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(RDS_IN_XMIT, &cp->cp_flags) == 0;
109 static void release_in_xmit(struct rds_conn_path *cp)
111 clear_bit(RDS_IN_XMIT, &cp->cp_flags);
112 smp_mb__after_atomic();
114 * We don't use wait_on_bit()/wake_up_bit() because our waking is in a
115 * hot path and finding waiters is very rare. We don't want to walk
116 * the system-wide hashed waitqueue buckets in the fast path only to
117 * almost never find waiters.
119 if (waitqueue_active(&cp->cp_waitq))
120 wake_up_all(&cp->cp_waitq);
124 * We're making the conscious trade-off here to only send one message
125 * down the connection at a time.
127 * - tx queueing is a simple fifo list
128 * - reassembly is optional and easily done by transports per conn
129 * - no per flow rx lookup at all, straight to the socket
130 * - less per-frag memory and wire overhead
132 * - queued acks can be delayed behind large messages
134 * - small message latency is higher behind queued large messages
135 * - large message latency isn't starved by intervening small sends
137 int rds_send_xmit(struct rds_conn_path *cp)
139 struct rds_connection *conn = cp->cp_conn;
140 struct rds_message *rm;
143 struct scatterlist *sg;
145 LIST_HEAD(to_be_dropped);
147 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);
166 * we record the send generation after doing the xmit acquire.
167 * if someone else manages to jump in and do some work, we'll use
168 * this to avoid a goto restart farther down.
170 * The acquire_in_xmit() check above ensures that only one
171 * caller can increment c_send_gen at any time.
174 send_gen = cp->cp_send_gen;
177 * rds_conn_shutdown() sets the conn state and then tests RDS_IN_XMIT,
178 * we do the opposite to avoid races.
180 if (!rds_conn_path_up(cp)) {
186 if (conn->c_trans->xmit_path_prepare)
187 conn->c_trans->xmit_path_prepare(cp);
190 * spin trying to push headers and data down the connection until
191 * the connection doesn't make forward progress.
198 * If between sending messages, we can send a pending congestion
201 if (!rm && test_and_clear_bit(0, &conn->c_map_queued)) {
202 rm = rds_cong_update_alloc(conn);
207 rm->data.op_active = 1;
208 rm->m_inc.i_conn_path = cp;
209 rm->m_inc.i_conn = cp->cp_conn;
215 * If not already working on one, grab the next message.
217 * cp_xmit_rm holds a ref while we're sending this message down
218 * the connction. We can use this ref while holding the
219 * send_sem.. rds_send_reset() is serialized with it.
226 /* we want to process as big a batch as we can, but
227 * we also want to avoid softlockups. If we've been
228 * through a lot of messages, lets back off and see
229 * if anyone else jumps in
231 if (batch_count >= send_batch_count)
234 spin_lock_irqsave(&cp->cp_lock, flags);
236 if (!list_empty(&cp->cp_send_queue)) {
237 rm = list_entry(cp->cp_send_queue.next,
240 rds_message_addref(rm);
243 * Move the message from the send queue to the retransmit
246 list_move_tail(&rm->m_conn_item,
250 spin_unlock_irqrestore(&cp->cp_lock, flags);
255 /* Unfortunately, the way Infiniband deals with
256 * RDMA to a bad MR key is by moving the entire
257 * queue pair to error state. We cold possibly
258 * recover from that, but right now we drop the
260 * Therefore, we never retransmit messages with RDMA ops.
262 if (rm->rdma.op_active &&
263 test_bit(RDS_MSG_RETRANSMITTED, &rm->m_flags)) {
264 spin_lock_irqsave(&cp->cp_lock, flags);
265 if (test_and_clear_bit(RDS_MSG_ON_CONN, &rm->m_flags))
266 list_move(&rm->m_conn_item, &to_be_dropped);
267 spin_unlock_irqrestore(&cp->cp_lock, flags);
271 /* Require an ACK every once in a while */
272 len = ntohl(rm->m_inc.i_hdr.h_len);
273 if (cp->cp_unacked_packets == 0 ||
274 cp->cp_unacked_bytes < len) {
275 __set_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags);
277 cp->cp_unacked_packets =
278 rds_sysctl_max_unacked_packets;
279 cp->cp_unacked_bytes =
280 rds_sysctl_max_unacked_bytes;
281 rds_stats_inc(s_send_ack_required);
283 cp->cp_unacked_bytes -= len;
284 cp->cp_unacked_packets--;
290 /* The transport either sends the whole rdma or none of it */
291 if (rm->rdma.op_active && !cp->cp_xmit_rdma_sent) {
292 rm->m_final_op = &rm->rdma;
293 /* The transport owns the mapped memory for now.
294 * You can't unmap it while it's on the send queue
296 set_bit(RDS_MSG_MAPPED, &rm->m_flags);
297 ret = conn->c_trans->xmit_rdma(conn, &rm->rdma);
299 clear_bit(RDS_MSG_MAPPED, &rm->m_flags);
300 wake_up_interruptible(&rm->m_flush_wait);
303 cp->cp_xmit_rdma_sent = 1;
307 if (rm->atomic.op_active && !cp->cp_xmit_atomic_sent) {
308 rm->m_final_op = &rm->atomic;
309 /* The transport owns the mapped memory for now.
310 * You can't unmap it while it's on the send queue
312 set_bit(RDS_MSG_MAPPED, &rm->m_flags);
313 ret = conn->c_trans->xmit_atomic(conn, &rm->atomic);
315 clear_bit(RDS_MSG_MAPPED, &rm->m_flags);
316 wake_up_interruptible(&rm->m_flush_wait);
319 cp->cp_xmit_atomic_sent = 1;
324 * A number of cases require an RDS header to be sent
325 * even if there is no data.
326 * We permit 0-byte sends; rds-ping depends on this.
327 * However, if there are exclusively attached silent ops,
328 * we skip the hdr/data send, to enable silent operation.
330 if (rm->data.op_nents == 0) {
332 int all_ops_are_silent = 1;
334 ops_present = (rm->atomic.op_active || rm->rdma.op_active);
335 if (rm->atomic.op_active && !rm->atomic.op_silent)
336 all_ops_are_silent = 0;
337 if (rm->rdma.op_active && !rm->rdma.op_silent)
338 all_ops_are_silent = 0;
340 if (ops_present && all_ops_are_silent
341 && !rm->m_rdma_cookie)
342 rm->data.op_active = 0;
345 if (rm->data.op_active && !cp->cp_xmit_data_sent) {
346 rm->m_final_op = &rm->data;
348 ret = conn->c_trans->xmit(conn, rm,
351 cp->cp_xmit_data_off);
355 if (cp->cp_xmit_hdr_off < sizeof(struct rds_header)) {
356 tmp = min_t(int, ret,
357 sizeof(struct rds_header) -
358 cp->cp_xmit_hdr_off);
359 cp->cp_xmit_hdr_off += tmp;
363 sg = &rm->data.op_sg[cp->cp_xmit_sg];
365 tmp = min_t(int, ret, sg->length -
366 cp->cp_xmit_data_off);
367 cp->cp_xmit_data_off += tmp;
369 if (cp->cp_xmit_data_off == sg->length) {
370 cp->cp_xmit_data_off = 0;
373 BUG_ON(ret != 0 && cp->cp_xmit_sg ==
378 if (cp->cp_xmit_hdr_off == sizeof(struct rds_header) &&
379 (cp->cp_xmit_sg == rm->data.op_nents))
380 cp->cp_xmit_data_sent = 1;
384 * A rm will only take multiple times through this loop
385 * if there is a data op. Thus, if the data is sent (or there was
386 * none), then we're done with the rm.
388 if (!rm->data.op_active || cp->cp_xmit_data_sent) {
389 cp->cp_xmit_rm = NULL;
391 cp->cp_xmit_hdr_off = 0;
392 cp->cp_xmit_data_off = 0;
393 cp->cp_xmit_rdma_sent = 0;
394 cp->cp_xmit_atomic_sent = 0;
395 cp->cp_xmit_data_sent = 0;
402 if (conn->c_trans->xmit_path_complete)
403 conn->c_trans->xmit_path_complete(cp);
406 /* Nuke any messages we decided not to retransmit. */
407 if (!list_empty(&to_be_dropped)) {
408 /* irqs on here, so we can put(), unlike above */
409 list_for_each_entry(rm, &to_be_dropped, m_conn_item)
411 rds_send_remove_from_sock(&to_be_dropped, RDS_RDMA_DROPPED);
415 * Other senders can queue a message after we last test the send queue
416 * but before we clear RDS_IN_XMIT. In that case they'd back off and
417 * not try and send their newly queued message. We need to check the
418 * send queue after having cleared RDS_IN_XMIT so that their message
419 * doesn't get stuck on the send queue.
421 * If the transport cannot continue (i.e ret != 0), then it must
422 * call us when more room is available, such as from the tx
423 * completion handler.
425 * We have an extra generation check here so that if someone manages
426 * to jump in after our release_in_xmit, we'll see that they have done
427 * some work and we will skip our goto
431 if ((test_bit(0, &conn->c_map_queued) ||
432 !list_empty(&cp->cp_send_queue)) &&
433 send_gen == cp->cp_send_gen) {
434 rds_stats_inc(s_send_lock_queue_raced);
435 if (batch_count < send_batch_count)
437 queue_delayed_work(rds_wq, &cp->cp_send_w, 1);
443 EXPORT_SYMBOL_GPL(rds_send_xmit);
445 static void rds_send_sndbuf_remove(struct rds_sock *rs, struct rds_message *rm)
447 u32 len = be32_to_cpu(rm->m_inc.i_hdr.h_len);
449 assert_spin_locked(&rs->rs_lock);
451 BUG_ON(rs->rs_snd_bytes < len);
452 rs->rs_snd_bytes -= len;
454 if (rs->rs_snd_bytes == 0)
455 rds_stats_inc(s_send_queue_empty);
458 static inline int rds_send_is_acked(struct rds_message *rm, u64 ack,
459 is_acked_func is_acked)
462 return is_acked(rm, ack);
463 return be64_to_cpu(rm->m_inc.i_hdr.h_sequence) <= ack;
467 * This is pretty similar to what happens below in the ACK
468 * handling code - except that we call here as soon as we get
469 * the IB send completion on the RDMA op and the accompanying
472 void rds_rdma_send_complete(struct rds_message *rm, int status)
474 struct rds_sock *rs = NULL;
475 struct rm_rdma_op *ro;
476 struct rds_notifier *notifier;
478 unsigned int notify = 0;
480 spin_lock_irqsave(&rm->m_rs_lock, flags);
482 notify = rm->rdma.op_notify | rm->data.op_notify;
484 if (test_bit(RDS_MSG_ON_SOCK, &rm->m_flags) &&
485 ro->op_active && notify && ro->op_notifier) {
486 notifier = ro->op_notifier;
488 sock_hold(rds_rs_to_sk(rs));
490 notifier->n_status = status;
491 spin_lock(&rs->rs_lock);
492 list_add_tail(¬ifier->n_list, &rs->rs_notify_queue);
493 spin_unlock(&rs->rs_lock);
495 ro->op_notifier = NULL;
498 spin_unlock_irqrestore(&rm->m_rs_lock, flags);
501 rds_wake_sk_sleep(rs);
502 sock_put(rds_rs_to_sk(rs));
505 EXPORT_SYMBOL_GPL(rds_rdma_send_complete);
508 * Just like above, except looks at atomic op
510 void rds_atomic_send_complete(struct rds_message *rm, int status)
512 struct rds_sock *rs = NULL;
513 struct rm_atomic_op *ao;
514 struct rds_notifier *notifier;
517 spin_lock_irqsave(&rm->m_rs_lock, flags);
520 if (test_bit(RDS_MSG_ON_SOCK, &rm->m_flags)
521 && ao->op_active && ao->op_notify && ao->op_notifier) {
522 notifier = ao->op_notifier;
524 sock_hold(rds_rs_to_sk(rs));
526 notifier->n_status = status;
527 spin_lock(&rs->rs_lock);
528 list_add_tail(¬ifier->n_list, &rs->rs_notify_queue);
529 spin_unlock(&rs->rs_lock);
531 ao->op_notifier = NULL;
534 spin_unlock_irqrestore(&rm->m_rs_lock, flags);
537 rds_wake_sk_sleep(rs);
538 sock_put(rds_rs_to_sk(rs));
541 EXPORT_SYMBOL_GPL(rds_atomic_send_complete);
544 * This is the same as rds_rdma_send_complete except we
545 * don't do any locking - we have all the ingredients (message,
546 * socket, socket lock) and can just move the notifier.
549 __rds_send_complete(struct rds_sock *rs, struct rds_message *rm, int status)
551 struct rm_rdma_op *ro;
552 struct rm_atomic_op *ao;
555 if (ro->op_active && ro->op_notify && ro->op_notifier) {
556 ro->op_notifier->n_status = status;
557 list_add_tail(&ro->op_notifier->n_list, &rs->rs_notify_queue);
558 ro->op_notifier = NULL;
562 if (ao->op_active && ao->op_notify && ao->op_notifier) {
563 ao->op_notifier->n_status = status;
564 list_add_tail(&ao->op_notifier->n_list, &rs->rs_notify_queue);
565 ao->op_notifier = NULL;
568 /* No need to wake the app - caller does this */
572 * This removes messages from the socket's list if they're on it. The list
573 * argument must be private to the caller, we must be able to modify it
574 * without locks. The messages must have a reference held for their
575 * position on the list. This function will drop that reference after
576 * removing the messages from the 'messages' list regardless of if it found
577 * the messages on the socket list or not.
579 static void rds_send_remove_from_sock(struct list_head *messages, int status)
582 struct rds_sock *rs = NULL;
583 struct rds_message *rm;
585 while (!list_empty(messages)) {
588 rm = list_entry(messages->next, struct rds_message,
590 list_del_init(&rm->m_conn_item);
593 * If we see this flag cleared then we're *sure* that someone
594 * else beat us to removing it from the sock. If we race
595 * with their flag update we'll get the lock and then really
596 * see that the flag has been cleared.
598 * The message spinlock makes sure nobody clears rm->m_rs
599 * while we're messing with it. It does not prevent the
600 * message from being removed from the socket, though.
602 spin_lock_irqsave(&rm->m_rs_lock, flags);
603 if (!test_bit(RDS_MSG_ON_SOCK, &rm->m_flags))
604 goto unlock_and_drop;
606 if (rs != rm->m_rs) {
608 rds_wake_sk_sleep(rs);
609 sock_put(rds_rs_to_sk(rs));
613 sock_hold(rds_rs_to_sk(rs));
616 goto unlock_and_drop;
617 spin_lock(&rs->rs_lock);
619 if (test_and_clear_bit(RDS_MSG_ON_SOCK, &rm->m_flags)) {
620 struct rm_rdma_op *ro = &rm->rdma;
621 struct rds_notifier *notifier;
623 list_del_init(&rm->m_sock_item);
624 rds_send_sndbuf_remove(rs, rm);
626 if (ro->op_active && ro->op_notifier &&
627 (ro->op_notify || (ro->op_recverr && status))) {
628 notifier = ro->op_notifier;
629 list_add_tail(¬ifier->n_list,
630 &rs->rs_notify_queue);
631 if (!notifier->n_status)
632 notifier->n_status = status;
633 rm->rdma.op_notifier = NULL;
638 spin_unlock(&rs->rs_lock);
641 spin_unlock_irqrestore(&rm->m_rs_lock, flags);
648 rds_wake_sk_sleep(rs);
649 sock_put(rds_rs_to_sk(rs));
654 * Transports call here when they've determined that the receiver queued
655 * messages up to, and including, the given sequence number. Messages are
656 * moved to the retrans queue when rds_send_xmit picks them off the send
657 * queue. This means that in the TCP case, the message may not have been
658 * assigned the m_ack_seq yet - but that's fine as long as tcp_is_acked
659 * checks the RDS_MSG_HAS_ACK_SEQ bit.
661 void rds_send_path_drop_acked(struct rds_conn_path *cp, u64 ack,
662 is_acked_func is_acked)
664 struct rds_message *rm, *tmp;
668 spin_lock_irqsave(&cp->cp_lock, flags);
670 list_for_each_entry_safe(rm, tmp, &cp->cp_retrans, m_conn_item) {
671 if (!rds_send_is_acked(rm, ack, is_acked))
674 list_move(&rm->m_conn_item, &list);
675 clear_bit(RDS_MSG_ON_CONN, &rm->m_flags);
678 /* order flag updates with spin locks */
679 if (!list_empty(&list))
680 smp_mb__after_atomic();
682 spin_unlock_irqrestore(&cp->cp_lock, flags);
684 /* now remove the messages from the sock list as needed */
685 rds_send_remove_from_sock(&list, RDS_RDMA_SUCCESS);
687 EXPORT_SYMBOL_GPL(rds_send_path_drop_acked);
689 void rds_send_drop_acked(struct rds_connection *conn, u64 ack,
690 is_acked_func is_acked)
692 WARN_ON(conn->c_trans->t_mp_capable);
693 rds_send_path_drop_acked(&conn->c_path[0], ack, is_acked);
695 EXPORT_SYMBOL_GPL(rds_send_drop_acked);
697 void rds_send_drop_to(struct rds_sock *rs, struct sockaddr_in *dest)
699 struct rds_message *rm, *tmp;
700 struct rds_connection *conn;
701 struct rds_conn_path *cp;
705 /* get all the messages we're dropping under the rs lock */
706 spin_lock_irqsave(&rs->rs_lock, flags);
708 list_for_each_entry_safe(rm, tmp, &rs->rs_send_queue, m_sock_item) {
709 if (dest && (dest->sin_addr.s_addr != rm->m_daddr ||
710 dest->sin_port != rm->m_inc.i_hdr.h_dport))
713 list_move(&rm->m_sock_item, &list);
714 rds_send_sndbuf_remove(rs, rm);
715 clear_bit(RDS_MSG_ON_SOCK, &rm->m_flags);
718 /* order flag updates with the rs lock */
719 smp_mb__after_atomic();
721 spin_unlock_irqrestore(&rs->rs_lock, flags);
723 if (list_empty(&list))
726 /* Remove the messages from the conn */
727 list_for_each_entry(rm, &list, m_sock_item) {
729 conn = rm->m_inc.i_conn;
730 if (conn->c_trans->t_mp_capable)
731 cp = rm->m_inc.i_conn_path;
733 cp = &conn->c_path[0];
735 spin_lock_irqsave(&cp->cp_lock, flags);
737 * Maybe someone else beat us to removing rm from the conn.
738 * If we race with their flag update we'll get the lock and
739 * then really see that the flag has been cleared.
741 if (!test_and_clear_bit(RDS_MSG_ON_CONN, &rm->m_flags)) {
742 spin_unlock_irqrestore(&cp->cp_lock, flags);
743 spin_lock_irqsave(&rm->m_rs_lock, flags);
745 spin_unlock_irqrestore(&rm->m_rs_lock, flags);
748 list_del_init(&rm->m_conn_item);
749 spin_unlock_irqrestore(&cp->cp_lock, flags);
752 * Couldn't grab m_rs_lock in top loop (lock ordering),
755 spin_lock_irqsave(&rm->m_rs_lock, flags);
757 spin_lock(&rs->rs_lock);
758 __rds_send_complete(rs, rm, RDS_RDMA_CANCELED);
759 spin_unlock(&rs->rs_lock);
762 spin_unlock_irqrestore(&rm->m_rs_lock, flags);
767 rds_wake_sk_sleep(rs);
769 while (!list_empty(&list)) {
770 rm = list_entry(list.next, struct rds_message, m_sock_item);
771 list_del_init(&rm->m_sock_item);
772 rds_message_wait(rm);
774 /* just in case the code above skipped this message
775 * because RDS_MSG_ON_CONN wasn't set, run it again here
776 * taking m_rs_lock is the only thing that keeps us
777 * from racing with ack processing.
779 spin_lock_irqsave(&rm->m_rs_lock, flags);
781 spin_lock(&rs->rs_lock);
782 __rds_send_complete(rs, rm, RDS_RDMA_CANCELED);
783 spin_unlock(&rs->rs_lock);
786 spin_unlock_irqrestore(&rm->m_rs_lock, flags);
793 * we only want this to fire once so we use the callers 'queued'. It's
794 * possible that another thread can race with us and remove the
795 * message from the flow with RDS_CANCEL_SENT_TO.
797 static int rds_send_queue_rm(struct rds_sock *rs, struct rds_connection *conn,
798 struct rds_conn_path *cp,
799 struct rds_message *rm, __be16 sport,
800 __be16 dport, int *queued)
808 len = be32_to_cpu(rm->m_inc.i_hdr.h_len);
810 /* this is the only place which holds both the socket's rs_lock
811 * and the connection's c_lock */
812 spin_lock_irqsave(&rs->rs_lock, flags);
815 * If there is a little space in sndbuf, we don't queue anything,
816 * and userspace gets -EAGAIN. But poll() indicates there's send
817 * room. This can lead to bad behavior (spinning) if snd_bytes isn't
818 * freed up by incoming acks. So we check the *old* value of
819 * rs_snd_bytes here to allow the last msg to exceed the buffer,
820 * and poll() now knows no more data can be sent.
822 if (rs->rs_snd_bytes < rds_sk_sndbuf(rs)) {
823 rs->rs_snd_bytes += len;
825 /* let recv side know we are close to send space exhaustion.
826 * This is probably not the optimal way to do it, as this
827 * means we set the flag on *all* messages as soon as our
828 * throughput hits a certain threshold.
830 if (rs->rs_snd_bytes >= rds_sk_sndbuf(rs) / 2)
831 __set_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags);
833 list_add_tail(&rm->m_sock_item, &rs->rs_send_queue);
834 set_bit(RDS_MSG_ON_SOCK, &rm->m_flags);
835 rds_message_addref(rm);
838 /* The code ordering is a little weird, but we're
839 trying to minimize the time we hold c_lock */
840 rds_message_populate_header(&rm->m_inc.i_hdr, sport, dport, 0);
841 rm->m_inc.i_conn = conn;
842 rm->m_inc.i_conn_path = cp;
843 rds_message_addref(rm);
845 spin_lock(&cp->cp_lock);
846 rm->m_inc.i_hdr.h_sequence = cpu_to_be64(cp->cp_next_tx_seq++);
847 list_add_tail(&rm->m_conn_item, &cp->cp_send_queue);
848 set_bit(RDS_MSG_ON_CONN, &rm->m_flags);
849 spin_unlock(&cp->cp_lock);
851 rdsdebug("queued msg %p len %d, rs %p bytes %d seq %llu\n",
852 rm, len, rs, rs->rs_snd_bytes,
853 (unsigned long long)be64_to_cpu(rm->m_inc.i_hdr.h_sequence));
858 spin_unlock_irqrestore(&rs->rs_lock, flags);
864 * rds_message is getting to be quite complicated, and we'd like to allocate
865 * it all in one go. This figures out how big it needs to be up front.
867 static int rds_rm_size(struct msghdr *msg, int data_len)
869 struct cmsghdr *cmsg;
874 for_each_cmsghdr(cmsg, msg) {
875 if (!CMSG_OK(msg, cmsg))
878 if (cmsg->cmsg_level != SOL_RDS)
881 switch (cmsg->cmsg_type) {
882 case RDS_CMSG_RDMA_ARGS:
884 retval = rds_rdma_extra_size(CMSG_DATA(cmsg));
891 case RDS_CMSG_RDMA_DEST:
892 case RDS_CMSG_RDMA_MAP:
894 /* these are valid but do no add any size */
897 case RDS_CMSG_ATOMIC_CSWP:
898 case RDS_CMSG_ATOMIC_FADD:
899 case RDS_CMSG_MASKED_ATOMIC_CSWP:
900 case RDS_CMSG_MASKED_ATOMIC_FADD:
902 size += sizeof(struct scatterlist);
911 size += ceil(data_len, PAGE_SIZE) * sizeof(struct scatterlist);
913 /* Ensure (DEST, MAP) are never used with (ARGS, ATOMIC) */
914 if (cmsg_groups == 3)
920 static int rds_cmsg_send(struct rds_sock *rs, struct rds_message *rm,
921 struct msghdr *msg, int *allocated_mr)
923 struct cmsghdr *cmsg;
926 for_each_cmsghdr(cmsg, msg) {
927 if (!CMSG_OK(msg, cmsg))
930 if (cmsg->cmsg_level != SOL_RDS)
933 /* As a side effect, RDMA_DEST and RDMA_MAP will set
934 * rm->rdma.m_rdma_cookie and rm->rdma.m_rdma_mr.
936 switch (cmsg->cmsg_type) {
937 case RDS_CMSG_RDMA_ARGS:
938 ret = rds_cmsg_rdma_args(rs, rm, cmsg);
941 case RDS_CMSG_RDMA_DEST:
942 ret = rds_cmsg_rdma_dest(rs, rm, cmsg);
945 case RDS_CMSG_RDMA_MAP:
946 ret = rds_cmsg_rdma_map(rs, rm, cmsg);
949 else if (ret == -ENODEV)
950 /* Accommodate the get_mr() case which can fail
951 * if connection isn't established yet.
955 case RDS_CMSG_ATOMIC_CSWP:
956 case RDS_CMSG_ATOMIC_FADD:
957 case RDS_CMSG_MASKED_ATOMIC_CSWP:
958 case RDS_CMSG_MASKED_ATOMIC_FADD:
959 ret = rds_cmsg_atomic(rs, rm, cmsg);
973 static void rds_send_ping(struct rds_connection *conn);
975 static int rds_send_mprds_hash(struct rds_sock *rs, struct rds_connection *conn)
979 if (conn->c_npaths == 0)
980 hash = RDS_MPATH_HASH(rs, RDS_MPATH_WORKERS);
982 hash = RDS_MPATH_HASH(rs, conn->c_npaths);
983 if (conn->c_npaths == 0 && hash != 0) {
986 /* The underlying connection is not up yet. Need to wait
987 * until it is up to be sure that the non-zero c_path can be
988 * used. But if we are interrupted, we have to use the zero
989 * c_path in case the connection ends up being non-MP capable.
991 if (conn->c_npaths == 0)
992 if (wait_event_interruptible(conn->c_hs_waitq,
993 conn->c_npaths != 0))
995 if (conn->c_npaths == 1)
1001 static int rds_rdma_bytes(struct msghdr *msg, size_t *rdma_bytes)
1003 struct rds_rdma_args *args;
1004 struct cmsghdr *cmsg;
1006 for_each_cmsghdr(cmsg, msg) {
1007 if (!CMSG_OK(msg, cmsg))
1010 if (cmsg->cmsg_level != SOL_RDS)
1013 if (cmsg->cmsg_type == RDS_CMSG_RDMA_ARGS) {
1014 if (cmsg->cmsg_len <
1015 CMSG_LEN(sizeof(struct rds_rdma_args)))
1017 args = CMSG_DATA(cmsg);
1018 *rdma_bytes += args->remote_vec.bytes;
1024 int rds_sendmsg(struct socket *sock, struct msghdr *msg, size_t payload_len)
1026 struct sock *sk = sock->sk;
1027 struct rds_sock *rs = rds_sk_to_rs(sk);
1028 DECLARE_SOCKADDR(struct sockaddr_in *, usin, msg->msg_name);
1031 struct rds_message *rm = NULL;
1032 struct rds_connection *conn;
1034 int queued = 0, allocated_mr = 0;
1035 int nonblock = msg->msg_flags & MSG_DONTWAIT;
1036 long timeo = sock_sndtimeo(sk, nonblock);
1037 struct rds_conn_path *cpath;
1038 size_t total_payload_len = payload_len, rdma_payload_len = 0;
1040 /* Mirror Linux UDP mirror of BSD error message compatibility */
1041 /* XXX: Perhaps MSG_MORE someday */
1042 if (msg->msg_flags & ~(MSG_DONTWAIT | MSG_CMSG_COMPAT)) {
1047 if (msg->msg_namelen) {
1048 /* XXX fail non-unicast destination IPs? */
1049 if (msg->msg_namelen < sizeof(*usin) || usin->sin_family != AF_INET) {
1053 daddr = usin->sin_addr.s_addr;
1054 dport = usin->sin_port;
1056 /* We only care about consistency with ->connect() */
1058 daddr = rs->rs_conn_addr;
1059 dport = rs->rs_conn_port;
1064 if (daddr == 0 || rs->rs_bound_addr == 0) {
1066 ret = -ENOTCONN; /* XXX not a great errno */
1071 ret = rds_rdma_bytes(msg, &rdma_payload_len);
1075 total_payload_len += rdma_payload_len;
1076 if (max_t(size_t, payload_len, rdma_payload_len) > RDS_MAX_MSG_SIZE) {
1081 if (payload_len > rds_sk_sndbuf(rs)) {
1086 /* size of rm including all sgs */
1087 ret = rds_rm_size(msg, payload_len);
1091 rm = rds_message_alloc(ret, GFP_KERNEL);
1097 /* Attach data to the rm */
1099 rm->data.op_sg = rds_message_alloc_sgs(rm, ceil(payload_len, PAGE_SIZE));
1100 if (!rm->data.op_sg) {
1104 ret = rds_message_copy_from_user(rm, &msg->msg_iter);
1108 rm->data.op_active = 1;
1110 rm->m_daddr = daddr;
1112 /* rds_conn_create has a spinlock that runs with IRQ off.
1113 * Caching the conn in the socket helps a lot. */
1114 if (rs->rs_conn && rs->rs_conn->c_faddr == daddr)
1117 conn = rds_conn_create_outgoing(sock_net(sock->sk),
1118 rs->rs_bound_addr, daddr,
1120 sock->sk->sk_allocation);
1122 ret = PTR_ERR(conn);
1128 /* Parse any control messages the user may have included. */
1129 ret = rds_cmsg_send(rs, rm, msg, &allocated_mr);
1131 /* Trigger connection so that its ready for the next retry */
1133 rds_conn_connect_if_down(conn);
1137 if (rm->rdma.op_active && !conn->c_trans->xmit_rdma) {
1138 printk_ratelimited(KERN_NOTICE "rdma_op %p conn xmit_rdma %p\n",
1139 &rm->rdma, conn->c_trans->xmit_rdma);
1144 if (rm->atomic.op_active && !conn->c_trans->xmit_atomic) {
1145 printk_ratelimited(KERN_NOTICE "atomic_op %p conn xmit_atomic %p\n",
1146 &rm->atomic, conn->c_trans->xmit_atomic);
1151 if (conn->c_trans->t_mp_capable)
1152 cpath = &conn->c_path[rds_send_mprds_hash(rs, conn)];
1154 cpath = &conn->c_path[0];
1156 rds_conn_path_connect_if_down(cpath);
1158 ret = rds_cong_wait(conn->c_fcong, dport, nonblock, rs);
1160 rs->rs_seen_congestion = 1;
1163 while (!rds_send_queue_rm(rs, conn, cpath, rm, rs->rs_bound_port,
1165 rds_stats_inc(s_send_queue_full);
1172 timeo = wait_event_interruptible_timeout(*sk_sleep(sk),
1173 rds_send_queue_rm(rs, conn, cpath, rm,
1178 rdsdebug("sendmsg woke queued %d timeo %ld\n", queued, timeo);
1179 if (timeo > 0 || timeo == MAX_SCHEDULE_TIMEOUT)
1189 * By now we've committed to the send. We reuse rds_send_worker()
1190 * to retry sends in the rds thread if the transport asks us to.
1192 rds_stats_inc(s_send_queued);
1194 ret = rds_send_xmit(cpath);
1195 if (ret == -ENOMEM || ret == -EAGAIN)
1196 queue_delayed_work(rds_wq, &cpath->cp_send_w, 1);
1198 rds_message_put(rm);
1202 /* If the user included a RDMA_MAP cmsg, we allocated a MR on the fly.
1203 * If the sendmsg goes through, we keep the MR. If it fails with EAGAIN
1204 * or in any other way, we need to destroy the MR again */
1206 rds_rdma_unuse(rs, rds_rdma_cookie_key(rm->m_rdma_cookie), 1);
1209 rds_message_put(rm);
1214 * send out a probe. Can be shared by rds_send_ping,
1215 * rds_send_pong, rds_send_hb.
1216 * rds_send_hb should use h_flags
1217 * RDS_FLAG_HB_PING|RDS_FLAG_ACK_REQUIRED
1219 * RDS_FLAG_HB_PONG|RDS_FLAG_ACK_REQUIRED
1222 rds_send_probe(struct rds_conn_path *cp, __be16 sport,
1223 __be16 dport, u8 h_flags)
1225 struct rds_message *rm;
1226 unsigned long flags;
1229 rm = rds_message_alloc(0, GFP_ATOMIC);
1235 rm->m_daddr = cp->cp_conn->c_faddr;
1236 rm->data.op_active = 1;
1238 rds_conn_path_connect_if_down(cp);
1240 ret = rds_cong_wait(cp->cp_conn->c_fcong, dport, 1, NULL);
1244 spin_lock_irqsave(&cp->cp_lock, flags);
1245 list_add_tail(&rm->m_conn_item, &cp->cp_send_queue);
1246 set_bit(RDS_MSG_ON_CONN, &rm->m_flags);
1247 rds_message_addref(rm);
1248 rm->m_inc.i_conn = cp->cp_conn;
1249 rm->m_inc.i_conn_path = cp;
1251 rds_message_populate_header(&rm->m_inc.i_hdr, sport, dport,
1252 cp->cp_next_tx_seq);
1253 rm->m_inc.i_hdr.h_flags |= h_flags;
1254 cp->cp_next_tx_seq++;
1256 if (RDS_HS_PROBE(sport, dport) && cp->cp_conn->c_trans->t_mp_capable) {
1257 u16 npaths = RDS_MPATH_WORKERS;
1259 rds_message_add_extension(&rm->m_inc.i_hdr,
1260 RDS_EXTHDR_NPATHS, &npaths,
1263 spin_unlock_irqrestore(&cp->cp_lock, flags);
1265 rds_stats_inc(s_send_queued);
1266 rds_stats_inc(s_send_pong);
1268 /* schedule the send work on rds_wq */
1269 queue_delayed_work(rds_wq, &cp->cp_send_w, 1);
1271 rds_message_put(rm);
1276 rds_message_put(rm);
1281 rds_send_pong(struct rds_conn_path *cp, __be16 dport)
1283 return rds_send_probe(cp, 0, dport, 0);
1287 rds_send_ping(struct rds_connection *conn)
1289 unsigned long flags;
1290 struct rds_conn_path *cp = &conn->c_path[0];
1292 spin_lock_irqsave(&cp->cp_lock, flags);
1293 if (conn->c_ping_triggered) {
1294 spin_unlock_irqrestore(&cp->cp_lock, flags);
1297 conn->c_ping_triggered = 1;
1298 spin_unlock_irqrestore(&cp->cp_lock, flags);
1299 rds_send_probe(&conn->c_path[0], RDS_FLAG_PROBE_PORT, 0, 0);