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;
76 conn
->c_xmit_atomic_sent
= 0;
78 conn
->c_map_queued
= 0;
80 conn
->c_unacked_packets
= rds_sysctl_max_unacked_packets
;
81 conn
->c_unacked_bytes
= rds_sysctl_max_unacked_bytes
;
83 /* Mark messages as retransmissions, and move them to the send q */
84 spin_lock_irqsave(&conn
->c_lock
, flags
);
85 list_for_each_entry_safe(rm
, tmp
, &conn
->c_retrans
, m_conn_item
) {
86 set_bit(RDS_MSG_ACK_REQUIRED
, &rm
->m_flags
);
87 set_bit(RDS_MSG_RETRANSMITTED
, &rm
->m_flags
);
89 list_splice_init(&conn
->c_retrans
, &conn
->c_send_queue
);
90 spin_unlock_irqrestore(&conn
->c_lock
, flags
);
94 * We're making the concious trade-off here to only send one message
95 * down the connection at a time.
97 * - tx queueing is a simple fifo list
98 * - reassembly is optional and easily done by transports per conn
99 * - no per flow rx lookup at all, straight to the socket
100 * - less per-frag memory and wire overhead
102 * - queued acks can be delayed behind large messages
104 * - small message latency is higher behind queued large messages
105 * - large message latency isn't starved by intervening small sends
107 int rds_send_xmit(struct rds_connection
*conn
)
109 struct rds_message
*rm
;
112 unsigned int send_quota
= send_batch_count
;
113 struct scatterlist
*sg
;
116 LIST_HEAD(to_be_dropped
);
119 * sendmsg calls here after having queued its message on the send
120 * queue. We only have one task feeding the connection at a time. If
121 * another thread is already feeding the queue then we back off. This
122 * avoids blocking the caller and trading per-connection data between
123 * caches per message.
125 * The sem holder will issue a retry if they notice that someone queued
126 * a message after they stopped walking the send queue but before they
129 if (!mutex_trylock(&conn
->c_send_lock
)) {
130 rds_stats_inc(s_send_sem_contention
);
135 if (conn
->c_trans
->xmit_prepare
)
136 conn
->c_trans
->xmit_prepare(conn
);
139 * spin trying to push headers and data down the connection until
140 * the connection doens't make forward progress.
142 while (--send_quota
) {
144 * See if need to send a congestion map update if we're
145 * between sending messages. The send_sem protects our sole
146 * use of c_map_offset and _bytes.
147 * Note this is used only by transports that define a special
148 * xmit_cong_map function. For all others, we create allocate
149 * a cong_map message and treat it just like any other send.
151 if (conn
->c_map_bytes
) {
152 ret
= conn
->c_trans
->xmit_cong_map(conn
, conn
->c_lcong
,
157 conn
->c_map_offset
+= ret
;
158 conn
->c_map_bytes
-= ret
;
159 if (conn
->c_map_bytes
)
163 /* If we're done sending the current message, clear the
164 * offset and S/G temporaries.
166 rm
= conn
->c_xmit_rm
;
168 conn
->c_xmit_hdr_off
== sizeof(struct rds_header
) &&
169 conn
->c_xmit_sg
== rm
->data
.op_nents
) {
170 conn
->c_xmit_rm
= NULL
;
172 conn
->c_xmit_hdr_off
= 0;
173 conn
->c_xmit_data_off
= 0;
174 conn
->c_xmit_rdma_sent
= 0;
175 conn
->c_xmit_atomic_sent
= 0;
177 /* Release the reference to the previous message. */
182 /* If we're asked to send a cong map update, do so.
184 if (!rm
&& test_and_clear_bit(0, &conn
->c_map_queued
)) {
185 if (conn
->c_trans
->xmit_cong_map
) {
186 conn
->c_map_offset
= 0;
187 conn
->c_map_bytes
= sizeof(struct rds_header
) +
192 rm
= rds_cong_update_alloc(conn
);
198 conn
->c_xmit_rm
= rm
;
202 * Grab the next message from the send queue, if there is one.
204 * c_xmit_rm holds a ref while we're sending this message down
205 * the connction. We can use this ref while holding the
206 * send_sem.. rds_send_reset() is serialized with it.
211 spin_lock_irqsave(&conn
->c_lock
, flags
);
213 if (!list_empty(&conn
->c_send_queue
)) {
214 rm
= list_entry(conn
->c_send_queue
.next
,
217 rds_message_addref(rm
);
220 * Move the message from the send queue to the retransmit
223 list_move_tail(&rm
->m_conn_item
, &conn
->c_retrans
);
226 spin_unlock_irqrestore(&conn
->c_lock
, flags
);
233 /* Unfortunately, the way Infiniband deals with
234 * RDMA to a bad MR key is by moving the entire
235 * queue pair to error state. We cold possibly
236 * recover from that, but right now we drop the
238 * Therefore, we never retransmit messages with RDMA ops.
240 if (rm
->rdma
.op_active
&&
241 test_bit(RDS_MSG_RETRANSMITTED
, &rm
->m_flags
)) {
242 spin_lock_irqsave(&conn
->c_lock
, flags
);
243 if (test_and_clear_bit(RDS_MSG_ON_CONN
, &rm
->m_flags
))
244 list_move(&rm
->m_conn_item
, &to_be_dropped
);
245 spin_unlock_irqrestore(&conn
->c_lock
, flags
);
250 /* Require an ACK every once in a while */
251 len
= ntohl(rm
->m_inc
.i_hdr
.h_len
);
252 if (conn
->c_unacked_packets
== 0 ||
253 conn
->c_unacked_bytes
< len
) {
254 __set_bit(RDS_MSG_ACK_REQUIRED
, &rm
->m_flags
);
256 conn
->c_unacked_packets
= rds_sysctl_max_unacked_packets
;
257 conn
->c_unacked_bytes
= rds_sysctl_max_unacked_bytes
;
258 rds_stats_inc(s_send_ack_required
);
260 conn
->c_unacked_bytes
-= len
;
261 conn
->c_unacked_packets
--;
264 conn
->c_xmit_rm
= rm
;
268 if (rm
->atomic
.op_active
&& !conn
->c_xmit_atomic_sent
) {
269 ret
= conn
->c_trans
->xmit_atomic(conn
, rm
);
272 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 * Try and send an rdma message. Let's see if we can
280 * keep this simple and require that the transport either
281 * send the whole rdma or none of it.
283 if (rm
->rdma
.op_active
&& !conn
->c_xmit_rdma_sent
) {
284 ret
= conn
->c_trans
->xmit_rdma(conn
, &rm
->rdma
);
287 conn
->c_xmit_rdma_sent
= 1;
289 /* rdmas need data sent, even if just the header */
290 rm
->data
.op_active
= 1;
292 /* The transport owns the mapped memory for now.
293 * You can't unmap it while it's on the send queue */
294 set_bit(RDS_MSG_MAPPED
, &rm
->m_flags
);
297 if (rm
->data
.op_active
298 && (conn
->c_xmit_hdr_off
< sizeof(struct rds_header
) ||
299 conn
->c_xmit_sg
< rm
->data
.op_nents
)) {
300 ret
= conn
->c_trans
->xmit(conn
, rm
,
301 conn
->c_xmit_hdr_off
,
303 conn
->c_xmit_data_off
);
307 if (conn
->c_xmit_hdr_off
< sizeof(struct rds_header
)) {
308 tmp
= min_t(int, ret
,
309 sizeof(struct rds_header
) -
310 conn
->c_xmit_hdr_off
);
311 conn
->c_xmit_hdr_off
+= tmp
;
315 sg
= &rm
->data
.op_sg
[conn
->c_xmit_sg
];
317 tmp
= min_t(int, ret
, sg
->length
-
318 conn
->c_xmit_data_off
);
319 conn
->c_xmit_data_off
+= tmp
;
321 if (conn
->c_xmit_data_off
== sg
->length
) {
322 conn
->c_xmit_data_off
= 0;
326 conn
->c_xmit_sg
== rm
->data
.op_nents
);
332 /* Nuke any messages we decided not to retransmit. */
333 if (!list_empty(&to_be_dropped
))
334 rds_send_remove_from_sock(&to_be_dropped
, RDS_RDMA_DROPPED
);
336 if (conn
->c_trans
->xmit_complete
)
337 conn
->c_trans
->xmit_complete(conn
);
340 * We might be racing with another sender who queued a message but
341 * backed off on noticing that we held the c_send_lock. If we check
342 * for queued messages after dropping the sem then either we'll
343 * see the queued message or the queuer will get the sem. If we
344 * notice the queued message then we trigger an immediate retry.
346 * We need to be careful only to do this when we stopped processing
347 * the send queue because it was empty. It's the only way we
348 * stop processing the loop when the transport hasn't taken
349 * responsibility for forward progress.
351 mutex_unlock(&conn
->c_send_lock
);
353 if (conn
->c_map_bytes
|| (send_quota
== 0 && !was_empty
)) {
354 /* We exhausted the send quota, but there's work left to
355 * do. Return and (re-)schedule the send worker.
360 if (ret
== 0 && was_empty
) {
361 /* A simple bit test would be way faster than taking the
363 spin_lock_irqsave(&conn
->c_lock
, flags
);
364 if (!list_empty(&conn
->c_send_queue
)) {
365 rds_stats_inc(s_send_sem_queue_raced
);
368 spin_unlock_irqrestore(&conn
->c_lock
, flags
);
374 static void rds_send_sndbuf_remove(struct rds_sock
*rs
, struct rds_message
*rm
)
376 u32 len
= be32_to_cpu(rm
->m_inc
.i_hdr
.h_len
);
378 assert_spin_locked(&rs
->rs_lock
);
380 BUG_ON(rs
->rs_snd_bytes
< len
);
381 rs
->rs_snd_bytes
-= len
;
383 if (rs
->rs_snd_bytes
== 0)
384 rds_stats_inc(s_send_queue_empty
);
387 static inline int rds_send_is_acked(struct rds_message
*rm
, u64 ack
,
388 is_acked_func is_acked
)
391 return is_acked(rm
, ack
);
392 return be64_to_cpu(rm
->m_inc
.i_hdr
.h_sequence
) <= ack
;
396 * Returns true if there are no messages on the send and retransmit queues
397 * which have a sequence number greater than or equal to the given sequence
400 int rds_send_acked_before(struct rds_connection
*conn
, u64 seq
)
402 struct rds_message
*rm
, *tmp
;
405 spin_lock(&conn
->c_lock
);
407 list_for_each_entry_safe(rm
, tmp
, &conn
->c_retrans
, m_conn_item
) {
408 if (be64_to_cpu(rm
->m_inc
.i_hdr
.h_sequence
) < seq
)
413 list_for_each_entry_safe(rm
, tmp
, &conn
->c_send_queue
, m_conn_item
) {
414 if (be64_to_cpu(rm
->m_inc
.i_hdr
.h_sequence
) < seq
)
419 spin_unlock(&conn
->c_lock
);
425 * This is pretty similar to what happens below in the ACK
426 * handling code - except that we call here as soon as we get
427 * the IB send completion on the RDMA op and the accompanying
430 void rds_rdma_send_complete(struct rds_message
*rm
, int status
)
432 struct rds_sock
*rs
= NULL
;
433 struct rm_rdma_op
*ro
;
434 struct rds_notifier
*notifier
;
437 spin_lock_irqsave(&rm
->m_rs_lock
, flags
);
440 if (test_bit(RDS_MSG_ON_SOCK
, &rm
->m_flags
) &&
441 ro
->op_active
&& ro
->op_notify
&& ro
->op_notifier
) {
442 notifier
= ro
->op_notifier
;
444 sock_hold(rds_rs_to_sk(rs
));
446 notifier
->n_status
= status
;
447 spin_lock(&rs
->rs_lock
);
448 list_add_tail(¬ifier
->n_list
, &rs
->rs_notify_queue
);
449 spin_unlock(&rs
->rs_lock
);
451 ro
->op_notifier
= NULL
;
454 spin_unlock_irqrestore(&rm
->m_rs_lock
, flags
);
457 rds_wake_sk_sleep(rs
);
458 sock_put(rds_rs_to_sk(rs
));
461 EXPORT_SYMBOL_GPL(rds_rdma_send_complete
);
464 * Just like above, except looks at atomic op
466 void rds_atomic_send_complete(struct rds_message
*rm
, int status
)
468 struct rds_sock
*rs
= NULL
;
469 struct rm_atomic_op
*ao
;
470 struct rds_notifier
*notifier
;
472 spin_lock(&rm
->m_rs_lock
);
475 if (test_bit(RDS_MSG_ON_SOCK
, &rm
->m_flags
)
476 && ao
->op_active
&& ao
->op_notify
&& ao
->op_notifier
) {
477 notifier
= ao
->op_notifier
;
479 sock_hold(rds_rs_to_sk(rs
));
481 notifier
->n_status
= status
;
482 spin_lock(&rs
->rs_lock
);
483 list_add_tail(¬ifier
->n_list
, &rs
->rs_notify_queue
);
484 spin_unlock(&rs
->rs_lock
);
486 ao
->op_notifier
= NULL
;
489 spin_unlock(&rm
->m_rs_lock
);
492 rds_wake_sk_sleep(rs
);
493 sock_put(rds_rs_to_sk(rs
));
496 EXPORT_SYMBOL_GPL(rds_atomic_send_complete
);
499 * This is the same as rds_rdma_send_complete except we
500 * don't do any locking - we have all the ingredients (message,
501 * socket, socket lock) and can just move the notifier.
504 __rds_rdma_send_complete(struct rds_sock
*rs
, struct rds_message
*rm
, int status
)
506 struct rm_rdma_op
*ro
;
509 if (ro
->op_active
&& ro
->op_notify
&& ro
->op_notifier
) {
510 ro
->op_notifier
->n_status
= status
;
511 list_add_tail(&ro
->op_notifier
->n_list
, &rs
->rs_notify_queue
);
512 ro
->op_notifier
= NULL
;
515 /* No need to wake the app - caller does this */
519 * This is called from the IB send completion when we detect
520 * a RDMA operation that failed with remote access error.
521 * So speed is not an issue here.
523 struct rds_message
*rds_send_get_message(struct rds_connection
*conn
,
524 struct rm_rdma_op
*op
)
526 struct rds_message
*rm
, *tmp
, *found
= NULL
;
529 spin_lock_irqsave(&conn
->c_lock
, flags
);
531 list_for_each_entry_safe(rm
, tmp
, &conn
->c_retrans
, m_conn_item
) {
532 if (&rm
->rdma
== op
) {
533 atomic_inc(&rm
->m_refcount
);
539 list_for_each_entry_safe(rm
, tmp
, &conn
->c_send_queue
, m_conn_item
) {
540 if (&rm
->rdma
== op
) {
541 atomic_inc(&rm
->m_refcount
);
548 spin_unlock_irqrestore(&conn
->c_lock
, flags
);
552 EXPORT_SYMBOL_GPL(rds_send_get_message
);
555 * This removes messages from the socket's list if they're on it. The list
556 * argument must be private to the caller, we must be able to modify it
557 * without locks. The messages must have a reference held for their
558 * position on the list. This function will drop that reference after
559 * removing the messages from the 'messages' list regardless of if it found
560 * the messages on the socket list or not.
562 void rds_send_remove_from_sock(struct list_head
*messages
, int status
)
565 struct rds_sock
*rs
= NULL
;
566 struct rds_message
*rm
;
568 while (!list_empty(messages
)) {
571 rm
= list_entry(messages
->next
, struct rds_message
,
573 list_del_init(&rm
->m_conn_item
);
576 * If we see this flag cleared then we're *sure* that someone
577 * else beat us to removing it from the sock. If we race
578 * with their flag update we'll get the lock and then really
579 * see that the flag has been cleared.
581 * The message spinlock makes sure nobody clears rm->m_rs
582 * while we're messing with it. It does not prevent the
583 * message from being removed from the socket, though.
585 spin_lock_irqsave(&rm
->m_rs_lock
, flags
);
586 if (!test_bit(RDS_MSG_ON_SOCK
, &rm
->m_flags
))
587 goto unlock_and_drop
;
589 if (rs
!= rm
->m_rs
) {
591 rds_wake_sk_sleep(rs
);
592 sock_put(rds_rs_to_sk(rs
));
595 sock_hold(rds_rs_to_sk(rs
));
597 spin_lock(&rs
->rs_lock
);
599 if (test_and_clear_bit(RDS_MSG_ON_SOCK
, &rm
->m_flags
)) {
600 struct rm_rdma_op
*ro
= &rm
->rdma
;
601 struct rds_notifier
*notifier
;
603 list_del_init(&rm
->m_sock_item
);
604 rds_send_sndbuf_remove(rs
, rm
);
606 if (ro
->op_active
&& ro
->op_notifier
&&
607 (ro
->op_notify
|| (ro
->op_recverr
&& status
))) {
608 notifier
= ro
->op_notifier
;
609 list_add_tail(¬ifier
->n_list
,
610 &rs
->rs_notify_queue
);
611 if (!notifier
->n_status
)
612 notifier
->n_status
= status
;
613 rm
->rdma
.op_notifier
= NULL
;
618 spin_unlock(&rs
->rs_lock
);
621 spin_unlock_irqrestore(&rm
->m_rs_lock
, flags
);
628 rds_wake_sk_sleep(rs
);
629 sock_put(rds_rs_to_sk(rs
));
634 * Transports call here when they've determined that the receiver queued
635 * messages up to, and including, the given sequence number. Messages are
636 * moved to the retrans queue when rds_send_xmit picks them off the send
637 * queue. This means that in the TCP case, the message may not have been
638 * assigned the m_ack_seq yet - but that's fine as long as tcp_is_acked
639 * checks the RDS_MSG_HAS_ACK_SEQ bit.
641 * XXX It's not clear to me how this is safely serialized with socket
642 * destruction. Maybe it should bail if it sees SOCK_DEAD.
644 void rds_send_drop_acked(struct rds_connection
*conn
, u64 ack
,
645 is_acked_func is_acked
)
647 struct rds_message
*rm
, *tmp
;
651 spin_lock_irqsave(&conn
->c_lock
, flags
);
653 list_for_each_entry_safe(rm
, tmp
, &conn
->c_retrans
, m_conn_item
) {
654 if (!rds_send_is_acked(rm
, ack
, is_acked
))
657 list_move(&rm
->m_conn_item
, &list
);
658 clear_bit(RDS_MSG_ON_CONN
, &rm
->m_flags
);
661 /* order flag updates with spin locks */
662 if (!list_empty(&list
))
663 smp_mb__after_clear_bit();
665 spin_unlock_irqrestore(&conn
->c_lock
, flags
);
667 /* now remove the messages from the sock list as needed */
668 rds_send_remove_from_sock(&list
, RDS_RDMA_SUCCESS
);
670 EXPORT_SYMBOL_GPL(rds_send_drop_acked
);
672 void rds_send_drop_to(struct rds_sock
*rs
, struct sockaddr_in
*dest
)
674 struct rds_message
*rm
, *tmp
;
675 struct rds_connection
*conn
;
679 /* get all the messages we're dropping under the rs lock */
680 spin_lock_irqsave(&rs
->rs_lock
, flags
);
682 list_for_each_entry_safe(rm
, tmp
, &rs
->rs_send_queue
, m_sock_item
) {
683 if (dest
&& (dest
->sin_addr
.s_addr
!= rm
->m_daddr
||
684 dest
->sin_port
!= rm
->m_inc
.i_hdr
.h_dport
))
687 list_move(&rm
->m_sock_item
, &list
);
688 rds_send_sndbuf_remove(rs
, rm
);
689 clear_bit(RDS_MSG_ON_SOCK
, &rm
->m_flags
);
692 /* order flag updates with the rs lock */
693 smp_mb__after_clear_bit();
695 spin_unlock_irqrestore(&rs
->rs_lock
, flags
);
697 if (list_empty(&list
))
700 /* Remove the messages from the conn */
701 list_for_each_entry(rm
, &list
, m_sock_item
) {
703 conn
= rm
->m_inc
.i_conn
;
705 spin_lock_irqsave(&conn
->c_lock
, flags
);
707 * Maybe someone else beat us to removing rm from the conn.
708 * If we race with their flag update we'll get the lock and
709 * then really see that the flag has been cleared.
711 if (!test_and_clear_bit(RDS_MSG_ON_CONN
, &rm
->m_flags
)) {
712 spin_unlock_irqrestore(&conn
->c_lock
, flags
);
715 list_del_init(&rm
->m_conn_item
);
716 spin_unlock_irqrestore(&conn
->c_lock
, flags
);
719 * Couldn't grab m_rs_lock in top loop (lock ordering),
722 spin_lock_irqsave(&rm
->m_rs_lock
, flags
);
724 spin_lock(&rs
->rs_lock
);
725 __rds_rdma_send_complete(rs
, rm
, RDS_RDMA_CANCELED
);
726 spin_unlock(&rs
->rs_lock
);
729 spin_unlock_irqrestore(&rm
->m_rs_lock
, flags
);
734 rds_wake_sk_sleep(rs
);
736 while (!list_empty(&list
)) {
737 rm
= list_entry(list
.next
, struct rds_message
, m_sock_item
);
738 list_del_init(&rm
->m_sock_item
);
740 rds_message_wait(rm
);
746 * we only want this to fire once so we use the callers 'queued'. It's
747 * possible that another thread can race with us and remove the
748 * message from the flow with RDS_CANCEL_SENT_TO.
750 static int rds_send_queue_rm(struct rds_sock
*rs
, struct rds_connection
*conn
,
751 struct rds_message
*rm
, __be16 sport
,
752 __be16 dport
, int *queued
)
760 len
= be32_to_cpu(rm
->m_inc
.i_hdr
.h_len
);
762 /* this is the only place which holds both the socket's rs_lock
763 * and the connection's c_lock */
764 spin_lock_irqsave(&rs
->rs_lock
, flags
);
767 * If there is a little space in sndbuf, we don't queue anything,
768 * and userspace gets -EAGAIN. But poll() indicates there's send
769 * room. This can lead to bad behavior (spinning) if snd_bytes isn't
770 * freed up by incoming acks. So we check the *old* value of
771 * rs_snd_bytes here to allow the last msg to exceed the buffer,
772 * and poll() now knows no more data can be sent.
774 if (rs
->rs_snd_bytes
< rds_sk_sndbuf(rs
)) {
775 rs
->rs_snd_bytes
+= len
;
777 /* let recv side know we are close to send space exhaustion.
778 * This is probably not the optimal way to do it, as this
779 * means we set the flag on *all* messages as soon as our
780 * throughput hits a certain threshold.
782 if (rs
->rs_snd_bytes
>= rds_sk_sndbuf(rs
) / 2)
783 __set_bit(RDS_MSG_ACK_REQUIRED
, &rm
->m_flags
);
785 list_add_tail(&rm
->m_sock_item
, &rs
->rs_send_queue
);
786 set_bit(RDS_MSG_ON_SOCK
, &rm
->m_flags
);
787 rds_message_addref(rm
);
790 /* The code ordering is a little weird, but we're
791 trying to minimize the time we hold c_lock */
792 rds_message_populate_header(&rm
->m_inc
.i_hdr
, sport
, dport
, 0);
793 rm
->m_inc
.i_conn
= conn
;
794 rds_message_addref(rm
);
796 spin_lock(&conn
->c_lock
);
797 rm
->m_inc
.i_hdr
.h_sequence
= cpu_to_be64(conn
->c_next_tx_seq
++);
798 list_add_tail(&rm
->m_conn_item
, &conn
->c_send_queue
);
799 set_bit(RDS_MSG_ON_CONN
, &rm
->m_flags
);
800 spin_unlock(&conn
->c_lock
);
802 rdsdebug("queued msg %p len %d, rs %p bytes %d seq %llu\n",
803 rm
, len
, rs
, rs
->rs_snd_bytes
,
804 (unsigned long long)be64_to_cpu(rm
->m_inc
.i_hdr
.h_sequence
));
809 spin_unlock_irqrestore(&rs
->rs_lock
, flags
);
815 * rds_message is getting to be quite complicated, and we'd like to allocate
816 * it all in one go. This figures out how big it needs to be up front.
818 static int rds_rm_size(struct msghdr
*msg
, int data_len
)
820 struct cmsghdr
*cmsg
;
824 for (cmsg
= CMSG_FIRSTHDR(msg
); cmsg
; cmsg
= CMSG_NXTHDR(msg
, cmsg
)) {
825 if (!CMSG_OK(msg
, cmsg
))
828 if (cmsg
->cmsg_level
!= SOL_RDS
)
831 switch (cmsg
->cmsg_type
) {
832 case RDS_CMSG_RDMA_ARGS
:
833 retval
= rds_rdma_extra_size(CMSG_DATA(cmsg
));
839 case RDS_CMSG_RDMA_DEST
:
840 case RDS_CMSG_RDMA_MAP
:
841 /* these are valid but do no add any size */
844 case RDS_CMSG_ATOMIC_CSWP
:
845 case RDS_CMSG_ATOMIC_FADD
:
846 size
+= sizeof(struct scatterlist
);
855 size
+= ceil(data_len
, PAGE_SIZE
) * sizeof(struct scatterlist
);
860 static int rds_cmsg_send(struct rds_sock
*rs
, struct rds_message
*rm
,
861 struct msghdr
*msg
, int *allocated_mr
)
863 struct cmsghdr
*cmsg
;
866 for (cmsg
= CMSG_FIRSTHDR(msg
); cmsg
; cmsg
= CMSG_NXTHDR(msg
, cmsg
)) {
867 if (!CMSG_OK(msg
, cmsg
))
870 if (cmsg
->cmsg_level
!= SOL_RDS
)
873 /* As a side effect, RDMA_DEST and RDMA_MAP will set
874 * rm->rdma.m_rdma_cookie and rm->rdma.m_rdma_mr.
876 switch (cmsg
->cmsg_type
) {
877 case RDS_CMSG_RDMA_ARGS
:
878 ret
= rds_cmsg_rdma_args(rs
, rm
, cmsg
);
881 case RDS_CMSG_RDMA_DEST
:
882 ret
= rds_cmsg_rdma_dest(rs
, rm
, cmsg
);
885 case RDS_CMSG_RDMA_MAP
:
886 ret
= rds_cmsg_rdma_map(rs
, rm
, cmsg
);
890 case RDS_CMSG_ATOMIC_CSWP
:
891 case RDS_CMSG_ATOMIC_FADD
:
892 ret
= rds_cmsg_atomic(rs
, rm
, cmsg
);
906 int rds_sendmsg(struct kiocb
*iocb
, struct socket
*sock
, struct msghdr
*msg
,
909 struct sock
*sk
= sock
->sk
;
910 struct rds_sock
*rs
= rds_sk_to_rs(sk
);
911 struct sockaddr_in
*usin
= (struct sockaddr_in
*)msg
->msg_name
;
914 struct rds_message
*rm
= NULL
;
915 struct rds_connection
*conn
;
917 int queued
= 0, allocated_mr
= 0;
918 int nonblock
= msg
->msg_flags
& MSG_DONTWAIT
;
919 long timeo
= sock_sndtimeo(sk
, nonblock
);
921 /* Mirror Linux UDP mirror of BSD error message compatibility */
922 /* XXX: Perhaps MSG_MORE someday */
923 if (msg
->msg_flags
& ~(MSG_DONTWAIT
| MSG_CMSG_COMPAT
)) {
924 printk(KERN_INFO
"msg_flags 0x%08X\n", msg
->msg_flags
);
929 if (msg
->msg_namelen
) {
930 /* XXX fail non-unicast destination IPs? */
931 if (msg
->msg_namelen
< sizeof(*usin
) || usin
->sin_family
!= AF_INET
) {
935 daddr
= usin
->sin_addr
.s_addr
;
936 dport
= usin
->sin_port
;
938 /* We only care about consistency with ->connect() */
940 daddr
= rs
->rs_conn_addr
;
941 dport
= rs
->rs_conn_port
;
945 /* racing with another thread binding seems ok here */
946 if (daddr
== 0 || rs
->rs_bound_addr
== 0) {
947 ret
= -ENOTCONN
; /* XXX not a great errno */
951 /* size of rm including all sgs */
952 ret
= rds_rm_size(msg
, payload_len
);
956 rm
= rds_message_alloc(ret
, GFP_KERNEL
);
962 rm
->data
.op_sg
= rds_message_alloc_sgs(rm
, ceil(payload_len
, PAGE_SIZE
));
963 /* XXX fix this to not allocate memory */
964 ret
= rds_message_copy_from_user(rm
, msg
->msg_iov
, payload_len
);
970 /* rds_conn_create has a spinlock that runs with IRQ off.
971 * Caching the conn in the socket helps a lot. */
972 if (rs
->rs_conn
&& rs
->rs_conn
->c_faddr
== daddr
)
975 conn
= rds_conn_create_outgoing(rs
->rs_bound_addr
, daddr
,
977 sock
->sk
->sk_allocation
);
985 /* Parse any control messages the user may have included. */
986 ret
= rds_cmsg_send(rs
, rm
, msg
, &allocated_mr
);
990 if ((rm
->m_rdma_cookie
|| rm
->rdma
.op_active
) &&
991 !conn
->c_trans
->xmit_rdma
) {
992 if (printk_ratelimit())
993 printk(KERN_NOTICE
"rdma_op %p conn xmit_rdma %p\n",
994 &rm
->rdma
, conn
->c_trans
->xmit_rdma
);
999 if (rm
->atomic
.op_active
&& !conn
->c_trans
->xmit_atomic
) {
1000 if (printk_ratelimit())
1001 printk(KERN_NOTICE
"atomic_op %p conn xmit_atomic %p\n",
1002 &rm
->atomic
, conn
->c_trans
->xmit_atomic
);
1007 /* If the connection is down, trigger a connect. We may
1008 * have scheduled a delayed reconnect however - in this case
1009 * we should not interfere.
1011 if (rds_conn_state(conn
) == RDS_CONN_DOWN
&&
1012 !test_and_set_bit(RDS_RECONNECT_PENDING
, &conn
->c_flags
))
1013 queue_delayed_work(rds_wq
, &conn
->c_conn_w
, 0);
1015 ret
= rds_cong_wait(conn
->c_fcong
, dport
, nonblock
, rs
);
1017 rs
->rs_seen_congestion
= 1;
1021 while (!rds_send_queue_rm(rs
, conn
, rm
, rs
->rs_bound_port
,
1023 rds_stats_inc(s_send_queue_full
);
1024 /* XXX make sure this is reasonable */
1025 if (payload_len
> rds_sk_sndbuf(rs
)) {
1034 timeo
= wait_event_interruptible_timeout(*sk_sleep(sk
),
1035 rds_send_queue_rm(rs
, conn
, rm
,
1040 rdsdebug("sendmsg woke queued %d timeo %ld\n", queued
, timeo
);
1041 if (timeo
> 0 || timeo
== MAX_SCHEDULE_TIMEOUT
)
1051 * By now we've committed to the send. We reuse rds_send_worker()
1052 * to retry sends in the rds thread if the transport asks us to.
1054 rds_stats_inc(s_send_queued
);
1056 if (!test_bit(RDS_LL_SEND_FULL
, &conn
->c_flags
))
1057 rds_send_worker(&conn
->c_send_w
.work
);
1059 rds_message_put(rm
);
1063 /* If the user included a RDMA_MAP cmsg, we allocated a MR on the fly.
1064 * If the sendmsg goes through, we keep the MR. If it fails with EAGAIN
1065 * or in any other way, we need to destroy the MR again */
1067 rds_rdma_unuse(rs
, rds_rdma_cookie_key(rm
->m_rdma_cookie
), 1);
1070 rds_message_put(rm
);
1075 * Reply to a ping packet.
1078 rds_send_pong(struct rds_connection
*conn
, __be16 dport
)
1080 struct rds_message
*rm
;
1081 unsigned long flags
;
1084 rm
= rds_message_alloc(0, GFP_ATOMIC
);
1090 rm
->m_daddr
= conn
->c_faddr
;
1092 /* If the connection is down, trigger a connect. We may
1093 * have scheduled a delayed reconnect however - in this case
1094 * we should not interfere.
1096 if (rds_conn_state(conn
) == RDS_CONN_DOWN
&&
1097 !test_and_set_bit(RDS_RECONNECT_PENDING
, &conn
->c_flags
))
1098 queue_delayed_work(rds_wq
, &conn
->c_conn_w
, 0);
1100 ret
= rds_cong_wait(conn
->c_fcong
, dport
, 1, NULL
);
1104 spin_lock_irqsave(&conn
->c_lock
, flags
);
1105 list_add_tail(&rm
->m_conn_item
, &conn
->c_send_queue
);
1106 set_bit(RDS_MSG_ON_CONN
, &rm
->m_flags
);
1107 rds_message_addref(rm
);
1108 rm
->m_inc
.i_conn
= conn
;
1110 rds_message_populate_header(&rm
->m_inc
.i_hdr
, 0, dport
,
1111 conn
->c_next_tx_seq
);
1112 conn
->c_next_tx_seq
++;
1113 spin_unlock_irqrestore(&conn
->c_lock
, flags
);
1115 rds_stats_inc(s_send_queued
);
1116 rds_stats_inc(s_send_pong
);
1118 queue_delayed_work(rds_wq
, &conn
->c_send_w
, 0);
1119 rds_message_put(rm
);
1124 rds_message_put(rm
);