Loading...
1/*
2 * Copyright (c) 2006 Oracle. All rights reserved.
3 *
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:
9 *
10 * Redistribution and use in source and binary forms, with or
11 * without modification, are permitted provided that the following
12 * conditions are met:
13 *
14 * - Redistributions of source code must retain the above
15 * copyright notice, this list of conditions and the following
16 * disclaimer.
17 *
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.
22 *
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
30 * SOFTWARE.
31 *
32 */
33#include <linux/kernel.h>
34#include <linux/moduleparam.h>
35#include <linux/gfp.h>
36#include <net/sock.h>
37#include <linux/in.h>
38#include <linux/list.h>
39#include <linux/ratelimit.h>
40#include <linux/export.h>
41#include <linux/sizes.h>
42
43#include "rds.h"
44
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
47 * will kick our shin.
48 * Also, it seems fairer to not let one busy connection stall all the
49 * others.
50 *
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
53 * drained the queue).
54 */
55static int send_batch_count = SZ_1K;
56module_param(send_batch_count, int, 0444);
57MODULE_PARM_DESC(send_batch_count, " batch factor when working the send queue");
58
59static void rds_send_remove_from_sock(struct list_head *messages, int status);
60
61/*
62 * Reset the send state. Callers must ensure that this doesn't race with
63 * rds_send_xmit().
64 */
65void rds_send_reset(struct rds_connection *conn)
66{
67 struct rds_message *rm, *tmp;
68 unsigned long flags;
69
70 if (conn->c_xmit_rm) {
71 rm = 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);
78 rds_message_put(rm);
79 }
80
81 conn->c_xmit_sg = 0;
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;
87
88 conn->c_map_queued = 0;
89
90 conn->c_unacked_packets = rds_sysctl_max_unacked_packets;
91 conn->c_unacked_bytes = rds_sysctl_max_unacked_bytes;
92
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);
98 }
99 list_splice_init(&conn->c_retrans, &conn->c_send_queue);
100 spin_unlock_irqrestore(&conn->c_lock, flags);
101}
102
103static int acquire_in_xmit(struct rds_connection *conn)
104{
105 return test_and_set_bit(RDS_IN_XMIT, &conn->c_flags) == 0;
106}
107
108static void release_in_xmit(struct rds_connection *conn)
109{
110 clear_bit(RDS_IN_XMIT, &conn->c_flags);
111 smp_mb__after_atomic();
112 /*
113 * We don't use wait_on_bit()/wake_up_bit() because our waking is in a
114 * hot path and finding waiters is very rare. We don't want to walk
115 * the system-wide hashed waitqueue buckets in the fast path only to
116 * almost never find waiters.
117 */
118 if (waitqueue_active(&conn->c_waitq))
119 wake_up_all(&conn->c_waitq);
120}
121
122/*
123 * We're making the conscious trade-off here to only send one message
124 * down the connection at a time.
125 * Pro:
126 * - tx queueing is a simple fifo list
127 * - reassembly is optional and easily done by transports per conn
128 * - no per flow rx lookup at all, straight to the socket
129 * - less per-frag memory and wire overhead
130 * Con:
131 * - queued acks can be delayed behind large messages
132 * Depends:
133 * - small message latency is higher behind queued large messages
134 * - large message latency isn't starved by intervening small sends
135 */
136int rds_send_xmit(struct rds_connection *conn)
137{
138 struct rds_message *rm;
139 unsigned long flags;
140 unsigned int tmp;
141 struct scatterlist *sg;
142 int ret = 0;
143 LIST_HEAD(to_be_dropped);
144 int batch_count;
145 unsigned long send_gen = 0;
146
147restart:
148 batch_count = 0;
149
150 /*
151 * sendmsg calls here after having queued its message on the send
152 * queue. We only have one task feeding the connection at a time. If
153 * another thread is already feeding the queue then we back off. This
154 * avoids blocking the caller and trading per-connection data between
155 * caches per message.
156 */
157 if (!acquire_in_xmit(conn)) {
158 rds_stats_inc(s_send_lock_contention);
159 ret = -ENOMEM;
160 goto out;
161 }
162
163 /*
164 * we record the send generation after doing the xmit acquire.
165 * if someone else manages to jump in and do some work, we'll use
166 * this to avoid a goto restart farther down.
167 *
168 * The acquire_in_xmit() check above ensures that only one
169 * caller can increment c_send_gen at any time.
170 */
171 conn->c_send_gen++;
172 send_gen = conn->c_send_gen;
173
174 /*
175 * rds_conn_shutdown() sets the conn state and then tests RDS_IN_XMIT,
176 * we do the opposite to avoid races.
177 */
178 if (!rds_conn_up(conn)) {
179 release_in_xmit(conn);
180 ret = 0;
181 goto out;
182 }
183
184 if (conn->c_trans->xmit_prepare)
185 conn->c_trans->xmit_prepare(conn);
186
187 /*
188 * spin trying to push headers and data down the connection until
189 * the connection doesn't make forward progress.
190 */
191 while (1) {
192
193 rm = conn->c_xmit_rm;
194
195 /*
196 * If between sending messages, we can send a pending congestion
197 * map update.
198 */
199 if (!rm && test_and_clear_bit(0, &conn->c_map_queued)) {
200 rm = rds_cong_update_alloc(conn);
201 if (IS_ERR(rm)) {
202 ret = PTR_ERR(rm);
203 break;
204 }
205 rm->data.op_active = 1;
206
207 conn->c_xmit_rm = rm;
208 }
209
210 /*
211 * If not already working on one, grab the next message.
212 *
213 * c_xmit_rm holds a ref while we're sending this message down
214 * the connction. We can use this ref while holding the
215 * send_sem.. rds_send_reset() is serialized with it.
216 */
217 if (!rm) {
218 unsigned int len;
219
220 batch_count++;
221
222 /* we want to process as big a batch as we can, but
223 * we also want to avoid softlockups. If we've been
224 * through a lot of messages, lets back off and see
225 * if anyone else jumps in
226 */
227 if (batch_count >= send_batch_count)
228 goto over_batch;
229
230 spin_lock_irqsave(&conn->c_lock, flags);
231
232 if (!list_empty(&conn->c_send_queue)) {
233 rm = list_entry(conn->c_send_queue.next,
234 struct rds_message,
235 m_conn_item);
236 rds_message_addref(rm);
237
238 /*
239 * Move the message from the send queue to the retransmit
240 * list right away.
241 */
242 list_move_tail(&rm->m_conn_item, &conn->c_retrans);
243 }
244
245 spin_unlock_irqrestore(&conn->c_lock, flags);
246
247 if (!rm)
248 break;
249
250 /* Unfortunately, the way Infiniband deals with
251 * RDMA to a bad MR key is by moving the entire
252 * queue pair to error state. We cold possibly
253 * recover from that, but right now we drop the
254 * connection.
255 * Therefore, we never retransmit messages with RDMA ops.
256 */
257 if (rm->rdma.op_active &&
258 test_bit(RDS_MSG_RETRANSMITTED, &rm->m_flags)) {
259 spin_lock_irqsave(&conn->c_lock, flags);
260 if (test_and_clear_bit(RDS_MSG_ON_CONN, &rm->m_flags))
261 list_move(&rm->m_conn_item, &to_be_dropped);
262 spin_unlock_irqrestore(&conn->c_lock, flags);
263 continue;
264 }
265
266 /* Require an ACK every once in a while */
267 len = ntohl(rm->m_inc.i_hdr.h_len);
268 if (conn->c_unacked_packets == 0 ||
269 conn->c_unacked_bytes < len) {
270 __set_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags);
271
272 conn->c_unacked_packets = rds_sysctl_max_unacked_packets;
273 conn->c_unacked_bytes = rds_sysctl_max_unacked_bytes;
274 rds_stats_inc(s_send_ack_required);
275 } else {
276 conn->c_unacked_bytes -= len;
277 conn->c_unacked_packets--;
278 }
279
280 conn->c_xmit_rm = rm;
281 }
282
283 /* The transport either sends the whole rdma or none of it */
284 if (rm->rdma.op_active && !conn->c_xmit_rdma_sent) {
285 rm->m_final_op = &rm->rdma;
286 /* The transport owns the mapped memory for now.
287 * You can't unmap it while it's on the send queue
288 */
289 set_bit(RDS_MSG_MAPPED, &rm->m_flags);
290 ret = conn->c_trans->xmit_rdma(conn, &rm->rdma);
291 if (ret) {
292 clear_bit(RDS_MSG_MAPPED, &rm->m_flags);
293 wake_up_interruptible(&rm->m_flush_wait);
294 break;
295 }
296 conn->c_xmit_rdma_sent = 1;
297
298 }
299
300 if (rm->atomic.op_active && !conn->c_xmit_atomic_sent) {
301 rm->m_final_op = &rm->atomic;
302 /* The transport owns the mapped memory for now.
303 * You can't unmap it while it's on the send queue
304 */
305 set_bit(RDS_MSG_MAPPED, &rm->m_flags);
306 ret = conn->c_trans->xmit_atomic(conn, &rm->atomic);
307 if (ret) {
308 clear_bit(RDS_MSG_MAPPED, &rm->m_flags);
309 wake_up_interruptible(&rm->m_flush_wait);
310 break;
311 }
312 conn->c_xmit_atomic_sent = 1;
313
314 }
315
316 /*
317 * A number of cases require an RDS header to be sent
318 * even if there is no data.
319 * We permit 0-byte sends; rds-ping depends on this.
320 * However, if there are exclusively attached silent ops,
321 * we skip the hdr/data send, to enable silent operation.
322 */
323 if (rm->data.op_nents == 0) {
324 int ops_present;
325 int all_ops_are_silent = 1;
326
327 ops_present = (rm->atomic.op_active || rm->rdma.op_active);
328 if (rm->atomic.op_active && !rm->atomic.op_silent)
329 all_ops_are_silent = 0;
330 if (rm->rdma.op_active && !rm->rdma.op_silent)
331 all_ops_are_silent = 0;
332
333 if (ops_present && all_ops_are_silent
334 && !rm->m_rdma_cookie)
335 rm->data.op_active = 0;
336 }
337
338 if (rm->data.op_active && !conn->c_xmit_data_sent) {
339 rm->m_final_op = &rm->data;
340 ret = conn->c_trans->xmit(conn, rm,
341 conn->c_xmit_hdr_off,
342 conn->c_xmit_sg,
343 conn->c_xmit_data_off);
344 if (ret <= 0)
345 break;
346
347 if (conn->c_xmit_hdr_off < sizeof(struct rds_header)) {
348 tmp = min_t(int, ret,
349 sizeof(struct rds_header) -
350 conn->c_xmit_hdr_off);
351 conn->c_xmit_hdr_off += tmp;
352 ret -= tmp;
353 }
354
355 sg = &rm->data.op_sg[conn->c_xmit_sg];
356 while (ret) {
357 tmp = min_t(int, ret, sg->length -
358 conn->c_xmit_data_off);
359 conn->c_xmit_data_off += tmp;
360 ret -= tmp;
361 if (conn->c_xmit_data_off == sg->length) {
362 conn->c_xmit_data_off = 0;
363 sg++;
364 conn->c_xmit_sg++;
365 BUG_ON(ret != 0 &&
366 conn->c_xmit_sg == rm->data.op_nents);
367 }
368 }
369
370 if (conn->c_xmit_hdr_off == sizeof(struct rds_header) &&
371 (conn->c_xmit_sg == rm->data.op_nents))
372 conn->c_xmit_data_sent = 1;
373 }
374
375 /*
376 * A rm will only take multiple times through this loop
377 * if there is a data op. Thus, if the data is sent (or there was
378 * none), then we're done with the rm.
379 */
380 if (!rm->data.op_active || conn->c_xmit_data_sent) {
381 conn->c_xmit_rm = NULL;
382 conn->c_xmit_sg = 0;
383 conn->c_xmit_hdr_off = 0;
384 conn->c_xmit_data_off = 0;
385 conn->c_xmit_rdma_sent = 0;
386 conn->c_xmit_atomic_sent = 0;
387 conn->c_xmit_data_sent = 0;
388
389 rds_message_put(rm);
390 }
391 }
392
393over_batch:
394 if (conn->c_trans->xmit_complete)
395 conn->c_trans->xmit_complete(conn);
396 release_in_xmit(conn);
397
398 /* Nuke any messages we decided not to retransmit. */
399 if (!list_empty(&to_be_dropped)) {
400 /* irqs on here, so we can put(), unlike above */
401 list_for_each_entry(rm, &to_be_dropped, m_conn_item)
402 rds_message_put(rm);
403 rds_send_remove_from_sock(&to_be_dropped, RDS_RDMA_DROPPED);
404 }
405
406 /*
407 * Other senders can queue a message after we last test the send queue
408 * but before we clear RDS_IN_XMIT. In that case they'd back off and
409 * not try and send their newly queued message. We need to check the
410 * send queue after having cleared RDS_IN_XMIT so that their message
411 * doesn't get stuck on the send queue.
412 *
413 * If the transport cannot continue (i.e ret != 0), then it must
414 * call us when more room is available, such as from the tx
415 * completion handler.
416 *
417 * We have an extra generation check here so that if someone manages
418 * to jump in after our release_in_xmit, we'll see that they have done
419 * some work and we will skip our goto
420 */
421 if (ret == 0) {
422 smp_mb();
423 if ((test_bit(0, &conn->c_map_queued) ||
424 !list_empty(&conn->c_send_queue)) &&
425 send_gen == conn->c_send_gen) {
426 rds_stats_inc(s_send_lock_queue_raced);
427 if (batch_count < send_batch_count)
428 goto restart;
429 queue_delayed_work(rds_wq, &conn->c_send_w, 1);
430 }
431 }
432out:
433 return ret;
434}
435EXPORT_SYMBOL_GPL(rds_send_xmit);
436
437static void rds_send_sndbuf_remove(struct rds_sock *rs, struct rds_message *rm)
438{
439 u32 len = be32_to_cpu(rm->m_inc.i_hdr.h_len);
440
441 assert_spin_locked(&rs->rs_lock);
442
443 BUG_ON(rs->rs_snd_bytes < len);
444 rs->rs_snd_bytes -= len;
445
446 if (rs->rs_snd_bytes == 0)
447 rds_stats_inc(s_send_queue_empty);
448}
449
450static inline int rds_send_is_acked(struct rds_message *rm, u64 ack,
451 is_acked_func is_acked)
452{
453 if (is_acked)
454 return is_acked(rm, ack);
455 return be64_to_cpu(rm->m_inc.i_hdr.h_sequence) <= ack;
456}
457
458/*
459 * This is pretty similar to what happens below in the ACK
460 * handling code - except that we call here as soon as we get
461 * the IB send completion on the RDMA op and the accompanying
462 * message.
463 */
464void rds_rdma_send_complete(struct rds_message *rm, int status)
465{
466 struct rds_sock *rs = NULL;
467 struct rm_rdma_op *ro;
468 struct rds_notifier *notifier;
469 unsigned long flags;
470
471 spin_lock_irqsave(&rm->m_rs_lock, flags);
472
473 ro = &rm->rdma;
474 if (test_bit(RDS_MSG_ON_SOCK, &rm->m_flags) &&
475 ro->op_active && ro->op_notify && ro->op_notifier) {
476 notifier = ro->op_notifier;
477 rs = rm->m_rs;
478 sock_hold(rds_rs_to_sk(rs));
479
480 notifier->n_status = status;
481 spin_lock(&rs->rs_lock);
482 list_add_tail(¬ifier->n_list, &rs->rs_notify_queue);
483 spin_unlock(&rs->rs_lock);
484
485 ro->op_notifier = NULL;
486 }
487
488 spin_unlock_irqrestore(&rm->m_rs_lock, flags);
489
490 if (rs) {
491 rds_wake_sk_sleep(rs);
492 sock_put(rds_rs_to_sk(rs));
493 }
494}
495EXPORT_SYMBOL_GPL(rds_rdma_send_complete);
496
497/*
498 * Just like above, except looks at atomic op
499 */
500void rds_atomic_send_complete(struct rds_message *rm, int status)
501{
502 struct rds_sock *rs = NULL;
503 struct rm_atomic_op *ao;
504 struct rds_notifier *notifier;
505 unsigned long flags;
506
507 spin_lock_irqsave(&rm->m_rs_lock, flags);
508
509 ao = &rm->atomic;
510 if (test_bit(RDS_MSG_ON_SOCK, &rm->m_flags)
511 && ao->op_active && ao->op_notify && ao->op_notifier) {
512 notifier = ao->op_notifier;
513 rs = rm->m_rs;
514 sock_hold(rds_rs_to_sk(rs));
515
516 notifier->n_status = status;
517 spin_lock(&rs->rs_lock);
518 list_add_tail(¬ifier->n_list, &rs->rs_notify_queue);
519 spin_unlock(&rs->rs_lock);
520
521 ao->op_notifier = NULL;
522 }
523
524 spin_unlock_irqrestore(&rm->m_rs_lock, flags);
525
526 if (rs) {
527 rds_wake_sk_sleep(rs);
528 sock_put(rds_rs_to_sk(rs));
529 }
530}
531EXPORT_SYMBOL_GPL(rds_atomic_send_complete);
532
533/*
534 * This is the same as rds_rdma_send_complete except we
535 * don't do any locking - we have all the ingredients (message,
536 * socket, socket lock) and can just move the notifier.
537 */
538static inline void
539__rds_send_complete(struct rds_sock *rs, struct rds_message *rm, int status)
540{
541 struct rm_rdma_op *ro;
542 struct rm_atomic_op *ao;
543
544 ro = &rm->rdma;
545 if (ro->op_active && ro->op_notify && ro->op_notifier) {
546 ro->op_notifier->n_status = status;
547 list_add_tail(&ro->op_notifier->n_list, &rs->rs_notify_queue);
548 ro->op_notifier = NULL;
549 }
550
551 ao = &rm->atomic;
552 if (ao->op_active && ao->op_notify && ao->op_notifier) {
553 ao->op_notifier->n_status = status;
554 list_add_tail(&ao->op_notifier->n_list, &rs->rs_notify_queue);
555 ao->op_notifier = NULL;
556 }
557
558 /* No need to wake the app - caller does this */
559}
560
561/*
562 * This is called from the IB send completion when we detect
563 * a RDMA operation that failed with remote access error.
564 * So speed is not an issue here.
565 */
566struct rds_message *rds_send_get_message(struct rds_connection *conn,
567 struct rm_rdma_op *op)
568{
569 struct rds_message *rm, *tmp, *found = NULL;
570 unsigned long flags;
571
572 spin_lock_irqsave(&conn->c_lock, flags);
573
574 list_for_each_entry_safe(rm, tmp, &conn->c_retrans, m_conn_item) {
575 if (&rm->rdma == op) {
576 atomic_inc(&rm->m_refcount);
577 found = rm;
578 goto out;
579 }
580 }
581
582 list_for_each_entry_safe(rm, tmp, &conn->c_send_queue, m_conn_item) {
583 if (&rm->rdma == op) {
584 atomic_inc(&rm->m_refcount);
585 found = rm;
586 break;
587 }
588 }
589
590out:
591 spin_unlock_irqrestore(&conn->c_lock, flags);
592
593 return found;
594}
595EXPORT_SYMBOL_GPL(rds_send_get_message);
596
597/*
598 * This removes messages from the socket's list if they're on it. The list
599 * argument must be private to the caller, we must be able to modify it
600 * without locks. The messages must have a reference held for their
601 * position on the list. This function will drop that reference after
602 * removing the messages from the 'messages' list regardless of if it found
603 * the messages on the socket list or not.
604 */
605static void rds_send_remove_from_sock(struct list_head *messages, int status)
606{
607 unsigned long flags;
608 struct rds_sock *rs = NULL;
609 struct rds_message *rm;
610
611 while (!list_empty(messages)) {
612 int was_on_sock = 0;
613
614 rm = list_entry(messages->next, struct rds_message,
615 m_conn_item);
616 list_del_init(&rm->m_conn_item);
617
618 /*
619 * If we see this flag cleared then we're *sure* that someone
620 * else beat us to removing it from the sock. If we race
621 * with their flag update we'll get the lock and then really
622 * see that the flag has been cleared.
623 *
624 * The message spinlock makes sure nobody clears rm->m_rs
625 * while we're messing with it. It does not prevent the
626 * message from being removed from the socket, though.
627 */
628 spin_lock_irqsave(&rm->m_rs_lock, flags);
629 if (!test_bit(RDS_MSG_ON_SOCK, &rm->m_flags))
630 goto unlock_and_drop;
631
632 if (rs != rm->m_rs) {
633 if (rs) {
634 rds_wake_sk_sleep(rs);
635 sock_put(rds_rs_to_sk(rs));
636 }
637 rs = rm->m_rs;
638 if (rs)
639 sock_hold(rds_rs_to_sk(rs));
640 }
641 if (!rs)
642 goto unlock_and_drop;
643 spin_lock(&rs->rs_lock);
644
645 if (test_and_clear_bit(RDS_MSG_ON_SOCK, &rm->m_flags)) {
646 struct rm_rdma_op *ro = &rm->rdma;
647 struct rds_notifier *notifier;
648
649 list_del_init(&rm->m_sock_item);
650 rds_send_sndbuf_remove(rs, rm);
651
652 if (ro->op_active && ro->op_notifier &&
653 (ro->op_notify || (ro->op_recverr && status))) {
654 notifier = ro->op_notifier;
655 list_add_tail(¬ifier->n_list,
656 &rs->rs_notify_queue);
657 if (!notifier->n_status)
658 notifier->n_status = status;
659 rm->rdma.op_notifier = NULL;
660 }
661 was_on_sock = 1;
662 rm->m_rs = NULL;
663 }
664 spin_unlock(&rs->rs_lock);
665
666unlock_and_drop:
667 spin_unlock_irqrestore(&rm->m_rs_lock, flags);
668 rds_message_put(rm);
669 if (was_on_sock)
670 rds_message_put(rm);
671 }
672
673 if (rs) {
674 rds_wake_sk_sleep(rs);
675 sock_put(rds_rs_to_sk(rs));
676 }
677}
678
679/*
680 * Transports call here when they've determined that the receiver queued
681 * messages up to, and including, the given sequence number. Messages are
682 * moved to the retrans queue when rds_send_xmit picks them off the send
683 * queue. This means that in the TCP case, the message may not have been
684 * assigned the m_ack_seq yet - but that's fine as long as tcp_is_acked
685 * checks the RDS_MSG_HAS_ACK_SEQ bit.
686 */
687void rds_send_drop_acked(struct rds_connection *conn, u64 ack,
688 is_acked_func is_acked)
689{
690 struct rds_message *rm, *tmp;
691 unsigned long flags;
692 LIST_HEAD(list);
693
694 spin_lock_irqsave(&conn->c_lock, flags);
695
696 list_for_each_entry_safe(rm, tmp, &conn->c_retrans, m_conn_item) {
697 if (!rds_send_is_acked(rm, ack, is_acked))
698 break;
699
700 list_move(&rm->m_conn_item, &list);
701 clear_bit(RDS_MSG_ON_CONN, &rm->m_flags);
702 }
703
704 /* order flag updates with spin locks */
705 if (!list_empty(&list))
706 smp_mb__after_atomic();
707
708 spin_unlock_irqrestore(&conn->c_lock, flags);
709
710 /* now remove the messages from the sock list as needed */
711 rds_send_remove_from_sock(&list, RDS_RDMA_SUCCESS);
712}
713EXPORT_SYMBOL_GPL(rds_send_drop_acked);
714
715void rds_send_drop_to(struct rds_sock *rs, struct sockaddr_in *dest)
716{
717 struct rds_message *rm, *tmp;
718 struct rds_connection *conn;
719 unsigned long flags;
720 LIST_HEAD(list);
721
722 /* get all the messages we're dropping under the rs lock */
723 spin_lock_irqsave(&rs->rs_lock, flags);
724
725 list_for_each_entry_safe(rm, tmp, &rs->rs_send_queue, m_sock_item) {
726 if (dest && (dest->sin_addr.s_addr != rm->m_daddr ||
727 dest->sin_port != rm->m_inc.i_hdr.h_dport))
728 continue;
729
730 list_move(&rm->m_sock_item, &list);
731 rds_send_sndbuf_remove(rs, rm);
732 clear_bit(RDS_MSG_ON_SOCK, &rm->m_flags);
733 }
734
735 /* order flag updates with the rs lock */
736 smp_mb__after_atomic();
737
738 spin_unlock_irqrestore(&rs->rs_lock, flags);
739
740 if (list_empty(&list))
741 return;
742
743 /* Remove the messages from the conn */
744 list_for_each_entry(rm, &list, m_sock_item) {
745
746 conn = rm->m_inc.i_conn;
747
748 spin_lock_irqsave(&conn->c_lock, flags);
749 /*
750 * Maybe someone else beat us to removing rm from the conn.
751 * If we race with their flag update we'll get the lock and
752 * then really see that the flag has been cleared.
753 */
754 if (!test_and_clear_bit(RDS_MSG_ON_CONN, &rm->m_flags)) {
755 spin_unlock_irqrestore(&conn->c_lock, flags);
756 spin_lock_irqsave(&rm->m_rs_lock, flags);
757 rm->m_rs = NULL;
758 spin_unlock_irqrestore(&rm->m_rs_lock, flags);
759 continue;
760 }
761 list_del_init(&rm->m_conn_item);
762 spin_unlock_irqrestore(&conn->c_lock, flags);
763
764 /*
765 * Couldn't grab m_rs_lock in top loop (lock ordering),
766 * but we can now.
767 */
768 spin_lock_irqsave(&rm->m_rs_lock, flags);
769
770 spin_lock(&rs->rs_lock);
771 __rds_send_complete(rs, rm, RDS_RDMA_CANCELED);
772 spin_unlock(&rs->rs_lock);
773
774 rm->m_rs = NULL;
775 spin_unlock_irqrestore(&rm->m_rs_lock, flags);
776
777 rds_message_put(rm);
778 }
779
780 rds_wake_sk_sleep(rs);
781
782 while (!list_empty(&list)) {
783 rm = list_entry(list.next, struct rds_message, m_sock_item);
784 list_del_init(&rm->m_sock_item);
785 rds_message_wait(rm);
786
787 /* just in case the code above skipped this message
788 * because RDS_MSG_ON_CONN wasn't set, run it again here
789 * taking m_rs_lock is the only thing that keeps us
790 * from racing with ack processing.
791 */
792 spin_lock_irqsave(&rm->m_rs_lock, flags);
793
794 spin_lock(&rs->rs_lock);
795 __rds_send_complete(rs, rm, RDS_RDMA_CANCELED);
796 spin_unlock(&rs->rs_lock);
797
798 rm->m_rs = NULL;
799 spin_unlock_irqrestore(&rm->m_rs_lock, flags);
800
801 rds_message_put(rm);
802 }
803}
804
805/*
806 * we only want this to fire once so we use the callers 'queued'. It's
807 * possible that another thread can race with us and remove the
808 * message from the flow with RDS_CANCEL_SENT_TO.
809 */
810static int rds_send_queue_rm(struct rds_sock *rs, struct rds_connection *conn,
811 struct rds_message *rm, __be16 sport,
812 __be16 dport, int *queued)
813{
814 unsigned long flags;
815 u32 len;
816
817 if (*queued)
818 goto out;
819
820 len = be32_to_cpu(rm->m_inc.i_hdr.h_len);
821
822 /* this is the only place which holds both the socket's rs_lock
823 * and the connection's c_lock */
824 spin_lock_irqsave(&rs->rs_lock, flags);
825
826 /*
827 * If there is a little space in sndbuf, we don't queue anything,
828 * and userspace gets -EAGAIN. But poll() indicates there's send
829 * room. This can lead to bad behavior (spinning) if snd_bytes isn't
830 * freed up by incoming acks. So we check the *old* value of
831 * rs_snd_bytes here to allow the last msg to exceed the buffer,
832 * and poll() now knows no more data can be sent.
833 */
834 if (rs->rs_snd_bytes < rds_sk_sndbuf(rs)) {
835 rs->rs_snd_bytes += len;
836
837 /* let recv side know we are close to send space exhaustion.
838 * This is probably not the optimal way to do it, as this
839 * means we set the flag on *all* messages as soon as our
840 * throughput hits a certain threshold.
841 */
842 if (rs->rs_snd_bytes >= rds_sk_sndbuf(rs) / 2)
843 __set_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags);
844
845 list_add_tail(&rm->m_sock_item, &rs->rs_send_queue);
846 set_bit(RDS_MSG_ON_SOCK, &rm->m_flags);
847 rds_message_addref(rm);
848 rm->m_rs = rs;
849
850 /* The code ordering is a little weird, but we're
851 trying to minimize the time we hold c_lock */
852 rds_message_populate_header(&rm->m_inc.i_hdr, sport, dport, 0);
853 rm->m_inc.i_conn = conn;
854 rds_message_addref(rm);
855
856 spin_lock(&conn->c_lock);
857 rm->m_inc.i_hdr.h_sequence = cpu_to_be64(conn->c_next_tx_seq++);
858 list_add_tail(&rm->m_conn_item, &conn->c_send_queue);
859 set_bit(RDS_MSG_ON_CONN, &rm->m_flags);
860 spin_unlock(&conn->c_lock);
861
862 rdsdebug("queued msg %p len %d, rs %p bytes %d seq %llu\n",
863 rm, len, rs, rs->rs_snd_bytes,
864 (unsigned long long)be64_to_cpu(rm->m_inc.i_hdr.h_sequence));
865
866 *queued = 1;
867 }
868
869 spin_unlock_irqrestore(&rs->rs_lock, flags);
870out:
871 return *queued;
872}
873
874/*
875 * rds_message is getting to be quite complicated, and we'd like to allocate
876 * it all in one go. This figures out how big it needs to be up front.
877 */
878static int rds_rm_size(struct msghdr *msg, int data_len)
879{
880 struct cmsghdr *cmsg;
881 int size = 0;
882 int cmsg_groups = 0;
883 int retval;
884
885 for_each_cmsghdr(cmsg, msg) {
886 if (!CMSG_OK(msg, cmsg))
887 return -EINVAL;
888
889 if (cmsg->cmsg_level != SOL_RDS)
890 continue;
891
892 switch (cmsg->cmsg_type) {
893 case RDS_CMSG_RDMA_ARGS:
894 cmsg_groups |= 1;
895 retval = rds_rdma_extra_size(CMSG_DATA(cmsg));
896 if (retval < 0)
897 return retval;
898 size += retval;
899
900 break;
901
902 case RDS_CMSG_RDMA_DEST:
903 case RDS_CMSG_RDMA_MAP:
904 cmsg_groups |= 2;
905 /* these are valid but do no add any size */
906 break;
907
908 case RDS_CMSG_ATOMIC_CSWP:
909 case RDS_CMSG_ATOMIC_FADD:
910 case RDS_CMSG_MASKED_ATOMIC_CSWP:
911 case RDS_CMSG_MASKED_ATOMIC_FADD:
912 cmsg_groups |= 1;
913 size += sizeof(struct scatterlist);
914 break;
915
916 default:
917 return -EINVAL;
918 }
919
920 }
921
922 size += ceil(data_len, PAGE_SIZE) * sizeof(struct scatterlist);
923
924 /* Ensure (DEST, MAP) are never used with (ARGS, ATOMIC) */
925 if (cmsg_groups == 3)
926 return -EINVAL;
927
928 return size;
929}
930
931static int rds_cmsg_send(struct rds_sock *rs, struct rds_message *rm,
932 struct msghdr *msg, int *allocated_mr)
933{
934 struct cmsghdr *cmsg;
935 int ret = 0;
936
937 for_each_cmsghdr(cmsg, msg) {
938 if (!CMSG_OK(msg, cmsg))
939 return -EINVAL;
940
941 if (cmsg->cmsg_level != SOL_RDS)
942 continue;
943
944 /* As a side effect, RDMA_DEST and RDMA_MAP will set
945 * rm->rdma.m_rdma_cookie and rm->rdma.m_rdma_mr.
946 */
947 switch (cmsg->cmsg_type) {
948 case RDS_CMSG_RDMA_ARGS:
949 ret = rds_cmsg_rdma_args(rs, rm, cmsg);
950 break;
951
952 case RDS_CMSG_RDMA_DEST:
953 ret = rds_cmsg_rdma_dest(rs, rm, cmsg);
954 break;
955
956 case RDS_CMSG_RDMA_MAP:
957 ret = rds_cmsg_rdma_map(rs, rm, cmsg);
958 if (!ret)
959 *allocated_mr = 1;
960 break;
961 case RDS_CMSG_ATOMIC_CSWP:
962 case RDS_CMSG_ATOMIC_FADD:
963 case RDS_CMSG_MASKED_ATOMIC_CSWP:
964 case RDS_CMSG_MASKED_ATOMIC_FADD:
965 ret = rds_cmsg_atomic(rs, rm, cmsg);
966 break;
967
968 default:
969 return -EINVAL;
970 }
971
972 if (ret)
973 break;
974 }
975
976 return ret;
977}
978
979int rds_sendmsg(struct socket *sock, struct msghdr *msg, size_t payload_len)
980{
981 struct sock *sk = sock->sk;
982 struct rds_sock *rs = rds_sk_to_rs(sk);
983 DECLARE_SOCKADDR(struct sockaddr_in *, usin, msg->msg_name);
984 __be32 daddr;
985 __be16 dport;
986 struct rds_message *rm = NULL;
987 struct rds_connection *conn;
988 int ret = 0;
989 int queued = 0, allocated_mr = 0;
990 int nonblock = msg->msg_flags & MSG_DONTWAIT;
991 long timeo = sock_sndtimeo(sk, nonblock);
992
993 /* Mirror Linux UDP mirror of BSD error message compatibility */
994 /* XXX: Perhaps MSG_MORE someday */
995 if (msg->msg_flags & ~(MSG_DONTWAIT | MSG_CMSG_COMPAT)) {
996 ret = -EOPNOTSUPP;
997 goto out;
998 }
999
1000 if (msg->msg_namelen) {
1001 /* XXX fail non-unicast destination IPs? */
1002 if (msg->msg_namelen < sizeof(*usin) || usin->sin_family != AF_INET) {
1003 ret = -EINVAL;
1004 goto out;
1005 }
1006 daddr = usin->sin_addr.s_addr;
1007 dport = usin->sin_port;
1008 } else {
1009 /* We only care about consistency with ->connect() */
1010 lock_sock(sk);
1011 daddr = rs->rs_conn_addr;
1012 dport = rs->rs_conn_port;
1013 release_sock(sk);
1014 }
1015
1016 lock_sock(sk);
1017 if (daddr == 0 || rs->rs_bound_addr == 0) {
1018 release_sock(sk);
1019 ret = -ENOTCONN; /* XXX not a great errno */
1020 goto out;
1021 }
1022 release_sock(sk);
1023
1024 if (payload_len > rds_sk_sndbuf(rs)) {
1025 ret = -EMSGSIZE;
1026 goto out;
1027 }
1028
1029 /* size of rm including all sgs */
1030 ret = rds_rm_size(msg, payload_len);
1031 if (ret < 0)
1032 goto out;
1033
1034 rm = rds_message_alloc(ret, GFP_KERNEL);
1035 if (!rm) {
1036 ret = -ENOMEM;
1037 goto out;
1038 }
1039
1040 /* Attach data to the rm */
1041 if (payload_len) {
1042 rm->data.op_sg = rds_message_alloc_sgs(rm, ceil(payload_len, PAGE_SIZE));
1043 if (!rm->data.op_sg) {
1044 ret = -ENOMEM;
1045 goto out;
1046 }
1047 ret = rds_message_copy_from_user(rm, &msg->msg_iter);
1048 if (ret)
1049 goto out;
1050 }
1051 rm->data.op_active = 1;
1052
1053 rm->m_daddr = daddr;
1054
1055 /* rds_conn_create has a spinlock that runs with IRQ off.
1056 * Caching the conn in the socket helps a lot. */
1057 if (rs->rs_conn && rs->rs_conn->c_faddr == daddr)
1058 conn = rs->rs_conn;
1059 else {
1060 conn = rds_conn_create_outgoing(sock_net(sock->sk),
1061 rs->rs_bound_addr, daddr,
1062 rs->rs_transport,
1063 sock->sk->sk_allocation);
1064 if (IS_ERR(conn)) {
1065 ret = PTR_ERR(conn);
1066 goto out;
1067 }
1068 rs->rs_conn = conn;
1069 }
1070
1071 /* Parse any control messages the user may have included. */
1072 ret = rds_cmsg_send(rs, rm, msg, &allocated_mr);
1073 if (ret)
1074 goto out;
1075
1076 if (rm->rdma.op_active && !conn->c_trans->xmit_rdma) {
1077 printk_ratelimited(KERN_NOTICE "rdma_op %p conn xmit_rdma %p\n",
1078 &rm->rdma, conn->c_trans->xmit_rdma);
1079 ret = -EOPNOTSUPP;
1080 goto out;
1081 }
1082
1083 if (rm->atomic.op_active && !conn->c_trans->xmit_atomic) {
1084 printk_ratelimited(KERN_NOTICE "atomic_op %p conn xmit_atomic %p\n",
1085 &rm->atomic, conn->c_trans->xmit_atomic);
1086 ret = -EOPNOTSUPP;
1087 goto out;
1088 }
1089
1090 rds_conn_connect_if_down(conn);
1091
1092 ret = rds_cong_wait(conn->c_fcong, dport, nonblock, rs);
1093 if (ret) {
1094 rs->rs_seen_congestion = 1;
1095 goto out;
1096 }
1097
1098 while (!rds_send_queue_rm(rs, conn, rm, rs->rs_bound_port,
1099 dport, &queued)) {
1100 rds_stats_inc(s_send_queue_full);
1101
1102 if (nonblock) {
1103 ret = -EAGAIN;
1104 goto out;
1105 }
1106
1107 timeo = wait_event_interruptible_timeout(*sk_sleep(sk),
1108 rds_send_queue_rm(rs, conn, rm,
1109 rs->rs_bound_port,
1110 dport,
1111 &queued),
1112 timeo);
1113 rdsdebug("sendmsg woke queued %d timeo %ld\n", queued, timeo);
1114 if (timeo > 0 || timeo == MAX_SCHEDULE_TIMEOUT)
1115 continue;
1116
1117 ret = timeo;
1118 if (ret == 0)
1119 ret = -ETIMEDOUT;
1120 goto out;
1121 }
1122
1123 /*
1124 * By now we've committed to the send. We reuse rds_send_worker()
1125 * to retry sends in the rds thread if the transport asks us to.
1126 */
1127 rds_stats_inc(s_send_queued);
1128
1129 ret = rds_send_xmit(conn);
1130 if (ret == -ENOMEM || ret == -EAGAIN)
1131 queue_delayed_work(rds_wq, &conn->c_send_w, 1);
1132
1133 rds_message_put(rm);
1134 return payload_len;
1135
1136out:
1137 /* If the user included a RDMA_MAP cmsg, we allocated a MR on the fly.
1138 * If the sendmsg goes through, we keep the MR. If it fails with EAGAIN
1139 * or in any other way, we need to destroy the MR again */
1140 if (allocated_mr)
1141 rds_rdma_unuse(rs, rds_rdma_cookie_key(rm->m_rdma_cookie), 1);
1142
1143 if (rm)
1144 rds_message_put(rm);
1145 return ret;
1146}
1147
1148/*
1149 * Reply to a ping packet.
1150 */
1151int
1152rds_send_pong(struct rds_connection *conn, __be16 dport)
1153{
1154 struct rds_message *rm;
1155 unsigned long flags;
1156 int ret = 0;
1157
1158 rm = rds_message_alloc(0, GFP_ATOMIC);
1159 if (!rm) {
1160 ret = -ENOMEM;
1161 goto out;
1162 }
1163
1164 rm->m_daddr = conn->c_faddr;
1165 rm->data.op_active = 1;
1166
1167 rds_conn_connect_if_down(conn);
1168
1169 ret = rds_cong_wait(conn->c_fcong, dport, 1, NULL);
1170 if (ret)
1171 goto out;
1172
1173 spin_lock_irqsave(&conn->c_lock, flags);
1174 list_add_tail(&rm->m_conn_item, &conn->c_send_queue);
1175 set_bit(RDS_MSG_ON_CONN, &rm->m_flags);
1176 rds_message_addref(rm);
1177 rm->m_inc.i_conn = conn;
1178
1179 rds_message_populate_header(&rm->m_inc.i_hdr, 0, dport,
1180 conn->c_next_tx_seq);
1181 conn->c_next_tx_seq++;
1182 spin_unlock_irqrestore(&conn->c_lock, flags);
1183
1184 rds_stats_inc(s_send_queued);
1185 rds_stats_inc(s_send_pong);
1186
1187 /* schedule the send work on rds_wq */
1188 queue_delayed_work(rds_wq, &conn->c_send_w, 1);
1189
1190 rds_message_put(rm);
1191 return 0;
1192
1193out:
1194 if (rm)
1195 rds_message_put(rm);
1196 return ret;
1197}
1/*
2 * Copyright (c) 2006, 2018 Oracle and/or its affiliates. All rights reserved.
3 *
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:
9 *
10 * Redistribution and use in source and binary forms, with or
11 * without modification, are permitted provided that the following
12 * conditions are met:
13 *
14 * - Redistributions of source code must retain the above
15 * copyright notice, this list of conditions and the following
16 * disclaimer.
17 *
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.
22 *
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
30 * SOFTWARE.
31 *
32 */
33#include <linux/kernel.h>
34#include <linux/moduleparam.h>
35#include <linux/gfp.h>
36#include <net/sock.h>
37#include <linux/in.h>
38#include <linux/list.h>
39#include <linux/ratelimit.h>
40#include <linux/export.h>
41#include <linux/sizes.h>
42
43#include "rds.h"
44
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
47 * will kick our shin.
48 * Also, it seems fairer to not let one busy connection stall all the
49 * others.
50 *
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
53 * drained the queue).
54 */
55static int send_batch_count = SZ_1K;
56module_param(send_batch_count, int, 0444);
57MODULE_PARM_DESC(send_batch_count, " batch factor when working the send queue");
58
59static void rds_send_remove_from_sock(struct list_head *messages, int status);
60
61/*
62 * Reset the send state. Callers must ensure that this doesn't race with
63 * rds_send_xmit().
64 */
65void rds_send_path_reset(struct rds_conn_path *cp)
66{
67 struct rds_message *rm, *tmp;
68 unsigned long flags;
69
70 if (cp->cp_xmit_rm) {
71 rm = cp->cp_xmit_rm;
72 cp->cp_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);
78 rds_message_put(rm);
79 }
80
81 cp->cp_xmit_sg = 0;
82 cp->cp_xmit_hdr_off = 0;
83 cp->cp_xmit_data_off = 0;
84 cp->cp_xmit_atomic_sent = 0;
85 cp->cp_xmit_rdma_sent = 0;
86 cp->cp_xmit_data_sent = 0;
87
88 cp->cp_conn->c_map_queued = 0;
89
90 cp->cp_unacked_packets = rds_sysctl_max_unacked_packets;
91 cp->cp_unacked_bytes = rds_sysctl_max_unacked_bytes;
92
93 /* Mark messages as retransmissions, and move them to the send q */
94 spin_lock_irqsave(&cp->cp_lock, flags);
95 list_for_each_entry_safe(rm, tmp, &cp->cp_retrans, m_conn_item) {
96 set_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags);
97 set_bit(RDS_MSG_RETRANSMITTED, &rm->m_flags);
98 }
99 list_splice_init(&cp->cp_retrans, &cp->cp_send_queue);
100 spin_unlock_irqrestore(&cp->cp_lock, flags);
101}
102EXPORT_SYMBOL_GPL(rds_send_path_reset);
103
104static int acquire_in_xmit(struct rds_conn_path *cp)
105{
106 return test_and_set_bit(RDS_IN_XMIT, &cp->cp_flags) == 0;
107}
108
109static void release_in_xmit(struct rds_conn_path *cp)
110{
111 clear_bit(RDS_IN_XMIT, &cp->cp_flags);
112 smp_mb__after_atomic();
113 /*
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.
118 */
119 if (waitqueue_active(&cp->cp_waitq))
120 wake_up_all(&cp->cp_waitq);
121}
122
123/*
124 * We're making the conscious trade-off here to only send one message
125 * down the connection at a time.
126 * Pro:
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
131 * Con:
132 * - queued acks can be delayed behind large messages
133 * Depends:
134 * - small message latency is higher behind queued large messages
135 * - large message latency isn't starved by intervening small sends
136 */
137int rds_send_xmit(struct rds_conn_path *cp)
138{
139 struct rds_connection *conn = cp->cp_conn;
140 struct rds_message *rm;
141 unsigned long flags;
142 unsigned int tmp;
143 struct scatterlist *sg;
144 int ret = 0;
145 LIST_HEAD(to_be_dropped);
146 int batch_count;
147 unsigned long send_gen = 0;
148
149restart:
150 batch_count = 0;
151
152 /*
153 * sendmsg calls here after having queued its message on the send
154 * queue. We only have one task feeding the connection at a time. If
155 * another thread is already feeding the queue then we back off. This
156 * avoids blocking the caller and trading per-connection data between
157 * caches per message.
158 */
159 if (!acquire_in_xmit(cp)) {
160 rds_stats_inc(s_send_lock_contention);
161 ret = -ENOMEM;
162 goto out;
163 }
164
165 if (rds_destroy_pending(cp->cp_conn)) {
166 release_in_xmit(cp);
167 ret = -ENETUNREACH; /* dont requeue send work */
168 goto out;
169 }
170
171 /*
172 * we record the send generation after doing the xmit acquire.
173 * if someone else manages to jump in and do some work, we'll use
174 * this to avoid a goto restart farther down.
175 *
176 * The acquire_in_xmit() check above ensures that only one
177 * caller can increment c_send_gen at any time.
178 */
179 send_gen = READ_ONCE(cp->cp_send_gen) + 1;
180 WRITE_ONCE(cp->cp_send_gen, send_gen);
181
182 /*
183 * rds_conn_shutdown() sets the conn state and then tests RDS_IN_XMIT,
184 * we do the opposite to avoid races.
185 */
186 if (!rds_conn_path_up(cp)) {
187 release_in_xmit(cp);
188 ret = 0;
189 goto out;
190 }
191
192 if (conn->c_trans->xmit_path_prepare)
193 conn->c_trans->xmit_path_prepare(cp);
194
195 /*
196 * spin trying to push headers and data down the connection until
197 * the connection doesn't make forward progress.
198 */
199 while (1) {
200
201 rm = cp->cp_xmit_rm;
202
203 /*
204 * If between sending messages, we can send a pending congestion
205 * map update.
206 */
207 if (!rm && test_and_clear_bit(0, &conn->c_map_queued)) {
208 rm = rds_cong_update_alloc(conn);
209 if (IS_ERR(rm)) {
210 ret = PTR_ERR(rm);
211 break;
212 }
213 rm->data.op_active = 1;
214 rm->m_inc.i_conn_path = cp;
215 rm->m_inc.i_conn = cp->cp_conn;
216
217 cp->cp_xmit_rm = rm;
218 }
219
220 /*
221 * If not already working on one, grab the next message.
222 *
223 * cp_xmit_rm holds a ref while we're sending this message down
224 * the connction. We can use this ref while holding the
225 * send_sem.. rds_send_reset() is serialized with it.
226 */
227 if (!rm) {
228 unsigned int len;
229
230 batch_count++;
231
232 /* we want to process as big a batch as we can, but
233 * we also want to avoid softlockups. If we've been
234 * through a lot of messages, lets back off and see
235 * if anyone else jumps in
236 */
237 if (batch_count >= send_batch_count)
238 goto over_batch;
239
240 spin_lock_irqsave(&cp->cp_lock, flags);
241
242 if (!list_empty(&cp->cp_send_queue)) {
243 rm = list_entry(cp->cp_send_queue.next,
244 struct rds_message,
245 m_conn_item);
246 rds_message_addref(rm);
247
248 /*
249 * Move the message from the send queue to the retransmit
250 * list right away.
251 */
252 list_move_tail(&rm->m_conn_item,
253 &cp->cp_retrans);
254 }
255
256 spin_unlock_irqrestore(&cp->cp_lock, flags);
257
258 if (!rm)
259 break;
260
261 /* Unfortunately, the way Infiniband deals with
262 * RDMA to a bad MR key is by moving the entire
263 * queue pair to error state. We cold possibly
264 * recover from that, but right now we drop the
265 * connection.
266 * Therefore, we never retransmit messages with RDMA ops.
267 */
268 if (test_bit(RDS_MSG_FLUSH, &rm->m_flags) ||
269 (rm->rdma.op_active &&
270 test_bit(RDS_MSG_RETRANSMITTED, &rm->m_flags))) {
271 spin_lock_irqsave(&cp->cp_lock, flags);
272 if (test_and_clear_bit(RDS_MSG_ON_CONN, &rm->m_flags))
273 list_move(&rm->m_conn_item, &to_be_dropped);
274 spin_unlock_irqrestore(&cp->cp_lock, flags);
275 continue;
276 }
277
278 /* Require an ACK every once in a while */
279 len = ntohl(rm->m_inc.i_hdr.h_len);
280 if (cp->cp_unacked_packets == 0 ||
281 cp->cp_unacked_bytes < len) {
282 set_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags);
283
284 cp->cp_unacked_packets =
285 rds_sysctl_max_unacked_packets;
286 cp->cp_unacked_bytes =
287 rds_sysctl_max_unacked_bytes;
288 rds_stats_inc(s_send_ack_required);
289 } else {
290 cp->cp_unacked_bytes -= len;
291 cp->cp_unacked_packets--;
292 }
293
294 cp->cp_xmit_rm = rm;
295 }
296
297 /* The transport either sends the whole rdma or none of it */
298 if (rm->rdma.op_active && !cp->cp_xmit_rdma_sent) {
299 rm->m_final_op = &rm->rdma;
300 /* The transport owns the mapped memory for now.
301 * You can't unmap it while it's on the send queue
302 */
303 set_bit(RDS_MSG_MAPPED, &rm->m_flags);
304 ret = conn->c_trans->xmit_rdma(conn, &rm->rdma);
305 if (ret) {
306 clear_bit(RDS_MSG_MAPPED, &rm->m_flags);
307 wake_up_interruptible(&rm->m_flush_wait);
308 break;
309 }
310 cp->cp_xmit_rdma_sent = 1;
311
312 }
313
314 if (rm->atomic.op_active && !cp->cp_xmit_atomic_sent) {
315 rm->m_final_op = &rm->atomic;
316 /* The transport owns the mapped memory for now.
317 * You can't unmap it while it's on the send queue
318 */
319 set_bit(RDS_MSG_MAPPED, &rm->m_flags);
320 ret = conn->c_trans->xmit_atomic(conn, &rm->atomic);
321 if (ret) {
322 clear_bit(RDS_MSG_MAPPED, &rm->m_flags);
323 wake_up_interruptible(&rm->m_flush_wait);
324 break;
325 }
326 cp->cp_xmit_atomic_sent = 1;
327
328 }
329
330 /*
331 * A number of cases require an RDS header to be sent
332 * even if there is no data.
333 * We permit 0-byte sends; rds-ping depends on this.
334 * However, if there are exclusively attached silent ops,
335 * we skip the hdr/data send, to enable silent operation.
336 */
337 if (rm->data.op_nents == 0) {
338 int ops_present;
339 int all_ops_are_silent = 1;
340
341 ops_present = (rm->atomic.op_active || rm->rdma.op_active);
342 if (rm->atomic.op_active && !rm->atomic.op_silent)
343 all_ops_are_silent = 0;
344 if (rm->rdma.op_active && !rm->rdma.op_silent)
345 all_ops_are_silent = 0;
346
347 if (ops_present && all_ops_are_silent
348 && !rm->m_rdma_cookie)
349 rm->data.op_active = 0;
350 }
351
352 if (rm->data.op_active && !cp->cp_xmit_data_sent) {
353 rm->m_final_op = &rm->data;
354
355 ret = conn->c_trans->xmit(conn, rm,
356 cp->cp_xmit_hdr_off,
357 cp->cp_xmit_sg,
358 cp->cp_xmit_data_off);
359 if (ret <= 0)
360 break;
361
362 if (cp->cp_xmit_hdr_off < sizeof(struct rds_header)) {
363 tmp = min_t(int, ret,
364 sizeof(struct rds_header) -
365 cp->cp_xmit_hdr_off);
366 cp->cp_xmit_hdr_off += tmp;
367 ret -= tmp;
368 }
369
370 sg = &rm->data.op_sg[cp->cp_xmit_sg];
371 while (ret) {
372 tmp = min_t(int, ret, sg->length -
373 cp->cp_xmit_data_off);
374 cp->cp_xmit_data_off += tmp;
375 ret -= tmp;
376 if (cp->cp_xmit_data_off == sg->length) {
377 cp->cp_xmit_data_off = 0;
378 sg++;
379 cp->cp_xmit_sg++;
380 BUG_ON(ret != 0 && cp->cp_xmit_sg ==
381 rm->data.op_nents);
382 }
383 }
384
385 if (cp->cp_xmit_hdr_off == sizeof(struct rds_header) &&
386 (cp->cp_xmit_sg == rm->data.op_nents))
387 cp->cp_xmit_data_sent = 1;
388 }
389
390 /*
391 * A rm will only take multiple times through this loop
392 * if there is a data op. Thus, if the data is sent (or there was
393 * none), then we're done with the rm.
394 */
395 if (!rm->data.op_active || cp->cp_xmit_data_sent) {
396 cp->cp_xmit_rm = NULL;
397 cp->cp_xmit_sg = 0;
398 cp->cp_xmit_hdr_off = 0;
399 cp->cp_xmit_data_off = 0;
400 cp->cp_xmit_rdma_sent = 0;
401 cp->cp_xmit_atomic_sent = 0;
402 cp->cp_xmit_data_sent = 0;
403
404 rds_message_put(rm);
405 }
406 }
407
408over_batch:
409 if (conn->c_trans->xmit_path_complete)
410 conn->c_trans->xmit_path_complete(cp);
411 release_in_xmit(cp);
412
413 /* Nuke any messages we decided not to retransmit. */
414 if (!list_empty(&to_be_dropped)) {
415 /* irqs on here, so we can put(), unlike above */
416 list_for_each_entry(rm, &to_be_dropped, m_conn_item)
417 rds_message_put(rm);
418 rds_send_remove_from_sock(&to_be_dropped, RDS_RDMA_DROPPED);
419 }
420
421 /*
422 * Other senders can queue a message after we last test the send queue
423 * but before we clear RDS_IN_XMIT. In that case they'd back off and
424 * not try and send their newly queued message. We need to check the
425 * send queue after having cleared RDS_IN_XMIT so that their message
426 * doesn't get stuck on the send queue.
427 *
428 * If the transport cannot continue (i.e ret != 0), then it must
429 * call us when more room is available, such as from the tx
430 * completion handler.
431 *
432 * We have an extra generation check here so that if someone manages
433 * to jump in after our release_in_xmit, we'll see that they have done
434 * some work and we will skip our goto
435 */
436 if (ret == 0) {
437 bool raced;
438
439 smp_mb();
440 raced = send_gen != READ_ONCE(cp->cp_send_gen);
441
442 if ((test_bit(0, &conn->c_map_queued) ||
443 !list_empty(&cp->cp_send_queue)) && !raced) {
444 if (batch_count < send_batch_count)
445 goto restart;
446 rcu_read_lock();
447 if (rds_destroy_pending(cp->cp_conn))
448 ret = -ENETUNREACH;
449 else
450 queue_delayed_work(rds_wq, &cp->cp_send_w, 1);
451 rcu_read_unlock();
452 } else if (raced) {
453 rds_stats_inc(s_send_lock_queue_raced);
454 }
455 }
456out:
457 return ret;
458}
459EXPORT_SYMBOL_GPL(rds_send_xmit);
460
461static void rds_send_sndbuf_remove(struct rds_sock *rs, struct rds_message *rm)
462{
463 u32 len = be32_to_cpu(rm->m_inc.i_hdr.h_len);
464
465 assert_spin_locked(&rs->rs_lock);
466
467 BUG_ON(rs->rs_snd_bytes < len);
468 rs->rs_snd_bytes -= len;
469
470 if (rs->rs_snd_bytes == 0)
471 rds_stats_inc(s_send_queue_empty);
472}
473
474static inline int rds_send_is_acked(struct rds_message *rm, u64 ack,
475 is_acked_func is_acked)
476{
477 if (is_acked)
478 return is_acked(rm, ack);
479 return be64_to_cpu(rm->m_inc.i_hdr.h_sequence) <= ack;
480}
481
482/*
483 * This is pretty similar to what happens below in the ACK
484 * handling code - except that we call here as soon as we get
485 * the IB send completion on the RDMA op and the accompanying
486 * message.
487 */
488void rds_rdma_send_complete(struct rds_message *rm, int status)
489{
490 struct rds_sock *rs = NULL;
491 struct rm_rdma_op *ro;
492 struct rds_notifier *notifier;
493 unsigned long flags;
494 unsigned int notify = 0;
495
496 spin_lock_irqsave(&rm->m_rs_lock, flags);
497
498 notify = rm->rdma.op_notify | rm->data.op_notify;
499 ro = &rm->rdma;
500 if (test_bit(RDS_MSG_ON_SOCK, &rm->m_flags) &&
501 ro->op_active && notify && ro->op_notifier) {
502 notifier = ro->op_notifier;
503 rs = rm->m_rs;
504 sock_hold(rds_rs_to_sk(rs));
505
506 notifier->n_status = status;
507 spin_lock(&rs->rs_lock);
508 list_add_tail(¬ifier->n_list, &rs->rs_notify_queue);
509 spin_unlock(&rs->rs_lock);
510
511 ro->op_notifier = NULL;
512 }
513
514 spin_unlock_irqrestore(&rm->m_rs_lock, flags);
515
516 if (rs) {
517 rds_wake_sk_sleep(rs);
518 sock_put(rds_rs_to_sk(rs));
519 }
520}
521EXPORT_SYMBOL_GPL(rds_rdma_send_complete);
522
523/*
524 * Just like above, except looks at atomic op
525 */
526void rds_atomic_send_complete(struct rds_message *rm, int status)
527{
528 struct rds_sock *rs = NULL;
529 struct rm_atomic_op *ao;
530 struct rds_notifier *notifier;
531 unsigned long flags;
532
533 spin_lock_irqsave(&rm->m_rs_lock, flags);
534
535 ao = &rm->atomic;
536 if (test_bit(RDS_MSG_ON_SOCK, &rm->m_flags)
537 && ao->op_active && ao->op_notify && ao->op_notifier) {
538 notifier = ao->op_notifier;
539 rs = rm->m_rs;
540 sock_hold(rds_rs_to_sk(rs));
541
542 notifier->n_status = status;
543 spin_lock(&rs->rs_lock);
544 list_add_tail(¬ifier->n_list, &rs->rs_notify_queue);
545 spin_unlock(&rs->rs_lock);
546
547 ao->op_notifier = NULL;
548 }
549
550 spin_unlock_irqrestore(&rm->m_rs_lock, flags);
551
552 if (rs) {
553 rds_wake_sk_sleep(rs);
554 sock_put(rds_rs_to_sk(rs));
555 }
556}
557EXPORT_SYMBOL_GPL(rds_atomic_send_complete);
558
559/*
560 * This is the same as rds_rdma_send_complete except we
561 * don't do any locking - we have all the ingredients (message,
562 * socket, socket lock) and can just move the notifier.
563 */
564static inline void
565__rds_send_complete(struct rds_sock *rs, struct rds_message *rm, int status)
566{
567 struct rm_rdma_op *ro;
568 struct rm_atomic_op *ao;
569
570 ro = &rm->rdma;
571 if (ro->op_active && ro->op_notify && ro->op_notifier) {
572 ro->op_notifier->n_status = status;
573 list_add_tail(&ro->op_notifier->n_list, &rs->rs_notify_queue);
574 ro->op_notifier = NULL;
575 }
576
577 ao = &rm->atomic;
578 if (ao->op_active && ao->op_notify && ao->op_notifier) {
579 ao->op_notifier->n_status = status;
580 list_add_tail(&ao->op_notifier->n_list, &rs->rs_notify_queue);
581 ao->op_notifier = NULL;
582 }
583
584 /* No need to wake the app - caller does this */
585}
586
587/*
588 * This removes messages from the socket's list if they're on it. The list
589 * argument must be private to the caller, we must be able to modify it
590 * without locks. The messages must have a reference held for their
591 * position on the list. This function will drop that reference after
592 * removing the messages from the 'messages' list regardless of if it found
593 * the messages on the socket list or not.
594 */
595static void rds_send_remove_from_sock(struct list_head *messages, int status)
596{
597 unsigned long flags;
598 struct rds_sock *rs = NULL;
599 struct rds_message *rm;
600
601 while (!list_empty(messages)) {
602 int was_on_sock = 0;
603
604 rm = list_entry(messages->next, struct rds_message,
605 m_conn_item);
606 list_del_init(&rm->m_conn_item);
607
608 /*
609 * If we see this flag cleared then we're *sure* that someone
610 * else beat us to removing it from the sock. If we race
611 * with their flag update we'll get the lock and then really
612 * see that the flag has been cleared.
613 *
614 * The message spinlock makes sure nobody clears rm->m_rs
615 * while we're messing with it. It does not prevent the
616 * message from being removed from the socket, though.
617 */
618 spin_lock_irqsave(&rm->m_rs_lock, flags);
619 if (!test_bit(RDS_MSG_ON_SOCK, &rm->m_flags))
620 goto unlock_and_drop;
621
622 if (rs != rm->m_rs) {
623 if (rs) {
624 rds_wake_sk_sleep(rs);
625 sock_put(rds_rs_to_sk(rs));
626 }
627 rs = rm->m_rs;
628 if (rs)
629 sock_hold(rds_rs_to_sk(rs));
630 }
631 if (!rs)
632 goto unlock_and_drop;
633 spin_lock(&rs->rs_lock);
634
635 if (test_and_clear_bit(RDS_MSG_ON_SOCK, &rm->m_flags)) {
636 struct rm_rdma_op *ro = &rm->rdma;
637 struct rds_notifier *notifier;
638
639 list_del_init(&rm->m_sock_item);
640 rds_send_sndbuf_remove(rs, rm);
641
642 if (ro->op_active && ro->op_notifier &&
643 (ro->op_notify || (ro->op_recverr && status))) {
644 notifier = ro->op_notifier;
645 list_add_tail(¬ifier->n_list,
646 &rs->rs_notify_queue);
647 if (!notifier->n_status)
648 notifier->n_status = status;
649 rm->rdma.op_notifier = NULL;
650 }
651 was_on_sock = 1;
652 }
653 spin_unlock(&rs->rs_lock);
654
655unlock_and_drop:
656 spin_unlock_irqrestore(&rm->m_rs_lock, flags);
657 rds_message_put(rm);
658 if (was_on_sock)
659 rds_message_put(rm);
660 }
661
662 if (rs) {
663 rds_wake_sk_sleep(rs);
664 sock_put(rds_rs_to_sk(rs));
665 }
666}
667
668/*
669 * Transports call here when they've determined that the receiver queued
670 * messages up to, and including, the given sequence number. Messages are
671 * moved to the retrans queue when rds_send_xmit picks them off the send
672 * queue. This means that in the TCP case, the message may not have been
673 * assigned the m_ack_seq yet - but that's fine as long as tcp_is_acked
674 * checks the RDS_MSG_HAS_ACK_SEQ bit.
675 */
676void rds_send_path_drop_acked(struct rds_conn_path *cp, u64 ack,
677 is_acked_func is_acked)
678{
679 struct rds_message *rm, *tmp;
680 unsigned long flags;
681 LIST_HEAD(list);
682
683 spin_lock_irqsave(&cp->cp_lock, flags);
684
685 list_for_each_entry_safe(rm, tmp, &cp->cp_retrans, m_conn_item) {
686 if (!rds_send_is_acked(rm, ack, is_acked))
687 break;
688
689 list_move(&rm->m_conn_item, &list);
690 clear_bit(RDS_MSG_ON_CONN, &rm->m_flags);
691 }
692
693 /* order flag updates with spin locks */
694 if (!list_empty(&list))
695 smp_mb__after_atomic();
696
697 spin_unlock_irqrestore(&cp->cp_lock, flags);
698
699 /* now remove the messages from the sock list as needed */
700 rds_send_remove_from_sock(&list, RDS_RDMA_SUCCESS);
701}
702EXPORT_SYMBOL_GPL(rds_send_path_drop_acked);
703
704void rds_send_drop_acked(struct rds_connection *conn, u64 ack,
705 is_acked_func is_acked)
706{
707 WARN_ON(conn->c_trans->t_mp_capable);
708 rds_send_path_drop_acked(&conn->c_path[0], ack, is_acked);
709}
710EXPORT_SYMBOL_GPL(rds_send_drop_acked);
711
712void rds_send_drop_to(struct rds_sock *rs, struct sockaddr_in *dest)
713{
714 struct rds_message *rm, *tmp;
715 struct rds_connection *conn;
716 struct rds_conn_path *cp;
717 unsigned long flags;
718 LIST_HEAD(list);
719
720 /* get all the messages we're dropping under the rs lock */
721 spin_lock_irqsave(&rs->rs_lock, flags);
722
723 list_for_each_entry_safe(rm, tmp, &rs->rs_send_queue, m_sock_item) {
724 if (dest && (dest->sin_addr.s_addr != rm->m_daddr ||
725 dest->sin_port != rm->m_inc.i_hdr.h_dport))
726 continue;
727
728 list_move(&rm->m_sock_item, &list);
729 rds_send_sndbuf_remove(rs, rm);
730 clear_bit(RDS_MSG_ON_SOCK, &rm->m_flags);
731 }
732
733 /* order flag updates with the rs lock */
734 smp_mb__after_atomic();
735
736 spin_unlock_irqrestore(&rs->rs_lock, flags);
737
738 if (list_empty(&list))
739 return;
740
741 /* Remove the messages from the conn */
742 list_for_each_entry(rm, &list, m_sock_item) {
743
744 conn = rm->m_inc.i_conn;
745 if (conn->c_trans->t_mp_capable)
746 cp = rm->m_inc.i_conn_path;
747 else
748 cp = &conn->c_path[0];
749
750 spin_lock_irqsave(&cp->cp_lock, flags);
751 /*
752 * Maybe someone else beat us to removing rm from the conn.
753 * If we race with their flag update we'll get the lock and
754 * then really see that the flag has been cleared.
755 */
756 if (!test_and_clear_bit(RDS_MSG_ON_CONN, &rm->m_flags)) {
757 spin_unlock_irqrestore(&cp->cp_lock, flags);
758 continue;
759 }
760 list_del_init(&rm->m_conn_item);
761 spin_unlock_irqrestore(&cp->cp_lock, flags);
762
763 /*
764 * Couldn't grab m_rs_lock in top loop (lock ordering),
765 * but we can now.
766 */
767 spin_lock_irqsave(&rm->m_rs_lock, flags);
768
769 spin_lock(&rs->rs_lock);
770 __rds_send_complete(rs, rm, RDS_RDMA_CANCELED);
771 spin_unlock(&rs->rs_lock);
772
773 spin_unlock_irqrestore(&rm->m_rs_lock, flags);
774
775 rds_message_put(rm);
776 }
777
778 rds_wake_sk_sleep(rs);
779
780 while (!list_empty(&list)) {
781 rm = list_entry(list.next, struct rds_message, m_sock_item);
782 list_del_init(&rm->m_sock_item);
783 rds_message_wait(rm);
784
785 /* just in case the code above skipped this message
786 * because RDS_MSG_ON_CONN wasn't set, run it again here
787 * taking m_rs_lock is the only thing that keeps us
788 * from racing with ack processing.
789 */
790 spin_lock_irqsave(&rm->m_rs_lock, flags);
791
792 spin_lock(&rs->rs_lock);
793 __rds_send_complete(rs, rm, RDS_RDMA_CANCELED);
794 spin_unlock(&rs->rs_lock);
795
796 spin_unlock_irqrestore(&rm->m_rs_lock, flags);
797
798 rds_message_put(rm);
799 }
800}
801
802/*
803 * we only want this to fire once so we use the callers 'queued'. It's
804 * possible that another thread can race with us and remove the
805 * message from the flow with RDS_CANCEL_SENT_TO.
806 */
807static int rds_send_queue_rm(struct rds_sock *rs, struct rds_connection *conn,
808 struct rds_conn_path *cp,
809 struct rds_message *rm, __be16 sport,
810 __be16 dport, int *queued)
811{
812 unsigned long flags;
813 u32 len;
814
815 if (*queued)
816 goto out;
817
818 len = be32_to_cpu(rm->m_inc.i_hdr.h_len);
819
820 /* this is the only place which holds both the socket's rs_lock
821 * and the connection's c_lock */
822 spin_lock_irqsave(&rs->rs_lock, flags);
823
824 /*
825 * If there is a little space in sndbuf, we don't queue anything,
826 * and userspace gets -EAGAIN. But poll() indicates there's send
827 * room. This can lead to bad behavior (spinning) if snd_bytes isn't
828 * freed up by incoming acks. So we check the *old* value of
829 * rs_snd_bytes here to allow the last msg to exceed the buffer,
830 * and poll() now knows no more data can be sent.
831 */
832 if (rs->rs_snd_bytes < rds_sk_sndbuf(rs)) {
833 rs->rs_snd_bytes += len;
834
835 /* let recv side know we are close to send space exhaustion.
836 * This is probably not the optimal way to do it, as this
837 * means we set the flag on *all* messages as soon as our
838 * throughput hits a certain threshold.
839 */
840 if (rs->rs_snd_bytes >= rds_sk_sndbuf(rs) / 2)
841 set_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags);
842
843 list_add_tail(&rm->m_sock_item, &rs->rs_send_queue);
844 set_bit(RDS_MSG_ON_SOCK, &rm->m_flags);
845 rds_message_addref(rm);
846 sock_hold(rds_rs_to_sk(rs));
847 rm->m_rs = rs;
848
849 /* The code ordering is a little weird, but we're
850 trying to minimize the time we hold c_lock */
851 rds_message_populate_header(&rm->m_inc.i_hdr, sport, dport, 0);
852 rm->m_inc.i_conn = conn;
853 rm->m_inc.i_conn_path = cp;
854 rds_message_addref(rm);
855
856 spin_lock(&cp->cp_lock);
857 rm->m_inc.i_hdr.h_sequence = cpu_to_be64(cp->cp_next_tx_seq++);
858 list_add_tail(&rm->m_conn_item, &cp->cp_send_queue);
859 set_bit(RDS_MSG_ON_CONN, &rm->m_flags);
860 spin_unlock(&cp->cp_lock);
861
862 rdsdebug("queued msg %p len %d, rs %p bytes %d seq %llu\n",
863 rm, len, rs, rs->rs_snd_bytes,
864 (unsigned long long)be64_to_cpu(rm->m_inc.i_hdr.h_sequence));
865
866 *queued = 1;
867 }
868
869 spin_unlock_irqrestore(&rs->rs_lock, flags);
870out:
871 return *queued;
872}
873
874/*
875 * rds_message is getting to be quite complicated, and we'd like to allocate
876 * it all in one go. This figures out how big it needs to be up front.
877 */
878static int rds_rm_size(struct msghdr *msg, int num_sgs)
879{
880 struct cmsghdr *cmsg;
881 int size = 0;
882 int cmsg_groups = 0;
883 int retval;
884 bool zcopy_cookie = false;
885
886 for_each_cmsghdr(cmsg, msg) {
887 if (!CMSG_OK(msg, cmsg))
888 return -EINVAL;
889
890 if (cmsg->cmsg_level != SOL_RDS)
891 continue;
892
893 switch (cmsg->cmsg_type) {
894 case RDS_CMSG_RDMA_ARGS:
895 cmsg_groups |= 1;
896 retval = rds_rdma_extra_size(CMSG_DATA(cmsg));
897 if (retval < 0)
898 return retval;
899 size += retval;
900
901 break;
902
903 case RDS_CMSG_ZCOPY_COOKIE:
904 zcopy_cookie = true;
905 /* fall through */
906
907 case RDS_CMSG_RDMA_DEST:
908 case RDS_CMSG_RDMA_MAP:
909 cmsg_groups |= 2;
910 /* these are valid but do no add any size */
911 break;
912
913 case RDS_CMSG_ATOMIC_CSWP:
914 case RDS_CMSG_ATOMIC_FADD:
915 case RDS_CMSG_MASKED_ATOMIC_CSWP:
916 case RDS_CMSG_MASKED_ATOMIC_FADD:
917 cmsg_groups |= 1;
918 size += sizeof(struct scatterlist);
919 break;
920
921 default:
922 return -EINVAL;
923 }
924
925 }
926
927 if ((msg->msg_flags & MSG_ZEROCOPY) && !zcopy_cookie)
928 return -EINVAL;
929
930 size += num_sgs * sizeof(struct scatterlist);
931
932 /* Ensure (DEST, MAP) are never used with (ARGS, ATOMIC) */
933 if (cmsg_groups == 3)
934 return -EINVAL;
935
936 return size;
937}
938
939static int rds_cmsg_zcopy(struct rds_sock *rs, struct rds_message *rm,
940 struct cmsghdr *cmsg)
941{
942 u32 *cookie;
943
944 if (cmsg->cmsg_len < CMSG_LEN(sizeof(*cookie)) ||
945 !rm->data.op_mmp_znotifier)
946 return -EINVAL;
947 cookie = CMSG_DATA(cmsg);
948 rm->data.op_mmp_znotifier->z_cookie = *cookie;
949 return 0;
950}
951
952static int rds_cmsg_send(struct rds_sock *rs, struct rds_message *rm,
953 struct msghdr *msg, int *allocated_mr)
954{
955 struct cmsghdr *cmsg;
956 int ret = 0;
957
958 for_each_cmsghdr(cmsg, msg) {
959 if (!CMSG_OK(msg, cmsg))
960 return -EINVAL;
961
962 if (cmsg->cmsg_level != SOL_RDS)
963 continue;
964
965 /* As a side effect, RDMA_DEST and RDMA_MAP will set
966 * rm->rdma.m_rdma_cookie and rm->rdma.m_rdma_mr.
967 */
968 switch (cmsg->cmsg_type) {
969 case RDS_CMSG_RDMA_ARGS:
970 ret = rds_cmsg_rdma_args(rs, rm, cmsg);
971 break;
972
973 case RDS_CMSG_RDMA_DEST:
974 ret = rds_cmsg_rdma_dest(rs, rm, cmsg);
975 break;
976
977 case RDS_CMSG_RDMA_MAP:
978 ret = rds_cmsg_rdma_map(rs, rm, cmsg);
979 if (!ret)
980 *allocated_mr = 1;
981 else if (ret == -ENODEV)
982 /* Accommodate the get_mr() case which can fail
983 * if connection isn't established yet.
984 */
985 ret = -EAGAIN;
986 break;
987 case RDS_CMSG_ATOMIC_CSWP:
988 case RDS_CMSG_ATOMIC_FADD:
989 case RDS_CMSG_MASKED_ATOMIC_CSWP:
990 case RDS_CMSG_MASKED_ATOMIC_FADD:
991 ret = rds_cmsg_atomic(rs, rm, cmsg);
992 break;
993
994 case RDS_CMSG_ZCOPY_COOKIE:
995 ret = rds_cmsg_zcopy(rs, rm, cmsg);
996 break;
997
998 default:
999 return -EINVAL;
1000 }
1001
1002 if (ret)
1003 break;
1004 }
1005
1006 return ret;
1007}
1008
1009static int rds_send_mprds_hash(struct rds_sock *rs, struct rds_connection *conn)
1010{
1011 int hash;
1012
1013 if (conn->c_npaths == 0)
1014 hash = RDS_MPATH_HASH(rs, RDS_MPATH_WORKERS);
1015 else
1016 hash = RDS_MPATH_HASH(rs, conn->c_npaths);
1017 if (conn->c_npaths == 0 && hash != 0) {
1018 rds_send_ping(conn, 0);
1019
1020 /* The underlying connection is not up yet. Need to wait
1021 * until it is up to be sure that the non-zero c_path can be
1022 * used. But if we are interrupted, we have to use the zero
1023 * c_path in case the connection ends up being non-MP capable.
1024 */
1025 if (conn->c_npaths == 0)
1026 if (wait_event_interruptible(conn->c_hs_waitq,
1027 conn->c_npaths != 0))
1028 hash = 0;
1029 if (conn->c_npaths == 1)
1030 hash = 0;
1031 }
1032 return hash;
1033}
1034
1035static int rds_rdma_bytes(struct msghdr *msg, size_t *rdma_bytes)
1036{
1037 struct rds_rdma_args *args;
1038 struct cmsghdr *cmsg;
1039
1040 for_each_cmsghdr(cmsg, msg) {
1041 if (!CMSG_OK(msg, cmsg))
1042 return -EINVAL;
1043
1044 if (cmsg->cmsg_level != SOL_RDS)
1045 continue;
1046
1047 if (cmsg->cmsg_type == RDS_CMSG_RDMA_ARGS) {
1048 if (cmsg->cmsg_len <
1049 CMSG_LEN(sizeof(struct rds_rdma_args)))
1050 return -EINVAL;
1051 args = CMSG_DATA(cmsg);
1052 *rdma_bytes += args->remote_vec.bytes;
1053 }
1054 }
1055 return 0;
1056}
1057
1058int rds_sendmsg(struct socket *sock, struct msghdr *msg, size_t payload_len)
1059{
1060 struct sock *sk = sock->sk;
1061 struct rds_sock *rs = rds_sk_to_rs(sk);
1062 DECLARE_SOCKADDR(struct sockaddr_in *, usin, msg->msg_name);
1063 __be32 daddr;
1064 __be16 dport;
1065 struct rds_message *rm = NULL;
1066 struct rds_connection *conn;
1067 int ret = 0;
1068 int queued = 0, allocated_mr = 0;
1069 int nonblock = msg->msg_flags & MSG_DONTWAIT;
1070 long timeo = sock_sndtimeo(sk, nonblock);
1071 struct rds_conn_path *cpath;
1072 size_t total_payload_len = payload_len, rdma_payload_len = 0;
1073 bool zcopy = ((msg->msg_flags & MSG_ZEROCOPY) &&
1074 sock_flag(rds_rs_to_sk(rs), SOCK_ZEROCOPY));
1075 int num_sgs = ceil(payload_len, PAGE_SIZE);
1076
1077 /* Mirror Linux UDP mirror of BSD error message compatibility */
1078 /* XXX: Perhaps MSG_MORE someday */
1079 if (msg->msg_flags & ~(MSG_DONTWAIT | MSG_CMSG_COMPAT | MSG_ZEROCOPY)) {
1080 ret = -EOPNOTSUPP;
1081 goto out;
1082 }
1083
1084 if (msg->msg_namelen) {
1085 /* XXX fail non-unicast destination IPs? */
1086 if (msg->msg_namelen < sizeof(*usin) || usin->sin_family != AF_INET) {
1087 ret = -EINVAL;
1088 goto out;
1089 }
1090 daddr = usin->sin_addr.s_addr;
1091 dport = usin->sin_port;
1092 } else {
1093 /* We only care about consistency with ->connect() */
1094 lock_sock(sk);
1095 daddr = rs->rs_conn_addr;
1096 dport = rs->rs_conn_port;
1097 release_sock(sk);
1098 }
1099
1100 lock_sock(sk);
1101 if (daddr == 0 || rs->rs_bound_addr == 0) {
1102 release_sock(sk);
1103 ret = -ENOTCONN; /* XXX not a great errno */
1104 goto out;
1105 }
1106 release_sock(sk);
1107
1108 ret = rds_rdma_bytes(msg, &rdma_payload_len);
1109 if (ret)
1110 goto out;
1111
1112 total_payload_len += rdma_payload_len;
1113 if (max_t(size_t, payload_len, rdma_payload_len) > RDS_MAX_MSG_SIZE) {
1114 ret = -EMSGSIZE;
1115 goto out;
1116 }
1117
1118 if (payload_len > rds_sk_sndbuf(rs)) {
1119 ret = -EMSGSIZE;
1120 goto out;
1121 }
1122
1123 if (zcopy) {
1124 if (rs->rs_transport->t_type != RDS_TRANS_TCP) {
1125 ret = -EOPNOTSUPP;
1126 goto out;
1127 }
1128 num_sgs = iov_iter_npages(&msg->msg_iter, INT_MAX);
1129 }
1130 /* size of rm including all sgs */
1131 ret = rds_rm_size(msg, num_sgs);
1132 if (ret < 0)
1133 goto out;
1134
1135 rm = rds_message_alloc(ret, GFP_KERNEL);
1136 if (!rm) {
1137 ret = -ENOMEM;
1138 goto out;
1139 }
1140
1141 /* Attach data to the rm */
1142 if (payload_len) {
1143 rm->data.op_sg = rds_message_alloc_sgs(rm, num_sgs);
1144 if (!rm->data.op_sg) {
1145 ret = -ENOMEM;
1146 goto out;
1147 }
1148 ret = rds_message_copy_from_user(rm, &msg->msg_iter, zcopy);
1149 if (ret)
1150 goto out;
1151 }
1152 rm->data.op_active = 1;
1153
1154 rm->m_daddr = daddr;
1155
1156 /* rds_conn_create has a spinlock that runs with IRQ off.
1157 * Caching the conn in the socket helps a lot. */
1158 if (rs->rs_conn && rs->rs_conn->c_faddr == daddr)
1159 conn = rs->rs_conn;
1160 else {
1161 conn = rds_conn_create_outgoing(sock_net(sock->sk),
1162 rs->rs_bound_addr, daddr,
1163 rs->rs_transport,
1164 sock->sk->sk_allocation);
1165 if (IS_ERR(conn)) {
1166 ret = PTR_ERR(conn);
1167 goto out;
1168 }
1169 rs->rs_conn = conn;
1170 }
1171
1172 /* Parse any control messages the user may have included. */
1173 ret = rds_cmsg_send(rs, rm, msg, &allocated_mr);
1174 if (ret) {
1175 /* Trigger connection so that its ready for the next retry */
1176 if (ret == -EAGAIN)
1177 rds_conn_connect_if_down(conn);
1178 goto out;
1179 }
1180
1181 if (rm->rdma.op_active && !conn->c_trans->xmit_rdma) {
1182 printk_ratelimited(KERN_NOTICE "rdma_op %p conn xmit_rdma %p\n",
1183 &rm->rdma, conn->c_trans->xmit_rdma);
1184 ret = -EOPNOTSUPP;
1185 goto out;
1186 }
1187
1188 if (rm->atomic.op_active && !conn->c_trans->xmit_atomic) {
1189 printk_ratelimited(KERN_NOTICE "atomic_op %p conn xmit_atomic %p\n",
1190 &rm->atomic, conn->c_trans->xmit_atomic);
1191 ret = -EOPNOTSUPP;
1192 goto out;
1193 }
1194
1195 if (conn->c_trans->t_mp_capable)
1196 cpath = &conn->c_path[rds_send_mprds_hash(rs, conn)];
1197 else
1198 cpath = &conn->c_path[0];
1199
1200 if (rds_destroy_pending(conn)) {
1201 ret = -EAGAIN;
1202 goto out;
1203 }
1204
1205 rds_conn_path_connect_if_down(cpath);
1206
1207 ret = rds_cong_wait(conn->c_fcong, dport, nonblock, rs);
1208 if (ret) {
1209 rs->rs_seen_congestion = 1;
1210 goto out;
1211 }
1212 while (!rds_send_queue_rm(rs, conn, cpath, rm, rs->rs_bound_port,
1213 dport, &queued)) {
1214 rds_stats_inc(s_send_queue_full);
1215
1216 if (nonblock) {
1217 ret = -EAGAIN;
1218 goto out;
1219 }
1220
1221 timeo = wait_event_interruptible_timeout(*sk_sleep(sk),
1222 rds_send_queue_rm(rs, conn, cpath, rm,
1223 rs->rs_bound_port,
1224 dport,
1225 &queued),
1226 timeo);
1227 rdsdebug("sendmsg woke queued %d timeo %ld\n", queued, timeo);
1228 if (timeo > 0 || timeo == MAX_SCHEDULE_TIMEOUT)
1229 continue;
1230
1231 ret = timeo;
1232 if (ret == 0)
1233 ret = -ETIMEDOUT;
1234 goto out;
1235 }
1236
1237 /*
1238 * By now we've committed to the send. We reuse rds_send_worker()
1239 * to retry sends in the rds thread if the transport asks us to.
1240 */
1241 rds_stats_inc(s_send_queued);
1242
1243 ret = rds_send_xmit(cpath);
1244 if (ret == -ENOMEM || ret == -EAGAIN) {
1245 ret = 0;
1246 rcu_read_lock();
1247 if (rds_destroy_pending(cpath->cp_conn))
1248 ret = -ENETUNREACH;
1249 else
1250 queue_delayed_work(rds_wq, &cpath->cp_send_w, 1);
1251 rcu_read_unlock();
1252 }
1253 if (ret)
1254 goto out;
1255 rds_message_put(rm);
1256 return payload_len;
1257
1258out:
1259 /* If the user included a RDMA_MAP cmsg, we allocated a MR on the fly.
1260 * If the sendmsg goes through, we keep the MR. If it fails with EAGAIN
1261 * or in any other way, we need to destroy the MR again */
1262 if (allocated_mr)
1263 rds_rdma_unuse(rs, rds_rdma_cookie_key(rm->m_rdma_cookie), 1);
1264
1265 if (rm)
1266 rds_message_put(rm);
1267 return ret;
1268}
1269
1270/*
1271 * send out a probe. Can be shared by rds_send_ping,
1272 * rds_send_pong, rds_send_hb.
1273 * rds_send_hb should use h_flags
1274 * RDS_FLAG_HB_PING|RDS_FLAG_ACK_REQUIRED
1275 * or
1276 * RDS_FLAG_HB_PONG|RDS_FLAG_ACK_REQUIRED
1277 */
1278static int
1279rds_send_probe(struct rds_conn_path *cp, __be16 sport,
1280 __be16 dport, u8 h_flags)
1281{
1282 struct rds_message *rm;
1283 unsigned long flags;
1284 int ret = 0;
1285
1286 rm = rds_message_alloc(0, GFP_ATOMIC);
1287 if (!rm) {
1288 ret = -ENOMEM;
1289 goto out;
1290 }
1291
1292 rm->m_daddr = cp->cp_conn->c_faddr;
1293 rm->data.op_active = 1;
1294
1295 rds_conn_path_connect_if_down(cp);
1296
1297 ret = rds_cong_wait(cp->cp_conn->c_fcong, dport, 1, NULL);
1298 if (ret)
1299 goto out;
1300
1301 spin_lock_irqsave(&cp->cp_lock, flags);
1302 list_add_tail(&rm->m_conn_item, &cp->cp_send_queue);
1303 set_bit(RDS_MSG_ON_CONN, &rm->m_flags);
1304 rds_message_addref(rm);
1305 rm->m_inc.i_conn = cp->cp_conn;
1306 rm->m_inc.i_conn_path = cp;
1307
1308 rds_message_populate_header(&rm->m_inc.i_hdr, sport, dport,
1309 cp->cp_next_tx_seq);
1310 rm->m_inc.i_hdr.h_flags |= h_flags;
1311 cp->cp_next_tx_seq++;
1312
1313 if (RDS_HS_PROBE(be16_to_cpu(sport), be16_to_cpu(dport)) &&
1314 cp->cp_conn->c_trans->t_mp_capable) {
1315 u16 npaths = cpu_to_be16(RDS_MPATH_WORKERS);
1316 u32 my_gen_num = cpu_to_be32(cp->cp_conn->c_my_gen_num);
1317
1318 rds_message_add_extension(&rm->m_inc.i_hdr,
1319 RDS_EXTHDR_NPATHS, &npaths,
1320 sizeof(npaths));
1321 rds_message_add_extension(&rm->m_inc.i_hdr,
1322 RDS_EXTHDR_GEN_NUM,
1323 &my_gen_num,
1324 sizeof(u32));
1325 }
1326 spin_unlock_irqrestore(&cp->cp_lock, flags);
1327
1328 rds_stats_inc(s_send_queued);
1329 rds_stats_inc(s_send_pong);
1330
1331 /* schedule the send work on rds_wq */
1332 rcu_read_lock();
1333 if (!rds_destroy_pending(cp->cp_conn))
1334 queue_delayed_work(rds_wq, &cp->cp_send_w, 1);
1335 rcu_read_unlock();
1336
1337 rds_message_put(rm);
1338 return 0;
1339
1340out:
1341 if (rm)
1342 rds_message_put(rm);
1343 return ret;
1344}
1345
1346int
1347rds_send_pong(struct rds_conn_path *cp, __be16 dport)
1348{
1349 return rds_send_probe(cp, 0, dport, 0);
1350}
1351
1352void
1353rds_send_ping(struct rds_connection *conn, int cp_index)
1354{
1355 unsigned long flags;
1356 struct rds_conn_path *cp = &conn->c_path[cp_index];
1357
1358 spin_lock_irqsave(&cp->cp_lock, flags);
1359 if (conn->c_ping_triggered) {
1360 spin_unlock_irqrestore(&cp->cp_lock, flags);
1361 return;
1362 }
1363 conn->c_ping_triggered = 1;
1364 spin_unlock_irqrestore(&cp->cp_lock, flags);
1365 rds_send_probe(cp, cpu_to_be16(RDS_FLAG_PROBE_PORT), 0, 0);
1366}
1367EXPORT_SYMBOL_GPL(rds_send_ping);