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/moduleparam.h>
35 #include <linux/gfp.h>
38 #include <linux/list.h>
39 #include <linux/ratelimit.h>
40 #include <linux/export.h>
41 #include <linux/sizes.h>
45 /* When transmitting messages in rds_send_xmit, we need to emerge from
46 * time to time and briefly release the CPU. Otherwise the softlock watchdog
48 * Also, it seems fairer to not let one busy connection stall all the
51 * send_batch_count is the number of times we'll loop in send_xmit. Setting
52 * it to 0 will restore the old behavior (where we looped until we had
55 static int send_batch_count = SZ_1K;
56 module_param(send_batch_count, int, 0444);
57 MODULE_PARM_DESC(send_batch_count, " batch factor when working the send queue");
59 static void rds_send_remove_from_sock(struct list_head *messages, int status);
62 * Reset the send state. Callers must ensure that this doesn't race with
65 void rds_send_reset(struct rds_connection *conn)
67 struct rds_message *rm, *tmp;
70 if (conn->c_xmit_rm) {
72 conn->c_xmit_rm = NULL;
73 /* Tell the user the RDMA op is no longer mapped by the
74 * transport. This isn't entirely true (it's flushed out
75 * independently) but as the connection is down, there's
76 * no ongoing RDMA to/from that memory */
77 rds_message_unmapped(rm);
82 conn->c_xmit_hdr_off = 0;
83 conn->c_xmit_data_off = 0;
84 conn->c_xmit_atomic_sent = 0;
85 conn->c_xmit_rdma_sent = 0;
86 conn->c_xmit_data_sent = 0;
88 conn->c_map_queued = 0;
90 conn->c_unacked_packets = rds_sysctl_max_unacked_packets;
91 conn->c_unacked_bytes = rds_sysctl_max_unacked_bytes;
93 /* Mark messages as retransmissions, and move them to the send q */
94 spin_lock_irqsave(&conn->c_lock, flags);
95 list_for_each_entry_safe(rm, tmp, &conn->c_retrans, m_conn_item) {
96 set_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags);
97 set_bit(RDS_MSG_RETRANSMITTED, &rm->m_flags);
99 list_splice_init(&conn->c_retrans, &conn->c_send_queue);
100 spin_unlock_irqrestore(&conn->c_lock, flags);
102 EXPORT_SYMBOL_GPL(rds_send_reset);
104 static int acquire_in_xmit(struct rds_connection *conn)
106 return test_and_set_bit(RDS_IN_XMIT, &conn->c_flags) == 0;
109 static void release_in_xmit(struct rds_connection *conn)
111 clear_bit(RDS_IN_XMIT, &conn->c_flags);
112 smp_mb__after_atomic();
114 * We don't use wait_on_bit()/wake_up_bit() because our waking is in a
115 * hot path and finding waiters is very rare. We don't want to walk
116 * the system-wide hashed waitqueue buckets in the fast path only to
117 * almost never find waiters.
119 if (waitqueue_active(&conn->c_waitq))
120 wake_up_all(&conn->c_waitq);
124 * We're making the conscious trade-off here to only send one message
125 * down the connection at a time.
127 * - tx queueing is a simple fifo list
128 * - reassembly is optional and easily done by transports per conn
129 * - no per flow rx lookup at all, straight to the socket
130 * - less per-frag memory and wire overhead
132 * - queued acks can be delayed behind large messages
134 * - small message latency is higher behind queued large messages
135 * - large message latency isn't starved by intervening small sends
137 int rds_send_xmit(struct rds_connection *conn)
139 struct rds_message *rm;
142 struct scatterlist *sg;
144 LIST_HEAD(to_be_dropped);
146 unsigned long send_gen = 0;
152 * sendmsg calls here after having queued its message on the send
153 * queue. We only have one task feeding the connection at a time. If
154 * another thread is already feeding the queue then we back off. This
155 * avoids blocking the caller and trading per-connection data between
156 * caches per message.
158 if (!acquire_in_xmit(conn)) {
159 rds_stats_inc(s_send_lock_contention);
165 * we record the send generation after doing the xmit acquire.
166 * if someone else manages to jump in and do some work, we'll use
167 * this to avoid a goto restart farther down.
169 * The acquire_in_xmit() check above ensures that only one
170 * caller can increment c_send_gen at any time.
173 send_gen = conn->c_send_gen;
176 * rds_conn_shutdown() sets the conn state and then tests RDS_IN_XMIT,
177 * we do the opposite to avoid races.
179 if (!rds_conn_up(conn)) {
180 release_in_xmit(conn);
185 if (conn->c_trans->xmit_prepare)
186 conn->c_trans->xmit_prepare(conn);
189 * spin trying to push headers and data down the connection until
190 * the connection doesn't make forward progress.
194 rm = conn->c_xmit_rm;
197 * If between sending messages, we can send a pending congestion
200 if (!rm && test_and_clear_bit(0, &conn->c_map_queued)) {
201 rm = rds_cong_update_alloc(conn);
206 rm->data.op_active = 1;
208 conn->c_xmit_rm = rm;
212 * If not already working on one, grab the next message.
214 * c_xmit_rm holds a ref while we're sending this message down
215 * the connction. We can use this ref while holding the
216 * send_sem.. rds_send_reset() is serialized with it.
223 /* we want to process as big a batch as we can, but
224 * we also want to avoid softlockups. If we've been
225 * through a lot of messages, lets back off and see
226 * if anyone else jumps in
228 if (batch_count >= send_batch_count)
231 spin_lock_irqsave(&conn->c_lock, flags);
233 if (!list_empty(&conn->c_send_queue)) {
234 rm = list_entry(conn->c_send_queue.next,
237 rds_message_addref(rm);
240 * Move the message from the send queue to the retransmit
243 list_move_tail(&rm->m_conn_item, &conn->c_retrans);
246 spin_unlock_irqrestore(&conn->c_lock, flags);
251 /* Unfortunately, the way Infiniband deals with
252 * RDMA to a bad MR key is by moving the entire
253 * queue pair to error state. We cold possibly
254 * recover from that, but right now we drop the
256 * Therefore, we never retransmit messages with RDMA ops.
258 if (rm->rdma.op_active &&
259 test_bit(RDS_MSG_RETRANSMITTED, &rm->m_flags)) {
260 spin_lock_irqsave(&conn->c_lock, flags);
261 if (test_and_clear_bit(RDS_MSG_ON_CONN, &rm->m_flags))
262 list_move(&rm->m_conn_item, &to_be_dropped);
263 spin_unlock_irqrestore(&conn->c_lock, flags);
267 /* Require an ACK every once in a while */
268 len = ntohl(rm->m_inc.i_hdr.h_len);
269 if (conn->c_unacked_packets == 0 ||
270 conn->c_unacked_bytes < len) {
271 __set_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags);
273 conn->c_unacked_packets = rds_sysctl_max_unacked_packets;
274 conn->c_unacked_bytes = rds_sysctl_max_unacked_bytes;
275 rds_stats_inc(s_send_ack_required);
277 conn->c_unacked_bytes -= len;
278 conn->c_unacked_packets--;
281 conn->c_xmit_rm = rm;
284 /* The transport either sends the whole rdma or none of it */
285 if (rm->rdma.op_active && !conn->c_xmit_rdma_sent) {
286 rm->m_final_op = &rm->rdma;
287 /* The transport owns the mapped memory for now.
288 * You can't unmap it while it's on the send queue
290 set_bit(RDS_MSG_MAPPED, &rm->m_flags);
291 ret = conn->c_trans->xmit_rdma(conn, &rm->rdma);
293 clear_bit(RDS_MSG_MAPPED, &rm->m_flags);
294 wake_up_interruptible(&rm->m_flush_wait);
297 conn->c_xmit_rdma_sent = 1;
301 if (rm->atomic.op_active && !conn->c_xmit_atomic_sent) {
302 rm->m_final_op = &rm->atomic;
303 /* The transport owns the mapped memory for now.
304 * You can't unmap it while it's on the send queue
306 set_bit(RDS_MSG_MAPPED, &rm->m_flags);
307 ret = conn->c_trans->xmit_atomic(conn, &rm->atomic);
309 clear_bit(RDS_MSG_MAPPED, &rm->m_flags);
310 wake_up_interruptible(&rm->m_flush_wait);
313 conn->c_xmit_atomic_sent = 1;
318 * A number of cases require an RDS header to be sent
319 * even if there is no data.
320 * We permit 0-byte sends; rds-ping depends on this.
321 * However, if there are exclusively attached silent ops,
322 * we skip the hdr/data send, to enable silent operation.
324 if (rm->data.op_nents == 0) {
326 int all_ops_are_silent = 1;
328 ops_present = (rm->atomic.op_active || rm->rdma.op_active);
329 if (rm->atomic.op_active && !rm->atomic.op_silent)
330 all_ops_are_silent = 0;
331 if (rm->rdma.op_active && !rm->rdma.op_silent)
332 all_ops_are_silent = 0;
334 if (ops_present && all_ops_are_silent
335 && !rm->m_rdma_cookie)
336 rm->data.op_active = 0;
339 if (rm->data.op_active && !conn->c_xmit_data_sent) {
340 rm->m_final_op = &rm->data;
341 ret = conn->c_trans->xmit(conn, rm,
342 conn->c_xmit_hdr_off,
344 conn->c_xmit_data_off);
348 if (conn->c_xmit_hdr_off < sizeof(struct rds_header)) {
349 tmp = min_t(int, ret,
350 sizeof(struct rds_header) -
351 conn->c_xmit_hdr_off);
352 conn->c_xmit_hdr_off += tmp;
356 sg = &rm->data.op_sg[conn->c_xmit_sg];
358 tmp = min_t(int, ret, sg->length -
359 conn->c_xmit_data_off);
360 conn->c_xmit_data_off += tmp;
362 if (conn->c_xmit_data_off == sg->length) {
363 conn->c_xmit_data_off = 0;
367 conn->c_xmit_sg == rm->data.op_nents);
371 if (conn->c_xmit_hdr_off == sizeof(struct rds_header) &&
372 (conn->c_xmit_sg == rm->data.op_nents))
373 conn->c_xmit_data_sent = 1;
377 * A rm will only take multiple times through this loop
378 * if there is a data op. Thus, if the data is sent (or there was
379 * none), then we're done with the rm.
381 if (!rm->data.op_active || conn->c_xmit_data_sent) {
382 conn->c_xmit_rm = NULL;
384 conn->c_xmit_hdr_off = 0;
385 conn->c_xmit_data_off = 0;
386 conn->c_xmit_rdma_sent = 0;
387 conn->c_xmit_atomic_sent = 0;
388 conn->c_xmit_data_sent = 0;
395 if (conn->c_trans->xmit_complete)
396 conn->c_trans->xmit_complete(conn);
397 release_in_xmit(conn);
399 /* Nuke any messages we decided not to retransmit. */
400 if (!list_empty(&to_be_dropped)) {
401 /* irqs on here, so we can put(), unlike above */
402 list_for_each_entry(rm, &to_be_dropped, m_conn_item)
404 rds_send_remove_from_sock(&to_be_dropped, RDS_RDMA_DROPPED);
408 * Other senders can queue a message after we last test the send queue
409 * but before we clear RDS_IN_XMIT. In that case they'd back off and
410 * not try and send their newly queued message. We need to check the
411 * send queue after having cleared RDS_IN_XMIT so that their message
412 * doesn't get stuck on the send queue.
414 * If the transport cannot continue (i.e ret != 0), then it must
415 * call us when more room is available, such as from the tx
416 * completion handler.
418 * We have an extra generation check here so that if someone manages
419 * to jump in after our release_in_xmit, we'll see that they have done
420 * some work and we will skip our goto
424 if ((test_bit(0, &conn->c_map_queued) ||
425 !list_empty(&conn->c_send_queue)) &&
426 send_gen == conn->c_send_gen) {
427 rds_stats_inc(s_send_lock_queue_raced);
428 if (batch_count < send_batch_count)
430 queue_delayed_work(rds_wq, &conn->c_send_w, 1);
436 EXPORT_SYMBOL_GPL(rds_send_xmit);
438 static void rds_send_sndbuf_remove(struct rds_sock *rs, struct rds_message *rm)
440 u32 len = be32_to_cpu(rm->m_inc.i_hdr.h_len);
442 assert_spin_locked(&rs->rs_lock);
444 BUG_ON(rs->rs_snd_bytes < len);
445 rs->rs_snd_bytes -= len;
447 if (rs->rs_snd_bytes == 0)
448 rds_stats_inc(s_send_queue_empty);
451 static inline int rds_send_is_acked(struct rds_message *rm, u64 ack,
452 is_acked_func is_acked)
455 return is_acked(rm, ack);
456 return be64_to_cpu(rm->m_inc.i_hdr.h_sequence) <= ack;
460 * This is pretty similar to what happens below in the ACK
461 * handling code - except that we call here as soon as we get
462 * the IB send completion on the RDMA op and the accompanying
465 void rds_rdma_send_complete(struct rds_message *rm, int status)
467 struct rds_sock *rs = NULL;
468 struct rm_rdma_op *ro;
469 struct rds_notifier *notifier;
472 spin_lock_irqsave(&rm->m_rs_lock, flags);
475 if (test_bit(RDS_MSG_ON_SOCK, &rm->m_flags) &&
476 ro->op_active && ro->op_notify && ro->op_notifier) {
477 notifier = ro->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 ro->op_notifier = NULL;
489 spin_unlock_irqrestore(&rm->m_rs_lock, flags);
492 rds_wake_sk_sleep(rs);
493 sock_put(rds_rs_to_sk(rs));
496 EXPORT_SYMBOL_GPL(rds_rdma_send_complete);
499 * Just like above, except looks at atomic op
501 void rds_atomic_send_complete(struct rds_message *rm, int status)
503 struct rds_sock *rs = NULL;
504 struct rm_atomic_op *ao;
505 struct rds_notifier *notifier;
508 spin_lock_irqsave(&rm->m_rs_lock, flags);
511 if (test_bit(RDS_MSG_ON_SOCK, &rm->m_flags)
512 && ao->op_active && ao->op_notify && ao->op_notifier) {
513 notifier = ao->op_notifier;
515 sock_hold(rds_rs_to_sk(rs));
517 notifier->n_status = status;
518 spin_lock(&rs->rs_lock);
519 list_add_tail(¬ifier->n_list, &rs->rs_notify_queue);
520 spin_unlock(&rs->rs_lock);
522 ao->op_notifier = NULL;
525 spin_unlock_irqrestore(&rm->m_rs_lock, flags);
528 rds_wake_sk_sleep(rs);
529 sock_put(rds_rs_to_sk(rs));
532 EXPORT_SYMBOL_GPL(rds_atomic_send_complete);
535 * This is the same as rds_rdma_send_complete except we
536 * don't do any locking - we have all the ingredients (message,
537 * socket, socket lock) and can just move the notifier.
540 __rds_send_complete(struct rds_sock *rs, struct rds_message *rm, int status)
542 struct rm_rdma_op *ro;
543 struct rm_atomic_op *ao;
546 if (ro->op_active && ro->op_notify && ro->op_notifier) {
547 ro->op_notifier->n_status = status;
548 list_add_tail(&ro->op_notifier->n_list, &rs->rs_notify_queue);
549 ro->op_notifier = NULL;
553 if (ao->op_active && ao->op_notify && ao->op_notifier) {
554 ao->op_notifier->n_status = status;
555 list_add_tail(&ao->op_notifier->n_list, &rs->rs_notify_queue);
556 ao->op_notifier = NULL;
559 /* No need to wake the app - caller does this */
563 * This is called from the IB send completion when we detect
564 * a RDMA operation that failed with remote access error.
565 * So speed is not an issue here.
567 struct rds_message *rds_send_get_message(struct rds_connection *conn,
568 struct rm_rdma_op *op)
570 struct rds_message *rm, *tmp, *found = NULL;
573 spin_lock_irqsave(&conn->c_lock, flags);
575 list_for_each_entry_safe(rm, tmp, &conn->c_retrans, m_conn_item) {
576 if (&rm->rdma == op) {
577 atomic_inc(&rm->m_refcount);
583 list_for_each_entry_safe(rm, tmp, &conn->c_send_queue, m_conn_item) {
584 if (&rm->rdma == op) {
585 atomic_inc(&rm->m_refcount);
592 spin_unlock_irqrestore(&conn->c_lock, flags);
596 EXPORT_SYMBOL_GPL(rds_send_get_message);
599 * This removes messages from the socket's list if they're on it. The list
600 * argument must be private to the caller, we must be able to modify it
601 * without locks. The messages must have a reference held for their
602 * position on the list. This function will drop that reference after
603 * removing the messages from the 'messages' list regardless of if it found
604 * the messages on the socket list or not.
606 static void rds_send_remove_from_sock(struct list_head *messages, int status)
609 struct rds_sock *rs = NULL;
610 struct rds_message *rm;
612 while (!list_empty(messages)) {
615 rm = list_entry(messages->next, struct rds_message,
617 list_del_init(&rm->m_conn_item);
620 * If we see this flag cleared then we're *sure* that someone
621 * else beat us to removing it from the sock. If we race
622 * with their flag update we'll get the lock and then really
623 * see that the flag has been cleared.
625 * The message spinlock makes sure nobody clears rm->m_rs
626 * while we're messing with it. It does not prevent the
627 * message from being removed from the socket, though.
629 spin_lock_irqsave(&rm->m_rs_lock, flags);
630 if (!test_bit(RDS_MSG_ON_SOCK, &rm->m_flags))
631 goto unlock_and_drop;
633 if (rs != rm->m_rs) {
635 rds_wake_sk_sleep(rs);
636 sock_put(rds_rs_to_sk(rs));
640 sock_hold(rds_rs_to_sk(rs));
643 goto unlock_and_drop;
644 spin_lock(&rs->rs_lock);
646 if (test_and_clear_bit(RDS_MSG_ON_SOCK, &rm->m_flags)) {
647 struct rm_rdma_op *ro = &rm->rdma;
648 struct rds_notifier *notifier;
650 list_del_init(&rm->m_sock_item);
651 rds_send_sndbuf_remove(rs, rm);
653 if (ro->op_active && ro->op_notifier &&
654 (ro->op_notify || (ro->op_recverr && status))) {
655 notifier = ro->op_notifier;
656 list_add_tail(¬ifier->n_list,
657 &rs->rs_notify_queue);
658 if (!notifier->n_status)
659 notifier->n_status = status;
660 rm->rdma.op_notifier = NULL;
665 spin_unlock(&rs->rs_lock);
668 spin_unlock_irqrestore(&rm->m_rs_lock, flags);
675 rds_wake_sk_sleep(rs);
676 sock_put(rds_rs_to_sk(rs));
681 * Transports call here when they've determined that the receiver queued
682 * messages up to, and including, the given sequence number. Messages are
683 * moved to the retrans queue when rds_send_xmit picks them off the send
684 * queue. This means that in the TCP case, the message may not have been
685 * assigned the m_ack_seq yet - but that's fine as long as tcp_is_acked
686 * checks the RDS_MSG_HAS_ACK_SEQ bit.
688 void rds_send_drop_acked(struct rds_connection *conn, u64 ack,
689 is_acked_func is_acked)
691 struct rds_message *rm, *tmp;
695 spin_lock_irqsave(&conn->c_lock, flags);
697 list_for_each_entry_safe(rm, tmp, &conn->c_retrans, m_conn_item) {
698 if (!rds_send_is_acked(rm, ack, is_acked))
701 list_move(&rm->m_conn_item, &list);
702 clear_bit(RDS_MSG_ON_CONN, &rm->m_flags);
705 /* order flag updates with spin locks */
706 if (!list_empty(&list))
707 smp_mb__after_atomic();
709 spin_unlock_irqrestore(&conn->c_lock, flags);
711 /* now remove the messages from the sock list as needed */
712 rds_send_remove_from_sock(&list, RDS_RDMA_SUCCESS);
714 EXPORT_SYMBOL_GPL(rds_send_drop_acked);
716 void rds_send_drop_to(struct rds_sock *rs, struct sockaddr_in *dest)
718 struct rds_message *rm, *tmp;
719 struct rds_connection *conn;
723 /* get all the messages we're dropping under the rs lock */
724 spin_lock_irqsave(&rs->rs_lock, flags);
726 list_for_each_entry_safe(rm, tmp, &rs->rs_send_queue, m_sock_item) {
727 if (dest && (dest->sin_addr.s_addr != rm->m_daddr ||
728 dest->sin_port != rm->m_inc.i_hdr.h_dport))
731 list_move(&rm->m_sock_item, &list);
732 rds_send_sndbuf_remove(rs, rm);
733 clear_bit(RDS_MSG_ON_SOCK, &rm->m_flags);
736 /* order flag updates with the rs lock */
737 smp_mb__after_atomic();
739 spin_unlock_irqrestore(&rs->rs_lock, flags);
741 if (list_empty(&list))
744 /* Remove the messages from the conn */
745 list_for_each_entry(rm, &list, m_sock_item) {
747 conn = rm->m_inc.i_conn;
749 spin_lock_irqsave(&conn->c_lock, flags);
751 * Maybe someone else beat us to removing rm from the conn.
752 * If we race with their flag update we'll get the lock and
753 * then really see that the flag has been cleared.
755 if (!test_and_clear_bit(RDS_MSG_ON_CONN, &rm->m_flags)) {
756 spin_unlock_irqrestore(&conn->c_lock, flags);
757 spin_lock_irqsave(&rm->m_rs_lock, flags);
759 spin_unlock_irqrestore(&rm->m_rs_lock, flags);
762 list_del_init(&rm->m_conn_item);
763 spin_unlock_irqrestore(&conn->c_lock, flags);
766 * Couldn't grab m_rs_lock in top loop (lock ordering),
769 spin_lock_irqsave(&rm->m_rs_lock, flags);
771 spin_lock(&rs->rs_lock);
772 __rds_send_complete(rs, rm, RDS_RDMA_CANCELED);
773 spin_unlock(&rs->rs_lock);
776 spin_unlock_irqrestore(&rm->m_rs_lock, flags);
781 rds_wake_sk_sleep(rs);
783 while (!list_empty(&list)) {
784 rm = list_entry(list.next, struct rds_message, m_sock_item);
785 list_del_init(&rm->m_sock_item);
786 rds_message_wait(rm);
788 /* just in case the code above skipped this message
789 * because RDS_MSG_ON_CONN wasn't set, run it again here
790 * taking m_rs_lock is the only thing that keeps us
791 * from racing with ack processing.
793 spin_lock_irqsave(&rm->m_rs_lock, flags);
795 spin_lock(&rs->rs_lock);
796 __rds_send_complete(rs, rm, RDS_RDMA_CANCELED);
797 spin_unlock(&rs->rs_lock);
800 spin_unlock_irqrestore(&rm->m_rs_lock, flags);
807 * we only want this to fire once so we use the callers 'queued'. It's
808 * possible that another thread can race with us and remove the
809 * message from the flow with RDS_CANCEL_SENT_TO.
811 static int rds_send_queue_rm(struct rds_sock *rs, struct rds_connection *conn,
812 struct rds_message *rm, __be16 sport,
813 __be16 dport, int *queued)
821 len = be32_to_cpu(rm->m_inc.i_hdr.h_len);
823 /* this is the only place which holds both the socket's rs_lock
824 * and the connection's c_lock */
825 spin_lock_irqsave(&rs->rs_lock, flags);
828 * If there is a little space in sndbuf, we don't queue anything,
829 * and userspace gets -EAGAIN. But poll() indicates there's send
830 * room. This can lead to bad behavior (spinning) if snd_bytes isn't
831 * freed up by incoming acks. So we check the *old* value of
832 * rs_snd_bytes here to allow the last msg to exceed the buffer,
833 * and poll() now knows no more data can be sent.
835 if (rs->rs_snd_bytes < rds_sk_sndbuf(rs)) {
836 rs->rs_snd_bytes += len;
838 /* let recv side know we are close to send space exhaustion.
839 * This is probably not the optimal way to do it, as this
840 * means we set the flag on *all* messages as soon as our
841 * throughput hits a certain threshold.
843 if (rs->rs_snd_bytes >= rds_sk_sndbuf(rs) / 2)
844 __set_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags);
846 list_add_tail(&rm->m_sock_item, &rs->rs_send_queue);
847 set_bit(RDS_MSG_ON_SOCK, &rm->m_flags);
848 rds_message_addref(rm);
851 /* The code ordering is a little weird, but we're
852 trying to minimize the time we hold c_lock */
853 rds_message_populate_header(&rm->m_inc.i_hdr, sport, dport, 0);
854 rm->m_inc.i_conn = conn;
855 rds_message_addref(rm);
857 spin_lock(&conn->c_lock);
858 rm->m_inc.i_hdr.h_sequence = cpu_to_be64(conn->c_next_tx_seq++);
859 list_add_tail(&rm->m_conn_item, &conn->c_send_queue);
860 set_bit(RDS_MSG_ON_CONN, &rm->m_flags);
861 spin_unlock(&conn->c_lock);
863 rdsdebug("queued msg %p len %d, rs %p bytes %d seq %llu\n",
864 rm, len, rs, rs->rs_snd_bytes,
865 (unsigned long long)be64_to_cpu(rm->m_inc.i_hdr.h_sequence));
870 spin_unlock_irqrestore(&rs->rs_lock, flags);
876 * rds_message is getting to be quite complicated, and we'd like to allocate
877 * it all in one go. This figures out how big it needs to be up front.
879 static int rds_rm_size(struct msghdr *msg, int data_len)
881 struct cmsghdr *cmsg;
886 for_each_cmsghdr(cmsg, msg) {
887 if (!CMSG_OK(msg, cmsg))
890 if (cmsg->cmsg_level != SOL_RDS)
893 switch (cmsg->cmsg_type) {
894 case RDS_CMSG_RDMA_ARGS:
896 retval = rds_rdma_extra_size(CMSG_DATA(cmsg));
903 case RDS_CMSG_RDMA_DEST:
904 case RDS_CMSG_RDMA_MAP:
906 /* these are valid but do no add any size */
909 case RDS_CMSG_ATOMIC_CSWP:
910 case RDS_CMSG_ATOMIC_FADD:
911 case RDS_CMSG_MASKED_ATOMIC_CSWP:
912 case RDS_CMSG_MASKED_ATOMIC_FADD:
914 size += sizeof(struct scatterlist);
923 size += ceil(data_len, PAGE_SIZE) * sizeof(struct scatterlist);
925 /* Ensure (DEST, MAP) are never used with (ARGS, ATOMIC) */
926 if (cmsg_groups == 3)
932 static int rds_cmsg_send(struct rds_sock *rs, struct rds_message *rm,
933 struct msghdr *msg, int *allocated_mr)
935 struct cmsghdr *cmsg;
938 for_each_cmsghdr(cmsg, msg) {
939 if (!CMSG_OK(msg, cmsg))
942 if (cmsg->cmsg_level != SOL_RDS)
945 /* As a side effect, RDMA_DEST and RDMA_MAP will set
946 * rm->rdma.m_rdma_cookie and rm->rdma.m_rdma_mr.
948 switch (cmsg->cmsg_type) {
949 case RDS_CMSG_RDMA_ARGS:
950 ret = rds_cmsg_rdma_args(rs, rm, cmsg);
953 case RDS_CMSG_RDMA_DEST:
954 ret = rds_cmsg_rdma_dest(rs, rm, cmsg);
957 case RDS_CMSG_RDMA_MAP:
958 ret = rds_cmsg_rdma_map(rs, rm, cmsg);
962 case RDS_CMSG_ATOMIC_CSWP:
963 case RDS_CMSG_ATOMIC_FADD:
964 case RDS_CMSG_MASKED_ATOMIC_CSWP:
965 case RDS_CMSG_MASKED_ATOMIC_FADD:
966 ret = rds_cmsg_atomic(rs, rm, cmsg);
980 int rds_sendmsg(struct socket *sock, struct msghdr *msg, size_t payload_len)
982 struct sock *sk = sock->sk;
983 struct rds_sock *rs = rds_sk_to_rs(sk);
984 DECLARE_SOCKADDR(struct sockaddr_in *, usin, msg->msg_name);
987 struct rds_message *rm = NULL;
988 struct rds_connection *conn;
990 int queued = 0, allocated_mr = 0;
991 int nonblock = msg->msg_flags & MSG_DONTWAIT;
992 long timeo = sock_sndtimeo(sk, nonblock);
994 /* Mirror Linux UDP mirror of BSD error message compatibility */
995 /* XXX: Perhaps MSG_MORE someday */
996 if (msg->msg_flags & ~(MSG_DONTWAIT | MSG_CMSG_COMPAT)) {
1001 if (msg->msg_namelen) {
1002 /* XXX fail non-unicast destination IPs? */
1003 if (msg->msg_namelen < sizeof(*usin) || usin->sin_family != AF_INET) {
1007 daddr = usin->sin_addr.s_addr;
1008 dport = usin->sin_port;
1010 /* We only care about consistency with ->connect() */
1012 daddr = rs->rs_conn_addr;
1013 dport = rs->rs_conn_port;
1018 if (daddr == 0 || rs->rs_bound_addr == 0) {
1020 ret = -ENOTCONN; /* XXX not a great errno */
1025 if (payload_len > rds_sk_sndbuf(rs)) {
1030 /* size of rm including all sgs */
1031 ret = rds_rm_size(msg, payload_len);
1035 rm = rds_message_alloc(ret, GFP_KERNEL);
1041 /* Attach data to the rm */
1043 rm->data.op_sg = rds_message_alloc_sgs(rm, ceil(payload_len, PAGE_SIZE));
1044 if (!rm->data.op_sg) {
1048 ret = rds_message_copy_from_user(rm, &msg->msg_iter);
1052 rm->data.op_active = 1;
1054 rm->m_daddr = daddr;
1056 /* rds_conn_create has a spinlock that runs with IRQ off.
1057 * Caching the conn in the socket helps a lot. */
1058 if (rs->rs_conn && rs->rs_conn->c_faddr == daddr)
1061 conn = rds_conn_create_outgoing(sock_net(sock->sk),
1062 rs->rs_bound_addr, daddr,
1064 sock->sk->sk_allocation);
1066 ret = PTR_ERR(conn);
1072 /* Parse any control messages the user may have included. */
1073 ret = rds_cmsg_send(rs, rm, msg, &allocated_mr);
1077 if (rm->rdma.op_active && !conn->c_trans->xmit_rdma) {
1078 printk_ratelimited(KERN_NOTICE "rdma_op %p conn xmit_rdma %p\n",
1079 &rm->rdma, conn->c_trans->xmit_rdma);
1084 if (rm->atomic.op_active && !conn->c_trans->xmit_atomic) {
1085 printk_ratelimited(KERN_NOTICE "atomic_op %p conn xmit_atomic %p\n",
1086 &rm->atomic, conn->c_trans->xmit_atomic);
1091 rds_conn_connect_if_down(conn);
1093 ret = rds_cong_wait(conn->c_fcong, dport, nonblock, rs);
1095 rs->rs_seen_congestion = 1;
1099 while (!rds_send_queue_rm(rs, conn, rm, rs->rs_bound_port,
1101 rds_stats_inc(s_send_queue_full);
1108 timeo = wait_event_interruptible_timeout(*sk_sleep(sk),
1109 rds_send_queue_rm(rs, conn, rm,
1114 rdsdebug("sendmsg woke queued %d timeo %ld\n", queued, timeo);
1115 if (timeo > 0 || timeo == MAX_SCHEDULE_TIMEOUT)
1125 * By now we've committed to the send. We reuse rds_send_worker()
1126 * to retry sends in the rds thread if the transport asks us to.
1128 rds_stats_inc(s_send_queued);
1130 ret = rds_send_xmit(conn);
1131 if (ret == -ENOMEM || ret == -EAGAIN)
1132 queue_delayed_work(rds_wq, &conn->c_send_w, 1);
1134 rds_message_put(rm);
1138 /* If the user included a RDMA_MAP cmsg, we allocated a MR on the fly.
1139 * If the sendmsg goes through, we keep the MR. If it fails with EAGAIN
1140 * or in any other way, we need to destroy the MR again */
1142 rds_rdma_unuse(rs, rds_rdma_cookie_key(rm->m_rdma_cookie), 1);
1145 rds_message_put(rm);
1150 * Reply to a ping packet.
1153 rds_send_pong(struct rds_connection *conn, __be16 dport)
1155 struct rds_message *rm;
1156 unsigned long flags;
1159 rm = rds_message_alloc(0, GFP_ATOMIC);
1165 rm->m_daddr = conn->c_faddr;
1166 rm->data.op_active = 1;
1168 rds_conn_connect_if_down(conn);
1170 ret = rds_cong_wait(conn->c_fcong, dport, 1, NULL);
1174 spin_lock_irqsave(&conn->c_lock, flags);
1175 list_add_tail(&rm->m_conn_item, &conn->c_send_queue);
1176 set_bit(RDS_MSG_ON_CONN, &rm->m_flags);
1177 rds_message_addref(rm);
1178 rm->m_inc.i_conn = conn;
1180 rds_message_populate_header(&rm->m_inc.i_hdr, 0, dport,
1181 conn->c_next_tx_seq);
1182 conn->c_next_tx_seq++;
1183 spin_unlock_irqrestore(&conn->c_lock, flags);
1185 rds_stats_inc(s_send_queued);
1186 rds_stats_inc(s_send_pong);
1188 /* schedule the send work on rds_wq */
1189 queue_delayed_work(rds_wq, &conn->c_send_w, 1);
1191 rds_message_put(rm);
1196 rds_message_put(rm);