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