2 * Copyright (c) 2006 Oracle. 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.
173 send_gen = READ_ONCE(cp->cp_send_gen) + 1;
174 WRITE_ONCE(cp->cp_send_gen, 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 (test_bit(RDS_MSG_FLUSH, &rm->m_flags) ||
263 (rm->rdma.op_active &&
264 test_bit(RDS_MSG_RETRANSMITTED, &rm->m_flags))) {
265 spin_lock_irqsave(&cp->cp_lock, flags);
266 if (test_and_clear_bit(RDS_MSG_ON_CONN, &rm->m_flags))
267 list_move(&rm->m_conn_item, &to_be_dropped);
268 spin_unlock_irqrestore(&cp->cp_lock, flags);
272 /* Require an ACK every once in a while */
273 len = ntohl(rm->m_inc.i_hdr.h_len);
274 if (cp->cp_unacked_packets == 0 ||
275 cp->cp_unacked_bytes < len) {
276 set_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags);
278 cp->cp_unacked_packets =
279 rds_sysctl_max_unacked_packets;
280 cp->cp_unacked_bytes =
281 rds_sysctl_max_unacked_bytes;
282 rds_stats_inc(s_send_ack_required);
284 cp->cp_unacked_bytes -= len;
285 cp->cp_unacked_packets--;
291 /* The transport either sends the whole rdma or none of it */
292 if (rm->rdma.op_active && !cp->cp_xmit_rdma_sent) {
293 rm->m_final_op = &rm->rdma;
294 /* The transport owns the mapped memory for now.
295 * You can't unmap it while it's on the send queue
297 set_bit(RDS_MSG_MAPPED, &rm->m_flags);
298 ret = conn->c_trans->xmit_rdma(conn, &rm->rdma);
300 clear_bit(RDS_MSG_MAPPED, &rm->m_flags);
301 wake_up_interruptible(&rm->m_flush_wait);
304 cp->cp_xmit_rdma_sent = 1;
308 if (rm->atomic.op_active && !cp->cp_xmit_atomic_sent) {
309 rm->m_final_op = &rm->atomic;
310 /* The transport owns the mapped memory for now.
311 * You can't unmap it while it's on the send queue
313 set_bit(RDS_MSG_MAPPED, &rm->m_flags);
314 ret = conn->c_trans->xmit_atomic(conn, &rm->atomic);
316 clear_bit(RDS_MSG_MAPPED, &rm->m_flags);
317 wake_up_interruptible(&rm->m_flush_wait);
320 cp->cp_xmit_atomic_sent = 1;
325 * A number of cases require an RDS header to be sent
326 * even if there is no data.
327 * We permit 0-byte sends; rds-ping depends on this.
328 * However, if there are exclusively attached silent ops,
329 * we skip the hdr/data send, to enable silent operation.
331 if (rm->data.op_nents == 0) {
333 int all_ops_are_silent = 1;
335 ops_present = (rm->atomic.op_active || rm->rdma.op_active);
336 if (rm->atomic.op_active && !rm->atomic.op_silent)
337 all_ops_are_silent = 0;
338 if (rm->rdma.op_active && !rm->rdma.op_silent)
339 all_ops_are_silent = 0;
341 if (ops_present && all_ops_are_silent
342 && !rm->m_rdma_cookie)
343 rm->data.op_active = 0;
346 if (rm->data.op_active && !cp->cp_xmit_data_sent) {
347 rm->m_final_op = &rm->data;
349 ret = conn->c_trans->xmit(conn, rm,
352 cp->cp_xmit_data_off);
356 if (cp->cp_xmit_hdr_off < sizeof(struct rds_header)) {
357 tmp = min_t(int, ret,
358 sizeof(struct rds_header) -
359 cp->cp_xmit_hdr_off);
360 cp->cp_xmit_hdr_off += tmp;
364 sg = &rm->data.op_sg[cp->cp_xmit_sg];
366 tmp = min_t(int, ret, sg->length -
367 cp->cp_xmit_data_off);
368 cp->cp_xmit_data_off += tmp;
370 if (cp->cp_xmit_data_off == sg->length) {
371 cp->cp_xmit_data_off = 0;
374 BUG_ON(ret != 0 && cp->cp_xmit_sg ==
379 if (cp->cp_xmit_hdr_off == sizeof(struct rds_header) &&
380 (cp->cp_xmit_sg == rm->data.op_nents))
381 cp->cp_xmit_data_sent = 1;
385 * A rm will only take multiple times through this loop
386 * if there is a data op. Thus, if the data is sent (or there was
387 * none), then we're done with the rm.
389 if (!rm->data.op_active || cp->cp_xmit_data_sent) {
390 cp->cp_xmit_rm = NULL;
392 cp->cp_xmit_hdr_off = 0;
393 cp->cp_xmit_data_off = 0;
394 cp->cp_xmit_rdma_sent = 0;
395 cp->cp_xmit_atomic_sent = 0;
396 cp->cp_xmit_data_sent = 0;
403 if (conn->c_trans->xmit_path_complete)
404 conn->c_trans->xmit_path_complete(cp);
407 /* Nuke any messages we decided not to retransmit. */
408 if (!list_empty(&to_be_dropped)) {
409 /* irqs on here, so we can put(), unlike above */
410 list_for_each_entry(rm, &to_be_dropped, m_conn_item)
412 rds_send_remove_from_sock(&to_be_dropped, RDS_RDMA_DROPPED);
416 * Other senders can queue a message after we last test the send queue
417 * but before we clear RDS_IN_XMIT. In that case they'd back off and
418 * not try and send their newly queued message. We need to check the
419 * send queue after having cleared RDS_IN_XMIT so that their message
420 * doesn't get stuck on the send queue.
422 * If the transport cannot continue (i.e ret != 0), then it must
423 * call us when more room is available, such as from the tx
424 * completion handler.
426 * We have an extra generation check here so that if someone manages
427 * to jump in after our release_in_xmit, we'll see that they have done
428 * some work and we will skip our goto
434 raced = send_gen != READ_ONCE(cp->cp_send_gen);
436 if ((test_bit(0, &conn->c_map_queued) ||
437 !list_empty(&cp->cp_send_queue)) && !raced) {
438 if (batch_count < send_batch_count)
440 queue_delayed_work(rds_wq, &cp->cp_send_w, 1);
442 rds_stats_inc(s_send_lock_queue_raced);
448 EXPORT_SYMBOL_GPL(rds_send_xmit);
450 static void rds_send_sndbuf_remove(struct rds_sock *rs, struct rds_message *rm)
452 u32 len = be32_to_cpu(rm->m_inc.i_hdr.h_len);
454 assert_spin_locked(&rs->rs_lock);
456 BUG_ON(rs->rs_snd_bytes < len);
457 rs->rs_snd_bytes -= len;
459 if (rs->rs_snd_bytes == 0)
460 rds_stats_inc(s_send_queue_empty);
463 static inline int rds_send_is_acked(struct rds_message *rm, u64 ack,
464 is_acked_func is_acked)
467 return is_acked(rm, ack);
468 return be64_to_cpu(rm->m_inc.i_hdr.h_sequence) <= ack;
472 * This is pretty similar to what happens below in the ACK
473 * handling code - except that we call here as soon as we get
474 * the IB send completion on the RDMA op and the accompanying
477 void rds_rdma_send_complete(struct rds_message *rm, int status)
479 struct rds_sock *rs = NULL;
480 struct rm_rdma_op *ro;
481 struct rds_notifier *notifier;
483 unsigned int notify = 0;
485 spin_lock_irqsave(&rm->m_rs_lock, flags);
487 notify = rm->rdma.op_notify | rm->data.op_notify;
489 if (test_bit(RDS_MSG_ON_SOCK, &rm->m_flags) &&
490 ro->op_active && notify && ro->op_notifier) {
491 notifier = ro->op_notifier;
493 sock_hold(rds_rs_to_sk(rs));
495 notifier->n_status = status;
496 spin_lock(&rs->rs_lock);
497 list_add_tail(¬ifier->n_list, &rs->rs_notify_queue);
498 spin_unlock(&rs->rs_lock);
500 ro->op_notifier = NULL;
503 spin_unlock_irqrestore(&rm->m_rs_lock, flags);
506 rds_wake_sk_sleep(rs);
507 sock_put(rds_rs_to_sk(rs));
510 EXPORT_SYMBOL_GPL(rds_rdma_send_complete);
513 * Just like above, except looks at atomic op
515 void rds_atomic_send_complete(struct rds_message *rm, int status)
517 struct rds_sock *rs = NULL;
518 struct rm_atomic_op *ao;
519 struct rds_notifier *notifier;
522 spin_lock_irqsave(&rm->m_rs_lock, flags);
525 if (test_bit(RDS_MSG_ON_SOCK, &rm->m_flags)
526 && ao->op_active && ao->op_notify && ao->op_notifier) {
527 notifier = ao->op_notifier;
529 sock_hold(rds_rs_to_sk(rs));
531 notifier->n_status = status;
532 spin_lock(&rs->rs_lock);
533 list_add_tail(¬ifier->n_list, &rs->rs_notify_queue);
534 spin_unlock(&rs->rs_lock);
536 ao->op_notifier = NULL;
539 spin_unlock_irqrestore(&rm->m_rs_lock, flags);
542 rds_wake_sk_sleep(rs);
543 sock_put(rds_rs_to_sk(rs));
546 EXPORT_SYMBOL_GPL(rds_atomic_send_complete);
549 * This is the same as rds_rdma_send_complete except we
550 * don't do any locking - we have all the ingredients (message,
551 * socket, socket lock) and can just move the notifier.
554 __rds_send_complete(struct rds_sock *rs, struct rds_message *rm, int status)
556 struct rm_rdma_op *ro;
557 struct rm_atomic_op *ao;
560 if (ro->op_active && ro->op_notify && ro->op_notifier) {
561 ro->op_notifier->n_status = status;
562 list_add_tail(&ro->op_notifier->n_list, &rs->rs_notify_queue);
563 ro->op_notifier = NULL;
567 if (ao->op_active && ao->op_notify && ao->op_notifier) {
568 ao->op_notifier->n_status = status;
569 list_add_tail(&ao->op_notifier->n_list, &rs->rs_notify_queue);
570 ao->op_notifier = NULL;
573 /* No need to wake the app - caller does this */
577 * This removes messages from the socket's list if they're on it. The list
578 * argument must be private to the caller, we must be able to modify it
579 * without locks. The messages must have a reference held for their
580 * position on the list. This function will drop that reference after
581 * removing the messages from the 'messages' list regardless of if it found
582 * the messages on the socket list or not.
584 static void rds_send_remove_from_sock(struct list_head *messages, int status)
587 struct rds_sock *rs = NULL;
588 struct rds_message *rm;
590 while (!list_empty(messages)) {
593 rm = list_entry(messages->next, struct rds_message,
595 list_del_init(&rm->m_conn_item);
598 * If we see this flag cleared then we're *sure* that someone
599 * else beat us to removing it from the sock. If we race
600 * with their flag update we'll get the lock and then really
601 * see that the flag has been cleared.
603 * The message spinlock makes sure nobody clears rm->m_rs
604 * while we're messing with it. It does not prevent the
605 * message from being removed from the socket, though.
607 spin_lock_irqsave(&rm->m_rs_lock, flags);
608 if (!test_bit(RDS_MSG_ON_SOCK, &rm->m_flags))
609 goto unlock_and_drop;
611 if (rs != rm->m_rs) {
613 rds_wake_sk_sleep(rs);
614 sock_put(rds_rs_to_sk(rs));
618 sock_hold(rds_rs_to_sk(rs));
621 goto unlock_and_drop;
622 spin_lock(&rs->rs_lock);
624 if (test_and_clear_bit(RDS_MSG_ON_SOCK, &rm->m_flags)) {
625 struct rm_rdma_op *ro = &rm->rdma;
626 struct rds_notifier *notifier;
628 list_del_init(&rm->m_sock_item);
629 rds_send_sndbuf_remove(rs, rm);
631 if (ro->op_active && ro->op_notifier &&
632 (ro->op_notify || (ro->op_recverr && status))) {
633 notifier = ro->op_notifier;
634 list_add_tail(¬ifier->n_list,
635 &rs->rs_notify_queue);
636 if (!notifier->n_status)
637 notifier->n_status = status;
638 rm->rdma.op_notifier = NULL;
643 spin_unlock(&rs->rs_lock);
646 spin_unlock_irqrestore(&rm->m_rs_lock, flags);
653 rds_wake_sk_sleep(rs);
654 sock_put(rds_rs_to_sk(rs));
659 * Transports call here when they've determined that the receiver queued
660 * messages up to, and including, the given sequence number. Messages are
661 * moved to the retrans queue when rds_send_xmit picks them off the send
662 * queue. This means that in the TCP case, the message may not have been
663 * assigned the m_ack_seq yet - but that's fine as long as tcp_is_acked
664 * checks the RDS_MSG_HAS_ACK_SEQ bit.
666 void rds_send_path_drop_acked(struct rds_conn_path *cp, u64 ack,
667 is_acked_func is_acked)
669 struct rds_message *rm, *tmp;
673 spin_lock_irqsave(&cp->cp_lock, flags);
675 list_for_each_entry_safe(rm, tmp, &cp->cp_retrans, m_conn_item) {
676 if (!rds_send_is_acked(rm, ack, is_acked))
679 list_move(&rm->m_conn_item, &list);
680 clear_bit(RDS_MSG_ON_CONN, &rm->m_flags);
683 /* order flag updates with spin locks */
684 if (!list_empty(&list))
685 smp_mb__after_atomic();
687 spin_unlock_irqrestore(&cp->cp_lock, flags);
689 /* now remove the messages from the sock list as needed */
690 rds_send_remove_from_sock(&list, RDS_RDMA_SUCCESS);
692 EXPORT_SYMBOL_GPL(rds_send_path_drop_acked);
694 void rds_send_drop_acked(struct rds_connection *conn, u64 ack,
695 is_acked_func is_acked)
697 WARN_ON(conn->c_trans->t_mp_capable);
698 rds_send_path_drop_acked(&conn->c_path[0], ack, is_acked);
700 EXPORT_SYMBOL_GPL(rds_send_drop_acked);
702 void rds_send_drop_to(struct rds_sock *rs, struct sockaddr_in *dest)
704 struct rds_message *rm, *tmp;
705 struct rds_connection *conn;
706 struct rds_conn_path *cp;
710 /* get all the messages we're dropping under the rs lock */
711 spin_lock_irqsave(&rs->rs_lock, flags);
713 list_for_each_entry_safe(rm, tmp, &rs->rs_send_queue, m_sock_item) {
714 if (dest && (dest->sin_addr.s_addr != rm->m_daddr ||
715 dest->sin_port != rm->m_inc.i_hdr.h_dport))
718 list_move(&rm->m_sock_item, &list);
719 rds_send_sndbuf_remove(rs, rm);
720 clear_bit(RDS_MSG_ON_SOCK, &rm->m_flags);
723 /* order flag updates with the rs lock */
724 smp_mb__after_atomic();
726 spin_unlock_irqrestore(&rs->rs_lock, flags);
728 if (list_empty(&list))
731 /* Remove the messages from the conn */
732 list_for_each_entry(rm, &list, m_sock_item) {
734 conn = rm->m_inc.i_conn;
735 if (conn->c_trans->t_mp_capable)
736 cp = rm->m_inc.i_conn_path;
738 cp = &conn->c_path[0];
740 spin_lock_irqsave(&cp->cp_lock, flags);
742 * Maybe someone else beat us to removing rm from the conn.
743 * If we race with their flag update we'll get the lock and
744 * then really see that the flag has been cleared.
746 if (!test_and_clear_bit(RDS_MSG_ON_CONN, &rm->m_flags)) {
747 spin_unlock_irqrestore(&cp->cp_lock, flags);
748 spin_lock_irqsave(&rm->m_rs_lock, flags);
750 spin_unlock_irqrestore(&rm->m_rs_lock, flags);
753 list_del_init(&rm->m_conn_item);
754 spin_unlock_irqrestore(&cp->cp_lock, flags);
757 * Couldn't grab m_rs_lock in top loop (lock ordering),
760 spin_lock_irqsave(&rm->m_rs_lock, flags);
762 spin_lock(&rs->rs_lock);
763 __rds_send_complete(rs, rm, RDS_RDMA_CANCELED);
764 spin_unlock(&rs->rs_lock);
767 spin_unlock_irqrestore(&rm->m_rs_lock, flags);
772 rds_wake_sk_sleep(rs);
774 while (!list_empty(&list)) {
775 rm = list_entry(list.next, struct rds_message, m_sock_item);
776 list_del_init(&rm->m_sock_item);
777 rds_message_wait(rm);
779 /* just in case the code above skipped this message
780 * because RDS_MSG_ON_CONN wasn't set, run it again here
781 * taking m_rs_lock is the only thing that keeps us
782 * from racing with ack processing.
784 spin_lock_irqsave(&rm->m_rs_lock, flags);
786 spin_lock(&rs->rs_lock);
787 __rds_send_complete(rs, rm, RDS_RDMA_CANCELED);
788 spin_unlock(&rs->rs_lock);
791 spin_unlock_irqrestore(&rm->m_rs_lock, flags);
798 * we only want this to fire once so we use the callers 'queued'. It's
799 * possible that another thread can race with us and remove the
800 * message from the flow with RDS_CANCEL_SENT_TO.
802 static int rds_send_queue_rm(struct rds_sock *rs, struct rds_connection *conn,
803 struct rds_conn_path *cp,
804 struct rds_message *rm, __be16 sport,
805 __be16 dport, int *queued)
813 len = be32_to_cpu(rm->m_inc.i_hdr.h_len);
815 /* this is the only place which holds both the socket's rs_lock
816 * and the connection's c_lock */
817 spin_lock_irqsave(&rs->rs_lock, flags);
820 * If there is a little space in sndbuf, we don't queue anything,
821 * and userspace gets -EAGAIN. But poll() indicates there's send
822 * room. This can lead to bad behavior (spinning) if snd_bytes isn't
823 * freed up by incoming acks. So we check the *old* value of
824 * rs_snd_bytes here to allow the last msg to exceed the buffer,
825 * and poll() now knows no more data can be sent.
827 if (rs->rs_snd_bytes < rds_sk_sndbuf(rs)) {
828 rs->rs_snd_bytes += len;
830 /* let recv side know we are close to send space exhaustion.
831 * This is probably not the optimal way to do it, as this
832 * means we set the flag on *all* messages as soon as our
833 * throughput hits a certain threshold.
835 if (rs->rs_snd_bytes >= rds_sk_sndbuf(rs) / 2)
836 set_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags);
838 list_add_tail(&rm->m_sock_item, &rs->rs_send_queue);
839 set_bit(RDS_MSG_ON_SOCK, &rm->m_flags);
840 rds_message_addref(rm);
843 /* The code ordering is a little weird, but we're
844 trying to minimize the time we hold c_lock */
845 rds_message_populate_header(&rm->m_inc.i_hdr, sport, dport, 0);
846 rm->m_inc.i_conn = conn;
847 rm->m_inc.i_conn_path = cp;
848 rds_message_addref(rm);
850 spin_lock(&cp->cp_lock);
851 rm->m_inc.i_hdr.h_sequence = cpu_to_be64(cp->cp_next_tx_seq++);
852 list_add_tail(&rm->m_conn_item, &cp->cp_send_queue);
853 set_bit(RDS_MSG_ON_CONN, &rm->m_flags);
854 spin_unlock(&cp->cp_lock);
856 rdsdebug("queued msg %p len %d, rs %p bytes %d seq %llu\n",
857 rm, len, rs, rs->rs_snd_bytes,
858 (unsigned long long)be64_to_cpu(rm->m_inc.i_hdr.h_sequence));
863 spin_unlock_irqrestore(&rs->rs_lock, flags);
869 * rds_message is getting to be quite complicated, and we'd like to allocate
870 * it all in one go. This figures out how big it needs to be up front.
872 static int rds_rm_size(struct msghdr *msg, int data_len)
874 struct cmsghdr *cmsg;
879 for_each_cmsghdr(cmsg, msg) {
880 if (!CMSG_OK(msg, cmsg))
883 if (cmsg->cmsg_level != SOL_RDS)
886 switch (cmsg->cmsg_type) {
887 case RDS_CMSG_RDMA_ARGS:
889 retval = rds_rdma_extra_size(CMSG_DATA(cmsg));
896 case RDS_CMSG_RDMA_DEST:
897 case RDS_CMSG_RDMA_MAP:
899 /* these are valid but do no add any size */
902 case RDS_CMSG_ATOMIC_CSWP:
903 case RDS_CMSG_ATOMIC_FADD:
904 case RDS_CMSG_MASKED_ATOMIC_CSWP:
905 case RDS_CMSG_MASKED_ATOMIC_FADD:
907 size += sizeof(struct scatterlist);
916 size += ceil(data_len, PAGE_SIZE) * sizeof(struct scatterlist);
918 /* Ensure (DEST, MAP) are never used with (ARGS, ATOMIC) */
919 if (cmsg_groups == 3)
925 static int rds_cmsg_send(struct rds_sock *rs, struct rds_message *rm,
926 struct msghdr *msg, int *allocated_mr)
928 struct cmsghdr *cmsg;
931 for_each_cmsghdr(cmsg, msg) {
932 if (!CMSG_OK(msg, cmsg))
935 if (cmsg->cmsg_level != SOL_RDS)
938 /* As a side effect, RDMA_DEST and RDMA_MAP will set
939 * rm->rdma.m_rdma_cookie and rm->rdma.m_rdma_mr.
941 switch (cmsg->cmsg_type) {
942 case RDS_CMSG_RDMA_ARGS:
943 ret = rds_cmsg_rdma_args(rs, rm, cmsg);
946 case RDS_CMSG_RDMA_DEST:
947 ret = rds_cmsg_rdma_dest(rs, rm, cmsg);
950 case RDS_CMSG_RDMA_MAP:
951 ret = rds_cmsg_rdma_map(rs, rm, cmsg);
954 else if (ret == -ENODEV)
955 /* Accommodate the get_mr() case which can fail
956 * if connection isn't established yet.
960 case RDS_CMSG_ATOMIC_CSWP:
961 case RDS_CMSG_ATOMIC_FADD:
962 case RDS_CMSG_MASKED_ATOMIC_CSWP:
963 case RDS_CMSG_MASKED_ATOMIC_FADD:
964 ret = rds_cmsg_atomic(rs, rm, cmsg);
978 static int rds_send_mprds_hash(struct rds_sock *rs, struct rds_connection *conn)
982 if (conn->c_npaths == 0)
983 hash = RDS_MPATH_HASH(rs, RDS_MPATH_WORKERS);
985 hash = RDS_MPATH_HASH(rs, conn->c_npaths);
986 if (conn->c_npaths == 0 && hash != 0) {
987 rds_send_ping(conn, 0);
989 if (conn->c_npaths == 0) {
990 wait_event_interruptible(conn->c_hs_waitq,
991 (conn->c_npaths != 0));
993 if (conn->c_npaths == 1)
999 static int rds_rdma_bytes(struct msghdr *msg, size_t *rdma_bytes)
1001 struct rds_rdma_args *args;
1002 struct cmsghdr *cmsg;
1004 for_each_cmsghdr(cmsg, msg) {
1005 if (!CMSG_OK(msg, cmsg))
1008 if (cmsg->cmsg_level != SOL_RDS)
1011 if (cmsg->cmsg_type == RDS_CMSG_RDMA_ARGS) {
1012 if (cmsg->cmsg_len <
1013 CMSG_LEN(sizeof(struct rds_rdma_args)))
1015 args = CMSG_DATA(cmsg);
1016 *rdma_bytes += args->remote_vec.bytes;
1022 int rds_sendmsg(struct socket *sock, struct msghdr *msg, size_t payload_len)
1024 struct sock *sk = sock->sk;
1025 struct rds_sock *rs = rds_sk_to_rs(sk);
1026 DECLARE_SOCKADDR(struct sockaddr_in *, usin, msg->msg_name);
1029 struct rds_message *rm = NULL;
1030 struct rds_connection *conn;
1032 int queued = 0, allocated_mr = 0;
1033 int nonblock = msg->msg_flags & MSG_DONTWAIT;
1034 long timeo = sock_sndtimeo(sk, nonblock);
1035 struct rds_conn_path *cpath;
1036 size_t total_payload_len = payload_len, rdma_payload_len = 0;
1038 /* Mirror Linux UDP mirror of BSD error message compatibility */
1039 /* XXX: Perhaps MSG_MORE someday */
1040 if (msg->msg_flags & ~(MSG_DONTWAIT | MSG_CMSG_COMPAT)) {
1045 if (msg->msg_namelen) {
1046 /* XXX fail non-unicast destination IPs? */
1047 if (msg->msg_namelen < sizeof(*usin) || usin->sin_family != AF_INET) {
1051 daddr = usin->sin_addr.s_addr;
1052 dport = usin->sin_port;
1054 /* We only care about consistency with ->connect() */
1056 daddr = rs->rs_conn_addr;
1057 dport = rs->rs_conn_port;
1062 if (daddr == 0 || rs->rs_bound_addr == 0) {
1064 ret = -ENOTCONN; /* XXX not a great errno */
1069 ret = rds_rdma_bytes(msg, &rdma_payload_len);
1073 total_payload_len += rdma_payload_len;
1074 if (max_t(size_t, payload_len, rdma_payload_len) > RDS_MAX_MSG_SIZE) {
1079 if (payload_len > rds_sk_sndbuf(rs)) {
1084 /* size of rm including all sgs */
1085 ret = rds_rm_size(msg, payload_len);
1089 rm = rds_message_alloc(ret, GFP_KERNEL);
1095 /* Attach data to the rm */
1097 rm->data.op_sg = rds_message_alloc_sgs(rm, ceil(payload_len, PAGE_SIZE));
1098 if (!rm->data.op_sg) {
1102 ret = rds_message_copy_from_user(rm, &msg->msg_iter);
1106 rm->data.op_active = 1;
1108 rm->m_daddr = daddr;
1110 /* rds_conn_create has a spinlock that runs with IRQ off.
1111 * Caching the conn in the socket helps a lot. */
1112 if (rs->rs_conn && rs->rs_conn->c_faddr == daddr)
1115 conn = rds_conn_create_outgoing(sock_net(sock->sk),
1116 rs->rs_bound_addr, daddr,
1118 sock->sk->sk_allocation);
1120 ret = PTR_ERR(conn);
1126 /* Parse any control messages the user may have included. */
1127 ret = rds_cmsg_send(rs, rm, msg, &allocated_mr);
1129 /* Trigger connection so that its ready for the next retry */
1131 rds_conn_connect_if_down(conn);
1135 if (rm->rdma.op_active && !conn->c_trans->xmit_rdma) {
1136 printk_ratelimited(KERN_NOTICE "rdma_op %p conn xmit_rdma %p\n",
1137 &rm->rdma, conn->c_trans->xmit_rdma);
1142 if (rm->atomic.op_active && !conn->c_trans->xmit_atomic) {
1143 printk_ratelimited(KERN_NOTICE "atomic_op %p conn xmit_atomic %p\n",
1144 &rm->atomic, conn->c_trans->xmit_atomic);
1149 if (conn->c_trans->t_mp_capable)
1150 cpath = &conn->c_path[rds_send_mprds_hash(rs, conn)];
1152 cpath = &conn->c_path[0];
1154 rds_conn_path_connect_if_down(cpath);
1156 ret = rds_cong_wait(conn->c_fcong, dport, nonblock, rs);
1158 rs->rs_seen_congestion = 1;
1161 while (!rds_send_queue_rm(rs, conn, cpath, rm, rs->rs_bound_port,
1163 rds_stats_inc(s_send_queue_full);
1170 timeo = wait_event_interruptible_timeout(*sk_sleep(sk),
1171 rds_send_queue_rm(rs, conn, cpath, rm,
1176 rdsdebug("sendmsg woke queued %d timeo %ld\n", queued, timeo);
1177 if (timeo > 0 || timeo == MAX_SCHEDULE_TIMEOUT)
1187 * By now we've committed to the send. We reuse rds_send_worker()
1188 * to retry sends in the rds thread if the transport asks us to.
1190 rds_stats_inc(s_send_queued);
1192 ret = rds_send_xmit(cpath);
1193 if (ret == -ENOMEM || ret == -EAGAIN)
1194 queue_delayed_work(rds_wq, &cpath->cp_send_w, 1);
1196 rds_message_put(rm);
1200 /* If the user included a RDMA_MAP cmsg, we allocated a MR on the fly.
1201 * If the sendmsg goes through, we keep the MR. If it fails with EAGAIN
1202 * or in any other way, we need to destroy the MR again */
1204 rds_rdma_unuse(rs, rds_rdma_cookie_key(rm->m_rdma_cookie), 1);
1207 rds_message_put(rm);
1212 * send out a probe. Can be shared by rds_send_ping,
1213 * rds_send_pong, rds_send_hb.
1214 * rds_send_hb should use h_flags
1215 * RDS_FLAG_HB_PING|RDS_FLAG_ACK_REQUIRED
1217 * RDS_FLAG_HB_PONG|RDS_FLAG_ACK_REQUIRED
1220 rds_send_probe(struct rds_conn_path *cp, __be16 sport,
1221 __be16 dport, u8 h_flags)
1223 struct rds_message *rm;
1224 unsigned long flags;
1227 rm = rds_message_alloc(0, GFP_ATOMIC);
1233 rm->m_daddr = cp->cp_conn->c_faddr;
1234 rm->data.op_active = 1;
1236 rds_conn_path_connect_if_down(cp);
1238 ret = rds_cong_wait(cp->cp_conn->c_fcong, dport, 1, NULL);
1242 spin_lock_irqsave(&cp->cp_lock, flags);
1243 list_add_tail(&rm->m_conn_item, &cp->cp_send_queue);
1244 set_bit(RDS_MSG_ON_CONN, &rm->m_flags);
1245 rds_message_addref(rm);
1246 rm->m_inc.i_conn = cp->cp_conn;
1247 rm->m_inc.i_conn_path = cp;
1249 rds_message_populate_header(&rm->m_inc.i_hdr, sport, dport,
1250 cp->cp_next_tx_seq);
1251 rm->m_inc.i_hdr.h_flags |= h_flags;
1252 cp->cp_next_tx_seq++;
1254 if (RDS_HS_PROBE(be16_to_cpu(sport), be16_to_cpu(dport)) &&
1255 cp->cp_conn->c_trans->t_mp_capable) {
1256 u16 npaths = cpu_to_be16(RDS_MPATH_WORKERS);
1257 u32 my_gen_num = cpu_to_be32(cp->cp_conn->c_my_gen_num);
1259 rds_message_add_extension(&rm->m_inc.i_hdr,
1260 RDS_EXTHDR_NPATHS, &npaths,
1262 rds_message_add_extension(&rm->m_inc.i_hdr,
1267 spin_unlock_irqrestore(&cp->cp_lock, flags);
1269 rds_stats_inc(s_send_queued);
1270 rds_stats_inc(s_send_pong);
1272 /* schedule the send work on rds_wq */
1273 queue_delayed_work(rds_wq, &cp->cp_send_w, 1);
1275 rds_message_put(rm);
1280 rds_message_put(rm);
1285 rds_send_pong(struct rds_conn_path *cp, __be16 dport)
1287 return rds_send_probe(cp, 0, dport, 0);
1291 rds_send_ping(struct rds_connection *conn, int cp_index)
1293 unsigned long flags;
1294 struct rds_conn_path *cp = &conn->c_path[cp_index];
1296 spin_lock_irqsave(&cp->cp_lock, flags);
1297 if (conn->c_ping_triggered) {
1298 spin_unlock_irqrestore(&cp->cp_lock, flags);
1301 conn->c_ping_triggered = 1;
1302 spin_unlock_irqrestore(&cp->cp_lock, flags);
1303 rds_send_probe(cp, cpu_to_be16(RDS_FLAG_PROBE_PORT), 0, 0);
1305 EXPORT_SYMBOL_GPL(rds_send_ping);