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 /* The transport either sends the whole rdma or none of it */
254 if (rm->rdma.op_active && !conn->c_xmit_rdma_sent) {
255 rm->m_final_op = &rm->rdma;
256 ret = conn->c_trans->xmit_rdma(conn, &rm->rdma);
259 conn->c_xmit_rdma_sent = 1;
261 /* The transport owns the mapped memory for now.
262 * You can't unmap it while it's on the send queue */
263 set_bit(RDS_MSG_MAPPED, &rm->m_flags);
266 if (rm->atomic.op_active && !conn->c_xmit_atomic_sent) {
267 rm->m_final_op = &rm->atomic;
268 ret = conn->c_trans->xmit_atomic(conn, &rm->atomic);
271 conn->c_xmit_atomic_sent = 1;
273 /* The transport owns the mapped memory for now.
274 * You can't unmap it while it's on the send queue */
275 set_bit(RDS_MSG_MAPPED, &rm->m_flags);
279 * A number of cases require an RDS header to be sent
280 * even if there is no data.
281 * We permit 0-byte sends; rds-ping depends on this.
282 * However, if there are exclusively attached silent ops,
283 * we skip the hdr/data send, to enable silent operation.
285 if (rm->data.op_nents == 0) {
287 int all_ops_are_silent = 1;
289 ops_present = (rm->atomic.op_active || rm->rdma.op_active);
290 if (rm->atomic.op_active && !rm->atomic.op_silent)
291 all_ops_are_silent = 0;
292 if (rm->rdma.op_active && !rm->rdma.op_silent)
293 all_ops_are_silent = 0;
295 if (ops_present && all_ops_are_silent
296 && !rm->m_rdma_cookie)
297 rm->data.op_active = 0;
300 if (rm->data.op_active && !conn->c_xmit_data_sent) {
301 rm->m_final_op = &rm->data;
302 ret = conn->c_trans->xmit(conn, rm,
303 conn->c_xmit_hdr_off,
305 conn->c_xmit_data_off);
309 if (conn->c_xmit_hdr_off < sizeof(struct rds_header)) {
310 tmp = min_t(int, ret,
311 sizeof(struct rds_header) -
312 conn->c_xmit_hdr_off);
313 conn->c_xmit_hdr_off += tmp;
317 sg = &rm->data.op_sg[conn->c_xmit_sg];
319 tmp = min_t(int, ret, sg->length -
320 conn->c_xmit_data_off);
321 conn->c_xmit_data_off += tmp;
323 if (conn->c_xmit_data_off == sg->length) {
324 conn->c_xmit_data_off = 0;
328 conn->c_xmit_sg == rm->data.op_nents);
332 if (conn->c_xmit_hdr_off == sizeof(struct rds_header) &&
333 (conn->c_xmit_sg == rm->data.op_nents))
334 conn->c_xmit_data_sent = 1;
338 * A rm will only take multiple times through this loop
339 * if there is a data op. Thus, if the data is sent (or there was
340 * none), then we're done with the rm.
342 if (!rm->data.op_active || conn->c_xmit_data_sent) {
343 conn->c_xmit_rm = NULL;
345 conn->c_xmit_hdr_off = 0;
346 conn->c_xmit_data_off = 0;
347 conn->c_xmit_rdma_sent = 0;
348 conn->c_xmit_atomic_sent = 0;
349 conn->c_xmit_data_sent = 0;
355 /* Nuke any messages we decided not to retransmit. */
356 if (!list_empty(&to_be_dropped))
357 rds_send_remove_from_sock(&to_be_dropped, RDS_RDMA_DROPPED);
359 if (conn->c_trans->xmit_complete)
360 conn->c_trans->xmit_complete(conn);
363 * We might be racing with another sender who queued a message but
364 * backed off on noticing that we held the c_send_lock. If we check
365 * for queued messages after dropping the sem then either we'll
366 * see the queued message or the queuer will get the sem. If we
367 * notice the queued message then we trigger an immediate retry.
369 * We need to be careful only to do this when we stopped processing
370 * the send queue because it was empty. It's the only way we
371 * stop processing the loop when the transport hasn't taken
372 * responsibility for forward progress.
374 mutex_unlock(&conn->c_send_lock);
376 if (send_quota == 0 && !was_empty) {
377 /* We exhausted the send quota, but there's work left to
378 * do. Return and (re-)schedule the send worker.
383 if (ret == 0 && was_empty) {
384 /* A simple bit test would be way faster than taking the
386 spin_lock_irqsave(&conn->c_lock, flags);
387 if (!list_empty(&conn->c_send_queue)) {
388 rds_stats_inc(s_send_sem_queue_raced);
391 spin_unlock_irqrestore(&conn->c_lock, flags);
397 static void rds_send_sndbuf_remove(struct rds_sock *rs, struct rds_message *rm)
399 u32 len = be32_to_cpu(rm->m_inc.i_hdr.h_len);
401 assert_spin_locked(&rs->rs_lock);
403 BUG_ON(rs->rs_snd_bytes < len);
404 rs->rs_snd_bytes -= len;
406 if (rs->rs_snd_bytes == 0)
407 rds_stats_inc(s_send_queue_empty);
410 static inline int rds_send_is_acked(struct rds_message *rm, u64 ack,
411 is_acked_func is_acked)
414 return is_acked(rm, ack);
415 return be64_to_cpu(rm->m_inc.i_hdr.h_sequence) <= ack;
419 * Returns true if there are no messages on the send and retransmit queues
420 * which have a sequence number greater than or equal to the given sequence
423 int rds_send_acked_before(struct rds_connection *conn, u64 seq)
425 struct rds_message *rm, *tmp;
428 spin_lock(&conn->c_lock);
430 list_for_each_entry_safe(rm, tmp, &conn->c_retrans, m_conn_item) {
431 if (be64_to_cpu(rm->m_inc.i_hdr.h_sequence) < seq)
436 list_for_each_entry_safe(rm, tmp, &conn->c_send_queue, m_conn_item) {
437 if (be64_to_cpu(rm->m_inc.i_hdr.h_sequence) < seq)
442 spin_unlock(&conn->c_lock);
448 * This is pretty similar to what happens below in the ACK
449 * handling code - except that we call here as soon as we get
450 * the IB send completion on the RDMA op and the accompanying
453 void rds_rdma_send_complete(struct rds_message *rm, int status)
455 struct rds_sock *rs = NULL;
456 struct rm_rdma_op *ro;
457 struct rds_notifier *notifier;
460 spin_lock_irqsave(&rm->m_rs_lock, flags);
463 if (test_bit(RDS_MSG_ON_SOCK, &rm->m_flags) &&
464 ro->op_active && ro->op_notify && ro->op_notifier) {
465 notifier = ro->op_notifier;
467 sock_hold(rds_rs_to_sk(rs));
469 notifier->n_status = status;
470 spin_lock(&rs->rs_lock);
471 list_add_tail(¬ifier->n_list, &rs->rs_notify_queue);
472 spin_unlock(&rs->rs_lock);
474 ro->op_notifier = NULL;
477 spin_unlock_irqrestore(&rm->m_rs_lock, flags);
480 rds_wake_sk_sleep(rs);
481 sock_put(rds_rs_to_sk(rs));
484 EXPORT_SYMBOL_GPL(rds_rdma_send_complete);
487 * Just like above, except looks at atomic op
489 void rds_atomic_send_complete(struct rds_message *rm, int status)
491 struct rds_sock *rs = NULL;
492 struct rm_atomic_op *ao;
493 struct rds_notifier *notifier;
495 spin_lock(&rm->m_rs_lock);
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(&rm->m_rs_lock);
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 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));
626 sock_hold(rds_rs_to_sk(rs));
628 spin_lock(&rs->rs_lock);
630 if (test_and_clear_bit(RDS_MSG_ON_SOCK, &rm->m_flags)) {
631 struct rm_rdma_op *ro = &rm->rdma;
632 struct rds_notifier *notifier;
634 list_del_init(&rm->m_sock_item);
635 rds_send_sndbuf_remove(rs, rm);
637 if (ro->op_active && ro->op_notifier &&
638 (ro->op_notify || (ro->op_recverr && status))) {
639 notifier = ro->op_notifier;
640 list_add_tail(¬ifier->n_list,
641 &rs->rs_notify_queue);
642 if (!notifier->n_status)
643 notifier->n_status = status;
644 rm->rdma.op_notifier = NULL;
649 spin_unlock(&rs->rs_lock);
652 spin_unlock_irqrestore(&rm->m_rs_lock, flags);
659 rds_wake_sk_sleep(rs);
660 sock_put(rds_rs_to_sk(rs));
665 * Transports call here when they've determined that the receiver queued
666 * messages up to, and including, the given sequence number. Messages are
667 * moved to the retrans queue when rds_send_xmit picks them off the send
668 * queue. This means that in the TCP case, the message may not have been
669 * assigned the m_ack_seq yet - but that's fine as long as tcp_is_acked
670 * checks the RDS_MSG_HAS_ACK_SEQ bit.
672 * XXX It's not clear to me how this is safely serialized with socket
673 * destruction. Maybe it should bail if it sees SOCK_DEAD.
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_clear_bit();
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_clear_bit();
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);
746 list_del_init(&rm->m_conn_item);
747 spin_unlock_irqrestore(&conn->c_lock, flags);
750 * Couldn't grab m_rs_lock in top loop (lock ordering),
753 spin_lock_irqsave(&rm->m_rs_lock, flags);
755 spin_lock(&rs->rs_lock);
756 __rds_send_complete(rs, rm, RDS_RDMA_CANCELED);
757 spin_unlock(&rs->rs_lock);
760 spin_unlock_irqrestore(&rm->m_rs_lock, flags);
765 rds_wake_sk_sleep(rs);
767 while (!list_empty(&list)) {
768 rm = list_entry(list.next, struct rds_message, m_sock_item);
769 list_del_init(&rm->m_sock_item);
771 rds_message_wait(rm);
777 * we only want this to fire once so we use the callers 'queued'. It's
778 * possible that another thread can race with us and remove the
779 * message from the flow with RDS_CANCEL_SENT_TO.
781 static int rds_send_queue_rm(struct rds_sock *rs, struct rds_connection *conn,
782 struct rds_message *rm, __be16 sport,
783 __be16 dport, int *queued)
791 len = be32_to_cpu(rm->m_inc.i_hdr.h_len);
793 /* this is the only place which holds both the socket's rs_lock
794 * and the connection's c_lock */
795 spin_lock_irqsave(&rs->rs_lock, flags);
798 * If there is a little space in sndbuf, we don't queue anything,
799 * and userspace gets -EAGAIN. But poll() indicates there's send
800 * room. This can lead to bad behavior (spinning) if snd_bytes isn't
801 * freed up by incoming acks. So we check the *old* value of
802 * rs_snd_bytes here to allow the last msg to exceed the buffer,
803 * and poll() now knows no more data can be sent.
805 if (rs->rs_snd_bytes < rds_sk_sndbuf(rs)) {
806 rs->rs_snd_bytes += len;
808 /* let recv side know we are close to send space exhaustion.
809 * This is probably not the optimal way to do it, as this
810 * means we set the flag on *all* messages as soon as our
811 * throughput hits a certain threshold.
813 if (rs->rs_snd_bytes >= rds_sk_sndbuf(rs) / 2)
814 __set_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags);
816 list_add_tail(&rm->m_sock_item, &rs->rs_send_queue);
817 set_bit(RDS_MSG_ON_SOCK, &rm->m_flags);
818 rds_message_addref(rm);
821 /* The code ordering is a little weird, but we're
822 trying to minimize the time we hold c_lock */
823 rds_message_populate_header(&rm->m_inc.i_hdr, sport, dport, 0);
824 rm->m_inc.i_conn = conn;
825 rds_message_addref(rm);
827 spin_lock(&conn->c_lock);
828 rm->m_inc.i_hdr.h_sequence = cpu_to_be64(conn->c_next_tx_seq++);
829 list_add_tail(&rm->m_conn_item, &conn->c_send_queue);
830 set_bit(RDS_MSG_ON_CONN, &rm->m_flags);
831 spin_unlock(&conn->c_lock);
833 rdsdebug("queued msg %p len %d, rs %p bytes %d seq %llu\n",
834 rm, len, rs, rs->rs_snd_bytes,
835 (unsigned long long)be64_to_cpu(rm->m_inc.i_hdr.h_sequence));
840 spin_unlock_irqrestore(&rs->rs_lock, flags);
846 * rds_message is getting to be quite complicated, and we'd like to allocate
847 * it all in one go. This figures out how big it needs to be up front.
849 static int rds_rm_size(struct msghdr *msg, int data_len)
851 struct cmsghdr *cmsg;
856 for (cmsg = CMSG_FIRSTHDR(msg); cmsg; cmsg = CMSG_NXTHDR(msg, cmsg)) {
857 if (!CMSG_OK(msg, cmsg))
860 if (cmsg->cmsg_level != SOL_RDS)
863 switch (cmsg->cmsg_type) {
864 case RDS_CMSG_RDMA_ARGS:
866 retval = rds_rdma_extra_size(CMSG_DATA(cmsg));
873 case RDS_CMSG_RDMA_DEST:
874 case RDS_CMSG_RDMA_MAP:
876 /* these are valid but do no add any size */
879 case RDS_CMSG_ATOMIC_CSWP:
880 case RDS_CMSG_ATOMIC_FADD:
882 size += sizeof(struct scatterlist);
891 size += ceil(data_len, PAGE_SIZE) * sizeof(struct scatterlist);
893 /* Ensure (DEST, MAP) are never used with (ARGS, ATOMIC) */
894 if (cmsg_groups == 3)
900 static int rds_cmsg_send(struct rds_sock *rs, struct rds_message *rm,
901 struct msghdr *msg, int *allocated_mr)
903 struct cmsghdr *cmsg;
906 for (cmsg = CMSG_FIRSTHDR(msg); cmsg; cmsg = CMSG_NXTHDR(msg, cmsg)) {
907 if (!CMSG_OK(msg, cmsg))
910 if (cmsg->cmsg_level != SOL_RDS)
913 /* As a side effect, RDMA_DEST and RDMA_MAP will set
914 * rm->rdma.m_rdma_cookie and rm->rdma.m_rdma_mr.
916 switch (cmsg->cmsg_type) {
917 case RDS_CMSG_RDMA_ARGS:
918 ret = rds_cmsg_rdma_args(rs, rm, cmsg);
921 case RDS_CMSG_RDMA_DEST:
922 ret = rds_cmsg_rdma_dest(rs, rm, cmsg);
925 case RDS_CMSG_RDMA_MAP:
926 ret = rds_cmsg_rdma_map(rs, rm, cmsg);
930 case RDS_CMSG_ATOMIC_CSWP:
931 case RDS_CMSG_ATOMIC_FADD:
932 ret = rds_cmsg_atomic(rs, rm, cmsg);
946 int rds_sendmsg(struct kiocb *iocb, struct socket *sock, struct msghdr *msg,
949 struct sock *sk = sock->sk;
950 struct rds_sock *rs = rds_sk_to_rs(sk);
951 struct sockaddr_in *usin = (struct sockaddr_in *)msg->msg_name;
954 struct rds_message *rm = NULL;
955 struct rds_connection *conn;
957 int queued = 0, allocated_mr = 0;
958 int nonblock = msg->msg_flags & MSG_DONTWAIT;
959 long timeo = sock_sndtimeo(sk, nonblock);
961 /* Mirror Linux UDP mirror of BSD error message compatibility */
962 /* XXX: Perhaps MSG_MORE someday */
963 if (msg->msg_flags & ~(MSG_DONTWAIT | MSG_CMSG_COMPAT)) {
964 printk(KERN_INFO "msg_flags 0x%08X\n", msg->msg_flags);
969 if (msg->msg_namelen) {
970 /* XXX fail non-unicast destination IPs? */
971 if (msg->msg_namelen < sizeof(*usin) || usin->sin_family != AF_INET) {
975 daddr = usin->sin_addr.s_addr;
976 dport = usin->sin_port;
978 /* We only care about consistency with ->connect() */
980 daddr = rs->rs_conn_addr;
981 dport = rs->rs_conn_port;
985 /* racing with another thread binding seems ok here */
986 if (daddr == 0 || rs->rs_bound_addr == 0) {
987 ret = -ENOTCONN; /* XXX not a great errno */
991 /* size of rm including all sgs */
992 ret = rds_rm_size(msg, payload_len);
996 rm = rds_message_alloc(ret, GFP_KERNEL);
1002 /* Attach data to the rm */
1004 rm->data.op_sg = rds_message_alloc_sgs(rm, ceil(payload_len, PAGE_SIZE));
1005 ret = rds_message_copy_from_user(rm, msg->msg_iov, payload_len);
1009 rm->data.op_active = 1;
1011 rm->m_daddr = daddr;
1013 /* rds_conn_create has a spinlock that runs with IRQ off.
1014 * Caching the conn in the socket helps a lot. */
1015 if (rs->rs_conn && rs->rs_conn->c_faddr == daddr)
1018 conn = rds_conn_create_outgoing(rs->rs_bound_addr, daddr,
1020 sock->sk->sk_allocation);
1022 ret = PTR_ERR(conn);
1028 /* Parse any control messages the user may have included. */
1029 ret = rds_cmsg_send(rs, rm, msg, &allocated_mr);
1033 if (rm->rdma.op_active && !conn->c_trans->xmit_rdma) {
1034 if (printk_ratelimit())
1035 printk(KERN_NOTICE "rdma_op %p conn xmit_rdma %p\n",
1036 &rm->rdma, conn->c_trans->xmit_rdma);
1041 if (rm->atomic.op_active && !conn->c_trans->xmit_atomic) {
1042 if (printk_ratelimit())
1043 printk(KERN_NOTICE "atomic_op %p conn xmit_atomic %p\n",
1044 &rm->atomic, conn->c_trans->xmit_atomic);
1049 /* If the connection is down, trigger a connect. We may
1050 * have scheduled a delayed reconnect however - in this case
1051 * we should not interfere.
1053 if (rds_conn_state(conn) == RDS_CONN_DOWN &&
1054 !test_and_set_bit(RDS_RECONNECT_PENDING, &conn->c_flags))
1055 queue_delayed_work(rds_wq, &conn->c_conn_w, 0);
1057 ret = rds_cong_wait(conn->c_fcong, dport, nonblock, rs);
1059 rs->rs_seen_congestion = 1;
1063 while (!rds_send_queue_rm(rs, conn, rm, rs->rs_bound_port,
1065 rds_stats_inc(s_send_queue_full);
1066 /* XXX make sure this is reasonable */
1067 if (payload_len > rds_sk_sndbuf(rs)) {
1076 timeo = wait_event_interruptible_timeout(*sk_sleep(sk),
1077 rds_send_queue_rm(rs, conn, rm,
1082 rdsdebug("sendmsg woke queued %d timeo %ld\n", queued, timeo);
1083 if (timeo > 0 || timeo == MAX_SCHEDULE_TIMEOUT)
1093 * By now we've committed to the send. We reuse rds_send_worker()
1094 * to retry sends in the rds thread if the transport asks us to.
1096 rds_stats_inc(s_send_queued);
1098 if (!test_bit(RDS_LL_SEND_FULL, &conn->c_flags))
1099 rds_send_worker(&conn->c_send_w.work);
1101 rds_message_put(rm);
1105 /* If the user included a RDMA_MAP cmsg, we allocated a MR on the fly.
1106 * If the sendmsg goes through, we keep the MR. If it fails with EAGAIN
1107 * or in any other way, we need to destroy the MR again */
1109 rds_rdma_unuse(rs, rds_rdma_cookie_key(rm->m_rdma_cookie), 1);
1112 rds_message_put(rm);
1117 * Reply to a ping packet.
1120 rds_send_pong(struct rds_connection *conn, __be16 dport)
1122 struct rds_message *rm;
1123 unsigned long flags;
1126 rm = rds_message_alloc(0, GFP_ATOMIC);
1132 rm->m_daddr = conn->c_faddr;
1134 /* If the connection is down, trigger a connect. We may
1135 * have scheduled a delayed reconnect however - in this case
1136 * we should not interfere.
1138 if (rds_conn_state(conn) == RDS_CONN_DOWN &&
1139 !test_and_set_bit(RDS_RECONNECT_PENDING, &conn->c_flags))
1140 queue_delayed_work(rds_wq, &conn->c_conn_w, 0);
1142 ret = rds_cong_wait(conn->c_fcong, dport, 1, NULL);
1146 spin_lock_irqsave(&conn->c_lock, flags);
1147 list_add_tail(&rm->m_conn_item, &conn->c_send_queue);
1148 set_bit(RDS_MSG_ON_CONN, &rm->m_flags);
1149 rds_message_addref(rm);
1150 rm->m_inc.i_conn = conn;
1152 rds_message_populate_header(&rm->m_inc.i_hdr, 0, dport,
1153 conn->c_next_tx_seq);
1154 conn->c_next_tx_seq++;
1155 spin_unlock_irqrestore(&conn->c_lock, flags);
1157 rds_stats_inc(s_send_queued);
1158 rds_stats_inc(s_send_pong);
1160 queue_delayed_work(rds_wq, &conn->c_send_w, 0);
1161 rds_message_put(rm);
1166 rds_message_put(rm);