1 // SPDX-License-Identifier: GPL-2.0-or-later
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.
11 #define pr_fmt(fmt) KBUILD_MODNAME " L" __stringify(__LINE__) ": " fmt
13 #include <linux/module.h>
14 #include <linux/rculist.h>
15 #include <linux/random.h>
20 #define RTRS_CONNECT_TIMEOUT_MS 30000
22 * Wait a bit before trying to reconnect after a failure
23 * in order to give server time to finish clean up which
24 * leads to "false positives" failed reconnect attempts
26 #define RTRS_RECONNECT_BACKOFF 1000
28 * Wait for additional random time between 0 and 8 seconds
29 * before starting to reconnect to avoid clients reconnecting
30 * all at once in case of a major network outage
32 #define RTRS_RECONNECT_SEED 8
34 MODULE_DESCRIPTION("RDMA Transport Client");
35 MODULE_LICENSE("GPL");
37 static const struct rtrs_rdma_dev_pd_ops dev_pd_ops;
38 static struct rtrs_rdma_dev_pd dev_pd = {
42 static struct workqueue_struct *rtrs_wq;
43 static struct class *rtrs_clt_dev_class;
45 static inline bool rtrs_clt_is_connected(const struct rtrs_clt *clt)
47 struct rtrs_clt_sess *sess;
48 bool connected = false;
51 list_for_each_entry_rcu(sess, &clt->paths_list, s.entry)
52 connected |= READ_ONCE(sess->state) == RTRS_CLT_CONNECTED;
58 static struct rtrs_permit *
59 __rtrs_get_permit(struct rtrs_clt *clt, enum rtrs_clt_con_type con_type)
61 size_t max_depth = clt->queue_depth;
62 struct rtrs_permit *permit;
66 * Adapted from null_blk get_tag(). Callers from different cpus may
67 * grab the same bit, since find_first_zero_bit is not atomic.
68 * But then the test_and_set_bit_lock will fail for all the
69 * callers but one, so that they will loop again.
70 * This way an explicit spinlock is not required.
73 bit = find_first_zero_bit(clt->permits_map, max_depth);
74 if (unlikely(bit >= max_depth))
76 } while (unlikely(test_and_set_bit_lock(bit, clt->permits_map)));
78 permit = get_permit(clt, bit);
79 WARN_ON(permit->mem_id != bit);
80 permit->cpu_id = raw_smp_processor_id();
81 permit->con_type = con_type;
86 static inline void __rtrs_put_permit(struct rtrs_clt *clt,
87 struct rtrs_permit *permit)
89 clear_bit_unlock(permit->mem_id, clt->permits_map);
93 * rtrs_clt_get_permit() - allocates permit for future RDMA operation
94 * @clt: Current session
95 * @con_type: Type of connection to use with the permit
96 * @can_wait: Wait type
99 * Allocates permit for the following RDMA operation. Permit is used
100 * to preallocate all resources and to propagate memory pressure
104 * Can sleep if @wait == RTRS_TAG_WAIT
106 struct rtrs_permit *rtrs_clt_get_permit(struct rtrs_clt *clt,
107 enum rtrs_clt_con_type con_type,
110 struct rtrs_permit *permit;
113 permit = __rtrs_get_permit(clt, con_type);
114 if (likely(permit) || !can_wait)
118 prepare_to_wait(&clt->permits_wait, &wait,
119 TASK_UNINTERRUPTIBLE);
120 permit = __rtrs_get_permit(clt, con_type);
127 finish_wait(&clt->permits_wait, &wait);
131 EXPORT_SYMBOL(rtrs_clt_get_permit);
134 * rtrs_clt_put_permit() - puts allocated permit
135 * @clt: Current session
136 * @permit: Permit to be freed
141 void rtrs_clt_put_permit(struct rtrs_clt *clt, struct rtrs_permit *permit)
143 if (WARN_ON(!test_bit(permit->mem_id, clt->permits_map)))
146 __rtrs_put_permit(clt, permit);
149 * rtrs_clt_get_permit() adds itself to the &clt->permits_wait list
150 * before calling schedule(). So if rtrs_clt_get_permit() is sleeping
151 * it must have added itself to &clt->permits_wait before
152 * __rtrs_put_permit() finished.
153 * Hence it is safe to guard wake_up() with a waitqueue_active() test.
155 if (waitqueue_active(&clt->permits_wait))
156 wake_up(&clt->permits_wait);
158 EXPORT_SYMBOL(rtrs_clt_put_permit);
160 void *rtrs_permit_to_pdu(struct rtrs_permit *permit)
164 EXPORT_SYMBOL(rtrs_permit_to_pdu);
167 * rtrs_permit_to_clt_con() - returns RDMA connection pointer by the permit
168 * @sess: client session pointer
169 * @permit: permit for the allocation of the RDMA buffer
171 * IO connection starts from 1.
172 * 0 connection is for user messages.
175 struct rtrs_clt_con *rtrs_permit_to_clt_con(struct rtrs_clt_sess *sess,
176 struct rtrs_permit *permit)
180 if (likely(permit->con_type == RTRS_IO_CON))
181 id = (permit->cpu_id % (sess->s.con_num - 1)) + 1;
183 return to_clt_con(sess->s.con[id]);
187 * __rtrs_clt_change_state() - change the session state through session state
190 * @sess: client session to change the state of.
191 * @new_state: state to change to.
193 * returns true if successful, false if the requested state can not be set.
196 * state_wq lock must be hold.
198 static bool __rtrs_clt_change_state(struct rtrs_clt_sess *sess,
199 enum rtrs_clt_state new_state)
201 enum rtrs_clt_state old_state;
202 bool changed = false;
204 lockdep_assert_held(&sess->state_wq.lock);
206 old_state = sess->state;
208 case RTRS_CLT_CONNECTING:
210 case RTRS_CLT_RECONNECTING:
217 case RTRS_CLT_RECONNECTING:
219 case RTRS_CLT_CONNECTED:
220 case RTRS_CLT_CONNECTING_ERR:
221 case RTRS_CLT_CLOSED:
228 case RTRS_CLT_CONNECTED:
230 case RTRS_CLT_CONNECTING:
237 case RTRS_CLT_CONNECTING_ERR:
239 case RTRS_CLT_CONNECTING:
246 case RTRS_CLT_CLOSING:
248 case RTRS_CLT_CONNECTING:
249 case RTRS_CLT_CONNECTING_ERR:
250 case RTRS_CLT_RECONNECTING:
251 case RTRS_CLT_CONNECTED:
258 case RTRS_CLT_CLOSED:
260 case RTRS_CLT_CLOSING:
269 case RTRS_CLT_CLOSED:
280 sess->state = new_state;
281 wake_up_locked(&sess->state_wq);
287 static bool rtrs_clt_change_state_from_to(struct rtrs_clt_sess *sess,
288 enum rtrs_clt_state old_state,
289 enum rtrs_clt_state new_state)
291 bool changed = false;
293 spin_lock_irq(&sess->state_wq.lock);
294 if (sess->state == old_state)
295 changed = __rtrs_clt_change_state(sess, new_state);
296 spin_unlock_irq(&sess->state_wq.lock);
301 static void rtrs_rdma_error_recovery(struct rtrs_clt_con *con)
303 struct rtrs_clt_sess *sess = to_clt_sess(con->c.sess);
305 if (rtrs_clt_change_state_from_to(sess,
307 RTRS_CLT_RECONNECTING)) {
308 struct rtrs_clt *clt = sess->clt;
309 unsigned int delay_ms;
312 * Normal scenario, reconnect if we were successfully connected
314 delay_ms = clt->reconnect_delay_sec * 1000;
315 queue_delayed_work(rtrs_wq, &sess->reconnect_dwork,
316 msecs_to_jiffies(delay_ms +
317 prandom_u32() % RTRS_RECONNECT_SEED));
320 * Error can happen just on establishing new connection,
321 * so notify waiter with error state, waiter is responsible
322 * for cleaning the rest and reconnect if needed.
324 rtrs_clt_change_state_from_to(sess,
326 RTRS_CLT_CONNECTING_ERR);
330 static void rtrs_clt_fast_reg_done(struct ib_cq *cq, struct ib_wc *wc)
332 struct rtrs_clt_con *con = cq->cq_context;
334 if (unlikely(wc->status != IB_WC_SUCCESS)) {
335 rtrs_err(con->c.sess, "Failed IB_WR_REG_MR: %s\n",
336 ib_wc_status_msg(wc->status));
337 rtrs_rdma_error_recovery(con);
341 static struct ib_cqe fast_reg_cqe = {
342 .done = rtrs_clt_fast_reg_done
345 static void complete_rdma_req(struct rtrs_clt_io_req *req, int errno,
346 bool notify, bool can_wait);
348 static void rtrs_clt_inv_rkey_done(struct ib_cq *cq, struct ib_wc *wc)
350 struct rtrs_clt_io_req *req =
351 container_of(wc->wr_cqe, typeof(*req), inv_cqe);
352 struct rtrs_clt_con *con = cq->cq_context;
354 if (unlikely(wc->status != IB_WC_SUCCESS)) {
355 rtrs_err(con->c.sess, "Failed IB_WR_LOCAL_INV: %s\n",
356 ib_wc_status_msg(wc->status));
357 rtrs_rdma_error_recovery(con);
359 req->need_inv = false;
360 if (likely(req->need_inv_comp))
361 complete(&req->inv_comp);
363 /* Complete request from INV callback */
364 complete_rdma_req(req, req->inv_errno, true, false);
367 static int rtrs_inv_rkey(struct rtrs_clt_io_req *req)
369 struct rtrs_clt_con *con = req->con;
370 struct ib_send_wr wr = {
371 .opcode = IB_WR_LOCAL_INV,
372 .wr_cqe = &req->inv_cqe,
373 .send_flags = IB_SEND_SIGNALED,
374 .ex.invalidate_rkey = req->mr->rkey,
376 req->inv_cqe.done = rtrs_clt_inv_rkey_done;
378 return ib_post_send(con->c.qp, &wr, NULL);
381 static void complete_rdma_req(struct rtrs_clt_io_req *req, int errno,
382 bool notify, bool can_wait)
384 struct rtrs_clt_con *con = req->con;
385 struct rtrs_clt_sess *sess;
388 if (WARN_ON(!req->in_use))
390 if (WARN_ON(!req->con))
392 sess = to_clt_sess(con->c.sess);
395 if (unlikely(req->dir == DMA_FROM_DEVICE && req->need_inv)) {
397 * We are here to invalidate read requests
398 * ourselves. In normal scenario server should
399 * send INV for all read requests, but
400 * we are here, thus two things could happen:
402 * 1. this is failover, when errno != 0
405 * 2. something totally bad happened and
406 * server forgot to send INV, so we
407 * should do that ourselves.
410 if (likely(can_wait)) {
411 req->need_inv_comp = true;
413 /* This should be IO path, so always notify */
415 /* Save errno for INV callback */
416 req->inv_errno = errno;
419 err = rtrs_inv_rkey(req);
421 rtrs_err(con->c.sess, "Send INV WR key=%#x: %d\n",
423 } else if (likely(can_wait)) {
424 wait_for_completion(&req->inv_comp);
427 * Something went wrong, so request will be
428 * completed from INV callback.
435 ib_dma_unmap_sg(sess->s.dev->ib_dev, req->sglist,
436 req->sg_cnt, req->dir);
438 if (sess->clt->mp_policy == MP_POLICY_MIN_INFLIGHT)
439 atomic_dec(&sess->stats->inflight);
445 req->conf(req->priv, errno);
448 static int rtrs_post_send_rdma(struct rtrs_clt_con *con,
449 struct rtrs_clt_io_req *req,
450 struct rtrs_rbuf *rbuf, u32 off,
451 u32 imm, struct ib_send_wr *wr)
453 struct rtrs_clt_sess *sess = to_clt_sess(con->c.sess);
454 enum ib_send_flags flags;
457 if (unlikely(!req->sg_size)) {
458 rtrs_wrn(con->c.sess,
459 "Doing RDMA Write failed, no data supplied\n");
463 /* user data and user message in the first list element */
464 sge.addr = req->iu->dma_addr;
465 sge.length = req->sg_size;
466 sge.lkey = sess->s.dev->ib_pd->local_dma_lkey;
469 * From time to time we have to post signalled sends,
470 * or send queue will fill up and only QP reset can help.
472 flags = atomic_inc_return(&con->io_cnt) % sess->queue_depth ?
473 0 : IB_SEND_SIGNALED;
475 ib_dma_sync_single_for_device(sess->s.dev->ib_dev, req->iu->dma_addr,
476 req->sg_size, DMA_TO_DEVICE);
478 return rtrs_iu_post_rdma_write_imm(&con->c, req->iu, &sge, 1,
479 rbuf->rkey, rbuf->addr + off,
483 static void process_io_rsp(struct rtrs_clt_sess *sess, u32 msg_id,
484 s16 errno, bool w_inval)
486 struct rtrs_clt_io_req *req;
488 if (WARN_ON(msg_id >= sess->queue_depth))
491 req = &sess->reqs[msg_id];
492 /* Drop need_inv if server responded with send with invalidation */
493 req->need_inv &= !w_inval;
494 complete_rdma_req(req, errno, true, false);
497 static void rtrs_clt_recv_done(struct rtrs_clt_con *con, struct ib_wc *wc)
501 struct rtrs_clt_sess *sess = to_clt_sess(con->c.sess);
503 WARN_ON(sess->flags != RTRS_MSG_NEW_RKEY_F);
504 iu = container_of(wc->wr_cqe, struct rtrs_iu,
506 err = rtrs_iu_post_recv(&con->c, iu);
508 rtrs_err(con->c.sess, "post iu failed %d\n", err);
509 rtrs_rdma_error_recovery(con);
513 static void rtrs_clt_rkey_rsp_done(struct rtrs_clt_con *con, struct ib_wc *wc)
515 struct rtrs_clt_sess *sess = to_clt_sess(con->c.sess);
516 struct rtrs_msg_rkey_rsp *msg;
517 u32 imm_type, imm_payload;
518 bool w_inval = false;
523 WARN_ON(sess->flags != RTRS_MSG_NEW_RKEY_F);
525 iu = container_of(wc->wr_cqe, struct rtrs_iu, cqe);
527 if (unlikely(wc->byte_len < sizeof(*msg))) {
528 rtrs_err(con->c.sess, "rkey response is malformed: size %d\n",
532 ib_dma_sync_single_for_cpu(sess->s.dev->ib_dev, iu->dma_addr,
533 iu->size, DMA_FROM_DEVICE);
535 if (unlikely(le16_to_cpu(msg->type) != RTRS_MSG_RKEY_RSP)) {
536 rtrs_err(sess->clt, "rkey response is malformed: type %d\n",
537 le16_to_cpu(msg->type));
540 buf_id = le16_to_cpu(msg->buf_id);
541 if (WARN_ON(buf_id >= sess->queue_depth))
544 rtrs_from_imm(be32_to_cpu(wc->ex.imm_data), &imm_type, &imm_payload);
545 if (likely(imm_type == RTRS_IO_RSP_IMM ||
546 imm_type == RTRS_IO_RSP_W_INV_IMM)) {
549 w_inval = (imm_type == RTRS_IO_RSP_W_INV_IMM);
550 rtrs_from_io_rsp_imm(imm_payload, &msg_id, &err);
552 if (WARN_ON(buf_id != msg_id))
554 sess->rbufs[buf_id].rkey = le32_to_cpu(msg->rkey);
555 process_io_rsp(sess, msg_id, err, w_inval);
557 ib_dma_sync_single_for_device(sess->s.dev->ib_dev, iu->dma_addr,
558 iu->size, DMA_FROM_DEVICE);
559 return rtrs_clt_recv_done(con, wc);
561 rtrs_rdma_error_recovery(con);
564 static void rtrs_clt_rdma_done(struct ib_cq *cq, struct ib_wc *wc);
566 static struct ib_cqe io_comp_cqe = {
567 .done = rtrs_clt_rdma_done
571 * Post x2 empty WRs: first is for this RDMA with IMM,
572 * second is for RECV with INV, which happened earlier.
574 static int rtrs_post_recv_empty_x2(struct rtrs_con *con, struct ib_cqe *cqe)
576 struct ib_recv_wr wr_arr[2], *wr;
579 memset(wr_arr, 0, sizeof(wr_arr));
580 for (i = 0; i < ARRAY_SIZE(wr_arr); i++) {
584 /* Chain backwards */
585 wr->next = &wr_arr[i - 1];
588 return ib_post_recv(con->qp, wr, NULL);
591 static void rtrs_clt_rdma_done(struct ib_cq *cq, struct ib_wc *wc)
593 struct rtrs_clt_con *con = cq->cq_context;
594 struct rtrs_clt_sess *sess = to_clt_sess(con->c.sess);
595 u32 imm_type, imm_payload;
596 bool w_inval = false;
599 if (unlikely(wc->status != IB_WC_SUCCESS)) {
600 if (wc->status != IB_WC_WR_FLUSH_ERR) {
601 rtrs_err(sess->clt, "RDMA failed: %s\n",
602 ib_wc_status_msg(wc->status));
603 rtrs_rdma_error_recovery(con);
607 rtrs_clt_update_wc_stats(con);
609 switch (wc->opcode) {
610 case IB_WC_RECV_RDMA_WITH_IMM:
612 * post_recv() RDMA write completions of IO reqs (read/write)
615 if (WARN_ON(wc->wr_cqe->done != rtrs_clt_rdma_done))
617 rtrs_from_imm(be32_to_cpu(wc->ex.imm_data),
618 &imm_type, &imm_payload);
619 if (likely(imm_type == RTRS_IO_RSP_IMM ||
620 imm_type == RTRS_IO_RSP_W_INV_IMM)) {
623 w_inval = (imm_type == RTRS_IO_RSP_W_INV_IMM);
624 rtrs_from_io_rsp_imm(imm_payload, &msg_id, &err);
626 process_io_rsp(sess, msg_id, err, w_inval);
627 } else if (imm_type == RTRS_HB_MSG_IMM) {
629 rtrs_send_hb_ack(&sess->s);
630 if (sess->flags == RTRS_MSG_NEW_RKEY_F)
631 return rtrs_clt_recv_done(con, wc);
632 } else if (imm_type == RTRS_HB_ACK_IMM) {
634 sess->s.hb_missed_cnt = 0;
635 if (sess->flags == RTRS_MSG_NEW_RKEY_F)
636 return rtrs_clt_recv_done(con, wc);
638 rtrs_wrn(con->c.sess, "Unknown IMM type %u\n",
643 * Post x2 empty WRs: first is for this RDMA with IMM,
644 * second is for RECV with INV, which happened earlier.
646 err = rtrs_post_recv_empty_x2(&con->c, &io_comp_cqe);
648 err = rtrs_post_recv_empty(&con->c, &io_comp_cqe);
650 rtrs_err(con->c.sess, "rtrs_post_recv_empty(): %d\n",
652 rtrs_rdma_error_recovery(con);
658 * Key invalidations from server side
660 WARN_ON(!(wc->wc_flags & IB_WC_WITH_INVALIDATE ||
661 wc->wc_flags & IB_WC_WITH_IMM));
662 WARN_ON(wc->wr_cqe->done != rtrs_clt_rdma_done);
663 if (sess->flags == RTRS_MSG_NEW_RKEY_F) {
664 if (wc->wc_flags & IB_WC_WITH_INVALIDATE)
665 return rtrs_clt_recv_done(con, wc);
667 return rtrs_clt_rkey_rsp_done(con, wc);
670 case IB_WC_RDMA_WRITE:
672 * post_send() RDMA write completions of IO reqs (read/write)
678 rtrs_wrn(sess->clt, "Unexpected WC type: %d\n", wc->opcode);
683 static int post_recv_io(struct rtrs_clt_con *con, size_t q_size)
686 struct rtrs_clt_sess *sess = to_clt_sess(con->c.sess);
688 for (i = 0; i < q_size; i++) {
689 if (sess->flags == RTRS_MSG_NEW_RKEY_F) {
690 struct rtrs_iu *iu = &con->rsp_ius[i];
692 err = rtrs_iu_post_recv(&con->c, iu);
694 err = rtrs_post_recv_empty(&con->c, &io_comp_cqe);
703 static int post_recv_sess(struct rtrs_clt_sess *sess)
708 for (cid = 0; cid < sess->s.con_num; cid++) {
710 q_size = SERVICE_CON_QUEUE_DEPTH;
712 q_size = sess->queue_depth;
715 * x2 for RDMA read responses + FR key invalidations,
716 * RDMA writes do not require any FR registrations.
720 err = post_recv_io(to_clt_con(sess->s.con[cid]), q_size);
722 rtrs_err(sess->clt, "post_recv_io(), err: %d\n", err);
732 struct list_head skip_list;
733 struct rtrs_clt *clt;
734 struct rtrs_clt_sess *(*next_path)(struct path_it *it);
738 * list_next_or_null_rr_rcu - get next list element in round-robin fashion.
739 * @head: the head for the list.
740 * @ptr: the list head to take the next element from.
741 * @type: the type of the struct this is embedded in.
742 * @memb: the name of the list_head within the struct.
744 * Next element returned in round-robin fashion, i.e. head will be skipped,
745 * but if list is observed as empty, NULL will be returned.
747 * This primitive may safely run concurrently with the _rcu list-mutation
748 * primitives such as list_add_rcu() as long as it's guarded by rcu_read_lock().
750 #define list_next_or_null_rr_rcu(head, ptr, type, memb) \
752 list_next_or_null_rcu(head, ptr, type, memb) ?: \
753 list_next_or_null_rcu(head, READ_ONCE((ptr)->next), \
758 * get_next_path_rr() - Returns path in round-robin fashion.
759 * @it: the path pointer
761 * Related to @MP_POLICY_RR
764 * rcu_read_lock() must be hold.
766 static struct rtrs_clt_sess *get_next_path_rr(struct path_it *it)
768 struct rtrs_clt_sess __rcu **ppcpu_path;
769 struct rtrs_clt_sess *path;
770 struct rtrs_clt *clt;
775 * Here we use two RCU objects: @paths_list and @pcpu_path
776 * pointer. See rtrs_clt_remove_path_from_arr() for details
777 * how that is handled.
780 ppcpu_path = this_cpu_ptr(clt->pcpu_path);
781 path = rcu_dereference(*ppcpu_path);
783 path = list_first_or_null_rcu(&clt->paths_list,
784 typeof(*path), s.entry);
786 path = list_next_or_null_rr_rcu(&clt->paths_list,
790 rcu_assign_pointer(*ppcpu_path, path);
796 * get_next_path_min_inflight() - Returns path with minimal inflight count.
797 * @it: the path pointer
799 * Related to @MP_POLICY_MIN_INFLIGHT
802 * rcu_read_lock() must be hold.
804 static struct rtrs_clt_sess *get_next_path_min_inflight(struct path_it *it)
806 struct rtrs_clt_sess *min_path = NULL;
807 struct rtrs_clt *clt = it->clt;
808 struct rtrs_clt_sess *sess;
809 int min_inflight = INT_MAX;
812 list_for_each_entry_rcu(sess, &clt->paths_list, s.entry) {
813 if (unlikely(!list_empty(raw_cpu_ptr(sess->mp_skip_entry))))
816 inflight = atomic_read(&sess->stats->inflight);
818 if (inflight < min_inflight) {
819 min_inflight = inflight;
825 * add the path to the skip list, so that next time we can get
829 list_add(raw_cpu_ptr(min_path->mp_skip_entry), &it->skip_list);
834 static inline void path_it_init(struct path_it *it, struct rtrs_clt *clt)
836 INIT_LIST_HEAD(&it->skip_list);
840 if (clt->mp_policy == MP_POLICY_RR)
841 it->next_path = get_next_path_rr;
843 it->next_path = get_next_path_min_inflight;
846 static inline void path_it_deinit(struct path_it *it)
848 struct list_head *skip, *tmp;
850 * The skip_list is used only for the MIN_INFLIGHT policy.
851 * We need to remove paths from it, so that next IO can insert
852 * paths (->mp_skip_entry) into a skip_list again.
854 list_for_each_safe(skip, tmp, &it->skip_list)
859 * rtrs_clt_init_req() Initialize an rtrs_clt_io_req holding information
860 * about an inflight IO.
861 * The user buffer holding user control message (not data) is copied into
862 * the corresponding buffer of rtrs_iu (req->iu->buf), which later on will
863 * also hold the control message of rtrs.
864 * @req: an io request holding information about IO.
865 * @sess: client session
866 * @conf: conformation callback function to notify upper layer.
867 * @permit: permit for allocation of RDMA remote buffer
868 * @priv: private pointer
869 * @vec: kernel vector containing control message
870 * @usr_len: length of the user message
871 * @sg: scater list for IO data
872 * @sg_cnt: number of scater list entries
873 * @data_len: length of the IO data
874 * @dir: direction of the IO.
876 static void rtrs_clt_init_req(struct rtrs_clt_io_req *req,
877 struct rtrs_clt_sess *sess,
878 void (*conf)(void *priv, int errno),
879 struct rtrs_permit *permit, void *priv,
880 const struct kvec *vec, size_t usr_len,
881 struct scatterlist *sg, size_t sg_cnt,
882 size_t data_len, int dir)
884 struct iov_iter iter;
887 req->permit = permit;
889 req->usr_len = usr_len;
890 req->data_len = data_len;
892 req->sg_cnt = sg_cnt;
895 req->con = rtrs_permit_to_clt_con(sess, permit);
897 req->need_inv = false;
898 req->need_inv_comp = false;
901 iov_iter_kvec(&iter, READ, vec, 1, usr_len);
902 len = _copy_from_iter(req->iu->buf, usr_len, &iter);
903 WARN_ON(len != usr_len);
905 reinit_completion(&req->inv_comp);
908 static struct rtrs_clt_io_req *
909 rtrs_clt_get_req(struct rtrs_clt_sess *sess,
910 void (*conf)(void *priv, int errno),
911 struct rtrs_permit *permit, void *priv,
912 const struct kvec *vec, size_t usr_len,
913 struct scatterlist *sg, size_t sg_cnt,
914 size_t data_len, int dir)
916 struct rtrs_clt_io_req *req;
918 req = &sess->reqs[permit->mem_id];
919 rtrs_clt_init_req(req, sess, conf, permit, priv, vec, usr_len,
920 sg, sg_cnt, data_len, dir);
924 static struct rtrs_clt_io_req *
925 rtrs_clt_get_copy_req(struct rtrs_clt_sess *alive_sess,
926 struct rtrs_clt_io_req *fail_req)
928 struct rtrs_clt_io_req *req;
930 .iov_base = fail_req->iu->buf,
931 .iov_len = fail_req->usr_len
934 req = &alive_sess->reqs[fail_req->permit->mem_id];
935 rtrs_clt_init_req(req, alive_sess, fail_req->conf, fail_req->permit,
936 fail_req->priv, &vec, fail_req->usr_len,
937 fail_req->sglist, fail_req->sg_cnt,
938 fail_req->data_len, fail_req->dir);
942 static int rtrs_post_rdma_write_sg(struct rtrs_clt_con *con,
943 struct rtrs_clt_io_req *req,
944 struct rtrs_rbuf *rbuf,
947 struct rtrs_clt_sess *sess = to_clt_sess(con->c.sess);
948 struct ib_sge *sge = req->sge;
949 enum ib_send_flags flags;
950 struct scatterlist *sg;
954 for_each_sg(req->sglist, sg, req->sg_cnt, i) {
955 sge[i].addr = sg_dma_address(sg);
956 sge[i].length = sg_dma_len(sg);
957 sge[i].lkey = sess->s.dev->ib_pd->local_dma_lkey;
959 sge[i].addr = req->iu->dma_addr;
960 sge[i].length = size;
961 sge[i].lkey = sess->s.dev->ib_pd->local_dma_lkey;
963 num_sge = 1 + req->sg_cnt;
966 * From time to time we have to post signalled sends,
967 * or send queue will fill up and only QP reset can help.
969 flags = atomic_inc_return(&con->io_cnt) % sess->queue_depth ?
970 0 : IB_SEND_SIGNALED;
972 ib_dma_sync_single_for_device(sess->s.dev->ib_dev, req->iu->dma_addr,
973 size, DMA_TO_DEVICE);
975 return rtrs_iu_post_rdma_write_imm(&con->c, req->iu, sge, num_sge,
976 rbuf->rkey, rbuf->addr, imm,
980 static int rtrs_clt_write_req(struct rtrs_clt_io_req *req)
982 struct rtrs_clt_con *con = req->con;
983 struct rtrs_sess *s = con->c.sess;
984 struct rtrs_clt_sess *sess = to_clt_sess(s);
985 struct rtrs_msg_rdma_write *msg;
987 struct rtrs_rbuf *rbuf;
991 const size_t tsize = sizeof(*msg) + req->data_len + req->usr_len;
993 if (unlikely(tsize > sess->chunk_size)) {
994 rtrs_wrn(s, "Write request failed, size too big %zu > %d\n",
995 tsize, sess->chunk_size);
999 count = ib_dma_map_sg(sess->s.dev->ib_dev, req->sglist,
1000 req->sg_cnt, req->dir);
1001 if (unlikely(!count)) {
1002 rtrs_wrn(s, "Write request failed, map failed\n");
1006 /* put rtrs msg after sg and user message */
1007 msg = req->iu->buf + req->usr_len;
1008 msg->type = cpu_to_le16(RTRS_MSG_WRITE);
1009 msg->usr_len = cpu_to_le16(req->usr_len);
1011 /* rtrs message on server side will be after user data and message */
1012 imm = req->permit->mem_off + req->data_len + req->usr_len;
1013 imm = rtrs_to_io_req_imm(imm);
1014 buf_id = req->permit->mem_id;
1015 req->sg_size = tsize;
1016 rbuf = &sess->rbufs[buf_id];
1019 * Update stats now, after request is successfully sent it is not
1020 * safe anymore to touch it.
1022 rtrs_clt_update_all_stats(req, WRITE);
1024 ret = rtrs_post_rdma_write_sg(req->con, req, rbuf,
1025 req->usr_len + sizeof(*msg),
1027 if (unlikely(ret)) {
1028 rtrs_err(s, "Write request failed: %d\n", ret);
1029 if (sess->clt->mp_policy == MP_POLICY_MIN_INFLIGHT)
1030 atomic_dec(&sess->stats->inflight);
1032 ib_dma_unmap_sg(sess->s.dev->ib_dev, req->sglist,
1033 req->sg_cnt, req->dir);
1039 static int rtrs_map_sg_fr(struct rtrs_clt_io_req *req, size_t count)
1043 /* Align the MR to a 4K page size to match the block virt boundary */
1044 nr = ib_map_mr_sg(req->mr, req->sglist, count, NULL, SZ_4K);
1047 if (unlikely(nr < req->sg_cnt))
1049 ib_update_fast_reg_key(req->mr, ib_inc_rkey(req->mr->rkey));
1054 static int rtrs_clt_read_req(struct rtrs_clt_io_req *req)
1056 struct rtrs_clt_con *con = req->con;
1057 struct rtrs_sess *s = con->c.sess;
1058 struct rtrs_clt_sess *sess = to_clt_sess(s);
1059 struct rtrs_msg_rdma_read *msg;
1060 struct rtrs_ib_dev *dev;
1062 struct ib_reg_wr rwr;
1063 struct ib_send_wr *wr = NULL;
1068 const size_t tsize = sizeof(*msg) + req->data_len + req->usr_len;
1073 if (unlikely(tsize > sess->chunk_size)) {
1075 "Read request failed, message size is %zu, bigger than CHUNK_SIZE %d\n",
1076 tsize, sess->chunk_size);
1081 count = ib_dma_map_sg(dev->ib_dev, req->sglist, req->sg_cnt,
1083 if (unlikely(!count)) {
1085 "Read request failed, dma map failed\n");
1089 /* put our message into req->buf after user message*/
1090 msg = req->iu->buf + req->usr_len;
1091 msg->type = cpu_to_le16(RTRS_MSG_READ);
1092 msg->usr_len = cpu_to_le16(req->usr_len);
1095 ret = rtrs_map_sg_fr(req, count);
1098 "Read request failed, failed to map fast reg. data, err: %d\n",
1100 ib_dma_unmap_sg(dev->ib_dev, req->sglist, req->sg_cnt,
1104 rwr = (struct ib_reg_wr) {
1105 .wr.opcode = IB_WR_REG_MR,
1106 .wr.wr_cqe = &fast_reg_cqe,
1108 .key = req->mr->rkey,
1109 .access = (IB_ACCESS_LOCAL_WRITE |
1110 IB_ACCESS_REMOTE_WRITE),
1114 msg->sg_cnt = cpu_to_le16(1);
1115 msg->flags = cpu_to_le16(RTRS_MSG_NEED_INVAL_F);
1117 msg->desc[0].addr = cpu_to_le64(req->mr->iova);
1118 msg->desc[0].key = cpu_to_le32(req->mr->rkey);
1119 msg->desc[0].len = cpu_to_le32(req->mr->length);
1121 /* Further invalidation is required */
1122 req->need_inv = !!RTRS_MSG_NEED_INVAL_F;
1129 * rtrs message will be after the space reserved for disk data and
1132 imm = req->permit->mem_off + req->data_len + req->usr_len;
1133 imm = rtrs_to_io_req_imm(imm);
1134 buf_id = req->permit->mem_id;
1136 req->sg_size = sizeof(*msg);
1137 req->sg_size += le16_to_cpu(msg->sg_cnt) * sizeof(struct rtrs_sg_desc);
1138 req->sg_size += req->usr_len;
1141 * Update stats now, after request is successfully sent it is not
1142 * safe anymore to touch it.
1144 rtrs_clt_update_all_stats(req, READ);
1146 ret = rtrs_post_send_rdma(req->con, req, &sess->rbufs[buf_id],
1147 req->data_len, imm, wr);
1148 if (unlikely(ret)) {
1149 rtrs_err(s, "Read request failed: %d\n", ret);
1150 if (sess->clt->mp_policy == MP_POLICY_MIN_INFLIGHT)
1151 atomic_dec(&sess->stats->inflight);
1152 req->need_inv = false;
1154 ib_dma_unmap_sg(dev->ib_dev, req->sglist,
1155 req->sg_cnt, req->dir);
1162 * rtrs_clt_failover_req() Try to find an active path for a failed request
1164 * @fail_req: a failed io request.
1166 static int rtrs_clt_failover_req(struct rtrs_clt *clt,
1167 struct rtrs_clt_io_req *fail_req)
1169 struct rtrs_clt_sess *alive_sess;
1170 struct rtrs_clt_io_req *req;
1171 int err = -ECONNABORTED;
1175 for (path_it_init(&it, clt);
1176 (alive_sess = it.next_path(&it)) && it.i < it.clt->paths_num;
1178 if (unlikely(READ_ONCE(alive_sess->state) !=
1179 RTRS_CLT_CONNECTED))
1181 req = rtrs_clt_get_copy_req(alive_sess, fail_req);
1182 if (req->dir == DMA_TO_DEVICE)
1183 err = rtrs_clt_write_req(req);
1185 err = rtrs_clt_read_req(req);
1186 if (unlikely(err)) {
1187 req->in_use = false;
1191 rtrs_clt_inc_failover_cnt(alive_sess->stats);
1194 path_it_deinit(&it);
1200 static void fail_all_outstanding_reqs(struct rtrs_clt_sess *sess)
1202 struct rtrs_clt *clt = sess->clt;
1203 struct rtrs_clt_io_req *req;
1208 for (i = 0; i < sess->queue_depth; ++i) {
1209 req = &sess->reqs[i];
1214 * Safely (without notification) complete failed request.
1215 * After completion this request is still useble and can
1216 * be failovered to another path.
1218 complete_rdma_req(req, -ECONNABORTED, false, true);
1220 err = rtrs_clt_failover_req(clt, req);
1222 /* Failover failed, notify anyway */
1223 req->conf(req->priv, err);
1227 static void free_sess_reqs(struct rtrs_clt_sess *sess)
1229 struct rtrs_clt_io_req *req;
1234 for (i = 0; i < sess->queue_depth; ++i) {
1235 req = &sess->reqs[i];
1237 ib_dereg_mr(req->mr);
1239 rtrs_iu_free(req->iu, sess->s.dev->ib_dev, 1);
1245 static int alloc_sess_reqs(struct rtrs_clt_sess *sess)
1247 struct rtrs_clt_io_req *req;
1248 struct rtrs_clt *clt = sess->clt;
1249 int i, err = -ENOMEM;
1251 sess->reqs = kcalloc(sess->queue_depth, sizeof(*sess->reqs),
1256 for (i = 0; i < sess->queue_depth; ++i) {
1257 req = &sess->reqs[i];
1258 req->iu = rtrs_iu_alloc(1, sess->max_hdr_size, GFP_KERNEL,
1259 sess->s.dev->ib_dev,
1261 rtrs_clt_rdma_done);
1265 req->sge = kmalloc_array(clt->max_segments + 1,
1266 sizeof(*req->sge), GFP_KERNEL);
1270 req->mr = ib_alloc_mr(sess->s.dev->ib_pd, IB_MR_TYPE_MEM_REG,
1271 sess->max_pages_per_mr);
1272 if (IS_ERR(req->mr)) {
1273 err = PTR_ERR(req->mr);
1275 pr_err("Failed to alloc sess->max_pages_per_mr %d\n",
1276 sess->max_pages_per_mr);
1280 init_completion(&req->inv_comp);
1286 free_sess_reqs(sess);
1291 static int alloc_permits(struct rtrs_clt *clt)
1293 unsigned int chunk_bits;
1296 clt->permits_map = kcalloc(BITS_TO_LONGS(clt->queue_depth),
1297 sizeof(long), GFP_KERNEL);
1298 if (!clt->permits_map) {
1302 clt->permits = kcalloc(clt->queue_depth, permit_size(clt), GFP_KERNEL);
1303 if (!clt->permits) {
1307 chunk_bits = ilog2(clt->queue_depth - 1) + 1;
1308 for (i = 0; i < clt->queue_depth; i++) {
1309 struct rtrs_permit *permit;
1311 permit = get_permit(clt, i);
1313 permit->mem_off = i << (MAX_IMM_PAYL_BITS - chunk_bits);
1319 kfree(clt->permits_map);
1320 clt->permits_map = NULL;
1325 static void free_permits(struct rtrs_clt *clt)
1327 kfree(clt->permits_map);
1328 clt->permits_map = NULL;
1329 kfree(clt->permits);
1330 clt->permits = NULL;
1333 static void query_fast_reg_mode(struct rtrs_clt_sess *sess)
1335 struct ib_device *ib_dev;
1336 u64 max_pages_per_mr;
1339 ib_dev = sess->s.dev->ib_dev;
1342 * Use the smallest page size supported by the HCA, down to a
1343 * minimum of 4096 bytes. We're unlikely to build large sglists
1344 * out of smaller entries.
1346 mr_page_shift = max(12, ffs(ib_dev->attrs.page_size_cap) - 1);
1347 max_pages_per_mr = ib_dev->attrs.max_mr_size;
1348 do_div(max_pages_per_mr, (1ull << mr_page_shift));
1349 sess->max_pages_per_mr =
1350 min3(sess->max_pages_per_mr, (u32)max_pages_per_mr,
1351 ib_dev->attrs.max_fast_reg_page_list_len);
1352 sess->max_send_sge = ib_dev->attrs.max_send_sge;
1355 static bool rtrs_clt_change_state_get_old(struct rtrs_clt_sess *sess,
1356 enum rtrs_clt_state new_state,
1357 enum rtrs_clt_state *old_state)
1361 spin_lock_irq(&sess->state_wq.lock);
1362 *old_state = sess->state;
1363 changed = __rtrs_clt_change_state(sess, new_state);
1364 spin_unlock_irq(&sess->state_wq.lock);
1369 static bool rtrs_clt_change_state(struct rtrs_clt_sess *sess,
1370 enum rtrs_clt_state new_state)
1372 enum rtrs_clt_state old_state;
1374 return rtrs_clt_change_state_get_old(sess, new_state, &old_state);
1377 static void rtrs_clt_hb_err_handler(struct rtrs_con *c)
1379 struct rtrs_clt_con *con = container_of(c, typeof(*con), c);
1381 rtrs_rdma_error_recovery(con);
1384 static void rtrs_clt_init_hb(struct rtrs_clt_sess *sess)
1386 rtrs_init_hb(&sess->s, &io_comp_cqe,
1387 RTRS_HB_INTERVAL_MS,
1389 rtrs_clt_hb_err_handler,
1393 static void rtrs_clt_start_hb(struct rtrs_clt_sess *sess)
1395 rtrs_start_hb(&sess->s);
1398 static void rtrs_clt_stop_hb(struct rtrs_clt_sess *sess)
1400 rtrs_stop_hb(&sess->s);
1403 static void rtrs_clt_reconnect_work(struct work_struct *work);
1404 static void rtrs_clt_close_work(struct work_struct *work);
1406 static struct rtrs_clt_sess *alloc_sess(struct rtrs_clt *clt,
1407 const struct rtrs_addr *path,
1408 size_t con_num, u16 max_segments,
1409 size_t max_segment_size)
1411 struct rtrs_clt_sess *sess;
1415 sess = kzalloc(sizeof(*sess), GFP_KERNEL);
1419 /* Extra connection for user messages */
1422 sess->s.con = kcalloc(con_num, sizeof(*sess->s.con), GFP_KERNEL);
1426 sess->stats = kzalloc(sizeof(*sess->stats), GFP_KERNEL);
1430 mutex_init(&sess->init_mutex);
1431 uuid_gen(&sess->s.uuid);
1432 memcpy(&sess->s.dst_addr, path->dst,
1433 rdma_addr_size((struct sockaddr *)path->dst));
1436 * rdma_resolve_addr() passes src_addr to cma_bind_addr, which
1437 * checks the sa_family to be non-zero. If user passed src_addr=NULL
1438 * the sess->src_addr will contain only zeros, which is then fine.
1441 memcpy(&sess->s.src_addr, path->src,
1442 rdma_addr_size((struct sockaddr *)path->src));
1443 strlcpy(sess->s.sessname, clt->sessname, sizeof(sess->s.sessname));
1444 sess->s.con_num = con_num;
1446 sess->max_pages_per_mr = max_segments * max_segment_size >> 12;
1447 init_waitqueue_head(&sess->state_wq);
1448 sess->state = RTRS_CLT_CONNECTING;
1449 atomic_set(&sess->connected_cnt, 0);
1450 INIT_WORK(&sess->close_work, rtrs_clt_close_work);
1451 INIT_DELAYED_WORK(&sess->reconnect_dwork, rtrs_clt_reconnect_work);
1452 rtrs_clt_init_hb(sess);
1454 sess->mp_skip_entry = alloc_percpu(typeof(*sess->mp_skip_entry));
1455 if (!sess->mp_skip_entry)
1456 goto err_free_stats;
1458 for_each_possible_cpu(cpu)
1459 INIT_LIST_HEAD(per_cpu_ptr(sess->mp_skip_entry, cpu));
1461 err = rtrs_clt_init_stats(sess->stats);
1463 goto err_free_percpu;
1468 free_percpu(sess->mp_skip_entry);
1476 return ERR_PTR(err);
1479 void free_sess(struct rtrs_clt_sess *sess)
1481 free_percpu(sess->mp_skip_entry);
1482 mutex_destroy(&sess->init_mutex);
1488 static int create_con(struct rtrs_clt_sess *sess, unsigned int cid)
1490 struct rtrs_clt_con *con;
1492 con = kzalloc(sizeof(*con), GFP_KERNEL);
1496 /* Map first two connections to the first CPU */
1497 con->cpu = (cid ? cid - 1 : 0) % nr_cpu_ids;
1499 con->c.sess = &sess->s;
1500 atomic_set(&con->io_cnt, 0);
1501 mutex_init(&con->con_mutex);
1503 sess->s.con[cid] = &con->c;
1508 static void destroy_con(struct rtrs_clt_con *con)
1510 struct rtrs_clt_sess *sess = to_clt_sess(con->c.sess);
1512 sess->s.con[con->c.cid] = NULL;
1513 mutex_destroy(&con->con_mutex);
1517 static int create_con_cq_qp(struct rtrs_clt_con *con)
1519 struct rtrs_clt_sess *sess = to_clt_sess(con->c.sess);
1522 struct rtrs_msg_rkey_rsp *rsp;
1524 lockdep_assert_held(&con->con_mutex);
1525 if (con->c.cid == 0) {
1527 * One completion for each receive and two for each send
1528 * (send request + registration)
1529 * + 2 for drain and heartbeat
1530 * in case qp gets into error state
1532 wr_queue_size = SERVICE_CON_QUEUE_DEPTH * 3 + 2;
1533 /* We must be the first here */
1534 if (WARN_ON(sess->s.dev))
1538 * The whole session uses device from user connection.
1539 * Be careful not to close user connection before ib dev
1540 * is gracefully put.
1542 sess->s.dev = rtrs_ib_dev_find_or_add(con->c.cm_id->device,
1546 "rtrs_ib_dev_find_get_or_add(): no memory\n");
1549 sess->s.dev_ref = 1;
1550 query_fast_reg_mode(sess);
1553 * Here we assume that session members are correctly set.
1554 * This is always true if user connection (cid == 0) is
1555 * established first.
1557 if (WARN_ON(!sess->s.dev))
1559 if (WARN_ON(!sess->queue_depth))
1562 /* Shared between connections */
1565 min_t(int, sess->s.dev->ib_dev->attrs.max_qp_wr,
1566 /* QD * (REQ + RSP + FR REGS or INVS) + drain */
1567 sess->queue_depth * 3 + 1);
1569 /* alloc iu to recv new rkey reply when server reports flags set */
1570 if (sess->flags == RTRS_MSG_NEW_RKEY_F || con->c.cid == 0) {
1571 con->rsp_ius = rtrs_iu_alloc(wr_queue_size, sizeof(*rsp),
1572 GFP_KERNEL, sess->s.dev->ib_dev,
1574 rtrs_clt_rdma_done);
1577 con->queue_size = wr_queue_size;
1579 cq_vector = con->cpu % sess->s.dev->ib_dev->num_comp_vectors;
1580 err = rtrs_cq_qp_create(&sess->s, &con->c, sess->max_send_sge,
1581 cq_vector, wr_queue_size, wr_queue_size,
1584 * In case of error we do not bother to clean previous allocations,
1585 * since destroy_con_cq_qp() must be called.
1590 static void destroy_con_cq_qp(struct rtrs_clt_con *con)
1592 struct rtrs_clt_sess *sess = to_clt_sess(con->c.sess);
1595 * Be careful here: destroy_con_cq_qp() can be called even
1596 * create_con_cq_qp() failed, see comments there.
1598 lockdep_assert_held(&con->con_mutex);
1599 rtrs_cq_qp_destroy(&con->c);
1601 rtrs_iu_free(con->rsp_ius, sess->s.dev->ib_dev, con->queue_size);
1602 con->rsp_ius = NULL;
1603 con->queue_size = 0;
1605 if (sess->s.dev_ref && !--sess->s.dev_ref) {
1606 rtrs_ib_dev_put(sess->s.dev);
1611 static void stop_cm(struct rtrs_clt_con *con)
1613 rdma_disconnect(con->c.cm_id);
1615 ib_drain_qp(con->c.qp);
1618 static void destroy_cm(struct rtrs_clt_con *con)
1620 rdma_destroy_id(con->c.cm_id);
1621 con->c.cm_id = NULL;
1624 static int rtrs_rdma_addr_resolved(struct rtrs_clt_con *con)
1626 struct rtrs_sess *s = con->c.sess;
1629 mutex_lock(&con->con_mutex);
1630 err = create_con_cq_qp(con);
1631 mutex_unlock(&con->con_mutex);
1633 rtrs_err(s, "create_con_cq_qp(), err: %d\n", err);
1636 err = rdma_resolve_route(con->c.cm_id, RTRS_CONNECT_TIMEOUT_MS);
1638 rtrs_err(s, "Resolving route failed, err: %d\n", err);
1643 static int rtrs_rdma_route_resolved(struct rtrs_clt_con *con)
1645 struct rtrs_clt_sess *sess = to_clt_sess(con->c.sess);
1646 struct rtrs_clt *clt = sess->clt;
1647 struct rtrs_msg_conn_req msg;
1648 struct rdma_conn_param param;
1652 param = (struct rdma_conn_param) {
1654 .rnr_retry_count = 7,
1655 .private_data = &msg,
1656 .private_data_len = sizeof(msg),
1659 msg = (struct rtrs_msg_conn_req) {
1660 .magic = cpu_to_le16(RTRS_MAGIC),
1661 .version = cpu_to_le16(RTRS_PROTO_VER),
1662 .cid = cpu_to_le16(con->c.cid),
1663 .cid_num = cpu_to_le16(sess->s.con_num),
1664 .recon_cnt = cpu_to_le16(sess->s.recon_cnt),
1666 uuid_copy(&msg.sess_uuid, &sess->s.uuid);
1667 uuid_copy(&msg.paths_uuid, &clt->paths_uuid);
1669 err = rdma_connect_locked(con->c.cm_id, ¶m);
1671 rtrs_err(clt, "rdma_connect_locked(): %d\n", err);
1676 static int rtrs_rdma_conn_established(struct rtrs_clt_con *con,
1677 struct rdma_cm_event *ev)
1679 struct rtrs_clt_sess *sess = to_clt_sess(con->c.sess);
1680 struct rtrs_clt *clt = sess->clt;
1681 const struct rtrs_msg_conn_rsp *msg;
1682 u16 version, queue_depth;
1686 msg = ev->param.conn.private_data;
1687 len = ev->param.conn.private_data_len;
1688 if (len < sizeof(*msg)) {
1689 rtrs_err(clt, "Invalid RTRS connection response\n");
1692 if (le16_to_cpu(msg->magic) != RTRS_MAGIC) {
1693 rtrs_err(clt, "Invalid RTRS magic\n");
1696 version = le16_to_cpu(msg->version);
1697 if (version >> 8 != RTRS_PROTO_VER_MAJOR) {
1698 rtrs_err(clt, "Unsupported major RTRS version: %d, expected %d\n",
1699 version >> 8, RTRS_PROTO_VER_MAJOR);
1702 errno = le16_to_cpu(msg->errno);
1704 rtrs_err(clt, "Invalid RTRS message: errno %d\n",
1708 if (con->c.cid == 0) {
1709 queue_depth = le16_to_cpu(msg->queue_depth);
1711 if (queue_depth > MAX_SESS_QUEUE_DEPTH) {
1712 rtrs_err(clt, "Invalid RTRS message: queue=%d\n",
1716 if (!sess->rbufs || sess->queue_depth < queue_depth) {
1718 sess->rbufs = kcalloc(queue_depth, sizeof(*sess->rbufs),
1723 sess->queue_depth = queue_depth;
1724 sess->max_hdr_size = le32_to_cpu(msg->max_hdr_size);
1725 sess->max_io_size = le32_to_cpu(msg->max_io_size);
1726 sess->flags = le32_to_cpu(msg->flags);
1727 sess->chunk_size = sess->max_io_size + sess->max_hdr_size;
1730 * Global queue depth and IO size is always a minimum.
1731 * If while a reconnection server sends us a value a bit
1732 * higher - client does not care and uses cached minimum.
1734 * Since we can have several sessions (paths) restablishing
1735 * connections in parallel, use lock.
1737 mutex_lock(&clt->paths_mutex);
1738 clt->queue_depth = min_not_zero(sess->queue_depth,
1740 clt->max_io_size = min_not_zero(sess->max_io_size,
1742 mutex_unlock(&clt->paths_mutex);
1745 * Cache the hca_port and hca_name for sysfs
1747 sess->hca_port = con->c.cm_id->port_num;
1748 scnprintf(sess->hca_name, sizeof(sess->hca_name),
1749 sess->s.dev->ib_dev->name);
1750 sess->s.src_addr = con->c.cm_id->route.addr.src_addr;
1756 static inline void flag_success_on_conn(struct rtrs_clt_con *con)
1758 struct rtrs_clt_sess *sess = to_clt_sess(con->c.sess);
1760 atomic_inc(&sess->connected_cnt);
1764 static int rtrs_rdma_conn_rejected(struct rtrs_clt_con *con,
1765 struct rdma_cm_event *ev)
1767 struct rtrs_sess *s = con->c.sess;
1768 const struct rtrs_msg_conn_rsp *msg;
1769 const char *rej_msg;
1773 status = ev->status;
1774 rej_msg = rdma_reject_msg(con->c.cm_id, status);
1775 msg = rdma_consumer_reject_data(con->c.cm_id, ev, &data_len);
1777 if (msg && data_len >= sizeof(*msg)) {
1778 errno = (int16_t)le16_to_cpu(msg->errno);
1779 if (errno == -EBUSY)
1781 "Previous session is still exists on the server, please reconnect later\n");
1784 "Connect rejected: status %d (%s), rtrs errno %d\n",
1785 status, rej_msg, errno);
1788 "Connect rejected but with malformed message: status %d (%s)\n",
1795 static void rtrs_clt_close_conns(struct rtrs_clt_sess *sess, bool wait)
1797 if (rtrs_clt_change_state(sess, RTRS_CLT_CLOSING))
1798 queue_work(rtrs_wq, &sess->close_work);
1800 flush_work(&sess->close_work);
1803 static inline void flag_error_on_conn(struct rtrs_clt_con *con, int cm_err)
1805 if (con->cm_err == 1) {
1806 struct rtrs_clt_sess *sess;
1808 sess = to_clt_sess(con->c.sess);
1809 if (atomic_dec_and_test(&sess->connected_cnt))
1811 wake_up(&sess->state_wq);
1813 con->cm_err = cm_err;
1816 static int rtrs_clt_rdma_cm_handler(struct rdma_cm_id *cm_id,
1817 struct rdma_cm_event *ev)
1819 struct rtrs_clt_con *con = cm_id->context;
1820 struct rtrs_sess *s = con->c.sess;
1821 struct rtrs_clt_sess *sess = to_clt_sess(s);
1824 switch (ev->event) {
1825 case RDMA_CM_EVENT_ADDR_RESOLVED:
1826 cm_err = rtrs_rdma_addr_resolved(con);
1828 case RDMA_CM_EVENT_ROUTE_RESOLVED:
1829 cm_err = rtrs_rdma_route_resolved(con);
1831 case RDMA_CM_EVENT_ESTABLISHED:
1832 cm_err = rtrs_rdma_conn_established(con, ev);
1833 if (likely(!cm_err)) {
1835 * Report success and wake up. Here we abuse state_wq,
1836 * i.e. wake up without state change, but we set cm_err.
1838 flag_success_on_conn(con);
1839 wake_up(&sess->state_wq);
1843 case RDMA_CM_EVENT_REJECTED:
1844 cm_err = rtrs_rdma_conn_rejected(con, ev);
1846 case RDMA_CM_EVENT_DISCONNECTED:
1847 /* No message for disconnecting */
1848 cm_err = -ECONNRESET;
1850 case RDMA_CM_EVENT_CONNECT_ERROR:
1851 case RDMA_CM_EVENT_UNREACHABLE:
1852 case RDMA_CM_EVENT_ADDR_CHANGE:
1853 case RDMA_CM_EVENT_TIMEWAIT_EXIT:
1854 rtrs_wrn(s, "CM error event %d\n", ev->event);
1855 cm_err = -ECONNRESET;
1857 case RDMA_CM_EVENT_ADDR_ERROR:
1858 case RDMA_CM_EVENT_ROUTE_ERROR:
1859 rtrs_wrn(s, "CM error event %d\n", ev->event);
1860 cm_err = -EHOSTUNREACH;
1862 case RDMA_CM_EVENT_DEVICE_REMOVAL:
1864 * Device removal is a special case. Queue close and return 0.
1866 rtrs_clt_close_conns(sess, false);
1869 rtrs_err(s, "Unexpected RDMA CM event (%d)\n", ev->event);
1870 cm_err = -ECONNRESET;
1876 * cm error makes sense only on connection establishing,
1877 * in other cases we rely on normal procedure of reconnecting.
1879 flag_error_on_conn(con, cm_err);
1880 rtrs_rdma_error_recovery(con);
1886 static int create_cm(struct rtrs_clt_con *con)
1888 struct rtrs_sess *s = con->c.sess;
1889 struct rtrs_clt_sess *sess = to_clt_sess(s);
1890 struct rdma_cm_id *cm_id;
1893 cm_id = rdma_create_id(&init_net, rtrs_clt_rdma_cm_handler, con,
1894 sess->s.dst_addr.ss_family == AF_IB ?
1895 RDMA_PS_IB : RDMA_PS_TCP, IB_QPT_RC);
1896 if (IS_ERR(cm_id)) {
1897 err = PTR_ERR(cm_id);
1898 rtrs_err(s, "Failed to create CM ID, err: %d\n", err);
1902 con->c.cm_id = cm_id;
1904 /* allow the port to be reused */
1905 err = rdma_set_reuseaddr(cm_id, 1);
1907 rtrs_err(s, "Set address reuse failed, err: %d\n", err);
1910 err = rdma_resolve_addr(cm_id, (struct sockaddr *)&sess->s.src_addr,
1911 (struct sockaddr *)&sess->s.dst_addr,
1912 RTRS_CONNECT_TIMEOUT_MS);
1914 rtrs_err(s, "Failed to resolve address, err: %d\n", err);
1918 * Combine connection status and session events. This is needed
1919 * for waiting two possible cases: cm_err has something meaningful
1920 * or session state was really changed to error by device removal.
1922 err = wait_event_interruptible_timeout(
1924 con->cm_err || sess->state != RTRS_CLT_CONNECTING,
1925 msecs_to_jiffies(RTRS_CONNECT_TIMEOUT_MS));
1926 if (err == 0 || err == -ERESTARTSYS) {
1929 /* Timedout or interrupted */
1932 if (con->cm_err < 0) {
1936 if (READ_ONCE(sess->state) != RTRS_CLT_CONNECTING) {
1937 /* Device removal */
1938 err = -ECONNABORTED;
1946 mutex_lock(&con->con_mutex);
1947 destroy_con_cq_qp(con);
1948 mutex_unlock(&con->con_mutex);
1955 static void rtrs_clt_sess_up(struct rtrs_clt_sess *sess)
1957 struct rtrs_clt *clt = sess->clt;
1961 * We can fire RECONNECTED event only when all paths were
1962 * connected on rtrs_clt_open(), then each was disconnected
1963 * and the first one connected again. That's why this nasty
1964 * game with counter value.
1967 mutex_lock(&clt->paths_ev_mutex);
1968 up = ++clt->paths_up;
1970 * Here it is safe to access paths num directly since up counter
1971 * is greater than MAX_PATHS_NUM only while rtrs_clt_open() is
1972 * in progress, thus paths removals are impossible.
1974 if (up > MAX_PATHS_NUM && up == MAX_PATHS_NUM + clt->paths_num)
1975 clt->paths_up = clt->paths_num;
1977 clt->link_ev(clt->priv, RTRS_CLT_LINK_EV_RECONNECTED);
1978 mutex_unlock(&clt->paths_ev_mutex);
1980 /* Mark session as established */
1981 sess->established = true;
1982 sess->reconnect_attempts = 0;
1983 sess->stats->reconnects.successful_cnt++;
1986 static void rtrs_clt_sess_down(struct rtrs_clt_sess *sess)
1988 struct rtrs_clt *clt = sess->clt;
1990 if (!sess->established)
1993 sess->established = false;
1994 mutex_lock(&clt->paths_ev_mutex);
1995 WARN_ON(!clt->paths_up);
1996 if (--clt->paths_up == 0)
1997 clt->link_ev(clt->priv, RTRS_CLT_LINK_EV_DISCONNECTED);
1998 mutex_unlock(&clt->paths_ev_mutex);
2001 static void rtrs_clt_stop_and_destroy_conns(struct rtrs_clt_sess *sess)
2003 struct rtrs_clt_con *con;
2006 WARN_ON(READ_ONCE(sess->state) == RTRS_CLT_CONNECTED);
2009 * Possible race with rtrs_clt_open(), when DEVICE_REMOVAL comes
2010 * exactly in between. Start destroying after it finishes.
2012 mutex_lock(&sess->init_mutex);
2013 mutex_unlock(&sess->init_mutex);
2016 * All IO paths must observe !CONNECTED state before we
2021 rtrs_clt_stop_hb(sess);
2024 * The order it utterly crucial: firstly disconnect and complete all
2025 * rdma requests with error (thus set in_use=false for requests),
2026 * then fail outstanding requests checking in_use for each, and
2027 * eventually notify upper layer about session disconnection.
2030 for (cid = 0; cid < sess->s.con_num; cid++) {
2031 if (!sess->s.con[cid])
2033 con = to_clt_con(sess->s.con[cid]);
2036 fail_all_outstanding_reqs(sess);
2037 free_sess_reqs(sess);
2038 rtrs_clt_sess_down(sess);
2041 * Wait for graceful shutdown, namely when peer side invokes
2042 * rdma_disconnect(). 'connected_cnt' is decremented only on
2043 * CM events, thus if other side had crashed and hb has detected
2044 * something is wrong, here we will stuck for exactly timeout ms,
2045 * since CM does not fire anything. That is fine, we are not in
2048 wait_event_timeout(sess->state_wq, !atomic_read(&sess->connected_cnt),
2049 msecs_to_jiffies(RTRS_CONNECT_TIMEOUT_MS));
2051 for (cid = 0; cid < sess->s.con_num; cid++) {
2052 if (!sess->s.con[cid])
2054 con = to_clt_con(sess->s.con[cid]);
2055 mutex_lock(&con->con_mutex);
2056 destroy_con_cq_qp(con);
2057 mutex_unlock(&con->con_mutex);
2063 static inline bool xchg_sessions(struct rtrs_clt_sess __rcu **rcu_ppcpu_path,
2064 struct rtrs_clt_sess *sess,
2065 struct rtrs_clt_sess *next)
2067 struct rtrs_clt_sess **ppcpu_path;
2069 /* Call cmpxchg() without sparse warnings */
2070 ppcpu_path = (typeof(ppcpu_path))rcu_ppcpu_path;
2071 return sess == cmpxchg(ppcpu_path, sess, next);
2074 static void rtrs_clt_remove_path_from_arr(struct rtrs_clt_sess *sess)
2076 struct rtrs_clt *clt = sess->clt;
2077 struct rtrs_clt_sess *next;
2078 bool wait_for_grace = false;
2081 mutex_lock(&clt->paths_mutex);
2082 list_del_rcu(&sess->s.entry);
2084 /* Make sure everybody observes path removal. */
2088 * At this point nobody sees @sess in the list, but still we have
2089 * dangling pointer @pcpu_path which _can_ point to @sess. Since
2090 * nobody can observe @sess in the list, we guarantee that IO path
2091 * will not assign @sess to @pcpu_path, i.e. @pcpu_path can be equal
2092 * to @sess, but can never again become @sess.
2096 * Decrement paths number only after grace period, because
2097 * caller of do_each_path() must firstly observe list without
2098 * path and only then decremented paths number.
2100 * Otherwise there can be the following situation:
2101 * o Two paths exist and IO is coming.
2102 * o One path is removed:
2104 * do_each_path(): rtrs_clt_remove_path_from_arr():
2105 * path = get_next_path()
2106 * ^^^ list_del_rcu(path)
2107 * [!CONNECTED path] clt->paths_num--
2109 * load clt->paths_num from 2 to 1
2113 * path is observed as !CONNECTED, but do_each_path() loop
2114 * ends, because expression i < clt->paths_num is false.
2119 * Get @next connection from current @sess which is going to be
2120 * removed. If @sess is the last element, then @next is NULL.
2123 next = list_next_or_null_rr_rcu(&clt->paths_list, &sess->s.entry,
2124 typeof(*next), s.entry);
2128 * @pcpu paths can still point to the path which is going to be
2129 * removed, so change the pointer manually.
2131 for_each_possible_cpu(cpu) {
2132 struct rtrs_clt_sess __rcu **ppcpu_path;
2134 ppcpu_path = per_cpu_ptr(clt->pcpu_path, cpu);
2135 if (rcu_dereference_protected(*ppcpu_path,
2136 lockdep_is_held(&clt->paths_mutex)) != sess)
2138 * synchronize_rcu() was called just after deleting
2139 * entry from the list, thus IO code path cannot
2140 * change pointer back to the pointer which is going
2141 * to be removed, we are safe here.
2146 * We race with IO code path, which also changes pointer,
2147 * thus we have to be careful not to overwrite it.
2149 if (xchg_sessions(ppcpu_path, sess, next))
2151 * @ppcpu_path was successfully replaced with @next,
2152 * that means that someone could also pick up the
2153 * @sess and dereferencing it right now, so wait for
2154 * a grace period is required.
2156 wait_for_grace = true;
2161 mutex_unlock(&clt->paths_mutex);
2164 static void rtrs_clt_add_path_to_arr(struct rtrs_clt_sess *sess)
2166 struct rtrs_clt *clt = sess->clt;
2168 mutex_lock(&clt->paths_mutex);
2171 list_add_tail_rcu(&sess->s.entry, &clt->paths_list);
2172 mutex_unlock(&clt->paths_mutex);
2175 static void rtrs_clt_close_work(struct work_struct *work)
2177 struct rtrs_clt_sess *sess;
2179 sess = container_of(work, struct rtrs_clt_sess, close_work);
2181 cancel_delayed_work_sync(&sess->reconnect_dwork);
2182 rtrs_clt_stop_and_destroy_conns(sess);
2183 rtrs_clt_change_state(sess, RTRS_CLT_CLOSED);
2186 static int init_conns(struct rtrs_clt_sess *sess)
2192 * On every new session connections increase reconnect counter
2193 * to avoid clashes with previous sessions not yet closed
2194 * sessions on a server side.
2196 sess->s.recon_cnt++;
2198 /* Establish all RDMA connections */
2199 for (cid = 0; cid < sess->s.con_num; cid++) {
2200 err = create_con(sess, cid);
2204 err = create_cm(to_clt_con(sess->s.con[cid]));
2206 destroy_con(to_clt_con(sess->s.con[cid]));
2210 err = alloc_sess_reqs(sess);
2214 rtrs_clt_start_hb(sess);
2220 struct rtrs_clt_con *con = to_clt_con(sess->s.con[cid]);
2224 mutex_lock(&con->con_mutex);
2225 destroy_con_cq_qp(con);
2226 mutex_unlock(&con->con_mutex);
2231 * If we've never taken async path and got an error, say,
2232 * doing rdma_resolve_addr(), switch to CONNECTION_ERR state
2233 * manually to keep reconnecting.
2235 rtrs_clt_change_state(sess, RTRS_CLT_CONNECTING_ERR);
2240 static void rtrs_clt_info_req_done(struct ib_cq *cq, struct ib_wc *wc)
2242 struct rtrs_clt_con *con = cq->cq_context;
2243 struct rtrs_clt_sess *sess = to_clt_sess(con->c.sess);
2246 iu = container_of(wc->wr_cqe, struct rtrs_iu, cqe);
2247 rtrs_iu_free(iu, sess->s.dev->ib_dev, 1);
2249 if (unlikely(wc->status != IB_WC_SUCCESS)) {
2250 rtrs_err(sess->clt, "Sess info request send failed: %s\n",
2251 ib_wc_status_msg(wc->status));
2252 rtrs_clt_change_state(sess, RTRS_CLT_CONNECTING_ERR);
2256 rtrs_clt_update_wc_stats(con);
2259 static int process_info_rsp(struct rtrs_clt_sess *sess,
2260 const struct rtrs_msg_info_rsp *msg)
2262 unsigned int sg_cnt, total_len;
2265 sg_cnt = le16_to_cpu(msg->sg_cnt);
2266 if (unlikely(!sg_cnt || (sess->queue_depth % sg_cnt))) {
2267 rtrs_err(sess->clt, "Incorrect sg_cnt %d, is not multiple\n",
2273 * Check if IB immediate data size is enough to hold the mem_id and
2274 * the offset inside the memory chunk.
2276 if (unlikely((ilog2(sg_cnt - 1) + 1) +
2277 (ilog2(sess->chunk_size - 1) + 1) >
2278 MAX_IMM_PAYL_BITS)) {
2280 "RDMA immediate size (%db) not enough to encode %d buffers of size %dB\n",
2281 MAX_IMM_PAYL_BITS, sg_cnt, sess->chunk_size);
2285 for (sgi = 0, i = 0; sgi < sg_cnt && i < sess->queue_depth; sgi++) {
2286 const struct rtrs_sg_desc *desc = &msg->desc[sgi];
2290 addr = le64_to_cpu(desc->addr);
2291 rkey = le32_to_cpu(desc->key);
2292 len = le32_to_cpu(desc->len);
2296 if (unlikely(!len || (len % sess->chunk_size))) {
2297 rtrs_err(sess->clt, "Incorrect [%d].len %d\n", sgi,
2301 for ( ; len && i < sess->queue_depth; i++) {
2302 sess->rbufs[i].addr = addr;
2303 sess->rbufs[i].rkey = rkey;
2305 len -= sess->chunk_size;
2306 addr += sess->chunk_size;
2310 if (unlikely(sgi != sg_cnt || i != sess->queue_depth)) {
2311 rtrs_err(sess->clt, "Incorrect sg vector, not fully mapped\n");
2314 if (unlikely(total_len != sess->chunk_size * sess->queue_depth)) {
2315 rtrs_err(sess->clt, "Incorrect total_len %d\n", total_len);
2322 static void rtrs_clt_info_rsp_done(struct ib_cq *cq, struct ib_wc *wc)
2324 struct rtrs_clt_con *con = cq->cq_context;
2325 struct rtrs_clt_sess *sess = to_clt_sess(con->c.sess);
2326 struct rtrs_msg_info_rsp *msg;
2327 enum rtrs_clt_state state;
2332 state = RTRS_CLT_CONNECTING_ERR;
2334 WARN_ON(con->c.cid);
2335 iu = container_of(wc->wr_cqe, struct rtrs_iu, cqe);
2336 if (unlikely(wc->status != IB_WC_SUCCESS)) {
2337 rtrs_err(sess->clt, "Sess info response recv failed: %s\n",
2338 ib_wc_status_msg(wc->status));
2341 WARN_ON(wc->opcode != IB_WC_RECV);
2343 if (unlikely(wc->byte_len < sizeof(*msg))) {
2344 rtrs_err(sess->clt, "Sess info response is malformed: size %d\n",
2348 ib_dma_sync_single_for_cpu(sess->s.dev->ib_dev, iu->dma_addr,
2349 iu->size, DMA_FROM_DEVICE);
2351 if (unlikely(le16_to_cpu(msg->type) != RTRS_MSG_INFO_RSP)) {
2352 rtrs_err(sess->clt, "Sess info response is malformed: type %d\n",
2353 le16_to_cpu(msg->type));
2356 rx_sz = sizeof(*msg);
2357 rx_sz += sizeof(msg->desc[0]) * le16_to_cpu(msg->sg_cnt);
2358 if (unlikely(wc->byte_len < rx_sz)) {
2359 rtrs_err(sess->clt, "Sess info response is malformed: size %d\n",
2363 err = process_info_rsp(sess, msg);
2367 err = post_recv_sess(sess);
2371 state = RTRS_CLT_CONNECTED;
2374 rtrs_clt_update_wc_stats(con);
2375 rtrs_iu_free(iu, sess->s.dev->ib_dev, 1);
2376 rtrs_clt_change_state(sess, state);
2379 static int rtrs_send_sess_info(struct rtrs_clt_sess *sess)
2381 struct rtrs_clt_con *usr_con = to_clt_con(sess->s.con[0]);
2382 struct rtrs_msg_info_req *msg;
2383 struct rtrs_iu *tx_iu, *rx_iu;
2387 rx_sz = sizeof(struct rtrs_msg_info_rsp);
2388 rx_sz += sizeof(u64) * MAX_SESS_QUEUE_DEPTH;
2390 tx_iu = rtrs_iu_alloc(1, sizeof(struct rtrs_msg_info_req), GFP_KERNEL,
2391 sess->s.dev->ib_dev, DMA_TO_DEVICE,
2392 rtrs_clt_info_req_done);
2393 rx_iu = rtrs_iu_alloc(1, rx_sz, GFP_KERNEL, sess->s.dev->ib_dev,
2394 DMA_FROM_DEVICE, rtrs_clt_info_rsp_done);
2395 if (unlikely(!tx_iu || !rx_iu)) {
2399 /* Prepare for getting info response */
2400 err = rtrs_iu_post_recv(&usr_con->c, rx_iu);
2401 if (unlikely(err)) {
2402 rtrs_err(sess->clt, "rtrs_iu_post_recv(), err: %d\n", err);
2408 msg->type = cpu_to_le16(RTRS_MSG_INFO_REQ);
2409 memcpy(msg->sessname, sess->s.sessname, sizeof(msg->sessname));
2411 ib_dma_sync_single_for_device(sess->s.dev->ib_dev, tx_iu->dma_addr,
2412 tx_iu->size, DMA_TO_DEVICE);
2414 /* Send info request */
2415 err = rtrs_iu_post_send(&usr_con->c, tx_iu, sizeof(*msg), NULL);
2416 if (unlikely(err)) {
2417 rtrs_err(sess->clt, "rtrs_iu_post_send(), err: %d\n", err);
2422 /* Wait for state change */
2423 wait_event_interruptible_timeout(sess->state_wq,
2424 sess->state != RTRS_CLT_CONNECTING,
2426 RTRS_CONNECT_TIMEOUT_MS));
2427 if (unlikely(READ_ONCE(sess->state) != RTRS_CLT_CONNECTED)) {
2428 if (READ_ONCE(sess->state) == RTRS_CLT_CONNECTING_ERR)
2437 rtrs_iu_free(tx_iu, sess->s.dev->ib_dev, 1);
2439 rtrs_iu_free(rx_iu, sess->s.dev->ib_dev, 1);
2441 /* If we've never taken async path because of malloc problems */
2442 rtrs_clt_change_state(sess, RTRS_CLT_CONNECTING_ERR);
2448 * init_sess() - establishes all session connections and does handshake
2449 * @sess: client session.
2450 * In case of error full close or reconnect procedure should be taken,
2451 * because reconnect or close async works can be started.
2453 static int init_sess(struct rtrs_clt_sess *sess)
2457 mutex_lock(&sess->init_mutex);
2458 err = init_conns(sess);
2460 rtrs_err(sess->clt, "init_conns(), err: %d\n", err);
2463 err = rtrs_send_sess_info(sess);
2465 rtrs_err(sess->clt, "rtrs_send_sess_info(), err: %d\n", err);
2468 rtrs_clt_sess_up(sess);
2470 mutex_unlock(&sess->init_mutex);
2475 static void rtrs_clt_reconnect_work(struct work_struct *work)
2477 struct rtrs_clt_sess *sess;
2478 struct rtrs_clt *clt;
2479 unsigned int delay_ms;
2482 sess = container_of(to_delayed_work(work), struct rtrs_clt_sess,
2486 if (READ_ONCE(sess->state) != RTRS_CLT_RECONNECTING)
2489 if (sess->reconnect_attempts >= clt->max_reconnect_attempts) {
2490 /* Close a session completely if max attempts is reached */
2491 rtrs_clt_close_conns(sess, false);
2494 sess->reconnect_attempts++;
2496 /* Stop everything */
2497 rtrs_clt_stop_and_destroy_conns(sess);
2498 msleep(RTRS_RECONNECT_BACKOFF);
2499 if (rtrs_clt_change_state(sess, RTRS_CLT_CONNECTING)) {
2500 err = init_sess(sess);
2502 goto reconnect_again;
2508 if (rtrs_clt_change_state(sess, RTRS_CLT_RECONNECTING)) {
2509 sess->stats->reconnects.fail_cnt++;
2510 delay_ms = clt->reconnect_delay_sec * 1000;
2511 queue_delayed_work(rtrs_wq, &sess->reconnect_dwork,
2512 msecs_to_jiffies(delay_ms +
2514 RTRS_RECONNECT_SEED));
2518 static void rtrs_clt_dev_release(struct device *dev)
2520 struct rtrs_clt *clt = container_of(dev, struct rtrs_clt, dev);
2525 static struct rtrs_clt *alloc_clt(const char *sessname, size_t paths_num,
2526 u16 port, size_t pdu_sz, void *priv,
2527 void (*link_ev)(void *priv,
2528 enum rtrs_clt_link_ev ev),
2529 unsigned int max_segments,
2530 size_t max_segment_size,
2531 unsigned int reconnect_delay_sec,
2532 unsigned int max_reconnect_attempts)
2534 struct rtrs_clt *clt;
2537 if (!paths_num || paths_num > MAX_PATHS_NUM)
2538 return ERR_PTR(-EINVAL);
2540 if (strlen(sessname) >= sizeof(clt->sessname))
2541 return ERR_PTR(-EINVAL);
2543 clt = kzalloc(sizeof(*clt), GFP_KERNEL);
2545 return ERR_PTR(-ENOMEM);
2547 clt->pcpu_path = alloc_percpu(typeof(*clt->pcpu_path));
2548 if (!clt->pcpu_path) {
2550 return ERR_PTR(-ENOMEM);
2553 uuid_gen(&clt->paths_uuid);
2554 INIT_LIST_HEAD_RCU(&clt->paths_list);
2555 clt->paths_num = paths_num;
2556 clt->paths_up = MAX_PATHS_NUM;
2558 clt->pdu_sz = pdu_sz;
2559 clt->max_segments = max_segments;
2560 clt->max_segment_size = max_segment_size;
2561 clt->reconnect_delay_sec = reconnect_delay_sec;
2562 clt->max_reconnect_attempts = max_reconnect_attempts;
2564 clt->link_ev = link_ev;
2565 clt->mp_policy = MP_POLICY_MIN_INFLIGHT;
2566 strlcpy(clt->sessname, sessname, sizeof(clt->sessname));
2567 init_waitqueue_head(&clt->permits_wait);
2568 mutex_init(&clt->paths_ev_mutex);
2569 mutex_init(&clt->paths_mutex);
2571 clt->dev.class = rtrs_clt_dev_class;
2572 clt->dev.release = rtrs_clt_dev_release;
2573 err = dev_set_name(&clt->dev, "%s", sessname);
2575 free_percpu(clt->pcpu_path);
2577 return ERR_PTR(err);
2580 * Suppress user space notification until
2581 * sysfs files are created
2583 dev_set_uevent_suppress(&clt->dev, true);
2584 err = device_register(&clt->dev);
2586 free_percpu(clt->pcpu_path);
2587 put_device(&clt->dev);
2588 return ERR_PTR(err);
2591 clt->kobj_paths = kobject_create_and_add("paths", &clt->dev.kobj);
2592 if (!clt->kobj_paths) {
2593 free_percpu(clt->pcpu_path);
2594 device_unregister(&clt->dev);
2597 err = rtrs_clt_create_sysfs_root_files(clt);
2599 free_percpu(clt->pcpu_path);
2600 kobject_del(clt->kobj_paths);
2601 kobject_put(clt->kobj_paths);
2602 device_unregister(&clt->dev);
2603 return ERR_PTR(err);
2605 dev_set_uevent_suppress(&clt->dev, false);
2606 kobject_uevent(&clt->dev.kobj, KOBJ_ADD);
2611 static void wait_for_inflight_permits(struct rtrs_clt *clt)
2613 if (clt->permits_map) {
2614 size_t sz = clt->queue_depth;
2616 wait_event(clt->permits_wait,
2617 find_first_bit(clt->permits_map, sz) >= sz);
2621 static void free_clt(struct rtrs_clt *clt)
2623 wait_for_inflight_permits(clt);
2625 free_percpu(clt->pcpu_path);
2626 mutex_destroy(&clt->paths_ev_mutex);
2627 mutex_destroy(&clt->paths_mutex);
2628 /* release callback will free clt in last put */
2629 device_unregister(&clt->dev);
2633 * rtrs_clt_open() - Open a session to an RTRS server
2634 * @ops: holds the link event callback and the private pointer.
2635 * @sessname: name of the session
2636 * @paths: Paths to be established defined by their src and dst addresses
2637 * @paths_num: Number of elements in the @paths array
2638 * @port: port to be used by the RTRS session
2639 * @pdu_sz: Size of extra payload which can be accessed after permit allocation.
2640 * @reconnect_delay_sec: time between reconnect tries
2641 * @max_segments: Max. number of segments per IO request
2642 * @max_segment_size: Max. size of one segment
2643 * @max_reconnect_attempts: Number of times to reconnect on error before giving
2644 * up, 0 for * disabled, -1 for forever
2646 * Starts session establishment with the rtrs_server. The function can block
2647 * up to ~2000ms before it returns.
2649 * Return a valid pointer on success otherwise PTR_ERR.
2651 struct rtrs_clt *rtrs_clt_open(struct rtrs_clt_ops *ops,
2652 const char *sessname,
2653 const struct rtrs_addr *paths,
2654 size_t paths_num, u16 port,
2655 size_t pdu_sz, u8 reconnect_delay_sec,
2657 size_t max_segment_size,
2658 s16 max_reconnect_attempts)
2660 struct rtrs_clt_sess *sess, *tmp;
2661 struct rtrs_clt *clt;
2664 clt = alloc_clt(sessname, paths_num, port, pdu_sz, ops->priv,
2666 max_segments, max_segment_size, reconnect_delay_sec,
2667 max_reconnect_attempts);
2672 for (i = 0; i < paths_num; i++) {
2673 struct rtrs_clt_sess *sess;
2675 sess = alloc_sess(clt, &paths[i], nr_cpu_ids,
2676 max_segments, max_segment_size);
2678 err = PTR_ERR(sess);
2679 goto close_all_sess;
2681 list_add_tail_rcu(&sess->s.entry, &clt->paths_list);
2683 err = init_sess(sess);
2685 list_del_rcu(&sess->s.entry);
2686 rtrs_clt_close_conns(sess, true);
2688 goto close_all_sess;
2691 err = rtrs_clt_create_sess_files(sess);
2693 list_del_rcu(&sess->s.entry);
2694 rtrs_clt_close_conns(sess, true);
2696 goto close_all_sess;
2699 err = alloc_permits(clt);
2701 goto close_all_sess;
2706 list_for_each_entry_safe(sess, tmp, &clt->paths_list, s.entry) {
2707 rtrs_clt_destroy_sess_files(sess, NULL);
2708 rtrs_clt_close_conns(sess, true);
2709 kobject_put(&sess->kobj);
2711 rtrs_clt_destroy_sysfs_root_files(clt);
2712 rtrs_clt_destroy_sysfs_root_folders(clt);
2716 return ERR_PTR(err);
2718 EXPORT_SYMBOL(rtrs_clt_open);
2721 * rtrs_clt_close() - Close a session
2722 * @clt: Session handle. Session is freed upon return.
2724 void rtrs_clt_close(struct rtrs_clt *clt)
2726 struct rtrs_clt_sess *sess, *tmp;
2728 /* Firstly forbid sysfs access */
2729 rtrs_clt_destroy_sysfs_root_files(clt);
2730 rtrs_clt_destroy_sysfs_root_folders(clt);
2732 /* Now it is safe to iterate over all paths without locks */
2733 list_for_each_entry_safe(sess, tmp, &clt->paths_list, s.entry) {
2734 rtrs_clt_destroy_sess_files(sess, NULL);
2735 rtrs_clt_close_conns(sess, true);
2736 kobject_put(&sess->kobj);
2740 EXPORT_SYMBOL(rtrs_clt_close);
2742 int rtrs_clt_reconnect_from_sysfs(struct rtrs_clt_sess *sess)
2744 enum rtrs_clt_state old_state;
2748 changed = rtrs_clt_change_state_get_old(sess, RTRS_CLT_RECONNECTING,
2751 sess->reconnect_attempts = 0;
2752 queue_delayed_work(rtrs_wq, &sess->reconnect_dwork, 0);
2754 if (changed || old_state == RTRS_CLT_RECONNECTING) {
2756 * flush_delayed_work() queues pending work for immediate
2757 * execution, so do the flush if we have queued something
2758 * right now or work is pending.
2760 flush_delayed_work(&sess->reconnect_dwork);
2761 err = (READ_ONCE(sess->state) ==
2762 RTRS_CLT_CONNECTED ? 0 : -ENOTCONN);
2768 int rtrs_clt_disconnect_from_sysfs(struct rtrs_clt_sess *sess)
2770 rtrs_clt_close_conns(sess, true);
2775 int rtrs_clt_remove_path_from_sysfs(struct rtrs_clt_sess *sess,
2776 const struct attribute *sysfs_self)
2778 enum rtrs_clt_state old_state;
2782 * Continue stopping path till state was changed to DEAD or
2783 * state was observed as DEAD:
2784 * 1. State was changed to DEAD - we were fast and nobody
2785 * invoked rtrs_clt_reconnect(), which can again start
2787 * 2. State was observed as DEAD - we have someone in parallel
2788 * removing the path.
2791 rtrs_clt_close_conns(sess, true);
2792 changed = rtrs_clt_change_state_get_old(sess,
2795 } while (!changed && old_state != RTRS_CLT_DEAD);
2797 if (likely(changed)) {
2798 rtrs_clt_destroy_sess_files(sess, sysfs_self);
2799 rtrs_clt_remove_path_from_arr(sess);
2800 kobject_put(&sess->kobj);
2806 void rtrs_clt_set_max_reconnect_attempts(struct rtrs_clt *clt, int value)
2808 clt->max_reconnect_attempts = (unsigned int)value;
2811 int rtrs_clt_get_max_reconnect_attempts(const struct rtrs_clt *clt)
2813 return (int)clt->max_reconnect_attempts;
2817 * rtrs_clt_request() - Request data transfer to/from server via RDMA.
2820 * @ops: callback function to be called as confirmation, and the pointer.
2822 * @permit: Preallocated permit
2823 * @vec: Message that is sent to server together with the request.
2824 * Sum of len of all @vec elements limited to <= IO_MSG_SIZE.
2825 * Since the msg is copied internally it can be allocated on stack.
2826 * @nr: Number of elements in @vec.
2827 * @data_len: length of data sent to/from server
2828 * @sg: Pages to be sent/received to/from server.
2829 * @sg_cnt: Number of elements in the @sg
2835 * On dir=READ rtrs client will request a data transfer from Server to client.
2836 * The data that the server will respond with will be stored in @sg when
2837 * the user receives an %RTRS_CLT_RDMA_EV_RDMA_REQUEST_WRITE_COMPL event.
2838 * On dir=WRITE rtrs client will rdma write data in sg to server side.
2840 int rtrs_clt_request(int dir, struct rtrs_clt_req_ops *ops,
2841 struct rtrs_clt *clt, struct rtrs_permit *permit,
2842 const struct kvec *vec, size_t nr, size_t data_len,
2843 struct scatterlist *sg, unsigned int sg_cnt)
2845 struct rtrs_clt_io_req *req;
2846 struct rtrs_clt_sess *sess;
2848 enum dma_data_direction dma_dir;
2849 int err = -ECONNABORTED, i;
2850 size_t usr_len, hdr_len;
2853 /* Get kvec length */
2854 for (i = 0, usr_len = 0; i < nr; i++)
2855 usr_len += vec[i].iov_len;
2858 hdr_len = sizeof(struct rtrs_msg_rdma_read) +
2859 sg_cnt * sizeof(struct rtrs_sg_desc);
2860 dma_dir = DMA_FROM_DEVICE;
2862 hdr_len = sizeof(struct rtrs_msg_rdma_write);
2863 dma_dir = DMA_TO_DEVICE;
2867 for (path_it_init(&it, clt);
2868 (sess = it.next_path(&it)) && it.i < it.clt->paths_num; it.i++) {
2869 if (unlikely(READ_ONCE(sess->state) != RTRS_CLT_CONNECTED))
2872 if (unlikely(usr_len + hdr_len > sess->max_hdr_size)) {
2873 rtrs_wrn_rl(sess->clt,
2874 "%s request failed, user message size is %zu and header length %zu, but max size is %u\n",
2875 dir == READ ? "Read" : "Write",
2876 usr_len, hdr_len, sess->max_hdr_size);
2880 req = rtrs_clt_get_req(sess, ops->conf_fn, permit, ops->priv,
2881 vec, usr_len, sg, sg_cnt, data_len,
2884 err = rtrs_clt_read_req(req);
2886 err = rtrs_clt_write_req(req);
2887 if (unlikely(err)) {
2888 req->in_use = false;
2894 path_it_deinit(&it);
2899 EXPORT_SYMBOL(rtrs_clt_request);
2902 * rtrs_clt_query() - queries RTRS session attributes
2903 *@clt: session pointer
2904 *@attr: query results for session attributes.
2907 * -ECOMM no connection to the server
2909 int rtrs_clt_query(struct rtrs_clt *clt, struct rtrs_attrs *attr)
2911 if (!rtrs_clt_is_connected(clt))
2914 attr->queue_depth = clt->queue_depth;
2915 attr->max_io_size = clt->max_io_size;
2916 attr->sess_kobj = &clt->dev.kobj;
2917 strlcpy(attr->sessname, clt->sessname, sizeof(attr->sessname));
2921 EXPORT_SYMBOL(rtrs_clt_query);
2923 int rtrs_clt_create_path_from_sysfs(struct rtrs_clt *clt,
2924 struct rtrs_addr *addr)
2926 struct rtrs_clt_sess *sess;
2929 sess = alloc_sess(clt, addr, nr_cpu_ids, clt->max_segments,
2930 clt->max_segment_size);
2932 return PTR_ERR(sess);
2935 * It is totally safe to add path in CONNECTING state: coming
2936 * IO will never grab it. Also it is very important to add
2937 * path before init, since init fires LINK_CONNECTED event.
2939 rtrs_clt_add_path_to_arr(sess);
2941 err = init_sess(sess);
2945 err = rtrs_clt_create_sess_files(sess);
2952 rtrs_clt_remove_path_from_arr(sess);
2953 rtrs_clt_close_conns(sess, true);
2959 static int rtrs_clt_ib_dev_init(struct rtrs_ib_dev *dev)
2961 if (!(dev->ib_dev->attrs.device_cap_flags &
2962 IB_DEVICE_MEM_MGT_EXTENSIONS)) {
2963 pr_err("Memory registrations not supported.\n");
2970 static const struct rtrs_rdma_dev_pd_ops dev_pd_ops = {
2971 .init = rtrs_clt_ib_dev_init
2974 static int __init rtrs_client_init(void)
2976 rtrs_rdma_dev_pd_init(0, &dev_pd);
2978 rtrs_clt_dev_class = class_create(THIS_MODULE, "rtrs-client");
2979 if (IS_ERR(rtrs_clt_dev_class)) {
2980 pr_err("Failed to create rtrs-client dev class\n");
2981 return PTR_ERR(rtrs_clt_dev_class);
2983 rtrs_wq = alloc_workqueue("rtrs_client_wq", 0, 0);
2985 class_destroy(rtrs_clt_dev_class);
2992 static void __exit rtrs_client_exit(void)
2994 destroy_workqueue(rtrs_wq);
2995 class_destroy(rtrs_clt_dev_class);
2996 rtrs_rdma_dev_pd_deinit(&dev_pd);
2999 module_init(rtrs_client_init);
3000 module_exit(rtrs_client_exit);