Loading...
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 int same_rm = 0;
149
150restart:
151 batch_count = 0;
152
153 /*
154 * sendmsg calls here after having queued its message on the send
155 * queue. We only have one task feeding the connection at a time. If
156 * another thread is already feeding the queue then we back off. This
157 * avoids blocking the caller and trading per-connection data between
158 * caches per message.
159 */
160 if (!acquire_in_xmit(cp)) {
161 rds_stats_inc(s_send_lock_contention);
162 ret = -ENOMEM;
163 goto out;
164 }
165
166 if (rds_destroy_pending(cp->cp_conn)) {
167 release_in_xmit(cp);
168 ret = -ENETUNREACH; /* dont requeue send work */
169 goto out;
170 }
171
172 /*
173 * we record the send generation after doing the xmit acquire.
174 * if someone else manages to jump in and do some work, we'll use
175 * this to avoid a goto restart farther down.
176 *
177 * The acquire_in_xmit() check above ensures that only one
178 * caller can increment c_send_gen at any time.
179 */
180 send_gen = READ_ONCE(cp->cp_send_gen) + 1;
181 WRITE_ONCE(cp->cp_send_gen, send_gen);
182
183 /*
184 * rds_conn_shutdown() sets the conn state and then tests RDS_IN_XMIT,
185 * we do the opposite to avoid races.
186 */
187 if (!rds_conn_path_up(cp)) {
188 release_in_xmit(cp);
189 ret = 0;
190 goto out;
191 }
192
193 if (conn->c_trans->xmit_path_prepare)
194 conn->c_trans->xmit_path_prepare(cp);
195
196 /*
197 * spin trying to push headers and data down the connection until
198 * the connection doesn't make forward progress.
199 */
200 while (1) {
201
202 rm = cp->cp_xmit_rm;
203
204 if (!rm) {
205 same_rm = 0;
206 } else {
207 same_rm++;
208 if (same_rm >= 4096) {
209 rds_stats_inc(s_send_stuck_rm);
210 ret = -EAGAIN;
211 break;
212 }
213 }
214
215 /*
216 * If between sending messages, we can send a pending congestion
217 * map update.
218 */
219 if (!rm && test_and_clear_bit(0, &conn->c_map_queued)) {
220 rm = rds_cong_update_alloc(conn);
221 if (IS_ERR(rm)) {
222 ret = PTR_ERR(rm);
223 break;
224 }
225 rm->data.op_active = 1;
226 rm->m_inc.i_conn_path = cp;
227 rm->m_inc.i_conn = cp->cp_conn;
228
229 cp->cp_xmit_rm = rm;
230 }
231
232 /*
233 * If not already working on one, grab the next message.
234 *
235 * cp_xmit_rm holds a ref while we're sending this message down
236 * the connction. We can use this ref while holding the
237 * send_sem.. rds_send_reset() is serialized with it.
238 */
239 if (!rm) {
240 unsigned int len;
241
242 batch_count++;
243
244 /* we want to process as big a batch as we can, but
245 * we also want to avoid softlockups. If we've been
246 * through a lot of messages, lets back off and see
247 * if anyone else jumps in
248 */
249 if (batch_count >= send_batch_count)
250 goto over_batch;
251
252 spin_lock_irqsave(&cp->cp_lock, flags);
253
254 if (!list_empty(&cp->cp_send_queue)) {
255 rm = list_entry(cp->cp_send_queue.next,
256 struct rds_message,
257 m_conn_item);
258 rds_message_addref(rm);
259
260 /*
261 * Move the message from the send queue to the retransmit
262 * list right away.
263 */
264 list_move_tail(&rm->m_conn_item,
265 &cp->cp_retrans);
266 }
267
268 spin_unlock_irqrestore(&cp->cp_lock, flags);
269
270 if (!rm)
271 break;
272
273 /* Unfortunately, the way Infiniband deals with
274 * RDMA to a bad MR key is by moving the entire
275 * queue pair to error state. We cold possibly
276 * recover from that, but right now we drop the
277 * connection.
278 * Therefore, we never retransmit messages with RDMA ops.
279 */
280 if (test_bit(RDS_MSG_FLUSH, &rm->m_flags) ||
281 (rm->rdma.op_active &&
282 test_bit(RDS_MSG_RETRANSMITTED, &rm->m_flags))) {
283 spin_lock_irqsave(&cp->cp_lock, flags);
284 if (test_and_clear_bit(RDS_MSG_ON_CONN, &rm->m_flags))
285 list_move(&rm->m_conn_item, &to_be_dropped);
286 spin_unlock_irqrestore(&cp->cp_lock, flags);
287 continue;
288 }
289
290 /* Require an ACK every once in a while */
291 len = ntohl(rm->m_inc.i_hdr.h_len);
292 if (cp->cp_unacked_packets == 0 ||
293 cp->cp_unacked_bytes < len) {
294 set_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags);
295
296 cp->cp_unacked_packets =
297 rds_sysctl_max_unacked_packets;
298 cp->cp_unacked_bytes =
299 rds_sysctl_max_unacked_bytes;
300 rds_stats_inc(s_send_ack_required);
301 } else {
302 cp->cp_unacked_bytes -= len;
303 cp->cp_unacked_packets--;
304 }
305
306 cp->cp_xmit_rm = rm;
307 }
308
309 /* The transport either sends the whole rdma or none of it */
310 if (rm->rdma.op_active && !cp->cp_xmit_rdma_sent) {
311 rm->m_final_op = &rm->rdma;
312 /* The transport owns the mapped memory for now.
313 * You can't unmap it while it's on the send queue
314 */
315 set_bit(RDS_MSG_MAPPED, &rm->m_flags);
316 ret = conn->c_trans->xmit_rdma(conn, &rm->rdma);
317 if (ret) {
318 clear_bit(RDS_MSG_MAPPED, &rm->m_flags);
319 wake_up_interruptible(&rm->m_flush_wait);
320 break;
321 }
322 cp->cp_xmit_rdma_sent = 1;
323
324 }
325
326 if (rm->atomic.op_active && !cp->cp_xmit_atomic_sent) {
327 rm->m_final_op = &rm->atomic;
328 /* The transport owns the mapped memory for now.
329 * You can't unmap it while it's on the send queue
330 */
331 set_bit(RDS_MSG_MAPPED, &rm->m_flags);
332 ret = conn->c_trans->xmit_atomic(conn, &rm->atomic);
333 if (ret) {
334 clear_bit(RDS_MSG_MAPPED, &rm->m_flags);
335 wake_up_interruptible(&rm->m_flush_wait);
336 break;
337 }
338 cp->cp_xmit_atomic_sent = 1;
339
340 }
341
342 /*
343 * A number of cases require an RDS header to be sent
344 * even if there is no data.
345 * We permit 0-byte sends; rds-ping depends on this.
346 * However, if there are exclusively attached silent ops,
347 * we skip the hdr/data send, to enable silent operation.
348 */
349 if (rm->data.op_nents == 0) {
350 int ops_present;
351 int all_ops_are_silent = 1;
352
353 ops_present = (rm->atomic.op_active || rm->rdma.op_active);
354 if (rm->atomic.op_active && !rm->atomic.op_silent)
355 all_ops_are_silent = 0;
356 if (rm->rdma.op_active && !rm->rdma.op_silent)
357 all_ops_are_silent = 0;
358
359 if (ops_present && all_ops_are_silent
360 && !rm->m_rdma_cookie)
361 rm->data.op_active = 0;
362 }
363
364 if (rm->data.op_active && !cp->cp_xmit_data_sent) {
365 rm->m_final_op = &rm->data;
366
367 ret = conn->c_trans->xmit(conn, rm,
368 cp->cp_xmit_hdr_off,
369 cp->cp_xmit_sg,
370 cp->cp_xmit_data_off);
371 if (ret <= 0)
372 break;
373
374 if (cp->cp_xmit_hdr_off < sizeof(struct rds_header)) {
375 tmp = min_t(int, ret,
376 sizeof(struct rds_header) -
377 cp->cp_xmit_hdr_off);
378 cp->cp_xmit_hdr_off += tmp;
379 ret -= tmp;
380 }
381
382 sg = &rm->data.op_sg[cp->cp_xmit_sg];
383 while (ret) {
384 tmp = min_t(int, ret, sg->length -
385 cp->cp_xmit_data_off);
386 cp->cp_xmit_data_off += tmp;
387 ret -= tmp;
388 if (cp->cp_xmit_data_off == sg->length) {
389 cp->cp_xmit_data_off = 0;
390 sg++;
391 cp->cp_xmit_sg++;
392 BUG_ON(ret != 0 && cp->cp_xmit_sg ==
393 rm->data.op_nents);
394 }
395 }
396
397 if (cp->cp_xmit_hdr_off == sizeof(struct rds_header) &&
398 (cp->cp_xmit_sg == rm->data.op_nents))
399 cp->cp_xmit_data_sent = 1;
400 }
401
402 /*
403 * A rm will only take multiple times through this loop
404 * if there is a data op. Thus, if the data is sent (or there was
405 * none), then we're done with the rm.
406 */
407 if (!rm->data.op_active || cp->cp_xmit_data_sent) {
408 cp->cp_xmit_rm = NULL;
409 cp->cp_xmit_sg = 0;
410 cp->cp_xmit_hdr_off = 0;
411 cp->cp_xmit_data_off = 0;
412 cp->cp_xmit_rdma_sent = 0;
413 cp->cp_xmit_atomic_sent = 0;
414 cp->cp_xmit_data_sent = 0;
415
416 rds_message_put(rm);
417 }
418 }
419
420over_batch:
421 if (conn->c_trans->xmit_path_complete)
422 conn->c_trans->xmit_path_complete(cp);
423 release_in_xmit(cp);
424
425 /* Nuke any messages we decided not to retransmit. */
426 if (!list_empty(&to_be_dropped)) {
427 /* irqs on here, so we can put(), unlike above */
428 list_for_each_entry(rm, &to_be_dropped, m_conn_item)
429 rds_message_put(rm);
430 rds_send_remove_from_sock(&to_be_dropped, RDS_RDMA_DROPPED);
431 }
432
433 /*
434 * Other senders can queue a message after we last test the send queue
435 * but before we clear RDS_IN_XMIT. In that case they'd back off and
436 * not try and send their newly queued message. We need to check the
437 * send queue after having cleared RDS_IN_XMIT so that their message
438 * doesn't get stuck on the send queue.
439 *
440 * If the transport cannot continue (i.e ret != 0), then it must
441 * call us when more room is available, such as from the tx
442 * completion handler.
443 *
444 * We have an extra generation check here so that if someone manages
445 * to jump in after our release_in_xmit, we'll see that they have done
446 * some work and we will skip our goto
447 */
448 if (ret == 0) {
449 bool raced;
450
451 smp_mb();
452 raced = send_gen != READ_ONCE(cp->cp_send_gen);
453
454 if ((test_bit(0, &conn->c_map_queued) ||
455 !list_empty(&cp->cp_send_queue)) && !raced) {
456 if (batch_count < send_batch_count)
457 goto restart;
458 rcu_read_lock();
459 if (rds_destroy_pending(cp->cp_conn))
460 ret = -ENETUNREACH;
461 else
462 queue_delayed_work(rds_wq, &cp->cp_send_w, 1);
463 rcu_read_unlock();
464 } else if (raced) {
465 rds_stats_inc(s_send_lock_queue_raced);
466 }
467 }
468out:
469 return ret;
470}
471EXPORT_SYMBOL_GPL(rds_send_xmit);
472
473static void rds_send_sndbuf_remove(struct rds_sock *rs, struct rds_message *rm)
474{
475 u32 len = be32_to_cpu(rm->m_inc.i_hdr.h_len);
476
477 assert_spin_locked(&rs->rs_lock);
478
479 BUG_ON(rs->rs_snd_bytes < len);
480 rs->rs_snd_bytes -= len;
481
482 if (rs->rs_snd_bytes == 0)
483 rds_stats_inc(s_send_queue_empty);
484}
485
486static inline int rds_send_is_acked(struct rds_message *rm, u64 ack,
487 is_acked_func is_acked)
488{
489 if (is_acked)
490 return is_acked(rm, ack);
491 return be64_to_cpu(rm->m_inc.i_hdr.h_sequence) <= ack;
492}
493
494/*
495 * This is pretty similar to what happens below in the ACK
496 * handling code - except that we call here as soon as we get
497 * the IB send completion on the RDMA op and the accompanying
498 * message.
499 */
500void rds_rdma_send_complete(struct rds_message *rm, int status)
501{
502 struct rds_sock *rs = NULL;
503 struct rm_rdma_op *ro;
504 struct rds_notifier *notifier;
505 unsigned long flags;
506
507 spin_lock_irqsave(&rm->m_rs_lock, flags);
508
509 ro = &rm->rdma;
510 if (test_bit(RDS_MSG_ON_SOCK, &rm->m_flags) &&
511 ro->op_active && ro->op_notify && ro->op_notifier) {
512 notifier = ro->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 ro->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_rdma_send_complete);
532
533/*
534 * Just like above, except looks at atomic op
535 */
536void rds_atomic_send_complete(struct rds_message *rm, int status)
537{
538 struct rds_sock *rs = NULL;
539 struct rm_atomic_op *ao;
540 struct rds_notifier *notifier;
541 unsigned long flags;
542
543 spin_lock_irqsave(&rm->m_rs_lock, flags);
544
545 ao = &rm->atomic;
546 if (test_bit(RDS_MSG_ON_SOCK, &rm->m_flags)
547 && ao->op_active && ao->op_notify && ao->op_notifier) {
548 notifier = ao->op_notifier;
549 rs = rm->m_rs;
550 sock_hold(rds_rs_to_sk(rs));
551
552 notifier->n_status = status;
553 spin_lock(&rs->rs_lock);
554 list_add_tail(¬ifier->n_list, &rs->rs_notify_queue);
555 spin_unlock(&rs->rs_lock);
556
557 ao->op_notifier = NULL;
558 }
559
560 spin_unlock_irqrestore(&rm->m_rs_lock, flags);
561
562 if (rs) {
563 rds_wake_sk_sleep(rs);
564 sock_put(rds_rs_to_sk(rs));
565 }
566}
567EXPORT_SYMBOL_GPL(rds_atomic_send_complete);
568
569/*
570 * This is the same as rds_rdma_send_complete except we
571 * don't do any locking - we have all the ingredients (message,
572 * socket, socket lock) and can just move the notifier.
573 */
574static inline void
575__rds_send_complete(struct rds_sock *rs, struct rds_message *rm, int status)
576{
577 struct rm_rdma_op *ro;
578 struct rm_atomic_op *ao;
579
580 ro = &rm->rdma;
581 if (ro->op_active && ro->op_notify && ro->op_notifier) {
582 ro->op_notifier->n_status = status;
583 list_add_tail(&ro->op_notifier->n_list, &rs->rs_notify_queue);
584 ro->op_notifier = NULL;
585 }
586
587 ao = &rm->atomic;
588 if (ao->op_active && ao->op_notify && ao->op_notifier) {
589 ao->op_notifier->n_status = status;
590 list_add_tail(&ao->op_notifier->n_list, &rs->rs_notify_queue);
591 ao->op_notifier = NULL;
592 }
593
594 /* No need to wake the app - caller does this */
595}
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 }
663 spin_unlock(&rs->rs_lock);
664
665unlock_and_drop:
666 spin_unlock_irqrestore(&rm->m_rs_lock, flags);
667 rds_message_put(rm);
668 if (was_on_sock)
669 rds_message_put(rm);
670 }
671
672 if (rs) {
673 rds_wake_sk_sleep(rs);
674 sock_put(rds_rs_to_sk(rs));
675 }
676}
677
678/*
679 * Transports call here when they've determined that the receiver queued
680 * messages up to, and including, the given sequence number. Messages are
681 * moved to the retrans queue when rds_send_xmit picks them off the send
682 * queue. This means that in the TCP case, the message may not have been
683 * assigned the m_ack_seq yet - but that's fine as long as tcp_is_acked
684 * checks the RDS_MSG_HAS_ACK_SEQ bit.
685 */
686void rds_send_path_drop_acked(struct rds_conn_path *cp, u64 ack,
687 is_acked_func is_acked)
688{
689 struct rds_message *rm, *tmp;
690 unsigned long flags;
691 LIST_HEAD(list);
692
693 spin_lock_irqsave(&cp->cp_lock, flags);
694
695 list_for_each_entry_safe(rm, tmp, &cp->cp_retrans, m_conn_item) {
696 if (!rds_send_is_acked(rm, ack, is_acked))
697 break;
698
699 list_move(&rm->m_conn_item, &list);
700 clear_bit(RDS_MSG_ON_CONN, &rm->m_flags);
701 }
702
703 /* order flag updates with spin locks */
704 if (!list_empty(&list))
705 smp_mb__after_atomic();
706
707 spin_unlock_irqrestore(&cp->cp_lock, flags);
708
709 /* now remove the messages from the sock list as needed */
710 rds_send_remove_from_sock(&list, RDS_RDMA_SUCCESS);
711}
712EXPORT_SYMBOL_GPL(rds_send_path_drop_acked);
713
714void rds_send_drop_acked(struct rds_connection *conn, u64 ack,
715 is_acked_func is_acked)
716{
717 WARN_ON(conn->c_trans->t_mp_capable);
718 rds_send_path_drop_acked(&conn->c_path[0], ack, is_acked);
719}
720EXPORT_SYMBOL_GPL(rds_send_drop_acked);
721
722void rds_send_drop_to(struct rds_sock *rs, struct sockaddr_in6 *dest)
723{
724 struct rds_message *rm, *tmp;
725 struct rds_connection *conn;
726 struct rds_conn_path *cp;
727 unsigned long flags;
728 LIST_HEAD(list);
729
730 /* get all the messages we're dropping under the rs lock */
731 spin_lock_irqsave(&rs->rs_lock, flags);
732
733 list_for_each_entry_safe(rm, tmp, &rs->rs_send_queue, m_sock_item) {
734 if (dest &&
735 (!ipv6_addr_equal(&dest->sin6_addr, &rm->m_daddr) ||
736 dest->sin6_port != rm->m_inc.i_hdr.h_dport))
737 continue;
738
739 list_move(&rm->m_sock_item, &list);
740 rds_send_sndbuf_remove(rs, rm);
741 clear_bit(RDS_MSG_ON_SOCK, &rm->m_flags);
742 }
743
744 /* order flag updates with the rs lock */
745 smp_mb__after_atomic();
746
747 spin_unlock_irqrestore(&rs->rs_lock, flags);
748
749 if (list_empty(&list))
750 return;
751
752 /* Remove the messages from the conn */
753 list_for_each_entry(rm, &list, m_sock_item) {
754
755 conn = rm->m_inc.i_conn;
756 if (conn->c_trans->t_mp_capable)
757 cp = rm->m_inc.i_conn_path;
758 else
759 cp = &conn->c_path[0];
760
761 spin_lock_irqsave(&cp->cp_lock, flags);
762 /*
763 * Maybe someone else beat us to removing rm from the conn.
764 * If we race with their flag update we'll get the lock and
765 * then really see that the flag has been cleared.
766 */
767 if (!test_and_clear_bit(RDS_MSG_ON_CONN, &rm->m_flags)) {
768 spin_unlock_irqrestore(&cp->cp_lock, flags);
769 continue;
770 }
771 list_del_init(&rm->m_conn_item);
772 spin_unlock_irqrestore(&cp->cp_lock, flags);
773
774 /*
775 * Couldn't grab m_rs_lock in top loop (lock ordering),
776 * but we can now.
777 */
778 spin_lock_irqsave(&rm->m_rs_lock, flags);
779
780 spin_lock(&rs->rs_lock);
781 __rds_send_complete(rs, rm, RDS_RDMA_CANCELED);
782 spin_unlock(&rs->rs_lock);
783
784 spin_unlock_irqrestore(&rm->m_rs_lock, flags);
785
786 rds_message_put(rm);
787 }
788
789 rds_wake_sk_sleep(rs);
790
791 while (!list_empty(&list)) {
792 rm = list_entry(list.next, struct rds_message, m_sock_item);
793 list_del_init(&rm->m_sock_item);
794 rds_message_wait(rm);
795
796 /* just in case the code above skipped this message
797 * because RDS_MSG_ON_CONN wasn't set, run it again here
798 * taking m_rs_lock is the only thing that keeps us
799 * from racing with ack processing.
800 */
801 spin_lock_irqsave(&rm->m_rs_lock, flags);
802
803 spin_lock(&rs->rs_lock);
804 __rds_send_complete(rs, rm, RDS_RDMA_CANCELED);
805 spin_unlock(&rs->rs_lock);
806
807 spin_unlock_irqrestore(&rm->m_rs_lock, flags);
808
809 rds_message_put(rm);
810 }
811}
812
813/*
814 * we only want this to fire once so we use the callers 'queued'. It's
815 * possible that another thread can race with us and remove the
816 * message from the flow with RDS_CANCEL_SENT_TO.
817 */
818static int rds_send_queue_rm(struct rds_sock *rs, struct rds_connection *conn,
819 struct rds_conn_path *cp,
820 struct rds_message *rm, __be16 sport,
821 __be16 dport, int *queued)
822{
823 unsigned long flags;
824 u32 len;
825
826 if (*queued)
827 goto out;
828
829 len = be32_to_cpu(rm->m_inc.i_hdr.h_len);
830
831 /* this is the only place which holds both the socket's rs_lock
832 * and the connection's c_lock */
833 spin_lock_irqsave(&rs->rs_lock, flags);
834
835 /*
836 * If there is a little space in sndbuf, we don't queue anything,
837 * and userspace gets -EAGAIN. But poll() indicates there's send
838 * room. This can lead to bad behavior (spinning) if snd_bytes isn't
839 * freed up by incoming acks. So we check the *old* value of
840 * rs_snd_bytes here to allow the last msg to exceed the buffer,
841 * and poll() now knows no more data can be sent.
842 */
843 if (rs->rs_snd_bytes < rds_sk_sndbuf(rs)) {
844 rs->rs_snd_bytes += len;
845
846 /* let recv side know we are close to send space exhaustion.
847 * This is probably not the optimal way to do it, as this
848 * means we set the flag on *all* messages as soon as our
849 * throughput hits a certain threshold.
850 */
851 if (rs->rs_snd_bytes >= rds_sk_sndbuf(rs) / 2)
852 set_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags);
853
854 list_add_tail(&rm->m_sock_item, &rs->rs_send_queue);
855 set_bit(RDS_MSG_ON_SOCK, &rm->m_flags);
856 rds_message_addref(rm);
857 sock_hold(rds_rs_to_sk(rs));
858 rm->m_rs = rs;
859
860 /* The code ordering is a little weird, but we're
861 trying to minimize the time we hold c_lock */
862 rds_message_populate_header(&rm->m_inc.i_hdr, sport, dport, 0);
863 rm->m_inc.i_conn = conn;
864 rm->m_inc.i_conn_path = cp;
865 rds_message_addref(rm);
866
867 spin_lock(&cp->cp_lock);
868 rm->m_inc.i_hdr.h_sequence = cpu_to_be64(cp->cp_next_tx_seq++);
869 list_add_tail(&rm->m_conn_item, &cp->cp_send_queue);
870 set_bit(RDS_MSG_ON_CONN, &rm->m_flags);
871 spin_unlock(&cp->cp_lock);
872
873 rdsdebug("queued msg %p len %d, rs %p bytes %d seq %llu\n",
874 rm, len, rs, rs->rs_snd_bytes,
875 (unsigned long long)be64_to_cpu(rm->m_inc.i_hdr.h_sequence));
876
877 *queued = 1;
878 }
879
880 spin_unlock_irqrestore(&rs->rs_lock, flags);
881out:
882 return *queued;
883}
884
885/*
886 * rds_message is getting to be quite complicated, and we'd like to allocate
887 * it all in one go. This figures out how big it needs to be up front.
888 */
889static int rds_rm_size(struct msghdr *msg, int num_sgs,
890 struct rds_iov_vector_arr *vct)
891{
892 struct cmsghdr *cmsg;
893 int size = 0;
894 int cmsg_groups = 0;
895 int retval;
896 bool zcopy_cookie = false;
897 struct rds_iov_vector *iov, *tmp_iov;
898
899 if (num_sgs < 0)
900 return -EINVAL;
901
902 for_each_cmsghdr(cmsg, msg) {
903 if (!CMSG_OK(msg, cmsg))
904 return -EINVAL;
905
906 if (cmsg->cmsg_level != SOL_RDS)
907 continue;
908
909 switch (cmsg->cmsg_type) {
910 case RDS_CMSG_RDMA_ARGS:
911 if (vct->indx >= vct->len) {
912 vct->len += vct->incr;
913 tmp_iov =
914 krealloc(vct->vec,
915 vct->len *
916 sizeof(struct rds_iov_vector),
917 GFP_KERNEL);
918 if (!tmp_iov) {
919 vct->len -= vct->incr;
920 return -ENOMEM;
921 }
922 vct->vec = tmp_iov;
923 }
924 iov = &vct->vec[vct->indx];
925 memset(iov, 0, sizeof(struct rds_iov_vector));
926 vct->indx++;
927 cmsg_groups |= 1;
928 retval = rds_rdma_extra_size(CMSG_DATA(cmsg), iov);
929 if (retval < 0)
930 return retval;
931 size += retval;
932
933 break;
934
935 case RDS_CMSG_ZCOPY_COOKIE:
936 zcopy_cookie = true;
937 fallthrough;
938
939 case RDS_CMSG_RDMA_DEST:
940 case RDS_CMSG_RDMA_MAP:
941 cmsg_groups |= 2;
942 /* these are valid but do no add any size */
943 break;
944
945 case RDS_CMSG_ATOMIC_CSWP:
946 case RDS_CMSG_ATOMIC_FADD:
947 case RDS_CMSG_MASKED_ATOMIC_CSWP:
948 case RDS_CMSG_MASKED_ATOMIC_FADD:
949 cmsg_groups |= 1;
950 size += sizeof(struct scatterlist);
951 break;
952
953 default:
954 return -EINVAL;
955 }
956
957 }
958
959 if ((msg->msg_flags & MSG_ZEROCOPY) && !zcopy_cookie)
960 return -EINVAL;
961
962 size += num_sgs * sizeof(struct scatterlist);
963
964 /* Ensure (DEST, MAP) are never used with (ARGS, ATOMIC) */
965 if (cmsg_groups == 3)
966 return -EINVAL;
967
968 return size;
969}
970
971static int rds_cmsg_zcopy(struct rds_sock *rs, struct rds_message *rm,
972 struct cmsghdr *cmsg)
973{
974 u32 *cookie;
975
976 if (cmsg->cmsg_len < CMSG_LEN(sizeof(*cookie)) ||
977 !rm->data.op_mmp_znotifier)
978 return -EINVAL;
979 cookie = CMSG_DATA(cmsg);
980 rm->data.op_mmp_znotifier->z_cookie = *cookie;
981 return 0;
982}
983
984static int rds_cmsg_send(struct rds_sock *rs, struct rds_message *rm,
985 struct msghdr *msg, int *allocated_mr,
986 struct rds_iov_vector_arr *vct)
987{
988 struct cmsghdr *cmsg;
989 int ret = 0, ind = 0;
990
991 for_each_cmsghdr(cmsg, msg) {
992 if (!CMSG_OK(msg, cmsg))
993 return -EINVAL;
994
995 if (cmsg->cmsg_level != SOL_RDS)
996 continue;
997
998 /* As a side effect, RDMA_DEST and RDMA_MAP will set
999 * rm->rdma.m_rdma_cookie and rm->rdma.m_rdma_mr.
1000 */
1001 switch (cmsg->cmsg_type) {
1002 case RDS_CMSG_RDMA_ARGS:
1003 if (ind >= vct->indx)
1004 return -ENOMEM;
1005 ret = rds_cmsg_rdma_args(rs, rm, cmsg, &vct->vec[ind]);
1006 ind++;
1007 break;
1008
1009 case RDS_CMSG_RDMA_DEST:
1010 ret = rds_cmsg_rdma_dest(rs, rm, cmsg);
1011 break;
1012
1013 case RDS_CMSG_RDMA_MAP:
1014 ret = rds_cmsg_rdma_map(rs, rm, cmsg);
1015 if (!ret)
1016 *allocated_mr = 1;
1017 else if (ret == -ENODEV)
1018 /* Accommodate the get_mr() case which can fail
1019 * if connection isn't established yet.
1020 */
1021 ret = -EAGAIN;
1022 break;
1023 case RDS_CMSG_ATOMIC_CSWP:
1024 case RDS_CMSG_ATOMIC_FADD:
1025 case RDS_CMSG_MASKED_ATOMIC_CSWP:
1026 case RDS_CMSG_MASKED_ATOMIC_FADD:
1027 ret = rds_cmsg_atomic(rs, rm, cmsg);
1028 break;
1029
1030 case RDS_CMSG_ZCOPY_COOKIE:
1031 ret = rds_cmsg_zcopy(rs, rm, cmsg);
1032 break;
1033
1034 default:
1035 return -EINVAL;
1036 }
1037
1038 if (ret)
1039 break;
1040 }
1041
1042 return ret;
1043}
1044
1045static int rds_send_mprds_hash(struct rds_sock *rs,
1046 struct rds_connection *conn, int nonblock)
1047{
1048 int hash;
1049
1050 if (conn->c_npaths == 0)
1051 hash = RDS_MPATH_HASH(rs, RDS_MPATH_WORKERS);
1052 else
1053 hash = RDS_MPATH_HASH(rs, conn->c_npaths);
1054 if (conn->c_npaths == 0 && hash != 0) {
1055 rds_send_ping(conn, 0);
1056
1057 /* The underlying connection is not up yet. Need to wait
1058 * until it is up to be sure that the non-zero c_path can be
1059 * used. But if we are interrupted, we have to use the zero
1060 * c_path in case the connection ends up being non-MP capable.
1061 */
1062 if (conn->c_npaths == 0) {
1063 /* Cannot wait for the connection be made, so just use
1064 * the base c_path.
1065 */
1066 if (nonblock)
1067 return 0;
1068 if (wait_event_interruptible(conn->c_hs_waitq,
1069 conn->c_npaths != 0))
1070 hash = 0;
1071 }
1072 if (conn->c_npaths == 1)
1073 hash = 0;
1074 }
1075 return hash;
1076}
1077
1078static int rds_rdma_bytes(struct msghdr *msg, size_t *rdma_bytes)
1079{
1080 struct rds_rdma_args *args;
1081 struct cmsghdr *cmsg;
1082
1083 for_each_cmsghdr(cmsg, msg) {
1084 if (!CMSG_OK(msg, cmsg))
1085 return -EINVAL;
1086
1087 if (cmsg->cmsg_level != SOL_RDS)
1088 continue;
1089
1090 if (cmsg->cmsg_type == RDS_CMSG_RDMA_ARGS) {
1091 if (cmsg->cmsg_len <
1092 CMSG_LEN(sizeof(struct rds_rdma_args)))
1093 return -EINVAL;
1094 args = CMSG_DATA(cmsg);
1095 *rdma_bytes += args->remote_vec.bytes;
1096 }
1097 }
1098 return 0;
1099}
1100
1101int rds_sendmsg(struct socket *sock, struct msghdr *msg, size_t payload_len)
1102{
1103 struct sock *sk = sock->sk;
1104 struct rds_sock *rs = rds_sk_to_rs(sk);
1105 DECLARE_SOCKADDR(struct sockaddr_in6 *, sin6, msg->msg_name);
1106 DECLARE_SOCKADDR(struct sockaddr_in *, usin, msg->msg_name);
1107 __be16 dport;
1108 struct rds_message *rm = NULL;
1109 struct rds_connection *conn;
1110 int ret = 0;
1111 int queued = 0, allocated_mr = 0;
1112 int nonblock = msg->msg_flags & MSG_DONTWAIT;
1113 long timeo = sock_sndtimeo(sk, nonblock);
1114 struct rds_conn_path *cpath;
1115 struct in6_addr daddr;
1116 __u32 scope_id = 0;
1117 size_t total_payload_len = payload_len, rdma_payload_len = 0;
1118 bool zcopy = ((msg->msg_flags & MSG_ZEROCOPY) &&
1119 sock_flag(rds_rs_to_sk(rs), SOCK_ZEROCOPY));
1120 int num_sgs = DIV_ROUND_UP(payload_len, PAGE_SIZE);
1121 int namelen;
1122 struct rds_iov_vector_arr vct;
1123 int ind;
1124
1125 memset(&vct, 0, sizeof(vct));
1126
1127 /* expect 1 RDMA CMSG per rds_sendmsg. can still grow if more needed. */
1128 vct.incr = 1;
1129
1130 /* Mirror Linux UDP mirror of BSD error message compatibility */
1131 /* XXX: Perhaps MSG_MORE someday */
1132 if (msg->msg_flags & ~(MSG_DONTWAIT | MSG_CMSG_COMPAT | MSG_ZEROCOPY)) {
1133 ret = -EOPNOTSUPP;
1134 goto out;
1135 }
1136
1137 namelen = msg->msg_namelen;
1138 if (namelen != 0) {
1139 if (namelen < sizeof(*usin)) {
1140 ret = -EINVAL;
1141 goto out;
1142 }
1143 switch (usin->sin_family) {
1144 case AF_INET:
1145 if (usin->sin_addr.s_addr == htonl(INADDR_ANY) ||
1146 usin->sin_addr.s_addr == htonl(INADDR_BROADCAST) ||
1147 ipv4_is_multicast(usin->sin_addr.s_addr)) {
1148 ret = -EINVAL;
1149 goto out;
1150 }
1151 ipv6_addr_set_v4mapped(usin->sin_addr.s_addr, &daddr);
1152 dport = usin->sin_port;
1153 break;
1154
1155#if IS_ENABLED(CONFIG_IPV6)
1156 case AF_INET6: {
1157 int addr_type;
1158
1159 if (namelen < sizeof(*sin6)) {
1160 ret = -EINVAL;
1161 goto out;
1162 }
1163 addr_type = ipv6_addr_type(&sin6->sin6_addr);
1164 if (!(addr_type & IPV6_ADDR_UNICAST)) {
1165 __be32 addr4;
1166
1167 if (!(addr_type & IPV6_ADDR_MAPPED)) {
1168 ret = -EINVAL;
1169 goto out;
1170 }
1171
1172 /* It is a mapped address. Need to do some
1173 * sanity checks.
1174 */
1175 addr4 = sin6->sin6_addr.s6_addr32[3];
1176 if (addr4 == htonl(INADDR_ANY) ||
1177 addr4 == htonl(INADDR_BROADCAST) ||
1178 ipv4_is_multicast(addr4)) {
1179 ret = -EINVAL;
1180 goto out;
1181 }
1182 }
1183 if (addr_type & IPV6_ADDR_LINKLOCAL) {
1184 if (sin6->sin6_scope_id == 0) {
1185 ret = -EINVAL;
1186 goto out;
1187 }
1188 scope_id = sin6->sin6_scope_id;
1189 }
1190
1191 daddr = sin6->sin6_addr;
1192 dport = sin6->sin6_port;
1193 break;
1194 }
1195#endif
1196
1197 default:
1198 ret = -EINVAL;
1199 goto out;
1200 }
1201 } else {
1202 /* We only care about consistency with ->connect() */
1203 lock_sock(sk);
1204 daddr = rs->rs_conn_addr;
1205 dport = rs->rs_conn_port;
1206 scope_id = rs->rs_bound_scope_id;
1207 release_sock(sk);
1208 }
1209
1210 lock_sock(sk);
1211 if (ipv6_addr_any(&rs->rs_bound_addr) || ipv6_addr_any(&daddr)) {
1212 release_sock(sk);
1213 ret = -ENOTCONN;
1214 goto out;
1215 } else if (namelen != 0) {
1216 /* Cannot send to an IPv4 address using an IPv6 source
1217 * address and cannot send to an IPv6 address using an
1218 * IPv4 source address.
1219 */
1220 if (ipv6_addr_v4mapped(&daddr) ^
1221 ipv6_addr_v4mapped(&rs->rs_bound_addr)) {
1222 release_sock(sk);
1223 ret = -EOPNOTSUPP;
1224 goto out;
1225 }
1226 /* If the socket is already bound to a link local address,
1227 * it can only send to peers on the same link. But allow
1228 * communicating beween link local and non-link local address.
1229 */
1230 if (scope_id != rs->rs_bound_scope_id) {
1231 if (!scope_id) {
1232 scope_id = rs->rs_bound_scope_id;
1233 } else if (rs->rs_bound_scope_id) {
1234 release_sock(sk);
1235 ret = -EINVAL;
1236 goto out;
1237 }
1238 }
1239 }
1240 release_sock(sk);
1241
1242 ret = rds_rdma_bytes(msg, &rdma_payload_len);
1243 if (ret)
1244 goto out;
1245
1246 total_payload_len += rdma_payload_len;
1247 if (max_t(size_t, payload_len, rdma_payload_len) > RDS_MAX_MSG_SIZE) {
1248 ret = -EMSGSIZE;
1249 goto out;
1250 }
1251
1252 if (payload_len > rds_sk_sndbuf(rs)) {
1253 ret = -EMSGSIZE;
1254 goto out;
1255 }
1256
1257 if (zcopy) {
1258 if (rs->rs_transport->t_type != RDS_TRANS_TCP) {
1259 ret = -EOPNOTSUPP;
1260 goto out;
1261 }
1262 num_sgs = iov_iter_npages(&msg->msg_iter, INT_MAX);
1263 }
1264 /* size of rm including all sgs */
1265 ret = rds_rm_size(msg, num_sgs, &vct);
1266 if (ret < 0)
1267 goto out;
1268
1269 rm = rds_message_alloc(ret, GFP_KERNEL);
1270 if (!rm) {
1271 ret = -ENOMEM;
1272 goto out;
1273 }
1274
1275 /* Attach data to the rm */
1276 if (payload_len) {
1277 rm->data.op_sg = rds_message_alloc_sgs(rm, num_sgs);
1278 if (IS_ERR(rm->data.op_sg)) {
1279 ret = PTR_ERR(rm->data.op_sg);
1280 goto out;
1281 }
1282 ret = rds_message_copy_from_user(rm, &msg->msg_iter, zcopy);
1283 if (ret)
1284 goto out;
1285 }
1286 rm->data.op_active = 1;
1287
1288 rm->m_daddr = daddr;
1289
1290 /* rds_conn_create has a spinlock that runs with IRQ off.
1291 * Caching the conn in the socket helps a lot. */
1292 if (rs->rs_conn && ipv6_addr_equal(&rs->rs_conn->c_faddr, &daddr) &&
1293 rs->rs_tos == rs->rs_conn->c_tos) {
1294 conn = rs->rs_conn;
1295 } else {
1296 conn = rds_conn_create_outgoing(sock_net(sock->sk),
1297 &rs->rs_bound_addr, &daddr,
1298 rs->rs_transport, rs->rs_tos,
1299 sock->sk->sk_allocation,
1300 scope_id);
1301 if (IS_ERR(conn)) {
1302 ret = PTR_ERR(conn);
1303 goto out;
1304 }
1305 rs->rs_conn = conn;
1306 }
1307
1308 if (conn->c_trans->t_mp_capable)
1309 cpath = &conn->c_path[rds_send_mprds_hash(rs, conn, nonblock)];
1310 else
1311 cpath = &conn->c_path[0];
1312
1313 rm->m_conn_path = cpath;
1314
1315 /* Parse any control messages the user may have included. */
1316 ret = rds_cmsg_send(rs, rm, msg, &allocated_mr, &vct);
1317 if (ret) {
1318 /* Trigger connection so that its ready for the next retry */
1319 if (ret == -EAGAIN)
1320 rds_conn_connect_if_down(conn);
1321 goto out;
1322 }
1323
1324 if (rm->rdma.op_active && !conn->c_trans->xmit_rdma) {
1325 printk_ratelimited(KERN_NOTICE "rdma_op %p conn xmit_rdma %p\n",
1326 &rm->rdma, conn->c_trans->xmit_rdma);
1327 ret = -EOPNOTSUPP;
1328 goto out;
1329 }
1330
1331 if (rm->atomic.op_active && !conn->c_trans->xmit_atomic) {
1332 printk_ratelimited(KERN_NOTICE "atomic_op %p conn xmit_atomic %p\n",
1333 &rm->atomic, conn->c_trans->xmit_atomic);
1334 ret = -EOPNOTSUPP;
1335 goto out;
1336 }
1337
1338 if (rds_destroy_pending(conn)) {
1339 ret = -EAGAIN;
1340 goto out;
1341 }
1342
1343 if (rds_conn_path_down(cpath))
1344 rds_check_all_paths(conn);
1345
1346 ret = rds_cong_wait(conn->c_fcong, dport, nonblock, rs);
1347 if (ret) {
1348 rs->rs_seen_congestion = 1;
1349 goto out;
1350 }
1351 while (!rds_send_queue_rm(rs, conn, cpath, rm, rs->rs_bound_port,
1352 dport, &queued)) {
1353 rds_stats_inc(s_send_queue_full);
1354
1355 if (nonblock) {
1356 ret = -EAGAIN;
1357 goto out;
1358 }
1359
1360 timeo = wait_event_interruptible_timeout(*sk_sleep(sk),
1361 rds_send_queue_rm(rs, conn, cpath, rm,
1362 rs->rs_bound_port,
1363 dport,
1364 &queued),
1365 timeo);
1366 rdsdebug("sendmsg woke queued %d timeo %ld\n", queued, timeo);
1367 if (timeo > 0 || timeo == MAX_SCHEDULE_TIMEOUT)
1368 continue;
1369
1370 ret = timeo;
1371 if (ret == 0)
1372 ret = -ETIMEDOUT;
1373 goto out;
1374 }
1375
1376 /*
1377 * By now we've committed to the send. We reuse rds_send_worker()
1378 * to retry sends in the rds thread if the transport asks us to.
1379 */
1380 rds_stats_inc(s_send_queued);
1381
1382 ret = rds_send_xmit(cpath);
1383 if (ret == -ENOMEM || ret == -EAGAIN) {
1384 ret = 0;
1385 rcu_read_lock();
1386 if (rds_destroy_pending(cpath->cp_conn))
1387 ret = -ENETUNREACH;
1388 else
1389 queue_delayed_work(rds_wq, &cpath->cp_send_w, 1);
1390 rcu_read_unlock();
1391 }
1392 if (ret)
1393 goto out;
1394 rds_message_put(rm);
1395
1396 for (ind = 0; ind < vct.indx; ind++)
1397 kfree(vct.vec[ind].iov);
1398 kfree(vct.vec);
1399
1400 return payload_len;
1401
1402out:
1403 for (ind = 0; ind < vct.indx; ind++)
1404 kfree(vct.vec[ind].iov);
1405 kfree(vct.vec);
1406
1407 /* If the user included a RDMA_MAP cmsg, we allocated a MR on the fly.
1408 * If the sendmsg goes through, we keep the MR. If it fails with EAGAIN
1409 * or in any other way, we need to destroy the MR again */
1410 if (allocated_mr)
1411 rds_rdma_unuse(rs, rds_rdma_cookie_key(rm->m_rdma_cookie), 1);
1412
1413 if (rm)
1414 rds_message_put(rm);
1415 return ret;
1416}
1417
1418/*
1419 * send out a probe. Can be shared by rds_send_ping,
1420 * rds_send_pong, rds_send_hb.
1421 * rds_send_hb should use h_flags
1422 * RDS_FLAG_HB_PING|RDS_FLAG_ACK_REQUIRED
1423 * or
1424 * RDS_FLAG_HB_PONG|RDS_FLAG_ACK_REQUIRED
1425 */
1426static int
1427rds_send_probe(struct rds_conn_path *cp, __be16 sport,
1428 __be16 dport, u8 h_flags)
1429{
1430 struct rds_message *rm;
1431 unsigned long flags;
1432 int ret = 0;
1433
1434 rm = rds_message_alloc(0, GFP_ATOMIC);
1435 if (!rm) {
1436 ret = -ENOMEM;
1437 goto out;
1438 }
1439
1440 rm->m_daddr = cp->cp_conn->c_faddr;
1441 rm->data.op_active = 1;
1442
1443 rds_conn_path_connect_if_down(cp);
1444
1445 ret = rds_cong_wait(cp->cp_conn->c_fcong, dport, 1, NULL);
1446 if (ret)
1447 goto out;
1448
1449 spin_lock_irqsave(&cp->cp_lock, flags);
1450 list_add_tail(&rm->m_conn_item, &cp->cp_send_queue);
1451 set_bit(RDS_MSG_ON_CONN, &rm->m_flags);
1452 rds_message_addref(rm);
1453 rm->m_inc.i_conn = cp->cp_conn;
1454 rm->m_inc.i_conn_path = cp;
1455
1456 rds_message_populate_header(&rm->m_inc.i_hdr, sport, dport,
1457 cp->cp_next_tx_seq);
1458 rm->m_inc.i_hdr.h_flags |= h_flags;
1459 cp->cp_next_tx_seq++;
1460
1461 if (RDS_HS_PROBE(be16_to_cpu(sport), be16_to_cpu(dport)) &&
1462 cp->cp_conn->c_trans->t_mp_capable) {
1463 u16 npaths = cpu_to_be16(RDS_MPATH_WORKERS);
1464 u32 my_gen_num = cpu_to_be32(cp->cp_conn->c_my_gen_num);
1465
1466 rds_message_add_extension(&rm->m_inc.i_hdr,
1467 RDS_EXTHDR_NPATHS, &npaths,
1468 sizeof(npaths));
1469 rds_message_add_extension(&rm->m_inc.i_hdr,
1470 RDS_EXTHDR_GEN_NUM,
1471 &my_gen_num,
1472 sizeof(u32));
1473 }
1474 spin_unlock_irqrestore(&cp->cp_lock, flags);
1475
1476 rds_stats_inc(s_send_queued);
1477 rds_stats_inc(s_send_pong);
1478
1479 /* schedule the send work on rds_wq */
1480 rcu_read_lock();
1481 if (!rds_destroy_pending(cp->cp_conn))
1482 queue_delayed_work(rds_wq, &cp->cp_send_w, 1);
1483 rcu_read_unlock();
1484
1485 rds_message_put(rm);
1486 return 0;
1487
1488out:
1489 if (rm)
1490 rds_message_put(rm);
1491 return ret;
1492}
1493
1494int
1495rds_send_pong(struct rds_conn_path *cp, __be16 dport)
1496{
1497 return rds_send_probe(cp, 0, dport, 0);
1498}
1499
1500void
1501rds_send_ping(struct rds_connection *conn, int cp_index)
1502{
1503 unsigned long flags;
1504 struct rds_conn_path *cp = &conn->c_path[cp_index];
1505
1506 spin_lock_irqsave(&cp->cp_lock, flags);
1507 if (conn->c_ping_triggered) {
1508 spin_unlock_irqrestore(&cp->cp_lock, flags);
1509 return;
1510 }
1511 conn->c_ping_triggered = 1;
1512 spin_unlock_irqrestore(&cp->cp_lock, flags);
1513 rds_send_probe(cp, cpu_to_be16(RDS_FLAG_PROBE_PORT), 0, 0);
1514}
1515EXPORT_SYMBOL_GPL(rds_send_ping);
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
42#include "rds.h"
43
44/* When transmitting messages in rds_send_xmit, we need to emerge from
45 * time to time and briefly release the CPU. Otherwise the softlock watchdog
46 * will kick our shin.
47 * Also, it seems fairer to not let one busy connection stall all the
48 * others.
49 *
50 * send_batch_count is the number of times we'll loop in send_xmit. Setting
51 * it to 0 will restore the old behavior (where we looped until we had
52 * drained the queue).
53 */
54static int send_batch_count = 64;
55module_param(send_batch_count, int, 0444);
56MODULE_PARM_DESC(send_batch_count, " batch factor when working the send queue");
57
58static void rds_send_remove_from_sock(struct list_head *messages, int status);
59
60/*
61 * Reset the send state. Callers must ensure that this doesn't race with
62 * rds_send_xmit().
63 */
64void rds_send_reset(struct rds_connection *conn)
65{
66 struct rds_message *rm, *tmp;
67 unsigned long flags;
68
69 if (conn->c_xmit_rm) {
70 rm = conn->c_xmit_rm;
71 conn->c_xmit_rm = NULL;
72 /* Tell the user the RDMA op is no longer mapped by the
73 * transport. This isn't entirely true (it's flushed out
74 * independently) but as the connection is down, there's
75 * no ongoing RDMA to/from that memory */
76 rds_message_unmapped(rm);
77 rds_message_put(rm);
78 }
79
80 conn->c_xmit_sg = 0;
81 conn->c_xmit_hdr_off = 0;
82 conn->c_xmit_data_off = 0;
83 conn->c_xmit_atomic_sent = 0;
84 conn->c_xmit_rdma_sent = 0;
85 conn->c_xmit_data_sent = 0;
86
87 conn->c_map_queued = 0;
88
89 conn->c_unacked_packets = rds_sysctl_max_unacked_packets;
90 conn->c_unacked_bytes = rds_sysctl_max_unacked_bytes;
91
92 /* Mark messages as retransmissions, and move them to the send q */
93 spin_lock_irqsave(&conn->c_lock, flags);
94 list_for_each_entry_safe(rm, tmp, &conn->c_retrans, m_conn_item) {
95 set_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags);
96 set_bit(RDS_MSG_RETRANSMITTED, &rm->m_flags);
97 }
98 list_splice_init(&conn->c_retrans, &conn->c_send_queue);
99 spin_unlock_irqrestore(&conn->c_lock, flags);
100}
101
102static int acquire_in_xmit(struct rds_connection *conn)
103{
104 return test_and_set_bit(RDS_IN_XMIT, &conn->c_flags) == 0;
105}
106
107static void release_in_xmit(struct rds_connection *conn)
108{
109 clear_bit(RDS_IN_XMIT, &conn->c_flags);
110 smp_mb__after_clear_bit();
111 /*
112 * We don't use wait_on_bit()/wake_up_bit() because our waking is in a
113 * hot path and finding waiters is very rare. We don't want to walk
114 * the system-wide hashed waitqueue buckets in the fast path only to
115 * almost never find waiters.
116 */
117 if (waitqueue_active(&conn->c_waitq))
118 wake_up_all(&conn->c_waitq);
119}
120
121/*
122 * We're making the conscious trade-off here to only send one message
123 * down the connection at a time.
124 * Pro:
125 * - tx queueing is a simple fifo list
126 * - reassembly is optional and easily done by transports per conn
127 * - no per flow rx lookup at all, straight to the socket
128 * - less per-frag memory and wire overhead
129 * Con:
130 * - queued acks can be delayed behind large messages
131 * Depends:
132 * - small message latency is higher behind queued large messages
133 * - large message latency isn't starved by intervening small sends
134 */
135int rds_send_xmit(struct rds_connection *conn)
136{
137 struct rds_message *rm;
138 unsigned long flags;
139 unsigned int tmp;
140 struct scatterlist *sg;
141 int ret = 0;
142 LIST_HEAD(to_be_dropped);
143
144restart:
145
146 /*
147 * sendmsg calls here after having queued its message on the send
148 * queue. We only have one task feeding the connection at a time. If
149 * another thread is already feeding the queue then we back off. This
150 * avoids blocking the caller and trading per-connection data between
151 * caches per message.
152 */
153 if (!acquire_in_xmit(conn)) {
154 rds_stats_inc(s_send_lock_contention);
155 ret = -ENOMEM;
156 goto out;
157 }
158
159 /*
160 * rds_conn_shutdown() sets the conn state and then tests RDS_IN_XMIT,
161 * we do the opposite to avoid races.
162 */
163 if (!rds_conn_up(conn)) {
164 release_in_xmit(conn);
165 ret = 0;
166 goto out;
167 }
168
169 if (conn->c_trans->xmit_prepare)
170 conn->c_trans->xmit_prepare(conn);
171
172 /*
173 * spin trying to push headers and data down the connection until
174 * the connection doesn't make forward progress.
175 */
176 while (1) {
177
178 rm = conn->c_xmit_rm;
179
180 /*
181 * If between sending messages, we can send a pending congestion
182 * map update.
183 */
184 if (!rm && test_and_clear_bit(0, &conn->c_map_queued)) {
185 rm = rds_cong_update_alloc(conn);
186 if (IS_ERR(rm)) {
187 ret = PTR_ERR(rm);
188 break;
189 }
190 rm->data.op_active = 1;
191
192 conn->c_xmit_rm = rm;
193 }
194
195 /*
196 * If not already working on one, grab the next message.
197 *
198 * c_xmit_rm holds a ref while we're sending this message down
199 * the connction. We can use this ref while holding the
200 * send_sem.. rds_send_reset() is serialized with it.
201 */
202 if (!rm) {
203 unsigned int len;
204
205 spin_lock_irqsave(&conn->c_lock, flags);
206
207 if (!list_empty(&conn->c_send_queue)) {
208 rm = list_entry(conn->c_send_queue.next,
209 struct rds_message,
210 m_conn_item);
211 rds_message_addref(rm);
212
213 /*
214 * Move the message from the send queue to the retransmit
215 * list right away.
216 */
217 list_move_tail(&rm->m_conn_item, &conn->c_retrans);
218 }
219
220 spin_unlock_irqrestore(&conn->c_lock, flags);
221
222 if (!rm)
223 break;
224
225 /* Unfortunately, the way Infiniband deals with
226 * RDMA to a bad MR key is by moving the entire
227 * queue pair to error state. We cold possibly
228 * recover from that, but right now we drop the
229 * connection.
230 * Therefore, we never retransmit messages with RDMA ops.
231 */
232 if (rm->rdma.op_active &&
233 test_bit(RDS_MSG_RETRANSMITTED, &rm->m_flags)) {
234 spin_lock_irqsave(&conn->c_lock, flags);
235 if (test_and_clear_bit(RDS_MSG_ON_CONN, &rm->m_flags))
236 list_move(&rm->m_conn_item, &to_be_dropped);
237 spin_unlock_irqrestore(&conn->c_lock, flags);
238 continue;
239 }
240
241 /* Require an ACK every once in a while */
242 len = ntohl(rm->m_inc.i_hdr.h_len);
243 if (conn->c_unacked_packets == 0 ||
244 conn->c_unacked_bytes < len) {
245 __set_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags);
246
247 conn->c_unacked_packets = rds_sysctl_max_unacked_packets;
248 conn->c_unacked_bytes = rds_sysctl_max_unacked_bytes;
249 rds_stats_inc(s_send_ack_required);
250 } else {
251 conn->c_unacked_bytes -= len;
252 conn->c_unacked_packets--;
253 }
254
255 conn->c_xmit_rm = rm;
256 }
257
258 /* The transport either sends the whole rdma or none of it */
259 if (rm->rdma.op_active && !conn->c_xmit_rdma_sent) {
260 rm->m_final_op = &rm->rdma;
261 ret = conn->c_trans->xmit_rdma(conn, &rm->rdma);
262 if (ret)
263 break;
264 conn->c_xmit_rdma_sent = 1;
265
266 /* The transport owns the mapped memory for now.
267 * You can't unmap it while it's on the send queue */
268 set_bit(RDS_MSG_MAPPED, &rm->m_flags);
269 }
270
271 if (rm->atomic.op_active && !conn->c_xmit_atomic_sent) {
272 rm->m_final_op = &rm->atomic;
273 ret = conn->c_trans->xmit_atomic(conn, &rm->atomic);
274 if (ret)
275 break;
276 conn->c_xmit_atomic_sent = 1;
277
278 /* The transport owns the mapped memory for now.
279 * You can't unmap it while it's on the send queue */
280 set_bit(RDS_MSG_MAPPED, &rm->m_flags);
281 }
282
283 /*
284 * A number of cases require an RDS header to be sent
285 * even if there is no data.
286 * We permit 0-byte sends; rds-ping depends on this.
287 * However, if there are exclusively attached silent ops,
288 * we skip the hdr/data send, to enable silent operation.
289 */
290 if (rm->data.op_nents == 0) {
291 int ops_present;
292 int all_ops_are_silent = 1;
293
294 ops_present = (rm->atomic.op_active || rm->rdma.op_active);
295 if (rm->atomic.op_active && !rm->atomic.op_silent)
296 all_ops_are_silent = 0;
297 if (rm->rdma.op_active && !rm->rdma.op_silent)
298 all_ops_are_silent = 0;
299
300 if (ops_present && all_ops_are_silent
301 && !rm->m_rdma_cookie)
302 rm->data.op_active = 0;
303 }
304
305 if (rm->data.op_active && !conn->c_xmit_data_sent) {
306 rm->m_final_op = &rm->data;
307 ret = conn->c_trans->xmit(conn, rm,
308 conn->c_xmit_hdr_off,
309 conn->c_xmit_sg,
310 conn->c_xmit_data_off);
311 if (ret <= 0)
312 break;
313
314 if (conn->c_xmit_hdr_off < sizeof(struct rds_header)) {
315 tmp = min_t(int, ret,
316 sizeof(struct rds_header) -
317 conn->c_xmit_hdr_off);
318 conn->c_xmit_hdr_off += tmp;
319 ret -= tmp;
320 }
321
322 sg = &rm->data.op_sg[conn->c_xmit_sg];
323 while (ret) {
324 tmp = min_t(int, ret, sg->length -
325 conn->c_xmit_data_off);
326 conn->c_xmit_data_off += tmp;
327 ret -= tmp;
328 if (conn->c_xmit_data_off == sg->length) {
329 conn->c_xmit_data_off = 0;
330 sg++;
331 conn->c_xmit_sg++;
332 BUG_ON(ret != 0 &&
333 conn->c_xmit_sg == rm->data.op_nents);
334 }
335 }
336
337 if (conn->c_xmit_hdr_off == sizeof(struct rds_header) &&
338 (conn->c_xmit_sg == rm->data.op_nents))
339 conn->c_xmit_data_sent = 1;
340 }
341
342 /*
343 * A rm will only take multiple times through this loop
344 * if there is a data op. Thus, if the data is sent (or there was
345 * none), then we're done with the rm.
346 */
347 if (!rm->data.op_active || conn->c_xmit_data_sent) {
348 conn->c_xmit_rm = NULL;
349 conn->c_xmit_sg = 0;
350 conn->c_xmit_hdr_off = 0;
351 conn->c_xmit_data_off = 0;
352 conn->c_xmit_rdma_sent = 0;
353 conn->c_xmit_atomic_sent = 0;
354 conn->c_xmit_data_sent = 0;
355
356 rds_message_put(rm);
357 }
358 }
359
360 if (conn->c_trans->xmit_complete)
361 conn->c_trans->xmit_complete(conn);
362
363 release_in_xmit(conn);
364
365 /* Nuke any messages we decided not to retransmit. */
366 if (!list_empty(&to_be_dropped)) {
367 /* irqs on here, so we can put(), unlike above */
368 list_for_each_entry(rm, &to_be_dropped, m_conn_item)
369 rds_message_put(rm);
370 rds_send_remove_from_sock(&to_be_dropped, RDS_RDMA_DROPPED);
371 }
372
373 /*
374 * Other senders can queue a message after we last test the send queue
375 * but before we clear RDS_IN_XMIT. In that case they'd back off and
376 * not try and send their newly queued message. We need to check the
377 * send queue after having cleared RDS_IN_XMIT so that their message
378 * doesn't get stuck on the send queue.
379 *
380 * If the transport cannot continue (i.e ret != 0), then it must
381 * call us when more room is available, such as from the tx
382 * completion handler.
383 */
384 if (ret == 0) {
385 smp_mb();
386 if (!list_empty(&conn->c_send_queue)) {
387 rds_stats_inc(s_send_lock_queue_raced);
388 goto restart;
389 }
390 }
391out:
392 return ret;
393}
394
395static void rds_send_sndbuf_remove(struct rds_sock *rs, struct rds_message *rm)
396{
397 u32 len = be32_to_cpu(rm->m_inc.i_hdr.h_len);
398
399 assert_spin_locked(&rs->rs_lock);
400
401 BUG_ON(rs->rs_snd_bytes < len);
402 rs->rs_snd_bytes -= len;
403
404 if (rs->rs_snd_bytes == 0)
405 rds_stats_inc(s_send_queue_empty);
406}
407
408static inline int rds_send_is_acked(struct rds_message *rm, u64 ack,
409 is_acked_func is_acked)
410{
411 if (is_acked)
412 return is_acked(rm, ack);
413 return be64_to_cpu(rm->m_inc.i_hdr.h_sequence) <= ack;
414}
415
416/*
417 * This is pretty similar to what happens below in the ACK
418 * handling code - except that we call here as soon as we get
419 * the IB send completion on the RDMA op and the accompanying
420 * message.
421 */
422void rds_rdma_send_complete(struct rds_message *rm, int status)
423{
424 struct rds_sock *rs = NULL;
425 struct rm_rdma_op *ro;
426 struct rds_notifier *notifier;
427 unsigned long flags;
428
429 spin_lock_irqsave(&rm->m_rs_lock, flags);
430
431 ro = &rm->rdma;
432 if (test_bit(RDS_MSG_ON_SOCK, &rm->m_flags) &&
433 ro->op_active && ro->op_notify && ro->op_notifier) {
434 notifier = ro->op_notifier;
435 rs = rm->m_rs;
436 sock_hold(rds_rs_to_sk(rs));
437
438 notifier->n_status = status;
439 spin_lock(&rs->rs_lock);
440 list_add_tail(¬ifier->n_list, &rs->rs_notify_queue);
441 spin_unlock(&rs->rs_lock);
442
443 ro->op_notifier = NULL;
444 }
445
446 spin_unlock_irqrestore(&rm->m_rs_lock, flags);
447
448 if (rs) {
449 rds_wake_sk_sleep(rs);
450 sock_put(rds_rs_to_sk(rs));
451 }
452}
453EXPORT_SYMBOL_GPL(rds_rdma_send_complete);
454
455/*
456 * Just like above, except looks at atomic op
457 */
458void rds_atomic_send_complete(struct rds_message *rm, int status)
459{
460 struct rds_sock *rs = NULL;
461 struct rm_atomic_op *ao;
462 struct rds_notifier *notifier;
463 unsigned long flags;
464
465 spin_lock_irqsave(&rm->m_rs_lock, flags);
466
467 ao = &rm->atomic;
468 if (test_bit(RDS_MSG_ON_SOCK, &rm->m_flags)
469 && ao->op_active && ao->op_notify && ao->op_notifier) {
470 notifier = ao->op_notifier;
471 rs = rm->m_rs;
472 sock_hold(rds_rs_to_sk(rs));
473
474 notifier->n_status = status;
475 spin_lock(&rs->rs_lock);
476 list_add_tail(¬ifier->n_list, &rs->rs_notify_queue);
477 spin_unlock(&rs->rs_lock);
478
479 ao->op_notifier = NULL;
480 }
481
482 spin_unlock_irqrestore(&rm->m_rs_lock, flags);
483
484 if (rs) {
485 rds_wake_sk_sleep(rs);
486 sock_put(rds_rs_to_sk(rs));
487 }
488}
489EXPORT_SYMBOL_GPL(rds_atomic_send_complete);
490
491/*
492 * This is the same as rds_rdma_send_complete except we
493 * don't do any locking - we have all the ingredients (message,
494 * socket, socket lock) and can just move the notifier.
495 */
496static inline void
497__rds_send_complete(struct rds_sock *rs, struct rds_message *rm, int status)
498{
499 struct rm_rdma_op *ro;
500 struct rm_atomic_op *ao;
501
502 ro = &rm->rdma;
503 if (ro->op_active && ro->op_notify && ro->op_notifier) {
504 ro->op_notifier->n_status = status;
505 list_add_tail(&ro->op_notifier->n_list, &rs->rs_notify_queue);
506 ro->op_notifier = NULL;
507 }
508
509 ao = &rm->atomic;
510 if (ao->op_active && ao->op_notify && ao->op_notifier) {
511 ao->op_notifier->n_status = status;
512 list_add_tail(&ao->op_notifier->n_list, &rs->rs_notify_queue);
513 ao->op_notifier = NULL;
514 }
515
516 /* No need to wake the app - caller does this */
517}
518
519/*
520 * This is called from the IB send completion when we detect
521 * a RDMA operation that failed with remote access error.
522 * So speed is not an issue here.
523 */
524struct rds_message *rds_send_get_message(struct rds_connection *conn,
525 struct rm_rdma_op *op)
526{
527 struct rds_message *rm, *tmp, *found = NULL;
528 unsigned long flags;
529
530 spin_lock_irqsave(&conn->c_lock, flags);
531
532 list_for_each_entry_safe(rm, tmp, &conn->c_retrans, m_conn_item) {
533 if (&rm->rdma == op) {
534 atomic_inc(&rm->m_refcount);
535 found = rm;
536 goto out;
537 }
538 }
539
540 list_for_each_entry_safe(rm, tmp, &conn->c_send_queue, m_conn_item) {
541 if (&rm->rdma == op) {
542 atomic_inc(&rm->m_refcount);
543 found = rm;
544 break;
545 }
546 }
547
548out:
549 spin_unlock_irqrestore(&conn->c_lock, flags);
550
551 return found;
552}
553EXPORT_SYMBOL_GPL(rds_send_get_message);
554
555/*
556 * This removes messages from the socket's list if they're on it. The list
557 * argument must be private to the caller, we must be able to modify it
558 * without locks. The messages must have a reference held for their
559 * position on the list. This function will drop that reference after
560 * removing the messages from the 'messages' list regardless of if it found
561 * the messages on the socket list or not.
562 */
563static void rds_send_remove_from_sock(struct list_head *messages, int status)
564{
565 unsigned long flags;
566 struct rds_sock *rs = NULL;
567 struct rds_message *rm;
568
569 while (!list_empty(messages)) {
570 int was_on_sock = 0;
571
572 rm = list_entry(messages->next, struct rds_message,
573 m_conn_item);
574 list_del_init(&rm->m_conn_item);
575
576 /*
577 * If we see this flag cleared then we're *sure* that someone
578 * else beat us to removing it from the sock. If we race
579 * with their flag update we'll get the lock and then really
580 * see that the flag has been cleared.
581 *
582 * The message spinlock makes sure nobody clears rm->m_rs
583 * while we're messing with it. It does not prevent the
584 * message from being removed from the socket, though.
585 */
586 spin_lock_irqsave(&rm->m_rs_lock, flags);
587 if (!test_bit(RDS_MSG_ON_SOCK, &rm->m_flags))
588 goto unlock_and_drop;
589
590 if (rs != rm->m_rs) {
591 if (rs) {
592 rds_wake_sk_sleep(rs);
593 sock_put(rds_rs_to_sk(rs));
594 }
595 rs = rm->m_rs;
596 sock_hold(rds_rs_to_sk(rs));
597 }
598 spin_lock(&rs->rs_lock);
599
600 if (test_and_clear_bit(RDS_MSG_ON_SOCK, &rm->m_flags)) {
601 struct rm_rdma_op *ro = &rm->rdma;
602 struct rds_notifier *notifier;
603
604 list_del_init(&rm->m_sock_item);
605 rds_send_sndbuf_remove(rs, rm);
606
607 if (ro->op_active && ro->op_notifier &&
608 (ro->op_notify || (ro->op_recverr && status))) {
609 notifier = ro->op_notifier;
610 list_add_tail(¬ifier->n_list,
611 &rs->rs_notify_queue);
612 if (!notifier->n_status)
613 notifier->n_status = status;
614 rm->rdma.op_notifier = NULL;
615 }
616 was_on_sock = 1;
617 rm->m_rs = NULL;
618 }
619 spin_unlock(&rs->rs_lock);
620
621unlock_and_drop:
622 spin_unlock_irqrestore(&rm->m_rs_lock, flags);
623 rds_message_put(rm);
624 if (was_on_sock)
625 rds_message_put(rm);
626 }
627
628 if (rs) {
629 rds_wake_sk_sleep(rs);
630 sock_put(rds_rs_to_sk(rs));
631 }
632}
633
634/*
635 * Transports call here when they've determined that the receiver queued
636 * messages up to, and including, the given sequence number. Messages are
637 * moved to the retrans queue when rds_send_xmit picks them off the send
638 * queue. This means that in the TCP case, the message may not have been
639 * assigned the m_ack_seq yet - but that's fine as long as tcp_is_acked
640 * checks the RDS_MSG_HAS_ACK_SEQ bit.
641 *
642 * XXX It's not clear to me how this is safely serialized with socket
643 * destruction. Maybe it should bail if it sees SOCK_DEAD.
644 */
645void rds_send_drop_acked(struct rds_connection *conn, u64 ack,
646 is_acked_func is_acked)
647{
648 struct rds_message *rm, *tmp;
649 unsigned long flags;
650 LIST_HEAD(list);
651
652 spin_lock_irqsave(&conn->c_lock, flags);
653
654 list_for_each_entry_safe(rm, tmp, &conn->c_retrans, m_conn_item) {
655 if (!rds_send_is_acked(rm, ack, is_acked))
656 break;
657
658 list_move(&rm->m_conn_item, &list);
659 clear_bit(RDS_MSG_ON_CONN, &rm->m_flags);
660 }
661
662 /* order flag updates with spin locks */
663 if (!list_empty(&list))
664 smp_mb__after_clear_bit();
665
666 spin_unlock_irqrestore(&conn->c_lock, flags);
667
668 /* now remove the messages from the sock list as needed */
669 rds_send_remove_from_sock(&list, RDS_RDMA_SUCCESS);
670}
671EXPORT_SYMBOL_GPL(rds_send_drop_acked);
672
673void rds_send_drop_to(struct rds_sock *rs, struct sockaddr_in *dest)
674{
675 struct rds_message *rm, *tmp;
676 struct rds_connection *conn;
677 unsigned long flags;
678 LIST_HEAD(list);
679
680 /* get all the messages we're dropping under the rs lock */
681 spin_lock_irqsave(&rs->rs_lock, flags);
682
683 list_for_each_entry_safe(rm, tmp, &rs->rs_send_queue, m_sock_item) {
684 if (dest && (dest->sin_addr.s_addr != rm->m_daddr ||
685 dest->sin_port != rm->m_inc.i_hdr.h_dport))
686 continue;
687
688 list_move(&rm->m_sock_item, &list);
689 rds_send_sndbuf_remove(rs, rm);
690 clear_bit(RDS_MSG_ON_SOCK, &rm->m_flags);
691 }
692
693 /* order flag updates with the rs lock */
694 smp_mb__after_clear_bit();
695
696 spin_unlock_irqrestore(&rs->rs_lock, flags);
697
698 if (list_empty(&list))
699 return;
700
701 /* Remove the messages from the conn */
702 list_for_each_entry(rm, &list, m_sock_item) {
703
704 conn = rm->m_inc.i_conn;
705
706 spin_lock_irqsave(&conn->c_lock, flags);
707 /*
708 * Maybe someone else beat us to removing rm from the conn.
709 * If we race with their flag update we'll get the lock and
710 * then really see that the flag has been cleared.
711 */
712 if (!test_and_clear_bit(RDS_MSG_ON_CONN, &rm->m_flags)) {
713 spin_unlock_irqrestore(&conn->c_lock, flags);
714 continue;
715 }
716 list_del_init(&rm->m_conn_item);
717 spin_unlock_irqrestore(&conn->c_lock, flags);
718
719 /*
720 * Couldn't grab m_rs_lock in top loop (lock ordering),
721 * but we can now.
722 */
723 spin_lock_irqsave(&rm->m_rs_lock, flags);
724
725 spin_lock(&rs->rs_lock);
726 __rds_send_complete(rs, rm, RDS_RDMA_CANCELED);
727 spin_unlock(&rs->rs_lock);
728
729 rm->m_rs = NULL;
730 spin_unlock_irqrestore(&rm->m_rs_lock, flags);
731
732 rds_message_put(rm);
733 }
734
735 rds_wake_sk_sleep(rs);
736
737 while (!list_empty(&list)) {
738 rm = list_entry(list.next, struct rds_message, m_sock_item);
739 list_del_init(&rm->m_sock_item);
740
741 rds_message_wait(rm);
742 rds_message_put(rm);
743 }
744}
745
746/*
747 * we only want this to fire once so we use the callers 'queued'. It's
748 * possible that another thread can race with us and remove the
749 * message from the flow with RDS_CANCEL_SENT_TO.
750 */
751static int rds_send_queue_rm(struct rds_sock *rs, struct rds_connection *conn,
752 struct rds_message *rm, __be16 sport,
753 __be16 dport, int *queued)
754{
755 unsigned long flags;
756 u32 len;
757
758 if (*queued)
759 goto out;
760
761 len = be32_to_cpu(rm->m_inc.i_hdr.h_len);
762
763 /* this is the only place which holds both the socket's rs_lock
764 * and the connection's c_lock */
765 spin_lock_irqsave(&rs->rs_lock, flags);
766
767 /*
768 * If there is a little space in sndbuf, we don't queue anything,
769 * and userspace gets -EAGAIN. But poll() indicates there's send
770 * room. This can lead to bad behavior (spinning) if snd_bytes isn't
771 * freed up by incoming acks. So we check the *old* value of
772 * rs_snd_bytes here to allow the last msg to exceed the buffer,
773 * and poll() now knows no more data can be sent.
774 */
775 if (rs->rs_snd_bytes < rds_sk_sndbuf(rs)) {
776 rs->rs_snd_bytes += len;
777
778 /* let recv side know we are close to send space exhaustion.
779 * This is probably not the optimal way to do it, as this
780 * means we set the flag on *all* messages as soon as our
781 * throughput hits a certain threshold.
782 */
783 if (rs->rs_snd_bytes >= rds_sk_sndbuf(rs) / 2)
784 __set_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags);
785
786 list_add_tail(&rm->m_sock_item, &rs->rs_send_queue);
787 set_bit(RDS_MSG_ON_SOCK, &rm->m_flags);
788 rds_message_addref(rm);
789 rm->m_rs = rs;
790
791 /* The code ordering is a little weird, but we're
792 trying to minimize the time we hold c_lock */
793 rds_message_populate_header(&rm->m_inc.i_hdr, sport, dport, 0);
794 rm->m_inc.i_conn = conn;
795 rds_message_addref(rm);
796
797 spin_lock(&conn->c_lock);
798 rm->m_inc.i_hdr.h_sequence = cpu_to_be64(conn->c_next_tx_seq++);
799 list_add_tail(&rm->m_conn_item, &conn->c_send_queue);
800 set_bit(RDS_MSG_ON_CONN, &rm->m_flags);
801 spin_unlock(&conn->c_lock);
802
803 rdsdebug("queued msg %p len %d, rs %p bytes %d seq %llu\n",
804 rm, len, rs, rs->rs_snd_bytes,
805 (unsigned long long)be64_to_cpu(rm->m_inc.i_hdr.h_sequence));
806
807 *queued = 1;
808 }
809
810 spin_unlock_irqrestore(&rs->rs_lock, flags);
811out:
812 return *queued;
813}
814
815/*
816 * rds_message is getting to be quite complicated, and we'd like to allocate
817 * it all in one go. This figures out how big it needs to be up front.
818 */
819static int rds_rm_size(struct msghdr *msg, int data_len)
820{
821 struct cmsghdr *cmsg;
822 int size = 0;
823 int cmsg_groups = 0;
824 int retval;
825
826 for (cmsg = CMSG_FIRSTHDR(msg); cmsg; cmsg = CMSG_NXTHDR(msg, cmsg)) {
827 if (!CMSG_OK(msg, cmsg))
828 return -EINVAL;
829
830 if (cmsg->cmsg_level != SOL_RDS)
831 continue;
832
833 switch (cmsg->cmsg_type) {
834 case RDS_CMSG_RDMA_ARGS:
835 cmsg_groups |= 1;
836 retval = rds_rdma_extra_size(CMSG_DATA(cmsg));
837 if (retval < 0)
838 return retval;
839 size += retval;
840
841 break;
842
843 case RDS_CMSG_RDMA_DEST:
844 case RDS_CMSG_RDMA_MAP:
845 cmsg_groups |= 2;
846 /* these are valid but do no add any size */
847 break;
848
849 case RDS_CMSG_ATOMIC_CSWP:
850 case RDS_CMSG_ATOMIC_FADD:
851 case RDS_CMSG_MASKED_ATOMIC_CSWP:
852 case RDS_CMSG_MASKED_ATOMIC_FADD:
853 cmsg_groups |= 1;
854 size += sizeof(struct scatterlist);
855 break;
856
857 default:
858 return -EINVAL;
859 }
860
861 }
862
863 size += ceil(data_len, PAGE_SIZE) * sizeof(struct scatterlist);
864
865 /* Ensure (DEST, MAP) are never used with (ARGS, ATOMIC) */
866 if (cmsg_groups == 3)
867 return -EINVAL;
868
869 return size;
870}
871
872static int rds_cmsg_send(struct rds_sock *rs, struct rds_message *rm,
873 struct msghdr *msg, int *allocated_mr)
874{
875 struct cmsghdr *cmsg;
876 int ret = 0;
877
878 for (cmsg = CMSG_FIRSTHDR(msg); cmsg; cmsg = CMSG_NXTHDR(msg, cmsg)) {
879 if (!CMSG_OK(msg, cmsg))
880 return -EINVAL;
881
882 if (cmsg->cmsg_level != SOL_RDS)
883 continue;
884
885 /* As a side effect, RDMA_DEST and RDMA_MAP will set
886 * rm->rdma.m_rdma_cookie and rm->rdma.m_rdma_mr.
887 */
888 switch (cmsg->cmsg_type) {
889 case RDS_CMSG_RDMA_ARGS:
890 ret = rds_cmsg_rdma_args(rs, rm, cmsg);
891 break;
892
893 case RDS_CMSG_RDMA_DEST:
894 ret = rds_cmsg_rdma_dest(rs, rm, cmsg);
895 break;
896
897 case RDS_CMSG_RDMA_MAP:
898 ret = rds_cmsg_rdma_map(rs, rm, cmsg);
899 if (!ret)
900 *allocated_mr = 1;
901 break;
902 case RDS_CMSG_ATOMIC_CSWP:
903 case RDS_CMSG_ATOMIC_FADD:
904 case RDS_CMSG_MASKED_ATOMIC_CSWP:
905 case RDS_CMSG_MASKED_ATOMIC_FADD:
906 ret = rds_cmsg_atomic(rs, rm, cmsg);
907 break;
908
909 default:
910 return -EINVAL;
911 }
912
913 if (ret)
914 break;
915 }
916
917 return ret;
918}
919
920int rds_sendmsg(struct kiocb *iocb, struct socket *sock, struct msghdr *msg,
921 size_t payload_len)
922{
923 struct sock *sk = sock->sk;
924 struct rds_sock *rs = rds_sk_to_rs(sk);
925 struct sockaddr_in *usin = (struct sockaddr_in *)msg->msg_name;
926 __be32 daddr;
927 __be16 dport;
928 struct rds_message *rm = NULL;
929 struct rds_connection *conn;
930 int ret = 0;
931 int queued = 0, allocated_mr = 0;
932 int nonblock = msg->msg_flags & MSG_DONTWAIT;
933 long timeo = sock_sndtimeo(sk, nonblock);
934
935 /* Mirror Linux UDP mirror of BSD error message compatibility */
936 /* XXX: Perhaps MSG_MORE someday */
937 if (msg->msg_flags & ~(MSG_DONTWAIT | MSG_CMSG_COMPAT)) {
938 ret = -EOPNOTSUPP;
939 goto out;
940 }
941
942 if (msg->msg_namelen) {
943 /* XXX fail non-unicast destination IPs? */
944 if (msg->msg_namelen < sizeof(*usin) || usin->sin_family != AF_INET) {
945 ret = -EINVAL;
946 goto out;
947 }
948 daddr = usin->sin_addr.s_addr;
949 dport = usin->sin_port;
950 } else {
951 /* We only care about consistency with ->connect() */
952 lock_sock(sk);
953 daddr = rs->rs_conn_addr;
954 dport = rs->rs_conn_port;
955 release_sock(sk);
956 }
957
958 /* racing with another thread binding seems ok here */
959 if (daddr == 0 || rs->rs_bound_addr == 0) {
960 ret = -ENOTCONN; /* XXX not a great errno */
961 goto out;
962 }
963
964 /* size of rm including all sgs */
965 ret = rds_rm_size(msg, payload_len);
966 if (ret < 0)
967 goto out;
968
969 rm = rds_message_alloc(ret, GFP_KERNEL);
970 if (!rm) {
971 ret = -ENOMEM;
972 goto out;
973 }
974
975 /* Attach data to the rm */
976 if (payload_len) {
977 rm->data.op_sg = rds_message_alloc_sgs(rm, ceil(payload_len, PAGE_SIZE));
978 if (!rm->data.op_sg) {
979 ret = -ENOMEM;
980 goto out;
981 }
982 ret = rds_message_copy_from_user(rm, msg->msg_iov, payload_len);
983 if (ret)
984 goto out;
985 }
986 rm->data.op_active = 1;
987
988 rm->m_daddr = daddr;
989
990 /* rds_conn_create has a spinlock that runs with IRQ off.
991 * Caching the conn in the socket helps a lot. */
992 if (rs->rs_conn && rs->rs_conn->c_faddr == daddr)
993 conn = rs->rs_conn;
994 else {
995 conn = rds_conn_create_outgoing(rs->rs_bound_addr, daddr,
996 rs->rs_transport,
997 sock->sk->sk_allocation);
998 if (IS_ERR(conn)) {
999 ret = PTR_ERR(conn);
1000 goto out;
1001 }
1002 rs->rs_conn = conn;
1003 }
1004
1005 /* Parse any control messages the user may have included. */
1006 ret = rds_cmsg_send(rs, rm, msg, &allocated_mr);
1007 if (ret)
1008 goto out;
1009
1010 if (rm->rdma.op_active && !conn->c_trans->xmit_rdma) {
1011 printk_ratelimited(KERN_NOTICE "rdma_op %p conn xmit_rdma %p\n",
1012 &rm->rdma, conn->c_trans->xmit_rdma);
1013 ret = -EOPNOTSUPP;
1014 goto out;
1015 }
1016
1017 if (rm->atomic.op_active && !conn->c_trans->xmit_atomic) {
1018 printk_ratelimited(KERN_NOTICE "atomic_op %p conn xmit_atomic %p\n",
1019 &rm->atomic, conn->c_trans->xmit_atomic);
1020 ret = -EOPNOTSUPP;
1021 goto out;
1022 }
1023
1024 rds_conn_connect_if_down(conn);
1025
1026 ret = rds_cong_wait(conn->c_fcong, dport, nonblock, rs);
1027 if (ret) {
1028 rs->rs_seen_congestion = 1;
1029 goto out;
1030 }
1031
1032 while (!rds_send_queue_rm(rs, conn, rm, rs->rs_bound_port,
1033 dport, &queued)) {
1034 rds_stats_inc(s_send_queue_full);
1035 /* XXX make sure this is reasonable */
1036 if (payload_len > rds_sk_sndbuf(rs)) {
1037 ret = -EMSGSIZE;
1038 goto out;
1039 }
1040 if (nonblock) {
1041 ret = -EAGAIN;
1042 goto out;
1043 }
1044
1045 timeo = wait_event_interruptible_timeout(*sk_sleep(sk),
1046 rds_send_queue_rm(rs, conn, rm,
1047 rs->rs_bound_port,
1048 dport,
1049 &queued),
1050 timeo);
1051 rdsdebug("sendmsg woke queued %d timeo %ld\n", queued, timeo);
1052 if (timeo > 0 || timeo == MAX_SCHEDULE_TIMEOUT)
1053 continue;
1054
1055 ret = timeo;
1056 if (ret == 0)
1057 ret = -ETIMEDOUT;
1058 goto out;
1059 }
1060
1061 /*
1062 * By now we've committed to the send. We reuse rds_send_worker()
1063 * to retry sends in the rds thread if the transport asks us to.
1064 */
1065 rds_stats_inc(s_send_queued);
1066
1067 if (!test_bit(RDS_LL_SEND_FULL, &conn->c_flags))
1068 rds_send_xmit(conn);
1069
1070 rds_message_put(rm);
1071 return payload_len;
1072
1073out:
1074 /* If the user included a RDMA_MAP cmsg, we allocated a MR on the fly.
1075 * If the sendmsg goes through, we keep the MR. If it fails with EAGAIN
1076 * or in any other way, we need to destroy the MR again */
1077 if (allocated_mr)
1078 rds_rdma_unuse(rs, rds_rdma_cookie_key(rm->m_rdma_cookie), 1);
1079
1080 if (rm)
1081 rds_message_put(rm);
1082 return ret;
1083}
1084
1085/*
1086 * Reply to a ping packet.
1087 */
1088int
1089rds_send_pong(struct rds_connection *conn, __be16 dport)
1090{
1091 struct rds_message *rm;
1092 unsigned long flags;
1093 int ret = 0;
1094
1095 rm = rds_message_alloc(0, GFP_ATOMIC);
1096 if (!rm) {
1097 ret = -ENOMEM;
1098 goto out;
1099 }
1100
1101 rm->m_daddr = conn->c_faddr;
1102 rm->data.op_active = 1;
1103
1104 rds_conn_connect_if_down(conn);
1105
1106 ret = rds_cong_wait(conn->c_fcong, dport, 1, NULL);
1107 if (ret)
1108 goto out;
1109
1110 spin_lock_irqsave(&conn->c_lock, flags);
1111 list_add_tail(&rm->m_conn_item, &conn->c_send_queue);
1112 set_bit(RDS_MSG_ON_CONN, &rm->m_flags);
1113 rds_message_addref(rm);
1114 rm->m_inc.i_conn = conn;
1115
1116 rds_message_populate_header(&rm->m_inc.i_hdr, 0, dport,
1117 conn->c_next_tx_seq);
1118 conn->c_next_tx_seq++;
1119 spin_unlock_irqrestore(&conn->c_lock, flags);
1120
1121 rds_stats_inc(s_send_queued);
1122 rds_stats_inc(s_send_pong);
1123
1124 if (!test_bit(RDS_LL_SEND_FULL, &conn->c_flags))
1125 rds_send_xmit(conn);
1126
1127 rds_message_put(rm);
1128 return 0;
1129
1130out:
1131 if (rm)
1132 rds_message_put(rm);
1133 return ret;
1134}