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 spin_lock_irqsave(&conn
->c_send_lock
, flags
);
64 if (conn
->c_xmit_rm
) {
66 conn
->c_xmit_rm
= NULL
;
67 /* Tell the user the RDMA op is no longer mapped by the
68 * transport. This isn't entirely true (it's flushed out
69 * independently) but as the connection is down, there's
70 * no ongoing RDMA to/from that memory */
71 rds_message_unmapped(rm
);
72 spin_unlock_irqrestore(&conn
->c_send_lock
, flags
);
76 spin_unlock_irqrestore(&conn
->c_send_lock
, flags
);
80 conn
->c_xmit_hdr_off
= 0;
81 conn
->c_xmit_data_off
= 0;
82 conn
->c_xmit_atomic_sent
= 0;
83 conn
->c_xmit_rdma_sent
= 0;
84 conn
->c_xmit_data_sent
= 0;
86 conn
->c_map_queued
= 0;
88 conn
->c_unacked_packets
= rds_sysctl_max_unacked_packets
;
89 conn
->c_unacked_bytes
= rds_sysctl_max_unacked_bytes
;
91 /* Mark messages as retransmissions, and move them to the send q */
92 spin_lock_irqsave(&conn
->c_lock
, flags
);
93 list_for_each_entry_safe(rm
, tmp
, &conn
->c_retrans
, m_conn_item
) {
94 set_bit(RDS_MSG_ACK_REQUIRED
, &rm
->m_flags
);
95 set_bit(RDS_MSG_RETRANSMITTED
, &rm
->m_flags
);
97 list_splice_init(&conn
->c_retrans
, &conn
->c_send_queue
);
98 spin_unlock_irqrestore(&conn
->c_lock
, flags
);
102 * We're making the concious trade-off here to only send one message
103 * down the connection at a time.
105 * - tx queueing is a simple fifo list
106 * - reassembly is optional and easily done by transports per conn
107 * - no per flow rx lookup at all, straight to the socket
108 * - less per-frag memory and wire overhead
110 * - queued acks can be delayed behind large messages
112 * - small message latency is higher behind queued large messages
113 * - large message latency isn't starved by intervening small sends
115 int rds_send_xmit(struct rds_connection
*conn
)
117 struct rds_message
*rm
;
120 struct scatterlist
*sg
;
123 LIST_HEAD(to_be_dropped
);
126 if (!rds_conn_up(conn
))
130 * sendmsg calls here after having queued its message on the send
131 * queue. We only have one task feeding the connection at a time. If
132 * another thread is already feeding the queue then we back off. This
133 * avoids blocking the caller and trading per-connection data between
134 * caches per message.
136 if (!spin_trylock_irqsave(&conn
->c_send_lock
, flags
)) {
137 rds_stats_inc(s_send_lock_contention
);
141 atomic_inc(&conn
->c_senders
);
143 if (conn
->c_trans
->xmit_prepare
)
144 conn
->c_trans
->xmit_prepare(conn
);
146 gen
= atomic_inc_return(&conn
->c_send_generation
);
149 * spin trying to push headers and data down the connection until
150 * the connection doesn't make forward progress.
154 rm
= conn
->c_xmit_rm
;
157 * If between sending messages, we can send a pending congestion
160 if (!rm
&& test_and_clear_bit(0, &conn
->c_map_queued
)) {
161 rm
= rds_cong_update_alloc(conn
);
166 rm
->data
.op_active
= 1;
168 conn
->c_xmit_rm
= rm
;
172 * If not already working on one, grab the next message.
174 * c_xmit_rm holds a ref while we're sending this message down
175 * the connction. We can use this ref while holding the
176 * send_sem.. rds_send_reset() is serialized with it.
181 spin_lock(&conn
->c_lock
);
183 if (!list_empty(&conn
->c_send_queue
)) {
184 rm
= list_entry(conn
->c_send_queue
.next
,
187 rds_message_addref(rm
);
190 * Move the message from the send queue to the retransmit
193 list_move_tail(&rm
->m_conn_item
, &conn
->c_retrans
);
196 spin_unlock(&conn
->c_lock
);
201 /* Unfortunately, the way Infiniband deals with
202 * RDMA to a bad MR key is by moving the entire
203 * queue pair to error state. We cold possibly
204 * recover from that, but right now we drop the
206 * Therefore, we never retransmit messages with RDMA ops.
208 if (rm
->rdma
.op_active
&&
209 test_bit(RDS_MSG_RETRANSMITTED
, &rm
->m_flags
)) {
210 spin_lock(&conn
->c_lock
);
211 if (test_and_clear_bit(RDS_MSG_ON_CONN
, &rm
->m_flags
))
212 list_move(&rm
->m_conn_item
, &to_be_dropped
);
213 spin_unlock(&conn
->c_lock
);
217 /* Require an ACK every once in a while */
218 len
= ntohl(rm
->m_inc
.i_hdr
.h_len
);
219 if (conn
->c_unacked_packets
== 0 ||
220 conn
->c_unacked_bytes
< len
) {
221 __set_bit(RDS_MSG_ACK_REQUIRED
, &rm
->m_flags
);
223 conn
->c_unacked_packets
= rds_sysctl_max_unacked_packets
;
224 conn
->c_unacked_bytes
= rds_sysctl_max_unacked_bytes
;
225 rds_stats_inc(s_send_ack_required
);
227 conn
->c_unacked_bytes
-= len
;
228 conn
->c_unacked_packets
--;
231 conn
->c_xmit_rm
= rm
;
234 /* The transport either sends the whole rdma or none of it */
235 if (rm
->rdma
.op_active
&& !conn
->c_xmit_rdma_sent
) {
236 rm
->m_final_op
= &rm
->rdma
;
237 ret
= conn
->c_trans
->xmit_rdma(conn
, &rm
->rdma
);
240 conn
->c_xmit_rdma_sent
= 1;
242 /* The transport owns the mapped memory for now.
243 * You can't unmap it while it's on the send queue */
244 set_bit(RDS_MSG_MAPPED
, &rm
->m_flags
);
247 if (rm
->atomic
.op_active
&& !conn
->c_xmit_atomic_sent
) {
248 rm
->m_final_op
= &rm
->atomic
;
249 ret
= conn
->c_trans
->xmit_atomic(conn
, &rm
->atomic
);
252 conn
->c_xmit_atomic_sent
= 1;
254 /* The transport owns the mapped memory for now.
255 * You can't unmap it while it's on the send queue */
256 set_bit(RDS_MSG_MAPPED
, &rm
->m_flags
);
260 * A number of cases require an RDS header to be sent
261 * even if there is no data.
262 * We permit 0-byte sends; rds-ping depends on this.
263 * However, if there are exclusively attached silent ops,
264 * we skip the hdr/data send, to enable silent operation.
266 if (rm
->data
.op_nents
== 0) {
268 int all_ops_are_silent
= 1;
270 ops_present
= (rm
->atomic
.op_active
|| rm
->rdma
.op_active
);
271 if (rm
->atomic
.op_active
&& !rm
->atomic
.op_silent
)
272 all_ops_are_silent
= 0;
273 if (rm
->rdma
.op_active
&& !rm
->rdma
.op_silent
)
274 all_ops_are_silent
= 0;
276 if (ops_present
&& all_ops_are_silent
277 && !rm
->m_rdma_cookie
)
278 rm
->data
.op_active
= 0;
281 if (rm
->data
.op_active
&& !conn
->c_xmit_data_sent
) {
282 rm
->m_final_op
= &rm
->data
;
283 ret
= conn
->c_trans
->xmit(conn
, rm
,
284 conn
->c_xmit_hdr_off
,
286 conn
->c_xmit_data_off
);
290 if (conn
->c_xmit_hdr_off
< sizeof(struct rds_header
)) {
291 tmp
= min_t(int, ret
,
292 sizeof(struct rds_header
) -
293 conn
->c_xmit_hdr_off
);
294 conn
->c_xmit_hdr_off
+= tmp
;
298 sg
= &rm
->data
.op_sg
[conn
->c_xmit_sg
];
300 tmp
= min_t(int, ret
, sg
->length
-
301 conn
->c_xmit_data_off
);
302 conn
->c_xmit_data_off
+= tmp
;
304 if (conn
->c_xmit_data_off
== sg
->length
) {
305 conn
->c_xmit_data_off
= 0;
309 conn
->c_xmit_sg
== rm
->data
.op_nents
);
313 if (conn
->c_xmit_hdr_off
== sizeof(struct rds_header
) &&
314 (conn
->c_xmit_sg
== rm
->data
.op_nents
))
315 conn
->c_xmit_data_sent
= 1;
319 * A rm will only take multiple times through this loop
320 * if there is a data op. Thus, if the data is sent (or there was
321 * none), then we're done with the rm.
323 if (!rm
->data
.op_active
|| conn
->c_xmit_data_sent
) {
324 conn
->c_xmit_rm
= NULL
;
326 conn
->c_xmit_hdr_off
= 0;
327 conn
->c_xmit_data_off
= 0;
328 conn
->c_xmit_rdma_sent
= 0;
329 conn
->c_xmit_atomic_sent
= 0;
330 conn
->c_xmit_data_sent
= 0;
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 spin_unlock_irqrestore(&conn
->c_send_lock
, flags
);
353 /* Nuke any messages we decided not to retransmit. */
354 if (!list_empty(&to_be_dropped
)) {
355 /* irqs on here, so we can put(), unlike above */
356 list_for_each_entry(rm
, &to_be_dropped
, m_conn_item
)
358 rds_send_remove_from_sock(&to_be_dropped
, RDS_RDMA_DROPPED
);
361 atomic_dec(&conn
->c_senders
);
364 * Other senders will see we have c_send_lock and exit. We
365 * need to recheck the send queue and race again for c_send_lock
366 * to make sure messages don't just sit on the send queue, if
367 * somebody hasn't already beat us into the loop.
369 * If the transport cannot continue (i.e ret != 0), then it must
370 * call us when more room is available, such as from the tx
371 * completion handler.
375 if (!list_empty(&conn
->c_send_queue
)) {
376 rds_stats_inc(s_send_lock_queue_raced
);
377 if (gen
== atomic_read(&conn
->c_send_generation
)) {
386 static void rds_send_sndbuf_remove(struct rds_sock
*rs
, struct rds_message
*rm
)
388 u32 len
= be32_to_cpu(rm
->m_inc
.i_hdr
.h_len
);
390 assert_spin_locked(&rs
->rs_lock
);
392 BUG_ON(rs
->rs_snd_bytes
< len
);
393 rs
->rs_snd_bytes
-= len
;
395 if (rs
->rs_snd_bytes
== 0)
396 rds_stats_inc(s_send_queue_empty
);
399 static inline int rds_send_is_acked(struct rds_message
*rm
, u64 ack
,
400 is_acked_func is_acked
)
403 return is_acked(rm
, ack
);
404 return be64_to_cpu(rm
->m_inc
.i_hdr
.h_sequence
) <= ack
;
408 * Returns true if there are no messages on the send and retransmit queues
409 * which have a sequence number greater than or equal to the given sequence
412 int rds_send_acked_before(struct rds_connection
*conn
, u64 seq
)
414 struct rds_message
*rm
, *tmp
;
417 spin_lock(&conn
->c_lock
);
419 list_for_each_entry_safe(rm
, tmp
, &conn
->c_retrans
, m_conn_item
) {
420 if (be64_to_cpu(rm
->m_inc
.i_hdr
.h_sequence
) < seq
)
425 list_for_each_entry_safe(rm
, tmp
, &conn
->c_send_queue
, m_conn_item
) {
426 if (be64_to_cpu(rm
->m_inc
.i_hdr
.h_sequence
) < seq
)
431 spin_unlock(&conn
->c_lock
);
437 * This is pretty similar to what happens below in the ACK
438 * handling code - except that we call here as soon as we get
439 * the IB send completion on the RDMA op and the accompanying
442 void rds_rdma_send_complete(struct rds_message
*rm
, int status
)
444 struct rds_sock
*rs
= NULL
;
445 struct rm_rdma_op
*ro
;
446 struct rds_notifier
*notifier
;
449 spin_lock_irqsave(&rm
->m_rs_lock
, flags
);
452 if (test_bit(RDS_MSG_ON_SOCK
, &rm
->m_flags
) &&
453 ro
->op_active
&& ro
->op_notify
&& ro
->op_notifier
) {
454 notifier
= ro
->op_notifier
;
456 sock_hold(rds_rs_to_sk(rs
));
458 notifier
->n_status
= status
;
459 spin_lock(&rs
->rs_lock
);
460 list_add_tail(¬ifier
->n_list
, &rs
->rs_notify_queue
);
461 spin_unlock(&rs
->rs_lock
);
463 ro
->op_notifier
= NULL
;
466 spin_unlock_irqrestore(&rm
->m_rs_lock
, flags
);
469 rds_wake_sk_sleep(rs
);
470 sock_put(rds_rs_to_sk(rs
));
473 EXPORT_SYMBOL_GPL(rds_rdma_send_complete
);
476 * Just like above, except looks at atomic op
478 void rds_atomic_send_complete(struct rds_message
*rm
, int status
)
480 struct rds_sock
*rs
= NULL
;
481 struct rm_atomic_op
*ao
;
482 struct rds_notifier
*notifier
;
485 spin_lock_irqsave(&rm
->m_rs_lock
, flags
);
488 if (test_bit(RDS_MSG_ON_SOCK
, &rm
->m_flags
)
489 && ao
->op_active
&& ao
->op_notify
&& ao
->op_notifier
) {
490 notifier
= ao
->op_notifier
;
492 sock_hold(rds_rs_to_sk(rs
));
494 notifier
->n_status
= status
;
495 spin_lock(&rs
->rs_lock
);
496 list_add_tail(¬ifier
->n_list
, &rs
->rs_notify_queue
);
497 spin_unlock(&rs
->rs_lock
);
499 ao
->op_notifier
= NULL
;
502 spin_unlock_irqrestore(&rm
->m_rs_lock
, flags
);
505 rds_wake_sk_sleep(rs
);
506 sock_put(rds_rs_to_sk(rs
));
509 EXPORT_SYMBOL_GPL(rds_atomic_send_complete
);
512 * This is the same as rds_rdma_send_complete except we
513 * don't do any locking - we have all the ingredients (message,
514 * socket, socket lock) and can just move the notifier.
517 __rds_send_complete(struct rds_sock
*rs
, struct rds_message
*rm
, int status
)
519 struct rm_rdma_op
*ro
;
520 struct rm_atomic_op
*ao
;
523 if (ro
->op_active
&& ro
->op_notify
&& ro
->op_notifier
) {
524 ro
->op_notifier
->n_status
= status
;
525 list_add_tail(&ro
->op_notifier
->n_list
, &rs
->rs_notify_queue
);
526 ro
->op_notifier
= NULL
;
530 if (ao
->op_active
&& ao
->op_notify
&& ao
->op_notifier
) {
531 ao
->op_notifier
->n_status
= status
;
532 list_add_tail(&ao
->op_notifier
->n_list
, &rs
->rs_notify_queue
);
533 ao
->op_notifier
= NULL
;
536 /* No need to wake the app - caller does this */
540 * This is called from the IB send completion when we detect
541 * a RDMA operation that failed with remote access error.
542 * So speed is not an issue here.
544 struct rds_message
*rds_send_get_message(struct rds_connection
*conn
,
545 struct rm_rdma_op
*op
)
547 struct rds_message
*rm
, *tmp
, *found
= NULL
;
550 spin_lock_irqsave(&conn
->c_lock
, flags
);
552 list_for_each_entry_safe(rm
, tmp
, &conn
->c_retrans
, m_conn_item
) {
553 if (&rm
->rdma
== op
) {
554 atomic_inc(&rm
->m_refcount
);
560 list_for_each_entry_safe(rm
, tmp
, &conn
->c_send_queue
, m_conn_item
) {
561 if (&rm
->rdma
== op
) {
562 atomic_inc(&rm
->m_refcount
);
569 spin_unlock_irqrestore(&conn
->c_lock
, flags
);
573 EXPORT_SYMBOL_GPL(rds_send_get_message
);
576 * This removes messages from the socket's list if they're on it. The list
577 * argument must be private to the caller, we must be able to modify it
578 * without locks. The messages must have a reference held for their
579 * position on the list. This function will drop that reference after
580 * removing the messages from the 'messages' list regardless of if it found
581 * the messages on the socket list or not.
583 void rds_send_remove_from_sock(struct list_head
*messages
, int status
)
586 struct rds_sock
*rs
= NULL
;
587 struct rds_message
*rm
;
589 while (!list_empty(messages
)) {
592 rm
= list_entry(messages
->next
, struct rds_message
,
594 list_del_init(&rm
->m_conn_item
);
597 * If we see this flag cleared then we're *sure* that someone
598 * else beat us to removing it from the sock. If we race
599 * with their flag update we'll get the lock and then really
600 * see that the flag has been cleared.
602 * The message spinlock makes sure nobody clears rm->m_rs
603 * while we're messing with it. It does not prevent the
604 * message from being removed from the socket, though.
606 spin_lock_irqsave(&rm
->m_rs_lock
, flags
);
607 if (!test_bit(RDS_MSG_ON_SOCK
, &rm
->m_flags
))
608 goto unlock_and_drop
;
610 if (rs
!= rm
->m_rs
) {
612 rds_wake_sk_sleep(rs
);
613 sock_put(rds_rs_to_sk(rs
));
616 sock_hold(rds_rs_to_sk(rs
));
618 spin_lock(&rs
->rs_lock
);
620 if (test_and_clear_bit(RDS_MSG_ON_SOCK
, &rm
->m_flags
)) {
621 struct rm_rdma_op
*ro
= &rm
->rdma
;
622 struct rds_notifier
*notifier
;
624 list_del_init(&rm
->m_sock_item
);
625 rds_send_sndbuf_remove(rs
, rm
);
627 if (ro
->op_active
&& ro
->op_notifier
&&
628 (ro
->op_notify
|| (ro
->op_recverr
&& status
))) {
629 notifier
= ro
->op_notifier
;
630 list_add_tail(¬ifier
->n_list
,
631 &rs
->rs_notify_queue
);
632 if (!notifier
->n_status
)
633 notifier
->n_status
= status
;
634 rm
->rdma
.op_notifier
= NULL
;
639 spin_unlock(&rs
->rs_lock
);
642 spin_unlock_irqrestore(&rm
->m_rs_lock
, flags
);
649 rds_wake_sk_sleep(rs
);
650 sock_put(rds_rs_to_sk(rs
));
655 * Transports call here when they've determined that the receiver queued
656 * messages up to, and including, the given sequence number. Messages are
657 * moved to the retrans queue when rds_send_xmit picks them off the send
658 * queue. This means that in the TCP case, the message may not have been
659 * assigned the m_ack_seq yet - but that's fine as long as tcp_is_acked
660 * checks the RDS_MSG_HAS_ACK_SEQ bit.
662 * XXX It's not clear to me how this is safely serialized with socket
663 * destruction. Maybe it should bail if it sees SOCK_DEAD.
665 void rds_send_drop_acked(struct rds_connection
*conn
, u64 ack
,
666 is_acked_func is_acked
)
668 struct rds_message
*rm
, *tmp
;
672 spin_lock_irqsave(&conn
->c_lock
, flags
);
674 list_for_each_entry_safe(rm
, tmp
, &conn
->c_retrans
, m_conn_item
) {
675 if (!rds_send_is_acked(rm
, ack
, is_acked
))
678 list_move(&rm
->m_conn_item
, &list
);
679 clear_bit(RDS_MSG_ON_CONN
, &rm
->m_flags
);
682 /* order flag updates with spin locks */
683 if (!list_empty(&list
))
684 smp_mb__after_clear_bit();
686 spin_unlock_irqrestore(&conn
->c_lock
, flags
);
688 /* now remove the messages from the sock list as needed */
689 rds_send_remove_from_sock(&list
, RDS_RDMA_SUCCESS
);
691 EXPORT_SYMBOL_GPL(rds_send_drop_acked
);
693 void rds_send_drop_to(struct rds_sock
*rs
, struct sockaddr_in
*dest
)
695 struct rds_message
*rm
, *tmp
;
696 struct rds_connection
*conn
;
700 /* get all the messages we're dropping under the rs lock */
701 spin_lock_irqsave(&rs
->rs_lock
, flags
);
703 list_for_each_entry_safe(rm
, tmp
, &rs
->rs_send_queue
, m_sock_item
) {
704 if (dest
&& (dest
->sin_addr
.s_addr
!= rm
->m_daddr
||
705 dest
->sin_port
!= rm
->m_inc
.i_hdr
.h_dport
))
708 list_move(&rm
->m_sock_item
, &list
);
709 rds_send_sndbuf_remove(rs
, rm
);
710 clear_bit(RDS_MSG_ON_SOCK
, &rm
->m_flags
);
713 /* order flag updates with the rs lock */
714 smp_mb__after_clear_bit();
716 spin_unlock_irqrestore(&rs
->rs_lock
, flags
);
718 if (list_empty(&list
))
721 /* Remove the messages from the conn */
722 list_for_each_entry(rm
, &list
, m_sock_item
) {
724 conn
= rm
->m_inc
.i_conn
;
726 spin_lock_irqsave(&conn
->c_lock
, flags
);
728 * Maybe someone else beat us to removing rm from the conn.
729 * If we race with their flag update we'll get the lock and
730 * then really see that the flag has been cleared.
732 if (!test_and_clear_bit(RDS_MSG_ON_CONN
, &rm
->m_flags
)) {
733 spin_unlock_irqrestore(&conn
->c_lock
, flags
);
736 list_del_init(&rm
->m_conn_item
);
737 spin_unlock_irqrestore(&conn
->c_lock
, flags
);
740 * Couldn't grab m_rs_lock in top loop (lock ordering),
743 spin_lock_irqsave(&rm
->m_rs_lock
, flags
);
745 spin_lock(&rs
->rs_lock
);
746 __rds_send_complete(rs
, rm
, RDS_RDMA_CANCELED
);
747 spin_unlock(&rs
->rs_lock
);
750 spin_unlock_irqrestore(&rm
->m_rs_lock
, flags
);
755 rds_wake_sk_sleep(rs
);
757 while (!list_empty(&list
)) {
758 rm
= list_entry(list
.next
, struct rds_message
, m_sock_item
);
759 list_del_init(&rm
->m_sock_item
);
761 rds_message_wait(rm
);
767 * we only want this to fire once so we use the callers 'queued'. It's
768 * possible that another thread can race with us and remove the
769 * message from the flow with RDS_CANCEL_SENT_TO.
771 static int rds_send_queue_rm(struct rds_sock
*rs
, struct rds_connection
*conn
,
772 struct rds_message
*rm
, __be16 sport
,
773 __be16 dport
, int *queued
)
781 len
= be32_to_cpu(rm
->m_inc
.i_hdr
.h_len
);
783 /* this is the only place which holds both the socket's rs_lock
784 * and the connection's c_lock */
785 spin_lock_irqsave(&rs
->rs_lock
, flags
);
788 * If there is a little space in sndbuf, we don't queue anything,
789 * and userspace gets -EAGAIN. But poll() indicates there's send
790 * room. This can lead to bad behavior (spinning) if snd_bytes isn't
791 * freed up by incoming acks. So we check the *old* value of
792 * rs_snd_bytes here to allow the last msg to exceed the buffer,
793 * and poll() now knows no more data can be sent.
795 if (rs
->rs_snd_bytes
< rds_sk_sndbuf(rs
)) {
796 rs
->rs_snd_bytes
+= len
;
798 /* let recv side know we are close to send space exhaustion.
799 * This is probably not the optimal way to do it, as this
800 * means we set the flag on *all* messages as soon as our
801 * throughput hits a certain threshold.
803 if (rs
->rs_snd_bytes
>= rds_sk_sndbuf(rs
) / 2)
804 __set_bit(RDS_MSG_ACK_REQUIRED
, &rm
->m_flags
);
806 list_add_tail(&rm
->m_sock_item
, &rs
->rs_send_queue
);
807 set_bit(RDS_MSG_ON_SOCK
, &rm
->m_flags
);
808 rds_message_addref(rm
);
811 /* The code ordering is a little weird, but we're
812 trying to minimize the time we hold c_lock */
813 rds_message_populate_header(&rm
->m_inc
.i_hdr
, sport
, dport
, 0);
814 rm
->m_inc
.i_conn
= conn
;
815 rds_message_addref(rm
);
817 spin_lock(&conn
->c_lock
);
818 rm
->m_inc
.i_hdr
.h_sequence
= cpu_to_be64(conn
->c_next_tx_seq
++);
819 list_add_tail(&rm
->m_conn_item
, &conn
->c_send_queue
);
820 set_bit(RDS_MSG_ON_CONN
, &rm
->m_flags
);
821 spin_unlock(&conn
->c_lock
);
823 rdsdebug("queued msg %p len %d, rs %p bytes %d seq %llu\n",
824 rm
, len
, rs
, rs
->rs_snd_bytes
,
825 (unsigned long long)be64_to_cpu(rm
->m_inc
.i_hdr
.h_sequence
));
830 spin_unlock_irqrestore(&rs
->rs_lock
, flags
);
836 * rds_message is getting to be quite complicated, and we'd like to allocate
837 * it all in one go. This figures out how big it needs to be up front.
839 static int rds_rm_size(struct msghdr
*msg
, int data_len
)
841 struct cmsghdr
*cmsg
;
846 for (cmsg
= CMSG_FIRSTHDR(msg
); cmsg
; cmsg
= CMSG_NXTHDR(msg
, cmsg
)) {
847 if (!CMSG_OK(msg
, cmsg
))
850 if (cmsg
->cmsg_level
!= SOL_RDS
)
853 switch (cmsg
->cmsg_type
) {
854 case RDS_CMSG_RDMA_ARGS
:
856 retval
= rds_rdma_extra_size(CMSG_DATA(cmsg
));
863 case RDS_CMSG_RDMA_DEST
:
864 case RDS_CMSG_RDMA_MAP
:
866 /* these are valid but do no add any size */
869 case RDS_CMSG_ATOMIC_CSWP
:
870 case RDS_CMSG_ATOMIC_FADD
:
872 size
+= sizeof(struct scatterlist
);
881 size
+= ceil(data_len
, PAGE_SIZE
) * sizeof(struct scatterlist
);
883 /* Ensure (DEST, MAP) are never used with (ARGS, ATOMIC) */
884 if (cmsg_groups
== 3)
890 static int rds_cmsg_send(struct rds_sock
*rs
, struct rds_message
*rm
,
891 struct msghdr
*msg
, int *allocated_mr
)
893 struct cmsghdr
*cmsg
;
896 for (cmsg
= CMSG_FIRSTHDR(msg
); cmsg
; cmsg
= CMSG_NXTHDR(msg
, cmsg
)) {
897 if (!CMSG_OK(msg
, cmsg
))
900 if (cmsg
->cmsg_level
!= SOL_RDS
)
903 /* As a side effect, RDMA_DEST and RDMA_MAP will set
904 * rm->rdma.m_rdma_cookie and rm->rdma.m_rdma_mr.
906 switch (cmsg
->cmsg_type
) {
907 case RDS_CMSG_RDMA_ARGS
:
908 ret
= rds_cmsg_rdma_args(rs
, rm
, cmsg
);
911 case RDS_CMSG_RDMA_DEST
:
912 ret
= rds_cmsg_rdma_dest(rs
, rm
, cmsg
);
915 case RDS_CMSG_RDMA_MAP
:
916 ret
= rds_cmsg_rdma_map(rs
, rm
, cmsg
);
920 case RDS_CMSG_ATOMIC_CSWP
:
921 case RDS_CMSG_ATOMIC_FADD
:
922 ret
= rds_cmsg_atomic(rs
, rm
, cmsg
);
936 int rds_sendmsg(struct kiocb
*iocb
, struct socket
*sock
, struct msghdr
*msg
,
939 struct sock
*sk
= sock
->sk
;
940 struct rds_sock
*rs
= rds_sk_to_rs(sk
);
941 struct sockaddr_in
*usin
= (struct sockaddr_in
*)msg
->msg_name
;
944 struct rds_message
*rm
= NULL
;
945 struct rds_connection
*conn
;
947 int queued
= 0, allocated_mr
= 0;
948 int nonblock
= msg
->msg_flags
& MSG_DONTWAIT
;
949 long timeo
= sock_sndtimeo(sk
, nonblock
);
951 /* Mirror Linux UDP mirror of BSD error message compatibility */
952 /* XXX: Perhaps MSG_MORE someday */
953 if (msg
->msg_flags
& ~(MSG_DONTWAIT
| MSG_CMSG_COMPAT
)) {
954 printk(KERN_INFO
"msg_flags 0x%08X\n", msg
->msg_flags
);
959 if (msg
->msg_namelen
) {
960 /* XXX fail non-unicast destination IPs? */
961 if (msg
->msg_namelen
< sizeof(*usin
) || usin
->sin_family
!= AF_INET
) {
965 daddr
= usin
->sin_addr
.s_addr
;
966 dport
= usin
->sin_port
;
968 /* We only care about consistency with ->connect() */
970 daddr
= rs
->rs_conn_addr
;
971 dport
= rs
->rs_conn_port
;
975 /* racing with another thread binding seems ok here */
976 if (daddr
== 0 || rs
->rs_bound_addr
== 0) {
977 ret
= -ENOTCONN
; /* XXX not a great errno */
981 /* size of rm including all sgs */
982 ret
= rds_rm_size(msg
, payload_len
);
986 rm
= rds_message_alloc(ret
, GFP_KERNEL
);
992 /* Attach data to the rm */
994 rm
->data
.op_sg
= rds_message_alloc_sgs(rm
, ceil(payload_len
, PAGE_SIZE
));
995 ret
= rds_message_copy_from_user(rm
, msg
->msg_iov
, payload_len
);
999 rm
->data
.op_active
= 1;
1001 rm
->m_daddr
= daddr
;
1003 /* rds_conn_create has a spinlock that runs with IRQ off.
1004 * Caching the conn in the socket helps a lot. */
1005 if (rs
->rs_conn
&& rs
->rs_conn
->c_faddr
== daddr
)
1008 conn
= rds_conn_create_outgoing(rs
->rs_bound_addr
, daddr
,
1010 sock
->sk
->sk_allocation
);
1012 ret
= PTR_ERR(conn
);
1018 /* Parse any control messages the user may have included. */
1019 ret
= rds_cmsg_send(rs
, rm
, msg
, &allocated_mr
);
1023 if (rm
->rdma
.op_active
&& !conn
->c_trans
->xmit_rdma
) {
1024 if (printk_ratelimit())
1025 printk(KERN_NOTICE
"rdma_op %p conn xmit_rdma %p\n",
1026 &rm
->rdma
, conn
->c_trans
->xmit_rdma
);
1031 if (rm
->atomic
.op_active
&& !conn
->c_trans
->xmit_atomic
) {
1032 if (printk_ratelimit())
1033 printk(KERN_NOTICE
"atomic_op %p conn xmit_atomic %p\n",
1034 &rm
->atomic
, conn
->c_trans
->xmit_atomic
);
1039 /* If the connection is down, trigger a connect. We may
1040 * have scheduled a delayed reconnect however - in this case
1041 * we should not interfere.
1043 if (rds_conn_state(conn
) == RDS_CONN_DOWN
&&
1044 !test_and_set_bit(RDS_RECONNECT_PENDING
, &conn
->c_flags
))
1045 queue_delayed_work(rds_wq
, &conn
->c_conn_w
, 0);
1047 ret
= rds_cong_wait(conn
->c_fcong
, dport
, nonblock
, rs
);
1049 rs
->rs_seen_congestion
= 1;
1053 while (!rds_send_queue_rm(rs
, conn
, rm
, rs
->rs_bound_port
,
1055 rds_stats_inc(s_send_queue_full
);
1056 /* XXX make sure this is reasonable */
1057 if (payload_len
> rds_sk_sndbuf(rs
)) {
1066 timeo
= wait_event_interruptible_timeout(*sk_sleep(sk
),
1067 rds_send_queue_rm(rs
, conn
, rm
,
1072 rdsdebug("sendmsg woke queued %d timeo %ld\n", queued
, timeo
);
1073 if (timeo
> 0 || timeo
== MAX_SCHEDULE_TIMEOUT
)
1083 * By now we've committed to the send. We reuse rds_send_worker()
1084 * to retry sends in the rds thread if the transport asks us to.
1086 rds_stats_inc(s_send_queued
);
1088 if (!test_bit(RDS_LL_SEND_FULL
, &conn
->c_flags
))
1089 rds_send_xmit(conn
);
1091 rds_message_put(rm
);
1095 /* If the user included a RDMA_MAP cmsg, we allocated a MR on the fly.
1096 * If the sendmsg goes through, we keep the MR. If it fails with EAGAIN
1097 * or in any other way, we need to destroy the MR again */
1099 rds_rdma_unuse(rs
, rds_rdma_cookie_key(rm
->m_rdma_cookie
), 1);
1102 rds_message_put(rm
);
1107 * Reply to a ping packet.
1110 rds_send_pong(struct rds_connection
*conn
, __be16 dport
)
1112 struct rds_message
*rm
;
1113 unsigned long flags
;
1116 rm
= rds_message_alloc(0, GFP_ATOMIC
);
1122 rm
->m_daddr
= conn
->c_faddr
;
1123 rm
->data
.op_active
= 1;
1125 /* If the connection is down, trigger a connect. We may
1126 * have scheduled a delayed reconnect however - in this case
1127 * we should not interfere.
1129 if (rds_conn_state(conn
) == RDS_CONN_DOWN
&&
1130 !test_and_set_bit(RDS_RECONNECT_PENDING
, &conn
->c_flags
))
1131 queue_delayed_work(rds_wq
, &conn
->c_conn_w
, 0);
1133 ret
= rds_cong_wait(conn
->c_fcong
, dport
, 1, NULL
);
1137 spin_lock_irqsave(&conn
->c_lock
, flags
);
1138 list_add_tail(&rm
->m_conn_item
, &conn
->c_send_queue
);
1139 set_bit(RDS_MSG_ON_CONN
, &rm
->m_flags
);
1140 rds_message_addref(rm
);
1141 rm
->m_inc
.i_conn
= conn
;
1143 rds_message_populate_header(&rm
->m_inc
.i_hdr
, 0, dport
,
1144 conn
->c_next_tx_seq
);
1145 conn
->c_next_tx_seq
++;
1146 spin_unlock_irqrestore(&conn
->c_lock
, flags
);
1148 rds_stats_inc(s_send_queued
);
1149 rds_stats_inc(s_send_pong
);
1151 if (!test_bit(RDS_LL_SEND_FULL
, &conn
->c_flags
))
1152 rds_send_xmit(conn
);
1154 rds_message_put(rm
);
1159 rds_message_put(rm
);