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);
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.
204 * If between sending messages, we can send a pending congestion
207 if (!rm && test_and_clear_bit(0, &conn->c_map_queued)) {
208 rm = rds_cong_update_alloc(conn);
213 rm->data.op_active = 1;
214 rm->m_inc.i_conn_path = cp;
215 rm->m_inc.i_conn = cp->cp_conn;
221 * If not already working on one, grab the next message.
223 * cp_xmit_rm holds a ref while we're sending this message down
224 * the connction. We can use this ref while holding the
225 * send_sem.. rds_send_reset() is serialized with it.
232 /* we want to process as big a batch as we can, but
233 * we also want to avoid softlockups. If we've been
234 * through a lot of messages, lets back off and see
235 * if anyone else jumps in
237 if (batch_count >= send_batch_count)
240 spin_lock_irqsave(&cp->cp_lock, flags);
242 if (!list_empty(&cp->cp_send_queue)) {
243 rm = list_entry(cp->cp_send_queue.next,
246 rds_message_addref(rm);
249 * Move the message from the send queue to the retransmit
252 list_move_tail(&rm->m_conn_item,
256 spin_unlock_irqrestore(&cp->cp_lock, flags);
261 /* Unfortunately, the way Infiniband deals with
262 * RDMA to a bad MR key is by moving the entire
263 * queue pair to error state. We cold possibly
264 * recover from that, but right now we drop the
266 * Therefore, we never retransmit messages with RDMA ops.
268 if (test_bit(RDS_MSG_FLUSH, &rm->m_flags) ||
269 (rm->rdma.op_active &&
270 test_bit(RDS_MSG_RETRANSMITTED, &rm->m_flags))) {
271 spin_lock_irqsave(&cp->cp_lock, flags);
272 if (test_and_clear_bit(RDS_MSG_ON_CONN, &rm->m_flags))
273 list_move(&rm->m_conn_item, &to_be_dropped);
274 spin_unlock_irqrestore(&cp->cp_lock, flags);
278 /* Require an ACK every once in a while */
279 len = ntohl(rm->m_inc.i_hdr.h_len);
280 if (cp->cp_unacked_packets == 0 ||
281 cp->cp_unacked_bytes < len) {
282 set_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags);
284 cp->cp_unacked_packets =
285 rds_sysctl_max_unacked_packets;
286 cp->cp_unacked_bytes =
287 rds_sysctl_max_unacked_bytes;
288 rds_stats_inc(s_send_ack_required);
290 cp->cp_unacked_bytes -= len;
291 cp->cp_unacked_packets--;
297 /* The transport either sends the whole rdma or none of it */
298 if (rm->rdma.op_active && !cp->cp_xmit_rdma_sent) {
299 rm->m_final_op = &rm->rdma;
300 /* The transport owns the mapped memory for now.
301 * You can't unmap it while it's on the send queue
303 set_bit(RDS_MSG_MAPPED, &rm->m_flags);
304 ret = conn->c_trans->xmit_rdma(conn, &rm->rdma);
306 clear_bit(RDS_MSG_MAPPED, &rm->m_flags);
307 wake_up_interruptible(&rm->m_flush_wait);
310 cp->cp_xmit_rdma_sent = 1;
314 if (rm->atomic.op_active && !cp->cp_xmit_atomic_sent) {
315 rm->m_final_op = &rm->atomic;
316 /* The transport owns the mapped memory for now.
317 * You can't unmap it while it's on the send queue
319 set_bit(RDS_MSG_MAPPED, &rm->m_flags);
320 ret = conn->c_trans->xmit_atomic(conn, &rm->atomic);
322 clear_bit(RDS_MSG_MAPPED, &rm->m_flags);
323 wake_up_interruptible(&rm->m_flush_wait);
326 cp->cp_xmit_atomic_sent = 1;
331 * A number of cases require an RDS header to be sent
332 * even if there is no data.
333 * We permit 0-byte sends; rds-ping depends on this.
334 * However, if there are exclusively attached silent ops,
335 * we skip the hdr/data send, to enable silent operation.
337 if (rm->data.op_nents == 0) {
339 int all_ops_are_silent = 1;
341 ops_present = (rm->atomic.op_active || rm->rdma.op_active);
342 if (rm->atomic.op_active && !rm->atomic.op_silent)
343 all_ops_are_silent = 0;
344 if (rm->rdma.op_active && !rm->rdma.op_silent)
345 all_ops_are_silent = 0;
347 if (ops_present && all_ops_are_silent
348 && !rm->m_rdma_cookie)
349 rm->data.op_active = 0;
352 if (rm->data.op_active && !cp->cp_xmit_data_sent) {
353 rm->m_final_op = &rm->data;
355 ret = conn->c_trans->xmit(conn, rm,
358 cp->cp_xmit_data_off);
362 if (cp->cp_xmit_hdr_off < sizeof(struct rds_header)) {
363 tmp = min_t(int, ret,
364 sizeof(struct rds_header) -
365 cp->cp_xmit_hdr_off);
366 cp->cp_xmit_hdr_off += tmp;
370 sg = &rm->data.op_sg[cp->cp_xmit_sg];
372 tmp = min_t(int, ret, sg->length -
373 cp->cp_xmit_data_off);
374 cp->cp_xmit_data_off += tmp;
376 if (cp->cp_xmit_data_off == sg->length) {
377 cp->cp_xmit_data_off = 0;
380 BUG_ON(ret != 0 && cp->cp_xmit_sg ==
385 if (cp->cp_xmit_hdr_off == sizeof(struct rds_header) &&
386 (cp->cp_xmit_sg == rm->data.op_nents))
387 cp->cp_xmit_data_sent = 1;
391 * A rm will only take multiple times through this loop
392 * if there is a data op. Thus, if the data is sent (or there was
393 * none), then we're done with the rm.
395 if (!rm->data.op_active || cp->cp_xmit_data_sent) {
396 cp->cp_xmit_rm = NULL;
398 cp->cp_xmit_hdr_off = 0;
399 cp->cp_xmit_data_off = 0;
400 cp->cp_xmit_rdma_sent = 0;
401 cp->cp_xmit_atomic_sent = 0;
402 cp->cp_xmit_data_sent = 0;
409 if (conn->c_trans->xmit_path_complete)
410 conn->c_trans->xmit_path_complete(cp);
413 /* Nuke any messages we decided not to retransmit. */
414 if (!list_empty(&to_be_dropped)) {
415 /* irqs on here, so we can put(), unlike above */
416 list_for_each_entry(rm, &to_be_dropped, m_conn_item)
418 rds_send_remove_from_sock(&to_be_dropped, RDS_RDMA_DROPPED);
422 * Other senders can queue a message after we last test the send queue
423 * but before we clear RDS_IN_XMIT. In that case they'd back off and
424 * not try and send their newly queued message. We need to check the
425 * send queue after having cleared RDS_IN_XMIT so that their message
426 * doesn't get stuck on the send queue.
428 * If the transport cannot continue (i.e ret != 0), then it must
429 * call us when more room is available, such as from the tx
430 * completion handler.
432 * We have an extra generation check here so that if someone manages
433 * to jump in after our release_in_xmit, we'll see that they have done
434 * some work and we will skip our goto
440 raced = send_gen != READ_ONCE(cp->cp_send_gen);
442 if ((test_bit(0, &conn->c_map_queued) ||
443 !list_empty(&cp->cp_send_queue)) && !raced) {
444 if (batch_count < send_batch_count)
447 if (rds_destroy_pending(cp->cp_conn))
450 queue_delayed_work(rds_wq, &cp->cp_send_w, 1);
453 rds_stats_inc(s_send_lock_queue_raced);
459 EXPORT_SYMBOL_GPL(rds_send_xmit);
461 static void rds_send_sndbuf_remove(struct rds_sock *rs, struct rds_message *rm)
463 u32 len = be32_to_cpu(rm->m_inc.i_hdr.h_len);
465 assert_spin_locked(&rs->rs_lock);
467 BUG_ON(rs->rs_snd_bytes < len);
468 rs->rs_snd_bytes -= len;
470 if (rs->rs_snd_bytes == 0)
471 rds_stats_inc(s_send_queue_empty);
474 static inline int rds_send_is_acked(struct rds_message *rm, u64 ack,
475 is_acked_func is_acked)
478 return is_acked(rm, ack);
479 return be64_to_cpu(rm->m_inc.i_hdr.h_sequence) <= ack;
483 * This is pretty similar to what happens below in the ACK
484 * handling code - except that we call here as soon as we get
485 * the IB send completion on the RDMA op and the accompanying
488 void rds_rdma_send_complete(struct rds_message *rm, int status)
490 struct rds_sock *rs = NULL;
491 struct rm_rdma_op *ro;
492 struct rds_notifier *notifier;
494 unsigned int notify = 0;
496 spin_lock_irqsave(&rm->m_rs_lock, flags);
498 notify = rm->rdma.op_notify | rm->data.op_notify;
500 if (test_bit(RDS_MSG_ON_SOCK, &rm->m_flags) &&
501 ro->op_active && notify && ro->op_notifier) {
502 notifier = ro->op_notifier;
504 sock_hold(rds_rs_to_sk(rs));
506 notifier->n_status = status;
507 spin_lock(&rs->rs_lock);
508 list_add_tail(¬ifier->n_list, &rs->rs_notify_queue);
509 spin_unlock(&rs->rs_lock);
511 ro->op_notifier = NULL;
514 spin_unlock_irqrestore(&rm->m_rs_lock, flags);
517 rds_wake_sk_sleep(rs);
518 sock_put(rds_rs_to_sk(rs));
521 EXPORT_SYMBOL_GPL(rds_rdma_send_complete);
524 * Just like above, except looks at atomic op
526 void rds_atomic_send_complete(struct rds_message *rm, int status)
528 struct rds_sock *rs = NULL;
529 struct rm_atomic_op *ao;
530 struct rds_notifier *notifier;
533 spin_lock_irqsave(&rm->m_rs_lock, flags);
536 if (test_bit(RDS_MSG_ON_SOCK, &rm->m_flags)
537 && ao->op_active && ao->op_notify && ao->op_notifier) {
538 notifier = ao->op_notifier;
540 sock_hold(rds_rs_to_sk(rs));
542 notifier->n_status = status;
543 spin_lock(&rs->rs_lock);
544 list_add_tail(¬ifier->n_list, &rs->rs_notify_queue);
545 spin_unlock(&rs->rs_lock);
547 ao->op_notifier = NULL;
550 spin_unlock_irqrestore(&rm->m_rs_lock, flags);
553 rds_wake_sk_sleep(rs);
554 sock_put(rds_rs_to_sk(rs));
557 EXPORT_SYMBOL_GPL(rds_atomic_send_complete);
560 * This is the same as rds_rdma_send_complete except we
561 * don't do any locking - we have all the ingredients (message,
562 * socket, socket lock) and can just move the notifier.
565 __rds_send_complete(struct rds_sock *rs, struct rds_message *rm, int status)
567 struct rm_rdma_op *ro;
568 struct rm_atomic_op *ao;
571 if (ro->op_active && ro->op_notify && ro->op_notifier) {
572 ro->op_notifier->n_status = status;
573 list_add_tail(&ro->op_notifier->n_list, &rs->rs_notify_queue);
574 ro->op_notifier = NULL;
578 if (ao->op_active && ao->op_notify && ao->op_notifier) {
579 ao->op_notifier->n_status = status;
580 list_add_tail(&ao->op_notifier->n_list, &rs->rs_notify_queue);
581 ao->op_notifier = NULL;
584 /* No need to wake the app - caller does this */
588 * This removes messages from the socket's list if they're on it. The list
589 * argument must be private to the caller, we must be able to modify it
590 * without locks. The messages must have a reference held for their
591 * position on the list. This function will drop that reference after
592 * removing the messages from the 'messages' list regardless of if it found
593 * the messages on the socket list or not.
595 static void rds_send_remove_from_sock(struct list_head *messages, int status)
598 struct rds_sock *rs = NULL;
599 struct rds_message *rm;
601 while (!list_empty(messages)) {
604 rm = list_entry(messages->next, struct rds_message,
606 list_del_init(&rm->m_conn_item);
609 * If we see this flag cleared then we're *sure* that someone
610 * else beat us to removing it from the sock. If we race
611 * with their flag update we'll get the lock and then really
612 * see that the flag has been cleared.
614 * The message spinlock makes sure nobody clears rm->m_rs
615 * while we're messing with it. It does not prevent the
616 * message from being removed from the socket, though.
618 spin_lock_irqsave(&rm->m_rs_lock, flags);
619 if (!test_bit(RDS_MSG_ON_SOCK, &rm->m_flags))
620 goto unlock_and_drop;
622 if (rs != rm->m_rs) {
624 rds_wake_sk_sleep(rs);
625 sock_put(rds_rs_to_sk(rs));
629 sock_hold(rds_rs_to_sk(rs));
632 goto unlock_and_drop;
633 spin_lock(&rs->rs_lock);
635 if (test_and_clear_bit(RDS_MSG_ON_SOCK, &rm->m_flags)) {
636 struct rm_rdma_op *ro = &rm->rdma;
637 struct rds_notifier *notifier;
639 list_del_init(&rm->m_sock_item);
640 rds_send_sndbuf_remove(rs, rm);
642 if (ro->op_active && ro->op_notifier &&
643 (ro->op_notify || (ro->op_recverr && status))) {
644 notifier = ro->op_notifier;
645 list_add_tail(¬ifier->n_list,
646 &rs->rs_notify_queue);
647 if (!notifier->n_status)
648 notifier->n_status = status;
649 rm->rdma.op_notifier = NULL;
653 spin_unlock(&rs->rs_lock);
656 spin_unlock_irqrestore(&rm->m_rs_lock, flags);
663 rds_wake_sk_sleep(rs);
664 sock_put(rds_rs_to_sk(rs));
669 * Transports call here when they've determined that the receiver queued
670 * messages up to, and including, the given sequence number. Messages are
671 * moved to the retrans queue when rds_send_xmit picks them off the send
672 * queue. This means that in the TCP case, the message may not have been
673 * assigned the m_ack_seq yet - but that's fine as long as tcp_is_acked
674 * checks the RDS_MSG_HAS_ACK_SEQ bit.
676 void rds_send_path_drop_acked(struct rds_conn_path *cp, u64 ack,
677 is_acked_func is_acked)
679 struct rds_message *rm, *tmp;
683 spin_lock_irqsave(&cp->cp_lock, flags);
685 list_for_each_entry_safe(rm, tmp, &cp->cp_retrans, m_conn_item) {
686 if (!rds_send_is_acked(rm, ack, is_acked))
689 list_move(&rm->m_conn_item, &list);
690 clear_bit(RDS_MSG_ON_CONN, &rm->m_flags);
693 /* order flag updates with spin locks */
694 if (!list_empty(&list))
695 smp_mb__after_atomic();
697 spin_unlock_irqrestore(&cp->cp_lock, flags);
699 /* now remove the messages from the sock list as needed */
700 rds_send_remove_from_sock(&list, RDS_RDMA_SUCCESS);
702 EXPORT_SYMBOL_GPL(rds_send_path_drop_acked);
704 void rds_send_drop_acked(struct rds_connection *conn, u64 ack,
705 is_acked_func is_acked)
707 WARN_ON(conn->c_trans->t_mp_capable);
708 rds_send_path_drop_acked(&conn->c_path[0], ack, is_acked);
710 EXPORT_SYMBOL_GPL(rds_send_drop_acked);
712 void rds_send_drop_to(struct rds_sock *rs, struct sockaddr_in *dest)
714 struct rds_message *rm, *tmp;
715 struct rds_connection *conn;
716 struct rds_conn_path *cp;
720 /* get all the messages we're dropping under the rs lock */
721 spin_lock_irqsave(&rs->rs_lock, flags);
723 list_for_each_entry_safe(rm, tmp, &rs->rs_send_queue, m_sock_item) {
724 if (dest && (dest->sin_addr.s_addr != rm->m_daddr ||
725 dest->sin_port != rm->m_inc.i_hdr.h_dport))
728 list_move(&rm->m_sock_item, &list);
729 rds_send_sndbuf_remove(rs, rm);
730 clear_bit(RDS_MSG_ON_SOCK, &rm->m_flags);
733 /* order flag updates with the rs lock */
734 smp_mb__after_atomic();
736 spin_unlock_irqrestore(&rs->rs_lock, flags);
738 if (list_empty(&list))
741 /* Remove the messages from the conn */
742 list_for_each_entry(rm, &list, m_sock_item) {
744 conn = rm->m_inc.i_conn;
745 if (conn->c_trans->t_mp_capable)
746 cp = rm->m_inc.i_conn_path;
748 cp = &conn->c_path[0];
750 spin_lock_irqsave(&cp->cp_lock, flags);
752 * Maybe someone else beat us to removing rm from the conn.
753 * If we race with their flag update we'll get the lock and
754 * then really see that the flag has been cleared.
756 if (!test_and_clear_bit(RDS_MSG_ON_CONN, &rm->m_flags)) {
757 spin_unlock_irqrestore(&cp->cp_lock, flags);
760 list_del_init(&rm->m_conn_item);
761 spin_unlock_irqrestore(&cp->cp_lock, flags);
764 * Couldn't grab m_rs_lock in top loop (lock ordering),
767 spin_lock_irqsave(&rm->m_rs_lock, flags);
769 spin_lock(&rs->rs_lock);
770 __rds_send_complete(rs, rm, RDS_RDMA_CANCELED);
771 spin_unlock(&rs->rs_lock);
773 spin_unlock_irqrestore(&rm->m_rs_lock, flags);
778 rds_wake_sk_sleep(rs);
780 while (!list_empty(&list)) {
781 rm = list_entry(list.next, struct rds_message, m_sock_item);
782 list_del_init(&rm->m_sock_item);
783 rds_message_wait(rm);
785 /* just in case the code above skipped this message
786 * because RDS_MSG_ON_CONN wasn't set, run it again here
787 * taking m_rs_lock is the only thing that keeps us
788 * from racing with ack processing.
790 spin_lock_irqsave(&rm->m_rs_lock, flags);
792 spin_lock(&rs->rs_lock);
793 __rds_send_complete(rs, rm, RDS_RDMA_CANCELED);
794 spin_unlock(&rs->rs_lock);
796 spin_unlock_irqrestore(&rm->m_rs_lock, flags);
803 * we only want this to fire once so we use the callers 'queued'. It's
804 * possible that another thread can race with us and remove the
805 * message from the flow with RDS_CANCEL_SENT_TO.
807 static int rds_send_queue_rm(struct rds_sock *rs, struct rds_connection *conn,
808 struct rds_conn_path *cp,
809 struct rds_message *rm, __be16 sport,
810 __be16 dport, int *queued)
818 len = be32_to_cpu(rm->m_inc.i_hdr.h_len);
820 /* this is the only place which holds both the socket's rs_lock
821 * and the connection's c_lock */
822 spin_lock_irqsave(&rs->rs_lock, flags);
825 * If there is a little space in sndbuf, we don't queue anything,
826 * and userspace gets -EAGAIN. But poll() indicates there's send
827 * room. This can lead to bad behavior (spinning) if snd_bytes isn't
828 * freed up by incoming acks. So we check the *old* value of
829 * rs_snd_bytes here to allow the last msg to exceed the buffer,
830 * and poll() now knows no more data can be sent.
832 if (rs->rs_snd_bytes < rds_sk_sndbuf(rs)) {
833 rs->rs_snd_bytes += len;
835 /* let recv side know we are close to send space exhaustion.
836 * This is probably not the optimal way to do it, as this
837 * means we set the flag on *all* messages as soon as our
838 * throughput hits a certain threshold.
840 if (rs->rs_snd_bytes >= rds_sk_sndbuf(rs) / 2)
841 set_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags);
843 list_add_tail(&rm->m_sock_item, &rs->rs_send_queue);
844 set_bit(RDS_MSG_ON_SOCK, &rm->m_flags);
845 rds_message_addref(rm);
846 sock_hold(rds_rs_to_sk(rs));
849 /* The code ordering is a little weird, but we're
850 trying to minimize the time we hold c_lock */
851 rds_message_populate_header(&rm->m_inc.i_hdr, sport, dport, 0);
852 rm->m_inc.i_conn = conn;
853 rm->m_inc.i_conn_path = cp;
854 rds_message_addref(rm);
856 spin_lock(&cp->cp_lock);
857 rm->m_inc.i_hdr.h_sequence = cpu_to_be64(cp->cp_next_tx_seq++);
858 list_add_tail(&rm->m_conn_item, &cp->cp_send_queue);
859 set_bit(RDS_MSG_ON_CONN, &rm->m_flags);
860 spin_unlock(&cp->cp_lock);
862 rdsdebug("queued msg %p len %d, rs %p bytes %d seq %llu\n",
863 rm, len, rs, rs->rs_snd_bytes,
864 (unsigned long long)be64_to_cpu(rm->m_inc.i_hdr.h_sequence));
869 spin_unlock_irqrestore(&rs->rs_lock, flags);
875 * rds_message is getting to be quite complicated, and we'd like to allocate
876 * it all in one go. This figures out how big it needs to be up front.
878 static int rds_rm_size(struct msghdr *msg, int num_sgs)
880 struct cmsghdr *cmsg;
884 bool zcopy_cookie = false;
886 for_each_cmsghdr(cmsg, msg) {
887 if (!CMSG_OK(msg, cmsg))
890 if (cmsg->cmsg_level != SOL_RDS)
893 switch (cmsg->cmsg_type) {
894 case RDS_CMSG_RDMA_ARGS:
896 retval = rds_rdma_extra_size(CMSG_DATA(cmsg));
903 case RDS_CMSG_ZCOPY_COOKIE:
907 case RDS_CMSG_RDMA_DEST:
908 case RDS_CMSG_RDMA_MAP:
910 /* these are valid but do no add any size */
913 case RDS_CMSG_ATOMIC_CSWP:
914 case RDS_CMSG_ATOMIC_FADD:
915 case RDS_CMSG_MASKED_ATOMIC_CSWP:
916 case RDS_CMSG_MASKED_ATOMIC_FADD:
918 size += sizeof(struct scatterlist);
927 if ((msg->msg_flags & MSG_ZEROCOPY) && !zcopy_cookie)
930 size += num_sgs * sizeof(struct scatterlist);
932 /* Ensure (DEST, MAP) are never used with (ARGS, ATOMIC) */
933 if (cmsg_groups == 3)
939 static int rds_cmsg_zcopy(struct rds_sock *rs, struct rds_message *rm,
940 struct cmsghdr *cmsg)
944 if (cmsg->cmsg_len < CMSG_LEN(sizeof(*cookie)) ||
945 !rm->data.op_mmp_znotifier)
947 cookie = CMSG_DATA(cmsg);
948 rm->data.op_mmp_znotifier->z_cookie = *cookie;
952 static int rds_cmsg_send(struct rds_sock *rs, struct rds_message *rm,
953 struct msghdr *msg, int *allocated_mr)
955 struct cmsghdr *cmsg;
958 for_each_cmsghdr(cmsg, msg) {
959 if (!CMSG_OK(msg, cmsg))
962 if (cmsg->cmsg_level != SOL_RDS)
965 /* As a side effect, RDMA_DEST and RDMA_MAP will set
966 * rm->rdma.m_rdma_cookie and rm->rdma.m_rdma_mr.
968 switch (cmsg->cmsg_type) {
969 case RDS_CMSG_RDMA_ARGS:
970 ret = rds_cmsg_rdma_args(rs, rm, cmsg);
973 case RDS_CMSG_RDMA_DEST:
974 ret = rds_cmsg_rdma_dest(rs, rm, cmsg);
977 case RDS_CMSG_RDMA_MAP:
978 ret = rds_cmsg_rdma_map(rs, rm, cmsg);
981 else if (ret == -ENODEV)
982 /* Accommodate the get_mr() case which can fail
983 * if connection isn't established yet.
987 case RDS_CMSG_ATOMIC_CSWP:
988 case RDS_CMSG_ATOMIC_FADD:
989 case RDS_CMSG_MASKED_ATOMIC_CSWP:
990 case RDS_CMSG_MASKED_ATOMIC_FADD:
991 ret = rds_cmsg_atomic(rs, rm, cmsg);
994 case RDS_CMSG_ZCOPY_COOKIE:
995 ret = rds_cmsg_zcopy(rs, rm, cmsg);
1009 static int rds_send_mprds_hash(struct rds_sock *rs, struct rds_connection *conn)
1013 if (conn->c_npaths == 0)
1014 hash = RDS_MPATH_HASH(rs, RDS_MPATH_WORKERS);
1016 hash = RDS_MPATH_HASH(rs, conn->c_npaths);
1017 if (conn->c_npaths == 0 && hash != 0) {
1018 rds_send_ping(conn, 0);
1020 /* The underlying connection is not up yet. Need to wait
1021 * until it is up to be sure that the non-zero c_path can be
1022 * used. But if we are interrupted, we have to use the zero
1023 * c_path in case the connection ends up being non-MP capable.
1025 if (conn->c_npaths == 0)
1026 if (wait_event_interruptible(conn->c_hs_waitq,
1027 conn->c_npaths != 0))
1029 if (conn->c_npaths == 1)
1035 static int rds_rdma_bytes(struct msghdr *msg, size_t *rdma_bytes)
1037 struct rds_rdma_args *args;
1038 struct cmsghdr *cmsg;
1040 for_each_cmsghdr(cmsg, msg) {
1041 if (!CMSG_OK(msg, cmsg))
1044 if (cmsg->cmsg_level != SOL_RDS)
1047 if (cmsg->cmsg_type == RDS_CMSG_RDMA_ARGS) {
1048 if (cmsg->cmsg_len <
1049 CMSG_LEN(sizeof(struct rds_rdma_args)))
1051 args = CMSG_DATA(cmsg);
1052 *rdma_bytes += args->remote_vec.bytes;
1058 int rds_sendmsg(struct socket *sock, struct msghdr *msg, size_t payload_len)
1060 struct sock *sk = sock->sk;
1061 struct rds_sock *rs = rds_sk_to_rs(sk);
1062 DECLARE_SOCKADDR(struct sockaddr_in *, usin, msg->msg_name);
1065 struct rds_message *rm = NULL;
1066 struct rds_connection *conn;
1068 int queued = 0, allocated_mr = 0;
1069 int nonblock = msg->msg_flags & MSG_DONTWAIT;
1070 long timeo = sock_sndtimeo(sk, nonblock);
1071 struct rds_conn_path *cpath;
1072 size_t total_payload_len = payload_len, rdma_payload_len = 0;
1073 bool zcopy = ((msg->msg_flags & MSG_ZEROCOPY) &&
1074 sock_flag(rds_rs_to_sk(rs), SOCK_ZEROCOPY));
1075 int num_sgs = ceil(payload_len, PAGE_SIZE);
1077 /* Mirror Linux UDP mirror of BSD error message compatibility */
1078 /* XXX: Perhaps MSG_MORE someday */
1079 if (msg->msg_flags & ~(MSG_DONTWAIT | MSG_CMSG_COMPAT | MSG_ZEROCOPY)) {
1084 if (msg->msg_namelen) {
1085 /* XXX fail non-unicast destination IPs? */
1086 if (msg->msg_namelen < sizeof(*usin) || usin->sin_family != AF_INET) {
1090 daddr = usin->sin_addr.s_addr;
1091 dport = usin->sin_port;
1093 /* We only care about consistency with ->connect() */
1095 daddr = rs->rs_conn_addr;
1096 dport = rs->rs_conn_port;
1101 if (daddr == 0 || rs->rs_bound_addr == 0) {
1103 ret = -ENOTCONN; /* XXX not a great errno */
1108 ret = rds_rdma_bytes(msg, &rdma_payload_len);
1112 total_payload_len += rdma_payload_len;
1113 if (max_t(size_t, payload_len, rdma_payload_len) > RDS_MAX_MSG_SIZE) {
1118 if (payload_len > rds_sk_sndbuf(rs)) {
1124 if (rs->rs_transport->t_type != RDS_TRANS_TCP) {
1128 num_sgs = iov_iter_npages(&msg->msg_iter, INT_MAX);
1130 /* size of rm including all sgs */
1131 ret = rds_rm_size(msg, num_sgs);
1135 rm = rds_message_alloc(ret, GFP_KERNEL);
1141 /* Attach data to the rm */
1143 rm->data.op_sg = rds_message_alloc_sgs(rm, num_sgs);
1144 if (!rm->data.op_sg) {
1148 ret = rds_message_copy_from_user(rm, &msg->msg_iter, zcopy);
1152 rm->data.op_active = 1;
1154 rm->m_daddr = daddr;
1156 /* rds_conn_create has a spinlock that runs with IRQ off.
1157 * Caching the conn in the socket helps a lot. */
1158 if (rs->rs_conn && rs->rs_conn->c_faddr == daddr)
1161 conn = rds_conn_create_outgoing(sock_net(sock->sk),
1162 rs->rs_bound_addr, daddr,
1164 sock->sk->sk_allocation);
1166 ret = PTR_ERR(conn);
1172 /* Parse any control messages the user may have included. */
1173 ret = rds_cmsg_send(rs, rm, msg, &allocated_mr);
1175 /* Trigger connection so that its ready for the next retry */
1177 rds_conn_connect_if_down(conn);
1181 if (rm->rdma.op_active && !conn->c_trans->xmit_rdma) {
1182 printk_ratelimited(KERN_NOTICE "rdma_op %p conn xmit_rdma %p\n",
1183 &rm->rdma, conn->c_trans->xmit_rdma);
1188 if (rm->atomic.op_active && !conn->c_trans->xmit_atomic) {
1189 printk_ratelimited(KERN_NOTICE "atomic_op %p conn xmit_atomic %p\n",
1190 &rm->atomic, conn->c_trans->xmit_atomic);
1195 if (conn->c_trans->t_mp_capable)
1196 cpath = &conn->c_path[rds_send_mprds_hash(rs, conn)];
1198 cpath = &conn->c_path[0];
1200 if (rds_destroy_pending(conn)) {
1205 rds_conn_path_connect_if_down(cpath);
1207 ret = rds_cong_wait(conn->c_fcong, dport, nonblock, rs);
1209 rs->rs_seen_congestion = 1;
1212 while (!rds_send_queue_rm(rs, conn, cpath, rm, rs->rs_bound_port,
1214 rds_stats_inc(s_send_queue_full);
1221 timeo = wait_event_interruptible_timeout(*sk_sleep(sk),
1222 rds_send_queue_rm(rs, conn, cpath, rm,
1227 rdsdebug("sendmsg woke queued %d timeo %ld\n", queued, timeo);
1228 if (timeo > 0 || timeo == MAX_SCHEDULE_TIMEOUT)
1238 * By now we've committed to the send. We reuse rds_send_worker()
1239 * to retry sends in the rds thread if the transport asks us to.
1241 rds_stats_inc(s_send_queued);
1243 ret = rds_send_xmit(cpath);
1244 if (ret == -ENOMEM || ret == -EAGAIN) {
1247 if (rds_destroy_pending(cpath->cp_conn))
1250 queue_delayed_work(rds_wq, &cpath->cp_send_w, 1);
1255 rds_message_put(rm);
1259 /* If the user included a RDMA_MAP cmsg, we allocated a MR on the fly.
1260 * If the sendmsg goes through, we keep the MR. If it fails with EAGAIN
1261 * or in any other way, we need to destroy the MR again */
1263 rds_rdma_unuse(rs, rds_rdma_cookie_key(rm->m_rdma_cookie), 1);
1266 rds_message_put(rm);
1271 * send out a probe. Can be shared by rds_send_ping,
1272 * rds_send_pong, rds_send_hb.
1273 * rds_send_hb should use h_flags
1274 * RDS_FLAG_HB_PING|RDS_FLAG_ACK_REQUIRED
1276 * RDS_FLAG_HB_PONG|RDS_FLAG_ACK_REQUIRED
1279 rds_send_probe(struct rds_conn_path *cp, __be16 sport,
1280 __be16 dport, u8 h_flags)
1282 struct rds_message *rm;
1283 unsigned long flags;
1286 rm = rds_message_alloc(0, GFP_ATOMIC);
1292 rm->m_daddr = cp->cp_conn->c_faddr;
1293 rm->data.op_active = 1;
1295 rds_conn_path_connect_if_down(cp);
1297 ret = rds_cong_wait(cp->cp_conn->c_fcong, dport, 1, NULL);
1301 spin_lock_irqsave(&cp->cp_lock, flags);
1302 list_add_tail(&rm->m_conn_item, &cp->cp_send_queue);
1303 set_bit(RDS_MSG_ON_CONN, &rm->m_flags);
1304 rds_message_addref(rm);
1305 rm->m_inc.i_conn = cp->cp_conn;
1306 rm->m_inc.i_conn_path = cp;
1308 rds_message_populate_header(&rm->m_inc.i_hdr, sport, dport,
1309 cp->cp_next_tx_seq);
1310 rm->m_inc.i_hdr.h_flags |= h_flags;
1311 cp->cp_next_tx_seq++;
1313 if (RDS_HS_PROBE(be16_to_cpu(sport), be16_to_cpu(dport)) &&
1314 cp->cp_conn->c_trans->t_mp_capable) {
1315 u16 npaths = cpu_to_be16(RDS_MPATH_WORKERS);
1316 u32 my_gen_num = cpu_to_be32(cp->cp_conn->c_my_gen_num);
1318 rds_message_add_extension(&rm->m_inc.i_hdr,
1319 RDS_EXTHDR_NPATHS, &npaths,
1321 rds_message_add_extension(&rm->m_inc.i_hdr,
1326 spin_unlock_irqrestore(&cp->cp_lock, flags);
1328 rds_stats_inc(s_send_queued);
1329 rds_stats_inc(s_send_pong);
1331 /* schedule the send work on rds_wq */
1333 if (!rds_destroy_pending(cp->cp_conn))
1334 queue_delayed_work(rds_wq, &cp->cp_send_w, 1);
1337 rds_message_put(rm);
1342 rds_message_put(rm);
1347 rds_send_pong(struct rds_conn_path *cp, __be16 dport)
1349 return rds_send_probe(cp, 0, dport, 0);
1353 rds_send_ping(struct rds_connection *conn, int cp_index)
1355 unsigned long flags;
1356 struct rds_conn_path *cp = &conn->c_path[cp_index];
1358 spin_lock_irqsave(&cp->cp_lock, flags);
1359 if (conn->c_ping_triggered) {
1360 spin_unlock_irqrestore(&cp->cp_lock, flags);
1363 conn->c_ping_triggered = 1;
1364 spin_unlock_irqrestore(&cp->cp_lock, flags);
1365 rds_send_probe(cp, cpu_to_be16(RDS_FLAG_PROBE_PORT), 0, 0);
1367 EXPORT_SYMBOL_GPL(rds_send_ping);