2 * Copyright (c) 2006, 2018 Oracle and/or its affiliates. All rights reserved.
4 * This software is available to you under a choice of one of two
5 * licenses. You may choose to be licensed under the terms of the GNU
6 * General Public License (GPL) Version 2, available from the file
7 * COPYING in the main directory of this source tree, or the
8 * OpenIB.org BSD license below:
10 * Redistribution and use in source and binary forms, with or
11 * without modification, are permitted provided that the following
14 * - Redistributions of source code must retain the above
15 * copyright notice, this list of conditions and the following
18 * - Redistributions in binary form must reproduce the above
19 * copyright notice, this list of conditions and the following
20 * disclaimer in the documentation and/or other materials
21 * provided with the distribution.
23 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
24 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
25 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
26 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
27 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
28 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
29 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
33 #include <linux/kernel.h>
34 #include <linux/moduleparam.h>
35 #include <linux/gfp.h>
38 #include <linux/list.h>
39 #include <linux/ratelimit.h>
40 #include <linux/export.h>
41 #include <linux/sizes.h>
45 /* When transmitting messages in rds_send_xmit, we need to emerge from
46 * time to time and briefly release the CPU. Otherwise the softlock watchdog
48 * Also, it seems fairer to not let one busy connection stall all the
51 * send_batch_count is the number of times we'll loop in send_xmit. Setting
52 * it to 0 will restore the old behavior (where we looped until we had
55 static int send_batch_count = SZ_1K;
56 module_param(send_batch_count, int, 0444);
57 MODULE_PARM_DESC(send_batch_count, " batch factor when working the send queue");
59 static void rds_send_remove_from_sock(struct list_head *messages, int status);
62 * Reset the send state. Callers must ensure that this doesn't race with
65 void rds_send_path_reset(struct rds_conn_path *cp)
67 struct rds_message *rm, *tmp;
72 cp->cp_xmit_rm = NULL;
73 /* Tell the user the RDMA op is no longer mapped by the
74 * transport. This isn't entirely true (it's flushed out
75 * independently) but as the connection is down, there's
76 * no ongoing RDMA to/from that memory */
77 rds_message_unmapped(rm);
82 cp->cp_xmit_hdr_off = 0;
83 cp->cp_xmit_data_off = 0;
84 cp->cp_xmit_atomic_sent = 0;
85 cp->cp_xmit_rdma_sent = 0;
86 cp->cp_xmit_data_sent = 0;
88 cp->cp_conn->c_map_queued = 0;
90 cp->cp_unacked_packets = rds_sysctl_max_unacked_packets;
91 cp->cp_unacked_bytes = rds_sysctl_max_unacked_bytes;
93 /* Mark messages as retransmissions, and move them to the send q */
94 spin_lock_irqsave(&cp->cp_lock, flags);
95 list_for_each_entry_safe(rm, tmp, &cp->cp_retrans, m_conn_item) {
96 set_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags);
97 set_bit(RDS_MSG_RETRANSMITTED, &rm->m_flags);
99 list_splice_init(&cp->cp_retrans, &cp->cp_send_queue);
100 spin_unlock_irqrestore(&cp->cp_lock, flags);
102 EXPORT_SYMBOL_GPL(rds_send_path_reset);
104 static int acquire_in_xmit(struct rds_conn_path *cp)
106 return test_and_set_bit_lock(RDS_IN_XMIT, &cp->cp_flags) == 0;
109 static void release_in_xmit(struct rds_conn_path *cp)
111 clear_bit_unlock(RDS_IN_XMIT, &cp->cp_flags);
113 * We don't use wait_on_bit()/wake_up_bit() because our waking is in a
114 * hot path and finding waiters is very rare. We don't want to walk
115 * the system-wide hashed waitqueue buckets in the fast path only to
116 * almost never find waiters.
118 if (waitqueue_active(&cp->cp_waitq))
119 wake_up_all(&cp->cp_waitq);
123 * We're making the conscious trade-off here to only send one message
124 * down the connection at a time.
126 * - tx queueing is a simple fifo list
127 * - reassembly is optional and easily done by transports per conn
128 * - no per flow rx lookup at all, straight to the socket
129 * - less per-frag memory and wire overhead
131 * - queued acks can be delayed behind large messages
133 * - small message latency is higher behind queued large messages
134 * - large message latency isn't starved by intervening small sends
136 int rds_send_xmit(struct rds_conn_path *cp)
138 struct rds_connection *conn = cp->cp_conn;
139 struct rds_message *rm;
142 struct scatterlist *sg;
144 LIST_HEAD(to_be_dropped);
146 unsigned long send_gen = 0;
153 * sendmsg calls here after having queued its message on the send
154 * queue. We only have one task feeding the connection at a time. If
155 * another thread is already feeding the queue then we back off. This
156 * avoids blocking the caller and trading per-connection data between
157 * caches per message.
159 if (!acquire_in_xmit(cp)) {
160 rds_stats_inc(s_send_lock_contention);
165 if (rds_destroy_pending(cp->cp_conn)) {
167 ret = -ENETUNREACH; /* dont requeue send work */
172 * we record the send generation after doing the xmit acquire.
173 * if someone else manages to jump in and do some work, we'll use
174 * this to avoid a goto restart farther down.
176 * The acquire_in_xmit() check above ensures that only one
177 * caller can increment c_send_gen at any time.
179 send_gen = READ_ONCE(cp->cp_send_gen) + 1;
180 WRITE_ONCE(cp->cp_send_gen, send_gen);
183 * rds_conn_shutdown() sets the conn state and then tests RDS_IN_XMIT,
184 * we do the opposite to avoid races.
186 if (!rds_conn_path_up(cp)) {
192 if (conn->c_trans->xmit_path_prepare)
193 conn->c_trans->xmit_path_prepare(cp);
196 * spin trying to push headers and data down the connection until
197 * the connection doesn't make forward progress.
207 if (same_rm >= 4096) {
208 rds_stats_inc(s_send_stuck_rm);
215 * If between sending messages, we can send a pending congestion
218 if (!rm && test_and_clear_bit(0, &conn->c_map_queued)) {
219 rm = rds_cong_update_alloc(conn);
224 rm->data.op_active = 1;
225 rm->m_inc.i_conn_path = cp;
226 rm->m_inc.i_conn = cp->cp_conn;
232 * If not already working on one, grab the next message.
234 * cp_xmit_rm holds a ref while we're sending this message down
235 * the connction. We can use this ref while holding the
236 * send_sem.. rds_send_reset() is serialized with it.
243 /* we want to process as big a batch as we can, but
244 * we also want to avoid softlockups. If we've been
245 * through a lot of messages, lets back off and see
246 * if anyone else jumps in
248 if (batch_count >= send_batch_count)
251 spin_lock_irqsave(&cp->cp_lock, flags);
253 if (!list_empty(&cp->cp_send_queue)) {
254 rm = list_entry(cp->cp_send_queue.next,
257 rds_message_addref(rm);
260 * Move the message from the send queue to the retransmit
263 list_move_tail(&rm->m_conn_item,
267 spin_unlock_irqrestore(&cp->cp_lock, flags);
272 /* Unfortunately, the way Infiniband deals with
273 * RDMA to a bad MR key is by moving the entire
274 * queue pair to error state. We cold possibly
275 * recover from that, but right now we drop the
277 * Therefore, we never retransmit messages with RDMA ops.
279 if (test_bit(RDS_MSG_FLUSH, &rm->m_flags) ||
280 (rm->rdma.op_active &&
281 test_bit(RDS_MSG_RETRANSMITTED, &rm->m_flags))) {
282 spin_lock_irqsave(&cp->cp_lock, flags);
283 if (test_and_clear_bit(RDS_MSG_ON_CONN, &rm->m_flags))
284 list_move(&rm->m_conn_item, &to_be_dropped);
285 spin_unlock_irqrestore(&cp->cp_lock, flags);
289 /* Require an ACK every once in a while */
290 len = ntohl(rm->m_inc.i_hdr.h_len);
291 if (cp->cp_unacked_packets == 0 ||
292 cp->cp_unacked_bytes < len) {
293 set_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags);
295 cp->cp_unacked_packets =
296 rds_sysctl_max_unacked_packets;
297 cp->cp_unacked_bytes =
298 rds_sysctl_max_unacked_bytes;
299 rds_stats_inc(s_send_ack_required);
301 cp->cp_unacked_bytes -= len;
302 cp->cp_unacked_packets--;
308 /* The transport either sends the whole rdma or none of it */
309 if (rm->rdma.op_active && !cp->cp_xmit_rdma_sent) {
310 rm->m_final_op = &rm->rdma;
311 /* The transport owns the mapped memory for now.
312 * You can't unmap it while it's on the send queue
314 set_bit(RDS_MSG_MAPPED, &rm->m_flags);
315 ret = conn->c_trans->xmit_rdma(conn, &rm->rdma);
317 clear_bit(RDS_MSG_MAPPED, &rm->m_flags);
318 wake_up_interruptible(&rm->m_flush_wait);
321 cp->cp_xmit_rdma_sent = 1;
325 if (rm->atomic.op_active && !cp->cp_xmit_atomic_sent) {
326 rm->m_final_op = &rm->atomic;
327 /* The transport owns the mapped memory for now.
328 * You can't unmap it while it's on the send queue
330 set_bit(RDS_MSG_MAPPED, &rm->m_flags);
331 ret = conn->c_trans->xmit_atomic(conn, &rm->atomic);
333 clear_bit(RDS_MSG_MAPPED, &rm->m_flags);
334 wake_up_interruptible(&rm->m_flush_wait);
337 cp->cp_xmit_atomic_sent = 1;
342 * A number of cases require an RDS header to be sent
343 * even if there is no data.
344 * We permit 0-byte sends; rds-ping depends on this.
345 * However, if there are exclusively attached silent ops,
346 * we skip the hdr/data send, to enable silent operation.
348 if (rm->data.op_nents == 0) {
350 int all_ops_are_silent = 1;
352 ops_present = (rm->atomic.op_active || rm->rdma.op_active);
353 if (rm->atomic.op_active && !rm->atomic.op_silent)
354 all_ops_are_silent = 0;
355 if (rm->rdma.op_active && !rm->rdma.op_silent)
356 all_ops_are_silent = 0;
358 if (ops_present && all_ops_are_silent
359 && !rm->m_rdma_cookie)
360 rm->data.op_active = 0;
363 if (rm->data.op_active && !cp->cp_xmit_data_sent) {
364 rm->m_final_op = &rm->data;
366 ret = conn->c_trans->xmit(conn, rm,
369 cp->cp_xmit_data_off);
373 if (cp->cp_xmit_hdr_off < sizeof(struct rds_header)) {
374 tmp = min_t(int, ret,
375 sizeof(struct rds_header) -
376 cp->cp_xmit_hdr_off);
377 cp->cp_xmit_hdr_off += tmp;
381 sg = &rm->data.op_sg[cp->cp_xmit_sg];
383 tmp = min_t(int, ret, sg->length -
384 cp->cp_xmit_data_off);
385 cp->cp_xmit_data_off += tmp;
387 if (cp->cp_xmit_data_off == sg->length) {
388 cp->cp_xmit_data_off = 0;
391 BUG_ON(ret != 0 && cp->cp_xmit_sg ==
396 if (cp->cp_xmit_hdr_off == sizeof(struct rds_header) &&
397 (cp->cp_xmit_sg == rm->data.op_nents))
398 cp->cp_xmit_data_sent = 1;
402 * A rm will only take multiple times through this loop
403 * if there is a data op. Thus, if the data is sent (or there was
404 * none), then we're done with the rm.
406 if (!rm->data.op_active || cp->cp_xmit_data_sent) {
407 cp->cp_xmit_rm = NULL;
409 cp->cp_xmit_hdr_off = 0;
410 cp->cp_xmit_data_off = 0;
411 cp->cp_xmit_rdma_sent = 0;
412 cp->cp_xmit_atomic_sent = 0;
413 cp->cp_xmit_data_sent = 0;
420 if (conn->c_trans->xmit_path_complete)
421 conn->c_trans->xmit_path_complete(cp);
424 /* Nuke any messages we decided not to retransmit. */
425 if (!list_empty(&to_be_dropped)) {
426 /* irqs on here, so we can put(), unlike above */
427 list_for_each_entry(rm, &to_be_dropped, m_conn_item)
429 rds_send_remove_from_sock(&to_be_dropped, RDS_RDMA_DROPPED);
433 * Other senders can queue a message after we last test the send queue
434 * but before we clear RDS_IN_XMIT. In that case they'd back off and
435 * not try and send their newly queued message. We need to check the
436 * send queue after having cleared RDS_IN_XMIT so that their message
437 * doesn't get stuck on the send queue.
439 * If the transport cannot continue (i.e ret != 0), then it must
440 * call us when more room is available, such as from the tx
441 * completion handler.
443 * We have an extra generation check here so that if someone manages
444 * to jump in after our release_in_xmit, we'll see that they have done
445 * some work and we will skip our goto
451 raced = send_gen != READ_ONCE(cp->cp_send_gen);
453 if ((test_bit(0, &conn->c_map_queued) ||
454 !list_empty(&cp->cp_send_queue)) && !raced) {
455 if (batch_count < send_batch_count)
458 if (rds_destroy_pending(cp->cp_conn))
461 queue_delayed_work(rds_wq, &cp->cp_send_w, 1);
464 rds_stats_inc(s_send_lock_queue_raced);
470 EXPORT_SYMBOL_GPL(rds_send_xmit);
472 static void rds_send_sndbuf_remove(struct rds_sock *rs, struct rds_message *rm)
474 u32 len = be32_to_cpu(rm->m_inc.i_hdr.h_len);
476 assert_spin_locked(&rs->rs_lock);
478 BUG_ON(rs->rs_snd_bytes < len);
479 rs->rs_snd_bytes -= len;
481 if (rs->rs_snd_bytes == 0)
482 rds_stats_inc(s_send_queue_empty);
485 static inline int rds_send_is_acked(struct rds_message *rm, u64 ack,
486 is_acked_func is_acked)
489 return is_acked(rm, ack);
490 return be64_to_cpu(rm->m_inc.i_hdr.h_sequence) <= ack;
494 * This is pretty similar to what happens below in the ACK
495 * handling code - except that we call here as soon as we get
496 * the IB send completion on the RDMA op and the accompanying
499 void rds_rdma_send_complete(struct rds_message *rm, int status)
501 struct rds_sock *rs = NULL;
502 struct rm_rdma_op *ro;
503 struct rds_notifier *notifier;
506 spin_lock_irqsave(&rm->m_rs_lock, flags);
509 if (test_bit(RDS_MSG_ON_SOCK, &rm->m_flags) &&
510 ro->op_active && ro->op_notify && ro->op_notifier) {
511 notifier = ro->op_notifier;
513 sock_hold(rds_rs_to_sk(rs));
515 notifier->n_status = status;
516 spin_lock(&rs->rs_lock);
517 list_add_tail(¬ifier->n_list, &rs->rs_notify_queue);
518 spin_unlock(&rs->rs_lock);
520 ro->op_notifier = NULL;
523 spin_unlock_irqrestore(&rm->m_rs_lock, flags);
526 rds_wake_sk_sleep(rs);
527 sock_put(rds_rs_to_sk(rs));
530 EXPORT_SYMBOL_GPL(rds_rdma_send_complete);
533 * Just like above, except looks at atomic op
535 void rds_atomic_send_complete(struct rds_message *rm, int status)
537 struct rds_sock *rs = NULL;
538 struct rm_atomic_op *ao;
539 struct rds_notifier *notifier;
542 spin_lock_irqsave(&rm->m_rs_lock, flags);
545 if (test_bit(RDS_MSG_ON_SOCK, &rm->m_flags)
546 && ao->op_active && ao->op_notify && ao->op_notifier) {
547 notifier = ao->op_notifier;
549 sock_hold(rds_rs_to_sk(rs));
551 notifier->n_status = status;
552 spin_lock(&rs->rs_lock);
553 list_add_tail(¬ifier->n_list, &rs->rs_notify_queue);
554 spin_unlock(&rs->rs_lock);
556 ao->op_notifier = NULL;
559 spin_unlock_irqrestore(&rm->m_rs_lock, flags);
562 rds_wake_sk_sleep(rs);
563 sock_put(rds_rs_to_sk(rs));
566 EXPORT_SYMBOL_GPL(rds_atomic_send_complete);
569 * This is the same as rds_rdma_send_complete except we
570 * don't do any locking - we have all the ingredients (message,
571 * socket, socket lock) and can just move the notifier.
574 __rds_send_complete(struct rds_sock *rs, struct rds_message *rm, int status)
576 struct rm_rdma_op *ro;
577 struct rm_atomic_op *ao;
580 if (ro->op_active && ro->op_notify && ro->op_notifier) {
581 ro->op_notifier->n_status = status;
582 list_add_tail(&ro->op_notifier->n_list, &rs->rs_notify_queue);
583 ro->op_notifier = NULL;
587 if (ao->op_active && ao->op_notify && ao->op_notifier) {
588 ao->op_notifier->n_status = status;
589 list_add_tail(&ao->op_notifier->n_list, &rs->rs_notify_queue);
590 ao->op_notifier = NULL;
593 /* No need to wake the app - caller does this */
597 * This removes messages from the socket's list if they're on it. The list
598 * argument must be private to the caller, we must be able to modify it
599 * without locks. The messages must have a reference held for their
600 * position on the list. This function will drop that reference after
601 * removing the messages from the 'messages' list regardless of if it found
602 * the messages on the socket list or not.
604 static void rds_send_remove_from_sock(struct list_head *messages, int status)
607 struct rds_sock *rs = NULL;
608 struct rds_message *rm;
610 while (!list_empty(messages)) {
613 rm = list_entry(messages->next, struct rds_message,
615 list_del_init(&rm->m_conn_item);
618 * If we see this flag cleared then we're *sure* that someone
619 * else beat us to removing it from the sock. If we race
620 * with their flag update we'll get the lock and then really
621 * see that the flag has been cleared.
623 * The message spinlock makes sure nobody clears rm->m_rs
624 * while we're messing with it. It does not prevent the
625 * message from being removed from the socket, though.
627 spin_lock_irqsave(&rm->m_rs_lock, flags);
628 if (!test_bit(RDS_MSG_ON_SOCK, &rm->m_flags))
629 goto unlock_and_drop;
631 if (rs != rm->m_rs) {
633 rds_wake_sk_sleep(rs);
634 sock_put(rds_rs_to_sk(rs));
638 sock_hold(rds_rs_to_sk(rs));
641 goto unlock_and_drop;
642 spin_lock(&rs->rs_lock);
644 if (test_and_clear_bit(RDS_MSG_ON_SOCK, &rm->m_flags)) {
645 struct rm_rdma_op *ro = &rm->rdma;
646 struct rds_notifier *notifier;
648 list_del_init(&rm->m_sock_item);
649 rds_send_sndbuf_remove(rs, rm);
651 if (ro->op_active && ro->op_notifier &&
652 (ro->op_notify || (ro->op_recverr && status))) {
653 notifier = ro->op_notifier;
654 list_add_tail(¬ifier->n_list,
655 &rs->rs_notify_queue);
656 if (!notifier->n_status)
657 notifier->n_status = status;
658 rm->rdma.op_notifier = NULL;
662 spin_unlock(&rs->rs_lock);
665 spin_unlock_irqrestore(&rm->m_rs_lock, flags);
672 rds_wake_sk_sleep(rs);
673 sock_put(rds_rs_to_sk(rs));
678 * Transports call here when they've determined that the receiver queued
679 * messages up to, and including, the given sequence number. Messages are
680 * moved to the retrans queue when rds_send_xmit picks them off the send
681 * queue. This means that in the TCP case, the message may not have been
682 * assigned the m_ack_seq yet - but that's fine as long as tcp_is_acked
683 * checks the RDS_MSG_HAS_ACK_SEQ bit.
685 void rds_send_path_drop_acked(struct rds_conn_path *cp, u64 ack,
686 is_acked_func is_acked)
688 struct rds_message *rm, *tmp;
692 spin_lock_irqsave(&cp->cp_lock, flags);
694 list_for_each_entry_safe(rm, tmp, &cp->cp_retrans, m_conn_item) {
695 if (!rds_send_is_acked(rm, ack, is_acked))
698 list_move(&rm->m_conn_item, &list);
699 clear_bit(RDS_MSG_ON_CONN, &rm->m_flags);
702 /* order flag updates with spin locks */
703 if (!list_empty(&list))
704 smp_mb__after_atomic();
706 spin_unlock_irqrestore(&cp->cp_lock, flags);
708 /* now remove the messages from the sock list as needed */
709 rds_send_remove_from_sock(&list, RDS_RDMA_SUCCESS);
711 EXPORT_SYMBOL_GPL(rds_send_path_drop_acked);
713 void rds_send_drop_acked(struct rds_connection *conn, u64 ack,
714 is_acked_func is_acked)
716 WARN_ON(conn->c_trans->t_mp_capable);
717 rds_send_path_drop_acked(&conn->c_path[0], ack, is_acked);
719 EXPORT_SYMBOL_GPL(rds_send_drop_acked);
721 void rds_send_drop_to(struct rds_sock *rs, struct sockaddr_in6 *dest)
723 struct rds_message *rm, *tmp;
724 struct rds_connection *conn;
725 struct rds_conn_path *cp;
729 /* get all the messages we're dropping under the rs lock */
730 spin_lock_irqsave(&rs->rs_lock, flags);
732 list_for_each_entry_safe(rm, tmp, &rs->rs_send_queue, m_sock_item) {
734 (!ipv6_addr_equal(&dest->sin6_addr, &rm->m_daddr) ||
735 dest->sin6_port != rm->m_inc.i_hdr.h_dport))
738 list_move(&rm->m_sock_item, &list);
739 rds_send_sndbuf_remove(rs, rm);
740 clear_bit(RDS_MSG_ON_SOCK, &rm->m_flags);
743 /* order flag updates with the rs lock */
744 smp_mb__after_atomic();
746 spin_unlock_irqrestore(&rs->rs_lock, flags);
748 if (list_empty(&list))
751 /* Remove the messages from the conn */
752 list_for_each_entry(rm, &list, m_sock_item) {
754 conn = rm->m_inc.i_conn;
755 if (conn->c_trans->t_mp_capable)
756 cp = rm->m_inc.i_conn_path;
758 cp = &conn->c_path[0];
760 spin_lock_irqsave(&cp->cp_lock, flags);
762 * Maybe someone else beat us to removing rm from the conn.
763 * If we race with their flag update we'll get the lock and
764 * then really see that the flag has been cleared.
766 if (!test_and_clear_bit(RDS_MSG_ON_CONN, &rm->m_flags)) {
767 spin_unlock_irqrestore(&cp->cp_lock, flags);
770 list_del_init(&rm->m_conn_item);
771 spin_unlock_irqrestore(&cp->cp_lock, flags);
774 * Couldn't grab m_rs_lock in top loop (lock ordering),
777 spin_lock_irqsave(&rm->m_rs_lock, flags);
779 spin_lock(&rs->rs_lock);
780 __rds_send_complete(rs, rm, RDS_RDMA_CANCELED);
781 spin_unlock(&rs->rs_lock);
783 spin_unlock_irqrestore(&rm->m_rs_lock, flags);
788 rds_wake_sk_sleep(rs);
790 while (!list_empty(&list)) {
791 rm = list_entry(list.next, struct rds_message, m_sock_item);
792 list_del_init(&rm->m_sock_item);
793 rds_message_wait(rm);
795 /* just in case the code above skipped this message
796 * because RDS_MSG_ON_CONN wasn't set, run it again here
797 * taking m_rs_lock is the only thing that keeps us
798 * from racing with ack processing.
800 spin_lock_irqsave(&rm->m_rs_lock, flags);
802 spin_lock(&rs->rs_lock);
803 __rds_send_complete(rs, rm, RDS_RDMA_CANCELED);
804 spin_unlock(&rs->rs_lock);
806 spin_unlock_irqrestore(&rm->m_rs_lock, flags);
813 * we only want this to fire once so we use the callers 'queued'. It's
814 * possible that another thread can race with us and remove the
815 * message from the flow with RDS_CANCEL_SENT_TO.
817 static int rds_send_queue_rm(struct rds_sock *rs, struct rds_connection *conn,
818 struct rds_conn_path *cp,
819 struct rds_message *rm, __be16 sport,
820 __be16 dport, int *queued)
828 len = be32_to_cpu(rm->m_inc.i_hdr.h_len);
830 /* this is the only place which holds both the socket's rs_lock
831 * and the connection's c_lock */
832 spin_lock_irqsave(&rs->rs_lock, flags);
835 * If there is a little space in sndbuf, we don't queue anything,
836 * and userspace gets -EAGAIN. But poll() indicates there's send
837 * room. This can lead to bad behavior (spinning) if snd_bytes isn't
838 * freed up by incoming acks. So we check the *old* value of
839 * rs_snd_bytes here to allow the last msg to exceed the buffer,
840 * and poll() now knows no more data can be sent.
842 if (rs->rs_snd_bytes < rds_sk_sndbuf(rs)) {
843 rs->rs_snd_bytes += len;
845 /* let recv side know we are close to send space exhaustion.
846 * This is probably not the optimal way to do it, as this
847 * means we set the flag on *all* messages as soon as our
848 * throughput hits a certain threshold.
850 if (rs->rs_snd_bytes >= rds_sk_sndbuf(rs) / 2)
851 set_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags);
853 list_add_tail(&rm->m_sock_item, &rs->rs_send_queue);
854 set_bit(RDS_MSG_ON_SOCK, &rm->m_flags);
855 rds_message_addref(rm);
856 sock_hold(rds_rs_to_sk(rs));
859 /* The code ordering is a little weird, but we're
860 trying to minimize the time we hold c_lock */
861 rds_message_populate_header(&rm->m_inc.i_hdr, sport, dport, 0);
862 rm->m_inc.i_conn = conn;
863 rm->m_inc.i_conn_path = cp;
864 rds_message_addref(rm);
866 spin_lock(&cp->cp_lock);
867 rm->m_inc.i_hdr.h_sequence = cpu_to_be64(cp->cp_next_tx_seq++);
868 list_add_tail(&rm->m_conn_item, &cp->cp_send_queue);
869 set_bit(RDS_MSG_ON_CONN, &rm->m_flags);
870 spin_unlock(&cp->cp_lock);
872 rdsdebug("queued msg %p len %d, rs %p bytes %d seq %llu\n",
873 rm, len, rs, rs->rs_snd_bytes,
874 (unsigned long long)be64_to_cpu(rm->m_inc.i_hdr.h_sequence));
879 spin_unlock_irqrestore(&rs->rs_lock, flags);
885 * rds_message is getting to be quite complicated, and we'd like to allocate
886 * it all in one go. This figures out how big it needs to be up front.
888 static int rds_rm_size(struct msghdr *msg, int num_sgs,
889 struct rds_iov_vector_arr *vct)
891 struct cmsghdr *cmsg;
895 bool zcopy_cookie = false;
896 struct rds_iov_vector *iov, *tmp_iov;
901 for_each_cmsghdr(cmsg, msg) {
902 if (!CMSG_OK(msg, cmsg))
905 if (cmsg->cmsg_level != SOL_RDS)
908 switch (cmsg->cmsg_type) {
909 case RDS_CMSG_RDMA_ARGS:
910 if (vct->indx >= vct->len) {
911 vct->len += vct->incr;
915 sizeof(struct rds_iov_vector),
918 vct->len -= vct->incr;
923 iov = &vct->vec[vct->indx];
924 memset(iov, 0, sizeof(struct rds_iov_vector));
927 retval = rds_rdma_extra_size(CMSG_DATA(cmsg), iov);
934 case RDS_CMSG_ZCOPY_COOKIE:
938 case RDS_CMSG_RDMA_DEST:
939 case RDS_CMSG_RDMA_MAP:
941 /* these are valid but do no add any size */
944 case RDS_CMSG_ATOMIC_CSWP:
945 case RDS_CMSG_ATOMIC_FADD:
946 case RDS_CMSG_MASKED_ATOMIC_CSWP:
947 case RDS_CMSG_MASKED_ATOMIC_FADD:
949 size += sizeof(struct scatterlist);
958 if ((msg->msg_flags & MSG_ZEROCOPY) && !zcopy_cookie)
961 size += num_sgs * sizeof(struct scatterlist);
963 /* Ensure (DEST, MAP) are never used with (ARGS, ATOMIC) */
964 if (cmsg_groups == 3)
970 static int rds_cmsg_zcopy(struct rds_sock *rs, struct rds_message *rm,
971 struct cmsghdr *cmsg)
975 if (cmsg->cmsg_len < CMSG_LEN(sizeof(*cookie)) ||
976 !rm->data.op_mmp_znotifier)
978 cookie = CMSG_DATA(cmsg);
979 rm->data.op_mmp_znotifier->z_cookie = *cookie;
983 static int rds_cmsg_send(struct rds_sock *rs, struct rds_message *rm,
984 struct msghdr *msg, int *allocated_mr,
985 struct rds_iov_vector_arr *vct)
987 struct cmsghdr *cmsg;
988 int ret = 0, ind = 0;
990 for_each_cmsghdr(cmsg, msg) {
991 if (!CMSG_OK(msg, cmsg))
994 if (cmsg->cmsg_level != SOL_RDS)
997 /* As a side effect, RDMA_DEST and RDMA_MAP will set
998 * rm->rdma.m_rdma_cookie and rm->rdma.m_rdma_mr.
1000 switch (cmsg->cmsg_type) {
1001 case RDS_CMSG_RDMA_ARGS:
1002 if (ind >= vct->indx)
1004 ret = rds_cmsg_rdma_args(rs, rm, cmsg, &vct->vec[ind]);
1008 case RDS_CMSG_RDMA_DEST:
1009 ret = rds_cmsg_rdma_dest(rs, rm, cmsg);
1012 case RDS_CMSG_RDMA_MAP:
1013 ret = rds_cmsg_rdma_map(rs, rm, cmsg);
1016 else if (ret == -ENODEV)
1017 /* Accommodate the get_mr() case which can fail
1018 * if connection isn't established yet.
1022 case RDS_CMSG_ATOMIC_CSWP:
1023 case RDS_CMSG_ATOMIC_FADD:
1024 case RDS_CMSG_MASKED_ATOMIC_CSWP:
1025 case RDS_CMSG_MASKED_ATOMIC_FADD:
1026 ret = rds_cmsg_atomic(rs, rm, cmsg);
1029 case RDS_CMSG_ZCOPY_COOKIE:
1030 ret = rds_cmsg_zcopy(rs, rm, cmsg);
1044 static int rds_send_mprds_hash(struct rds_sock *rs,
1045 struct rds_connection *conn, int nonblock)
1049 if (conn->c_npaths == 0)
1050 hash = RDS_MPATH_HASH(rs, RDS_MPATH_WORKERS);
1052 hash = RDS_MPATH_HASH(rs, conn->c_npaths);
1053 if (conn->c_npaths == 0 && hash != 0) {
1054 rds_send_ping(conn, 0);
1056 /* The underlying connection is not up yet. Need to wait
1057 * until it is up to be sure that the non-zero c_path can be
1058 * used. But if we are interrupted, we have to use the zero
1059 * c_path in case the connection ends up being non-MP capable.
1061 if (conn->c_npaths == 0) {
1062 /* Cannot wait for the connection be made, so just use
1067 if (wait_event_interruptible(conn->c_hs_waitq,
1068 conn->c_npaths != 0))
1071 if (conn->c_npaths == 1)
1077 static int rds_rdma_bytes(struct msghdr *msg, size_t *rdma_bytes)
1079 struct rds_rdma_args *args;
1080 struct cmsghdr *cmsg;
1082 for_each_cmsghdr(cmsg, msg) {
1083 if (!CMSG_OK(msg, cmsg))
1086 if (cmsg->cmsg_level != SOL_RDS)
1089 if (cmsg->cmsg_type == RDS_CMSG_RDMA_ARGS) {
1090 if (cmsg->cmsg_len <
1091 CMSG_LEN(sizeof(struct rds_rdma_args)))
1093 args = CMSG_DATA(cmsg);
1094 *rdma_bytes += args->remote_vec.bytes;
1100 int rds_sendmsg(struct socket *sock, struct msghdr *msg, size_t payload_len)
1102 struct sock *sk = sock->sk;
1103 struct rds_sock *rs = rds_sk_to_rs(sk);
1104 DECLARE_SOCKADDR(struct sockaddr_in6 *, sin6, msg->msg_name);
1105 DECLARE_SOCKADDR(struct sockaddr_in *, usin, msg->msg_name);
1107 struct rds_message *rm = NULL;
1108 struct rds_connection *conn;
1110 int queued = 0, allocated_mr = 0;
1111 int nonblock = msg->msg_flags & MSG_DONTWAIT;
1112 long timeo = sock_sndtimeo(sk, nonblock);
1113 struct rds_conn_path *cpath;
1114 struct in6_addr daddr;
1116 size_t total_payload_len = payload_len, rdma_payload_len = 0;
1117 bool zcopy = ((msg->msg_flags & MSG_ZEROCOPY) &&
1118 sock_flag(rds_rs_to_sk(rs), SOCK_ZEROCOPY));
1119 int num_sgs = DIV_ROUND_UP(payload_len, PAGE_SIZE);
1121 struct rds_iov_vector_arr vct;
1124 memset(&vct, 0, sizeof(vct));
1126 /* expect 1 RDMA CMSG per rds_sendmsg. can still grow if more needed. */
1129 /* Mirror Linux UDP mirror of BSD error message compatibility */
1130 /* XXX: Perhaps MSG_MORE someday */
1131 if (msg->msg_flags & ~(MSG_DONTWAIT | MSG_CMSG_COMPAT | MSG_ZEROCOPY)) {
1136 namelen = msg->msg_namelen;
1138 if (namelen < sizeof(*usin)) {
1142 switch (usin->sin_family) {
1144 if (usin->sin_addr.s_addr == htonl(INADDR_ANY) ||
1145 usin->sin_addr.s_addr == htonl(INADDR_BROADCAST) ||
1146 ipv4_is_multicast(usin->sin_addr.s_addr)) {
1150 ipv6_addr_set_v4mapped(usin->sin_addr.s_addr, &daddr);
1151 dport = usin->sin_port;
1154 #if IS_ENABLED(CONFIG_IPV6)
1158 if (namelen < sizeof(*sin6)) {
1162 addr_type = ipv6_addr_type(&sin6->sin6_addr);
1163 if (!(addr_type & IPV6_ADDR_UNICAST)) {
1166 if (!(addr_type & IPV6_ADDR_MAPPED)) {
1171 /* It is a mapped address. Need to do some
1174 addr4 = sin6->sin6_addr.s6_addr32[3];
1175 if (addr4 == htonl(INADDR_ANY) ||
1176 addr4 == htonl(INADDR_BROADCAST) ||
1177 ipv4_is_multicast(addr4)) {
1182 if (addr_type & IPV6_ADDR_LINKLOCAL) {
1183 if (sin6->sin6_scope_id == 0) {
1187 scope_id = sin6->sin6_scope_id;
1190 daddr = sin6->sin6_addr;
1191 dport = sin6->sin6_port;
1201 /* We only care about consistency with ->connect() */
1203 daddr = rs->rs_conn_addr;
1204 dport = rs->rs_conn_port;
1205 scope_id = rs->rs_bound_scope_id;
1210 if (ipv6_addr_any(&rs->rs_bound_addr) || ipv6_addr_any(&daddr)) {
1214 } else if (namelen != 0) {
1215 /* Cannot send to an IPv4 address using an IPv6 source
1216 * address and cannot send to an IPv6 address using an
1217 * IPv4 source address.
1219 if (ipv6_addr_v4mapped(&daddr) ^
1220 ipv6_addr_v4mapped(&rs->rs_bound_addr)) {
1225 /* If the socket is already bound to a link local address,
1226 * it can only send to peers on the same link. But allow
1227 * communicating beween link local and non-link local address.
1229 if (scope_id != rs->rs_bound_scope_id) {
1231 scope_id = rs->rs_bound_scope_id;
1232 } else if (rs->rs_bound_scope_id) {
1241 ret = rds_rdma_bytes(msg, &rdma_payload_len);
1245 total_payload_len += rdma_payload_len;
1246 if (max_t(size_t, payload_len, rdma_payload_len) > RDS_MAX_MSG_SIZE) {
1251 if (payload_len > rds_sk_sndbuf(rs)) {
1257 if (rs->rs_transport->t_type != RDS_TRANS_TCP) {
1261 num_sgs = iov_iter_npages(&msg->msg_iter, INT_MAX);
1263 /* size of rm including all sgs */
1264 ret = rds_rm_size(msg, num_sgs, &vct);
1268 rm = rds_message_alloc(ret, GFP_KERNEL);
1274 /* Attach data to the rm */
1276 rm->data.op_sg = rds_message_alloc_sgs(rm, num_sgs);
1277 if (IS_ERR(rm->data.op_sg)) {
1278 ret = PTR_ERR(rm->data.op_sg);
1281 ret = rds_message_copy_from_user(rm, &msg->msg_iter, zcopy);
1285 rm->data.op_active = 1;
1287 rm->m_daddr = daddr;
1289 /* rds_conn_create has a spinlock that runs with IRQ off.
1290 * Caching the conn in the socket helps a lot. */
1291 if (rs->rs_conn && ipv6_addr_equal(&rs->rs_conn->c_faddr, &daddr) &&
1292 rs->rs_tos == rs->rs_conn->c_tos) {
1295 conn = rds_conn_create_outgoing(sock_net(sock->sk),
1296 &rs->rs_bound_addr, &daddr,
1297 rs->rs_transport, rs->rs_tos,
1298 sock->sk->sk_allocation,
1301 ret = PTR_ERR(conn);
1307 if (conn->c_trans->t_mp_capable)
1308 cpath = &conn->c_path[rds_send_mprds_hash(rs, conn, nonblock)];
1310 cpath = &conn->c_path[0];
1312 rm->m_conn_path = cpath;
1314 /* Parse any control messages the user may have included. */
1315 ret = rds_cmsg_send(rs, rm, msg, &allocated_mr, &vct);
1319 if (rm->rdma.op_active && !conn->c_trans->xmit_rdma) {
1320 printk_ratelimited(KERN_NOTICE "rdma_op %p conn xmit_rdma %p\n",
1321 &rm->rdma, conn->c_trans->xmit_rdma);
1326 if (rm->atomic.op_active && !conn->c_trans->xmit_atomic) {
1327 printk_ratelimited(KERN_NOTICE "atomic_op %p conn xmit_atomic %p\n",
1328 &rm->atomic, conn->c_trans->xmit_atomic);
1333 if (rds_destroy_pending(conn)) {
1338 if (rds_conn_path_down(cpath))
1339 rds_check_all_paths(conn);
1341 ret = rds_cong_wait(conn->c_fcong, dport, nonblock, rs);
1343 rs->rs_seen_congestion = 1;
1346 while (!rds_send_queue_rm(rs, conn, cpath, rm, rs->rs_bound_port,
1348 rds_stats_inc(s_send_queue_full);
1355 timeo = wait_event_interruptible_timeout(*sk_sleep(sk),
1356 rds_send_queue_rm(rs, conn, cpath, rm,
1361 rdsdebug("sendmsg woke queued %d timeo %ld\n", queued, timeo);
1362 if (timeo > 0 || timeo == MAX_SCHEDULE_TIMEOUT)
1372 * By now we've committed to the send. We reuse rds_send_worker()
1373 * to retry sends in the rds thread if the transport asks us to.
1375 rds_stats_inc(s_send_queued);
1377 ret = rds_send_xmit(cpath);
1378 if (ret == -ENOMEM || ret == -EAGAIN) {
1381 if (rds_destroy_pending(cpath->cp_conn))
1384 queue_delayed_work(rds_wq, &cpath->cp_send_w, 1);
1389 rds_message_put(rm);
1391 for (ind = 0; ind < vct.indx; ind++)
1392 kfree(vct.vec[ind].iov);
1398 for (ind = 0; ind < vct.indx; ind++)
1399 kfree(vct.vec[ind].iov);
1402 /* If the user included a RDMA_MAP cmsg, we allocated a MR on the fly.
1403 * If the sendmsg goes through, we keep the MR. If it fails with EAGAIN
1404 * or in any other way, we need to destroy the MR again */
1406 rds_rdma_unuse(rs, rds_rdma_cookie_key(rm->m_rdma_cookie), 1);
1409 rds_message_put(rm);
1414 * send out a probe. Can be shared by rds_send_ping,
1415 * rds_send_pong, rds_send_hb.
1416 * rds_send_hb should use h_flags
1417 * RDS_FLAG_HB_PING|RDS_FLAG_ACK_REQUIRED
1419 * RDS_FLAG_HB_PONG|RDS_FLAG_ACK_REQUIRED
1422 rds_send_probe(struct rds_conn_path *cp, __be16 sport,
1423 __be16 dport, u8 h_flags)
1425 struct rds_message *rm;
1426 unsigned long flags;
1429 rm = rds_message_alloc(0, GFP_ATOMIC);
1435 rm->m_daddr = cp->cp_conn->c_faddr;
1436 rm->data.op_active = 1;
1438 rds_conn_path_connect_if_down(cp);
1440 ret = rds_cong_wait(cp->cp_conn->c_fcong, dport, 1, NULL);
1444 spin_lock_irqsave(&cp->cp_lock, flags);
1445 list_add_tail(&rm->m_conn_item, &cp->cp_send_queue);
1446 set_bit(RDS_MSG_ON_CONN, &rm->m_flags);
1447 rds_message_addref(rm);
1448 rm->m_inc.i_conn = cp->cp_conn;
1449 rm->m_inc.i_conn_path = cp;
1451 rds_message_populate_header(&rm->m_inc.i_hdr, sport, dport,
1452 cp->cp_next_tx_seq);
1453 rm->m_inc.i_hdr.h_flags |= h_flags;
1454 cp->cp_next_tx_seq++;
1456 if (RDS_HS_PROBE(be16_to_cpu(sport), be16_to_cpu(dport)) &&
1457 cp->cp_conn->c_trans->t_mp_capable) {
1458 u16 npaths = cpu_to_be16(RDS_MPATH_WORKERS);
1459 u32 my_gen_num = cpu_to_be32(cp->cp_conn->c_my_gen_num);
1461 rds_message_add_extension(&rm->m_inc.i_hdr,
1462 RDS_EXTHDR_NPATHS, &npaths,
1464 rds_message_add_extension(&rm->m_inc.i_hdr,
1469 spin_unlock_irqrestore(&cp->cp_lock, flags);
1471 rds_stats_inc(s_send_queued);
1472 rds_stats_inc(s_send_pong);
1474 /* schedule the send work on rds_wq */
1476 if (!rds_destroy_pending(cp->cp_conn))
1477 queue_delayed_work(rds_wq, &cp->cp_send_w, 1);
1480 rds_message_put(rm);
1485 rds_message_put(rm);
1490 rds_send_pong(struct rds_conn_path *cp, __be16 dport)
1492 return rds_send_probe(cp, 0, dport, 0);
1496 rds_send_ping(struct rds_connection *conn, int cp_index)
1498 unsigned long flags;
1499 struct rds_conn_path *cp = &conn->c_path[cp_index];
1501 spin_lock_irqsave(&cp->cp_lock, flags);
1502 if (conn->c_ping_triggered) {
1503 spin_unlock_irqrestore(&cp->cp_lock, flags);
1506 conn->c_ping_triggered = 1;
1507 spin_unlock_irqrestore(&cp->cp_lock, flags);
1508 rds_send_probe(cp, cpu_to_be16(RDS_FLAG_PROBE_PORT), 0, 0);
1510 EXPORT_SYMBOL_GPL(rds_send_ping);