto err_put;
2734
2735 clt->kobj_paths = kobject_create_and_add("paths", &clt->dev.kobj);
2736 if (!clt->kobj_paths) {
2737 err = -ENOMEM;
2738 goto err_del;
2739 }
2740 err = rtrs_clt_create_sysfs_root_files(clt);
2741 if (err) {
2742 kobject_del(clt->kobj_paths);
2743 kobject_put(clt->kobj_paths);
2744 goto err_del;
2745 }
2746 dev_set_uevent_suppress(&clt->dev, false);
2747 kobject_uevent(&clt->dev.kobj, KOBJ_ADD);
2748
2749 return clt;
2750 err_del:
2751 device_del(&clt->dev);
2752 err_put:
2753 free_percpu(clt->pcpu_path);
2754 put_device(&clt->dev);
2755 return ERR_PTR(err);
2756 }
2757
free_clt(struct rtrs_clt * clt)2758 static void free_clt(struct rtrs_clt *clt)
2759 {
2760 free_percpu(clt->pcpu_path);
2761
2762 /*
2763 * release callback will free clt and destroy mutexes in last put
2764 */
2765 device_unregister(&clt->dev);
2766 }
2767
2768 /**
2769 * rtrs_clt_open() - Open a path to an RTRS server
2770 * @ops: holds the link event callback and the private pointer.
2771 * @sessname: name of the session
2772 * @paths: Paths to be established defined by their src and dst addresses
2773 * @paths_num: Number of elements in the @paths array
2774 * @port: port to be used by the RTRS session
2775 * @pdu_sz: Size of extra payload which can be accessed after permit allocation.
2776 * @reconnect_delay_sec: time between reconnect tries
2777 * @max_reconnect_attempts: Number of times to reconnect on error before giving
2778 * up, 0 for * disabled, -1 for forever
2779 * @nr_poll_queues: number of polling mode connection using IB_POLL_DIRECT flag
2780 *
2781 * Starts session establishment with the rtrs_server. The function can block
2782 * up to ~2000ms before it returns.
2783 *
2784 * Return a valid pointer on success otherwise PTR_ERR.
2785 */
rtrs_clt_open(struct rtrs_clt_ops * ops,const char * pathname,const struct rtrs_addr * paths,size_t paths_num,u16 port,size_t pdu_sz,u8 reconnect_delay_sec,s16 max_reconnect_attempts,u32 nr_poll_queues)2786 struct rtrs_clt *rtrs_clt_open(struct rtrs_clt_ops *ops,
2787 const char *pathname,
2788 const struct rtrs_addr *paths,
2789 size_t paths_num, u16 port,
2790 size_t pdu_sz, u8 reconnect_delay_sec,
2791 s16 max_reconnect_attempts, u32 nr_poll_queues)
2792 {
2793 struct rtrs_clt_path *clt_path, *tmp;
2794 struct rtrs_clt *clt;
2795 int err, i;
2796
2797 if (strchr(pathname, '/') || strchr(pathname, '.')) {
2798 pr_err("pathname cannot contain / and .\n");
2799 err = -EINVAL;
2800 goto out;
2801 }
2802
2803 clt = alloc_clt(pathname, paths_num, port, pdu_sz, ops->priv,
2804 ops->link_ev,
2805 reconnect_delay_sec,
2806 max_reconnect_attempts);
2807 if (IS_ERR(clt)) {
2808 err = PTR_ERR(clt);
2809 goto out;
2810 }
2811 for (i = 0; i < paths_num; i++) {
2812 struct rtrs_clt_path *clt_path;
2813
2814 clt_path = alloc_path(clt, &paths[i], nr_cpu_ids,
2815 nr_poll_queues);
2816 if (IS_ERR(clt_path)) {
2817 err = PTR_ERR(clt_path);
2818 goto close_all_path;
2819 }
2820 if (!i)
2821 clt_path->for_new_clt = 1;
2822 list_add_tail_rcu(&clt_path->s.entry, &clt->paths_list);
2823
2824 err = init_path(clt_path);
2825 if (err) {
2826 list_del_rcu(&clt_path->s.entry);
2827 rtrs_clt_close_conns(clt_path, true);
2828 free_percpu(clt_path->stats->pcpu_stats);
2829 kfree(clt_path->stats);
2830 free_path(clt_path);
2831 goto close_all_path;
2832 }
2833
2834 err = rtrs_clt_create_path_files(clt_path);
2835 if (err) {
2836 list_del_rcu(&clt_path->s.entry);
2837 rtrs_clt_close_conns(clt_path, true);
2838 free_percpu(clt_path->stats->pcpu_stats);
2839 kfree(clt_path->stats);
2840 free_path(clt_path);
2841 goto close_all_path;
2842 }
2843 }
2844 err = alloc_permits(clt);
2845 if (err)
2846 goto close_all_path;
2847
2848 return clt;
2849
2850 close_all_path:
2851 list_for_each_entry_safe(clt_path, tmp, &clt->paths_list, s.entry) {
2852 rtrs_clt_destroy_path_files(clt_path, NULL);
2853 rtrs_clt_close_conns(clt_path, true);
2854 kobject_put(&clt_path->kobj);
2855 }
2856 rtrs_clt_destroy_sysfs_root(clt);
2857 free_clt(clt);
2858
2859 out:
2860 return ERR_PTR(err);
2861 }
2862 EXPORT_SYMBOL(rtrs_clt_open);
2863
2864 /**
2865 * rtrs_clt_close() - Close a path
2866 * @clt: Session handle. Session is freed upon return.
2867 */
rtrs_clt_close(struct rtrs_clt * clt)2868 void rtrs_clt_close(struct rtrs_clt *clt)
2869 {
2870 struct rtrs_clt_path *clt_path, *tmp;
2871
2872 /* Firstly forbid sysfs access */
2873 rtrs_clt_destroy_sysfs_root(clt);
2874
2875 /* Now it is safe to iterate over all paths without locks */
2876 list_for_each_entry_safe(clt_path, tmp, &clt->paths_list, s.entry) {
2877 rtrs_clt_close_conns(clt_path, true);
2878 rtrs_clt_destroy_path_files(clt_path, NULL);
2879 kobject_put(&clt_path->kobj);
2880 }
2881 free_permits(clt);
2882 free_clt(clt);
2883 }
2884 EXPORT_SYMBOL(rtrs_clt_close);
2885
rtrs_clt_reconnect_from_sysfs(struct rtrs_clt_path * clt_path)2886 int rtrs_clt_reconnect_from_sysfs(struct rtrs_clt_path *clt_path)
2887 {
2888 enum rtrs_clt_state old_state;
2889 int err = -EBUSY;
2890 bool changed;
2891
2892 changed = rtrs_clt_change_state_get_old(clt_path,
2893 RTRS_CLT_RECONNECTING,
2894 &old_state);
2895 if (changed) {
2896 clt_path->reconnect_attempts = 0;
2897 queue_delayed_work(rtrs_wq, &clt_path->reconnect_dwork, 0);
2898 }
2899 if (changed || old_state == RTRS_CLT_RECONNECTING) {
2900 /*
2901 * flush_delayed_work() queues pending work for immediate
2902 * execution, so do the flush if we have queued something
2903 * right now or work is pending.
2904 */
2905 flush_delayed_work(&clt_path->reconnect_dwork);
2906 err = (READ_ONCE(clt_path->state) ==
2907 RTRS_CLT_CONNECTED ? 0 : -ENOTCONN);
2908 }
2909
2910 return err;
2911 }
2912
rtrs_clt_remove_path_from_sysfs(struct rtrs_clt_path * clt_path,const struct attribute * sysfs_self)2913 int rtrs_clt_remove_path_from_sysfs(struct rtrs_clt_path *clt_path,
2914 const struct attribute *sysfs_self)
2915 {
2916 enum rtrs_clt_state old_state;
2917 bool changed;
2918
2919 /*
2920 * Continue stopping path till state was changed to DEAD or
2921 * state was observed as DEAD:
2922 * 1. State was changed to DEAD - we were fast and nobody
2923 * invoked rtrs_clt_reconnect(), which can again start
2924 * reconnecting.
2925 * 2. State was observed as DEAD - we have someone in parallel
2926 * removing the path.
2927 */
2928 do {
2929 rtrs_clt_close_conns(clt_path, true);
2930 changed = rtrs_clt_change_state_get_old(clt_path,
2931 RTRS_CLT_DEAD,
2932 &old_state);
2933 } while (!changed && old_state != RTRS_CLT_DEAD);
2934
2935 if (changed) {
2936 rtrs_clt_remove_path_from_arr(clt_path);
2937 rtrs_clt_destroy_path_files(clt_path, sysfs_self);
2938 kobject_put(&clt_path->kobj);
2939 }
2940
2941 return 0;
2942 }
2943
rtrs_clt_set_max_reconnect_attempts(struct rtrs_clt * clt,int value)2944 void rtrs_clt_set_max_reconnect_attempts(struct rtrs_clt *clt, int value)
2945 {
2946 clt->max_reconnect_attempts = (unsigned int)value;
2947 }
2948
rtrs_clt_get_max_reconnect_attempts(const struct rtrs_clt * clt)2949 int rtrs_clt_get_max_reconnect_attempts(const struct rtrs_clt *clt)
2950 {
2951 return (int)clt->max_reconnect_attempts;
2952 }
2953
2954 /**
2955 * rtrs_clt_request() - Request data transfer to/from server via RDMA.
2956 *
2957 * @dir: READ/WRITE
2958 * @ops: callback function to be called as confirmation, and the pointer.
2959 * @clt: Session
2960 * @permit: Preallocated permit
2961 * @vec: Message that is sent to server together with the request.
2962 * Sum of len of all @vec elements limited to <= IO_MSG_SIZE.
2963 * Since the msg is copied internally it can be allocated on stack.
2964 * @nr: Number of elements in @vec.
2965 * @data_len: length of data sent to/from server
2966 * @sg: Pages to be sent/received to/from server.
2967 * @sg_cnt: Number of elements in the @sg
2968 *
2969 * Return:
2970 * 0: Success
2971 * <0: Error
2972 *
2973 * On dir=READ rtrs client will request a data transfer from Server to client.
2974 * The data that the server will respond with will be stored in @sg when
2975 * the user receives an %RTRS_CLT_RDMA_EV_RDMA_REQUEST_WRITE_COMPL event.
2976 * On dir=WRITE rtrs client will rdma write data in sg to server side.
2977 */
rtrs_clt_request(int dir,struct rtrs_clt_req_ops * ops,struct rtrs_clt * clt,struct rtrs_permit * permit,const struct kvec * vec,size_t nr,size_t data_len,struct scatterlist * sg,unsigned int sg_cnt)2978 int rtrs_clt_request(int dir, struct rtrs_clt_req_ops *ops,
2979 struct rtrs_clt *clt, struct rtrs_permit *permit,
2980 const struct kvec *vec, size_t nr, size_t data_len,
2981 struct scatterlist *sg, unsigned int sg_cnt)
2982 {
2983 struct rtrs_clt_io_req *req;
2984 struct rtrs_clt_path *clt_path;
2985
2986 enum dma_data_direction dma_dir;
2987 int err = -ECONNABORTED, i;
2988 size_t usr_len, hdr_len;
2989 struct path_it it;
2990
2991 /* Get kvec length */
2992 for (i = 0, usr_len = 0; i < nr; i++)
2993 usr_len += vec[i].iov_len;
2994
2995 if (dir == READ) {
2996 hdr_len = sizeof(struct rtrs_msg_rdma_read) +
2997 sg_cnt * sizeof(struct rtrs_sg_desc);
2998 dma_dir = DMA_FROM_DEVICE;
2999 } else {
3000 hdr_len = sizeof(struct rtrs_msg_rdma_write);
3001 dma_dir = DMA_TO_DEVICE;
3002 }
3003
3004 rcu_read_lock();
3005 for (path_it_init(&it, clt);
3006 (clt_path = it.next_path(&it)) && it.i < it.clt->paths_num; it.i++) {
3007 if (READ_ONCE(clt_path->state) != RTRS_CLT_CONNECTED)
3008 continue;
3009
3010 if (usr_len + hdr_len > clt_path->max_hdr_size) {
3011 rtrs_wrn_rl(clt_path->clt,
3012 "%s request failed, user message size is %zu and header length %zu, but max size is %u\n",
3013 dir == READ ? "Read" : "Write",
3014 usr_len, hdr_len, clt_path->max_hdr_size);
3015 err = -EMSGSIZE;
3016 break;
3017 }
3018 req = rtrs_clt_get_req(clt_path, ops->conf_fn, permit, ops->priv,
3019 vec, usr_len, sg, sg_cnt, data_len,
3020 dma_dir);
3021 if (dir == READ)
3022 err = rtrs_clt_read_req(req);
3023 else
3024 err = rtrs_clt_write_req(req);
3025 if (err) {
3026 req->in_use = false;
3027 continue;
3028 }
3029 /* Success path */
3030 break;
3031 }
3032 path_it_deinit(&it);
3033 rcu_read_unlock();
3034
3035 return err;
3036 }
3037 EXPORT_SYMBOL(rtrs_clt_request);
3038
rtrs_clt_rdma_cq_direct(struct rtrs_clt * clt,unsigned int index)3039 int rtrs_clt_rdma_cq_direct(struct rtrs_clt *clt, unsigned int index)
3040 {
3041 /* If no path, return -1 for block layer not to try again */
3042 int cnt = -1;
3043 struct rtrs_con *con;
3044 struct rtrs_clt_path *clt_path;
3045 struct path_it it;
3046
3047 rcu_read_lock();
3048 for (path_it_init(&it, clt);
3049 (clt_path = it.next_path(&it)) && it.i < it.clt->paths_num; it.i++) {
3050 if (READ_ONCE(clt_path->state) != RTRS_CLT_CONNECTED)
3051 continue;
3052
3053 con = clt_path->s.con[index + 1];
3054 cnt = ib_process_cq_direct(con->cq, -1);
3055 if (cnt)
3056 break;
3057 }
3058 path_it_deinit(&it);
3059 rcu_read_unlock();
3060
3061 return cnt;
3062 }
3063 EXPORT_SYMBOL(rtrs_clt_rdma_cq_direct);
3064
3065 /**
3066 * rtrs_clt_query() - queries RTRS session attributes
3067 *@clt: session pointer
3068 *@attr: query results for session attributes.
3069 * Returns:
3070 * 0 on success
3071 * -ECOMM no connection to the server
3072 */
rtrs_clt_query(struct rtrs_clt * clt,struct rtrs_attrs * attr)3073 int rtrs_clt_query(struct rtrs_clt *clt, struct rtrs_attrs *attr)
3074 {
3075 if (!rtrs_clt_is_connected(clt))
3076 return -ECOMM;
3077
3078 attr->queue_depth = clt->queue_depth;
3079 attr->max_segments = clt->max_segments;
3080 /* Cap max_io_size to min of remote buffer size and the fr pages */
3081 attr->max_io_size = min_t(int, clt->max_io_size,
3082 clt->max_segments * SZ_4K);
3083
3084 return 0;
3085 }
3086 EXPORT_SYMBOL(rtrs_clt_query);
3087
rtrs_clt_create_path_from_sysfs(struct rtrs_clt * clt,struct rtrs_addr * addr)3088 int rtrs_clt_create_path_from_sysfs(struct rtrs_clt *clt,
3089 struct rtrs_addr *addr)
3090 {
3091 struct rtrs_clt_path *clt_path;
3092 int err;
3093
3094 clt_path = alloc_path(clt, addr, nr_cpu_ids, 0);
3095 if (IS_ERR(clt_path))
3096 return PTR_ERR(clt_path);
3097
3098 mutex_lock(&clt->paths_mutex);
3099 if (clt->paths_num == 0) {
3100 /*
3101 * When all the paths are removed for a session,
3102 * the addition of the first path is like a new session for
3103 * the storage server
3104 */
3105 clt_path->for_new_clt = 1;
3106 }
3107
3108 mutex_unlock(&clt->paths_mutex);
3109
3110 /*
3111 * It is totally safe to add path in CONNECTING state: coming
3112 * IO will never grab it. Also it is very important to add
3113 * path before init, since init fires LINK_CONNECTED event.
3114 */
3115 rtrs_clt_add_path_to_arr(clt_path);
3116
3117 err = init_path(clt_path);
3118 if (err)
3119 goto close_path;
3120
3121 err = rtrs_clt_create_path_files(clt_path);
3122 if (err)
3123 goto close_path;
3124
3125 return 0;
3126
3127 close_path:
3128 rtrs_clt_remove_path_from_arr(clt_path);
3129 rtrs_clt_close_conns(clt_path, true);
3130 free_percpu(clt_path->stats->pcpu_stats);
3131 kfree(clt_path->stats);
3132 free_path(clt_path);
3133
3134 return err;
3135 }
3136
rtrs_clt_ib_dev_init(struct rtrs_ib_dev * dev)3137 static int rtrs_clt_ib_dev_init(struct rtrs_ib_dev *dev)
3138 {
3139 if (!(dev->ib_dev->attrs.device_cap_flags &
3140 IB_DEVICE_MEM_MGT_EXTENSIONS)) {
3141 pr_err("Memory registrations not supported.\n");
3142 return -ENOTSUPP;
3143 }
3144
3145 return 0;
3146 }
3147
3148 static const struct rtrs_rdma_dev_pd_ops dev_pd_ops = {
3149 .init = rtrs_clt_ib_dev_init
3150 };
3151
rtrs_client_init(void)3152 static int __init rtrs_client_init(void)
3153 {
3154 rtrs_rdma_dev_pd_init(0, &dev_pd);
3155
3156 rtrs_clt_dev_class = class_create(THIS_MODULE, "rtrs-client");
3157 if (IS_ERR(rtrs_clt_dev_class)) {
3158 pr_err("Failed to create rtrs-client dev class\n");
3159 return PTR_ERR(rtrs_clt_dev_class);
3160 }
3161 rtrs_wq = alloc_workqueue("rtrs_client_wq", 0, 0);
3162 if (!rtrs_wq) {
3163 class_destroy(rtrs_clt_dev_class);
3164 return -ENOMEM;
3165 }
3166
3167 return 0;
3168 }
3169
rtrs_client_exit(void)3170 static void __exit rtrs_client_exit(void)
3171 {
3172 destroy_workqueue(rtrs_wq);
3173 class_destroy(rtrs_clt_dev_class);
3174 rtrs_rdma_dev_pd_deinit(&dev_pd);
3175 }
3176
3177 module_init(rtrs_client_init);
3178 module_exit(rtrs_client_exit);
3179