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>
44 /* When transmitting messages in rds_send_xmit, we need to emerge from
45 * time to time and briefly release the CPU. Otherwise the softlock watchdog
47 * Also, it seems fairer to not let one busy connection stall all the
50 * send_batch_count is the number of times we'll loop in send_xmit. Setting
51 * it to 0 will restore the old behavior (where we looped until we had
54 static int send_batch_count = 64;
55 module_param(send_batch_count, int, 0444);
56 MODULE_PARM_DESC(send_batch_count, " batch factor when working the send queue");
58 static void rds_send_remove_from_sock(struct list_head *messages, int status);
61 * Reset the send state. Callers must ensure that this doesn't race with
64 void rds_send_reset(struct rds_connection *conn)
66 struct rds_message *rm, *tmp;
69 if (conn->c_xmit_rm) {
71 conn->c_xmit_rm = NULL;
72 /* Tell the user the RDMA op is no longer mapped by the
73 * transport. This isn't entirely true (it's flushed out
74 * independently) but as the connection is down, there's
75 * no ongoing RDMA to/from that memory */
76 rds_message_unmapped(rm);
81 conn->c_xmit_hdr_off = 0;
82 conn->c_xmit_data_off = 0;
83 conn->c_xmit_atomic_sent = 0;
84 conn->c_xmit_rdma_sent = 0;
85 conn->c_xmit_data_sent = 0;
87 conn->c_map_queued = 0;
89 conn->c_unacked_packets = rds_sysctl_max_unacked_packets;
90 conn->c_unacked_bytes = rds_sysctl_max_unacked_bytes;
92 /* Mark messages as retransmissions, and move them to the send q */
93 spin_lock_irqsave(&conn->c_lock, flags);
94 list_for_each_entry_safe(rm, tmp, &conn->c_retrans, m_conn_item) {
95 set_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags);
96 set_bit(RDS_MSG_RETRANSMITTED, &rm->m_flags);
98 list_splice_init(&conn->c_retrans, &conn->c_send_queue);
99 spin_unlock_irqrestore(&conn->c_lock, flags);
102 static int acquire_in_xmit(struct rds_connection *conn)
104 return test_and_set_bit(RDS_IN_XMIT, &conn->c_flags) == 0;
107 static void release_in_xmit(struct rds_connection *conn)
109 clear_bit(RDS_IN_XMIT, &conn->c_flags);
110 smp_mb__after_atomic();
112 * We don't use wait_on_bit()/wake_up_bit() because our waking is in a
113 * hot path and finding waiters is very rare. We don't want to walk
114 * the system-wide hashed waitqueue buckets in the fast path only to
115 * almost never find waiters.
117 if (waitqueue_active(&conn->c_waitq))
118 wake_up_all(&conn->c_waitq);
122 * We're making the conscious trade-off here to only send one message
123 * down the connection at a time.
125 * - tx queueing is a simple fifo list
126 * - reassembly is optional and easily done by transports per conn
127 * - no per flow rx lookup at all, straight to the socket
128 * - less per-frag memory and wire overhead
130 * - queued acks can be delayed behind large messages
132 * - small message latency is higher behind queued large messages
133 * - large message latency isn't starved by intervening small sends
135 int rds_send_xmit(struct rds_connection *conn)
137 struct rds_message *rm;
140 struct scatterlist *sg;
142 LIST_HEAD(to_be_dropped);
144 unsigned long send_gen = 0;
150 * sendmsg calls here after having queued its message on the send
151 * queue. We only have one task feeding the connection at a time. If
152 * another thread is already feeding the queue then we back off. This
153 * avoids blocking the caller and trading per-connection data between
154 * caches per message.
156 if (!acquire_in_xmit(conn)) {
157 rds_stats_inc(s_send_lock_contention);
163 * we record the send generation after doing the xmit acquire.
164 * if someone else manages to jump in and do some work, we'll use
165 * this to avoid a goto restart farther down.
167 * The acquire_in_xmit() check above ensures that only one
168 * caller can increment c_send_gen at any time.
171 send_gen = conn->c_send_gen;
174 * rds_conn_shutdown() sets the conn state and then tests RDS_IN_XMIT,
175 * we do the opposite to avoid races.
177 if (!rds_conn_up(conn)) {
178 release_in_xmit(conn);
183 if (conn->c_trans->xmit_prepare)
184 conn->c_trans->xmit_prepare(conn);
187 * spin trying to push headers and data down the connection until
188 * the connection doesn't make forward progress.
192 rm = conn->c_xmit_rm;
195 * If between sending messages, we can send a pending congestion
198 if (!rm && test_and_clear_bit(0, &conn->c_map_queued)) {
199 rm = rds_cong_update_alloc(conn);
204 rm->data.op_active = 1;
206 conn->c_xmit_rm = rm;
210 * If not already working on one, grab the next message.
212 * c_xmit_rm holds a ref while we're sending this message down
213 * the connction. We can use this ref while holding the
214 * send_sem.. rds_send_reset() is serialized with it.
221 /* we want to process as big a batch as we can, but
222 * we also want to avoid softlockups. If we've been
223 * through a lot of messages, lets back off and see
224 * if anyone else jumps in
226 if (batch_count >= 1024)
229 spin_lock_irqsave(&conn->c_lock, flags);
231 if (!list_empty(&conn->c_send_queue)) {
232 rm = list_entry(conn->c_send_queue.next,
235 rds_message_addref(rm);
238 * Move the message from the send queue to the retransmit
241 list_move_tail(&rm->m_conn_item, &conn->c_retrans);
244 spin_unlock_irqrestore(&conn->c_lock, flags);
249 /* Unfortunately, the way Infiniband deals with
250 * RDMA to a bad MR key is by moving the entire
251 * queue pair to error state. We cold possibly
252 * recover from that, but right now we drop the
254 * Therefore, we never retransmit messages with RDMA ops.
256 if (rm->rdma.op_active &&
257 test_bit(RDS_MSG_RETRANSMITTED, &rm->m_flags)) {
258 spin_lock_irqsave(&conn->c_lock, flags);
259 if (test_and_clear_bit(RDS_MSG_ON_CONN, &rm->m_flags))
260 list_move(&rm->m_conn_item, &to_be_dropped);
261 spin_unlock_irqrestore(&conn->c_lock, flags);
265 /* Require an ACK every once in a while */
266 len = ntohl(rm->m_inc.i_hdr.h_len);
267 if (conn->c_unacked_packets == 0 ||
268 conn->c_unacked_bytes < len) {
269 __set_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags);
271 conn->c_unacked_packets = rds_sysctl_max_unacked_packets;
272 conn->c_unacked_bytes = rds_sysctl_max_unacked_bytes;
273 rds_stats_inc(s_send_ack_required);
275 conn->c_unacked_bytes -= len;
276 conn->c_unacked_packets--;
279 conn->c_xmit_rm = rm;
282 /* The transport either sends the whole rdma or none of it */
283 if (rm->rdma.op_active && !conn->c_xmit_rdma_sent) {
284 rm->m_final_op = &rm->rdma;
285 ret = conn->c_trans->xmit_rdma(conn, &rm->rdma);
288 conn->c_xmit_rdma_sent = 1;
290 /* The transport owns the mapped memory for now.
291 * You can't unmap it while it's on the send queue */
292 set_bit(RDS_MSG_MAPPED, &rm->m_flags);
295 if (rm->atomic.op_active && !conn->c_xmit_atomic_sent) {
296 rm->m_final_op = &rm->atomic;
297 ret = conn->c_trans->xmit_atomic(conn, &rm->atomic);
300 conn->c_xmit_atomic_sent = 1;
302 /* The transport owns the mapped memory for now.
303 * You can't unmap it while it's on the send queue */
304 set_bit(RDS_MSG_MAPPED, &rm->m_flags);
308 * A number of cases require an RDS header to be sent
309 * even if there is no data.
310 * We permit 0-byte sends; rds-ping depends on this.
311 * However, if there are exclusively attached silent ops,
312 * we skip the hdr/data send, to enable silent operation.
314 if (rm->data.op_nents == 0) {
316 int all_ops_are_silent = 1;
318 ops_present = (rm->atomic.op_active || rm->rdma.op_active);
319 if (rm->atomic.op_active && !rm->atomic.op_silent)
320 all_ops_are_silent = 0;
321 if (rm->rdma.op_active && !rm->rdma.op_silent)
322 all_ops_are_silent = 0;
324 if (ops_present && all_ops_are_silent
325 && !rm->m_rdma_cookie)
326 rm->data.op_active = 0;
329 if (rm->data.op_active && !conn->c_xmit_data_sent) {
330 rm->m_final_op = &rm->data;
331 ret = conn->c_trans->xmit(conn, rm,
332 conn->c_xmit_hdr_off,
334 conn->c_xmit_data_off);
338 if (conn->c_xmit_hdr_off < sizeof(struct rds_header)) {
339 tmp = min_t(int, ret,
340 sizeof(struct rds_header) -
341 conn->c_xmit_hdr_off);
342 conn->c_xmit_hdr_off += tmp;
346 sg = &rm->data.op_sg[conn->c_xmit_sg];
348 tmp = min_t(int, ret, sg->length -
349 conn->c_xmit_data_off);
350 conn->c_xmit_data_off += tmp;
352 if (conn->c_xmit_data_off == sg->length) {
353 conn->c_xmit_data_off = 0;
357 conn->c_xmit_sg == rm->data.op_nents);
361 if (conn->c_xmit_hdr_off == sizeof(struct rds_header) &&
362 (conn->c_xmit_sg == rm->data.op_nents))
363 conn->c_xmit_data_sent = 1;
367 * A rm will only take multiple times through this loop
368 * if there is a data op. Thus, if the data is sent (or there was
369 * none), then we're done with the rm.
371 if (!rm->data.op_active || conn->c_xmit_data_sent) {
372 conn->c_xmit_rm = NULL;
374 conn->c_xmit_hdr_off = 0;
375 conn->c_xmit_data_off = 0;
376 conn->c_xmit_rdma_sent = 0;
377 conn->c_xmit_atomic_sent = 0;
378 conn->c_xmit_data_sent = 0;
385 if (conn->c_trans->xmit_complete)
386 conn->c_trans->xmit_complete(conn);
387 release_in_xmit(conn);
389 /* Nuke any messages we decided not to retransmit. */
390 if (!list_empty(&to_be_dropped)) {
391 /* irqs on here, so we can put(), unlike above */
392 list_for_each_entry(rm, &to_be_dropped, m_conn_item)
394 rds_send_remove_from_sock(&to_be_dropped, RDS_RDMA_DROPPED);
398 * Other senders can queue a message after we last test the send queue
399 * but before we clear RDS_IN_XMIT. In that case they'd back off and
400 * not try and send their newly queued message. We need to check the
401 * send queue after having cleared RDS_IN_XMIT so that their message
402 * doesn't get stuck on the send queue.
404 * If the transport cannot continue (i.e ret != 0), then it must
405 * call us when more room is available, such as from the tx
406 * completion handler.
408 * We have an extra generation check here so that if someone manages
409 * to jump in after our release_in_xmit, we'll see that they have done
410 * some work and we will skip our goto
414 if ((test_bit(0, &conn->c_map_queued) ||
415 !list_empty(&conn->c_send_queue)) &&
416 send_gen == conn->c_send_gen) {
417 rds_stats_inc(s_send_lock_queue_raced);
425 static void rds_send_sndbuf_remove(struct rds_sock *rs, struct rds_message *rm)
427 u32 len = be32_to_cpu(rm->m_inc.i_hdr.h_len);
429 assert_spin_locked(&rs->rs_lock);
431 BUG_ON(rs->rs_snd_bytes < len);
432 rs->rs_snd_bytes -= len;
434 if (rs->rs_snd_bytes == 0)
435 rds_stats_inc(s_send_queue_empty);
438 static inline int rds_send_is_acked(struct rds_message *rm, u64 ack,
439 is_acked_func is_acked)
442 return is_acked(rm, ack);
443 return be64_to_cpu(rm->m_inc.i_hdr.h_sequence) <= ack;
447 * This is pretty similar to what happens below in the ACK
448 * handling code - except that we call here as soon as we get
449 * the IB send completion on the RDMA op and the accompanying
452 void rds_rdma_send_complete(struct rds_message *rm, int status)
454 struct rds_sock *rs = NULL;
455 struct rm_rdma_op *ro;
456 struct rds_notifier *notifier;
459 spin_lock_irqsave(&rm->m_rs_lock, flags);
462 if (test_bit(RDS_MSG_ON_SOCK, &rm->m_flags) &&
463 ro->op_active && ro->op_notify && ro->op_notifier) {
464 notifier = ro->op_notifier;
466 sock_hold(rds_rs_to_sk(rs));
468 notifier->n_status = status;
469 spin_lock(&rs->rs_lock);
470 list_add_tail(¬ifier->n_list, &rs->rs_notify_queue);
471 spin_unlock(&rs->rs_lock);
473 ro->op_notifier = NULL;
476 spin_unlock_irqrestore(&rm->m_rs_lock, flags);
479 rds_wake_sk_sleep(rs);
480 sock_put(rds_rs_to_sk(rs));
483 EXPORT_SYMBOL_GPL(rds_rdma_send_complete);
486 * Just like above, except looks at atomic op
488 void rds_atomic_send_complete(struct rds_message *rm, int status)
490 struct rds_sock *rs = NULL;
491 struct rm_atomic_op *ao;
492 struct rds_notifier *notifier;
495 spin_lock_irqsave(&rm->m_rs_lock, flags);
498 if (test_bit(RDS_MSG_ON_SOCK, &rm->m_flags)
499 && ao->op_active && ao->op_notify && ao->op_notifier) {
500 notifier = ao->op_notifier;
502 sock_hold(rds_rs_to_sk(rs));
504 notifier->n_status = status;
505 spin_lock(&rs->rs_lock);
506 list_add_tail(¬ifier->n_list, &rs->rs_notify_queue);
507 spin_unlock(&rs->rs_lock);
509 ao->op_notifier = NULL;
512 spin_unlock_irqrestore(&rm->m_rs_lock, flags);
515 rds_wake_sk_sleep(rs);
516 sock_put(rds_rs_to_sk(rs));
519 EXPORT_SYMBOL_GPL(rds_atomic_send_complete);
522 * This is the same as rds_rdma_send_complete except we
523 * don't do any locking - we have all the ingredients (message,
524 * socket, socket lock) and can just move the notifier.
527 __rds_send_complete(struct rds_sock *rs, struct rds_message *rm, int status)
529 struct rm_rdma_op *ro;
530 struct rm_atomic_op *ao;
533 if (ro->op_active && ro->op_notify && ro->op_notifier) {
534 ro->op_notifier->n_status = status;
535 list_add_tail(&ro->op_notifier->n_list, &rs->rs_notify_queue);
536 ro->op_notifier = NULL;
540 if (ao->op_active && ao->op_notify && ao->op_notifier) {
541 ao->op_notifier->n_status = status;
542 list_add_tail(&ao->op_notifier->n_list, &rs->rs_notify_queue);
543 ao->op_notifier = NULL;
546 /* No need to wake the app - caller does this */
550 * This is called from the IB send completion when we detect
551 * a RDMA operation that failed with remote access error.
552 * So speed is not an issue here.
554 struct rds_message *rds_send_get_message(struct rds_connection *conn,
555 struct rm_rdma_op *op)
557 struct rds_message *rm, *tmp, *found = NULL;
560 spin_lock_irqsave(&conn->c_lock, flags);
562 list_for_each_entry_safe(rm, tmp, &conn->c_retrans, m_conn_item) {
563 if (&rm->rdma == op) {
564 atomic_inc(&rm->m_refcount);
570 list_for_each_entry_safe(rm, tmp, &conn->c_send_queue, m_conn_item) {
571 if (&rm->rdma == op) {
572 atomic_inc(&rm->m_refcount);
579 spin_unlock_irqrestore(&conn->c_lock, flags);
583 EXPORT_SYMBOL_GPL(rds_send_get_message);
586 * This removes messages from the socket's list if they're on it. The list
587 * argument must be private to the caller, we must be able to modify it
588 * without locks. The messages must have a reference held for their
589 * position on the list. This function will drop that reference after
590 * removing the messages from the 'messages' list regardless of if it found
591 * the messages on the socket list or not.
593 static void rds_send_remove_from_sock(struct list_head *messages, int status)
596 struct rds_sock *rs = NULL;
597 struct rds_message *rm;
599 while (!list_empty(messages)) {
602 rm = list_entry(messages->next, struct rds_message,
604 list_del_init(&rm->m_conn_item);
607 * If we see this flag cleared then we're *sure* that someone
608 * else beat us to removing it from the sock. If we race
609 * with their flag update we'll get the lock and then really
610 * see that the flag has been cleared.
612 * The message spinlock makes sure nobody clears rm->m_rs
613 * while we're messing with it. It does not prevent the
614 * message from being removed from the socket, though.
616 spin_lock_irqsave(&rm->m_rs_lock, flags);
617 if (!test_bit(RDS_MSG_ON_SOCK, &rm->m_flags))
618 goto unlock_and_drop;
620 if (rs != rm->m_rs) {
622 rds_wake_sk_sleep(rs);
623 sock_put(rds_rs_to_sk(rs));
627 sock_hold(rds_rs_to_sk(rs));
630 goto unlock_and_drop;
631 spin_lock(&rs->rs_lock);
633 if (test_and_clear_bit(RDS_MSG_ON_SOCK, &rm->m_flags)) {
634 struct rm_rdma_op *ro = &rm->rdma;
635 struct rds_notifier *notifier;
637 list_del_init(&rm->m_sock_item);
638 rds_send_sndbuf_remove(rs, rm);
640 if (ro->op_active && ro->op_notifier &&
641 (ro->op_notify || (ro->op_recverr && status))) {
642 notifier = ro->op_notifier;
643 list_add_tail(¬ifier->n_list,
644 &rs->rs_notify_queue);
645 if (!notifier->n_status)
646 notifier->n_status = status;
647 rm->rdma.op_notifier = NULL;
652 spin_unlock(&rs->rs_lock);
655 spin_unlock_irqrestore(&rm->m_rs_lock, flags);
662 rds_wake_sk_sleep(rs);
663 sock_put(rds_rs_to_sk(rs));
668 * Transports call here when they've determined that the receiver queued
669 * messages up to, and including, the given sequence number. Messages are
670 * moved to the retrans queue when rds_send_xmit picks them off the send
671 * queue. This means that in the TCP case, the message may not have been
672 * assigned the m_ack_seq yet - but that's fine as long as tcp_is_acked
673 * checks the RDS_MSG_HAS_ACK_SEQ bit.
675 void rds_send_drop_acked(struct rds_connection *conn, u64 ack,
676 is_acked_func is_acked)
678 struct rds_message *rm, *tmp;
682 spin_lock_irqsave(&conn->c_lock, flags);
684 list_for_each_entry_safe(rm, tmp, &conn->c_retrans, m_conn_item) {
685 if (!rds_send_is_acked(rm, ack, is_acked))
688 list_move(&rm->m_conn_item, &list);
689 clear_bit(RDS_MSG_ON_CONN, &rm->m_flags);
692 /* order flag updates with spin locks */
693 if (!list_empty(&list))
694 smp_mb__after_atomic();
696 spin_unlock_irqrestore(&conn->c_lock, flags);
698 /* now remove the messages from the sock list as needed */
699 rds_send_remove_from_sock(&list, RDS_RDMA_SUCCESS);
701 EXPORT_SYMBOL_GPL(rds_send_drop_acked);
703 void rds_send_drop_to(struct rds_sock *rs, struct sockaddr_in *dest)
705 struct rds_message *rm, *tmp;
706 struct rds_connection *conn;
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;
736 spin_lock_irqsave(&conn->c_lock, flags);
738 * Maybe someone else beat us to removing rm from the conn.
739 * If we race with their flag update we'll get the lock and
740 * then really see that the flag has been cleared.
742 if (!test_and_clear_bit(RDS_MSG_ON_CONN, &rm->m_flags)) {
743 spin_unlock_irqrestore(&conn->c_lock, flags);
744 spin_lock_irqsave(&rm->m_rs_lock, flags);
746 spin_unlock_irqrestore(&rm->m_rs_lock, flags);
749 list_del_init(&rm->m_conn_item);
750 spin_unlock_irqrestore(&conn->c_lock, flags);
753 * Couldn't grab m_rs_lock in top loop (lock ordering),
756 spin_lock_irqsave(&rm->m_rs_lock, flags);
758 spin_lock(&rs->rs_lock);
759 __rds_send_complete(rs, rm, RDS_RDMA_CANCELED);
760 spin_unlock(&rs->rs_lock);
763 spin_unlock_irqrestore(&rm->m_rs_lock, flags);
768 rds_wake_sk_sleep(rs);
770 while (!list_empty(&list)) {
771 rm = list_entry(list.next, struct rds_message, m_sock_item);
772 list_del_init(&rm->m_sock_item);
774 rds_message_wait(rm);
780 * we only want this to fire once so we use the callers 'queued'. It's
781 * possible that another thread can race with us and remove the
782 * message from the flow with RDS_CANCEL_SENT_TO.
784 static int rds_send_queue_rm(struct rds_sock *rs, struct rds_connection *conn,
785 struct rds_message *rm, __be16 sport,
786 __be16 dport, int *queued)
794 len = be32_to_cpu(rm->m_inc.i_hdr.h_len);
796 /* this is the only place which holds both the socket's rs_lock
797 * and the connection's c_lock */
798 spin_lock_irqsave(&rs->rs_lock, flags);
801 * If there is a little space in sndbuf, we don't queue anything,
802 * and userspace gets -EAGAIN. But poll() indicates there's send
803 * room. This can lead to bad behavior (spinning) if snd_bytes isn't
804 * freed up by incoming acks. So we check the *old* value of
805 * rs_snd_bytes here to allow the last msg to exceed the buffer,
806 * and poll() now knows no more data can be sent.
808 if (rs->rs_snd_bytes < rds_sk_sndbuf(rs)) {
809 rs->rs_snd_bytes += len;
811 /* let recv side know we are close to send space exhaustion.
812 * This is probably not the optimal way to do it, as this
813 * means we set the flag on *all* messages as soon as our
814 * throughput hits a certain threshold.
816 if (rs->rs_snd_bytes >= rds_sk_sndbuf(rs) / 2)
817 __set_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags);
819 list_add_tail(&rm->m_sock_item, &rs->rs_send_queue);
820 set_bit(RDS_MSG_ON_SOCK, &rm->m_flags);
821 rds_message_addref(rm);
824 /* The code ordering is a little weird, but we're
825 trying to minimize the time we hold c_lock */
826 rds_message_populate_header(&rm->m_inc.i_hdr, sport, dport, 0);
827 rm->m_inc.i_conn = conn;
828 rds_message_addref(rm);
830 spin_lock(&conn->c_lock);
831 rm->m_inc.i_hdr.h_sequence = cpu_to_be64(conn->c_next_tx_seq++);
832 list_add_tail(&rm->m_conn_item, &conn->c_send_queue);
833 set_bit(RDS_MSG_ON_CONN, &rm->m_flags);
834 spin_unlock(&conn->c_lock);
836 rdsdebug("queued msg %p len %d, rs %p bytes %d seq %llu\n",
837 rm, len, rs, rs->rs_snd_bytes,
838 (unsigned long long)be64_to_cpu(rm->m_inc.i_hdr.h_sequence));
843 spin_unlock_irqrestore(&rs->rs_lock, flags);
849 * rds_message is getting to be quite complicated, and we'd like to allocate
850 * it all in one go. This figures out how big it needs to be up front.
852 static int rds_rm_size(struct msghdr *msg, int data_len)
854 struct cmsghdr *cmsg;
859 for_each_cmsghdr(cmsg, msg) {
860 if (!CMSG_OK(msg, cmsg))
863 if (cmsg->cmsg_level != SOL_RDS)
866 switch (cmsg->cmsg_type) {
867 case RDS_CMSG_RDMA_ARGS:
869 retval = rds_rdma_extra_size(CMSG_DATA(cmsg));
876 case RDS_CMSG_RDMA_DEST:
877 case RDS_CMSG_RDMA_MAP:
879 /* these are valid but do no add any size */
882 case RDS_CMSG_ATOMIC_CSWP:
883 case RDS_CMSG_ATOMIC_FADD:
884 case RDS_CMSG_MASKED_ATOMIC_CSWP:
885 case RDS_CMSG_MASKED_ATOMIC_FADD:
887 size += sizeof(struct scatterlist);
896 size += ceil(data_len, PAGE_SIZE) * sizeof(struct scatterlist);
898 /* Ensure (DEST, MAP) are never used with (ARGS, ATOMIC) */
899 if (cmsg_groups == 3)
905 static int rds_cmsg_send(struct rds_sock *rs, struct rds_message *rm,
906 struct msghdr *msg, int *allocated_mr)
908 struct cmsghdr *cmsg;
911 for_each_cmsghdr(cmsg, msg) {
912 if (!CMSG_OK(msg, cmsg))
915 if (cmsg->cmsg_level != SOL_RDS)
918 /* As a side effect, RDMA_DEST and RDMA_MAP will set
919 * rm->rdma.m_rdma_cookie and rm->rdma.m_rdma_mr.
921 switch (cmsg->cmsg_type) {
922 case RDS_CMSG_RDMA_ARGS:
923 ret = rds_cmsg_rdma_args(rs, rm, cmsg);
926 case RDS_CMSG_RDMA_DEST:
927 ret = rds_cmsg_rdma_dest(rs, rm, cmsg);
930 case RDS_CMSG_RDMA_MAP:
931 ret = rds_cmsg_rdma_map(rs, rm, cmsg);
935 case RDS_CMSG_ATOMIC_CSWP:
936 case RDS_CMSG_ATOMIC_FADD:
937 case RDS_CMSG_MASKED_ATOMIC_CSWP:
938 case RDS_CMSG_MASKED_ATOMIC_FADD:
939 ret = rds_cmsg_atomic(rs, rm, cmsg);
953 int rds_sendmsg(struct socket *sock, struct msghdr *msg, size_t payload_len)
955 struct sock *sk = sock->sk;
956 struct rds_sock *rs = rds_sk_to_rs(sk);
957 DECLARE_SOCKADDR(struct sockaddr_in *, usin, msg->msg_name);
960 struct rds_message *rm = NULL;
961 struct rds_connection *conn;
963 int queued = 0, allocated_mr = 0;
964 int nonblock = msg->msg_flags & MSG_DONTWAIT;
965 long timeo = sock_sndtimeo(sk, nonblock);
967 /* Mirror Linux UDP mirror of BSD error message compatibility */
968 /* XXX: Perhaps MSG_MORE someday */
969 if (msg->msg_flags & ~(MSG_DONTWAIT | MSG_CMSG_COMPAT)) {
974 if (msg->msg_namelen) {
975 /* XXX fail non-unicast destination IPs? */
976 if (msg->msg_namelen < sizeof(*usin) || usin->sin_family != AF_INET) {
980 daddr = usin->sin_addr.s_addr;
981 dport = usin->sin_port;
983 /* We only care about consistency with ->connect() */
985 daddr = rs->rs_conn_addr;
986 dport = rs->rs_conn_port;
990 /* racing with another thread binding seems ok here */
991 if (daddr == 0 || rs->rs_bound_addr == 0) {
992 ret = -ENOTCONN; /* XXX not a great errno */
996 /* size of rm including all sgs */
997 ret = rds_rm_size(msg, payload_len);
1001 rm = rds_message_alloc(ret, GFP_KERNEL);
1007 /* Attach data to the rm */
1009 rm->data.op_sg = rds_message_alloc_sgs(rm, ceil(payload_len, PAGE_SIZE));
1010 if (!rm->data.op_sg) {
1014 ret = rds_message_copy_from_user(rm, &msg->msg_iter);
1018 rm->data.op_active = 1;
1020 rm->m_daddr = daddr;
1022 /* rds_conn_create has a spinlock that runs with IRQ off.
1023 * Caching the conn in the socket helps a lot. */
1024 if (rs->rs_conn && rs->rs_conn->c_faddr == daddr)
1027 conn = rds_conn_create_outgoing(sock_net(sock->sk),
1028 rs->rs_bound_addr, daddr,
1030 sock->sk->sk_allocation);
1032 ret = PTR_ERR(conn);
1038 /* Parse any control messages the user may have included. */
1039 ret = rds_cmsg_send(rs, rm, msg, &allocated_mr);
1043 if (rm->rdma.op_active && !conn->c_trans->xmit_rdma) {
1044 printk_ratelimited(KERN_NOTICE "rdma_op %p conn xmit_rdma %p\n",
1045 &rm->rdma, conn->c_trans->xmit_rdma);
1050 if (rm->atomic.op_active && !conn->c_trans->xmit_atomic) {
1051 printk_ratelimited(KERN_NOTICE "atomic_op %p conn xmit_atomic %p\n",
1052 &rm->atomic, conn->c_trans->xmit_atomic);
1057 rds_conn_connect_if_down(conn);
1059 ret = rds_cong_wait(conn->c_fcong, dport, nonblock, rs);
1061 rs->rs_seen_congestion = 1;
1065 while (!rds_send_queue_rm(rs, conn, rm, rs->rs_bound_port,
1067 rds_stats_inc(s_send_queue_full);
1068 /* XXX make sure this is reasonable */
1069 if (payload_len > rds_sk_sndbuf(rs)) {
1078 timeo = wait_event_interruptible_timeout(*sk_sleep(sk),
1079 rds_send_queue_rm(rs, conn, rm,
1084 rdsdebug("sendmsg woke queued %d timeo %ld\n", queued, timeo);
1085 if (timeo > 0 || timeo == MAX_SCHEDULE_TIMEOUT)
1095 * By now we've committed to the send. We reuse rds_send_worker()
1096 * to retry sends in the rds thread if the transport asks us to.
1098 rds_stats_inc(s_send_queued);
1100 if (!test_bit(RDS_LL_SEND_FULL, &conn->c_flags))
1101 rds_send_xmit(conn);
1103 rds_message_put(rm);
1107 /* If the user included a RDMA_MAP cmsg, we allocated a MR on the fly.
1108 * If the sendmsg goes through, we keep the MR. If it fails with EAGAIN
1109 * or in any other way, we need to destroy the MR again */
1111 rds_rdma_unuse(rs, rds_rdma_cookie_key(rm->m_rdma_cookie), 1);
1114 rds_message_put(rm);
1119 * Reply to a ping packet.
1122 rds_send_pong(struct rds_connection *conn, __be16 dport)
1124 struct rds_message *rm;
1125 unsigned long flags;
1128 rm = rds_message_alloc(0, GFP_ATOMIC);
1134 rm->m_daddr = conn->c_faddr;
1135 rm->data.op_active = 1;
1137 rds_conn_connect_if_down(conn);
1139 ret = rds_cong_wait(conn->c_fcong, dport, 1, NULL);
1143 spin_lock_irqsave(&conn->c_lock, flags);
1144 list_add_tail(&rm->m_conn_item, &conn->c_send_queue);
1145 set_bit(RDS_MSG_ON_CONN, &rm->m_flags);
1146 rds_message_addref(rm);
1147 rm->m_inc.i_conn = conn;
1149 rds_message_populate_header(&rm->m_inc.i_hdr, 0, dport,
1150 conn->c_next_tx_seq);
1151 conn->c_next_tx_seq++;
1152 spin_unlock_irqrestore(&conn->c_lock, flags);
1154 rds_stats_inc(s_send_queued);
1155 rds_stats_inc(s_send_pong);
1157 if (!test_bit(RDS_LL_SEND_FULL, &conn->c_flags))
1158 queue_delayed_work(rds_wq, &conn->c_send_w, 0);
1160 rds_message_put(rm);
1165 rds_message_put(rm);