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_rdma_sent
= 0;
77 conn
->c_map_queued
= 0;
79 conn
->c_unacked_packets
= rds_sysctl_max_unacked_packets
;
80 conn
->c_unacked_bytes
= rds_sysctl_max_unacked_bytes
;
82 /* Mark messages as retransmissions, and move them to the send q */
83 spin_lock_irqsave(&conn
->c_lock
, flags
);
84 list_for_each_entry_safe(rm
, tmp
, &conn
->c_retrans
, m_conn_item
) {
85 set_bit(RDS_MSG_ACK_REQUIRED
, &rm
->m_flags
);
86 set_bit(RDS_MSG_RETRANSMITTED
, &rm
->m_flags
);
88 list_splice_init(&conn
->c_retrans
, &conn
->c_send_queue
);
89 spin_unlock_irqrestore(&conn
->c_lock
, flags
);
93 * We're making the concious trade-off here to only send one message
94 * down the connection at a time.
96 * - tx queueing is a simple fifo list
97 * - reassembly is optional and easily done by transports per conn
98 * - no per flow rx lookup at all, straight to the socket
99 * - less per-frag memory and wire overhead
101 * - queued acks can be delayed behind large messages
103 * - small message latency is higher behind queued large messages
104 * - large message latency isn't starved by intervening small sends
106 int rds_send_xmit(struct rds_connection
*conn
)
108 struct rds_message
*rm
;
111 unsigned int send_quota
= send_batch_count
;
112 struct scatterlist
*sg
;
115 LIST_HEAD(to_be_dropped
);
118 * sendmsg calls here after having queued its message on the send
119 * queue. We only have one task feeding the connection at a time. If
120 * another thread is already feeding the queue then we back off. This
121 * avoids blocking the caller and trading per-connection data between
122 * caches per message.
124 * The sem holder will issue a retry if they notice that someone queued
125 * a message after they stopped walking the send queue but before they
128 if (!mutex_trylock(&conn
->c_send_lock
)) {
129 rds_stats_inc(s_send_sem_contention
);
134 if (conn
->c_trans
->xmit_prepare
)
135 conn
->c_trans
->xmit_prepare(conn
);
138 * spin trying to push headers and data down the connection until
139 * the connection doens't make forward progress.
141 while (--send_quota
) {
143 * See if need to send a congestion map update if we're
144 * between sending messages. The send_sem protects our sole
145 * use of c_map_offset and _bytes.
146 * Note this is used only by transports that define a special
147 * xmit_cong_map function. For all others, we create allocate
148 * a cong_map message and treat it just like any other send.
150 if (conn
->c_map_bytes
) {
151 ret
= conn
->c_trans
->xmit_cong_map(conn
, conn
->c_lcong
,
156 conn
->c_map_offset
+= ret
;
157 conn
->c_map_bytes
-= ret
;
158 if (conn
->c_map_bytes
)
162 /* If we're done sending the current message, clear the
163 * offset and S/G temporaries.
165 rm
= conn
->c_xmit_rm
;
167 conn
->c_xmit_hdr_off
== sizeof(struct rds_header
) &&
168 conn
->c_xmit_sg
== rm
->data
.m_nents
) {
169 conn
->c_xmit_rm
= NULL
;
171 conn
->c_xmit_hdr_off
= 0;
172 conn
->c_xmit_data_off
= 0;
173 conn
->c_xmit_rdma_sent
= 0;
175 /* Release the reference to the previous message. */
180 /* If we're asked to send a cong map update, do so.
182 if (!rm
&& test_and_clear_bit(0, &conn
->c_map_queued
)) {
183 if (conn
->c_trans
->xmit_cong_map
) {
184 conn
->c_map_offset
= 0;
185 conn
->c_map_bytes
= sizeof(struct rds_header
) +
190 rm
= rds_cong_update_alloc(conn
);
196 conn
->c_xmit_rm
= rm
;
200 * Grab the next message from the send queue, if there is one.
202 * c_xmit_rm holds a ref while we're sending this message down
203 * the connction. We can use this ref while holding the
204 * send_sem.. rds_send_reset() is serialized with it.
209 spin_lock_irqsave(&conn
->c_lock
, flags
);
211 if (!list_empty(&conn
->c_send_queue
)) {
212 rm
= list_entry(conn
->c_send_queue
.next
,
215 rds_message_addref(rm
);
218 * Move the message from the send queue to the retransmit
221 list_move_tail(&rm
->m_conn_item
, &conn
->c_retrans
);
224 spin_unlock_irqrestore(&conn
->c_lock
, flags
);
231 /* Unfortunately, the way Infiniband deals with
232 * RDMA to a bad MR key is by moving the entire
233 * queue pair to error state. We cold possibly
234 * recover from that, but right now we drop the
236 * Therefore, we never retransmit messages with RDMA ops.
238 if (rm
->rdma
.m_rdma_op
.r_active
&&
239 test_bit(RDS_MSG_RETRANSMITTED
, &rm
->m_flags
)) {
240 spin_lock_irqsave(&conn
->c_lock
, flags
);
241 if (test_and_clear_bit(RDS_MSG_ON_CONN
, &rm
->m_flags
))
242 list_move(&rm
->m_conn_item
, &to_be_dropped
);
243 spin_unlock_irqrestore(&conn
->c_lock
, flags
);
248 /* Require an ACK every once in a while */
249 len
= ntohl(rm
->m_inc
.i_hdr
.h_len
);
250 if (conn
->c_unacked_packets
== 0 ||
251 conn
->c_unacked_bytes
< len
) {
252 __set_bit(RDS_MSG_ACK_REQUIRED
, &rm
->m_flags
);
254 conn
->c_unacked_packets
= rds_sysctl_max_unacked_packets
;
255 conn
->c_unacked_bytes
= rds_sysctl_max_unacked_bytes
;
256 rds_stats_inc(s_send_ack_required
);
258 conn
->c_unacked_bytes
-= len
;
259 conn
->c_unacked_packets
--;
262 conn
->c_xmit_rm
= rm
;
266 * Try and send an rdma message. Let's see if we can
267 * keep this simple and require that the transport either
268 * send the whole rdma or none of it.
270 if (rm
->rdma
.m_rdma_op
.r_active
&& !conn
->c_xmit_rdma_sent
) {
271 ret
= conn
->c_trans
->xmit_rdma(conn
, &rm
->rdma
.m_rdma_op
);
274 conn
->c_xmit_rdma_sent
= 1;
275 /* The transport owns the mapped memory for now.
276 * You can't unmap it while it's on the send queue */
277 set_bit(RDS_MSG_MAPPED
, &rm
->m_flags
);
280 if (conn
->c_xmit_hdr_off
< sizeof(struct rds_header
) ||
281 conn
->c_xmit_sg
< rm
->data
.m_nents
) {
282 ret
= conn
->c_trans
->xmit(conn
, rm
,
283 conn
->c_xmit_hdr_off
,
285 conn
->c_xmit_data_off
);
289 if (conn
->c_xmit_hdr_off
< sizeof(struct rds_header
)) {
290 tmp
= min_t(int, ret
,
291 sizeof(struct rds_header
) -
292 conn
->c_xmit_hdr_off
);
293 conn
->c_xmit_hdr_off
+= tmp
;
297 sg
= &rm
->data
.m_sg
[conn
->c_xmit_sg
];
299 tmp
= min_t(int, ret
, sg
->length
-
300 conn
->c_xmit_data_off
);
301 conn
->c_xmit_data_off
+= tmp
;
303 if (conn
->c_xmit_data_off
== sg
->length
) {
304 conn
->c_xmit_data_off
= 0;
308 conn
->c_xmit_sg
== rm
->data
.m_nents
);
314 /* Nuke any messages we decided not to retransmit. */
315 if (!list_empty(&to_be_dropped
))
316 rds_send_remove_from_sock(&to_be_dropped
, RDS_RDMA_DROPPED
);
318 if (conn
->c_trans
->xmit_complete
)
319 conn
->c_trans
->xmit_complete(conn
);
322 * We might be racing with another sender who queued a message but
323 * backed off on noticing that we held the c_send_lock. If we check
324 * for queued messages after dropping the sem then either we'll
325 * see the queued message or the queuer will get the sem. If we
326 * notice the queued message then we trigger an immediate retry.
328 * We need to be careful only to do this when we stopped processing
329 * the send queue because it was empty. It's the only way we
330 * stop processing the loop when the transport hasn't taken
331 * responsibility for forward progress.
333 mutex_unlock(&conn
->c_send_lock
);
335 if (conn
->c_map_bytes
|| (send_quota
== 0 && !was_empty
)) {
336 /* We exhausted the send quota, but there's work left to
337 * do. Return and (re-)schedule the send worker.
342 if (ret
== 0 && was_empty
) {
343 /* A simple bit test would be way faster than taking the
345 spin_lock_irqsave(&conn
->c_lock
, flags
);
346 if (!list_empty(&conn
->c_send_queue
)) {
347 rds_stats_inc(s_send_sem_queue_raced
);
350 spin_unlock_irqrestore(&conn
->c_lock
, flags
);
356 static void rds_send_sndbuf_remove(struct rds_sock
*rs
, struct rds_message
*rm
)
358 u32 len
= be32_to_cpu(rm
->m_inc
.i_hdr
.h_len
);
360 assert_spin_locked(&rs
->rs_lock
);
362 BUG_ON(rs
->rs_snd_bytes
< len
);
363 rs
->rs_snd_bytes
-= len
;
365 if (rs
->rs_snd_bytes
== 0)
366 rds_stats_inc(s_send_queue_empty
);
369 static inline int rds_send_is_acked(struct rds_message
*rm
, u64 ack
,
370 is_acked_func is_acked
)
373 return is_acked(rm
, ack
);
374 return be64_to_cpu(rm
->m_inc
.i_hdr
.h_sequence
) <= ack
;
378 * Returns true if there are no messages on the send and retransmit queues
379 * which have a sequence number greater than or equal to the given sequence
382 int rds_send_acked_before(struct rds_connection
*conn
, u64 seq
)
384 struct rds_message
*rm
, *tmp
;
387 spin_lock(&conn
->c_lock
);
389 list_for_each_entry_safe(rm
, tmp
, &conn
->c_retrans
, m_conn_item
) {
390 if (be64_to_cpu(rm
->m_inc
.i_hdr
.h_sequence
) < seq
)
395 list_for_each_entry_safe(rm
, tmp
, &conn
->c_send_queue
, m_conn_item
) {
396 if (be64_to_cpu(rm
->m_inc
.i_hdr
.h_sequence
) < seq
)
401 spin_unlock(&conn
->c_lock
);
407 * This is pretty similar to what happens below in the ACK
408 * handling code - except that we call here as soon as we get
409 * the IB send completion on the RDMA op and the accompanying
412 void rds_rdma_send_complete(struct rds_message
*rm
, int status
)
414 struct rds_sock
*rs
= NULL
;
415 struct rds_rdma_op
*ro
;
416 struct rds_notifier
*notifier
;
419 spin_lock_irqsave(&rm
->m_rs_lock
, flags
);
421 ro
= &rm
->rdma
.m_rdma_op
;
422 if (test_bit(RDS_MSG_ON_SOCK
, &rm
->m_flags
) &&
423 ro
->r_active
&& ro
->r_notify
&& ro
->r_notifier
) {
424 notifier
= ro
->r_notifier
;
426 sock_hold(rds_rs_to_sk(rs
));
428 notifier
->n_status
= status
;
429 spin_lock(&rs
->rs_lock
);
430 list_add_tail(¬ifier
->n_list
, &rs
->rs_notify_queue
);
431 spin_unlock(&rs
->rs_lock
);
433 ro
->r_notifier
= NULL
;
436 spin_unlock_irqrestore(&rm
->m_rs_lock
, flags
);
439 rds_wake_sk_sleep(rs
);
440 sock_put(rds_rs_to_sk(rs
));
443 EXPORT_SYMBOL_GPL(rds_rdma_send_complete
);
446 * This is the same as rds_rdma_send_complete except we
447 * don't do any locking - we have all the ingredients (message,
448 * socket, socket lock) and can just move the notifier.
451 __rds_rdma_send_complete(struct rds_sock
*rs
, struct rds_message
*rm
, int status
)
453 struct rds_rdma_op
*ro
;
455 ro
= &rm
->rdma
.m_rdma_op
;
456 if (ro
->r_active
&& ro
->r_notify
&& ro
->r_notifier
) {
457 ro
->r_notifier
->n_status
= status
;
458 list_add_tail(&ro
->r_notifier
->n_list
, &rs
->rs_notify_queue
);
459 ro
->r_notifier
= NULL
;
462 /* No need to wake the app - caller does this */
466 * This is called from the IB send completion when we detect
467 * a RDMA operation that failed with remote access error.
468 * So speed is not an issue here.
470 struct rds_message
*rds_send_get_message(struct rds_connection
*conn
,
471 struct rds_rdma_op
*op
)
473 struct rds_message
*rm
, *tmp
, *found
= NULL
;
476 spin_lock_irqsave(&conn
->c_lock
, flags
);
478 list_for_each_entry_safe(rm
, tmp
, &conn
->c_retrans
, m_conn_item
) {
479 if (&rm
->rdma
.m_rdma_op
== op
) {
480 atomic_inc(&rm
->m_refcount
);
486 list_for_each_entry_safe(rm
, tmp
, &conn
->c_send_queue
, m_conn_item
) {
487 if (&rm
->rdma
.m_rdma_op
== op
) {
488 atomic_inc(&rm
->m_refcount
);
495 spin_unlock_irqrestore(&conn
->c_lock
, flags
);
499 EXPORT_SYMBOL_GPL(rds_send_get_message
);
502 * This removes messages from the socket's list if they're on it. The list
503 * argument must be private to the caller, we must be able to modify it
504 * without locks. The messages must have a reference held for their
505 * position on the list. This function will drop that reference after
506 * removing the messages from the 'messages' list regardless of if it found
507 * the messages on the socket list or not.
509 void rds_send_remove_from_sock(struct list_head
*messages
, int status
)
512 struct rds_sock
*rs
= NULL
;
513 struct rds_message
*rm
;
515 while (!list_empty(messages
)) {
518 rm
= list_entry(messages
->next
, struct rds_message
,
520 list_del_init(&rm
->m_conn_item
);
523 * If we see this flag cleared then we're *sure* that someone
524 * else beat us to removing it from the sock. If we race
525 * with their flag update we'll get the lock and then really
526 * see that the flag has been cleared.
528 * The message spinlock makes sure nobody clears rm->m_rs
529 * while we're messing with it. It does not prevent the
530 * message from being removed from the socket, though.
532 spin_lock_irqsave(&rm
->m_rs_lock
, flags
);
533 if (!test_bit(RDS_MSG_ON_SOCK
, &rm
->m_flags
))
534 goto unlock_and_drop
;
536 if (rs
!= rm
->m_rs
) {
538 rds_wake_sk_sleep(rs
);
539 sock_put(rds_rs_to_sk(rs
));
542 sock_hold(rds_rs_to_sk(rs
));
544 spin_lock(&rs
->rs_lock
);
546 if (test_and_clear_bit(RDS_MSG_ON_SOCK
, &rm
->m_flags
)) {
547 struct rds_rdma_op
*ro
= &rm
->rdma
.m_rdma_op
;
548 struct rds_notifier
*notifier
;
550 list_del_init(&rm
->m_sock_item
);
551 rds_send_sndbuf_remove(rs
, rm
);
553 if (ro
->r_active
&& ro
->r_notifier
&&
554 (status
|| ro
->r_notify
)) {
555 notifier
= ro
->r_notifier
;
556 list_add_tail(¬ifier
->n_list
,
557 &rs
->rs_notify_queue
);
558 if (!notifier
->n_status
)
559 notifier
->n_status
= status
;
560 rm
->rdma
.m_rdma_op
.r_notifier
= NULL
;
565 spin_unlock(&rs
->rs_lock
);
568 spin_unlock_irqrestore(&rm
->m_rs_lock
, flags
);
575 rds_wake_sk_sleep(rs
);
576 sock_put(rds_rs_to_sk(rs
));
581 * Transports call here when they've determined that the receiver queued
582 * messages up to, and including, the given sequence number. Messages are
583 * moved to the retrans queue when rds_send_xmit picks them off the send
584 * queue. This means that in the TCP case, the message may not have been
585 * assigned the m_ack_seq yet - but that's fine as long as tcp_is_acked
586 * checks the RDS_MSG_HAS_ACK_SEQ bit.
588 * XXX It's not clear to me how this is safely serialized with socket
589 * destruction. Maybe it should bail if it sees SOCK_DEAD.
591 void rds_send_drop_acked(struct rds_connection
*conn
, u64 ack
,
592 is_acked_func is_acked
)
594 struct rds_message
*rm
, *tmp
;
598 spin_lock_irqsave(&conn
->c_lock
, flags
);
600 list_for_each_entry_safe(rm
, tmp
, &conn
->c_retrans
, m_conn_item
) {
601 if (!rds_send_is_acked(rm
, ack
, is_acked
))
604 list_move(&rm
->m_conn_item
, &list
);
605 clear_bit(RDS_MSG_ON_CONN
, &rm
->m_flags
);
608 /* order flag updates with spin locks */
609 if (!list_empty(&list
))
610 smp_mb__after_clear_bit();
612 spin_unlock_irqrestore(&conn
->c_lock
, flags
);
614 /* now remove the messages from the sock list as needed */
615 rds_send_remove_from_sock(&list
, RDS_RDMA_SUCCESS
);
617 EXPORT_SYMBOL_GPL(rds_send_drop_acked
);
619 void rds_send_drop_to(struct rds_sock
*rs
, struct sockaddr_in
*dest
)
621 struct rds_message
*rm
, *tmp
;
622 struct rds_connection
*conn
;
626 /* get all the messages we're dropping under the rs lock */
627 spin_lock_irqsave(&rs
->rs_lock
, flags
);
629 list_for_each_entry_safe(rm
, tmp
, &rs
->rs_send_queue
, m_sock_item
) {
630 if (dest
&& (dest
->sin_addr
.s_addr
!= rm
->m_daddr
||
631 dest
->sin_port
!= rm
->m_inc
.i_hdr
.h_dport
))
634 list_move(&rm
->m_sock_item
, &list
);
635 rds_send_sndbuf_remove(rs
, rm
);
636 clear_bit(RDS_MSG_ON_SOCK
, &rm
->m_flags
);
639 /* order flag updates with the rs lock */
640 smp_mb__after_clear_bit();
642 spin_unlock_irqrestore(&rs
->rs_lock
, flags
);
644 if (list_empty(&list
))
647 /* Remove the messages from the conn */
648 list_for_each_entry(rm
, &list
, m_sock_item
) {
650 conn
= rm
->m_inc
.i_conn
;
652 spin_lock_irqsave(&conn
->c_lock
, flags
);
654 * Maybe someone else beat us to removing rm from the conn.
655 * If we race with their flag update we'll get the lock and
656 * then really see that the flag has been cleared.
658 if (!test_and_clear_bit(RDS_MSG_ON_CONN
, &rm
->m_flags
)) {
659 spin_unlock_irqrestore(&conn
->c_lock
, flags
);
662 list_del_init(&rm
->m_conn_item
);
663 spin_unlock_irqrestore(&conn
->c_lock
, flags
);
666 * Couldn't grab m_rs_lock in top loop (lock ordering),
669 spin_lock_irqsave(&rm
->m_rs_lock
, flags
);
671 spin_lock(&rs
->rs_lock
);
672 __rds_rdma_send_complete(rs
, rm
, RDS_RDMA_CANCELED
);
673 spin_unlock(&rs
->rs_lock
);
676 spin_unlock_irqrestore(&rm
->m_rs_lock
, flags
);
681 rds_wake_sk_sleep(rs
);
683 while (!list_empty(&list
)) {
684 rm
= list_entry(list
.next
, struct rds_message
, m_sock_item
);
685 list_del_init(&rm
->m_sock_item
);
687 rds_message_wait(rm
);
693 * we only want this to fire once so we use the callers 'queued'. It's
694 * possible that another thread can race with us and remove the
695 * message from the flow with RDS_CANCEL_SENT_TO.
697 static int rds_send_queue_rm(struct rds_sock
*rs
, struct rds_connection
*conn
,
698 struct rds_message
*rm
, __be16 sport
,
699 __be16 dport
, int *queued
)
707 len
= be32_to_cpu(rm
->m_inc
.i_hdr
.h_len
);
709 /* this is the only place which holds both the socket's rs_lock
710 * and the connection's c_lock */
711 spin_lock_irqsave(&rs
->rs_lock
, flags
);
714 * If there is a little space in sndbuf, we don't queue anything,
715 * and userspace gets -EAGAIN. But poll() indicates there's send
716 * room. This can lead to bad behavior (spinning) if snd_bytes isn't
717 * freed up by incoming acks. So we check the *old* value of
718 * rs_snd_bytes here to allow the last msg to exceed the buffer,
719 * and poll() now knows no more data can be sent.
721 if (rs
->rs_snd_bytes
< rds_sk_sndbuf(rs
)) {
722 rs
->rs_snd_bytes
+= len
;
724 /* let recv side know we are close to send space exhaustion.
725 * This is probably not the optimal way to do it, as this
726 * means we set the flag on *all* messages as soon as our
727 * throughput hits a certain threshold.
729 if (rs
->rs_snd_bytes
>= rds_sk_sndbuf(rs
) / 2)
730 __set_bit(RDS_MSG_ACK_REQUIRED
, &rm
->m_flags
);
732 list_add_tail(&rm
->m_sock_item
, &rs
->rs_send_queue
);
733 set_bit(RDS_MSG_ON_SOCK
, &rm
->m_flags
);
734 rds_message_addref(rm
);
737 /* The code ordering is a little weird, but we're
738 trying to minimize the time we hold c_lock */
739 rds_message_populate_header(&rm
->m_inc
.i_hdr
, sport
, dport
, 0);
740 rm
->m_inc
.i_conn
= conn
;
741 rds_message_addref(rm
);
743 spin_lock(&conn
->c_lock
);
744 rm
->m_inc
.i_hdr
.h_sequence
= cpu_to_be64(conn
->c_next_tx_seq
++);
745 list_add_tail(&rm
->m_conn_item
, &conn
->c_send_queue
);
746 set_bit(RDS_MSG_ON_CONN
, &rm
->m_flags
);
747 spin_unlock(&conn
->c_lock
);
749 rdsdebug("queued msg %p len %d, rs %p bytes %d seq %llu\n",
750 rm
, len
, rs
, rs
->rs_snd_bytes
,
751 (unsigned long long)be64_to_cpu(rm
->m_inc
.i_hdr
.h_sequence
));
756 spin_unlock_irqrestore(&rs
->rs_lock
, flags
);
762 * rds_message is getting to be quite complicated, and we'd like to allocate
763 * it all in one go. This figures out how big it needs to be up front.
765 static int rds_rm_size(struct msghdr
*msg
, int data_len
)
767 struct cmsghdr
*cmsg
;
771 for (cmsg
= CMSG_FIRSTHDR(msg
); cmsg
; cmsg
= CMSG_NXTHDR(msg
, cmsg
)) {
772 if (!CMSG_OK(msg
, cmsg
))
775 if (cmsg
->cmsg_level
!= SOL_RDS
)
778 switch (cmsg
->cmsg_type
) {
779 case RDS_CMSG_RDMA_ARGS
:
780 retval
= rds_rdma_extra_size(CMSG_DATA(cmsg
));
786 case RDS_CMSG_RDMA_DEST
:
787 case RDS_CMSG_RDMA_MAP
:
788 /* these are valid but do no add any size */
797 size
+= ceil(data_len
, PAGE_SIZE
) * sizeof(struct scatterlist
);
802 static int rds_cmsg_send(struct rds_sock
*rs
, struct rds_message
*rm
,
803 struct msghdr
*msg
, int *allocated_mr
)
805 struct cmsghdr
*cmsg
;
808 for (cmsg
= CMSG_FIRSTHDR(msg
); cmsg
; cmsg
= CMSG_NXTHDR(msg
, cmsg
)) {
809 if (!CMSG_OK(msg
, cmsg
))
812 if (cmsg
->cmsg_level
!= SOL_RDS
)
815 /* As a side effect, RDMA_DEST and RDMA_MAP will set
816 * rm->m_rdma_cookie and rm->m_rdma_mr.
818 switch (cmsg
->cmsg_type
) {
819 case RDS_CMSG_RDMA_ARGS
:
820 ret
= rds_cmsg_rdma_args(rs
, rm
, cmsg
);
823 case RDS_CMSG_RDMA_DEST
:
824 ret
= rds_cmsg_rdma_dest(rs
, rm
, cmsg
);
827 case RDS_CMSG_RDMA_MAP
:
828 ret
= rds_cmsg_rdma_map(rs
, rm
, cmsg
);
844 int rds_sendmsg(struct kiocb
*iocb
, struct socket
*sock
, struct msghdr
*msg
,
847 struct sock
*sk
= sock
->sk
;
848 struct rds_sock
*rs
= rds_sk_to_rs(sk
);
849 struct sockaddr_in
*usin
= (struct sockaddr_in
*)msg
->msg_name
;
852 struct rds_message
*rm
= NULL
;
853 struct rds_connection
*conn
;
855 int queued
= 0, allocated_mr
= 0;
856 int nonblock
= msg
->msg_flags
& MSG_DONTWAIT
;
857 long timeo
= sock_sndtimeo(sk
, nonblock
);
859 /* Mirror Linux UDP mirror of BSD error message compatibility */
860 /* XXX: Perhaps MSG_MORE someday */
861 if (msg
->msg_flags
& ~(MSG_DONTWAIT
| MSG_CMSG_COMPAT
)) {
862 printk(KERN_INFO
"msg_flags 0x%08X\n", msg
->msg_flags
);
867 if (msg
->msg_namelen
) {
868 /* XXX fail non-unicast destination IPs? */
869 if (msg
->msg_namelen
< sizeof(*usin
) || usin
->sin_family
!= AF_INET
) {
873 daddr
= usin
->sin_addr
.s_addr
;
874 dport
= usin
->sin_port
;
876 /* We only care about consistency with ->connect() */
878 daddr
= rs
->rs_conn_addr
;
879 dport
= rs
->rs_conn_port
;
883 /* racing with another thread binding seems ok here */
884 if (daddr
== 0 || rs
->rs_bound_addr
== 0) {
885 ret
= -ENOTCONN
; /* XXX not a great errno */
889 /* size of rm including all sgs */
890 ret
= rds_rm_size(msg
, payload_len
);
894 rm
= rds_message_alloc(ret
, GFP_KERNEL
);
900 rm
->data
.m_sg
= rds_message_alloc_sgs(rm
, ceil(payload_len
, PAGE_SIZE
));
901 /* XXX fix this to not allocate memory */
902 ret
= rds_message_copy_from_user(rm
, msg
->msg_iov
, payload_len
);
908 /* rds_conn_create has a spinlock that runs with IRQ off.
909 * Caching the conn in the socket helps a lot. */
910 if (rs
->rs_conn
&& rs
->rs_conn
->c_faddr
== daddr
)
913 conn
= rds_conn_create_outgoing(rs
->rs_bound_addr
, daddr
,
915 sock
->sk
->sk_allocation
);
923 /* Parse any control messages the user may have included. */
924 ret
= rds_cmsg_send(rs
, rm
, msg
, &allocated_mr
);
928 if ((rm
->m_rdma_cookie
|| rm
->rdma
.m_rdma_op
.r_active
) &&
929 !conn
->c_trans
->xmit_rdma
) {
930 if (printk_ratelimit())
931 printk(KERN_NOTICE
"rdma_op %p conn xmit_rdma %p\n",
932 &rm
->rdma
.m_rdma_op
, conn
->c_trans
->xmit_rdma
);
937 /* If the connection is down, trigger a connect. We may
938 * have scheduled a delayed reconnect however - in this case
939 * we should not interfere.
941 if (rds_conn_state(conn
) == RDS_CONN_DOWN
&&
942 !test_and_set_bit(RDS_RECONNECT_PENDING
, &conn
->c_flags
))
943 queue_delayed_work(rds_wq
, &conn
->c_conn_w
, 0);
945 ret
= rds_cong_wait(conn
->c_fcong
, dport
, nonblock
, rs
);
947 rs
->rs_seen_congestion
= 1;
951 while (!rds_send_queue_rm(rs
, conn
, rm
, rs
->rs_bound_port
,
953 rds_stats_inc(s_send_queue_full
);
954 /* XXX make sure this is reasonable */
955 if (payload_len
> rds_sk_sndbuf(rs
)) {
964 timeo
= wait_event_interruptible_timeout(*sk_sleep(sk
),
965 rds_send_queue_rm(rs
, conn
, rm
,
970 rdsdebug("sendmsg woke queued %d timeo %ld\n", queued
, timeo
);
971 if (timeo
> 0 || timeo
== MAX_SCHEDULE_TIMEOUT
)
981 * By now we've committed to the send. We reuse rds_send_worker()
982 * to retry sends in the rds thread if the transport asks us to.
984 rds_stats_inc(s_send_queued
);
986 if (!test_bit(RDS_LL_SEND_FULL
, &conn
->c_flags
))
987 rds_send_worker(&conn
->c_send_w
.work
);
993 /* If the user included a RDMA_MAP cmsg, we allocated a MR on the fly.
994 * If the sendmsg goes through, we keep the MR. If it fails with EAGAIN
995 * or in any other way, we need to destroy the MR again */
997 rds_rdma_unuse(rs
, rds_rdma_cookie_key(rm
->m_rdma_cookie
), 1);
1000 rds_message_put(rm
);
1005 * Reply to a ping packet.
1008 rds_send_pong(struct rds_connection
*conn
, __be16 dport
)
1010 struct rds_message
*rm
;
1011 unsigned long flags
;
1014 rm
= rds_message_alloc(0, GFP_ATOMIC
);
1020 rm
->m_daddr
= conn
->c_faddr
;
1022 /* If the connection is down, trigger a connect. We may
1023 * have scheduled a delayed reconnect however - in this case
1024 * we should not interfere.
1026 if (rds_conn_state(conn
) == RDS_CONN_DOWN
&&
1027 !test_and_set_bit(RDS_RECONNECT_PENDING
, &conn
->c_flags
))
1028 queue_delayed_work(rds_wq
, &conn
->c_conn_w
, 0);
1030 ret
= rds_cong_wait(conn
->c_fcong
, dport
, 1, NULL
);
1034 spin_lock_irqsave(&conn
->c_lock
, flags
);
1035 list_add_tail(&rm
->m_conn_item
, &conn
->c_send_queue
);
1036 set_bit(RDS_MSG_ON_CONN
, &rm
->m_flags
);
1037 rds_message_addref(rm
);
1038 rm
->m_inc
.i_conn
= conn
;
1040 rds_message_populate_header(&rm
->m_inc
.i_hdr
, 0, dport
,
1041 conn
->c_next_tx_seq
);
1042 conn
->c_next_tx_seq
++;
1043 spin_unlock_irqrestore(&conn
->c_lock
, flags
);
1045 rds_stats_inc(s_send_queued
);
1046 rds_stats_inc(s_send_pong
);
1048 queue_delayed_work(rds_wq
, &conn
->c_send_w
, 0);
1049 rds_message_put(rm
);
1054 rds_message_put(rm
);