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/gfp.h>
37 #include <linux/list.h>
41 /* When transmitting messages in rds_send_xmit, we need to emerge from
42 * time to time and briefly release the CPU. Otherwise the softlock watchdog
44 * Also, it seems fairer to not let one busy connection stall all the
47 * send_batch_count is the number of times we'll loop in send_xmit. Setting
48 * it to 0 will restore the old behavior (where we looped until we had
51 static int send_batch_count = 64;
52 module_param(send_batch_count, int, 0444);
53 MODULE_PARM_DESC(send_batch_count, " batch factor when working the send queue");
56 * Reset the send state. Caller must hold c_send_lock when calling here.
58 void rds_send_reset(struct rds_connection *conn)
60 struct rds_message *rm, *tmp;
63 if (conn->c_xmit_rm) {
64 /* Tell the user the RDMA op is no longer mapped by the
65 * transport. This isn't entirely true (it's flushed out
66 * independently) but as the connection is down, there's
67 * no ongoing RDMA to/from that memory */
68 rds_message_unmapped(conn->c_xmit_rm);
69 rds_message_put(conn->c_xmit_rm);
70 conn->c_xmit_rm = NULL;
73 conn->c_xmit_hdr_off = 0;
74 conn->c_xmit_data_off = 0;
75 conn->c_xmit_atomic_sent = 0;
76 conn->c_xmit_rdma_sent = 0;
77 conn->c_xmit_data_sent = 0;
79 conn->c_map_queued = 0;
81 conn->c_unacked_packets = rds_sysctl_max_unacked_packets;
82 conn->c_unacked_bytes = rds_sysctl_max_unacked_bytes;
84 /* Mark messages as retransmissions, and move them to the send q */
85 spin_lock_irqsave(&conn->c_lock, flags);
86 list_for_each_entry_safe(rm, tmp, &conn->c_retrans, m_conn_item) {
87 set_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags);
88 set_bit(RDS_MSG_RETRANSMITTED, &rm->m_flags);
90 list_splice_init(&conn->c_retrans, &conn->c_send_queue);
91 spin_unlock_irqrestore(&conn->c_lock, flags);
95 * We're making the concious trade-off here to only send one message
96 * down the connection at a time.
98 * - tx queueing is a simple fifo list
99 * - reassembly is optional and easily done by transports per conn
100 * - no per flow rx lookup at all, straight to the socket
101 * - less per-frag memory and wire overhead
103 * - queued acks can be delayed behind large messages
105 * - small message latency is higher behind queued large messages
106 * - large message latency isn't starved by intervening small sends
108 int rds_send_xmit(struct rds_connection *conn)
110 struct rds_message *rm;
113 unsigned int send_quota = send_batch_count;
114 struct scatterlist *sg;
117 LIST_HEAD(to_be_dropped);
120 * sendmsg calls here after having queued its message on the send
121 * queue. We only have one task feeding the connection at a time. If
122 * another thread is already feeding the queue then we back off. This
123 * avoids blocking the caller and trading per-connection data between
124 * caches per message.
126 * The sem holder will issue a retry if they notice that someone queued
127 * a message after they stopped walking the send queue but before they
130 if (!mutex_trylock(&conn->c_send_lock)) {
131 rds_stats_inc(s_send_sem_contention);
136 if (conn->c_trans->xmit_prepare)
137 conn->c_trans->xmit_prepare(conn);
140 * spin trying to push headers and data down the connection until
141 * the connection doesn't make forward progress.
143 while (--send_quota) {
145 rm = conn->c_xmit_rm;
148 * If between sending messages, we can send a pending congestion
151 * Transports either define a special xmit_cong_map function,
152 * or we allocate a cong_map message and treat it just like any
155 if (!rm && test_and_clear_bit(0, &conn->c_map_queued)) {
156 if (conn->c_trans->xmit_cong_map) {
157 unsigned long map_offset = 0;
158 unsigned long map_bytes = sizeof(struct rds_header) +
162 ret = conn->c_trans->xmit_cong_map(conn, conn->c_lcong,
165 /* too far down the rabbithole! */
166 mutex_unlock(&conn->c_send_lock);
167 rds_conn_error(conn, "Cong map xmit failed\n");
175 /* send cong update like a normal rm */
176 rm = rds_cong_update_alloc(conn);
181 rm->data.op_active = 1;
183 conn->c_xmit_rm = rm;
188 * If not already working on one, grab the next message.
190 * c_xmit_rm holds a ref while we're sending this message down
191 * the connction. We can use this ref while holding the
192 * send_sem.. rds_send_reset() is serialized with it.
197 spin_lock_irqsave(&conn->c_lock, flags);
199 if (!list_empty(&conn->c_send_queue)) {
200 rm = list_entry(conn->c_send_queue.next,
203 rds_message_addref(rm);
206 * Move the message from the send queue to the retransmit
209 list_move_tail(&rm->m_conn_item, &conn->c_retrans);
212 spin_unlock_irqrestore(&conn->c_lock, flags);
219 /* Unfortunately, the way Infiniband deals with
220 * RDMA to a bad MR key is by moving the entire
221 * queue pair to error state. We cold possibly
222 * recover from that, but right now we drop the
224 * Therefore, we never retransmit messages with RDMA ops.
226 if (rm->rdma.op_active &&
227 test_bit(RDS_MSG_RETRANSMITTED, &rm->m_flags)) {
228 spin_lock_irqsave(&conn->c_lock, flags);
229 if (test_and_clear_bit(RDS_MSG_ON_CONN, &rm->m_flags))
230 list_move(&rm->m_conn_item, &to_be_dropped);
231 spin_unlock_irqrestore(&conn->c_lock, flags);
236 /* Require an ACK every once in a while */
237 len = ntohl(rm->m_inc.i_hdr.h_len);
238 if (conn->c_unacked_packets == 0 ||
239 conn->c_unacked_bytes < len) {
240 __set_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags);
242 conn->c_unacked_packets = rds_sysctl_max_unacked_packets;
243 conn->c_unacked_bytes = rds_sysctl_max_unacked_bytes;
244 rds_stats_inc(s_send_ack_required);
246 conn->c_unacked_bytes -= len;
247 conn->c_unacked_packets--;
250 conn->c_xmit_rm = rm;
253 if (rm->atomic.op_active && !conn->c_xmit_atomic_sent) {
254 ret = conn->c_trans->xmit_atomic(conn, rm);
257 conn->c_xmit_atomic_sent = 1;
258 /* The transport owns the mapped memory for now.
259 * You can't unmap it while it's on the send queue */
260 set_bit(RDS_MSG_MAPPED, &rm->m_flags);
263 * This is evil, muahaha.
264 * We permit 0-byte sends. (rds-ping depends on this.)
265 * BUT if there is an atomic op and no sent data,
266 * we turn off sending the header, to achieve
268 * But see below; RDMA op might toggle this back on!
270 if (rm->data.op_nents == 0)
271 rm->data.op_active = 0;
274 /* The transport either sends the whole rdma or none of it */
275 if (rm->rdma.op_active && !conn->c_xmit_rdma_sent) {
276 ret = conn->c_trans->xmit_rdma(conn, &rm->rdma);
279 conn->c_xmit_rdma_sent = 1;
281 /* rdmas need data sent, even if just the header */
282 rm->data.op_active = 1;
284 /* The transport owns the mapped memory for now.
285 * You can't unmap it while it's on the send queue */
286 set_bit(RDS_MSG_MAPPED, &rm->m_flags);
289 if (rm->data.op_active && !conn->c_xmit_data_sent) {
290 ret = conn->c_trans->xmit(conn, rm,
291 conn->c_xmit_hdr_off,
293 conn->c_xmit_data_off);
297 if (conn->c_xmit_hdr_off < sizeof(struct rds_header)) {
298 tmp = min_t(int, ret,
299 sizeof(struct rds_header) -
300 conn->c_xmit_hdr_off);
301 conn->c_xmit_hdr_off += tmp;
305 sg = &rm->data.op_sg[conn->c_xmit_sg];
307 tmp = min_t(int, ret, sg->length -
308 conn->c_xmit_data_off);
309 conn->c_xmit_data_off += tmp;
311 if (conn->c_xmit_data_off == sg->length) {
312 conn->c_xmit_data_off = 0;
316 conn->c_xmit_sg == rm->data.op_nents);
320 if (conn->c_xmit_hdr_off == sizeof(struct rds_header) &&
321 (conn->c_xmit_sg == rm->data.op_nents))
322 conn->c_xmit_data_sent = 1;
326 * A rm will only take multiple times through this loop
327 * if there is a data op. Thus, if the data is sent (or there was
328 * none), then we're done with the rm.
330 if (!rm->data.op_active || conn->c_xmit_data_sent) {
331 conn->c_xmit_rm = NULL;
333 conn->c_xmit_hdr_off = 0;
334 conn->c_xmit_data_off = 0;
335 conn->c_xmit_rdma_sent = 0;
336 conn->c_xmit_atomic_sent = 0;
337 conn->c_xmit_data_sent = 0;
343 /* Nuke any messages we decided not to retransmit. */
344 if (!list_empty(&to_be_dropped))
345 rds_send_remove_from_sock(&to_be_dropped, RDS_RDMA_DROPPED);
347 if (conn->c_trans->xmit_complete)
348 conn->c_trans->xmit_complete(conn);
351 * We might be racing with another sender who queued a message but
352 * backed off on noticing that we held the c_send_lock. If we check
353 * for queued messages after dropping the sem then either we'll
354 * see the queued message or the queuer will get the sem. If we
355 * notice the queued message then we trigger an immediate retry.
357 * We need to be careful only to do this when we stopped processing
358 * the send queue because it was empty. It's the only way we
359 * stop processing the loop when the transport hasn't taken
360 * responsibility for forward progress.
362 mutex_unlock(&conn->c_send_lock);
364 if (send_quota == 0 && !was_empty) {
365 /* We exhausted the send quota, but there's work left to
366 * do. Return and (re-)schedule the send worker.
371 if (ret == 0 && was_empty) {
372 /* A simple bit test would be way faster than taking the
374 spin_lock_irqsave(&conn->c_lock, flags);
375 if (!list_empty(&conn->c_send_queue)) {
376 rds_stats_inc(s_send_sem_queue_raced);
379 spin_unlock_irqrestore(&conn->c_lock, flags);
385 static void rds_send_sndbuf_remove(struct rds_sock *rs, struct rds_message *rm)
387 u32 len = be32_to_cpu(rm->m_inc.i_hdr.h_len);
389 assert_spin_locked(&rs->rs_lock);
391 BUG_ON(rs->rs_snd_bytes < len);
392 rs->rs_snd_bytes -= len;
394 if (rs->rs_snd_bytes == 0)
395 rds_stats_inc(s_send_queue_empty);
398 static inline int rds_send_is_acked(struct rds_message *rm, u64 ack,
399 is_acked_func is_acked)
402 return is_acked(rm, ack);
403 return be64_to_cpu(rm->m_inc.i_hdr.h_sequence) <= ack;
407 * Returns true if there are no messages on the send and retransmit queues
408 * which have a sequence number greater than or equal to the given sequence
411 int rds_send_acked_before(struct rds_connection *conn, u64 seq)
413 struct rds_message *rm, *tmp;
416 spin_lock(&conn->c_lock);
418 list_for_each_entry_safe(rm, tmp, &conn->c_retrans, m_conn_item) {
419 if (be64_to_cpu(rm->m_inc.i_hdr.h_sequence) < seq)
424 list_for_each_entry_safe(rm, tmp, &conn->c_send_queue, m_conn_item) {
425 if (be64_to_cpu(rm->m_inc.i_hdr.h_sequence) < seq)
430 spin_unlock(&conn->c_lock);
436 * This is pretty similar to what happens below in the ACK
437 * handling code - except that we call here as soon as we get
438 * the IB send completion on the RDMA op and the accompanying
441 void rds_rdma_send_complete(struct rds_message *rm, int status)
443 struct rds_sock *rs = NULL;
444 struct rm_rdma_op *ro;
445 struct rds_notifier *notifier;
448 spin_lock_irqsave(&rm->m_rs_lock, flags);
451 if (test_bit(RDS_MSG_ON_SOCK, &rm->m_flags) &&
452 ro->op_active && ro->op_notify && ro->op_notifier) {
453 notifier = ro->op_notifier;
455 sock_hold(rds_rs_to_sk(rs));
457 notifier->n_status = status;
458 spin_lock(&rs->rs_lock);
459 list_add_tail(¬ifier->n_list, &rs->rs_notify_queue);
460 spin_unlock(&rs->rs_lock);
462 ro->op_notifier = NULL;
465 spin_unlock_irqrestore(&rm->m_rs_lock, flags);
468 rds_wake_sk_sleep(rs);
469 sock_put(rds_rs_to_sk(rs));
472 EXPORT_SYMBOL_GPL(rds_rdma_send_complete);
475 * Just like above, except looks at atomic op
477 void rds_atomic_send_complete(struct rds_message *rm, int status)
479 struct rds_sock *rs = NULL;
480 struct rm_atomic_op *ao;
481 struct rds_notifier *notifier;
483 spin_lock(&rm->m_rs_lock);
486 if (test_bit(RDS_MSG_ON_SOCK, &rm->m_flags)
487 && ao->op_active && ao->op_notify && ao->op_notifier) {
488 notifier = ao->op_notifier;
490 sock_hold(rds_rs_to_sk(rs));
492 notifier->n_status = status;
493 spin_lock(&rs->rs_lock);
494 list_add_tail(¬ifier->n_list, &rs->rs_notify_queue);
495 spin_unlock(&rs->rs_lock);
497 ao->op_notifier = NULL;
500 spin_unlock(&rm->m_rs_lock);
503 rds_wake_sk_sleep(rs);
504 sock_put(rds_rs_to_sk(rs));
507 EXPORT_SYMBOL_GPL(rds_atomic_send_complete);
510 * This is the same as rds_rdma_send_complete except we
511 * don't do any locking - we have all the ingredients (message,
512 * socket, socket lock) and can just move the notifier.
515 __rds_rdma_send_complete(struct rds_sock *rs, struct rds_message *rm, int status)
517 struct rm_rdma_op *ro;
520 if (ro->op_active && ro->op_notify && ro->op_notifier) {
521 ro->op_notifier->n_status = status;
522 list_add_tail(&ro->op_notifier->n_list, &rs->rs_notify_queue);
523 ro->op_notifier = NULL;
526 /* No need to wake the app - caller does this */
530 * This is called from the IB send completion when we detect
531 * a RDMA operation that failed with remote access error.
532 * So speed is not an issue here.
534 struct rds_message *rds_send_get_message(struct rds_connection *conn,
535 struct rm_rdma_op *op)
537 struct rds_message *rm, *tmp, *found = NULL;
540 spin_lock_irqsave(&conn->c_lock, flags);
542 list_for_each_entry_safe(rm, tmp, &conn->c_retrans, m_conn_item) {
543 if (&rm->rdma == op) {
544 atomic_inc(&rm->m_refcount);
550 list_for_each_entry_safe(rm, tmp, &conn->c_send_queue, m_conn_item) {
551 if (&rm->rdma == op) {
552 atomic_inc(&rm->m_refcount);
559 spin_unlock_irqrestore(&conn->c_lock, flags);
563 EXPORT_SYMBOL_GPL(rds_send_get_message);
566 * This removes messages from the socket's list if they're on it. The list
567 * argument must be private to the caller, we must be able to modify it
568 * without locks. The messages must have a reference held for their
569 * position on the list. This function will drop that reference after
570 * removing the messages from the 'messages' list regardless of if it found
571 * the messages on the socket list or not.
573 void rds_send_remove_from_sock(struct list_head *messages, int status)
576 struct rds_sock *rs = NULL;
577 struct rds_message *rm;
579 while (!list_empty(messages)) {
582 rm = list_entry(messages->next, struct rds_message,
584 list_del_init(&rm->m_conn_item);
587 * If we see this flag cleared then we're *sure* that someone
588 * else beat us to removing it from the sock. If we race
589 * with their flag update we'll get the lock and then really
590 * see that the flag has been cleared.
592 * The message spinlock makes sure nobody clears rm->m_rs
593 * while we're messing with it. It does not prevent the
594 * message from being removed from the socket, though.
596 spin_lock_irqsave(&rm->m_rs_lock, flags);
597 if (!test_bit(RDS_MSG_ON_SOCK, &rm->m_flags))
598 goto unlock_and_drop;
600 if (rs != rm->m_rs) {
602 rds_wake_sk_sleep(rs);
603 sock_put(rds_rs_to_sk(rs));
606 sock_hold(rds_rs_to_sk(rs));
608 spin_lock(&rs->rs_lock);
610 if (test_and_clear_bit(RDS_MSG_ON_SOCK, &rm->m_flags)) {
611 struct rm_rdma_op *ro = &rm->rdma;
612 struct rds_notifier *notifier;
614 list_del_init(&rm->m_sock_item);
615 rds_send_sndbuf_remove(rs, rm);
617 if (ro->op_active && ro->op_notifier &&
618 (ro->op_notify || (ro->op_recverr && status))) {
619 notifier = ro->op_notifier;
620 list_add_tail(¬ifier->n_list,
621 &rs->rs_notify_queue);
622 if (!notifier->n_status)
623 notifier->n_status = status;
624 rm->rdma.op_notifier = NULL;
629 spin_unlock(&rs->rs_lock);
632 spin_unlock_irqrestore(&rm->m_rs_lock, flags);
639 rds_wake_sk_sleep(rs);
640 sock_put(rds_rs_to_sk(rs));
645 * Transports call here when they've determined that the receiver queued
646 * messages up to, and including, the given sequence number. Messages are
647 * moved to the retrans queue when rds_send_xmit picks them off the send
648 * queue. This means that in the TCP case, the message may not have been
649 * assigned the m_ack_seq yet - but that's fine as long as tcp_is_acked
650 * checks the RDS_MSG_HAS_ACK_SEQ bit.
652 * XXX It's not clear to me how this is safely serialized with socket
653 * destruction. Maybe it should bail if it sees SOCK_DEAD.
655 void rds_send_drop_acked(struct rds_connection *conn, u64 ack,
656 is_acked_func is_acked)
658 struct rds_message *rm, *tmp;
662 spin_lock_irqsave(&conn->c_lock, flags);
664 list_for_each_entry_safe(rm, tmp, &conn->c_retrans, m_conn_item) {
665 if (!rds_send_is_acked(rm, ack, is_acked))
668 list_move(&rm->m_conn_item, &list);
669 clear_bit(RDS_MSG_ON_CONN, &rm->m_flags);
672 /* order flag updates with spin locks */
673 if (!list_empty(&list))
674 smp_mb__after_clear_bit();
676 spin_unlock_irqrestore(&conn->c_lock, flags);
678 /* now remove the messages from the sock list as needed */
679 rds_send_remove_from_sock(&list, RDS_RDMA_SUCCESS);
681 EXPORT_SYMBOL_GPL(rds_send_drop_acked);
683 void rds_send_drop_to(struct rds_sock *rs, struct sockaddr_in *dest)
685 struct rds_message *rm, *tmp;
686 struct rds_connection *conn;
690 /* get all the messages we're dropping under the rs lock */
691 spin_lock_irqsave(&rs->rs_lock, flags);
693 list_for_each_entry_safe(rm, tmp, &rs->rs_send_queue, m_sock_item) {
694 if (dest && (dest->sin_addr.s_addr != rm->m_daddr ||
695 dest->sin_port != rm->m_inc.i_hdr.h_dport))
698 list_move(&rm->m_sock_item, &list);
699 rds_send_sndbuf_remove(rs, rm);
700 clear_bit(RDS_MSG_ON_SOCK, &rm->m_flags);
703 /* order flag updates with the rs lock */
704 smp_mb__after_clear_bit();
706 spin_unlock_irqrestore(&rs->rs_lock, flags);
708 if (list_empty(&list))
711 /* Remove the messages from the conn */
712 list_for_each_entry(rm, &list, m_sock_item) {
714 conn = rm->m_inc.i_conn;
716 spin_lock_irqsave(&conn->c_lock, flags);
718 * Maybe someone else beat us to removing rm from the conn.
719 * If we race with their flag update we'll get the lock and
720 * then really see that the flag has been cleared.
722 if (!test_and_clear_bit(RDS_MSG_ON_CONN, &rm->m_flags)) {
723 spin_unlock_irqrestore(&conn->c_lock, flags);
726 list_del_init(&rm->m_conn_item);
727 spin_unlock_irqrestore(&conn->c_lock, flags);
730 * Couldn't grab m_rs_lock in top loop (lock ordering),
733 spin_lock_irqsave(&rm->m_rs_lock, flags);
735 spin_lock(&rs->rs_lock);
736 __rds_rdma_send_complete(rs, rm, RDS_RDMA_CANCELED);
737 spin_unlock(&rs->rs_lock);
740 spin_unlock_irqrestore(&rm->m_rs_lock, flags);
745 rds_wake_sk_sleep(rs);
747 while (!list_empty(&list)) {
748 rm = list_entry(list.next, struct rds_message, m_sock_item);
749 list_del_init(&rm->m_sock_item);
751 rds_message_wait(rm);
757 * we only want this to fire once so we use the callers 'queued'. It's
758 * possible that another thread can race with us and remove the
759 * message from the flow with RDS_CANCEL_SENT_TO.
761 static int rds_send_queue_rm(struct rds_sock *rs, struct rds_connection *conn,
762 struct rds_message *rm, __be16 sport,
763 __be16 dport, int *queued)
771 len = be32_to_cpu(rm->m_inc.i_hdr.h_len);
773 /* this is the only place which holds both the socket's rs_lock
774 * and the connection's c_lock */
775 spin_lock_irqsave(&rs->rs_lock, flags);
778 * If there is a little space in sndbuf, we don't queue anything,
779 * and userspace gets -EAGAIN. But poll() indicates there's send
780 * room. This can lead to bad behavior (spinning) if snd_bytes isn't
781 * freed up by incoming acks. So we check the *old* value of
782 * rs_snd_bytes here to allow the last msg to exceed the buffer,
783 * and poll() now knows no more data can be sent.
785 if (rs->rs_snd_bytes < rds_sk_sndbuf(rs)) {
786 rs->rs_snd_bytes += len;
788 /* let recv side know we are close to send space exhaustion.
789 * This is probably not the optimal way to do it, as this
790 * means we set the flag on *all* messages as soon as our
791 * throughput hits a certain threshold.
793 if (rs->rs_snd_bytes >= rds_sk_sndbuf(rs) / 2)
794 __set_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags);
796 list_add_tail(&rm->m_sock_item, &rs->rs_send_queue);
797 set_bit(RDS_MSG_ON_SOCK, &rm->m_flags);
798 rds_message_addref(rm);
801 /* The code ordering is a little weird, but we're
802 trying to minimize the time we hold c_lock */
803 rds_message_populate_header(&rm->m_inc.i_hdr, sport, dport, 0);
804 rm->m_inc.i_conn = conn;
805 rds_message_addref(rm);
807 spin_lock(&conn->c_lock);
808 rm->m_inc.i_hdr.h_sequence = cpu_to_be64(conn->c_next_tx_seq++);
809 list_add_tail(&rm->m_conn_item, &conn->c_send_queue);
810 set_bit(RDS_MSG_ON_CONN, &rm->m_flags);
811 spin_unlock(&conn->c_lock);
813 rdsdebug("queued msg %p len %d, rs %p bytes %d seq %llu\n",
814 rm, len, rs, rs->rs_snd_bytes,
815 (unsigned long long)be64_to_cpu(rm->m_inc.i_hdr.h_sequence));
820 spin_unlock_irqrestore(&rs->rs_lock, flags);
826 * rds_message is getting to be quite complicated, and we'd like to allocate
827 * it all in one go. This figures out how big it needs to be up front.
829 static int rds_rm_size(struct msghdr *msg, int data_len)
831 struct cmsghdr *cmsg;
835 for (cmsg = CMSG_FIRSTHDR(msg); cmsg; cmsg = CMSG_NXTHDR(msg, cmsg)) {
836 if (!CMSG_OK(msg, cmsg))
839 if (cmsg->cmsg_level != SOL_RDS)
842 switch (cmsg->cmsg_type) {
843 case RDS_CMSG_RDMA_ARGS:
844 retval = rds_rdma_extra_size(CMSG_DATA(cmsg));
850 case RDS_CMSG_RDMA_DEST:
851 case RDS_CMSG_RDMA_MAP:
852 /* these are valid but do no add any size */
855 case RDS_CMSG_ATOMIC_CSWP:
856 case RDS_CMSG_ATOMIC_FADD:
857 size += sizeof(struct scatterlist);
866 size += ceil(data_len, PAGE_SIZE) * sizeof(struct scatterlist);
871 static int rds_cmsg_send(struct rds_sock *rs, struct rds_message *rm,
872 struct msghdr *msg, int *allocated_mr)
874 struct cmsghdr *cmsg;
877 for (cmsg = CMSG_FIRSTHDR(msg); cmsg; cmsg = CMSG_NXTHDR(msg, cmsg)) {
878 if (!CMSG_OK(msg, cmsg))
881 if (cmsg->cmsg_level != SOL_RDS)
884 /* As a side effect, RDMA_DEST and RDMA_MAP will set
885 * rm->rdma.m_rdma_cookie and rm->rdma.m_rdma_mr.
887 switch (cmsg->cmsg_type) {
888 case RDS_CMSG_RDMA_ARGS:
889 ret = rds_cmsg_rdma_args(rs, rm, cmsg);
892 case RDS_CMSG_RDMA_DEST:
893 ret = rds_cmsg_rdma_dest(rs, rm, cmsg);
896 case RDS_CMSG_RDMA_MAP:
897 ret = rds_cmsg_rdma_map(rs, rm, cmsg);
901 case RDS_CMSG_ATOMIC_CSWP:
902 case RDS_CMSG_ATOMIC_FADD:
903 ret = rds_cmsg_atomic(rs, rm, cmsg);
917 int rds_sendmsg(struct kiocb *iocb, struct socket *sock, struct msghdr *msg,
920 struct sock *sk = sock->sk;
921 struct rds_sock *rs = rds_sk_to_rs(sk);
922 struct sockaddr_in *usin = (struct sockaddr_in *)msg->msg_name;
925 struct rds_message *rm = NULL;
926 struct rds_connection *conn;
928 int queued = 0, allocated_mr = 0;
929 int nonblock = msg->msg_flags & MSG_DONTWAIT;
930 long timeo = sock_sndtimeo(sk, nonblock);
932 /* Mirror Linux UDP mirror of BSD error message compatibility */
933 /* XXX: Perhaps MSG_MORE someday */
934 if (msg->msg_flags & ~(MSG_DONTWAIT | MSG_CMSG_COMPAT)) {
935 printk(KERN_INFO "msg_flags 0x%08X\n", msg->msg_flags);
940 if (msg->msg_namelen) {
941 /* XXX fail non-unicast destination IPs? */
942 if (msg->msg_namelen < sizeof(*usin) || usin->sin_family != AF_INET) {
946 daddr = usin->sin_addr.s_addr;
947 dport = usin->sin_port;
949 /* We only care about consistency with ->connect() */
951 daddr = rs->rs_conn_addr;
952 dport = rs->rs_conn_port;
956 /* racing with another thread binding seems ok here */
957 if (daddr == 0 || rs->rs_bound_addr == 0) {
958 ret = -ENOTCONN; /* XXX not a great errno */
962 /* size of rm including all sgs */
963 ret = rds_rm_size(msg, payload_len);
967 rm = rds_message_alloc(ret, GFP_KERNEL);
973 /* Attach data to the rm */
975 rm->data.op_sg = rds_message_alloc_sgs(rm, ceil(payload_len, PAGE_SIZE));
976 ret = rds_message_copy_from_user(rm, msg->msg_iov, payload_len);
980 rm->data.op_active = 1;
984 /* rds_conn_create has a spinlock that runs with IRQ off.
985 * Caching the conn in the socket helps a lot. */
986 if (rs->rs_conn && rs->rs_conn->c_faddr == daddr)
989 conn = rds_conn_create_outgoing(rs->rs_bound_addr, daddr,
991 sock->sk->sk_allocation);
999 /* Parse any control messages the user may have included. */
1000 ret = rds_cmsg_send(rs, rm, msg, &allocated_mr);
1004 if ((rm->m_rdma_cookie || rm->rdma.op_active) &&
1005 !conn->c_trans->xmit_rdma) {
1006 if (printk_ratelimit())
1007 printk(KERN_NOTICE "rdma_op %p conn xmit_rdma %p\n",
1008 &rm->rdma, conn->c_trans->xmit_rdma);
1013 if (rm->atomic.op_active && !conn->c_trans->xmit_atomic) {
1014 if (printk_ratelimit())
1015 printk(KERN_NOTICE "atomic_op %p conn xmit_atomic %p\n",
1016 &rm->atomic, conn->c_trans->xmit_atomic);
1021 /* If the connection is down, trigger a connect. We may
1022 * have scheduled a delayed reconnect however - in this case
1023 * we should not interfere.
1025 if (rds_conn_state(conn) == RDS_CONN_DOWN &&
1026 !test_and_set_bit(RDS_RECONNECT_PENDING, &conn->c_flags))
1027 queue_delayed_work(rds_wq, &conn->c_conn_w, 0);
1029 ret = rds_cong_wait(conn->c_fcong, dport, nonblock, rs);
1031 rs->rs_seen_congestion = 1;
1035 while (!rds_send_queue_rm(rs, conn, rm, rs->rs_bound_port,
1037 rds_stats_inc(s_send_queue_full);
1038 /* XXX make sure this is reasonable */
1039 if (payload_len > rds_sk_sndbuf(rs)) {
1048 timeo = wait_event_interruptible_timeout(*sk_sleep(sk),
1049 rds_send_queue_rm(rs, conn, rm,
1054 rdsdebug("sendmsg woke queued %d timeo %ld\n", queued, timeo);
1055 if (timeo > 0 || timeo == MAX_SCHEDULE_TIMEOUT)
1065 * By now we've committed to the send. We reuse rds_send_worker()
1066 * to retry sends in the rds thread if the transport asks us to.
1068 rds_stats_inc(s_send_queued);
1070 if (!test_bit(RDS_LL_SEND_FULL, &conn->c_flags))
1071 rds_send_worker(&conn->c_send_w.work);
1073 rds_message_put(rm);
1077 /* If the user included a RDMA_MAP cmsg, we allocated a MR on the fly.
1078 * If the sendmsg goes through, we keep the MR. If it fails with EAGAIN
1079 * or in any other way, we need to destroy the MR again */
1081 rds_rdma_unuse(rs, rds_rdma_cookie_key(rm->m_rdma_cookie), 1);
1084 rds_message_put(rm);
1089 * Reply to a ping packet.
1092 rds_send_pong(struct rds_connection *conn, __be16 dport)
1094 struct rds_message *rm;
1095 unsigned long flags;
1098 rm = rds_message_alloc(0, GFP_ATOMIC);
1104 rm->m_daddr = conn->c_faddr;
1106 /* If the connection is down, trigger a connect. We may
1107 * have scheduled a delayed reconnect however - in this case
1108 * we should not interfere.
1110 if (rds_conn_state(conn) == RDS_CONN_DOWN &&
1111 !test_and_set_bit(RDS_RECONNECT_PENDING, &conn->c_flags))
1112 queue_delayed_work(rds_wq, &conn->c_conn_w, 0);
1114 ret = rds_cong_wait(conn->c_fcong, dport, 1, NULL);
1118 spin_lock_irqsave(&conn->c_lock, flags);
1119 list_add_tail(&rm->m_conn_item, &conn->c_send_queue);
1120 set_bit(RDS_MSG_ON_CONN, &rm->m_flags);
1121 rds_message_addref(rm);
1122 rm->m_inc.i_conn = conn;
1124 rds_message_populate_header(&rm->m_inc.i_hdr, 0, dport,
1125 conn->c_next_tx_seq);
1126 conn->c_next_tx_seq++;
1127 spin_unlock_irqrestore(&conn->c_lock, flags);
1129 rds_stats_inc(s_send_queued);
1130 rds_stats_inc(s_send_pong);
1132 queue_delayed_work(rds_wq, &conn->c_send_w, 0);
1133 rds_message_put(rm);
1138 rds_message_put(rm);