Loading...
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#include "rtrs-clt-trace.h"
20
21#define RTRS_CONNECT_TIMEOUT_MS 30000
22/*
23 * Wait a bit before trying to reconnect after a failure
24 * in order to give server time to finish clean up which
25 * leads to "false positives" failed reconnect attempts
26 */
27#define RTRS_RECONNECT_BACKOFF 1000
28/*
29 * Wait for additional random time between 0 and 8 seconds
30 * before starting to reconnect to avoid clients reconnecting
31 * all at once in case of a major network outage
32 */
33#define RTRS_RECONNECT_SEED 8
34
35#define FIRST_CONN 0x01
36/* limit to 128 * 4k = 512k max IO */
37#define RTRS_MAX_SEGMENTS 128
38
39MODULE_DESCRIPTION("RDMA Transport Client");
40MODULE_LICENSE("GPL");
41
42static const struct rtrs_rdma_dev_pd_ops dev_pd_ops;
43static struct rtrs_rdma_dev_pd dev_pd = {
44 .ops = &dev_pd_ops
45};
46
47static struct workqueue_struct *rtrs_wq;
48static const struct class rtrs_clt_dev_class = {
49 .name = "rtrs-client",
50};
51
52static inline bool rtrs_clt_is_connected(const struct rtrs_clt_sess *clt)
53{
54 struct rtrs_clt_path *clt_path;
55 bool connected = false;
56
57 rcu_read_lock();
58 list_for_each_entry_rcu(clt_path, &clt->paths_list, s.entry)
59 if (READ_ONCE(clt_path->state) == RTRS_CLT_CONNECTED) {
60 connected = true;
61 break;
62 }
63 rcu_read_unlock();
64
65 return connected;
66}
67
68static struct rtrs_permit *
69__rtrs_get_permit(struct rtrs_clt_sess *clt, enum rtrs_clt_con_type con_type)
70{
71 size_t max_depth = clt->queue_depth;
72 struct rtrs_permit *permit;
73 int bit;
74
75 /*
76 * Adapted from null_blk get_tag(). Callers from different cpus may
77 * grab the same bit, since find_first_zero_bit is not atomic.
78 * But then the test_and_set_bit_lock will fail for all the
79 * callers but one, so that they will loop again.
80 * This way an explicit spinlock is not required.
81 */
82 do {
83 bit = find_first_zero_bit(clt->permits_map, max_depth);
84 if (bit >= max_depth)
85 return NULL;
86 } while (test_and_set_bit_lock(bit, clt->permits_map));
87
88 permit = get_permit(clt, bit);
89 WARN_ON(permit->mem_id != bit);
90 permit->cpu_id = raw_smp_processor_id();
91 permit->con_type = con_type;
92
93 return permit;
94}
95
96static inline void __rtrs_put_permit(struct rtrs_clt_sess *clt,
97 struct rtrs_permit *permit)
98{
99 clear_bit_unlock(permit->mem_id, clt->permits_map);
100}
101
102/**
103 * rtrs_clt_get_permit() - allocates permit for future RDMA operation
104 * @clt: Current session
105 * @con_type: Type of connection to use with the permit
106 * @can_wait: Wait type
107 *
108 * Description:
109 * Allocates permit for the following RDMA operation. Permit is used
110 * to preallocate all resources and to propagate memory pressure
111 * up earlier.
112 *
113 * Context:
114 * Can sleep if @wait == RTRS_PERMIT_WAIT
115 */
116struct rtrs_permit *rtrs_clt_get_permit(struct rtrs_clt_sess *clt,
117 enum rtrs_clt_con_type con_type,
118 enum wait_type can_wait)
119{
120 struct rtrs_permit *permit;
121 DEFINE_WAIT(wait);
122
123 permit = __rtrs_get_permit(clt, con_type);
124 if (permit || !can_wait)
125 return permit;
126
127 do {
128 prepare_to_wait(&clt->permits_wait, &wait,
129 TASK_UNINTERRUPTIBLE);
130 permit = __rtrs_get_permit(clt, con_type);
131 if (permit)
132 break;
133
134 io_schedule();
135 } while (1);
136
137 finish_wait(&clt->permits_wait, &wait);
138
139 return permit;
140}
141EXPORT_SYMBOL(rtrs_clt_get_permit);
142
143/**
144 * rtrs_clt_put_permit() - puts allocated permit
145 * @clt: Current session
146 * @permit: Permit to be freed
147 *
148 * Context:
149 * Does not matter
150 */
151void rtrs_clt_put_permit(struct rtrs_clt_sess *clt,
152 struct rtrs_permit *permit)
153{
154 if (WARN_ON(!test_bit(permit->mem_id, clt->permits_map)))
155 return;
156
157 __rtrs_put_permit(clt, permit);
158
159 /*
160 * rtrs_clt_get_permit() adds itself to the &clt->permits_wait list
161 * before calling schedule(). So if rtrs_clt_get_permit() is sleeping
162 * it must have added itself to &clt->permits_wait before
163 * __rtrs_put_permit() finished.
164 * Hence it is safe to guard wake_up() with a waitqueue_active() test.
165 */
166 if (waitqueue_active(&clt->permits_wait))
167 wake_up(&clt->permits_wait);
168}
169EXPORT_SYMBOL(rtrs_clt_put_permit);
170
171/**
172 * rtrs_permit_to_clt_con() - returns RDMA connection pointer by the permit
173 * @clt_path: client path pointer
174 * @permit: permit for the allocation of the RDMA buffer
175 * Note:
176 * IO connection starts from 1.
177 * 0 connection is for user messages.
178 */
179static
180struct rtrs_clt_con *rtrs_permit_to_clt_con(struct rtrs_clt_path *clt_path,
181 struct rtrs_permit *permit)
182{
183 int id = 0;
184
185 if (permit->con_type == RTRS_IO_CON)
186 id = (permit->cpu_id % (clt_path->s.irq_con_num - 1)) + 1;
187
188 return to_clt_con(clt_path->s.con[id]);
189}
190
191/**
192 * rtrs_clt_change_state() - change the session state through session state
193 * machine.
194 *
195 * @clt_path: client path to change the state of.
196 * @new_state: state to change to.
197 *
198 * returns true if sess's state is changed to new state, otherwise return false.
199 *
200 * Locks:
201 * state_wq lock must be hold.
202 */
203static bool rtrs_clt_change_state(struct rtrs_clt_path *clt_path,
204 enum rtrs_clt_state new_state)
205{
206 enum rtrs_clt_state old_state;
207 bool changed = false;
208
209 lockdep_assert_held(&clt_path->state_wq.lock);
210
211 old_state = clt_path->state;
212 switch (new_state) {
213 case RTRS_CLT_CONNECTING:
214 switch (old_state) {
215 case RTRS_CLT_RECONNECTING:
216 changed = true;
217 fallthrough;
218 default:
219 break;
220 }
221 break;
222 case RTRS_CLT_RECONNECTING:
223 switch (old_state) {
224 case RTRS_CLT_CONNECTED:
225 case RTRS_CLT_CONNECTING_ERR:
226 case RTRS_CLT_CLOSED:
227 changed = true;
228 fallthrough;
229 default:
230 break;
231 }
232 break;
233 case RTRS_CLT_CONNECTED:
234 switch (old_state) {
235 case RTRS_CLT_CONNECTING:
236 changed = true;
237 fallthrough;
238 default:
239 break;
240 }
241 break;
242 case RTRS_CLT_CONNECTING_ERR:
243 switch (old_state) {
244 case RTRS_CLT_CONNECTING:
245 changed = true;
246 fallthrough;
247 default:
248 break;
249 }
250 break;
251 case RTRS_CLT_CLOSING:
252 switch (old_state) {
253 case RTRS_CLT_CONNECTING:
254 case RTRS_CLT_CONNECTING_ERR:
255 case RTRS_CLT_RECONNECTING:
256 case RTRS_CLT_CONNECTED:
257 changed = true;
258 fallthrough;
259 default:
260 break;
261 }
262 break;
263 case RTRS_CLT_CLOSED:
264 switch (old_state) {
265 case RTRS_CLT_CLOSING:
266 changed = true;
267 fallthrough;
268 default:
269 break;
270 }
271 break;
272 case RTRS_CLT_DEAD:
273 switch (old_state) {
274 case RTRS_CLT_CLOSED:
275 changed = true;
276 fallthrough;
277 default:
278 break;
279 }
280 break;
281 default:
282 break;
283 }
284 if (changed) {
285 clt_path->state = new_state;
286 wake_up_locked(&clt_path->state_wq);
287 }
288
289 return changed;
290}
291
292static bool rtrs_clt_change_state_from_to(struct rtrs_clt_path *clt_path,
293 enum rtrs_clt_state old_state,
294 enum rtrs_clt_state new_state)
295{
296 bool changed = false;
297
298 spin_lock_irq(&clt_path->state_wq.lock);
299 if (clt_path->state == old_state)
300 changed = rtrs_clt_change_state(clt_path, new_state);
301 spin_unlock_irq(&clt_path->state_wq.lock);
302
303 return changed;
304}
305
306static void rtrs_clt_stop_and_destroy_conns(struct rtrs_clt_path *clt_path);
307static void rtrs_rdma_error_recovery(struct rtrs_clt_con *con)
308{
309 struct rtrs_clt_path *clt_path = to_clt_path(con->c.path);
310
311 trace_rtrs_rdma_error_recovery(clt_path);
312
313 if (rtrs_clt_change_state_from_to(clt_path,
314 RTRS_CLT_CONNECTED,
315 RTRS_CLT_RECONNECTING)) {
316 queue_work(rtrs_wq, &clt_path->err_recovery_work);
317 } else {
318 /*
319 * Error can happen just on establishing new connection,
320 * so notify waiter with error state, waiter is responsible
321 * for cleaning the rest and reconnect if needed.
322 */
323 rtrs_clt_change_state_from_to(clt_path,
324 RTRS_CLT_CONNECTING,
325 RTRS_CLT_CONNECTING_ERR);
326 }
327}
328
329static void rtrs_clt_fast_reg_done(struct ib_cq *cq, struct ib_wc *wc)
330{
331 struct rtrs_clt_con *con = to_clt_con(wc->qp->qp_context);
332
333 if (wc->status != IB_WC_SUCCESS) {
334 rtrs_err(con->c.path, "Failed IB_WR_REG_MR: %s\n",
335 ib_wc_status_msg(wc->status));
336 rtrs_rdma_error_recovery(con);
337 }
338}
339
340static struct ib_cqe fast_reg_cqe = {
341 .done = rtrs_clt_fast_reg_done
342};
343
344static void complete_rdma_req(struct rtrs_clt_io_req *req, int errno,
345 bool notify, bool can_wait);
346
347static void rtrs_clt_inv_rkey_done(struct ib_cq *cq, struct ib_wc *wc)
348{
349 struct rtrs_clt_io_req *req =
350 container_of(wc->wr_cqe, typeof(*req), inv_cqe);
351 struct rtrs_clt_con *con = to_clt_con(wc->qp->qp_context);
352
353 if (wc->status != IB_WC_SUCCESS) {
354 rtrs_err(con->c.path, "Failed IB_WR_LOCAL_INV: %s\n",
355 ib_wc_status_msg(wc->status));
356 rtrs_rdma_error_recovery(con);
357 }
358 req->need_inv = false;
359 if (req->need_inv_comp)
360 complete(&req->inv_comp);
361 else
362 /* Complete request from INV callback */
363 complete_rdma_req(req, req->inv_errno, true, false);
364}
365
366static int rtrs_inv_rkey(struct rtrs_clt_io_req *req)
367{
368 struct rtrs_clt_con *con = req->con;
369 struct ib_send_wr wr = {
370 .opcode = IB_WR_LOCAL_INV,
371 .wr_cqe = &req->inv_cqe,
372 .send_flags = IB_SEND_SIGNALED,
373 .ex.invalidate_rkey = req->mr->rkey,
374 };
375 req->inv_cqe.done = rtrs_clt_inv_rkey_done;
376
377 return ib_post_send(con->c.qp, &wr, NULL);
378}
379
380static void complete_rdma_req(struct rtrs_clt_io_req *req, int errno,
381 bool notify, bool can_wait)
382{
383 struct rtrs_clt_con *con = req->con;
384 struct rtrs_clt_path *clt_path;
385 int err;
386
387 if (!req->in_use)
388 return;
389 if (WARN_ON(!req->con))
390 return;
391 clt_path = to_clt_path(con->c.path);
392
393 if (req->sg_cnt) {
394 if (req->dir == DMA_FROM_DEVICE && req->need_inv) {
395 /*
396 * We are here to invalidate read requests
397 * ourselves. In normal scenario server should
398 * send INV for all read requests, but
399 * we are here, thus two things could happen:
400 *
401 * 1. this is failover, when errno != 0
402 * and can_wait == 1,
403 *
404 * 2. something totally bad happened and
405 * server forgot to send INV, so we
406 * should do that ourselves.
407 */
408
409 if (can_wait) {
410 req->need_inv_comp = true;
411 } else {
412 /* This should be IO path, so always notify */
413 WARN_ON(!notify);
414 /* Save errno for INV callback */
415 req->inv_errno = errno;
416 }
417
418 refcount_inc(&req->ref);
419 err = rtrs_inv_rkey(req);
420 if (err) {
421 rtrs_err(con->c.path, "Send INV WR key=%#x: %d\n",
422 req->mr->rkey, err);
423 } else if (can_wait) {
424 wait_for_completion(&req->inv_comp);
425 } else {
426 /*
427 * Something went wrong, so request will be
428 * completed from INV callback.
429 */
430 WARN_ON_ONCE(1);
431
432 return;
433 }
434 if (!refcount_dec_and_test(&req->ref))
435 return;
436 }
437 ib_dma_unmap_sg(clt_path->s.dev->ib_dev, req->sglist,
438 req->sg_cnt, req->dir);
439 }
440 if (!refcount_dec_and_test(&req->ref))
441 return;
442 if (req->mp_policy == MP_POLICY_MIN_INFLIGHT)
443 atomic_dec(&clt_path->stats->inflight);
444
445 req->in_use = false;
446 req->con = NULL;
447
448 if (errno) {
449 rtrs_err_rl(con->c.path, "IO request failed: error=%d path=%s [%s:%u] notify=%d\n",
450 errno, kobject_name(&clt_path->kobj), clt_path->hca_name,
451 clt_path->hca_port, notify);
452 }
453
454 if (notify)
455 req->conf(req->priv, errno);
456}
457
458static int rtrs_post_send_rdma(struct rtrs_clt_con *con,
459 struct rtrs_clt_io_req *req,
460 struct rtrs_rbuf *rbuf, u32 off,
461 u32 imm, struct ib_send_wr *wr)
462{
463 struct rtrs_clt_path *clt_path = to_clt_path(con->c.path);
464 enum ib_send_flags flags;
465 struct ib_sge sge;
466
467 if (!req->sg_size) {
468 rtrs_wrn(con->c.path,
469 "Doing RDMA Write failed, no data supplied\n");
470 return -EINVAL;
471 }
472
473 /* user data and user message in the first list element */
474 sge.addr = req->iu->dma_addr;
475 sge.length = req->sg_size;
476 sge.lkey = clt_path->s.dev->ib_pd->local_dma_lkey;
477
478 /*
479 * From time to time we have to post signalled sends,
480 * or send queue will fill up and only QP reset can help.
481 */
482 flags = atomic_inc_return(&con->c.wr_cnt) % clt_path->s.signal_interval ?
483 0 : IB_SEND_SIGNALED;
484
485 ib_dma_sync_single_for_device(clt_path->s.dev->ib_dev,
486 req->iu->dma_addr,
487 req->sg_size, DMA_TO_DEVICE);
488
489 return rtrs_iu_post_rdma_write_imm(&con->c, req->iu, &sge, 1,
490 rbuf->rkey, rbuf->addr + off,
491 imm, flags, wr, NULL);
492}
493
494static void process_io_rsp(struct rtrs_clt_path *clt_path, u32 msg_id,
495 s16 errno, bool w_inval)
496{
497 struct rtrs_clt_io_req *req;
498
499 if (WARN_ON(msg_id >= clt_path->queue_depth))
500 return;
501
502 req = &clt_path->reqs[msg_id];
503 /* Drop need_inv if server responded with send with invalidation */
504 req->need_inv &= !w_inval;
505 complete_rdma_req(req, errno, true, false);
506}
507
508static void rtrs_clt_recv_done(struct rtrs_clt_con *con, struct ib_wc *wc)
509{
510 struct rtrs_iu *iu;
511 int err;
512 struct rtrs_clt_path *clt_path = to_clt_path(con->c.path);
513
514 WARN_ON((clt_path->flags & RTRS_MSG_NEW_RKEY_F) == 0);
515 iu = container_of(wc->wr_cqe, struct rtrs_iu,
516 cqe);
517 err = rtrs_iu_post_recv(&con->c, iu);
518 if (err) {
519 rtrs_err(con->c.path, "post iu failed %d\n", err);
520 rtrs_rdma_error_recovery(con);
521 }
522}
523
524static void rtrs_clt_rkey_rsp_done(struct rtrs_clt_con *con, struct ib_wc *wc)
525{
526 struct rtrs_clt_path *clt_path = to_clt_path(con->c.path);
527 struct rtrs_msg_rkey_rsp *msg;
528 u32 imm_type, imm_payload;
529 bool w_inval = false;
530 struct rtrs_iu *iu;
531 u32 buf_id;
532 int err;
533
534 WARN_ON((clt_path->flags & RTRS_MSG_NEW_RKEY_F) == 0);
535
536 iu = container_of(wc->wr_cqe, struct rtrs_iu, cqe);
537
538 if (wc->byte_len < sizeof(*msg)) {
539 rtrs_err(con->c.path, "rkey response is malformed: size %d\n",
540 wc->byte_len);
541 goto out;
542 }
543 ib_dma_sync_single_for_cpu(clt_path->s.dev->ib_dev, iu->dma_addr,
544 iu->size, DMA_FROM_DEVICE);
545 msg = iu->buf;
546 if (le16_to_cpu(msg->type) != RTRS_MSG_RKEY_RSP) {
547 rtrs_err(clt_path->clt,
548 "rkey response is malformed: type %d\n",
549 le16_to_cpu(msg->type));
550 goto out;
551 }
552 buf_id = le16_to_cpu(msg->buf_id);
553 if (WARN_ON(buf_id >= clt_path->queue_depth))
554 goto out;
555
556 rtrs_from_imm(be32_to_cpu(wc->ex.imm_data), &imm_type, &imm_payload);
557 if (imm_type == RTRS_IO_RSP_IMM ||
558 imm_type == RTRS_IO_RSP_W_INV_IMM) {
559 u32 msg_id;
560
561 w_inval = (imm_type == RTRS_IO_RSP_W_INV_IMM);
562 rtrs_from_io_rsp_imm(imm_payload, &msg_id, &err);
563
564 if (WARN_ON(buf_id != msg_id))
565 goto out;
566 clt_path->rbufs[buf_id].rkey = le32_to_cpu(msg->rkey);
567 process_io_rsp(clt_path, msg_id, err, w_inval);
568 }
569 ib_dma_sync_single_for_device(clt_path->s.dev->ib_dev, iu->dma_addr,
570 iu->size, DMA_FROM_DEVICE);
571 return rtrs_clt_recv_done(con, wc);
572out:
573 rtrs_rdma_error_recovery(con);
574}
575
576static void rtrs_clt_rdma_done(struct ib_cq *cq, struct ib_wc *wc);
577
578static struct ib_cqe io_comp_cqe = {
579 .done = rtrs_clt_rdma_done
580};
581
582/*
583 * Post x2 empty WRs: first is for this RDMA with IMM,
584 * second is for RECV with INV, which happened earlier.
585 */
586static int rtrs_post_recv_empty_x2(struct rtrs_con *con, struct ib_cqe *cqe)
587{
588 struct ib_recv_wr wr_arr[2], *wr;
589 int i;
590
591 memset(wr_arr, 0, sizeof(wr_arr));
592 for (i = 0; i < ARRAY_SIZE(wr_arr); i++) {
593 wr = &wr_arr[i];
594 wr->wr_cqe = cqe;
595 if (i)
596 /* Chain backwards */
597 wr->next = &wr_arr[i - 1];
598 }
599
600 return ib_post_recv(con->qp, wr, NULL);
601}
602
603static void rtrs_clt_rdma_done(struct ib_cq *cq, struct ib_wc *wc)
604{
605 struct rtrs_clt_con *con = to_clt_con(wc->qp->qp_context);
606 struct rtrs_clt_path *clt_path = to_clt_path(con->c.path);
607 u32 imm_type, imm_payload;
608 bool w_inval = false;
609 int err;
610
611 if (wc->status != IB_WC_SUCCESS) {
612 if (wc->status != IB_WC_WR_FLUSH_ERR) {
613 rtrs_err(clt_path->clt, "RDMA failed: %s\n",
614 ib_wc_status_msg(wc->status));
615 rtrs_rdma_error_recovery(con);
616 }
617 return;
618 }
619 rtrs_clt_update_wc_stats(con);
620
621 switch (wc->opcode) {
622 case IB_WC_RECV_RDMA_WITH_IMM:
623 /*
624 * post_recv() RDMA write completions of IO reqs (read/write)
625 * and hb
626 */
627 if (WARN_ON(wc->wr_cqe->done != rtrs_clt_rdma_done))
628 return;
629 rtrs_from_imm(be32_to_cpu(wc->ex.imm_data),
630 &imm_type, &imm_payload);
631 if (imm_type == RTRS_IO_RSP_IMM ||
632 imm_type == RTRS_IO_RSP_W_INV_IMM) {
633 u32 msg_id;
634
635 w_inval = (imm_type == RTRS_IO_RSP_W_INV_IMM);
636 rtrs_from_io_rsp_imm(imm_payload, &msg_id, &err);
637
638 process_io_rsp(clt_path, msg_id, err, w_inval);
639 } else if (imm_type == RTRS_HB_MSG_IMM) {
640 WARN_ON(con->c.cid);
641 rtrs_send_hb_ack(&clt_path->s);
642 if (clt_path->flags & RTRS_MSG_NEW_RKEY_F)
643 return rtrs_clt_recv_done(con, wc);
644 } else if (imm_type == RTRS_HB_ACK_IMM) {
645 WARN_ON(con->c.cid);
646 clt_path->s.hb_missed_cnt = 0;
647 clt_path->s.hb_cur_latency =
648 ktime_sub(ktime_get(), clt_path->s.hb_last_sent);
649 if (clt_path->flags & RTRS_MSG_NEW_RKEY_F)
650 return rtrs_clt_recv_done(con, wc);
651 } else {
652 rtrs_wrn(con->c.path, "Unknown IMM type %u\n",
653 imm_type);
654 }
655 if (w_inval)
656 /*
657 * Post x2 empty WRs: first is for this RDMA with IMM,
658 * second is for RECV with INV, which happened earlier.
659 */
660 err = rtrs_post_recv_empty_x2(&con->c, &io_comp_cqe);
661 else
662 err = rtrs_post_recv_empty(&con->c, &io_comp_cqe);
663 if (err) {
664 rtrs_err(con->c.path, "rtrs_post_recv_empty(): %d\n",
665 err);
666 rtrs_rdma_error_recovery(con);
667 }
668 break;
669 case IB_WC_RECV:
670 /*
671 * Key invalidations from server side
672 */
673 WARN_ON(!(wc->wc_flags & IB_WC_WITH_INVALIDATE ||
674 wc->wc_flags & IB_WC_WITH_IMM));
675 WARN_ON(wc->wr_cqe->done != rtrs_clt_rdma_done);
676 if (clt_path->flags & RTRS_MSG_NEW_RKEY_F) {
677 if (wc->wc_flags & IB_WC_WITH_INVALIDATE)
678 return rtrs_clt_recv_done(con, wc);
679
680 return rtrs_clt_rkey_rsp_done(con, wc);
681 }
682 break;
683 case IB_WC_RDMA_WRITE:
684 /*
685 * post_send() RDMA write completions of IO reqs (read/write)
686 * and hb.
687 */
688 break;
689
690 default:
691 rtrs_wrn(clt_path->clt, "Unexpected WC type: %d\n", wc->opcode);
692 return;
693 }
694}
695
696static int post_recv_io(struct rtrs_clt_con *con, size_t q_size)
697{
698 int err, i;
699 struct rtrs_clt_path *clt_path = to_clt_path(con->c.path);
700
701 for (i = 0; i < q_size; i++) {
702 if (clt_path->flags & RTRS_MSG_NEW_RKEY_F) {
703 struct rtrs_iu *iu = &con->rsp_ius[i];
704
705 err = rtrs_iu_post_recv(&con->c, iu);
706 } else {
707 err = rtrs_post_recv_empty(&con->c, &io_comp_cqe);
708 }
709 if (err)
710 return err;
711 }
712
713 return 0;
714}
715
716static int post_recv_path(struct rtrs_clt_path *clt_path)
717{
718 size_t q_size = 0;
719 int err, cid;
720
721 for (cid = 0; cid < clt_path->s.con_num; cid++) {
722 if (cid == 0)
723 q_size = SERVICE_CON_QUEUE_DEPTH;
724 else
725 q_size = clt_path->queue_depth;
726
727 /*
728 * x2 for RDMA read responses + FR key invalidations,
729 * RDMA writes do not require any FR registrations.
730 */
731 q_size *= 2;
732
733 err = post_recv_io(to_clt_con(clt_path->s.con[cid]), q_size);
734 if (err) {
735 rtrs_err(clt_path->clt, "post_recv_io(), err: %d\n",
736 err);
737 return err;
738 }
739 }
740
741 return 0;
742}
743
744struct path_it {
745 int i;
746 struct list_head skip_list;
747 struct rtrs_clt_sess *clt;
748 struct rtrs_clt_path *(*next_path)(struct path_it *it);
749};
750
751/*
752 * rtrs_clt_get_next_path_or_null - get clt path from the list or return NULL
753 * @head: the head for the list.
754 * @clt_path: The element to take the next clt_path from.
755 *
756 * Next clt path returned in round-robin fashion, i.e. head will be skipped,
757 * but if list is observed as empty, NULL will be returned.
758 *
759 * This function may safely run concurrently with the _rcu list-mutation
760 * primitives such as list_add_rcu() as long as it's guarded by rcu_read_lock().
761 */
762static inline struct rtrs_clt_path *
763rtrs_clt_get_next_path_or_null(struct list_head *head, struct rtrs_clt_path *clt_path)
764{
765 return list_next_or_null_rcu(head, &clt_path->s.entry, typeof(*clt_path), s.entry) ?:
766 list_next_or_null_rcu(head,
767 READ_ONCE((&clt_path->s.entry)->next),
768 typeof(*clt_path), s.entry);
769}
770
771/**
772 * get_next_path_rr() - Returns path in round-robin fashion.
773 * @it: the path pointer
774 *
775 * Related to @MP_POLICY_RR
776 *
777 * Locks:
778 * rcu_read_lock() must be held.
779 */
780static struct rtrs_clt_path *get_next_path_rr(struct path_it *it)
781{
782 struct rtrs_clt_path __rcu **ppcpu_path;
783 struct rtrs_clt_path *path;
784 struct rtrs_clt_sess *clt;
785
786 /*
787 * Assert that rcu lock must be held
788 */
789 RCU_LOCKDEP_WARN(!rcu_read_lock_held(), "no rcu read lock held");
790
791 clt = it->clt;
792
793 /*
794 * Here we use two RCU objects: @paths_list and @pcpu_path
795 * pointer. See rtrs_clt_remove_path_from_arr() for details
796 * how that is handled.
797 */
798
799 ppcpu_path = this_cpu_ptr(clt->pcpu_path);
800 path = rcu_dereference(*ppcpu_path);
801 if (!path)
802 path = list_first_or_null_rcu(&clt->paths_list,
803 typeof(*path), s.entry);
804 else
805 path = rtrs_clt_get_next_path_or_null(&clt->paths_list, path);
806
807 rcu_assign_pointer(*ppcpu_path, path);
808
809 return path;
810}
811
812/**
813 * get_next_path_min_inflight() - Returns path with minimal inflight count.
814 * @it: the path pointer
815 *
816 * Related to @MP_POLICY_MIN_INFLIGHT
817 *
818 * Locks:
819 * rcu_read_lock() must be hold.
820 */
821static struct rtrs_clt_path *get_next_path_min_inflight(struct path_it *it)
822{
823 struct rtrs_clt_path *min_path = NULL;
824 struct rtrs_clt_sess *clt = it->clt;
825 struct rtrs_clt_path *clt_path;
826 int min_inflight = INT_MAX;
827 int inflight;
828
829 list_for_each_entry_rcu(clt_path, &clt->paths_list, s.entry) {
830 if (READ_ONCE(clt_path->state) != RTRS_CLT_CONNECTED)
831 continue;
832
833 if (!list_empty(raw_cpu_ptr(clt_path->mp_skip_entry)))
834 continue;
835
836 inflight = atomic_read(&clt_path->stats->inflight);
837
838 if (inflight < min_inflight) {
839 min_inflight = inflight;
840 min_path = clt_path;
841 }
842 }
843
844 /*
845 * add the path to the skip list, so that next time we can get
846 * a different one
847 */
848 if (min_path)
849 list_add(raw_cpu_ptr(min_path->mp_skip_entry), &it->skip_list);
850
851 return min_path;
852}
853
854/**
855 * get_next_path_min_latency() - Returns path with minimal latency.
856 * @it: the path pointer
857 *
858 * Return: a path with the lowest latency or NULL if all paths are tried
859 *
860 * Locks:
861 * rcu_read_lock() must be hold.
862 *
863 * Related to @MP_POLICY_MIN_LATENCY
864 *
865 * This DOES skip an already-tried path.
866 * There is a skip-list to skip a path if the path has tried but failed.
867 * It will try the minimum latency path and then the second minimum latency
868 * path and so on. Finally it will return NULL if all paths are tried.
869 * Therefore the caller MUST check the returned
870 * path is NULL and trigger the IO error.
871 */
872static struct rtrs_clt_path *get_next_path_min_latency(struct path_it *it)
873{
874 struct rtrs_clt_path *min_path = NULL;
875 struct rtrs_clt_sess *clt = it->clt;
876 struct rtrs_clt_path *clt_path;
877 ktime_t min_latency = KTIME_MAX;
878 ktime_t latency;
879
880 list_for_each_entry_rcu(clt_path, &clt->paths_list, s.entry) {
881 if (READ_ONCE(clt_path->state) != RTRS_CLT_CONNECTED)
882 continue;
883
884 if (!list_empty(raw_cpu_ptr(clt_path->mp_skip_entry)))
885 continue;
886
887 latency = clt_path->s.hb_cur_latency;
888
889 if (latency < min_latency) {
890 min_latency = latency;
891 min_path = clt_path;
892 }
893 }
894
895 /*
896 * add the path to the skip list, so that next time we can get
897 * a different one
898 */
899 if (min_path)
900 list_add(raw_cpu_ptr(min_path->mp_skip_entry), &it->skip_list);
901
902 return min_path;
903}
904
905static inline void path_it_init(struct path_it *it, struct rtrs_clt_sess *clt)
906{
907 INIT_LIST_HEAD(&it->skip_list);
908 it->clt = clt;
909 it->i = 0;
910
911 if (clt->mp_policy == MP_POLICY_RR)
912 it->next_path = get_next_path_rr;
913 else if (clt->mp_policy == MP_POLICY_MIN_INFLIGHT)
914 it->next_path = get_next_path_min_inflight;
915 else
916 it->next_path = get_next_path_min_latency;
917}
918
919static inline void path_it_deinit(struct path_it *it)
920{
921 struct list_head *skip, *tmp;
922 /*
923 * The skip_list is used only for the MIN_INFLIGHT and MIN_LATENCY policies.
924 * We need to remove paths from it, so that next IO can insert
925 * paths (->mp_skip_entry) into a skip_list again.
926 */
927 list_for_each_safe(skip, tmp, &it->skip_list)
928 list_del_init(skip);
929}
930
931/**
932 * rtrs_clt_init_req() - Initialize an rtrs_clt_io_req holding information
933 * about an inflight IO.
934 * The user buffer holding user control message (not data) is copied into
935 * the corresponding buffer of rtrs_iu (req->iu->buf), which later on will
936 * also hold the control message of rtrs.
937 * @req: an io request holding information about IO.
938 * @clt_path: client path
939 * @conf: conformation callback function to notify upper layer.
940 * @permit: permit for allocation of RDMA remote buffer
941 * @priv: private pointer
942 * @vec: kernel vector containing control message
943 * @usr_len: length of the user message
944 * @sg: scater list for IO data
945 * @sg_cnt: number of scater list entries
946 * @data_len: length of the IO data
947 * @dir: direction of the IO.
948 */
949static void rtrs_clt_init_req(struct rtrs_clt_io_req *req,
950 struct rtrs_clt_path *clt_path,
951 void (*conf)(void *priv, int errno),
952 struct rtrs_permit *permit, void *priv,
953 const struct kvec *vec, size_t usr_len,
954 struct scatterlist *sg, size_t sg_cnt,
955 size_t data_len, int dir)
956{
957 struct iov_iter iter;
958 size_t len;
959
960 req->permit = permit;
961 req->in_use = true;
962 req->usr_len = usr_len;
963 req->data_len = data_len;
964 req->sglist = sg;
965 req->sg_cnt = sg_cnt;
966 req->priv = priv;
967 req->dir = dir;
968 req->con = rtrs_permit_to_clt_con(clt_path, permit);
969 req->conf = conf;
970 req->need_inv = false;
971 req->need_inv_comp = false;
972 req->inv_errno = 0;
973 refcount_set(&req->ref, 1);
974 req->mp_policy = clt_path->clt->mp_policy;
975
976 iov_iter_kvec(&iter, ITER_SOURCE, vec, 1, usr_len);
977 len = _copy_from_iter(req->iu->buf, usr_len, &iter);
978 WARN_ON(len != usr_len);
979
980 reinit_completion(&req->inv_comp);
981}
982
983static struct rtrs_clt_io_req *
984rtrs_clt_get_req(struct rtrs_clt_path *clt_path,
985 void (*conf)(void *priv, int errno),
986 struct rtrs_permit *permit, void *priv,
987 const struct kvec *vec, size_t usr_len,
988 struct scatterlist *sg, size_t sg_cnt,
989 size_t data_len, int dir)
990{
991 struct rtrs_clt_io_req *req;
992
993 req = &clt_path->reqs[permit->mem_id];
994 rtrs_clt_init_req(req, clt_path, conf, permit, priv, vec, usr_len,
995 sg, sg_cnt, data_len, dir);
996 return req;
997}
998
999static struct rtrs_clt_io_req *
1000rtrs_clt_get_copy_req(struct rtrs_clt_path *alive_path,
1001 struct rtrs_clt_io_req *fail_req)
1002{
1003 struct rtrs_clt_io_req *req;
1004 struct kvec vec = {
1005 .iov_base = fail_req->iu->buf,
1006 .iov_len = fail_req->usr_len
1007 };
1008
1009 req = &alive_path->reqs[fail_req->permit->mem_id];
1010 rtrs_clt_init_req(req, alive_path, fail_req->conf, fail_req->permit,
1011 fail_req->priv, &vec, fail_req->usr_len,
1012 fail_req->sglist, fail_req->sg_cnt,
1013 fail_req->data_len, fail_req->dir);
1014 return req;
1015}
1016
1017static int rtrs_post_rdma_write_sg(struct rtrs_clt_con *con,
1018 struct rtrs_clt_io_req *req,
1019 struct rtrs_rbuf *rbuf, bool fr_en,
1020 u32 count, u32 size, u32 imm,
1021 struct ib_send_wr *wr,
1022 struct ib_send_wr *tail)
1023{
1024 struct rtrs_clt_path *clt_path = to_clt_path(con->c.path);
1025 struct ib_sge *sge = req->sge;
1026 enum ib_send_flags flags;
1027 struct scatterlist *sg;
1028 size_t num_sge;
1029 int i;
1030 struct ib_send_wr *ptail = NULL;
1031
1032 if (fr_en) {
1033 i = 0;
1034 sge[i].addr = req->mr->iova;
1035 sge[i].length = req->mr->length;
1036 sge[i].lkey = req->mr->lkey;
1037 i++;
1038 num_sge = 2;
1039 ptail = tail;
1040 } else {
1041 for_each_sg(req->sglist, sg, count, i) {
1042 sge[i].addr = sg_dma_address(sg);
1043 sge[i].length = sg_dma_len(sg);
1044 sge[i].lkey = clt_path->s.dev->ib_pd->local_dma_lkey;
1045 }
1046 num_sge = 1 + count;
1047 }
1048 sge[i].addr = req->iu->dma_addr;
1049 sge[i].length = size;
1050 sge[i].lkey = clt_path->s.dev->ib_pd->local_dma_lkey;
1051
1052 /*
1053 * From time to time we have to post signalled sends,
1054 * or send queue will fill up and only QP reset can help.
1055 */
1056 flags = atomic_inc_return(&con->c.wr_cnt) % clt_path->s.signal_interval ?
1057 0 : IB_SEND_SIGNALED;
1058
1059 ib_dma_sync_single_for_device(clt_path->s.dev->ib_dev,
1060 req->iu->dma_addr,
1061 size, DMA_TO_DEVICE);
1062
1063 return rtrs_iu_post_rdma_write_imm(&con->c, req->iu, sge, num_sge,
1064 rbuf->rkey, rbuf->addr, imm,
1065 flags, wr, ptail);
1066}
1067
1068static int rtrs_map_sg_fr(struct rtrs_clt_io_req *req, size_t count)
1069{
1070 int nr;
1071
1072 /* Align the MR to a 4K page size to match the block virt boundary */
1073 nr = ib_map_mr_sg(req->mr, req->sglist, count, NULL, SZ_4K);
1074 if (nr != count)
1075 return nr < 0 ? nr : -EINVAL;
1076 ib_update_fast_reg_key(req->mr, ib_inc_rkey(req->mr->rkey));
1077
1078 return nr;
1079}
1080
1081static int rtrs_clt_write_req(struct rtrs_clt_io_req *req)
1082{
1083 struct rtrs_clt_con *con = req->con;
1084 struct rtrs_path *s = con->c.path;
1085 struct rtrs_clt_path *clt_path = to_clt_path(s);
1086 struct rtrs_msg_rdma_write *msg;
1087
1088 struct rtrs_rbuf *rbuf;
1089 int ret, count = 0;
1090 u32 imm, buf_id;
1091 struct ib_reg_wr rwr;
1092 struct ib_send_wr inv_wr;
1093 struct ib_send_wr *wr = NULL;
1094 bool fr_en = false;
1095
1096 const size_t tsize = sizeof(*msg) + req->data_len + req->usr_len;
1097
1098 if (tsize > clt_path->chunk_size) {
1099 rtrs_wrn(s, "Write request failed, size too big %zu > %d\n",
1100 tsize, clt_path->chunk_size);
1101 return -EMSGSIZE;
1102 }
1103 if (req->sg_cnt) {
1104 count = ib_dma_map_sg(clt_path->s.dev->ib_dev, req->sglist,
1105 req->sg_cnt, req->dir);
1106 if (!count) {
1107 rtrs_wrn(s, "Write request failed, map failed\n");
1108 return -EINVAL;
1109 }
1110 }
1111 /* put rtrs msg after sg and user message */
1112 msg = req->iu->buf + req->usr_len;
1113 msg->type = cpu_to_le16(RTRS_MSG_WRITE);
1114 msg->usr_len = cpu_to_le16(req->usr_len);
1115
1116 /* rtrs message on server side will be after user data and message */
1117 imm = req->permit->mem_off + req->data_len + req->usr_len;
1118 imm = rtrs_to_io_req_imm(imm);
1119 buf_id = req->permit->mem_id;
1120 req->sg_size = tsize;
1121 rbuf = &clt_path->rbufs[buf_id];
1122
1123 if (count) {
1124 ret = rtrs_map_sg_fr(req, count);
1125 if (ret < 0) {
1126 rtrs_err_rl(s,
1127 "Write request failed, failed to map fast reg. data, err: %d\n",
1128 ret);
1129 ib_dma_unmap_sg(clt_path->s.dev->ib_dev, req->sglist,
1130 req->sg_cnt, req->dir);
1131 return ret;
1132 }
1133 inv_wr = (struct ib_send_wr) {
1134 .opcode = IB_WR_LOCAL_INV,
1135 .wr_cqe = &req->inv_cqe,
1136 .send_flags = IB_SEND_SIGNALED,
1137 .ex.invalidate_rkey = req->mr->rkey,
1138 };
1139 req->inv_cqe.done = rtrs_clt_inv_rkey_done;
1140 rwr = (struct ib_reg_wr) {
1141 .wr.opcode = IB_WR_REG_MR,
1142 .wr.wr_cqe = &fast_reg_cqe,
1143 .mr = req->mr,
1144 .key = req->mr->rkey,
1145 .access = (IB_ACCESS_LOCAL_WRITE),
1146 };
1147 wr = &rwr.wr;
1148 fr_en = true;
1149 refcount_inc(&req->ref);
1150 }
1151 /*
1152 * Update stats now, after request is successfully sent it is not
1153 * safe anymore to touch it.
1154 */
1155 rtrs_clt_update_all_stats(req, WRITE);
1156
1157 ret = rtrs_post_rdma_write_sg(req->con, req, rbuf, fr_en, count,
1158 req->usr_len + sizeof(*msg),
1159 imm, wr, &inv_wr);
1160 if (ret) {
1161 rtrs_err_rl(s,
1162 "Write request failed: error=%d path=%s [%s:%u]\n",
1163 ret, kobject_name(&clt_path->kobj), clt_path->hca_name,
1164 clt_path->hca_port);
1165 if (req->mp_policy == MP_POLICY_MIN_INFLIGHT)
1166 atomic_dec(&clt_path->stats->inflight);
1167 if (req->sg_cnt)
1168 ib_dma_unmap_sg(clt_path->s.dev->ib_dev, req->sglist,
1169 req->sg_cnt, req->dir);
1170 }
1171
1172 return ret;
1173}
1174
1175static int rtrs_clt_read_req(struct rtrs_clt_io_req *req)
1176{
1177 struct rtrs_clt_con *con = req->con;
1178 struct rtrs_path *s = con->c.path;
1179 struct rtrs_clt_path *clt_path = to_clt_path(s);
1180 struct rtrs_msg_rdma_read *msg;
1181 struct rtrs_ib_dev *dev = clt_path->s.dev;
1182
1183 struct ib_reg_wr rwr;
1184 struct ib_send_wr *wr = NULL;
1185
1186 int ret, count = 0;
1187 u32 imm, buf_id;
1188
1189 const size_t tsize = sizeof(*msg) + req->data_len + req->usr_len;
1190
1191 if (tsize > clt_path->chunk_size) {
1192 rtrs_wrn(s,
1193 "Read request failed, message size is %zu, bigger than CHUNK_SIZE %d\n",
1194 tsize, clt_path->chunk_size);
1195 return -EMSGSIZE;
1196 }
1197
1198 if (req->sg_cnt) {
1199 count = ib_dma_map_sg(dev->ib_dev, req->sglist, req->sg_cnt,
1200 req->dir);
1201 if (!count) {
1202 rtrs_wrn(s,
1203 "Read request failed, dma map failed\n");
1204 return -EINVAL;
1205 }
1206 }
1207 /* put our message into req->buf after user message*/
1208 msg = req->iu->buf + req->usr_len;
1209 msg->type = cpu_to_le16(RTRS_MSG_READ);
1210 msg->usr_len = cpu_to_le16(req->usr_len);
1211
1212 if (count) {
1213 ret = rtrs_map_sg_fr(req, count);
1214 if (ret < 0) {
1215 rtrs_err_rl(s,
1216 "Read request failed, failed to map fast reg. data, err: %d\n",
1217 ret);
1218 ib_dma_unmap_sg(dev->ib_dev, req->sglist, req->sg_cnt,
1219 req->dir);
1220 return ret;
1221 }
1222 rwr = (struct ib_reg_wr) {
1223 .wr.opcode = IB_WR_REG_MR,
1224 .wr.wr_cqe = &fast_reg_cqe,
1225 .mr = req->mr,
1226 .key = req->mr->rkey,
1227 .access = (IB_ACCESS_LOCAL_WRITE |
1228 IB_ACCESS_REMOTE_WRITE),
1229 };
1230 wr = &rwr.wr;
1231
1232 msg->sg_cnt = cpu_to_le16(1);
1233 msg->flags = cpu_to_le16(RTRS_MSG_NEED_INVAL_F);
1234
1235 msg->desc[0].addr = cpu_to_le64(req->mr->iova);
1236 msg->desc[0].key = cpu_to_le32(req->mr->rkey);
1237 msg->desc[0].len = cpu_to_le32(req->mr->length);
1238
1239 /* Further invalidation is required */
1240 req->need_inv = !!RTRS_MSG_NEED_INVAL_F;
1241
1242 } else {
1243 msg->sg_cnt = 0;
1244 msg->flags = 0;
1245 }
1246 /*
1247 * rtrs message will be after the space reserved for disk data and
1248 * user message
1249 */
1250 imm = req->permit->mem_off + req->data_len + req->usr_len;
1251 imm = rtrs_to_io_req_imm(imm);
1252 buf_id = req->permit->mem_id;
1253
1254 req->sg_size = sizeof(*msg);
1255 req->sg_size += le16_to_cpu(msg->sg_cnt) * sizeof(struct rtrs_sg_desc);
1256 req->sg_size += req->usr_len;
1257
1258 /*
1259 * Update stats now, after request is successfully sent it is not
1260 * safe anymore to touch it.
1261 */
1262 rtrs_clt_update_all_stats(req, READ);
1263
1264 ret = rtrs_post_send_rdma(req->con, req, &clt_path->rbufs[buf_id],
1265 req->data_len, imm, wr);
1266 if (ret) {
1267 rtrs_err_rl(s,
1268 "Read request failed: error=%d path=%s [%s:%u]\n",
1269 ret, kobject_name(&clt_path->kobj), clt_path->hca_name,
1270 clt_path->hca_port);
1271 if (req->mp_policy == MP_POLICY_MIN_INFLIGHT)
1272 atomic_dec(&clt_path->stats->inflight);
1273 req->need_inv = false;
1274 if (req->sg_cnt)
1275 ib_dma_unmap_sg(dev->ib_dev, req->sglist,
1276 req->sg_cnt, req->dir);
1277 }
1278
1279 return ret;
1280}
1281
1282/**
1283 * rtrs_clt_failover_req() - Try to find an active path for a failed request
1284 * @clt: clt context
1285 * @fail_req: a failed io request.
1286 */
1287static int rtrs_clt_failover_req(struct rtrs_clt_sess *clt,
1288 struct rtrs_clt_io_req *fail_req)
1289{
1290 struct rtrs_clt_path *alive_path;
1291 struct rtrs_clt_io_req *req;
1292 int err = -ECONNABORTED;
1293 struct path_it it;
1294
1295 rcu_read_lock();
1296 for (path_it_init(&it, clt);
1297 (alive_path = it.next_path(&it)) && it.i < it.clt->paths_num;
1298 it.i++) {
1299 if (READ_ONCE(alive_path->state) != RTRS_CLT_CONNECTED)
1300 continue;
1301 req = rtrs_clt_get_copy_req(alive_path, fail_req);
1302 if (req->dir == DMA_TO_DEVICE)
1303 err = rtrs_clt_write_req(req);
1304 else
1305 err = rtrs_clt_read_req(req);
1306 if (err) {
1307 req->in_use = false;
1308 continue;
1309 }
1310 /* Success path */
1311 rtrs_clt_inc_failover_cnt(alive_path->stats);
1312 break;
1313 }
1314 path_it_deinit(&it);
1315 rcu_read_unlock();
1316
1317 return err;
1318}
1319
1320static void fail_all_outstanding_reqs(struct rtrs_clt_path *clt_path)
1321{
1322 struct rtrs_clt_sess *clt = clt_path->clt;
1323 struct rtrs_clt_io_req *req;
1324 int i, err;
1325
1326 if (!clt_path->reqs)
1327 return;
1328 for (i = 0; i < clt_path->queue_depth; ++i) {
1329 req = &clt_path->reqs[i];
1330 if (!req->in_use)
1331 continue;
1332
1333 /*
1334 * Safely (without notification) complete failed request.
1335 * After completion this request is still useble and can
1336 * be failovered to another path.
1337 */
1338 complete_rdma_req(req, -ECONNABORTED, false, true);
1339
1340 err = rtrs_clt_failover_req(clt, req);
1341 if (err)
1342 /* Failover failed, notify anyway */
1343 req->conf(req->priv, err);
1344 }
1345}
1346
1347static void free_path_reqs(struct rtrs_clt_path *clt_path)
1348{
1349 struct rtrs_clt_io_req *req;
1350 int i;
1351
1352 if (!clt_path->reqs)
1353 return;
1354 for (i = 0; i < clt_path->queue_depth; ++i) {
1355 req = &clt_path->reqs[i];
1356 if (req->mr)
1357 ib_dereg_mr(req->mr);
1358 kfree(req->sge);
1359 rtrs_iu_free(req->iu, clt_path->s.dev->ib_dev, 1);
1360 }
1361 kfree(clt_path->reqs);
1362 clt_path->reqs = NULL;
1363}
1364
1365static int alloc_path_reqs(struct rtrs_clt_path *clt_path)
1366{
1367 struct rtrs_clt_io_req *req;
1368 int i, err = -ENOMEM;
1369
1370 clt_path->reqs = kcalloc(clt_path->queue_depth,
1371 sizeof(*clt_path->reqs),
1372 GFP_KERNEL);
1373 if (!clt_path->reqs)
1374 return -ENOMEM;
1375
1376 for (i = 0; i < clt_path->queue_depth; ++i) {
1377 req = &clt_path->reqs[i];
1378 req->iu = rtrs_iu_alloc(1, clt_path->max_hdr_size, GFP_KERNEL,
1379 clt_path->s.dev->ib_dev,
1380 DMA_TO_DEVICE,
1381 rtrs_clt_rdma_done);
1382 if (!req->iu)
1383 goto out;
1384
1385 req->sge = kcalloc(2, sizeof(*req->sge), GFP_KERNEL);
1386 if (!req->sge)
1387 goto out;
1388
1389 req->mr = ib_alloc_mr(clt_path->s.dev->ib_pd,
1390 IB_MR_TYPE_MEM_REG,
1391 clt_path->max_pages_per_mr);
1392 if (IS_ERR(req->mr)) {
1393 err = PTR_ERR(req->mr);
1394 pr_err("Failed to alloc clt_path->max_pages_per_mr %d: %pe\n",
1395 clt_path->max_pages_per_mr, req->mr);
1396 req->mr = NULL;
1397 goto out;
1398 }
1399
1400 init_completion(&req->inv_comp);
1401 }
1402
1403 return 0;
1404
1405out:
1406 free_path_reqs(clt_path);
1407
1408 return err;
1409}
1410
1411static int alloc_permits(struct rtrs_clt_sess *clt)
1412{
1413 unsigned int chunk_bits;
1414 int err, i;
1415
1416 clt->permits_map = bitmap_zalloc(clt->queue_depth, GFP_KERNEL);
1417 if (!clt->permits_map) {
1418 err = -ENOMEM;
1419 goto out_err;
1420 }
1421 clt->permits = kcalloc(clt->queue_depth, permit_size(clt), GFP_KERNEL);
1422 if (!clt->permits) {
1423 err = -ENOMEM;
1424 goto err_map;
1425 }
1426 chunk_bits = ilog2(clt->queue_depth - 1) + 1;
1427 for (i = 0; i < clt->queue_depth; i++) {
1428 struct rtrs_permit *permit;
1429
1430 permit = get_permit(clt, i);
1431 permit->mem_id = i;
1432 permit->mem_off = i << (MAX_IMM_PAYL_BITS - chunk_bits);
1433 }
1434
1435 return 0;
1436
1437err_map:
1438 bitmap_free(clt->permits_map);
1439 clt->permits_map = NULL;
1440out_err:
1441 return err;
1442}
1443
1444static void free_permits(struct rtrs_clt_sess *clt)
1445{
1446 if (clt->permits_map)
1447 wait_event(clt->permits_wait,
1448 bitmap_empty(clt->permits_map, clt->queue_depth));
1449
1450 bitmap_free(clt->permits_map);
1451 clt->permits_map = NULL;
1452 kfree(clt->permits);
1453 clt->permits = NULL;
1454}
1455
1456static void query_fast_reg_mode(struct rtrs_clt_path *clt_path)
1457{
1458 struct ib_device *ib_dev;
1459 u64 max_pages_per_mr;
1460 int mr_page_shift;
1461
1462 ib_dev = clt_path->s.dev->ib_dev;
1463
1464 /*
1465 * Use the smallest page size supported by the HCA, down to a
1466 * minimum of 4096 bytes. We're unlikely to build large sglists
1467 * out of smaller entries.
1468 */
1469 mr_page_shift = max(12, ffs(ib_dev->attrs.page_size_cap) - 1);
1470 max_pages_per_mr = ib_dev->attrs.max_mr_size;
1471 do_div(max_pages_per_mr, (1ull << mr_page_shift));
1472 clt_path->max_pages_per_mr =
1473 min3(clt_path->max_pages_per_mr, (u32)max_pages_per_mr,
1474 ib_dev->attrs.max_fast_reg_page_list_len);
1475 clt_path->clt->max_segments =
1476 min(clt_path->max_pages_per_mr, clt_path->clt->max_segments);
1477}
1478
1479static bool rtrs_clt_change_state_get_old(struct rtrs_clt_path *clt_path,
1480 enum rtrs_clt_state new_state,
1481 enum rtrs_clt_state *old_state)
1482{
1483 bool changed;
1484
1485 spin_lock_irq(&clt_path->state_wq.lock);
1486 if (old_state)
1487 *old_state = clt_path->state;
1488 changed = rtrs_clt_change_state(clt_path, new_state);
1489 spin_unlock_irq(&clt_path->state_wq.lock);
1490
1491 return changed;
1492}
1493
1494static void rtrs_clt_hb_err_handler(struct rtrs_con *c)
1495{
1496 struct rtrs_clt_con *con = container_of(c, typeof(*con), c);
1497
1498 rtrs_rdma_error_recovery(con);
1499}
1500
1501static void rtrs_clt_init_hb(struct rtrs_clt_path *clt_path)
1502{
1503 rtrs_init_hb(&clt_path->s, &io_comp_cqe,
1504 RTRS_HB_INTERVAL_MS,
1505 RTRS_HB_MISSED_MAX,
1506 rtrs_clt_hb_err_handler,
1507 rtrs_wq);
1508}
1509
1510static void rtrs_clt_reconnect_work(struct work_struct *work);
1511static void rtrs_clt_close_work(struct work_struct *work);
1512
1513static void rtrs_clt_err_recovery_work(struct work_struct *work)
1514{
1515 struct rtrs_clt_path *clt_path;
1516 struct rtrs_clt_sess *clt;
1517 int delay_ms;
1518
1519 clt_path = container_of(work, struct rtrs_clt_path, err_recovery_work);
1520 clt = clt_path->clt;
1521 delay_ms = clt->reconnect_delay_sec * 1000;
1522 rtrs_clt_stop_and_destroy_conns(clt_path);
1523 queue_delayed_work(rtrs_wq, &clt_path->reconnect_dwork,
1524 msecs_to_jiffies(delay_ms +
1525 get_random_u32_below(RTRS_RECONNECT_SEED)));
1526}
1527
1528static struct rtrs_clt_path *alloc_path(struct rtrs_clt_sess *clt,
1529 const struct rtrs_addr *path,
1530 size_t con_num, u32 nr_poll_queues)
1531{
1532 struct rtrs_clt_path *clt_path;
1533 int err = -ENOMEM;
1534 int cpu;
1535 size_t total_con;
1536
1537 clt_path = kzalloc(sizeof(*clt_path), GFP_KERNEL);
1538 if (!clt_path)
1539 goto err;
1540
1541 /*
1542 * irqmode and poll
1543 * +1: Extra connection for user messages
1544 */
1545 total_con = con_num + nr_poll_queues + 1;
1546 clt_path->s.con = kcalloc(total_con, sizeof(*clt_path->s.con),
1547 GFP_KERNEL);
1548 if (!clt_path->s.con)
1549 goto err_free_path;
1550
1551 clt_path->s.con_num = total_con;
1552 clt_path->s.irq_con_num = con_num + 1;
1553
1554 clt_path->stats = kzalloc(sizeof(*clt_path->stats), GFP_KERNEL);
1555 if (!clt_path->stats)
1556 goto err_free_con;
1557
1558 mutex_init(&clt_path->init_mutex);
1559 uuid_gen(&clt_path->s.uuid);
1560 memcpy(&clt_path->s.dst_addr, path->dst,
1561 rdma_addr_size((struct sockaddr *)path->dst));
1562
1563 /*
1564 * rdma_resolve_addr() passes src_addr to cma_bind_addr, which
1565 * checks the sa_family to be non-zero. If user passed src_addr=NULL
1566 * the sess->src_addr will contain only zeros, which is then fine.
1567 */
1568 if (path->src)
1569 memcpy(&clt_path->s.src_addr, path->src,
1570 rdma_addr_size((struct sockaddr *)path->src));
1571 strscpy(clt_path->s.sessname, clt->sessname,
1572 sizeof(clt_path->s.sessname));
1573 clt_path->clt = clt;
1574 clt_path->max_pages_per_mr = RTRS_MAX_SEGMENTS;
1575 init_waitqueue_head(&clt_path->state_wq);
1576 clt_path->state = RTRS_CLT_CONNECTING;
1577 atomic_set(&clt_path->connected_cnt, 0);
1578 INIT_WORK(&clt_path->close_work, rtrs_clt_close_work);
1579 INIT_WORK(&clt_path->err_recovery_work, rtrs_clt_err_recovery_work);
1580 INIT_DELAYED_WORK(&clt_path->reconnect_dwork, rtrs_clt_reconnect_work);
1581 rtrs_clt_init_hb(clt_path);
1582
1583 clt_path->mp_skip_entry = alloc_percpu(typeof(*clt_path->mp_skip_entry));
1584 if (!clt_path->mp_skip_entry)
1585 goto err_free_stats;
1586
1587 for_each_possible_cpu(cpu)
1588 INIT_LIST_HEAD(per_cpu_ptr(clt_path->mp_skip_entry, cpu));
1589
1590 err = rtrs_clt_init_stats(clt_path->stats);
1591 if (err)
1592 goto err_free_percpu;
1593
1594 return clt_path;
1595
1596err_free_percpu:
1597 free_percpu(clt_path->mp_skip_entry);
1598err_free_stats:
1599 kfree(clt_path->stats);
1600err_free_con:
1601 kfree(clt_path->s.con);
1602err_free_path:
1603 kfree(clt_path);
1604err:
1605 return ERR_PTR(err);
1606}
1607
1608void free_path(struct rtrs_clt_path *clt_path)
1609{
1610 free_percpu(clt_path->mp_skip_entry);
1611 mutex_destroy(&clt_path->init_mutex);
1612 kfree(clt_path->s.con);
1613 kfree(clt_path->rbufs);
1614 kfree(clt_path);
1615}
1616
1617static int create_con(struct rtrs_clt_path *clt_path, unsigned int cid)
1618{
1619 struct rtrs_clt_con *con;
1620
1621 con = kzalloc(sizeof(*con), GFP_KERNEL);
1622 if (!con)
1623 return -ENOMEM;
1624
1625 /* Map first two connections to the first CPU */
1626 con->cpu = (cid ? cid - 1 : 0) % nr_cpu_ids;
1627 con->c.cid = cid;
1628 con->c.path = &clt_path->s;
1629 /* Align with srv, init as 1 */
1630 atomic_set(&con->c.wr_cnt, 1);
1631 mutex_init(&con->con_mutex);
1632
1633 clt_path->s.con[cid] = &con->c;
1634
1635 return 0;
1636}
1637
1638static void destroy_con(struct rtrs_clt_con *con)
1639{
1640 struct rtrs_clt_path *clt_path = to_clt_path(con->c.path);
1641
1642 clt_path->s.con[con->c.cid] = NULL;
1643 mutex_destroy(&con->con_mutex);
1644 kfree(con);
1645}
1646
1647static int create_con_cq_qp(struct rtrs_clt_con *con)
1648{
1649 struct rtrs_clt_path *clt_path = to_clt_path(con->c.path);
1650 u32 max_send_wr, max_recv_wr, cq_num, max_send_sge, wr_limit;
1651 int err, cq_vector;
1652 struct rtrs_msg_rkey_rsp *rsp;
1653
1654 lockdep_assert_held(&con->con_mutex);
1655 if (con->c.cid == 0) {
1656 max_send_sge = 1;
1657 /* We must be the first here */
1658 if (WARN_ON(clt_path->s.dev))
1659 return -EINVAL;
1660
1661 /*
1662 * The whole session uses device from user connection.
1663 * Be careful not to close user connection before ib dev
1664 * is gracefully put.
1665 */
1666 clt_path->s.dev = rtrs_ib_dev_find_or_add(con->c.cm_id->device,
1667 &dev_pd);
1668 if (!clt_path->s.dev) {
1669 rtrs_wrn(clt_path->clt,
1670 "rtrs_ib_dev_find_get_or_add(): no memory\n");
1671 return -ENOMEM;
1672 }
1673 clt_path->s.dev_ref = 1;
1674 query_fast_reg_mode(clt_path);
1675 wr_limit = clt_path->s.dev->ib_dev->attrs.max_qp_wr;
1676 /*
1677 * Two (request + registration) completion for send
1678 * Two for recv if always_invalidate is set on server
1679 * or one for recv.
1680 * + 2 for drain and heartbeat
1681 * in case qp gets into error state.
1682 */
1683 max_send_wr =
1684 min_t(int, wr_limit, SERVICE_CON_QUEUE_DEPTH * 2 + 2);
1685 max_recv_wr = max_send_wr;
1686 } else {
1687 /*
1688 * Here we assume that session members are correctly set.
1689 * This is always true if user connection (cid == 0) is
1690 * established first.
1691 */
1692 if (WARN_ON(!clt_path->s.dev))
1693 return -EINVAL;
1694 if (WARN_ON(!clt_path->queue_depth))
1695 return -EINVAL;
1696
1697 wr_limit = clt_path->s.dev->ib_dev->attrs.max_qp_wr;
1698 /* Shared between connections */
1699 clt_path->s.dev_ref++;
1700 max_send_wr = min_t(int, wr_limit,
1701 /* QD * (REQ + RSP + FR REGS or INVS) + drain */
1702 clt_path->queue_depth * 4 + 1);
1703 max_recv_wr = min_t(int, wr_limit,
1704 clt_path->queue_depth * 3 + 1);
1705 max_send_sge = 2;
1706 }
1707 atomic_set(&con->c.sq_wr_avail, max_send_wr);
1708 cq_num = max_send_wr + max_recv_wr;
1709 /* alloc iu to recv new rkey reply when server reports flags set */
1710 if (clt_path->flags & RTRS_MSG_NEW_RKEY_F || con->c.cid == 0) {
1711 con->rsp_ius = rtrs_iu_alloc(cq_num, sizeof(*rsp),
1712 GFP_KERNEL,
1713 clt_path->s.dev->ib_dev,
1714 DMA_FROM_DEVICE,
1715 rtrs_clt_rdma_done);
1716 if (!con->rsp_ius)
1717 return -ENOMEM;
1718 con->queue_num = cq_num;
1719 }
1720 cq_vector = con->cpu % clt_path->s.dev->ib_dev->num_comp_vectors;
1721 if (con->c.cid >= clt_path->s.irq_con_num)
1722 err = rtrs_cq_qp_create(&clt_path->s, &con->c, max_send_sge,
1723 cq_vector, cq_num, max_send_wr,
1724 max_recv_wr, IB_POLL_DIRECT);
1725 else
1726 err = rtrs_cq_qp_create(&clt_path->s, &con->c, max_send_sge,
1727 cq_vector, cq_num, max_send_wr,
1728 max_recv_wr, IB_POLL_SOFTIRQ);
1729 /*
1730 * In case of error we do not bother to clean previous allocations,
1731 * since destroy_con_cq_qp() must be called.
1732 */
1733 return err;
1734}
1735
1736static void destroy_con_cq_qp(struct rtrs_clt_con *con)
1737{
1738 struct rtrs_clt_path *clt_path = to_clt_path(con->c.path);
1739
1740 /*
1741 * Be careful here: destroy_con_cq_qp() can be called even
1742 * create_con_cq_qp() failed, see comments there.
1743 */
1744 lockdep_assert_held(&con->con_mutex);
1745 rtrs_cq_qp_destroy(&con->c);
1746 if (con->rsp_ius) {
1747 rtrs_iu_free(con->rsp_ius, clt_path->s.dev->ib_dev,
1748 con->queue_num);
1749 con->rsp_ius = NULL;
1750 con->queue_num = 0;
1751 }
1752 if (clt_path->s.dev_ref && !--clt_path->s.dev_ref) {
1753 rtrs_ib_dev_put(clt_path->s.dev);
1754 clt_path->s.dev = NULL;
1755 }
1756}
1757
1758static void stop_cm(struct rtrs_clt_con *con)
1759{
1760 rdma_disconnect(con->c.cm_id);
1761 if (con->c.qp)
1762 ib_drain_qp(con->c.qp);
1763}
1764
1765static void destroy_cm(struct rtrs_clt_con *con)
1766{
1767 rdma_destroy_id(con->c.cm_id);
1768 con->c.cm_id = NULL;
1769}
1770
1771static int rtrs_rdma_addr_resolved(struct rtrs_clt_con *con)
1772{
1773 struct rtrs_path *s = con->c.path;
1774 int err;
1775
1776 mutex_lock(&con->con_mutex);
1777 err = create_con_cq_qp(con);
1778 mutex_unlock(&con->con_mutex);
1779 if (err) {
1780 rtrs_err(s, "create_con_cq_qp(), err: %d\n", err);
1781 return err;
1782 }
1783 err = rdma_resolve_route(con->c.cm_id, RTRS_CONNECT_TIMEOUT_MS);
1784 if (err)
1785 rtrs_err(s, "Resolving route failed, err: %d\n", err);
1786
1787 return err;
1788}
1789
1790static int rtrs_rdma_route_resolved(struct rtrs_clt_con *con)
1791{
1792 struct rtrs_clt_path *clt_path = to_clt_path(con->c.path);
1793 struct rtrs_clt_sess *clt = clt_path->clt;
1794 struct rtrs_msg_conn_req msg;
1795 struct rdma_conn_param param;
1796
1797 int err;
1798
1799 param = (struct rdma_conn_param) {
1800 .retry_count = 7,
1801 .rnr_retry_count = 7,
1802 .private_data = &msg,
1803 .private_data_len = sizeof(msg),
1804 };
1805
1806 msg = (struct rtrs_msg_conn_req) {
1807 .magic = cpu_to_le16(RTRS_MAGIC),
1808 .version = cpu_to_le16(RTRS_PROTO_VER),
1809 .cid = cpu_to_le16(con->c.cid),
1810 .cid_num = cpu_to_le16(clt_path->s.con_num),
1811 .recon_cnt = cpu_to_le16(clt_path->s.recon_cnt),
1812 };
1813 msg.first_conn = clt_path->for_new_clt ? FIRST_CONN : 0;
1814 uuid_copy(&msg.sess_uuid, &clt_path->s.uuid);
1815 uuid_copy(&msg.paths_uuid, &clt->paths_uuid);
1816
1817 err = rdma_connect_locked(con->c.cm_id, ¶m);
1818 if (err)
1819 rtrs_err(clt, "rdma_connect_locked(): %d\n", err);
1820
1821 return err;
1822}
1823
1824static int rtrs_rdma_conn_established(struct rtrs_clt_con *con,
1825 struct rdma_cm_event *ev)
1826{
1827 struct rtrs_clt_path *clt_path = to_clt_path(con->c.path);
1828 struct rtrs_clt_sess *clt = clt_path->clt;
1829 const struct rtrs_msg_conn_rsp *msg;
1830 u16 version, queue_depth;
1831 int errno;
1832 u8 len;
1833
1834 msg = ev->param.conn.private_data;
1835 len = ev->param.conn.private_data_len;
1836 if (len < sizeof(*msg)) {
1837 rtrs_err(clt, "Invalid RTRS connection response\n");
1838 return -ECONNRESET;
1839 }
1840 if (le16_to_cpu(msg->magic) != RTRS_MAGIC) {
1841 rtrs_err(clt, "Invalid RTRS magic\n");
1842 return -ECONNRESET;
1843 }
1844 version = le16_to_cpu(msg->version);
1845 if (version >> 8 != RTRS_PROTO_VER_MAJOR) {
1846 rtrs_err(clt, "Unsupported major RTRS version: %d, expected %d\n",
1847 version >> 8, RTRS_PROTO_VER_MAJOR);
1848 return -ECONNRESET;
1849 }
1850 errno = le16_to_cpu(msg->errno);
1851 if (errno) {
1852 rtrs_err(clt, "Invalid RTRS message: errno %d\n",
1853 errno);
1854 return -ECONNRESET;
1855 }
1856 if (con->c.cid == 0) {
1857 queue_depth = le16_to_cpu(msg->queue_depth);
1858
1859 if (clt_path->queue_depth > 0 && queue_depth != clt_path->queue_depth) {
1860 rtrs_err(clt, "Error: queue depth changed\n");
1861
1862 /*
1863 * Stop any more reconnection attempts
1864 */
1865 clt_path->reconnect_attempts = -1;
1866 rtrs_err(clt,
1867 "Disabling auto-reconnect. Trigger a manual reconnect after issue is resolved\n");
1868 return -ECONNRESET;
1869 }
1870
1871 if (!clt_path->rbufs) {
1872 clt_path->rbufs = kcalloc(queue_depth,
1873 sizeof(*clt_path->rbufs),
1874 GFP_KERNEL);
1875 if (!clt_path->rbufs)
1876 return -ENOMEM;
1877 }
1878 clt_path->queue_depth = queue_depth;
1879 clt_path->s.signal_interval = min_not_zero(queue_depth,
1880 (unsigned short) SERVICE_CON_QUEUE_DEPTH);
1881 clt_path->max_hdr_size = le32_to_cpu(msg->max_hdr_size);
1882 clt_path->max_io_size = le32_to_cpu(msg->max_io_size);
1883 clt_path->flags = le32_to_cpu(msg->flags);
1884 clt_path->chunk_size = clt_path->max_io_size + clt_path->max_hdr_size;
1885
1886 /*
1887 * Global IO size is always a minimum.
1888 * If while a reconnection server sends us a value a bit
1889 * higher - client does not care and uses cached minimum.
1890 *
1891 * Since we can have several sessions (paths) restablishing
1892 * connections in parallel, use lock.
1893 */
1894 mutex_lock(&clt->paths_mutex);
1895 clt->queue_depth = clt_path->queue_depth;
1896 clt->max_io_size = min_not_zero(clt_path->max_io_size,
1897 clt->max_io_size);
1898 mutex_unlock(&clt->paths_mutex);
1899
1900 /*
1901 * Cache the hca_port and hca_name for sysfs
1902 */
1903 clt_path->hca_port = con->c.cm_id->port_num;
1904 scnprintf(clt_path->hca_name, sizeof(clt_path->hca_name),
1905 clt_path->s.dev->ib_dev->name);
1906 clt_path->s.src_addr = con->c.cm_id->route.addr.src_addr;
1907 /* set for_new_clt, to allow future reconnect on any path */
1908 clt_path->for_new_clt = 1;
1909 }
1910
1911 return 0;
1912}
1913
1914static inline void flag_success_on_conn(struct rtrs_clt_con *con)
1915{
1916 struct rtrs_clt_path *clt_path = to_clt_path(con->c.path);
1917
1918 atomic_inc(&clt_path->connected_cnt);
1919 con->cm_err = 1;
1920}
1921
1922static int rtrs_rdma_conn_rejected(struct rtrs_clt_con *con,
1923 struct rdma_cm_event *ev)
1924{
1925 struct rtrs_path *s = con->c.path;
1926 const struct rtrs_msg_conn_rsp *msg;
1927 const char *rej_msg;
1928 int status, errno;
1929 u8 data_len;
1930
1931 status = ev->status;
1932 rej_msg = rdma_reject_msg(con->c.cm_id, status);
1933 msg = rdma_consumer_reject_data(con->c.cm_id, ev, &data_len);
1934
1935 if (msg && data_len >= sizeof(*msg)) {
1936 errno = (int16_t)le16_to_cpu(msg->errno);
1937 if (errno == -EBUSY)
1938 rtrs_err(s,
1939 "Previous session is still exists on the server, please reconnect later\n");
1940 else
1941 rtrs_err(s,
1942 "Connect rejected: status %d (%s), rtrs errno %d\n",
1943 status, rej_msg, errno);
1944 } else {
1945 rtrs_err(s,
1946 "Connect rejected but with malformed message: status %d (%s)\n",
1947 status, rej_msg);
1948 }
1949
1950 return -ECONNRESET;
1951}
1952
1953void rtrs_clt_close_conns(struct rtrs_clt_path *clt_path, bool wait)
1954{
1955 trace_rtrs_clt_close_conns(clt_path);
1956
1957 if (rtrs_clt_change_state_get_old(clt_path, RTRS_CLT_CLOSING, NULL))
1958 queue_work(rtrs_wq, &clt_path->close_work);
1959 if (wait)
1960 flush_work(&clt_path->close_work);
1961}
1962
1963static inline void flag_error_on_conn(struct rtrs_clt_con *con, int cm_err)
1964{
1965 if (con->cm_err == 1) {
1966 struct rtrs_clt_path *clt_path;
1967
1968 clt_path = to_clt_path(con->c.path);
1969 if (atomic_dec_and_test(&clt_path->connected_cnt))
1970
1971 wake_up(&clt_path->state_wq);
1972 }
1973 con->cm_err = cm_err;
1974}
1975
1976static int rtrs_clt_rdma_cm_handler(struct rdma_cm_id *cm_id,
1977 struct rdma_cm_event *ev)
1978{
1979 struct rtrs_clt_con *con = cm_id->context;
1980 struct rtrs_path *s = con->c.path;
1981 struct rtrs_clt_path *clt_path = to_clt_path(s);
1982 int cm_err = 0;
1983
1984 switch (ev->event) {
1985 case RDMA_CM_EVENT_ADDR_RESOLVED:
1986 cm_err = rtrs_rdma_addr_resolved(con);
1987 break;
1988 case RDMA_CM_EVENT_ROUTE_RESOLVED:
1989 cm_err = rtrs_rdma_route_resolved(con);
1990 break;
1991 case RDMA_CM_EVENT_ESTABLISHED:
1992 cm_err = rtrs_rdma_conn_established(con, ev);
1993 if (!cm_err) {
1994 /*
1995 * Report success and wake up. Here we abuse state_wq,
1996 * i.e. wake up without state change, but we set cm_err.
1997 */
1998 flag_success_on_conn(con);
1999 wake_up(&clt_path->state_wq);
2000 return 0;
2001 }
2002 break;
2003 case RDMA_CM_EVENT_REJECTED:
2004 cm_err = rtrs_rdma_conn_rejected(con, ev);
2005 break;
2006 case RDMA_CM_EVENT_DISCONNECTED:
2007 /* No message for disconnecting */
2008 cm_err = -ECONNRESET;
2009 break;
2010 case RDMA_CM_EVENT_CONNECT_ERROR:
2011 case RDMA_CM_EVENT_UNREACHABLE:
2012 case RDMA_CM_EVENT_ADDR_CHANGE:
2013 case RDMA_CM_EVENT_TIMEWAIT_EXIT:
2014 rtrs_wrn(s, "CM error (CM event: %s, err: %d)\n",
2015 rdma_event_msg(ev->event), ev->status);
2016 cm_err = -ECONNRESET;
2017 break;
2018 case RDMA_CM_EVENT_ADDR_ERROR:
2019 case RDMA_CM_EVENT_ROUTE_ERROR:
2020 rtrs_wrn(s, "CM error (CM event: %s, err: %d)\n",
2021 rdma_event_msg(ev->event), ev->status);
2022 cm_err = -EHOSTUNREACH;
2023 break;
2024 case RDMA_CM_EVENT_DEVICE_REMOVAL:
2025 /*
2026 * Device removal is a special case. Queue close and return 0.
2027 */
2028 rtrs_wrn_rl(s, "CM event: %s, status: %d\n", rdma_event_msg(ev->event),
2029 ev->status);
2030 rtrs_clt_close_conns(clt_path, false);
2031 return 0;
2032 default:
2033 rtrs_err(s, "Unexpected RDMA CM error (CM event: %s, err: %d)\n",
2034 rdma_event_msg(ev->event), ev->status);
2035 cm_err = -ECONNRESET;
2036 break;
2037 }
2038
2039 if (cm_err) {
2040 /*
2041 * cm error makes sense only on connection establishing,
2042 * in other cases we rely on normal procedure of reconnecting.
2043 */
2044 flag_error_on_conn(con, cm_err);
2045 rtrs_rdma_error_recovery(con);
2046 }
2047
2048 return 0;
2049}
2050
2051/* The caller should do the cleanup in case of error */
2052static int create_cm(struct rtrs_clt_con *con)
2053{
2054 struct rtrs_path *s = con->c.path;
2055 struct rtrs_clt_path *clt_path = to_clt_path(s);
2056 struct rdma_cm_id *cm_id;
2057 int err;
2058
2059 cm_id = rdma_create_id(&init_net, rtrs_clt_rdma_cm_handler, con,
2060 clt_path->s.dst_addr.ss_family == AF_IB ?
2061 RDMA_PS_IB : RDMA_PS_TCP, IB_QPT_RC);
2062 if (IS_ERR(cm_id)) {
2063 rtrs_err(s, "Failed to create CM ID, err: %pe\n", cm_id);
2064 return PTR_ERR(cm_id);
2065 }
2066 con->c.cm_id = cm_id;
2067 con->cm_err = 0;
2068 /* allow the port to be reused */
2069 err = rdma_set_reuseaddr(cm_id, 1);
2070 if (err != 0) {
2071 rtrs_err(s, "Set address reuse failed, err: %d\n", err);
2072 return err;
2073 }
2074 err = rdma_resolve_addr(cm_id, (struct sockaddr *)&clt_path->s.src_addr,
2075 (struct sockaddr *)&clt_path->s.dst_addr,
2076 RTRS_CONNECT_TIMEOUT_MS);
2077 if (err) {
2078 rtrs_err(s, "Failed to resolve address, err: %d\n", err);
2079 return err;
2080 }
2081 /*
2082 * Combine connection status and session events. This is needed
2083 * for waiting two possible cases: cm_err has something meaningful
2084 * or session state was really changed to error by device removal.
2085 */
2086 err = wait_event_interruptible_timeout(
2087 clt_path->state_wq,
2088 con->cm_err || clt_path->state != RTRS_CLT_CONNECTING,
2089 msecs_to_jiffies(RTRS_CONNECT_TIMEOUT_MS));
2090 if (err == 0 || err == -ERESTARTSYS) {
2091 if (err == 0)
2092 err = -ETIMEDOUT;
2093 /* Timedout or interrupted */
2094 return err;
2095 }
2096 if (con->cm_err < 0)
2097 return con->cm_err;
2098 if (READ_ONCE(clt_path->state) != RTRS_CLT_CONNECTING)
2099 /* Device removal */
2100 return -ECONNABORTED;
2101
2102 return 0;
2103}
2104
2105static void rtrs_clt_path_up(struct rtrs_clt_path *clt_path)
2106{
2107 struct rtrs_clt_sess *clt = clt_path->clt;
2108 int up;
2109
2110 /*
2111 * We can fire RECONNECTED event only when all paths were
2112 * connected on rtrs_clt_open(), then each was disconnected
2113 * and the first one connected again. That's why this nasty
2114 * game with counter value.
2115 */
2116
2117 mutex_lock(&clt->paths_ev_mutex);
2118 up = ++clt->paths_up;
2119 /*
2120 * Here it is safe to access paths num directly since up counter
2121 * is greater than MAX_PATHS_NUM only while rtrs_clt_open() is
2122 * in progress, thus paths removals are impossible.
2123 */
2124 if (up > MAX_PATHS_NUM && up == MAX_PATHS_NUM + clt->paths_num)
2125 clt->paths_up = clt->paths_num;
2126 else if (up == 1)
2127 clt->link_ev(clt->priv, RTRS_CLT_LINK_EV_RECONNECTED);
2128 mutex_unlock(&clt->paths_ev_mutex);
2129
2130 /* Mark session as established */
2131 clt_path->established = true;
2132 clt_path->reconnect_attempts = 0;
2133 clt_path->stats->reconnects.successful_cnt++;
2134}
2135
2136static void rtrs_clt_path_down(struct rtrs_clt_path *clt_path)
2137{
2138 struct rtrs_clt_sess *clt = clt_path->clt;
2139
2140 if (!clt_path->established)
2141 return;
2142
2143 clt_path->established = false;
2144 mutex_lock(&clt->paths_ev_mutex);
2145 WARN_ON(!clt->paths_up);
2146 if (--clt->paths_up == 0)
2147 clt->link_ev(clt->priv, RTRS_CLT_LINK_EV_DISCONNECTED);
2148 mutex_unlock(&clt->paths_ev_mutex);
2149}
2150
2151static void rtrs_clt_stop_and_destroy_conns(struct rtrs_clt_path *clt_path)
2152{
2153 struct rtrs_clt_con *con;
2154 unsigned int cid;
2155
2156 WARN_ON(READ_ONCE(clt_path->state) == RTRS_CLT_CONNECTED);
2157
2158 /*
2159 * Possible race with rtrs_clt_open(), when DEVICE_REMOVAL comes
2160 * exactly in between. Start destroying after it finishes.
2161 */
2162 mutex_lock(&clt_path->init_mutex);
2163 mutex_unlock(&clt_path->init_mutex);
2164
2165 /*
2166 * All IO paths must observe !CONNECTED state before we
2167 * free everything.
2168 */
2169 synchronize_rcu();
2170
2171 rtrs_stop_hb(&clt_path->s);
2172
2173 /*
2174 * The order it utterly crucial: firstly disconnect and complete all
2175 * rdma requests with error (thus set in_use=false for requests),
2176 * then fail outstanding requests checking in_use for each, and
2177 * eventually notify upper layer about session disconnection.
2178 */
2179
2180 for (cid = 0; cid < clt_path->s.con_num; cid++) {
2181 if (!clt_path->s.con[cid])
2182 break;
2183 con = to_clt_con(clt_path->s.con[cid]);
2184 stop_cm(con);
2185 }
2186 fail_all_outstanding_reqs(clt_path);
2187 free_path_reqs(clt_path);
2188 rtrs_clt_path_down(clt_path);
2189
2190 /*
2191 * Wait for graceful shutdown, namely when peer side invokes
2192 * rdma_disconnect(). 'connected_cnt' is decremented only on
2193 * CM events, thus if other side had crashed and hb has detected
2194 * something is wrong, here we will stuck for exactly timeout ms,
2195 * since CM does not fire anything. That is fine, we are not in
2196 * hurry.
2197 */
2198 wait_event_timeout(clt_path->state_wq,
2199 !atomic_read(&clt_path->connected_cnt),
2200 msecs_to_jiffies(RTRS_CONNECT_TIMEOUT_MS));
2201
2202 for (cid = 0; cid < clt_path->s.con_num; cid++) {
2203 if (!clt_path->s.con[cid])
2204 break;
2205 con = to_clt_con(clt_path->s.con[cid]);
2206 mutex_lock(&con->con_mutex);
2207 destroy_con_cq_qp(con);
2208 mutex_unlock(&con->con_mutex);
2209 destroy_cm(con);
2210 destroy_con(con);
2211 }
2212}
2213
2214static void rtrs_clt_remove_path_from_arr(struct rtrs_clt_path *clt_path)
2215{
2216 struct rtrs_clt_sess *clt = clt_path->clt;
2217 struct rtrs_clt_path *next;
2218 bool wait_for_grace = false;
2219 int cpu;
2220
2221 mutex_lock(&clt->paths_mutex);
2222 list_del_rcu(&clt_path->s.entry);
2223
2224 /* Make sure everybody observes path removal. */
2225 synchronize_rcu();
2226
2227 /*
2228 * At this point nobody sees @sess in the list, but still we have
2229 * dangling pointer @pcpu_path which _can_ point to @sess. Since
2230 * nobody can observe @sess in the list, we guarantee that IO path
2231 * will not assign @sess to @pcpu_path, i.e. @pcpu_path can be equal
2232 * to @sess, but can never again become @sess.
2233 */
2234
2235 /*
2236 * Decrement paths number only after grace period, because
2237 * caller of do_each_path() must firstly observe list without
2238 * path and only then decremented paths number.
2239 *
2240 * Otherwise there can be the following situation:
2241 * o Two paths exist and IO is coming.
2242 * o One path is removed:
2243 * CPU#0 CPU#1
2244 * do_each_path(): rtrs_clt_remove_path_from_arr():
2245 * path = get_next_path()
2246 * ^^^ list_del_rcu(path)
2247 * [!CONNECTED path] clt->paths_num--
2248 * ^^^^^^^^^
2249 * load clt->paths_num from 2 to 1
2250 * ^^^^^^^^^
2251 * sees 1
2252 *
2253 * path is observed as !CONNECTED, but do_each_path() loop
2254 * ends, because expression i < clt->paths_num is false.
2255 */
2256 clt->paths_num--;
2257
2258 /*
2259 * Get @next connection from current @sess which is going to be
2260 * removed. If @sess is the last element, then @next is NULL.
2261 */
2262 rcu_read_lock();
2263 next = rtrs_clt_get_next_path_or_null(&clt->paths_list, clt_path);
2264 rcu_read_unlock();
2265
2266 /*
2267 * @pcpu paths can still point to the path which is going to be
2268 * removed, so change the pointer manually.
2269 */
2270 for_each_possible_cpu(cpu) {
2271 struct rtrs_clt_path __rcu **ppcpu_path;
2272
2273 ppcpu_path = per_cpu_ptr(clt->pcpu_path, cpu);
2274 if (rcu_dereference_protected(*ppcpu_path,
2275 lockdep_is_held(&clt->paths_mutex)) != clt_path)
2276 /*
2277 * synchronize_rcu() was called just after deleting
2278 * entry from the list, thus IO code path cannot
2279 * change pointer back to the pointer which is going
2280 * to be removed, we are safe here.
2281 */
2282 continue;
2283
2284 /*
2285 * We race with IO code path, which also changes pointer,
2286 * thus we have to be careful not to overwrite it.
2287 */
2288 if (try_cmpxchg((struct rtrs_clt_path **)ppcpu_path, &clt_path,
2289 next))
2290 /*
2291 * @ppcpu_path was successfully replaced with @next,
2292 * that means that someone could also pick up the
2293 * @sess and dereferencing it right now, so wait for
2294 * a grace period is required.
2295 */
2296 wait_for_grace = true;
2297 }
2298 if (wait_for_grace)
2299 synchronize_rcu();
2300
2301 mutex_unlock(&clt->paths_mutex);
2302}
2303
2304static void rtrs_clt_add_path_to_arr(struct rtrs_clt_path *clt_path)
2305{
2306 struct rtrs_clt_sess *clt = clt_path->clt;
2307
2308 mutex_lock(&clt->paths_mutex);
2309 clt->paths_num++;
2310
2311 list_add_tail_rcu(&clt_path->s.entry, &clt->paths_list);
2312 mutex_unlock(&clt->paths_mutex);
2313}
2314
2315static void rtrs_clt_close_work(struct work_struct *work)
2316{
2317 struct rtrs_clt_path *clt_path;
2318
2319 clt_path = container_of(work, struct rtrs_clt_path, close_work);
2320
2321 cancel_work_sync(&clt_path->err_recovery_work);
2322 cancel_delayed_work_sync(&clt_path->reconnect_dwork);
2323 rtrs_clt_stop_and_destroy_conns(clt_path);
2324 rtrs_clt_change_state_get_old(clt_path, RTRS_CLT_CLOSED, NULL);
2325}
2326
2327static int init_conns(struct rtrs_clt_path *clt_path)
2328{
2329 unsigned int cid;
2330 int err, i;
2331
2332 /*
2333 * On every new session connections increase reconnect counter
2334 * to avoid clashes with previous sessions not yet closed
2335 * sessions on a server side.
2336 */
2337 clt_path->s.recon_cnt++;
2338
2339 /* Establish all RDMA connections */
2340 for (cid = 0; cid < clt_path->s.con_num; cid++) {
2341 err = create_con(clt_path, cid);
2342 if (err)
2343 goto destroy;
2344
2345 err = create_cm(to_clt_con(clt_path->s.con[cid]));
2346 if (err)
2347 goto destroy;
2348 }
2349 err = alloc_path_reqs(clt_path);
2350 if (err)
2351 goto destroy;
2352
2353 return 0;
2354
2355destroy:
2356 /* Make sure we do the cleanup in the order they are created */
2357 for (i = 0; i <= cid; i++) {
2358 struct rtrs_clt_con *con;
2359
2360 if (!clt_path->s.con[i])
2361 break;
2362
2363 con = to_clt_con(clt_path->s.con[i]);
2364 if (con->c.cm_id) {
2365 stop_cm(con);
2366 mutex_lock(&con->con_mutex);
2367 destroy_con_cq_qp(con);
2368 mutex_unlock(&con->con_mutex);
2369 destroy_cm(con);
2370 }
2371 destroy_con(con);
2372 }
2373 /*
2374 * If we've never taken async path and got an error, say,
2375 * doing rdma_resolve_addr(), switch to CONNECTION_ERR state
2376 * manually to keep reconnecting.
2377 */
2378 rtrs_clt_change_state_get_old(clt_path, RTRS_CLT_CONNECTING_ERR, NULL);
2379
2380 return err;
2381}
2382
2383static void rtrs_clt_info_req_done(struct ib_cq *cq, struct ib_wc *wc)
2384{
2385 struct rtrs_clt_con *con = to_clt_con(wc->qp->qp_context);
2386 struct rtrs_clt_path *clt_path = to_clt_path(con->c.path);
2387 struct rtrs_iu *iu;
2388
2389 iu = container_of(wc->wr_cqe, struct rtrs_iu, cqe);
2390 rtrs_iu_free(iu, clt_path->s.dev->ib_dev, 1);
2391
2392 if (wc->status != IB_WC_SUCCESS) {
2393 rtrs_err(clt_path->clt, "Path info request send failed: %s\n",
2394 ib_wc_status_msg(wc->status));
2395 rtrs_clt_change_state_get_old(clt_path, RTRS_CLT_CONNECTING_ERR, NULL);
2396 return;
2397 }
2398
2399 rtrs_clt_update_wc_stats(con);
2400}
2401
2402static int process_info_rsp(struct rtrs_clt_path *clt_path,
2403 const struct rtrs_msg_info_rsp *msg)
2404{
2405 unsigned int sg_cnt, total_len;
2406 int i, sgi;
2407
2408 sg_cnt = le16_to_cpu(msg->sg_cnt);
2409 if (!sg_cnt || (clt_path->queue_depth % sg_cnt)) {
2410 rtrs_err(clt_path->clt,
2411 "Incorrect sg_cnt %d, is not multiple\n",
2412 sg_cnt);
2413 return -EINVAL;
2414 }
2415
2416 /*
2417 * Check if IB immediate data size is enough to hold the mem_id and
2418 * the offset inside the memory chunk.
2419 */
2420 if ((ilog2(sg_cnt - 1) + 1) + (ilog2(clt_path->chunk_size - 1) + 1) >
2421 MAX_IMM_PAYL_BITS) {
2422 rtrs_err(clt_path->clt,
2423 "RDMA immediate size (%db) not enough to encode %d buffers of size %dB\n",
2424 MAX_IMM_PAYL_BITS, sg_cnt, clt_path->chunk_size);
2425 return -EINVAL;
2426 }
2427 total_len = 0;
2428 for (sgi = 0, i = 0; sgi < sg_cnt && i < clt_path->queue_depth; sgi++) {
2429 const struct rtrs_sg_desc *desc = &msg->desc[sgi];
2430 u32 len, rkey;
2431 u64 addr;
2432
2433 addr = le64_to_cpu(desc->addr);
2434 rkey = le32_to_cpu(desc->key);
2435 len = le32_to_cpu(desc->len);
2436
2437 total_len += len;
2438
2439 if (!len || (len % clt_path->chunk_size)) {
2440 rtrs_err(clt_path->clt, "Incorrect [%d].len %d\n",
2441 sgi,
2442 len);
2443 return -EINVAL;
2444 }
2445 for ( ; len && i < clt_path->queue_depth; i++) {
2446 clt_path->rbufs[i].addr = addr;
2447 clt_path->rbufs[i].rkey = rkey;
2448
2449 len -= clt_path->chunk_size;
2450 addr += clt_path->chunk_size;
2451 }
2452 }
2453 /* Sanity check */
2454 if (sgi != sg_cnt || i != clt_path->queue_depth) {
2455 rtrs_err(clt_path->clt,
2456 "Incorrect sg vector, not fully mapped\n");
2457 return -EINVAL;
2458 }
2459 if (total_len != clt_path->chunk_size * clt_path->queue_depth) {
2460 rtrs_err(clt_path->clt, "Incorrect total_len %d\n", total_len);
2461 return -EINVAL;
2462 }
2463
2464 return 0;
2465}
2466
2467static void rtrs_clt_info_rsp_done(struct ib_cq *cq, struct ib_wc *wc)
2468{
2469 struct rtrs_clt_con *con = to_clt_con(wc->qp->qp_context);
2470 struct rtrs_clt_path *clt_path = to_clt_path(con->c.path);
2471 struct rtrs_msg_info_rsp *msg;
2472 enum rtrs_clt_state state;
2473 struct rtrs_iu *iu;
2474 size_t rx_sz;
2475 int err;
2476
2477 state = RTRS_CLT_CONNECTING_ERR;
2478
2479 WARN_ON(con->c.cid);
2480 iu = container_of(wc->wr_cqe, struct rtrs_iu, cqe);
2481 if (wc->status != IB_WC_SUCCESS) {
2482 rtrs_err(clt_path->clt, "Path info response recv failed: %s\n",
2483 ib_wc_status_msg(wc->status));
2484 goto out;
2485 }
2486 WARN_ON(wc->opcode != IB_WC_RECV);
2487
2488 if (wc->byte_len < sizeof(*msg)) {
2489 rtrs_err(clt_path->clt, "Path info response is malformed: size %d\n",
2490 wc->byte_len);
2491 goto out;
2492 }
2493 ib_dma_sync_single_for_cpu(clt_path->s.dev->ib_dev, iu->dma_addr,
2494 iu->size, DMA_FROM_DEVICE);
2495 msg = iu->buf;
2496 if (le16_to_cpu(msg->type) != RTRS_MSG_INFO_RSP) {
2497 rtrs_err(clt_path->clt, "Path info response is malformed: type %d\n",
2498 le16_to_cpu(msg->type));
2499 goto out;
2500 }
2501 rx_sz = sizeof(*msg);
2502 rx_sz += sizeof(msg->desc[0]) * le16_to_cpu(msg->sg_cnt);
2503 if (wc->byte_len < rx_sz) {
2504 rtrs_err(clt_path->clt, "Path info response is malformed: size %d\n",
2505 wc->byte_len);
2506 goto out;
2507 }
2508 err = process_info_rsp(clt_path, msg);
2509 if (err)
2510 goto out;
2511
2512 err = post_recv_path(clt_path);
2513 if (err)
2514 goto out;
2515
2516 state = RTRS_CLT_CONNECTED;
2517
2518out:
2519 rtrs_clt_update_wc_stats(con);
2520 rtrs_iu_free(iu, clt_path->s.dev->ib_dev, 1);
2521 rtrs_clt_change_state_get_old(clt_path, state, NULL);
2522}
2523
2524static int rtrs_send_path_info(struct rtrs_clt_path *clt_path)
2525{
2526 struct rtrs_clt_con *usr_con = to_clt_con(clt_path->s.con[0]);
2527 struct rtrs_msg_info_req *msg;
2528 struct rtrs_iu *tx_iu, *rx_iu;
2529 size_t rx_sz;
2530 int err;
2531
2532 rx_sz = sizeof(struct rtrs_msg_info_rsp);
2533 rx_sz += sizeof(struct rtrs_sg_desc) * clt_path->queue_depth;
2534
2535 tx_iu = rtrs_iu_alloc(1, sizeof(struct rtrs_msg_info_req), GFP_KERNEL,
2536 clt_path->s.dev->ib_dev, DMA_TO_DEVICE,
2537 rtrs_clt_info_req_done);
2538 rx_iu = rtrs_iu_alloc(1, rx_sz, GFP_KERNEL, clt_path->s.dev->ib_dev,
2539 DMA_FROM_DEVICE, rtrs_clt_info_rsp_done);
2540 if (!tx_iu || !rx_iu) {
2541 err = -ENOMEM;
2542 goto out;
2543 }
2544 /* Prepare for getting info response */
2545 err = rtrs_iu_post_recv(&usr_con->c, rx_iu);
2546 if (err) {
2547 rtrs_err(clt_path->clt, "rtrs_iu_post_recv(), err: %d\n", err);
2548 goto out;
2549 }
2550 rx_iu = NULL;
2551
2552 msg = tx_iu->buf;
2553 msg->type = cpu_to_le16(RTRS_MSG_INFO_REQ);
2554 memcpy(msg->pathname, clt_path->s.sessname, sizeof(msg->pathname));
2555
2556 ib_dma_sync_single_for_device(clt_path->s.dev->ib_dev,
2557 tx_iu->dma_addr,
2558 tx_iu->size, DMA_TO_DEVICE);
2559
2560 /* Send info request */
2561 err = rtrs_iu_post_send(&usr_con->c, tx_iu, sizeof(*msg), NULL);
2562 if (err) {
2563 rtrs_err(clt_path->clt, "rtrs_iu_post_send(), err: %d\n", err);
2564 goto out;
2565 }
2566 tx_iu = NULL;
2567
2568 /* Wait for state change */
2569 wait_event_interruptible_timeout(clt_path->state_wq,
2570 clt_path->state != RTRS_CLT_CONNECTING,
2571 msecs_to_jiffies(
2572 RTRS_CONNECT_TIMEOUT_MS));
2573 if (READ_ONCE(clt_path->state) != RTRS_CLT_CONNECTED) {
2574 if (READ_ONCE(clt_path->state) == RTRS_CLT_CONNECTING_ERR)
2575 err = -ECONNRESET;
2576 else
2577 err = -ETIMEDOUT;
2578 }
2579
2580out:
2581 if (tx_iu)
2582 rtrs_iu_free(tx_iu, clt_path->s.dev->ib_dev, 1);
2583 if (rx_iu)
2584 rtrs_iu_free(rx_iu, clt_path->s.dev->ib_dev, 1);
2585 if (err)
2586 /* If we've never taken async path because of malloc problems */
2587 rtrs_clt_change_state_get_old(clt_path,
2588 RTRS_CLT_CONNECTING_ERR, NULL);
2589
2590 return err;
2591}
2592
2593/**
2594 * init_path() - establishes all path connections and does handshake
2595 * @clt_path: client path.
2596 * In case of error full close or reconnect procedure should be taken,
2597 * because reconnect or close async works can be started.
2598 */
2599static int init_path(struct rtrs_clt_path *clt_path)
2600{
2601 int err;
2602 char str[NAME_MAX];
2603 struct rtrs_addr path = {
2604 .src = &clt_path->s.src_addr,
2605 .dst = &clt_path->s.dst_addr,
2606 };
2607
2608 rtrs_addr_to_str(&path, str, sizeof(str));
2609
2610 mutex_lock(&clt_path->init_mutex);
2611 err = init_conns(clt_path);
2612 if (err) {
2613 rtrs_err(clt_path->clt,
2614 "init_conns() failed: err=%d path=%s [%s:%u]\n", err,
2615 str, clt_path->hca_name, clt_path->hca_port);
2616 goto out;
2617 }
2618 err = rtrs_send_path_info(clt_path);
2619 if (err) {
2620 rtrs_err(clt_path->clt,
2621 "rtrs_send_path_info() failed: err=%d path=%s [%s:%u]\n",
2622 err, str, clt_path->hca_name, clt_path->hca_port);
2623 goto out;
2624 }
2625 rtrs_clt_path_up(clt_path);
2626 rtrs_start_hb(&clt_path->s);
2627out:
2628 mutex_unlock(&clt_path->init_mutex);
2629
2630 return err;
2631}
2632
2633static void rtrs_clt_reconnect_work(struct work_struct *work)
2634{
2635 struct rtrs_clt_path *clt_path;
2636 struct rtrs_clt_sess *clt;
2637 int err;
2638
2639 clt_path = container_of(to_delayed_work(work), struct rtrs_clt_path,
2640 reconnect_dwork);
2641 clt = clt_path->clt;
2642
2643 trace_rtrs_clt_reconnect_work(clt_path);
2644
2645 if (READ_ONCE(clt_path->state) != RTRS_CLT_RECONNECTING)
2646 return;
2647
2648 if (clt_path->reconnect_attempts >= clt->max_reconnect_attempts) {
2649 /* Close a path completely if max attempts is reached */
2650 rtrs_clt_close_conns(clt_path, false);
2651 return;
2652 }
2653 clt_path->reconnect_attempts++;
2654
2655 msleep(RTRS_RECONNECT_BACKOFF);
2656 if (rtrs_clt_change_state_get_old(clt_path, RTRS_CLT_CONNECTING, NULL)) {
2657 err = init_path(clt_path);
2658 if (err)
2659 goto reconnect_again;
2660 }
2661
2662 return;
2663
2664reconnect_again:
2665 if (rtrs_clt_change_state_get_old(clt_path, RTRS_CLT_RECONNECTING, NULL)) {
2666 clt_path->stats->reconnects.fail_cnt++;
2667 queue_work(rtrs_wq, &clt_path->err_recovery_work);
2668 }
2669}
2670
2671static void rtrs_clt_dev_release(struct device *dev)
2672{
2673 struct rtrs_clt_sess *clt = container_of(dev, struct rtrs_clt_sess,
2674 dev);
2675
2676 mutex_destroy(&clt->paths_ev_mutex);
2677 mutex_destroy(&clt->paths_mutex);
2678 kfree(clt);
2679}
2680
2681static struct rtrs_clt_sess *alloc_clt(const char *sessname, size_t paths_num,
2682 u16 port, size_t pdu_sz, void *priv,
2683 void (*link_ev)(void *priv,
2684 enum rtrs_clt_link_ev ev),
2685 unsigned int reconnect_delay_sec,
2686 unsigned int max_reconnect_attempts)
2687{
2688 struct rtrs_clt_sess *clt;
2689 int err;
2690
2691 if (!paths_num || paths_num > MAX_PATHS_NUM)
2692 return ERR_PTR(-EINVAL);
2693
2694 if (strlen(sessname) >= sizeof(clt->sessname))
2695 return ERR_PTR(-EINVAL);
2696
2697 clt = kzalloc(sizeof(*clt), GFP_KERNEL);
2698 if (!clt)
2699 return ERR_PTR(-ENOMEM);
2700
2701 clt->pcpu_path = alloc_percpu(typeof(*clt->pcpu_path));
2702 if (!clt->pcpu_path) {
2703 kfree(clt);
2704 return ERR_PTR(-ENOMEM);
2705 }
2706
2707 clt->dev.class = &rtrs_clt_dev_class;
2708 clt->dev.release = rtrs_clt_dev_release;
2709 uuid_gen(&clt->paths_uuid);
2710 INIT_LIST_HEAD_RCU(&clt->paths_list);
2711 clt->paths_num = paths_num;
2712 clt->paths_up = MAX_PATHS_NUM;
2713 clt->port = port;
2714 clt->pdu_sz = pdu_sz;
2715 clt->max_segments = RTRS_MAX_SEGMENTS;
2716 clt->reconnect_delay_sec = reconnect_delay_sec;
2717 clt->max_reconnect_attempts = max_reconnect_attempts;
2718 clt->priv = priv;
2719 clt->link_ev = link_ev;
2720 clt->mp_policy = MP_POLICY_MIN_INFLIGHT;
2721 strscpy(clt->sessname, sessname, sizeof(clt->sessname));
2722 init_waitqueue_head(&clt->permits_wait);
2723 mutex_init(&clt->paths_ev_mutex);
2724 mutex_init(&clt->paths_mutex);
2725 device_initialize(&clt->dev);
2726
2727 err = dev_set_name(&clt->dev, "%s", sessname);
2728 if (err)
2729 goto err_put;
2730
2731 /*
2732 * Suppress user space notification until
2733 * sysfs files are created
2734 */
2735 dev_set_uevent_suppress(&clt->dev, true);
2736 err = device_add(&clt->dev);
2737 if (err)
2738 goto err_put;
2739
2740 clt->kobj_paths = kobject_create_and_add("paths", &clt->dev.kobj);
2741 if (!clt->kobj_paths) {
2742 err = -ENOMEM;
2743 goto err_del;
2744 }
2745 err = rtrs_clt_create_sysfs_root_files(clt);
2746 if (err) {
2747 kobject_del(clt->kobj_paths);
2748 kobject_put(clt->kobj_paths);
2749 goto err_del;
2750 }
2751 dev_set_uevent_suppress(&clt->dev, false);
2752 kobject_uevent(&clt->dev.kobj, KOBJ_ADD);
2753
2754 return clt;
2755err_del:
2756 device_del(&clt->dev);
2757err_put:
2758 free_percpu(clt->pcpu_path);
2759 put_device(&clt->dev);
2760 return ERR_PTR(err);
2761}
2762
2763static void free_clt(struct rtrs_clt_sess *clt)
2764{
2765 free_percpu(clt->pcpu_path);
2766
2767 /*
2768 * release callback will free clt and destroy mutexes in last put
2769 */
2770 device_unregister(&clt->dev);
2771}
2772
2773/**
2774 * rtrs_clt_open() - Open a path to an RTRS server
2775 * @ops: holds the link event callback and the private pointer.
2776 * @pathname: name of the path to an RTRS server
2777 * @paths: Paths to be established defined by their src and dst addresses
2778 * @paths_num: Number of elements in the @paths array
2779 * @port: port to be used by the RTRS session
2780 * @pdu_sz: Size of extra payload which can be accessed after permit allocation.
2781 * @reconnect_delay_sec: time between reconnect tries
2782 * @max_reconnect_attempts: Number of times to reconnect on error before giving
2783 * up, 0 for * disabled, -1 for forever
2784 * @nr_poll_queues: number of polling mode connection using IB_POLL_DIRECT flag
2785 *
2786 * Starts session establishment with the rtrs_server. The function can block
2787 * up to ~2000ms before it returns.
2788 *
2789 * Return a valid pointer on success otherwise PTR_ERR.
2790 */
2791struct rtrs_clt_sess *rtrs_clt_open(struct rtrs_clt_ops *ops,
2792 const char *pathname,
2793 const struct rtrs_addr *paths,
2794 size_t paths_num, u16 port,
2795 size_t pdu_sz, u8 reconnect_delay_sec,
2796 s16 max_reconnect_attempts, u32 nr_poll_queues)
2797{
2798 struct rtrs_clt_path *clt_path, *tmp;
2799 struct rtrs_clt_sess *clt;
2800 int err, i;
2801
2802 if (strchr(pathname, '/') || strchr(pathname, '.')) {
2803 pr_err("pathname cannot contain / and .\n");
2804 err = -EINVAL;
2805 goto out;
2806 }
2807
2808 clt = alloc_clt(pathname, paths_num, port, pdu_sz, ops->priv,
2809 ops->link_ev,
2810 reconnect_delay_sec,
2811 max_reconnect_attempts);
2812 if (IS_ERR(clt)) {
2813 err = PTR_ERR(clt);
2814 goto out;
2815 }
2816 for (i = 0; i < paths_num; i++) {
2817 struct rtrs_clt_path *clt_path;
2818
2819 clt_path = alloc_path(clt, &paths[i], nr_cpu_ids,
2820 nr_poll_queues);
2821 if (IS_ERR(clt_path)) {
2822 err = PTR_ERR(clt_path);
2823 goto close_all_path;
2824 }
2825 if (!i)
2826 clt_path->for_new_clt = 1;
2827 list_add_tail_rcu(&clt_path->s.entry, &clt->paths_list);
2828
2829 err = init_path(clt_path);
2830 if (err) {
2831 list_del_rcu(&clt_path->s.entry);
2832 rtrs_clt_close_conns(clt_path, true);
2833 free_percpu(clt_path->stats->pcpu_stats);
2834 kfree(clt_path->stats);
2835 free_path(clt_path);
2836 goto close_all_path;
2837 }
2838
2839 err = rtrs_clt_create_path_files(clt_path);
2840 if (err) {
2841 list_del_rcu(&clt_path->s.entry);
2842 rtrs_clt_close_conns(clt_path, true);
2843 free_percpu(clt_path->stats->pcpu_stats);
2844 kfree(clt_path->stats);
2845 free_path(clt_path);
2846 goto close_all_path;
2847 }
2848 }
2849 err = alloc_permits(clt);
2850 if (err)
2851 goto close_all_path;
2852
2853 return clt;
2854
2855close_all_path:
2856 list_for_each_entry_safe(clt_path, tmp, &clt->paths_list, s.entry) {
2857 rtrs_clt_destroy_path_files(clt_path, NULL);
2858 rtrs_clt_close_conns(clt_path, true);
2859 kobject_put(&clt_path->kobj);
2860 }
2861 rtrs_clt_destroy_sysfs_root(clt);
2862 free_clt(clt);
2863
2864out:
2865 return ERR_PTR(err);
2866}
2867EXPORT_SYMBOL(rtrs_clt_open);
2868
2869/**
2870 * rtrs_clt_close() - Close a path
2871 * @clt: Session handle. Session is freed upon return.
2872 */
2873void rtrs_clt_close(struct rtrs_clt_sess *clt)
2874{
2875 struct rtrs_clt_path *clt_path, *tmp;
2876
2877 /* Firstly forbid sysfs access */
2878 rtrs_clt_destroy_sysfs_root(clt);
2879
2880 /* Now it is safe to iterate over all paths without locks */
2881 list_for_each_entry_safe(clt_path, tmp, &clt->paths_list, s.entry) {
2882 rtrs_clt_close_conns(clt_path, true);
2883 rtrs_clt_destroy_path_files(clt_path, NULL);
2884 kobject_put(&clt_path->kobj);
2885 }
2886 free_permits(clt);
2887 free_clt(clt);
2888}
2889EXPORT_SYMBOL(rtrs_clt_close);
2890
2891int rtrs_clt_reconnect_from_sysfs(struct rtrs_clt_path *clt_path)
2892{
2893 enum rtrs_clt_state old_state;
2894 int err = -EBUSY;
2895 bool changed;
2896
2897 changed = rtrs_clt_change_state_get_old(clt_path,
2898 RTRS_CLT_RECONNECTING,
2899 &old_state);
2900 if (changed) {
2901 clt_path->reconnect_attempts = 0;
2902 rtrs_clt_stop_and_destroy_conns(clt_path);
2903 queue_delayed_work(rtrs_wq, &clt_path->reconnect_dwork, 0);
2904 }
2905 if (changed || old_state == RTRS_CLT_RECONNECTING) {
2906 /*
2907 * flush_delayed_work() queues pending work for immediate
2908 * execution, so do the flush if we have queued something
2909 * right now or work is pending.
2910 */
2911 flush_delayed_work(&clt_path->reconnect_dwork);
2912 err = (READ_ONCE(clt_path->state) ==
2913 RTRS_CLT_CONNECTED ? 0 : -ENOTCONN);
2914 }
2915
2916 return err;
2917}
2918
2919int rtrs_clt_remove_path_from_sysfs(struct rtrs_clt_path *clt_path,
2920 const struct attribute *sysfs_self)
2921{
2922 enum rtrs_clt_state old_state;
2923 bool changed;
2924
2925 /*
2926 * Continue stopping path till state was changed to DEAD or
2927 * state was observed as DEAD:
2928 * 1. State was changed to DEAD - we were fast and nobody
2929 * invoked rtrs_clt_reconnect(), which can again start
2930 * reconnecting.
2931 * 2. State was observed as DEAD - we have someone in parallel
2932 * removing the path.
2933 */
2934 do {
2935 rtrs_clt_close_conns(clt_path, true);
2936 changed = rtrs_clt_change_state_get_old(clt_path,
2937 RTRS_CLT_DEAD,
2938 &old_state);
2939 } while (!changed && old_state != RTRS_CLT_DEAD);
2940
2941 if (changed) {
2942 rtrs_clt_remove_path_from_arr(clt_path);
2943 rtrs_clt_destroy_path_files(clt_path, sysfs_self);
2944 kobject_put(&clt_path->kobj);
2945 }
2946
2947 return 0;
2948}
2949
2950void rtrs_clt_set_max_reconnect_attempts(struct rtrs_clt_sess *clt, int value)
2951{
2952 clt->max_reconnect_attempts = (unsigned int)value;
2953}
2954
2955int rtrs_clt_get_max_reconnect_attempts(const struct rtrs_clt_sess *clt)
2956{
2957 return (int)clt->max_reconnect_attempts;
2958}
2959
2960/**
2961 * rtrs_clt_request() - Request data transfer to/from server via RDMA.
2962 *
2963 * @dir: READ/WRITE
2964 * @ops: callback function to be called as confirmation, and the pointer.
2965 * @clt: Session
2966 * @permit: Preallocated permit
2967 * @vec: Message that is sent to server together with the request.
2968 * Sum of len of all @vec elements limited to <= IO_MSG_SIZE.
2969 * Since the msg is copied internally it can be allocated on stack.
2970 * @nr: Number of elements in @vec.
2971 * @data_len: length of data sent to/from server
2972 * @sg: Pages to be sent/received to/from server.
2973 * @sg_cnt: Number of elements in the @sg
2974 *
2975 * Return:
2976 * 0: Success
2977 * <0: Error
2978 *
2979 * On dir=READ rtrs client will request a data transfer from Server to client.
2980 * The data that the server will respond with will be stored in @sg when
2981 * the user receives an %RTRS_CLT_RDMA_EV_RDMA_REQUEST_WRITE_COMPL event.
2982 * On dir=WRITE rtrs client will rdma write data in sg to server side.
2983 */
2984int rtrs_clt_request(int dir, struct rtrs_clt_req_ops *ops,
2985 struct rtrs_clt_sess *clt, struct rtrs_permit *permit,
2986 const struct kvec *vec, size_t nr, size_t data_len,
2987 struct scatterlist *sg, unsigned int sg_cnt)
2988{
2989 struct rtrs_clt_io_req *req;
2990 struct rtrs_clt_path *clt_path;
2991
2992 enum dma_data_direction dma_dir;
2993 int err = -ECONNABORTED, i;
2994 size_t usr_len, hdr_len;
2995 struct path_it it;
2996
2997 /* Get kvec length */
2998 for (i = 0, usr_len = 0; i < nr; i++)
2999 usr_len += vec[i].iov_len;
3000
3001 if (dir == READ) {
3002 hdr_len = sizeof(struct rtrs_msg_rdma_read) +
3003 sg_cnt * sizeof(struct rtrs_sg_desc);
3004 dma_dir = DMA_FROM_DEVICE;
3005 } else {
3006 hdr_len = sizeof(struct rtrs_msg_rdma_write);
3007 dma_dir = DMA_TO_DEVICE;
3008 }
3009
3010 rcu_read_lock();
3011 for (path_it_init(&it, clt);
3012 (clt_path = it.next_path(&it)) && it.i < it.clt->paths_num; it.i++) {
3013 if (READ_ONCE(clt_path->state) != RTRS_CLT_CONNECTED)
3014 continue;
3015
3016 if (usr_len + hdr_len > clt_path->max_hdr_size) {
3017 rtrs_wrn_rl(clt_path->clt,
3018 "%s request failed, user message size is %zu and header length %zu, but max size is %u\n",
3019 dir == READ ? "Read" : "Write",
3020 usr_len, hdr_len, clt_path->max_hdr_size);
3021 err = -EMSGSIZE;
3022 break;
3023 }
3024 req = rtrs_clt_get_req(clt_path, ops->conf_fn, permit, ops->priv,
3025 vec, usr_len, sg, sg_cnt, data_len,
3026 dma_dir);
3027 if (dir == READ)
3028 err = rtrs_clt_read_req(req);
3029 else
3030 err = rtrs_clt_write_req(req);
3031 if (err) {
3032 req->in_use = false;
3033 continue;
3034 }
3035 /* Success path */
3036 break;
3037 }
3038 path_it_deinit(&it);
3039 rcu_read_unlock();
3040
3041 return err;
3042}
3043EXPORT_SYMBOL(rtrs_clt_request);
3044
3045int rtrs_clt_rdma_cq_direct(struct rtrs_clt_sess *clt, unsigned int index)
3046{
3047 /* If no path, return -1 for block layer not to try again */
3048 int cnt = -1;
3049 struct rtrs_con *con;
3050 struct rtrs_clt_path *clt_path;
3051 struct path_it it;
3052
3053 rcu_read_lock();
3054 for (path_it_init(&it, clt);
3055 (clt_path = it.next_path(&it)) && it.i < it.clt->paths_num; it.i++) {
3056 if (READ_ONCE(clt_path->state) != RTRS_CLT_CONNECTED)
3057 continue;
3058
3059 con = clt_path->s.con[index + 1];
3060 cnt = ib_process_cq_direct(con->cq, -1);
3061 if (cnt)
3062 break;
3063 }
3064 path_it_deinit(&it);
3065 rcu_read_unlock();
3066
3067 return cnt;
3068}
3069EXPORT_SYMBOL(rtrs_clt_rdma_cq_direct);
3070
3071/**
3072 * rtrs_clt_query() - queries RTRS session attributes
3073 *@clt: session pointer
3074 *@attr: query results for session attributes.
3075 * Returns:
3076 * 0 on success
3077 * -ECOMM no connection to the server
3078 */
3079int rtrs_clt_query(struct rtrs_clt_sess *clt, struct rtrs_attrs *attr)
3080{
3081 if (!rtrs_clt_is_connected(clt))
3082 return -ECOMM;
3083
3084 attr->queue_depth = clt->queue_depth;
3085 attr->max_segments = clt->max_segments;
3086 /* Cap max_io_size to min of remote buffer size and the fr pages */
3087 attr->max_io_size = min_t(int, clt->max_io_size,
3088 clt->max_segments * SZ_4K);
3089
3090 return 0;
3091}
3092EXPORT_SYMBOL(rtrs_clt_query);
3093
3094int rtrs_clt_create_path_from_sysfs(struct rtrs_clt_sess *clt,
3095 struct rtrs_addr *addr)
3096{
3097 struct rtrs_clt_path *clt_path;
3098 int err;
3099
3100 clt_path = alloc_path(clt, addr, nr_cpu_ids, 0);
3101 if (IS_ERR(clt_path))
3102 return PTR_ERR(clt_path);
3103
3104 mutex_lock(&clt->paths_mutex);
3105 if (clt->paths_num == 0) {
3106 /*
3107 * When all the paths are removed for a session,
3108 * the addition of the first path is like a new session for
3109 * the storage server
3110 */
3111 clt_path->for_new_clt = 1;
3112 }
3113
3114 mutex_unlock(&clt->paths_mutex);
3115
3116 /*
3117 * It is totally safe to add path in CONNECTING state: coming
3118 * IO will never grab it. Also it is very important to add
3119 * path before init, since init fires LINK_CONNECTED event.
3120 */
3121 rtrs_clt_add_path_to_arr(clt_path);
3122
3123 err = init_path(clt_path);
3124 if (err)
3125 goto close_path;
3126
3127 err = rtrs_clt_create_path_files(clt_path);
3128 if (err)
3129 goto close_path;
3130
3131 return 0;
3132
3133close_path:
3134 rtrs_clt_remove_path_from_arr(clt_path);
3135 rtrs_clt_close_conns(clt_path, true);
3136 free_percpu(clt_path->stats->pcpu_stats);
3137 kfree(clt_path->stats);
3138 free_path(clt_path);
3139
3140 return err;
3141}
3142
3143static int rtrs_clt_ib_dev_init(struct rtrs_ib_dev *dev)
3144{
3145 if (!(dev->ib_dev->attrs.device_cap_flags &
3146 IB_DEVICE_MEM_MGT_EXTENSIONS)) {
3147 pr_err("Memory registrations not supported.\n");
3148 return -ENOTSUPP;
3149 }
3150
3151 return 0;
3152}
3153
3154static const struct rtrs_rdma_dev_pd_ops dev_pd_ops = {
3155 .init = rtrs_clt_ib_dev_init
3156};
3157
3158static int __init rtrs_client_init(void)
3159{
3160 int ret = 0;
3161
3162 rtrs_rdma_dev_pd_init(0, &dev_pd);
3163 ret = class_register(&rtrs_clt_dev_class);
3164 if (ret) {
3165 pr_err("Failed to create rtrs-client dev class\n");
3166 return ret;
3167 }
3168 rtrs_wq = alloc_workqueue("rtrs_client_wq", 0, 0);
3169 if (!rtrs_wq) {
3170 class_unregister(&rtrs_clt_dev_class);
3171 return -ENOMEM;
3172 }
3173
3174 return 0;
3175}
3176
3177static void __exit rtrs_client_exit(void)
3178{
3179 destroy_workqueue(rtrs_wq);
3180 class_unregister(&rtrs_clt_dev_class);
3181 rtrs_rdma_dev_pd_deinit(&dev_pd);
3182}
3183
3184module_init(rtrs_client_init);
3185module_exit(rtrs_client_exit);
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#include "rtrs-clt-trace.h"
20
21#define RTRS_CONNECT_TIMEOUT_MS 30000
22/*
23 * Wait a bit before trying to reconnect after a failure
24 * in order to give server time to finish clean up which
25 * leads to "false positives" failed reconnect attempts
26 */
27#define RTRS_RECONNECT_BACKOFF 1000
28/*
29 * Wait for additional random time between 0 and 8 seconds
30 * before starting to reconnect to avoid clients reconnecting
31 * all at once in case of a major network outage
32 */
33#define RTRS_RECONNECT_SEED 8
34
35#define FIRST_CONN 0x01
36/* limit to 128 * 4k = 512k max IO */
37#define RTRS_MAX_SEGMENTS 128
38
39MODULE_DESCRIPTION("RDMA Transport Client");
40MODULE_LICENSE("GPL");
41
42static const struct rtrs_rdma_dev_pd_ops dev_pd_ops;
43static struct rtrs_rdma_dev_pd dev_pd = {
44 .ops = &dev_pd_ops
45};
46
47static struct workqueue_struct *rtrs_wq;
48static struct class *rtrs_clt_dev_class;
49
50static inline bool rtrs_clt_is_connected(const struct rtrs_clt_sess *clt)
51{
52 struct rtrs_clt_path *clt_path;
53 bool connected = false;
54
55 rcu_read_lock();
56 list_for_each_entry_rcu(clt_path, &clt->paths_list, s.entry)
57 if (READ_ONCE(clt_path->state) == RTRS_CLT_CONNECTED) {
58 connected = true;
59 break;
60 }
61 rcu_read_unlock();
62
63 return connected;
64}
65
66static struct rtrs_permit *
67__rtrs_get_permit(struct rtrs_clt_sess *clt, enum rtrs_clt_con_type con_type)
68{
69 size_t max_depth = clt->queue_depth;
70 struct rtrs_permit *permit;
71 int bit;
72
73 /*
74 * Adapted from null_blk get_tag(). Callers from different cpus may
75 * grab the same bit, since find_first_zero_bit is not atomic.
76 * But then the test_and_set_bit_lock will fail for all the
77 * callers but one, so that they will loop again.
78 * This way an explicit spinlock is not required.
79 */
80 do {
81 bit = find_first_zero_bit(clt->permits_map, max_depth);
82 if (bit >= max_depth)
83 return NULL;
84 } while (test_and_set_bit_lock(bit, clt->permits_map));
85
86 permit = get_permit(clt, bit);
87 WARN_ON(permit->mem_id != bit);
88 permit->cpu_id = raw_smp_processor_id();
89 permit->con_type = con_type;
90
91 return permit;
92}
93
94static inline void __rtrs_put_permit(struct rtrs_clt_sess *clt,
95 struct rtrs_permit *permit)
96{
97 clear_bit_unlock(permit->mem_id, clt->permits_map);
98}
99
100/**
101 * rtrs_clt_get_permit() - allocates permit for future RDMA operation
102 * @clt: Current session
103 * @con_type: Type of connection to use with the permit
104 * @can_wait: Wait type
105 *
106 * Description:
107 * Allocates permit for the following RDMA operation. Permit is used
108 * to preallocate all resources and to propagate memory pressure
109 * up earlier.
110 *
111 * Context:
112 * Can sleep if @wait == RTRS_PERMIT_WAIT
113 */
114struct rtrs_permit *rtrs_clt_get_permit(struct rtrs_clt_sess *clt,
115 enum rtrs_clt_con_type con_type,
116 enum wait_type can_wait)
117{
118 struct rtrs_permit *permit;
119 DEFINE_WAIT(wait);
120
121 permit = __rtrs_get_permit(clt, con_type);
122 if (permit || !can_wait)
123 return permit;
124
125 do {
126 prepare_to_wait(&clt->permits_wait, &wait,
127 TASK_UNINTERRUPTIBLE);
128 permit = __rtrs_get_permit(clt, con_type);
129 if (permit)
130 break;
131
132 io_schedule();
133 } while (1);
134
135 finish_wait(&clt->permits_wait, &wait);
136
137 return permit;
138}
139EXPORT_SYMBOL(rtrs_clt_get_permit);
140
141/**
142 * rtrs_clt_put_permit() - puts allocated permit
143 * @clt: Current session
144 * @permit: Permit to be freed
145 *
146 * Context:
147 * Does not matter
148 */
149void rtrs_clt_put_permit(struct rtrs_clt_sess *clt,
150 struct rtrs_permit *permit)
151{
152 if (WARN_ON(!test_bit(permit->mem_id, clt->permits_map)))
153 return;
154
155 __rtrs_put_permit(clt, permit);
156
157 /*
158 * rtrs_clt_get_permit() adds itself to the &clt->permits_wait list
159 * before calling schedule(). So if rtrs_clt_get_permit() is sleeping
160 * it must have added itself to &clt->permits_wait before
161 * __rtrs_put_permit() finished.
162 * Hence it is safe to guard wake_up() with a waitqueue_active() test.
163 */
164 if (waitqueue_active(&clt->permits_wait))
165 wake_up(&clt->permits_wait);
166}
167EXPORT_SYMBOL(rtrs_clt_put_permit);
168
169/**
170 * rtrs_permit_to_clt_con() - returns RDMA connection pointer by the permit
171 * @clt_path: client path pointer
172 * @permit: permit for the allocation of the RDMA buffer
173 * Note:
174 * IO connection starts from 1.
175 * 0 connection is for user messages.
176 */
177static
178struct rtrs_clt_con *rtrs_permit_to_clt_con(struct rtrs_clt_path *clt_path,
179 struct rtrs_permit *permit)
180{
181 int id = 0;
182
183 if (permit->con_type == RTRS_IO_CON)
184 id = (permit->cpu_id % (clt_path->s.irq_con_num - 1)) + 1;
185
186 return to_clt_con(clt_path->s.con[id]);
187}
188
189/**
190 * rtrs_clt_change_state() - change the session state through session state
191 * machine.
192 *
193 * @clt_path: client path to change the state of.
194 * @new_state: state to change to.
195 *
196 * returns true if sess's state is changed to new state, otherwise return false.
197 *
198 * Locks:
199 * state_wq lock must be hold.
200 */
201static bool rtrs_clt_change_state(struct rtrs_clt_path *clt_path,
202 enum rtrs_clt_state new_state)
203{
204 enum rtrs_clt_state old_state;
205 bool changed = false;
206
207 lockdep_assert_held(&clt_path->state_wq.lock);
208
209 old_state = clt_path->state;
210 switch (new_state) {
211 case RTRS_CLT_CONNECTING:
212 switch (old_state) {
213 case RTRS_CLT_RECONNECTING:
214 changed = true;
215 fallthrough;
216 default:
217 break;
218 }
219 break;
220 case RTRS_CLT_RECONNECTING:
221 switch (old_state) {
222 case RTRS_CLT_CONNECTED:
223 case RTRS_CLT_CONNECTING_ERR:
224 case RTRS_CLT_CLOSED:
225 changed = true;
226 fallthrough;
227 default:
228 break;
229 }
230 break;
231 case RTRS_CLT_CONNECTED:
232 switch (old_state) {
233 case RTRS_CLT_CONNECTING:
234 changed = true;
235 fallthrough;
236 default:
237 break;
238 }
239 break;
240 case RTRS_CLT_CONNECTING_ERR:
241 switch (old_state) {
242 case RTRS_CLT_CONNECTING:
243 changed = true;
244 fallthrough;
245 default:
246 break;
247 }
248 break;
249 case RTRS_CLT_CLOSING:
250 switch (old_state) {
251 case RTRS_CLT_CONNECTING:
252 case RTRS_CLT_CONNECTING_ERR:
253 case RTRS_CLT_RECONNECTING:
254 case RTRS_CLT_CONNECTED:
255 changed = true;
256 fallthrough;
257 default:
258 break;
259 }
260 break;
261 case RTRS_CLT_CLOSED:
262 switch (old_state) {
263 case RTRS_CLT_CLOSING:
264 changed = true;
265 fallthrough;
266 default:
267 break;
268 }
269 break;
270 case RTRS_CLT_DEAD:
271 switch (old_state) {
272 case RTRS_CLT_CLOSED:
273 changed = true;
274 fallthrough;
275 default:
276 break;
277 }
278 break;
279 default:
280 break;
281 }
282 if (changed) {
283 clt_path->state = new_state;
284 wake_up_locked(&clt_path->state_wq);
285 }
286
287 return changed;
288}
289
290static bool rtrs_clt_change_state_from_to(struct rtrs_clt_path *clt_path,
291 enum rtrs_clt_state old_state,
292 enum rtrs_clt_state new_state)
293{
294 bool changed = false;
295
296 spin_lock_irq(&clt_path->state_wq.lock);
297 if (clt_path->state == old_state)
298 changed = rtrs_clt_change_state(clt_path, new_state);
299 spin_unlock_irq(&clt_path->state_wq.lock);
300
301 return changed;
302}
303
304static void rtrs_clt_stop_and_destroy_conns(struct rtrs_clt_path *clt_path);
305static void rtrs_rdma_error_recovery(struct rtrs_clt_con *con)
306{
307 struct rtrs_clt_path *clt_path = to_clt_path(con->c.path);
308
309 trace_rtrs_rdma_error_recovery(clt_path);
310
311 if (rtrs_clt_change_state_from_to(clt_path,
312 RTRS_CLT_CONNECTED,
313 RTRS_CLT_RECONNECTING)) {
314 queue_work(rtrs_wq, &clt_path->err_recovery_work);
315 } else {
316 /*
317 * Error can happen just on establishing new connection,
318 * so notify waiter with error state, waiter is responsible
319 * for cleaning the rest and reconnect if needed.
320 */
321 rtrs_clt_change_state_from_to(clt_path,
322 RTRS_CLT_CONNECTING,
323 RTRS_CLT_CONNECTING_ERR);
324 }
325}
326
327static void rtrs_clt_fast_reg_done(struct ib_cq *cq, struct ib_wc *wc)
328{
329 struct rtrs_clt_con *con = to_clt_con(wc->qp->qp_context);
330
331 if (wc->status != IB_WC_SUCCESS) {
332 rtrs_err(con->c.path, "Failed IB_WR_REG_MR: %s\n",
333 ib_wc_status_msg(wc->status));
334 rtrs_rdma_error_recovery(con);
335 }
336}
337
338static struct ib_cqe fast_reg_cqe = {
339 .done = rtrs_clt_fast_reg_done
340};
341
342static void complete_rdma_req(struct rtrs_clt_io_req *req, int errno,
343 bool notify, bool can_wait);
344
345static void rtrs_clt_inv_rkey_done(struct ib_cq *cq, struct ib_wc *wc)
346{
347 struct rtrs_clt_io_req *req =
348 container_of(wc->wr_cqe, typeof(*req), inv_cqe);
349 struct rtrs_clt_con *con = to_clt_con(wc->qp->qp_context);
350
351 if (wc->status != IB_WC_SUCCESS) {
352 rtrs_err(con->c.path, "Failed IB_WR_LOCAL_INV: %s\n",
353 ib_wc_status_msg(wc->status));
354 rtrs_rdma_error_recovery(con);
355 }
356 req->need_inv = false;
357 if (req->need_inv_comp)
358 complete(&req->inv_comp);
359 else
360 /* Complete request from INV callback */
361 complete_rdma_req(req, req->inv_errno, true, false);
362}
363
364static int rtrs_inv_rkey(struct rtrs_clt_io_req *req)
365{
366 struct rtrs_clt_con *con = req->con;
367 struct ib_send_wr wr = {
368 .opcode = IB_WR_LOCAL_INV,
369 .wr_cqe = &req->inv_cqe,
370 .send_flags = IB_SEND_SIGNALED,
371 .ex.invalidate_rkey = req->mr->rkey,
372 };
373 req->inv_cqe.done = rtrs_clt_inv_rkey_done;
374
375 return ib_post_send(con->c.qp, &wr, NULL);
376}
377
378static void complete_rdma_req(struct rtrs_clt_io_req *req, int errno,
379 bool notify, bool can_wait)
380{
381 struct rtrs_clt_con *con = req->con;
382 struct rtrs_clt_path *clt_path;
383 int err;
384
385 if (WARN_ON(!req->in_use))
386 return;
387 if (WARN_ON(!req->con))
388 return;
389 clt_path = to_clt_path(con->c.path);
390
391 if (req->sg_cnt) {
392 if (req->dir == DMA_FROM_DEVICE && req->need_inv) {
393 /*
394 * We are here to invalidate read requests
395 * ourselves. In normal scenario server should
396 * send INV for all read requests, but
397 * we are here, thus two things could happen:
398 *
399 * 1. this is failover, when errno != 0
400 * and can_wait == 1,
401 *
402 * 2. something totally bad happened and
403 * server forgot to send INV, so we
404 * should do that ourselves.
405 */
406
407 if (can_wait) {
408 req->need_inv_comp = true;
409 } else {
410 /* This should be IO path, so always notify */
411 WARN_ON(!notify);
412 /* Save errno for INV callback */
413 req->inv_errno = errno;
414 }
415
416 refcount_inc(&req->ref);
417 err = rtrs_inv_rkey(req);
418 if (err) {
419 rtrs_err(con->c.path, "Send INV WR key=%#x: %d\n",
420 req->mr->rkey, err);
421 } else if (can_wait) {
422 wait_for_completion(&req->inv_comp);
423 } else {
424 /*
425 * Something went wrong, so request will be
426 * completed from INV callback.
427 */
428 WARN_ON_ONCE(1);
429
430 return;
431 }
432 if (!refcount_dec_and_test(&req->ref))
433 return;
434 }
435 ib_dma_unmap_sg(clt_path->s.dev->ib_dev, req->sglist,
436 req->sg_cnt, req->dir);
437 }
438 if (!refcount_dec_and_test(&req->ref))
439 return;
440 if (req->mp_policy == MP_POLICY_MIN_INFLIGHT)
441 atomic_dec(&clt_path->stats->inflight);
442
443 req->in_use = false;
444 req->con = NULL;
445
446 if (errno) {
447 rtrs_err_rl(con->c.path, "IO request failed: error=%d path=%s [%s:%u] notify=%d\n",
448 errno, kobject_name(&clt_path->kobj), clt_path->hca_name,
449 clt_path->hca_port, notify);
450 }
451
452 if (notify)
453 req->conf(req->priv, errno);
454}
455
456static int rtrs_post_send_rdma(struct rtrs_clt_con *con,
457 struct rtrs_clt_io_req *req,
458 struct rtrs_rbuf *rbuf, u32 off,
459 u32 imm, struct ib_send_wr *wr)
460{
461 struct rtrs_clt_path *clt_path = to_clt_path(con->c.path);
462 enum ib_send_flags flags;
463 struct ib_sge sge;
464
465 if (!req->sg_size) {
466 rtrs_wrn(con->c.path,
467 "Doing RDMA Write failed, no data supplied\n");
468 return -EINVAL;
469 }
470
471 /* user data and user message in the first list element */
472 sge.addr = req->iu->dma_addr;
473 sge.length = req->sg_size;
474 sge.lkey = clt_path->s.dev->ib_pd->local_dma_lkey;
475
476 /*
477 * From time to time we have to post signalled sends,
478 * or send queue will fill up and only QP reset can help.
479 */
480 flags = atomic_inc_return(&con->c.wr_cnt) % clt_path->s.signal_interval ?
481 0 : IB_SEND_SIGNALED;
482
483 ib_dma_sync_single_for_device(clt_path->s.dev->ib_dev,
484 req->iu->dma_addr,
485 req->sg_size, DMA_TO_DEVICE);
486
487 return rtrs_iu_post_rdma_write_imm(&con->c, req->iu, &sge, 1,
488 rbuf->rkey, rbuf->addr + off,
489 imm, flags, wr, NULL);
490}
491
492static void process_io_rsp(struct rtrs_clt_path *clt_path, u32 msg_id,
493 s16 errno, bool w_inval)
494{
495 struct rtrs_clt_io_req *req;
496
497 if (WARN_ON(msg_id >= clt_path->queue_depth))
498 return;
499
500 req = &clt_path->reqs[msg_id];
501 /* Drop need_inv if server responded with send with invalidation */
502 req->need_inv &= !w_inval;
503 complete_rdma_req(req, errno, true, false);
504}
505
506static void rtrs_clt_recv_done(struct rtrs_clt_con *con, struct ib_wc *wc)
507{
508 struct rtrs_iu *iu;
509 int err;
510 struct rtrs_clt_path *clt_path = to_clt_path(con->c.path);
511
512 WARN_ON((clt_path->flags & RTRS_MSG_NEW_RKEY_F) == 0);
513 iu = container_of(wc->wr_cqe, struct rtrs_iu,
514 cqe);
515 err = rtrs_iu_post_recv(&con->c, iu);
516 if (err) {
517 rtrs_err(con->c.path, "post iu failed %d\n", err);
518 rtrs_rdma_error_recovery(con);
519 }
520}
521
522static void rtrs_clt_rkey_rsp_done(struct rtrs_clt_con *con, struct ib_wc *wc)
523{
524 struct rtrs_clt_path *clt_path = to_clt_path(con->c.path);
525 struct rtrs_msg_rkey_rsp *msg;
526 u32 imm_type, imm_payload;
527 bool w_inval = false;
528 struct rtrs_iu *iu;
529 u32 buf_id;
530 int err;
531
532 WARN_ON((clt_path->flags & RTRS_MSG_NEW_RKEY_F) == 0);
533
534 iu = container_of(wc->wr_cqe, struct rtrs_iu, cqe);
535
536 if (wc->byte_len < sizeof(*msg)) {
537 rtrs_err(con->c.path, "rkey response is malformed: size %d\n",
538 wc->byte_len);
539 goto out;
540 }
541 ib_dma_sync_single_for_cpu(clt_path->s.dev->ib_dev, iu->dma_addr,
542 iu->size, DMA_FROM_DEVICE);
543 msg = iu->buf;
544 if (le16_to_cpu(msg->type) != RTRS_MSG_RKEY_RSP) {
545 rtrs_err(clt_path->clt,
546 "rkey response is malformed: type %d\n",
547 le16_to_cpu(msg->type));
548 goto out;
549 }
550 buf_id = le16_to_cpu(msg->buf_id);
551 if (WARN_ON(buf_id >= clt_path->queue_depth))
552 goto out;
553
554 rtrs_from_imm(be32_to_cpu(wc->ex.imm_data), &imm_type, &imm_payload);
555 if (imm_type == RTRS_IO_RSP_IMM ||
556 imm_type == RTRS_IO_RSP_W_INV_IMM) {
557 u32 msg_id;
558
559 w_inval = (imm_type == RTRS_IO_RSP_W_INV_IMM);
560 rtrs_from_io_rsp_imm(imm_payload, &msg_id, &err);
561
562 if (WARN_ON(buf_id != msg_id))
563 goto out;
564 clt_path->rbufs[buf_id].rkey = le32_to_cpu(msg->rkey);
565 process_io_rsp(clt_path, msg_id, err, w_inval);
566 }
567 ib_dma_sync_single_for_device(clt_path->s.dev->ib_dev, iu->dma_addr,
568 iu->size, DMA_FROM_DEVICE);
569 return rtrs_clt_recv_done(con, wc);
570out:
571 rtrs_rdma_error_recovery(con);
572}
573
574static void rtrs_clt_rdma_done(struct ib_cq *cq, struct ib_wc *wc);
575
576static struct ib_cqe io_comp_cqe = {
577 .done = rtrs_clt_rdma_done
578};
579
580/*
581 * Post x2 empty WRs: first is for this RDMA with IMM,
582 * second is for RECV with INV, which happened earlier.
583 */
584static int rtrs_post_recv_empty_x2(struct rtrs_con *con, struct ib_cqe *cqe)
585{
586 struct ib_recv_wr wr_arr[2], *wr;
587 int i;
588
589 memset(wr_arr, 0, sizeof(wr_arr));
590 for (i = 0; i < ARRAY_SIZE(wr_arr); i++) {
591 wr = &wr_arr[i];
592 wr->wr_cqe = cqe;
593 if (i)
594 /* Chain backwards */
595 wr->next = &wr_arr[i - 1];
596 }
597
598 return ib_post_recv(con->qp, wr, NULL);
599}
600
601static void rtrs_clt_rdma_done(struct ib_cq *cq, struct ib_wc *wc)
602{
603 struct rtrs_clt_con *con = to_clt_con(wc->qp->qp_context);
604 struct rtrs_clt_path *clt_path = to_clt_path(con->c.path);
605 u32 imm_type, imm_payload;
606 bool w_inval = false;
607 int err;
608
609 if (wc->status != IB_WC_SUCCESS) {
610 if (wc->status != IB_WC_WR_FLUSH_ERR) {
611 rtrs_err(clt_path->clt, "RDMA failed: %s\n",
612 ib_wc_status_msg(wc->status));
613 rtrs_rdma_error_recovery(con);
614 }
615 return;
616 }
617 rtrs_clt_update_wc_stats(con);
618
619 switch (wc->opcode) {
620 case IB_WC_RECV_RDMA_WITH_IMM:
621 /*
622 * post_recv() RDMA write completions of IO reqs (read/write)
623 * and hb
624 */
625 if (WARN_ON(wc->wr_cqe->done != rtrs_clt_rdma_done))
626 return;
627 rtrs_from_imm(be32_to_cpu(wc->ex.imm_data),
628 &imm_type, &imm_payload);
629 if (imm_type == RTRS_IO_RSP_IMM ||
630 imm_type == RTRS_IO_RSP_W_INV_IMM) {
631 u32 msg_id;
632
633 w_inval = (imm_type == RTRS_IO_RSP_W_INV_IMM);
634 rtrs_from_io_rsp_imm(imm_payload, &msg_id, &err);
635
636 process_io_rsp(clt_path, msg_id, err, w_inval);
637 } else if (imm_type == RTRS_HB_MSG_IMM) {
638 WARN_ON(con->c.cid);
639 rtrs_send_hb_ack(&clt_path->s);
640 if (clt_path->flags & RTRS_MSG_NEW_RKEY_F)
641 return rtrs_clt_recv_done(con, wc);
642 } else if (imm_type == RTRS_HB_ACK_IMM) {
643 WARN_ON(con->c.cid);
644 clt_path->s.hb_missed_cnt = 0;
645 clt_path->s.hb_cur_latency =
646 ktime_sub(ktime_get(), clt_path->s.hb_last_sent);
647 if (clt_path->flags & RTRS_MSG_NEW_RKEY_F)
648 return rtrs_clt_recv_done(con, wc);
649 } else {
650 rtrs_wrn(con->c.path, "Unknown IMM type %u\n",
651 imm_type);
652 }
653 if (w_inval)
654 /*
655 * Post x2 empty WRs: first is for this RDMA with IMM,
656 * second is for RECV with INV, which happened earlier.
657 */
658 err = rtrs_post_recv_empty_x2(&con->c, &io_comp_cqe);
659 else
660 err = rtrs_post_recv_empty(&con->c, &io_comp_cqe);
661 if (err) {
662 rtrs_err(con->c.path, "rtrs_post_recv_empty(): %d\n",
663 err);
664 rtrs_rdma_error_recovery(con);
665 }
666 break;
667 case IB_WC_RECV:
668 /*
669 * Key invalidations from server side
670 */
671 WARN_ON(!(wc->wc_flags & IB_WC_WITH_INVALIDATE ||
672 wc->wc_flags & IB_WC_WITH_IMM));
673 WARN_ON(wc->wr_cqe->done != rtrs_clt_rdma_done);
674 if (clt_path->flags & RTRS_MSG_NEW_RKEY_F) {
675 if (wc->wc_flags & IB_WC_WITH_INVALIDATE)
676 return rtrs_clt_recv_done(con, wc);
677
678 return rtrs_clt_rkey_rsp_done(con, wc);
679 }
680 break;
681 case IB_WC_RDMA_WRITE:
682 /*
683 * post_send() RDMA write completions of IO reqs (read/write)
684 * and hb.
685 */
686 break;
687
688 default:
689 rtrs_wrn(clt_path->clt, "Unexpected WC type: %d\n", wc->opcode);
690 return;
691 }
692}
693
694static int post_recv_io(struct rtrs_clt_con *con, size_t q_size)
695{
696 int err, i;
697 struct rtrs_clt_path *clt_path = to_clt_path(con->c.path);
698
699 for (i = 0; i < q_size; i++) {
700 if (clt_path->flags & RTRS_MSG_NEW_RKEY_F) {
701 struct rtrs_iu *iu = &con->rsp_ius[i];
702
703 err = rtrs_iu_post_recv(&con->c, iu);
704 } else {
705 err = rtrs_post_recv_empty(&con->c, &io_comp_cqe);
706 }
707 if (err)
708 return err;
709 }
710
711 return 0;
712}
713
714static int post_recv_path(struct rtrs_clt_path *clt_path)
715{
716 size_t q_size = 0;
717 int err, cid;
718
719 for (cid = 0; cid < clt_path->s.con_num; cid++) {
720 if (cid == 0)
721 q_size = SERVICE_CON_QUEUE_DEPTH;
722 else
723 q_size = clt_path->queue_depth;
724
725 /*
726 * x2 for RDMA read responses + FR key invalidations,
727 * RDMA writes do not require any FR registrations.
728 */
729 q_size *= 2;
730
731 err = post_recv_io(to_clt_con(clt_path->s.con[cid]), q_size);
732 if (err) {
733 rtrs_err(clt_path->clt, "post_recv_io(), err: %d\n",
734 err);
735 return err;
736 }
737 }
738
739 return 0;
740}
741
742struct path_it {
743 int i;
744 struct list_head skip_list;
745 struct rtrs_clt_sess *clt;
746 struct rtrs_clt_path *(*next_path)(struct path_it *it);
747};
748
749/*
750 * rtrs_clt_get_next_path_or_null - get clt path from the list or return NULL
751 * @head: the head for the list.
752 * @clt_path: The element to take the next clt_path from.
753 *
754 * Next clt path returned in round-robin fashion, i.e. head will be skipped,
755 * but if list is observed as empty, NULL will be returned.
756 *
757 * This function may safely run concurrently with the _rcu list-mutation
758 * primitives such as list_add_rcu() as long as it's guarded by rcu_read_lock().
759 */
760static inline struct rtrs_clt_path *
761rtrs_clt_get_next_path_or_null(struct list_head *head, struct rtrs_clt_path *clt_path)
762{
763 return list_next_or_null_rcu(head, &clt_path->s.entry, typeof(*clt_path), s.entry) ?:
764 list_next_or_null_rcu(head,
765 READ_ONCE((&clt_path->s.entry)->next),
766 typeof(*clt_path), s.entry);
767}
768
769/**
770 * get_next_path_rr() - Returns path in round-robin fashion.
771 * @it: the path pointer
772 *
773 * Related to @MP_POLICY_RR
774 *
775 * Locks:
776 * rcu_read_lock() must be hold.
777 */
778static struct rtrs_clt_path *get_next_path_rr(struct path_it *it)
779{
780 struct rtrs_clt_path __rcu **ppcpu_path;
781 struct rtrs_clt_path *path;
782 struct rtrs_clt_sess *clt;
783
784 clt = it->clt;
785
786 /*
787 * Here we use two RCU objects: @paths_list and @pcpu_path
788 * pointer. See rtrs_clt_remove_path_from_arr() for details
789 * how that is handled.
790 */
791
792 ppcpu_path = this_cpu_ptr(clt->pcpu_path);
793 path = rcu_dereference(*ppcpu_path);
794 if (!path)
795 path = list_first_or_null_rcu(&clt->paths_list,
796 typeof(*path), s.entry);
797 else
798 path = rtrs_clt_get_next_path_or_null(&clt->paths_list, path);
799
800 rcu_assign_pointer(*ppcpu_path, path);
801
802 return path;
803}
804
805/**
806 * get_next_path_min_inflight() - Returns path with minimal inflight count.
807 * @it: the path pointer
808 *
809 * Related to @MP_POLICY_MIN_INFLIGHT
810 *
811 * Locks:
812 * rcu_read_lock() must be hold.
813 */
814static struct rtrs_clt_path *get_next_path_min_inflight(struct path_it *it)
815{
816 struct rtrs_clt_path *min_path = NULL;
817 struct rtrs_clt_sess *clt = it->clt;
818 struct rtrs_clt_path *clt_path;
819 int min_inflight = INT_MAX;
820 int inflight;
821
822 list_for_each_entry_rcu(clt_path, &clt->paths_list, s.entry) {
823 if (READ_ONCE(clt_path->state) != RTRS_CLT_CONNECTED)
824 continue;
825
826 if (!list_empty(raw_cpu_ptr(clt_path->mp_skip_entry)))
827 continue;
828
829 inflight = atomic_read(&clt_path->stats->inflight);
830
831 if (inflight < min_inflight) {
832 min_inflight = inflight;
833 min_path = clt_path;
834 }
835 }
836
837 /*
838 * add the path to the skip list, so that next time we can get
839 * a different one
840 */
841 if (min_path)
842 list_add(raw_cpu_ptr(min_path->mp_skip_entry), &it->skip_list);
843
844 return min_path;
845}
846
847/**
848 * get_next_path_min_latency() - Returns path with minimal latency.
849 * @it: the path pointer
850 *
851 * Return: a path with the lowest latency or NULL if all paths are tried
852 *
853 * Locks:
854 * rcu_read_lock() must be hold.
855 *
856 * Related to @MP_POLICY_MIN_LATENCY
857 *
858 * This DOES skip an already-tried path.
859 * There is a skip-list to skip a path if the path has tried but failed.
860 * It will try the minimum latency path and then the second minimum latency
861 * path and so on. Finally it will return NULL if all paths are tried.
862 * Therefore the caller MUST check the returned
863 * path is NULL and trigger the IO error.
864 */
865static struct rtrs_clt_path *get_next_path_min_latency(struct path_it *it)
866{
867 struct rtrs_clt_path *min_path = NULL;
868 struct rtrs_clt_sess *clt = it->clt;
869 struct rtrs_clt_path *clt_path;
870 ktime_t min_latency = KTIME_MAX;
871 ktime_t latency;
872
873 list_for_each_entry_rcu(clt_path, &clt->paths_list, s.entry) {
874 if (READ_ONCE(clt_path->state) != RTRS_CLT_CONNECTED)
875 continue;
876
877 if (!list_empty(raw_cpu_ptr(clt_path->mp_skip_entry)))
878 continue;
879
880 latency = clt_path->s.hb_cur_latency;
881
882 if (latency < min_latency) {
883 min_latency = latency;
884 min_path = clt_path;
885 }
886 }
887
888 /*
889 * add the path to the skip list, so that next time we can get
890 * a different one
891 */
892 if (min_path)
893 list_add(raw_cpu_ptr(min_path->mp_skip_entry), &it->skip_list);
894
895 return min_path;
896}
897
898static inline void path_it_init(struct path_it *it, struct rtrs_clt_sess *clt)
899{
900 INIT_LIST_HEAD(&it->skip_list);
901 it->clt = clt;
902 it->i = 0;
903
904 if (clt->mp_policy == MP_POLICY_RR)
905 it->next_path = get_next_path_rr;
906 else if (clt->mp_policy == MP_POLICY_MIN_INFLIGHT)
907 it->next_path = get_next_path_min_inflight;
908 else
909 it->next_path = get_next_path_min_latency;
910}
911
912static inline void path_it_deinit(struct path_it *it)
913{
914 struct list_head *skip, *tmp;
915 /*
916 * The skip_list is used only for the MIN_INFLIGHT and MIN_LATENCY policies.
917 * We need to remove paths from it, so that next IO can insert
918 * paths (->mp_skip_entry) into a skip_list again.
919 */
920 list_for_each_safe(skip, tmp, &it->skip_list)
921 list_del_init(skip);
922}
923
924/**
925 * rtrs_clt_init_req() - Initialize an rtrs_clt_io_req holding information
926 * about an inflight IO.
927 * The user buffer holding user control message (not data) is copied into
928 * the corresponding buffer of rtrs_iu (req->iu->buf), which later on will
929 * also hold the control message of rtrs.
930 * @req: an io request holding information about IO.
931 * @clt_path: client path
932 * @conf: conformation callback function to notify upper layer.
933 * @permit: permit for allocation of RDMA remote buffer
934 * @priv: private pointer
935 * @vec: kernel vector containing control message
936 * @usr_len: length of the user message
937 * @sg: scater list for IO data
938 * @sg_cnt: number of scater list entries
939 * @data_len: length of the IO data
940 * @dir: direction of the IO.
941 */
942static void rtrs_clt_init_req(struct rtrs_clt_io_req *req,
943 struct rtrs_clt_path *clt_path,
944 void (*conf)(void *priv, int errno),
945 struct rtrs_permit *permit, void *priv,
946 const struct kvec *vec, size_t usr_len,
947 struct scatterlist *sg, size_t sg_cnt,
948 size_t data_len, int dir)
949{
950 struct iov_iter iter;
951 size_t len;
952
953 req->permit = permit;
954 req->in_use = true;
955 req->usr_len = usr_len;
956 req->data_len = data_len;
957 req->sglist = sg;
958 req->sg_cnt = sg_cnt;
959 req->priv = priv;
960 req->dir = dir;
961 req->con = rtrs_permit_to_clt_con(clt_path, permit);
962 req->conf = conf;
963 req->need_inv = false;
964 req->need_inv_comp = false;
965 req->inv_errno = 0;
966 refcount_set(&req->ref, 1);
967 req->mp_policy = clt_path->clt->mp_policy;
968
969 iov_iter_kvec(&iter, ITER_SOURCE, vec, 1, usr_len);
970 len = _copy_from_iter(req->iu->buf, usr_len, &iter);
971 WARN_ON(len != usr_len);
972
973 reinit_completion(&req->inv_comp);
974}
975
976static struct rtrs_clt_io_req *
977rtrs_clt_get_req(struct rtrs_clt_path *clt_path,
978 void (*conf)(void *priv, int errno),
979 struct rtrs_permit *permit, void *priv,
980 const struct kvec *vec, size_t usr_len,
981 struct scatterlist *sg, size_t sg_cnt,
982 size_t data_len, int dir)
983{
984 struct rtrs_clt_io_req *req;
985
986 req = &clt_path->reqs[permit->mem_id];
987 rtrs_clt_init_req(req, clt_path, conf, permit, priv, vec, usr_len,
988 sg, sg_cnt, data_len, dir);
989 return req;
990}
991
992static struct rtrs_clt_io_req *
993rtrs_clt_get_copy_req(struct rtrs_clt_path *alive_path,
994 struct rtrs_clt_io_req *fail_req)
995{
996 struct rtrs_clt_io_req *req;
997 struct kvec vec = {
998 .iov_base = fail_req->iu->buf,
999 .iov_len = fail_req->usr_len
1000 };
1001
1002 req = &alive_path->reqs[fail_req->permit->mem_id];
1003 rtrs_clt_init_req(req, alive_path, fail_req->conf, fail_req->permit,
1004 fail_req->priv, &vec, fail_req->usr_len,
1005 fail_req->sglist, fail_req->sg_cnt,
1006 fail_req->data_len, fail_req->dir);
1007 return req;
1008}
1009
1010static int rtrs_post_rdma_write_sg(struct rtrs_clt_con *con,
1011 struct rtrs_clt_io_req *req,
1012 struct rtrs_rbuf *rbuf, bool fr_en,
1013 u32 count, u32 size, u32 imm,
1014 struct ib_send_wr *wr,
1015 struct ib_send_wr *tail)
1016{
1017 struct rtrs_clt_path *clt_path = to_clt_path(con->c.path);
1018 struct ib_sge *sge = req->sge;
1019 enum ib_send_flags flags;
1020 struct scatterlist *sg;
1021 size_t num_sge;
1022 int i;
1023 struct ib_send_wr *ptail = NULL;
1024
1025 if (fr_en) {
1026 i = 0;
1027 sge[i].addr = req->mr->iova;
1028 sge[i].length = req->mr->length;
1029 sge[i].lkey = req->mr->lkey;
1030 i++;
1031 num_sge = 2;
1032 ptail = tail;
1033 } else {
1034 for_each_sg(req->sglist, sg, count, i) {
1035 sge[i].addr = sg_dma_address(sg);
1036 sge[i].length = sg_dma_len(sg);
1037 sge[i].lkey = clt_path->s.dev->ib_pd->local_dma_lkey;
1038 }
1039 num_sge = 1 + count;
1040 }
1041 sge[i].addr = req->iu->dma_addr;
1042 sge[i].length = size;
1043 sge[i].lkey = clt_path->s.dev->ib_pd->local_dma_lkey;
1044
1045 /*
1046 * From time to time we have to post signalled sends,
1047 * or send queue will fill up and only QP reset can help.
1048 */
1049 flags = atomic_inc_return(&con->c.wr_cnt) % clt_path->s.signal_interval ?
1050 0 : IB_SEND_SIGNALED;
1051
1052 ib_dma_sync_single_for_device(clt_path->s.dev->ib_dev,
1053 req->iu->dma_addr,
1054 size, DMA_TO_DEVICE);
1055
1056 return rtrs_iu_post_rdma_write_imm(&con->c, req->iu, sge, num_sge,
1057 rbuf->rkey, rbuf->addr, imm,
1058 flags, wr, ptail);
1059}
1060
1061static int rtrs_map_sg_fr(struct rtrs_clt_io_req *req, size_t count)
1062{
1063 int nr;
1064
1065 /* Align the MR to a 4K page size to match the block virt boundary */
1066 nr = ib_map_mr_sg(req->mr, req->sglist, count, NULL, SZ_4K);
1067 if (nr != count)
1068 return nr < 0 ? nr : -EINVAL;
1069 ib_update_fast_reg_key(req->mr, ib_inc_rkey(req->mr->rkey));
1070
1071 return nr;
1072}
1073
1074static int rtrs_clt_write_req(struct rtrs_clt_io_req *req)
1075{
1076 struct rtrs_clt_con *con = req->con;
1077 struct rtrs_path *s = con->c.path;
1078 struct rtrs_clt_path *clt_path = to_clt_path(s);
1079 struct rtrs_msg_rdma_write *msg;
1080
1081 struct rtrs_rbuf *rbuf;
1082 int ret, count = 0;
1083 u32 imm, buf_id;
1084 struct ib_reg_wr rwr;
1085 struct ib_send_wr inv_wr;
1086 struct ib_send_wr *wr = NULL;
1087 bool fr_en = false;
1088
1089 const size_t tsize = sizeof(*msg) + req->data_len + req->usr_len;
1090
1091 if (tsize > clt_path->chunk_size) {
1092 rtrs_wrn(s, "Write request failed, size too big %zu > %d\n",
1093 tsize, clt_path->chunk_size);
1094 return -EMSGSIZE;
1095 }
1096 if (req->sg_cnt) {
1097 count = ib_dma_map_sg(clt_path->s.dev->ib_dev, req->sglist,
1098 req->sg_cnt, req->dir);
1099 if (!count) {
1100 rtrs_wrn(s, "Write request failed, map failed\n");
1101 return -EINVAL;
1102 }
1103 }
1104 /* put rtrs msg after sg and user message */
1105 msg = req->iu->buf + req->usr_len;
1106 msg->type = cpu_to_le16(RTRS_MSG_WRITE);
1107 msg->usr_len = cpu_to_le16(req->usr_len);
1108
1109 /* rtrs message on server side will be after user data and message */
1110 imm = req->permit->mem_off + req->data_len + req->usr_len;
1111 imm = rtrs_to_io_req_imm(imm);
1112 buf_id = req->permit->mem_id;
1113 req->sg_size = tsize;
1114 rbuf = &clt_path->rbufs[buf_id];
1115
1116 if (count) {
1117 ret = rtrs_map_sg_fr(req, count);
1118 if (ret < 0) {
1119 rtrs_err_rl(s,
1120 "Write request failed, failed to map fast reg. data, err: %d\n",
1121 ret);
1122 ib_dma_unmap_sg(clt_path->s.dev->ib_dev, req->sglist,
1123 req->sg_cnt, req->dir);
1124 return ret;
1125 }
1126 inv_wr = (struct ib_send_wr) {
1127 .opcode = IB_WR_LOCAL_INV,
1128 .wr_cqe = &req->inv_cqe,
1129 .send_flags = IB_SEND_SIGNALED,
1130 .ex.invalidate_rkey = req->mr->rkey,
1131 };
1132 req->inv_cqe.done = rtrs_clt_inv_rkey_done;
1133 rwr = (struct ib_reg_wr) {
1134 .wr.opcode = IB_WR_REG_MR,
1135 .wr.wr_cqe = &fast_reg_cqe,
1136 .mr = req->mr,
1137 .key = req->mr->rkey,
1138 .access = (IB_ACCESS_LOCAL_WRITE),
1139 };
1140 wr = &rwr.wr;
1141 fr_en = true;
1142 refcount_inc(&req->ref);
1143 }
1144 /*
1145 * Update stats now, after request is successfully sent it is not
1146 * safe anymore to touch it.
1147 */
1148 rtrs_clt_update_all_stats(req, WRITE);
1149
1150 ret = rtrs_post_rdma_write_sg(req->con, req, rbuf, fr_en, count,
1151 req->usr_len + sizeof(*msg),
1152 imm, wr, &inv_wr);
1153 if (ret) {
1154 rtrs_err_rl(s,
1155 "Write request failed: error=%d path=%s [%s:%u]\n",
1156 ret, kobject_name(&clt_path->kobj), clt_path->hca_name,
1157 clt_path->hca_port);
1158 if (req->mp_policy == MP_POLICY_MIN_INFLIGHT)
1159 atomic_dec(&clt_path->stats->inflight);
1160 if (req->sg_cnt)
1161 ib_dma_unmap_sg(clt_path->s.dev->ib_dev, req->sglist,
1162 req->sg_cnt, req->dir);
1163 }
1164
1165 return ret;
1166}
1167
1168static int rtrs_clt_read_req(struct rtrs_clt_io_req *req)
1169{
1170 struct rtrs_clt_con *con = req->con;
1171 struct rtrs_path *s = con->c.path;
1172 struct rtrs_clt_path *clt_path = to_clt_path(s);
1173 struct rtrs_msg_rdma_read *msg;
1174 struct rtrs_ib_dev *dev = clt_path->s.dev;
1175
1176 struct ib_reg_wr rwr;
1177 struct ib_send_wr *wr = NULL;
1178
1179 int ret, count = 0;
1180 u32 imm, buf_id;
1181
1182 const size_t tsize = sizeof(*msg) + req->data_len + req->usr_len;
1183
1184 if (tsize > clt_path->chunk_size) {
1185 rtrs_wrn(s,
1186 "Read request failed, message size is %zu, bigger than CHUNK_SIZE %d\n",
1187 tsize, clt_path->chunk_size);
1188 return -EMSGSIZE;
1189 }
1190
1191 if (req->sg_cnt) {
1192 count = ib_dma_map_sg(dev->ib_dev, req->sglist, req->sg_cnt,
1193 req->dir);
1194 if (!count) {
1195 rtrs_wrn(s,
1196 "Read request failed, dma map failed\n");
1197 return -EINVAL;
1198 }
1199 }
1200 /* put our message into req->buf after user message*/
1201 msg = req->iu->buf + req->usr_len;
1202 msg->type = cpu_to_le16(RTRS_MSG_READ);
1203 msg->usr_len = cpu_to_le16(req->usr_len);
1204
1205 if (count) {
1206 ret = rtrs_map_sg_fr(req, count);
1207 if (ret < 0) {
1208 rtrs_err_rl(s,
1209 "Read request failed, failed to map fast reg. data, err: %d\n",
1210 ret);
1211 ib_dma_unmap_sg(dev->ib_dev, req->sglist, req->sg_cnt,
1212 req->dir);
1213 return ret;
1214 }
1215 rwr = (struct ib_reg_wr) {
1216 .wr.opcode = IB_WR_REG_MR,
1217 .wr.wr_cqe = &fast_reg_cqe,
1218 .mr = req->mr,
1219 .key = req->mr->rkey,
1220 .access = (IB_ACCESS_LOCAL_WRITE |
1221 IB_ACCESS_REMOTE_WRITE),
1222 };
1223 wr = &rwr.wr;
1224
1225 msg->sg_cnt = cpu_to_le16(1);
1226 msg->flags = cpu_to_le16(RTRS_MSG_NEED_INVAL_F);
1227
1228 msg->desc[0].addr = cpu_to_le64(req->mr->iova);
1229 msg->desc[0].key = cpu_to_le32(req->mr->rkey);
1230 msg->desc[0].len = cpu_to_le32(req->mr->length);
1231
1232 /* Further invalidation is required */
1233 req->need_inv = !!RTRS_MSG_NEED_INVAL_F;
1234
1235 } else {
1236 msg->sg_cnt = 0;
1237 msg->flags = 0;
1238 }
1239 /*
1240 * rtrs message will be after the space reserved for disk data and
1241 * user message
1242 */
1243 imm = req->permit->mem_off + req->data_len + req->usr_len;
1244 imm = rtrs_to_io_req_imm(imm);
1245 buf_id = req->permit->mem_id;
1246
1247 req->sg_size = sizeof(*msg);
1248 req->sg_size += le16_to_cpu(msg->sg_cnt) * sizeof(struct rtrs_sg_desc);
1249 req->sg_size += req->usr_len;
1250
1251 /*
1252 * Update stats now, after request is successfully sent it is not
1253 * safe anymore to touch it.
1254 */
1255 rtrs_clt_update_all_stats(req, READ);
1256
1257 ret = rtrs_post_send_rdma(req->con, req, &clt_path->rbufs[buf_id],
1258 req->data_len, imm, wr);
1259 if (ret) {
1260 rtrs_err_rl(s,
1261 "Read request failed: error=%d path=%s [%s:%u]\n",
1262 ret, kobject_name(&clt_path->kobj), clt_path->hca_name,
1263 clt_path->hca_port);
1264 if (req->mp_policy == MP_POLICY_MIN_INFLIGHT)
1265 atomic_dec(&clt_path->stats->inflight);
1266 req->need_inv = false;
1267 if (req->sg_cnt)
1268 ib_dma_unmap_sg(dev->ib_dev, req->sglist,
1269 req->sg_cnt, req->dir);
1270 }
1271
1272 return ret;
1273}
1274
1275/**
1276 * rtrs_clt_failover_req() - Try to find an active path for a failed request
1277 * @clt: clt context
1278 * @fail_req: a failed io request.
1279 */
1280static int rtrs_clt_failover_req(struct rtrs_clt_sess *clt,
1281 struct rtrs_clt_io_req *fail_req)
1282{
1283 struct rtrs_clt_path *alive_path;
1284 struct rtrs_clt_io_req *req;
1285 int err = -ECONNABORTED;
1286 struct path_it it;
1287
1288 rcu_read_lock();
1289 for (path_it_init(&it, clt);
1290 (alive_path = it.next_path(&it)) && it.i < it.clt->paths_num;
1291 it.i++) {
1292 if (READ_ONCE(alive_path->state) != RTRS_CLT_CONNECTED)
1293 continue;
1294 req = rtrs_clt_get_copy_req(alive_path, fail_req);
1295 if (req->dir == DMA_TO_DEVICE)
1296 err = rtrs_clt_write_req(req);
1297 else
1298 err = rtrs_clt_read_req(req);
1299 if (err) {
1300 req->in_use = false;
1301 continue;
1302 }
1303 /* Success path */
1304 rtrs_clt_inc_failover_cnt(alive_path->stats);
1305 break;
1306 }
1307 path_it_deinit(&it);
1308 rcu_read_unlock();
1309
1310 return err;
1311}
1312
1313static void fail_all_outstanding_reqs(struct rtrs_clt_path *clt_path)
1314{
1315 struct rtrs_clt_sess *clt = clt_path->clt;
1316 struct rtrs_clt_io_req *req;
1317 int i, err;
1318
1319 if (!clt_path->reqs)
1320 return;
1321 for (i = 0; i < clt_path->queue_depth; ++i) {
1322 req = &clt_path->reqs[i];
1323 if (!req->in_use)
1324 continue;
1325
1326 /*
1327 * Safely (without notification) complete failed request.
1328 * After completion this request is still useble and can
1329 * be failovered to another path.
1330 */
1331 complete_rdma_req(req, -ECONNABORTED, false, true);
1332
1333 err = rtrs_clt_failover_req(clt, req);
1334 if (err)
1335 /* Failover failed, notify anyway */
1336 req->conf(req->priv, err);
1337 }
1338}
1339
1340static void free_path_reqs(struct rtrs_clt_path *clt_path)
1341{
1342 struct rtrs_clt_io_req *req;
1343 int i;
1344
1345 if (!clt_path->reqs)
1346 return;
1347 for (i = 0; i < clt_path->queue_depth; ++i) {
1348 req = &clt_path->reqs[i];
1349 if (req->mr)
1350 ib_dereg_mr(req->mr);
1351 kfree(req->sge);
1352 rtrs_iu_free(req->iu, clt_path->s.dev->ib_dev, 1);
1353 }
1354 kfree(clt_path->reqs);
1355 clt_path->reqs = NULL;
1356}
1357
1358static int alloc_path_reqs(struct rtrs_clt_path *clt_path)
1359{
1360 struct rtrs_clt_io_req *req;
1361 int i, err = -ENOMEM;
1362
1363 clt_path->reqs = kcalloc(clt_path->queue_depth,
1364 sizeof(*clt_path->reqs),
1365 GFP_KERNEL);
1366 if (!clt_path->reqs)
1367 return -ENOMEM;
1368
1369 for (i = 0; i < clt_path->queue_depth; ++i) {
1370 req = &clt_path->reqs[i];
1371 req->iu = rtrs_iu_alloc(1, clt_path->max_hdr_size, GFP_KERNEL,
1372 clt_path->s.dev->ib_dev,
1373 DMA_TO_DEVICE,
1374 rtrs_clt_rdma_done);
1375 if (!req->iu)
1376 goto out;
1377
1378 req->sge = kcalloc(2, sizeof(*req->sge), GFP_KERNEL);
1379 if (!req->sge)
1380 goto out;
1381
1382 req->mr = ib_alloc_mr(clt_path->s.dev->ib_pd,
1383 IB_MR_TYPE_MEM_REG,
1384 clt_path->max_pages_per_mr);
1385 if (IS_ERR(req->mr)) {
1386 err = PTR_ERR(req->mr);
1387 req->mr = NULL;
1388 pr_err("Failed to alloc clt_path->max_pages_per_mr %d\n",
1389 clt_path->max_pages_per_mr);
1390 goto out;
1391 }
1392
1393 init_completion(&req->inv_comp);
1394 }
1395
1396 return 0;
1397
1398out:
1399 free_path_reqs(clt_path);
1400
1401 return err;
1402}
1403
1404static int alloc_permits(struct rtrs_clt_sess *clt)
1405{
1406 unsigned int chunk_bits;
1407 int err, i;
1408
1409 clt->permits_map = bitmap_zalloc(clt->queue_depth, GFP_KERNEL);
1410 if (!clt->permits_map) {
1411 err = -ENOMEM;
1412 goto out_err;
1413 }
1414 clt->permits = kcalloc(clt->queue_depth, permit_size(clt), GFP_KERNEL);
1415 if (!clt->permits) {
1416 err = -ENOMEM;
1417 goto err_map;
1418 }
1419 chunk_bits = ilog2(clt->queue_depth - 1) + 1;
1420 for (i = 0; i < clt->queue_depth; i++) {
1421 struct rtrs_permit *permit;
1422
1423 permit = get_permit(clt, i);
1424 permit->mem_id = i;
1425 permit->mem_off = i << (MAX_IMM_PAYL_BITS - chunk_bits);
1426 }
1427
1428 return 0;
1429
1430err_map:
1431 bitmap_free(clt->permits_map);
1432 clt->permits_map = NULL;
1433out_err:
1434 return err;
1435}
1436
1437static void free_permits(struct rtrs_clt_sess *clt)
1438{
1439 if (clt->permits_map)
1440 wait_event(clt->permits_wait,
1441 bitmap_empty(clt->permits_map, clt->queue_depth));
1442
1443 bitmap_free(clt->permits_map);
1444 clt->permits_map = NULL;
1445 kfree(clt->permits);
1446 clt->permits = NULL;
1447}
1448
1449static void query_fast_reg_mode(struct rtrs_clt_path *clt_path)
1450{
1451 struct ib_device *ib_dev;
1452 u64 max_pages_per_mr;
1453 int mr_page_shift;
1454
1455 ib_dev = clt_path->s.dev->ib_dev;
1456
1457 /*
1458 * Use the smallest page size supported by the HCA, down to a
1459 * minimum of 4096 bytes. We're unlikely to build large sglists
1460 * out of smaller entries.
1461 */
1462 mr_page_shift = max(12, ffs(ib_dev->attrs.page_size_cap) - 1);
1463 max_pages_per_mr = ib_dev->attrs.max_mr_size;
1464 do_div(max_pages_per_mr, (1ull << mr_page_shift));
1465 clt_path->max_pages_per_mr =
1466 min3(clt_path->max_pages_per_mr, (u32)max_pages_per_mr,
1467 ib_dev->attrs.max_fast_reg_page_list_len);
1468 clt_path->clt->max_segments =
1469 min(clt_path->max_pages_per_mr, clt_path->clt->max_segments);
1470}
1471
1472static bool rtrs_clt_change_state_get_old(struct rtrs_clt_path *clt_path,
1473 enum rtrs_clt_state new_state,
1474 enum rtrs_clt_state *old_state)
1475{
1476 bool changed;
1477
1478 spin_lock_irq(&clt_path->state_wq.lock);
1479 if (old_state)
1480 *old_state = clt_path->state;
1481 changed = rtrs_clt_change_state(clt_path, new_state);
1482 spin_unlock_irq(&clt_path->state_wq.lock);
1483
1484 return changed;
1485}
1486
1487static void rtrs_clt_hb_err_handler(struct rtrs_con *c)
1488{
1489 struct rtrs_clt_con *con = container_of(c, typeof(*con), c);
1490
1491 rtrs_rdma_error_recovery(con);
1492}
1493
1494static void rtrs_clt_init_hb(struct rtrs_clt_path *clt_path)
1495{
1496 rtrs_init_hb(&clt_path->s, &io_comp_cqe,
1497 RTRS_HB_INTERVAL_MS,
1498 RTRS_HB_MISSED_MAX,
1499 rtrs_clt_hb_err_handler,
1500 rtrs_wq);
1501}
1502
1503static void rtrs_clt_reconnect_work(struct work_struct *work);
1504static void rtrs_clt_close_work(struct work_struct *work);
1505
1506static void rtrs_clt_err_recovery_work(struct work_struct *work)
1507{
1508 struct rtrs_clt_path *clt_path;
1509 struct rtrs_clt_sess *clt;
1510 int delay_ms;
1511
1512 clt_path = container_of(work, struct rtrs_clt_path, err_recovery_work);
1513 clt = clt_path->clt;
1514 delay_ms = clt->reconnect_delay_sec * 1000;
1515 rtrs_clt_stop_and_destroy_conns(clt_path);
1516 queue_delayed_work(rtrs_wq, &clt_path->reconnect_dwork,
1517 msecs_to_jiffies(delay_ms +
1518 get_random_u32_below(RTRS_RECONNECT_SEED)));
1519}
1520
1521static struct rtrs_clt_path *alloc_path(struct rtrs_clt_sess *clt,
1522 const struct rtrs_addr *path,
1523 size_t con_num, u32 nr_poll_queues)
1524{
1525 struct rtrs_clt_path *clt_path;
1526 int err = -ENOMEM;
1527 int cpu;
1528 size_t total_con;
1529
1530 clt_path = kzalloc(sizeof(*clt_path), GFP_KERNEL);
1531 if (!clt_path)
1532 goto err;
1533
1534 /*
1535 * irqmode and poll
1536 * +1: Extra connection for user messages
1537 */
1538 total_con = con_num + nr_poll_queues + 1;
1539 clt_path->s.con = kcalloc(total_con, sizeof(*clt_path->s.con),
1540 GFP_KERNEL);
1541 if (!clt_path->s.con)
1542 goto err_free_path;
1543
1544 clt_path->s.con_num = total_con;
1545 clt_path->s.irq_con_num = con_num + 1;
1546
1547 clt_path->stats = kzalloc(sizeof(*clt_path->stats), GFP_KERNEL);
1548 if (!clt_path->stats)
1549 goto err_free_con;
1550
1551 mutex_init(&clt_path->init_mutex);
1552 uuid_gen(&clt_path->s.uuid);
1553 memcpy(&clt_path->s.dst_addr, path->dst,
1554 rdma_addr_size((struct sockaddr *)path->dst));
1555
1556 /*
1557 * rdma_resolve_addr() passes src_addr to cma_bind_addr, which
1558 * checks the sa_family to be non-zero. If user passed src_addr=NULL
1559 * the sess->src_addr will contain only zeros, which is then fine.
1560 */
1561 if (path->src)
1562 memcpy(&clt_path->s.src_addr, path->src,
1563 rdma_addr_size((struct sockaddr *)path->src));
1564 strscpy(clt_path->s.sessname, clt->sessname,
1565 sizeof(clt_path->s.sessname));
1566 clt_path->clt = clt;
1567 clt_path->max_pages_per_mr = RTRS_MAX_SEGMENTS;
1568 init_waitqueue_head(&clt_path->state_wq);
1569 clt_path->state = RTRS_CLT_CONNECTING;
1570 atomic_set(&clt_path->connected_cnt, 0);
1571 INIT_WORK(&clt_path->close_work, rtrs_clt_close_work);
1572 INIT_WORK(&clt_path->err_recovery_work, rtrs_clt_err_recovery_work);
1573 INIT_DELAYED_WORK(&clt_path->reconnect_dwork, rtrs_clt_reconnect_work);
1574 rtrs_clt_init_hb(clt_path);
1575
1576 clt_path->mp_skip_entry = alloc_percpu(typeof(*clt_path->mp_skip_entry));
1577 if (!clt_path->mp_skip_entry)
1578 goto err_free_stats;
1579
1580 for_each_possible_cpu(cpu)
1581 INIT_LIST_HEAD(per_cpu_ptr(clt_path->mp_skip_entry, cpu));
1582
1583 err = rtrs_clt_init_stats(clt_path->stats);
1584 if (err)
1585 goto err_free_percpu;
1586
1587 return clt_path;
1588
1589err_free_percpu:
1590 free_percpu(clt_path->mp_skip_entry);
1591err_free_stats:
1592 kfree(clt_path->stats);
1593err_free_con:
1594 kfree(clt_path->s.con);
1595err_free_path:
1596 kfree(clt_path);
1597err:
1598 return ERR_PTR(err);
1599}
1600
1601void free_path(struct rtrs_clt_path *clt_path)
1602{
1603 free_percpu(clt_path->mp_skip_entry);
1604 mutex_destroy(&clt_path->init_mutex);
1605 kfree(clt_path->s.con);
1606 kfree(clt_path->rbufs);
1607 kfree(clt_path);
1608}
1609
1610static int create_con(struct rtrs_clt_path *clt_path, unsigned int cid)
1611{
1612 struct rtrs_clt_con *con;
1613
1614 con = kzalloc(sizeof(*con), GFP_KERNEL);
1615 if (!con)
1616 return -ENOMEM;
1617
1618 /* Map first two connections to the first CPU */
1619 con->cpu = (cid ? cid - 1 : 0) % nr_cpu_ids;
1620 con->c.cid = cid;
1621 con->c.path = &clt_path->s;
1622 /* Align with srv, init as 1 */
1623 atomic_set(&con->c.wr_cnt, 1);
1624 mutex_init(&con->con_mutex);
1625
1626 clt_path->s.con[cid] = &con->c;
1627
1628 return 0;
1629}
1630
1631static void destroy_con(struct rtrs_clt_con *con)
1632{
1633 struct rtrs_clt_path *clt_path = to_clt_path(con->c.path);
1634
1635 clt_path->s.con[con->c.cid] = NULL;
1636 mutex_destroy(&con->con_mutex);
1637 kfree(con);
1638}
1639
1640static int create_con_cq_qp(struct rtrs_clt_con *con)
1641{
1642 struct rtrs_clt_path *clt_path = to_clt_path(con->c.path);
1643 u32 max_send_wr, max_recv_wr, cq_num, max_send_sge, wr_limit;
1644 int err, cq_vector;
1645 struct rtrs_msg_rkey_rsp *rsp;
1646
1647 lockdep_assert_held(&con->con_mutex);
1648 if (con->c.cid == 0) {
1649 max_send_sge = 1;
1650 /* We must be the first here */
1651 if (WARN_ON(clt_path->s.dev))
1652 return -EINVAL;
1653
1654 /*
1655 * The whole session uses device from user connection.
1656 * Be careful not to close user connection before ib dev
1657 * is gracefully put.
1658 */
1659 clt_path->s.dev = rtrs_ib_dev_find_or_add(con->c.cm_id->device,
1660 &dev_pd);
1661 if (!clt_path->s.dev) {
1662 rtrs_wrn(clt_path->clt,
1663 "rtrs_ib_dev_find_get_or_add(): no memory\n");
1664 return -ENOMEM;
1665 }
1666 clt_path->s.dev_ref = 1;
1667 query_fast_reg_mode(clt_path);
1668 wr_limit = clt_path->s.dev->ib_dev->attrs.max_qp_wr;
1669 /*
1670 * Two (request + registration) completion for send
1671 * Two for recv if always_invalidate is set on server
1672 * or one for recv.
1673 * + 2 for drain and heartbeat
1674 * in case qp gets into error state.
1675 */
1676 max_send_wr =
1677 min_t(int, wr_limit, SERVICE_CON_QUEUE_DEPTH * 2 + 2);
1678 max_recv_wr = max_send_wr;
1679 } else {
1680 /*
1681 * Here we assume that session members are correctly set.
1682 * This is always true if user connection (cid == 0) is
1683 * established first.
1684 */
1685 if (WARN_ON(!clt_path->s.dev))
1686 return -EINVAL;
1687 if (WARN_ON(!clt_path->queue_depth))
1688 return -EINVAL;
1689
1690 wr_limit = clt_path->s.dev->ib_dev->attrs.max_qp_wr;
1691 /* Shared between connections */
1692 clt_path->s.dev_ref++;
1693 max_send_wr = min_t(int, wr_limit,
1694 /* QD * (REQ + RSP + FR REGS or INVS) + drain */
1695 clt_path->queue_depth * 3 + 1);
1696 max_recv_wr = min_t(int, wr_limit,
1697 clt_path->queue_depth * 3 + 1);
1698 max_send_sge = 2;
1699 }
1700 atomic_set(&con->c.sq_wr_avail, max_send_wr);
1701 cq_num = max_send_wr + max_recv_wr;
1702 /* alloc iu to recv new rkey reply when server reports flags set */
1703 if (clt_path->flags & RTRS_MSG_NEW_RKEY_F || con->c.cid == 0) {
1704 con->rsp_ius = rtrs_iu_alloc(cq_num, sizeof(*rsp),
1705 GFP_KERNEL,
1706 clt_path->s.dev->ib_dev,
1707 DMA_FROM_DEVICE,
1708 rtrs_clt_rdma_done);
1709 if (!con->rsp_ius)
1710 return -ENOMEM;
1711 con->queue_num = cq_num;
1712 }
1713 cq_num = max_send_wr + max_recv_wr;
1714 cq_vector = con->cpu % clt_path->s.dev->ib_dev->num_comp_vectors;
1715 if (con->c.cid >= clt_path->s.irq_con_num)
1716 err = rtrs_cq_qp_create(&clt_path->s, &con->c, max_send_sge,
1717 cq_vector, cq_num, max_send_wr,
1718 max_recv_wr, IB_POLL_DIRECT);
1719 else
1720 err = rtrs_cq_qp_create(&clt_path->s, &con->c, max_send_sge,
1721 cq_vector, cq_num, max_send_wr,
1722 max_recv_wr, IB_POLL_SOFTIRQ);
1723 /*
1724 * In case of error we do not bother to clean previous allocations,
1725 * since destroy_con_cq_qp() must be called.
1726 */
1727 return err;
1728}
1729
1730static void destroy_con_cq_qp(struct rtrs_clt_con *con)
1731{
1732 struct rtrs_clt_path *clt_path = to_clt_path(con->c.path);
1733
1734 /*
1735 * Be careful here: destroy_con_cq_qp() can be called even
1736 * create_con_cq_qp() failed, see comments there.
1737 */
1738 lockdep_assert_held(&con->con_mutex);
1739 rtrs_cq_qp_destroy(&con->c);
1740 if (con->rsp_ius) {
1741 rtrs_iu_free(con->rsp_ius, clt_path->s.dev->ib_dev,
1742 con->queue_num);
1743 con->rsp_ius = NULL;
1744 con->queue_num = 0;
1745 }
1746 if (clt_path->s.dev_ref && !--clt_path->s.dev_ref) {
1747 rtrs_ib_dev_put(clt_path->s.dev);
1748 clt_path->s.dev = NULL;
1749 }
1750}
1751
1752static void stop_cm(struct rtrs_clt_con *con)
1753{
1754 rdma_disconnect(con->c.cm_id);
1755 if (con->c.qp)
1756 ib_drain_qp(con->c.qp);
1757}
1758
1759static void destroy_cm(struct rtrs_clt_con *con)
1760{
1761 rdma_destroy_id(con->c.cm_id);
1762 con->c.cm_id = NULL;
1763}
1764
1765static int rtrs_rdma_addr_resolved(struct rtrs_clt_con *con)
1766{
1767 struct rtrs_path *s = con->c.path;
1768 int err;
1769
1770 mutex_lock(&con->con_mutex);
1771 err = create_con_cq_qp(con);
1772 mutex_unlock(&con->con_mutex);
1773 if (err) {
1774 rtrs_err(s, "create_con_cq_qp(), err: %d\n", err);
1775 return err;
1776 }
1777 err = rdma_resolve_route(con->c.cm_id, RTRS_CONNECT_TIMEOUT_MS);
1778 if (err)
1779 rtrs_err(s, "Resolving route failed, err: %d\n", err);
1780
1781 return err;
1782}
1783
1784static int rtrs_rdma_route_resolved(struct rtrs_clt_con *con)
1785{
1786 struct rtrs_clt_path *clt_path = to_clt_path(con->c.path);
1787 struct rtrs_clt_sess *clt = clt_path->clt;
1788 struct rtrs_msg_conn_req msg;
1789 struct rdma_conn_param param;
1790
1791 int err;
1792
1793 param = (struct rdma_conn_param) {
1794 .retry_count = 7,
1795 .rnr_retry_count = 7,
1796 .private_data = &msg,
1797 .private_data_len = sizeof(msg),
1798 };
1799
1800 msg = (struct rtrs_msg_conn_req) {
1801 .magic = cpu_to_le16(RTRS_MAGIC),
1802 .version = cpu_to_le16(RTRS_PROTO_VER),
1803 .cid = cpu_to_le16(con->c.cid),
1804 .cid_num = cpu_to_le16(clt_path->s.con_num),
1805 .recon_cnt = cpu_to_le16(clt_path->s.recon_cnt),
1806 };
1807 msg.first_conn = clt_path->for_new_clt ? FIRST_CONN : 0;
1808 uuid_copy(&msg.sess_uuid, &clt_path->s.uuid);
1809 uuid_copy(&msg.paths_uuid, &clt->paths_uuid);
1810
1811 err = rdma_connect_locked(con->c.cm_id, ¶m);
1812 if (err)
1813 rtrs_err(clt, "rdma_connect_locked(): %d\n", err);
1814
1815 return err;
1816}
1817
1818static int rtrs_rdma_conn_established(struct rtrs_clt_con *con,
1819 struct rdma_cm_event *ev)
1820{
1821 struct rtrs_clt_path *clt_path = to_clt_path(con->c.path);
1822 struct rtrs_clt_sess *clt = clt_path->clt;
1823 const struct rtrs_msg_conn_rsp *msg;
1824 u16 version, queue_depth;
1825 int errno;
1826 u8 len;
1827
1828 msg = ev->param.conn.private_data;
1829 len = ev->param.conn.private_data_len;
1830 if (len < sizeof(*msg)) {
1831 rtrs_err(clt, "Invalid RTRS connection response\n");
1832 return -ECONNRESET;
1833 }
1834 if (le16_to_cpu(msg->magic) != RTRS_MAGIC) {
1835 rtrs_err(clt, "Invalid RTRS magic\n");
1836 return -ECONNRESET;
1837 }
1838 version = le16_to_cpu(msg->version);
1839 if (version >> 8 != RTRS_PROTO_VER_MAJOR) {
1840 rtrs_err(clt, "Unsupported major RTRS version: %d, expected %d\n",
1841 version >> 8, RTRS_PROTO_VER_MAJOR);
1842 return -ECONNRESET;
1843 }
1844 errno = le16_to_cpu(msg->errno);
1845 if (errno) {
1846 rtrs_err(clt, "Invalid RTRS message: errno %d\n",
1847 errno);
1848 return -ECONNRESET;
1849 }
1850 if (con->c.cid == 0) {
1851 queue_depth = le16_to_cpu(msg->queue_depth);
1852
1853 if (clt_path->queue_depth > 0 && queue_depth != clt_path->queue_depth) {
1854 rtrs_err(clt, "Error: queue depth changed\n");
1855
1856 /*
1857 * Stop any more reconnection attempts
1858 */
1859 clt_path->reconnect_attempts = -1;
1860 rtrs_err(clt,
1861 "Disabling auto-reconnect. Trigger a manual reconnect after issue is resolved\n");
1862 return -ECONNRESET;
1863 }
1864
1865 if (!clt_path->rbufs) {
1866 clt_path->rbufs = kcalloc(queue_depth,
1867 sizeof(*clt_path->rbufs),
1868 GFP_KERNEL);
1869 if (!clt_path->rbufs)
1870 return -ENOMEM;
1871 }
1872 clt_path->queue_depth = queue_depth;
1873 clt_path->s.signal_interval = min_not_zero(queue_depth,
1874 (unsigned short) SERVICE_CON_QUEUE_DEPTH);
1875 clt_path->max_hdr_size = le32_to_cpu(msg->max_hdr_size);
1876 clt_path->max_io_size = le32_to_cpu(msg->max_io_size);
1877 clt_path->flags = le32_to_cpu(msg->flags);
1878 clt_path->chunk_size = clt_path->max_io_size + clt_path->max_hdr_size;
1879
1880 /*
1881 * Global IO size is always a minimum.
1882 * If while a reconnection server sends us a value a bit
1883 * higher - client does not care and uses cached minimum.
1884 *
1885 * Since we can have several sessions (paths) restablishing
1886 * connections in parallel, use lock.
1887 */
1888 mutex_lock(&clt->paths_mutex);
1889 clt->queue_depth = clt_path->queue_depth;
1890 clt->max_io_size = min_not_zero(clt_path->max_io_size,
1891 clt->max_io_size);
1892 mutex_unlock(&clt->paths_mutex);
1893
1894 /*
1895 * Cache the hca_port and hca_name for sysfs
1896 */
1897 clt_path->hca_port = con->c.cm_id->port_num;
1898 scnprintf(clt_path->hca_name, sizeof(clt_path->hca_name),
1899 clt_path->s.dev->ib_dev->name);
1900 clt_path->s.src_addr = con->c.cm_id->route.addr.src_addr;
1901 /* set for_new_clt, to allow future reconnect on any path */
1902 clt_path->for_new_clt = 1;
1903 }
1904
1905 return 0;
1906}
1907
1908static inline void flag_success_on_conn(struct rtrs_clt_con *con)
1909{
1910 struct rtrs_clt_path *clt_path = to_clt_path(con->c.path);
1911
1912 atomic_inc(&clt_path->connected_cnt);
1913 con->cm_err = 1;
1914}
1915
1916static int rtrs_rdma_conn_rejected(struct rtrs_clt_con *con,
1917 struct rdma_cm_event *ev)
1918{
1919 struct rtrs_path *s = con->c.path;
1920 const struct rtrs_msg_conn_rsp *msg;
1921 const char *rej_msg;
1922 int status, errno;
1923 u8 data_len;
1924
1925 status = ev->status;
1926 rej_msg = rdma_reject_msg(con->c.cm_id, status);
1927 msg = rdma_consumer_reject_data(con->c.cm_id, ev, &data_len);
1928
1929 if (msg && data_len >= sizeof(*msg)) {
1930 errno = (int16_t)le16_to_cpu(msg->errno);
1931 if (errno == -EBUSY)
1932 rtrs_err(s,
1933 "Previous session is still exists on the server, please reconnect later\n");
1934 else
1935 rtrs_err(s,
1936 "Connect rejected: status %d (%s), rtrs errno %d\n",
1937 status, rej_msg, errno);
1938 } else {
1939 rtrs_err(s,
1940 "Connect rejected but with malformed message: status %d (%s)\n",
1941 status, rej_msg);
1942 }
1943
1944 return -ECONNRESET;
1945}
1946
1947void rtrs_clt_close_conns(struct rtrs_clt_path *clt_path, bool wait)
1948{
1949 trace_rtrs_clt_close_conns(clt_path);
1950
1951 if (rtrs_clt_change_state_get_old(clt_path, RTRS_CLT_CLOSING, NULL))
1952 queue_work(rtrs_wq, &clt_path->close_work);
1953 if (wait)
1954 flush_work(&clt_path->close_work);
1955}
1956
1957static inline void flag_error_on_conn(struct rtrs_clt_con *con, int cm_err)
1958{
1959 if (con->cm_err == 1) {
1960 struct rtrs_clt_path *clt_path;
1961
1962 clt_path = to_clt_path(con->c.path);
1963 if (atomic_dec_and_test(&clt_path->connected_cnt))
1964
1965 wake_up(&clt_path->state_wq);
1966 }
1967 con->cm_err = cm_err;
1968}
1969
1970static int rtrs_clt_rdma_cm_handler(struct rdma_cm_id *cm_id,
1971 struct rdma_cm_event *ev)
1972{
1973 struct rtrs_clt_con *con = cm_id->context;
1974 struct rtrs_path *s = con->c.path;
1975 struct rtrs_clt_path *clt_path = to_clt_path(s);
1976 int cm_err = 0;
1977
1978 switch (ev->event) {
1979 case RDMA_CM_EVENT_ADDR_RESOLVED:
1980 cm_err = rtrs_rdma_addr_resolved(con);
1981 break;
1982 case RDMA_CM_EVENT_ROUTE_RESOLVED:
1983 cm_err = rtrs_rdma_route_resolved(con);
1984 break;
1985 case RDMA_CM_EVENT_ESTABLISHED:
1986 cm_err = rtrs_rdma_conn_established(con, ev);
1987 if (!cm_err) {
1988 /*
1989 * Report success and wake up. Here we abuse state_wq,
1990 * i.e. wake up without state change, but we set cm_err.
1991 */
1992 flag_success_on_conn(con);
1993 wake_up(&clt_path->state_wq);
1994 return 0;
1995 }
1996 break;
1997 case RDMA_CM_EVENT_REJECTED:
1998 cm_err = rtrs_rdma_conn_rejected(con, ev);
1999 break;
2000 case RDMA_CM_EVENT_DISCONNECTED:
2001 /* No message for disconnecting */
2002 cm_err = -ECONNRESET;
2003 break;
2004 case RDMA_CM_EVENT_CONNECT_ERROR:
2005 case RDMA_CM_EVENT_UNREACHABLE:
2006 case RDMA_CM_EVENT_ADDR_CHANGE:
2007 case RDMA_CM_EVENT_TIMEWAIT_EXIT:
2008 rtrs_wrn(s, "CM error (CM event: %s, err: %d)\n",
2009 rdma_event_msg(ev->event), ev->status);
2010 cm_err = -ECONNRESET;
2011 break;
2012 case RDMA_CM_EVENT_ADDR_ERROR:
2013 case RDMA_CM_EVENT_ROUTE_ERROR:
2014 rtrs_wrn(s, "CM error (CM event: %s, err: %d)\n",
2015 rdma_event_msg(ev->event), ev->status);
2016 cm_err = -EHOSTUNREACH;
2017 break;
2018 case RDMA_CM_EVENT_DEVICE_REMOVAL:
2019 /*
2020 * Device removal is a special case. Queue close and return 0.
2021 */
2022 rtrs_clt_close_conns(clt_path, false);
2023 return 0;
2024 default:
2025 rtrs_err(s, "Unexpected RDMA CM error (CM event: %s, err: %d)\n",
2026 rdma_event_msg(ev->event), ev->status);
2027 cm_err = -ECONNRESET;
2028 break;
2029 }
2030
2031 if (cm_err) {
2032 /*
2033 * cm error makes sense only on connection establishing,
2034 * in other cases we rely on normal procedure of reconnecting.
2035 */
2036 flag_error_on_conn(con, cm_err);
2037 rtrs_rdma_error_recovery(con);
2038 }
2039
2040 return 0;
2041}
2042
2043static int create_cm(struct rtrs_clt_con *con)
2044{
2045 struct rtrs_path *s = con->c.path;
2046 struct rtrs_clt_path *clt_path = to_clt_path(s);
2047 struct rdma_cm_id *cm_id;
2048 int err;
2049
2050 cm_id = rdma_create_id(&init_net, rtrs_clt_rdma_cm_handler, con,
2051 clt_path->s.dst_addr.ss_family == AF_IB ?
2052 RDMA_PS_IB : RDMA_PS_TCP, IB_QPT_RC);
2053 if (IS_ERR(cm_id)) {
2054 err = PTR_ERR(cm_id);
2055 rtrs_err(s, "Failed to create CM ID, err: %d\n", err);
2056
2057 return err;
2058 }
2059 con->c.cm_id = cm_id;
2060 con->cm_err = 0;
2061 /* allow the port to be reused */
2062 err = rdma_set_reuseaddr(cm_id, 1);
2063 if (err != 0) {
2064 rtrs_err(s, "Set address reuse failed, err: %d\n", err);
2065 goto destroy_cm;
2066 }
2067 err = rdma_resolve_addr(cm_id, (struct sockaddr *)&clt_path->s.src_addr,
2068 (struct sockaddr *)&clt_path->s.dst_addr,
2069 RTRS_CONNECT_TIMEOUT_MS);
2070 if (err) {
2071 rtrs_err(s, "Failed to resolve address, err: %d\n", err);
2072 goto destroy_cm;
2073 }
2074 /*
2075 * Combine connection status and session events. This is needed
2076 * for waiting two possible cases: cm_err has something meaningful
2077 * or session state was really changed to error by device removal.
2078 */
2079 err = wait_event_interruptible_timeout(
2080 clt_path->state_wq,
2081 con->cm_err || clt_path->state != RTRS_CLT_CONNECTING,
2082 msecs_to_jiffies(RTRS_CONNECT_TIMEOUT_MS));
2083 if (err == 0 || err == -ERESTARTSYS) {
2084 if (err == 0)
2085 err = -ETIMEDOUT;
2086 /* Timedout or interrupted */
2087 goto errr;
2088 }
2089 if (con->cm_err < 0) {
2090 err = con->cm_err;
2091 goto errr;
2092 }
2093 if (READ_ONCE(clt_path->state) != RTRS_CLT_CONNECTING) {
2094 /* Device removal */
2095 err = -ECONNABORTED;
2096 goto errr;
2097 }
2098
2099 return 0;
2100
2101errr:
2102 stop_cm(con);
2103 mutex_lock(&con->con_mutex);
2104 destroy_con_cq_qp(con);
2105 mutex_unlock(&con->con_mutex);
2106destroy_cm:
2107 destroy_cm(con);
2108
2109 return err;
2110}
2111
2112static void rtrs_clt_path_up(struct rtrs_clt_path *clt_path)
2113{
2114 struct rtrs_clt_sess *clt = clt_path->clt;
2115 int up;
2116
2117 /*
2118 * We can fire RECONNECTED event only when all paths were
2119 * connected on rtrs_clt_open(), then each was disconnected
2120 * and the first one connected again. That's why this nasty
2121 * game with counter value.
2122 */
2123
2124 mutex_lock(&clt->paths_ev_mutex);
2125 up = ++clt->paths_up;
2126 /*
2127 * Here it is safe to access paths num directly since up counter
2128 * is greater than MAX_PATHS_NUM only while rtrs_clt_open() is
2129 * in progress, thus paths removals are impossible.
2130 */
2131 if (up > MAX_PATHS_NUM && up == MAX_PATHS_NUM + clt->paths_num)
2132 clt->paths_up = clt->paths_num;
2133 else if (up == 1)
2134 clt->link_ev(clt->priv, RTRS_CLT_LINK_EV_RECONNECTED);
2135 mutex_unlock(&clt->paths_ev_mutex);
2136
2137 /* Mark session as established */
2138 clt_path->established = true;
2139 clt_path->reconnect_attempts = 0;
2140 clt_path->stats->reconnects.successful_cnt++;
2141}
2142
2143static void rtrs_clt_path_down(struct rtrs_clt_path *clt_path)
2144{
2145 struct rtrs_clt_sess *clt = clt_path->clt;
2146
2147 if (!clt_path->established)
2148 return;
2149
2150 clt_path->established = false;
2151 mutex_lock(&clt->paths_ev_mutex);
2152 WARN_ON(!clt->paths_up);
2153 if (--clt->paths_up == 0)
2154 clt->link_ev(clt->priv, RTRS_CLT_LINK_EV_DISCONNECTED);
2155 mutex_unlock(&clt->paths_ev_mutex);
2156}
2157
2158static void rtrs_clt_stop_and_destroy_conns(struct rtrs_clt_path *clt_path)
2159{
2160 struct rtrs_clt_con *con;
2161 unsigned int cid;
2162
2163 WARN_ON(READ_ONCE(clt_path->state) == RTRS_CLT_CONNECTED);
2164
2165 /*
2166 * Possible race with rtrs_clt_open(), when DEVICE_REMOVAL comes
2167 * exactly in between. Start destroying after it finishes.
2168 */
2169 mutex_lock(&clt_path->init_mutex);
2170 mutex_unlock(&clt_path->init_mutex);
2171
2172 /*
2173 * All IO paths must observe !CONNECTED state before we
2174 * free everything.
2175 */
2176 synchronize_rcu();
2177
2178 rtrs_stop_hb(&clt_path->s);
2179
2180 /*
2181 * The order it utterly crucial: firstly disconnect and complete all
2182 * rdma requests with error (thus set in_use=false for requests),
2183 * then fail outstanding requests checking in_use for each, and
2184 * eventually notify upper layer about session disconnection.
2185 */
2186
2187 for (cid = 0; cid < clt_path->s.con_num; cid++) {
2188 if (!clt_path->s.con[cid])
2189 break;
2190 con = to_clt_con(clt_path->s.con[cid]);
2191 stop_cm(con);
2192 }
2193 fail_all_outstanding_reqs(clt_path);
2194 free_path_reqs(clt_path);
2195 rtrs_clt_path_down(clt_path);
2196
2197 /*
2198 * Wait for graceful shutdown, namely when peer side invokes
2199 * rdma_disconnect(). 'connected_cnt' is decremented only on
2200 * CM events, thus if other side had crashed and hb has detected
2201 * something is wrong, here we will stuck for exactly timeout ms,
2202 * since CM does not fire anything. That is fine, we are not in
2203 * hurry.
2204 */
2205 wait_event_timeout(clt_path->state_wq,
2206 !atomic_read(&clt_path->connected_cnt),
2207 msecs_to_jiffies(RTRS_CONNECT_TIMEOUT_MS));
2208
2209 for (cid = 0; cid < clt_path->s.con_num; cid++) {
2210 if (!clt_path->s.con[cid])
2211 break;
2212 con = to_clt_con(clt_path->s.con[cid]);
2213 mutex_lock(&con->con_mutex);
2214 destroy_con_cq_qp(con);
2215 mutex_unlock(&con->con_mutex);
2216 destroy_cm(con);
2217 destroy_con(con);
2218 }
2219}
2220
2221static void rtrs_clt_remove_path_from_arr(struct rtrs_clt_path *clt_path)
2222{
2223 struct rtrs_clt_sess *clt = clt_path->clt;
2224 struct rtrs_clt_path *next;
2225 bool wait_for_grace = false;
2226 int cpu;
2227
2228 mutex_lock(&clt->paths_mutex);
2229 list_del_rcu(&clt_path->s.entry);
2230
2231 /* Make sure everybody observes path removal. */
2232 synchronize_rcu();
2233
2234 /*
2235 * At this point nobody sees @sess in the list, but still we have
2236 * dangling pointer @pcpu_path which _can_ point to @sess. Since
2237 * nobody can observe @sess in the list, we guarantee that IO path
2238 * will not assign @sess to @pcpu_path, i.e. @pcpu_path can be equal
2239 * to @sess, but can never again become @sess.
2240 */
2241
2242 /*
2243 * Decrement paths number only after grace period, because
2244 * caller of do_each_path() must firstly observe list without
2245 * path and only then decremented paths number.
2246 *
2247 * Otherwise there can be the following situation:
2248 * o Two paths exist and IO is coming.
2249 * o One path is removed:
2250 * CPU#0 CPU#1
2251 * do_each_path(): rtrs_clt_remove_path_from_arr():
2252 * path = get_next_path()
2253 * ^^^ list_del_rcu(path)
2254 * [!CONNECTED path] clt->paths_num--
2255 * ^^^^^^^^^
2256 * load clt->paths_num from 2 to 1
2257 * ^^^^^^^^^
2258 * sees 1
2259 *
2260 * path is observed as !CONNECTED, but do_each_path() loop
2261 * ends, because expression i < clt->paths_num is false.
2262 */
2263 clt->paths_num--;
2264
2265 /*
2266 * Get @next connection from current @sess which is going to be
2267 * removed. If @sess is the last element, then @next is NULL.
2268 */
2269 rcu_read_lock();
2270 next = rtrs_clt_get_next_path_or_null(&clt->paths_list, clt_path);
2271 rcu_read_unlock();
2272
2273 /*
2274 * @pcpu paths can still point to the path which is going to be
2275 * removed, so change the pointer manually.
2276 */
2277 for_each_possible_cpu(cpu) {
2278 struct rtrs_clt_path __rcu **ppcpu_path;
2279
2280 ppcpu_path = per_cpu_ptr(clt->pcpu_path, cpu);
2281 if (rcu_dereference_protected(*ppcpu_path,
2282 lockdep_is_held(&clt->paths_mutex)) != clt_path)
2283 /*
2284 * synchronize_rcu() was called just after deleting
2285 * entry from the list, thus IO code path cannot
2286 * change pointer back to the pointer which is going
2287 * to be removed, we are safe here.
2288 */
2289 continue;
2290
2291 /*
2292 * We race with IO code path, which also changes pointer,
2293 * thus we have to be careful not to overwrite it.
2294 */
2295 if (try_cmpxchg((struct rtrs_clt_path **)ppcpu_path, &clt_path,
2296 next))
2297 /*
2298 * @ppcpu_path was successfully replaced with @next,
2299 * that means that someone could also pick up the
2300 * @sess and dereferencing it right now, so wait for
2301 * a grace period is required.
2302 */
2303 wait_for_grace = true;
2304 }
2305 if (wait_for_grace)
2306 synchronize_rcu();
2307
2308 mutex_unlock(&clt->paths_mutex);
2309}
2310
2311static void rtrs_clt_add_path_to_arr(struct rtrs_clt_path *clt_path)
2312{
2313 struct rtrs_clt_sess *clt = clt_path->clt;
2314
2315 mutex_lock(&clt->paths_mutex);
2316 clt->paths_num++;
2317
2318 list_add_tail_rcu(&clt_path->s.entry, &clt->paths_list);
2319 mutex_unlock(&clt->paths_mutex);
2320}
2321
2322static void rtrs_clt_close_work(struct work_struct *work)
2323{
2324 struct rtrs_clt_path *clt_path;
2325
2326 clt_path = container_of(work, struct rtrs_clt_path, close_work);
2327
2328 cancel_work_sync(&clt_path->err_recovery_work);
2329 cancel_delayed_work_sync(&clt_path->reconnect_dwork);
2330 rtrs_clt_stop_and_destroy_conns(clt_path);
2331 rtrs_clt_change_state_get_old(clt_path, RTRS_CLT_CLOSED, NULL);
2332}
2333
2334static int init_conns(struct rtrs_clt_path *clt_path)
2335{
2336 unsigned int cid;
2337 int err;
2338
2339 /*
2340 * On every new session connections increase reconnect counter
2341 * to avoid clashes with previous sessions not yet closed
2342 * sessions on a server side.
2343 */
2344 clt_path->s.recon_cnt++;
2345
2346 /* Establish all RDMA connections */
2347 for (cid = 0; cid < clt_path->s.con_num; cid++) {
2348 err = create_con(clt_path, cid);
2349 if (err)
2350 goto destroy;
2351
2352 err = create_cm(to_clt_con(clt_path->s.con[cid]));
2353 if (err) {
2354 destroy_con(to_clt_con(clt_path->s.con[cid]));
2355 goto destroy;
2356 }
2357 }
2358 err = alloc_path_reqs(clt_path);
2359 if (err)
2360 goto destroy;
2361
2362 rtrs_start_hb(&clt_path->s);
2363
2364 return 0;
2365
2366destroy:
2367 while (cid--) {
2368 struct rtrs_clt_con *con = to_clt_con(clt_path->s.con[cid]);
2369
2370 stop_cm(con);
2371
2372 mutex_lock(&con->con_mutex);
2373 destroy_con_cq_qp(con);
2374 mutex_unlock(&con->con_mutex);
2375 destroy_cm(con);
2376 destroy_con(con);
2377 }
2378 /*
2379 * If we've never taken async path and got an error, say,
2380 * doing rdma_resolve_addr(), switch to CONNECTION_ERR state
2381 * manually to keep reconnecting.
2382 */
2383 rtrs_clt_change_state_get_old(clt_path, RTRS_CLT_CONNECTING_ERR, NULL);
2384
2385 return err;
2386}
2387
2388static void rtrs_clt_info_req_done(struct ib_cq *cq, struct ib_wc *wc)
2389{
2390 struct rtrs_clt_con *con = to_clt_con(wc->qp->qp_context);
2391 struct rtrs_clt_path *clt_path = to_clt_path(con->c.path);
2392 struct rtrs_iu *iu;
2393
2394 iu = container_of(wc->wr_cqe, struct rtrs_iu, cqe);
2395 rtrs_iu_free(iu, clt_path->s.dev->ib_dev, 1);
2396
2397 if (wc->status != IB_WC_SUCCESS) {
2398 rtrs_err(clt_path->clt, "Path info request send failed: %s\n",
2399 ib_wc_status_msg(wc->status));
2400 rtrs_clt_change_state_get_old(clt_path, RTRS_CLT_CONNECTING_ERR, NULL);
2401 return;
2402 }
2403
2404 rtrs_clt_update_wc_stats(con);
2405}
2406
2407static int process_info_rsp(struct rtrs_clt_path *clt_path,
2408 const struct rtrs_msg_info_rsp *msg)
2409{
2410 unsigned int sg_cnt, total_len;
2411 int i, sgi;
2412
2413 sg_cnt = le16_to_cpu(msg->sg_cnt);
2414 if (!sg_cnt || (clt_path->queue_depth % sg_cnt)) {
2415 rtrs_err(clt_path->clt,
2416 "Incorrect sg_cnt %d, is not multiple\n",
2417 sg_cnt);
2418 return -EINVAL;
2419 }
2420
2421 /*
2422 * Check if IB immediate data size is enough to hold the mem_id and
2423 * the offset inside the memory chunk.
2424 */
2425 if ((ilog2(sg_cnt - 1) + 1) + (ilog2(clt_path->chunk_size - 1) + 1) >
2426 MAX_IMM_PAYL_BITS) {
2427 rtrs_err(clt_path->clt,
2428 "RDMA immediate size (%db) not enough to encode %d buffers of size %dB\n",
2429 MAX_IMM_PAYL_BITS, sg_cnt, clt_path->chunk_size);
2430 return -EINVAL;
2431 }
2432 total_len = 0;
2433 for (sgi = 0, i = 0; sgi < sg_cnt && i < clt_path->queue_depth; sgi++) {
2434 const struct rtrs_sg_desc *desc = &msg->desc[sgi];
2435 u32 len, rkey;
2436 u64 addr;
2437
2438 addr = le64_to_cpu(desc->addr);
2439 rkey = le32_to_cpu(desc->key);
2440 len = le32_to_cpu(desc->len);
2441
2442 total_len += len;
2443
2444 if (!len || (len % clt_path->chunk_size)) {
2445 rtrs_err(clt_path->clt, "Incorrect [%d].len %d\n",
2446 sgi,
2447 len);
2448 return -EINVAL;
2449 }
2450 for ( ; len && i < clt_path->queue_depth; i++) {
2451 clt_path->rbufs[i].addr = addr;
2452 clt_path->rbufs[i].rkey = rkey;
2453
2454 len -= clt_path->chunk_size;
2455 addr += clt_path->chunk_size;
2456 }
2457 }
2458 /* Sanity check */
2459 if (sgi != sg_cnt || i != clt_path->queue_depth) {
2460 rtrs_err(clt_path->clt,
2461 "Incorrect sg vector, not fully mapped\n");
2462 return -EINVAL;
2463 }
2464 if (total_len != clt_path->chunk_size * clt_path->queue_depth) {
2465 rtrs_err(clt_path->clt, "Incorrect total_len %d\n", total_len);
2466 return -EINVAL;
2467 }
2468
2469 return 0;
2470}
2471
2472static void rtrs_clt_info_rsp_done(struct ib_cq *cq, struct ib_wc *wc)
2473{
2474 struct rtrs_clt_con *con = to_clt_con(wc->qp->qp_context);
2475 struct rtrs_clt_path *clt_path = to_clt_path(con->c.path);
2476 struct rtrs_msg_info_rsp *msg;
2477 enum rtrs_clt_state state;
2478 struct rtrs_iu *iu;
2479 size_t rx_sz;
2480 int err;
2481
2482 state = RTRS_CLT_CONNECTING_ERR;
2483
2484 WARN_ON(con->c.cid);
2485 iu = container_of(wc->wr_cqe, struct rtrs_iu, cqe);
2486 if (wc->status != IB_WC_SUCCESS) {
2487 rtrs_err(clt_path->clt, "Path info response recv failed: %s\n",
2488 ib_wc_status_msg(wc->status));
2489 goto out;
2490 }
2491 WARN_ON(wc->opcode != IB_WC_RECV);
2492
2493 if (wc->byte_len < sizeof(*msg)) {
2494 rtrs_err(clt_path->clt, "Path info response is malformed: size %d\n",
2495 wc->byte_len);
2496 goto out;
2497 }
2498 ib_dma_sync_single_for_cpu(clt_path->s.dev->ib_dev, iu->dma_addr,
2499 iu->size, DMA_FROM_DEVICE);
2500 msg = iu->buf;
2501 if (le16_to_cpu(msg->type) != RTRS_MSG_INFO_RSP) {
2502 rtrs_err(clt_path->clt, "Path info response is malformed: type %d\n",
2503 le16_to_cpu(msg->type));
2504 goto out;
2505 }
2506 rx_sz = sizeof(*msg);
2507 rx_sz += sizeof(msg->desc[0]) * le16_to_cpu(msg->sg_cnt);
2508 if (wc->byte_len < rx_sz) {
2509 rtrs_err(clt_path->clt, "Path info response is malformed: size %d\n",
2510 wc->byte_len);
2511 goto out;
2512 }
2513 err = process_info_rsp(clt_path, msg);
2514 if (err)
2515 goto out;
2516
2517 err = post_recv_path(clt_path);
2518 if (err)
2519 goto out;
2520
2521 state = RTRS_CLT_CONNECTED;
2522
2523out:
2524 rtrs_clt_update_wc_stats(con);
2525 rtrs_iu_free(iu, clt_path->s.dev->ib_dev, 1);
2526 rtrs_clt_change_state_get_old(clt_path, state, NULL);
2527}
2528
2529static int rtrs_send_path_info(struct rtrs_clt_path *clt_path)
2530{
2531 struct rtrs_clt_con *usr_con = to_clt_con(clt_path->s.con[0]);
2532 struct rtrs_msg_info_req *msg;
2533 struct rtrs_iu *tx_iu, *rx_iu;
2534 size_t rx_sz;
2535 int err;
2536
2537 rx_sz = sizeof(struct rtrs_msg_info_rsp);
2538 rx_sz += sizeof(struct rtrs_sg_desc) * clt_path->queue_depth;
2539
2540 tx_iu = rtrs_iu_alloc(1, sizeof(struct rtrs_msg_info_req), GFP_KERNEL,
2541 clt_path->s.dev->ib_dev, DMA_TO_DEVICE,
2542 rtrs_clt_info_req_done);
2543 rx_iu = rtrs_iu_alloc(1, rx_sz, GFP_KERNEL, clt_path->s.dev->ib_dev,
2544 DMA_FROM_DEVICE, rtrs_clt_info_rsp_done);
2545 if (!tx_iu || !rx_iu) {
2546 err = -ENOMEM;
2547 goto out;
2548 }
2549 /* Prepare for getting info response */
2550 err = rtrs_iu_post_recv(&usr_con->c, rx_iu);
2551 if (err) {
2552 rtrs_err(clt_path->clt, "rtrs_iu_post_recv(), err: %d\n", err);
2553 goto out;
2554 }
2555 rx_iu = NULL;
2556
2557 msg = tx_iu->buf;
2558 msg->type = cpu_to_le16(RTRS_MSG_INFO_REQ);
2559 memcpy(msg->pathname, clt_path->s.sessname, sizeof(msg->pathname));
2560
2561 ib_dma_sync_single_for_device(clt_path->s.dev->ib_dev,
2562 tx_iu->dma_addr,
2563 tx_iu->size, DMA_TO_DEVICE);
2564
2565 /* Send info request */
2566 err = rtrs_iu_post_send(&usr_con->c, tx_iu, sizeof(*msg), NULL);
2567 if (err) {
2568 rtrs_err(clt_path->clt, "rtrs_iu_post_send(), err: %d\n", err);
2569 goto out;
2570 }
2571 tx_iu = NULL;
2572
2573 /* Wait for state change */
2574 wait_event_interruptible_timeout(clt_path->state_wq,
2575 clt_path->state != RTRS_CLT_CONNECTING,
2576 msecs_to_jiffies(
2577 RTRS_CONNECT_TIMEOUT_MS));
2578 if (READ_ONCE(clt_path->state) != RTRS_CLT_CONNECTED) {
2579 if (READ_ONCE(clt_path->state) == RTRS_CLT_CONNECTING_ERR)
2580 err = -ECONNRESET;
2581 else
2582 err = -ETIMEDOUT;
2583 }
2584
2585out:
2586 if (tx_iu)
2587 rtrs_iu_free(tx_iu, clt_path->s.dev->ib_dev, 1);
2588 if (rx_iu)
2589 rtrs_iu_free(rx_iu, clt_path->s.dev->ib_dev, 1);
2590 if (err)
2591 /* If we've never taken async path because of malloc problems */
2592 rtrs_clt_change_state_get_old(clt_path,
2593 RTRS_CLT_CONNECTING_ERR, NULL);
2594
2595 return err;
2596}
2597
2598/**
2599 * init_path() - establishes all path connections and does handshake
2600 * @clt_path: client path.
2601 * In case of error full close or reconnect procedure should be taken,
2602 * because reconnect or close async works can be started.
2603 */
2604static int init_path(struct rtrs_clt_path *clt_path)
2605{
2606 int err;
2607 char str[NAME_MAX];
2608 struct rtrs_addr path = {
2609 .src = &clt_path->s.src_addr,
2610 .dst = &clt_path->s.dst_addr,
2611 };
2612
2613 rtrs_addr_to_str(&path, str, sizeof(str));
2614
2615 mutex_lock(&clt_path->init_mutex);
2616 err = init_conns(clt_path);
2617 if (err) {
2618 rtrs_err(clt_path->clt,
2619 "init_conns() failed: err=%d path=%s [%s:%u]\n", err,
2620 str, clt_path->hca_name, clt_path->hca_port);
2621 goto out;
2622 }
2623 err = rtrs_send_path_info(clt_path);
2624 if (err) {
2625 rtrs_err(clt_path->clt,
2626 "rtrs_send_path_info() failed: err=%d path=%s [%s:%u]\n",
2627 err, str, clt_path->hca_name, clt_path->hca_port);
2628 goto out;
2629 }
2630 rtrs_clt_path_up(clt_path);
2631out:
2632 mutex_unlock(&clt_path->init_mutex);
2633
2634 return err;
2635}
2636
2637static void rtrs_clt_reconnect_work(struct work_struct *work)
2638{
2639 struct rtrs_clt_path *clt_path;
2640 struct rtrs_clt_sess *clt;
2641 int err;
2642
2643 clt_path = container_of(to_delayed_work(work), struct rtrs_clt_path,
2644 reconnect_dwork);
2645 clt = clt_path->clt;
2646
2647 trace_rtrs_clt_reconnect_work(clt_path);
2648
2649 if (READ_ONCE(clt_path->state) != RTRS_CLT_RECONNECTING)
2650 return;
2651
2652 if (clt_path->reconnect_attempts >= clt->max_reconnect_attempts) {
2653 /* Close a path completely if max attempts is reached */
2654 rtrs_clt_close_conns(clt_path, false);
2655 return;
2656 }
2657 clt_path->reconnect_attempts++;
2658
2659 msleep(RTRS_RECONNECT_BACKOFF);
2660 if (rtrs_clt_change_state_get_old(clt_path, RTRS_CLT_CONNECTING, NULL)) {
2661 err = init_path(clt_path);
2662 if (err)
2663 goto reconnect_again;
2664 }
2665
2666 return;
2667
2668reconnect_again:
2669 if (rtrs_clt_change_state_get_old(clt_path, RTRS_CLT_RECONNECTING, NULL)) {
2670 clt_path->stats->reconnects.fail_cnt++;
2671 queue_work(rtrs_wq, &clt_path->err_recovery_work);
2672 }
2673}
2674
2675static void rtrs_clt_dev_release(struct device *dev)
2676{
2677 struct rtrs_clt_sess *clt = container_of(dev, struct rtrs_clt_sess,
2678 dev);
2679
2680 mutex_destroy(&clt->paths_ev_mutex);
2681 mutex_destroy(&clt->paths_mutex);
2682 kfree(clt);
2683}
2684
2685static struct rtrs_clt_sess *alloc_clt(const char *sessname, size_t paths_num,
2686 u16 port, size_t pdu_sz, void *priv,
2687 void (*link_ev)(void *priv,
2688 enum rtrs_clt_link_ev ev),
2689 unsigned int reconnect_delay_sec,
2690 unsigned int max_reconnect_attempts)
2691{
2692 struct rtrs_clt_sess *clt;
2693 int err;
2694
2695 if (!paths_num || paths_num > MAX_PATHS_NUM)
2696 return ERR_PTR(-EINVAL);
2697
2698 if (strlen(sessname) >= sizeof(clt->sessname))
2699 return ERR_PTR(-EINVAL);
2700
2701 clt = kzalloc(sizeof(*clt), GFP_KERNEL);
2702 if (!clt)
2703 return ERR_PTR(-ENOMEM);
2704
2705 clt->pcpu_path = alloc_percpu(typeof(*clt->pcpu_path));
2706 if (!clt->pcpu_path) {
2707 kfree(clt);
2708 return ERR_PTR(-ENOMEM);
2709 }
2710
2711 clt->dev.class = rtrs_clt_dev_class;
2712 clt->dev.release = rtrs_clt_dev_release;
2713 uuid_gen(&clt->paths_uuid);
2714 INIT_LIST_HEAD_RCU(&clt->paths_list);
2715 clt->paths_num = paths_num;
2716 clt->paths_up = MAX_PATHS_NUM;
2717 clt->port = port;
2718 clt->pdu_sz = pdu_sz;
2719 clt->max_segments = RTRS_MAX_SEGMENTS;
2720 clt->reconnect_delay_sec = reconnect_delay_sec;
2721 clt->max_reconnect_attempts = max_reconnect_attempts;
2722 clt->priv = priv;
2723 clt->link_ev = link_ev;
2724 clt->mp_policy = MP_POLICY_MIN_INFLIGHT;
2725 strscpy(clt->sessname, sessname, sizeof(clt->sessname));
2726 init_waitqueue_head(&clt->permits_wait);
2727 mutex_init(&clt->paths_ev_mutex);
2728 mutex_init(&clt->paths_mutex);
2729 device_initialize(&clt->dev);
2730
2731 err = dev_set_name(&clt->dev, "%s", sessname);
2732 if (err)
2733 goto err_put;
2734
2735 /*
2736 * Suppress user space notification until
2737 * sysfs files are created
2738 */
2739 dev_set_uevent_suppress(&clt->dev, true);
2740 err = device_add(&clt->dev);
2741 if (err)
2742 goto err_put;
2743
2744 clt->kobj_paths = kobject_create_and_add("paths", &clt->dev.kobj);
2745 if (!clt->kobj_paths) {
2746 err = -ENOMEM;
2747 goto err_del;
2748 }
2749 err = rtrs_clt_create_sysfs_root_files(clt);
2750 if (err) {
2751 kobject_del(clt->kobj_paths);
2752 kobject_put(clt->kobj_paths);
2753 goto err_del;
2754 }
2755 dev_set_uevent_suppress(&clt->dev, false);
2756 kobject_uevent(&clt->dev.kobj, KOBJ_ADD);
2757
2758 return clt;
2759err_del:
2760 device_del(&clt->dev);
2761err_put:
2762 free_percpu(clt->pcpu_path);
2763 put_device(&clt->dev);
2764 return ERR_PTR(err);
2765}
2766
2767static void free_clt(struct rtrs_clt_sess *clt)
2768{
2769 free_percpu(clt->pcpu_path);
2770
2771 /*
2772 * release callback will free clt and destroy mutexes in last put
2773 */
2774 device_unregister(&clt->dev);
2775}
2776
2777/**
2778 * rtrs_clt_open() - Open a path to an RTRS server
2779 * @ops: holds the link event callback and the private pointer.
2780 * @pathname: name of the path to an RTRS server
2781 * @paths: Paths to be established defined by their src and dst addresses
2782 * @paths_num: Number of elements in the @paths array
2783 * @port: port to be used by the RTRS session
2784 * @pdu_sz: Size of extra payload which can be accessed after permit allocation.
2785 * @reconnect_delay_sec: time between reconnect tries
2786 * @max_reconnect_attempts: Number of times to reconnect on error before giving
2787 * up, 0 for * disabled, -1 for forever
2788 * @nr_poll_queues: number of polling mode connection using IB_POLL_DIRECT flag
2789 *
2790 * Starts session establishment with the rtrs_server. The function can block
2791 * up to ~2000ms before it returns.
2792 *
2793 * Return a valid pointer on success otherwise PTR_ERR.
2794 */
2795struct rtrs_clt_sess *rtrs_clt_open(struct rtrs_clt_ops *ops,
2796 const char *pathname,
2797 const struct rtrs_addr *paths,
2798 size_t paths_num, u16 port,
2799 size_t pdu_sz, u8 reconnect_delay_sec,
2800 s16 max_reconnect_attempts, u32 nr_poll_queues)
2801{
2802 struct rtrs_clt_path *clt_path, *tmp;
2803 struct rtrs_clt_sess *clt;
2804 int err, i;
2805
2806 if (strchr(pathname, '/') || strchr(pathname, '.')) {
2807 pr_err("pathname cannot contain / and .\n");
2808 err = -EINVAL;
2809 goto out;
2810 }
2811
2812 clt = alloc_clt(pathname, paths_num, port, pdu_sz, ops->priv,
2813 ops->link_ev,
2814 reconnect_delay_sec,
2815 max_reconnect_attempts);
2816 if (IS_ERR(clt)) {
2817 err = PTR_ERR(clt);
2818 goto out;
2819 }
2820 for (i = 0; i < paths_num; i++) {
2821 struct rtrs_clt_path *clt_path;
2822
2823 clt_path = alloc_path(clt, &paths[i], nr_cpu_ids,
2824 nr_poll_queues);
2825 if (IS_ERR(clt_path)) {
2826 err = PTR_ERR(clt_path);
2827 goto close_all_path;
2828 }
2829 if (!i)
2830 clt_path->for_new_clt = 1;
2831 list_add_tail_rcu(&clt_path->s.entry, &clt->paths_list);
2832
2833 err = init_path(clt_path);
2834 if (err) {
2835 list_del_rcu(&clt_path->s.entry);
2836 rtrs_clt_close_conns(clt_path, true);
2837 free_percpu(clt_path->stats->pcpu_stats);
2838 kfree(clt_path->stats);
2839 free_path(clt_path);
2840 goto close_all_path;
2841 }
2842
2843 err = rtrs_clt_create_path_files(clt_path);
2844 if (err) {
2845 list_del_rcu(&clt_path->s.entry);
2846 rtrs_clt_close_conns(clt_path, true);
2847 free_percpu(clt_path->stats->pcpu_stats);
2848 kfree(clt_path->stats);
2849 free_path(clt_path);
2850 goto close_all_path;
2851 }
2852 }
2853 err = alloc_permits(clt);
2854 if (err)
2855 goto close_all_path;
2856
2857 return clt;
2858
2859close_all_path:
2860 list_for_each_entry_safe(clt_path, tmp, &clt->paths_list, s.entry) {
2861 rtrs_clt_destroy_path_files(clt_path, NULL);
2862 rtrs_clt_close_conns(clt_path, true);
2863 kobject_put(&clt_path->kobj);
2864 }
2865 rtrs_clt_destroy_sysfs_root(clt);
2866 free_clt(clt);
2867
2868out:
2869 return ERR_PTR(err);
2870}
2871EXPORT_SYMBOL(rtrs_clt_open);
2872
2873/**
2874 * rtrs_clt_close() - Close a path
2875 * @clt: Session handle. Session is freed upon return.
2876 */
2877void rtrs_clt_close(struct rtrs_clt_sess *clt)
2878{
2879 struct rtrs_clt_path *clt_path, *tmp;
2880
2881 /* Firstly forbid sysfs access */
2882 rtrs_clt_destroy_sysfs_root(clt);
2883
2884 /* Now it is safe to iterate over all paths without locks */
2885 list_for_each_entry_safe(clt_path, tmp, &clt->paths_list, s.entry) {
2886 rtrs_clt_close_conns(clt_path, true);
2887 rtrs_clt_destroy_path_files(clt_path, NULL);
2888 kobject_put(&clt_path->kobj);
2889 }
2890 free_permits(clt);
2891 free_clt(clt);
2892}
2893EXPORT_SYMBOL(rtrs_clt_close);
2894
2895int rtrs_clt_reconnect_from_sysfs(struct rtrs_clt_path *clt_path)
2896{
2897 enum rtrs_clt_state old_state;
2898 int err = -EBUSY;
2899 bool changed;
2900
2901 changed = rtrs_clt_change_state_get_old(clt_path,
2902 RTRS_CLT_RECONNECTING,
2903 &old_state);
2904 if (changed) {
2905 clt_path->reconnect_attempts = 0;
2906 rtrs_clt_stop_and_destroy_conns(clt_path);
2907 queue_delayed_work(rtrs_wq, &clt_path->reconnect_dwork, 0);
2908 }
2909 if (changed || old_state == RTRS_CLT_RECONNECTING) {
2910 /*
2911 * flush_delayed_work() queues pending work for immediate
2912 * execution, so do the flush if we have queued something
2913 * right now or work is pending.
2914 */
2915 flush_delayed_work(&clt_path->reconnect_dwork);
2916 err = (READ_ONCE(clt_path->state) ==
2917 RTRS_CLT_CONNECTED ? 0 : -ENOTCONN);
2918 }
2919
2920 return err;
2921}
2922
2923int rtrs_clt_remove_path_from_sysfs(struct rtrs_clt_path *clt_path,
2924 const struct attribute *sysfs_self)
2925{
2926 enum rtrs_clt_state old_state;
2927 bool changed;
2928
2929 /*
2930 * Continue stopping path till state was changed to DEAD or
2931 * state was observed as DEAD:
2932 * 1. State was changed to DEAD - we were fast and nobody
2933 * invoked rtrs_clt_reconnect(), which can again start
2934 * reconnecting.
2935 * 2. State was observed as DEAD - we have someone in parallel
2936 * removing the path.
2937 */
2938 do {
2939 rtrs_clt_close_conns(clt_path, true);
2940 changed = rtrs_clt_change_state_get_old(clt_path,
2941 RTRS_CLT_DEAD,
2942 &old_state);
2943 } while (!changed && old_state != RTRS_CLT_DEAD);
2944
2945 if (changed) {
2946 rtrs_clt_remove_path_from_arr(clt_path);
2947 rtrs_clt_destroy_path_files(clt_path, sysfs_self);
2948 kobject_put(&clt_path->kobj);
2949 }
2950
2951 return 0;
2952}
2953
2954void rtrs_clt_set_max_reconnect_attempts(struct rtrs_clt_sess *clt, int value)
2955{
2956 clt->max_reconnect_attempts = (unsigned int)value;
2957}
2958
2959int rtrs_clt_get_max_reconnect_attempts(const struct rtrs_clt_sess *clt)
2960{
2961 return (int)clt->max_reconnect_attempts;
2962}
2963
2964/**
2965 * rtrs_clt_request() - Request data transfer to/from server via RDMA.
2966 *
2967 * @dir: READ/WRITE
2968 * @ops: callback function to be called as confirmation, and the pointer.
2969 * @clt: Session
2970 * @permit: Preallocated permit
2971 * @vec: Message that is sent to server together with the request.
2972 * Sum of len of all @vec elements limited to <= IO_MSG_SIZE.
2973 * Since the msg is copied internally it can be allocated on stack.
2974 * @nr: Number of elements in @vec.
2975 * @data_len: length of data sent to/from server
2976 * @sg: Pages to be sent/received to/from server.
2977 * @sg_cnt: Number of elements in the @sg
2978 *
2979 * Return:
2980 * 0: Success
2981 * <0: Error
2982 *
2983 * On dir=READ rtrs client will request a data transfer from Server to client.
2984 * The data that the server will respond with will be stored in @sg when
2985 * the user receives an %RTRS_CLT_RDMA_EV_RDMA_REQUEST_WRITE_COMPL event.
2986 * On dir=WRITE rtrs client will rdma write data in sg to server side.
2987 */
2988int rtrs_clt_request(int dir, struct rtrs_clt_req_ops *ops,
2989 struct rtrs_clt_sess *clt, struct rtrs_permit *permit,
2990 const struct kvec *vec, size_t nr, size_t data_len,
2991 struct scatterlist *sg, unsigned int sg_cnt)
2992{
2993 struct rtrs_clt_io_req *req;
2994 struct rtrs_clt_path *clt_path;
2995
2996 enum dma_data_direction dma_dir;
2997 int err = -ECONNABORTED, i;
2998 size_t usr_len, hdr_len;
2999 struct path_it it;
3000
3001 /* Get kvec length */
3002 for (i = 0, usr_len = 0; i < nr; i++)
3003 usr_len += vec[i].iov_len;
3004
3005 if (dir == READ) {
3006 hdr_len = sizeof(struct rtrs_msg_rdma_read) +
3007 sg_cnt * sizeof(struct rtrs_sg_desc);
3008 dma_dir = DMA_FROM_DEVICE;
3009 } else {
3010 hdr_len = sizeof(struct rtrs_msg_rdma_write);
3011 dma_dir = DMA_TO_DEVICE;
3012 }
3013
3014 rcu_read_lock();
3015 for (path_it_init(&it, clt);
3016 (clt_path = it.next_path(&it)) && it.i < it.clt->paths_num; it.i++) {
3017 if (READ_ONCE(clt_path->state) != RTRS_CLT_CONNECTED)
3018 continue;
3019
3020 if (usr_len + hdr_len > clt_path->max_hdr_size) {
3021 rtrs_wrn_rl(clt_path->clt,
3022 "%s request failed, user message size is %zu and header length %zu, but max size is %u\n",
3023 dir == READ ? "Read" : "Write",
3024 usr_len, hdr_len, clt_path->max_hdr_size);
3025 err = -EMSGSIZE;
3026 break;
3027 }
3028 req = rtrs_clt_get_req(clt_path, ops->conf_fn, permit, ops->priv,
3029 vec, usr_len, sg, sg_cnt, data_len,
3030 dma_dir);
3031 if (dir == READ)
3032 err = rtrs_clt_read_req(req);
3033 else
3034 err = rtrs_clt_write_req(req);
3035 if (err) {
3036 req->in_use = false;
3037 continue;
3038 }
3039 /* Success path */
3040 break;
3041 }
3042 path_it_deinit(&it);
3043 rcu_read_unlock();
3044
3045 return err;
3046}
3047EXPORT_SYMBOL(rtrs_clt_request);
3048
3049int rtrs_clt_rdma_cq_direct(struct rtrs_clt_sess *clt, unsigned int index)
3050{
3051 /* If no path, return -1 for block layer not to try again */
3052 int cnt = -1;
3053 struct rtrs_con *con;
3054 struct rtrs_clt_path *clt_path;
3055 struct path_it it;
3056
3057 rcu_read_lock();
3058 for (path_it_init(&it, clt);
3059 (clt_path = it.next_path(&it)) && it.i < it.clt->paths_num; it.i++) {
3060 if (READ_ONCE(clt_path->state) != RTRS_CLT_CONNECTED)
3061 continue;
3062
3063 con = clt_path->s.con[index + 1];
3064 cnt = ib_process_cq_direct(con->cq, -1);
3065 if (cnt)
3066 break;
3067 }
3068 path_it_deinit(&it);
3069 rcu_read_unlock();
3070
3071 return cnt;
3072}
3073EXPORT_SYMBOL(rtrs_clt_rdma_cq_direct);
3074
3075/**
3076 * rtrs_clt_query() - queries RTRS session attributes
3077 *@clt: session pointer
3078 *@attr: query results for session attributes.
3079 * Returns:
3080 * 0 on success
3081 * -ECOMM no connection to the server
3082 */
3083int rtrs_clt_query(struct rtrs_clt_sess *clt, struct rtrs_attrs *attr)
3084{
3085 if (!rtrs_clt_is_connected(clt))
3086 return -ECOMM;
3087
3088 attr->queue_depth = clt->queue_depth;
3089 attr->max_segments = clt->max_segments;
3090 /* Cap max_io_size to min of remote buffer size and the fr pages */
3091 attr->max_io_size = min_t(int, clt->max_io_size,
3092 clt->max_segments * SZ_4K);
3093
3094 return 0;
3095}
3096EXPORT_SYMBOL(rtrs_clt_query);
3097
3098int rtrs_clt_create_path_from_sysfs(struct rtrs_clt_sess *clt,
3099 struct rtrs_addr *addr)
3100{
3101 struct rtrs_clt_path *clt_path;
3102 int err;
3103
3104 clt_path = alloc_path(clt, addr, nr_cpu_ids, 0);
3105 if (IS_ERR(clt_path))
3106 return PTR_ERR(clt_path);
3107
3108 mutex_lock(&clt->paths_mutex);
3109 if (clt->paths_num == 0) {
3110 /*
3111 * When all the paths are removed for a session,
3112 * the addition of the first path is like a new session for
3113 * the storage server
3114 */
3115 clt_path->for_new_clt = 1;
3116 }
3117
3118 mutex_unlock(&clt->paths_mutex);
3119
3120 /*
3121 * It is totally safe to add path in CONNECTING state: coming
3122 * IO will never grab it. Also it is very important to add
3123 * path before init, since init fires LINK_CONNECTED event.
3124 */
3125 rtrs_clt_add_path_to_arr(clt_path);
3126
3127 err = init_path(clt_path);
3128 if (err)
3129 goto close_path;
3130
3131 err = rtrs_clt_create_path_files(clt_path);
3132 if (err)
3133 goto close_path;
3134
3135 return 0;
3136
3137close_path:
3138 rtrs_clt_remove_path_from_arr(clt_path);
3139 rtrs_clt_close_conns(clt_path, true);
3140 free_percpu(clt_path->stats->pcpu_stats);
3141 kfree(clt_path->stats);
3142 free_path(clt_path);
3143
3144 return err;
3145}
3146
3147static int rtrs_clt_ib_dev_init(struct rtrs_ib_dev *dev)
3148{
3149 if (!(dev->ib_dev->attrs.device_cap_flags &
3150 IB_DEVICE_MEM_MGT_EXTENSIONS)) {
3151 pr_err("Memory registrations not supported.\n");
3152 return -ENOTSUPP;
3153 }
3154
3155 return 0;
3156}
3157
3158static const struct rtrs_rdma_dev_pd_ops dev_pd_ops = {
3159 .init = rtrs_clt_ib_dev_init
3160};
3161
3162static int __init rtrs_client_init(void)
3163{
3164 rtrs_rdma_dev_pd_init(0, &dev_pd);
3165
3166 rtrs_clt_dev_class = class_create(THIS_MODULE, "rtrs-client");
3167 if (IS_ERR(rtrs_clt_dev_class)) {
3168 pr_err("Failed to create rtrs-client dev class\n");
3169 return PTR_ERR(rtrs_clt_dev_class);
3170 }
3171 rtrs_wq = alloc_workqueue("rtrs_client_wq", 0, 0);
3172 if (!rtrs_wq) {
3173 class_destroy(rtrs_clt_dev_class);
3174 return -ENOMEM;
3175 }
3176
3177 return 0;
3178}
3179
3180static void __exit rtrs_client_exit(void)
3181{
3182 destroy_workqueue(rtrs_wq);
3183 class_destroy(rtrs_clt_dev_class);
3184 rtrs_rdma_dev_pd_deinit(&dev_pd);
3185}
3186
3187module_init(rtrs_client_init);
3188module_exit(rtrs_client_exit);