blob: e9430f537f9c2bb23bbaeeb66933e1e85058bd34 [file] [log] [blame]
Andy Grover5c115592009-02-24 15:30:27 +00001/*
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>
Paul Gortmakerd9b93842011-09-18 13:21:27 -040034#include <linux/moduleparam.h>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090035#include <linux/gfp.h>
Andy Grover5c115592009-02-24 15:30:27 +000036#include <net/sock.h>
37#include <linux/in.h>
38#include <linux/list.h>
Manuel Zerpiescb0a6052011-06-16 02:09:57 +000039#include <linux/ratelimit.h>
Paul Gortmakerbc3b2d72011-07-15 11:47:34 -040040#include <linux/export.h>
Andy Grover5c115592009-02-24 15:30:27 +000041
42#include "rds.h"
Andy Grover5c115592009-02-24 15:30:27 +000043
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
stephen hemmingerff51bf82010-10-19 08:08:33 +000058static void rds_send_remove_from_sock(struct list_head *messages, int status);
59
Andy Grover5c115592009-02-24 15:30:27 +000060/*
Zach Brown0f4b1c72010-06-04 14:41:41 -070061 * Reset the send state. Callers must ensure that this doesn't race with
62 * rds_send_xmit().
Andy Grover5c115592009-02-24 15:30:27 +000063 */
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) {
Chris Mason7e3f2952010-05-11 15:11:11 -070070 rm = conn->c_xmit_rm;
71 conn->c_xmit_rm = NULL;
Andy Grover5c115592009-02-24 15:30:27 +000072 /* 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 */
Chris Mason7e3f2952010-05-11 15:11:11 -070076 rds_message_unmapped(rm);
Chris Mason7e3f2952010-05-11 15:11:11 -070077 rds_message_put(rm);
Andy Grover5c115592009-02-24 15:30:27 +000078 }
Chris Mason7e3f2952010-05-11 15:11:11 -070079
Andy Grover5c115592009-02-24 15:30:27 +000080 conn->c_xmit_sg = 0;
81 conn->c_xmit_hdr_off = 0;
82 conn->c_xmit_data_off = 0;
Andy Grover15133f62010-01-12 14:33:38 -080083 conn->c_xmit_atomic_sent = 0;
Andy Grover5b2366b2010-02-03 19:36:44 -080084 conn->c_xmit_rdma_sent = 0;
85 conn->c_xmit_data_sent = 0;
Andy Grover5c115592009-02-24 15:30:27 +000086
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
Zach Brown0f4b1c72010-06-04 14:41:41 -0700102static 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);
Peter Zijlstra4e857c52014-03-17 18:06:10 +0100110 smp_mb__after_atomic();
Zach Brown0f4b1c72010-06-04 14:41:41 -0700111 /*
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
Andy Grover5c115592009-02-24 15:30:27 +0000121/*
Lucas De Marchi25985ed2011-03-30 22:57:33 -0300122 * We're making the conscious trade-off here to only send one message
Andy Grover5c115592009-02-24 15:30:27 +0000123 * 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;
Andy Grover5c115592009-02-24 15:30:27 +0000140 struct scatterlist *sg;
141 int ret = 0;
Andy Grover5c115592009-02-24 15:30:27 +0000142 LIST_HEAD(to_be_dropped);
Sowmini Varadhan443be0e2015-04-08 12:33:47 -0400143 int batch_count;
144 unsigned long send_gen = 0;
Andy Grover5c115592009-02-24 15:30:27 +0000145
Andy Groverfcc54502010-03-29 17:08:49 -0700146restart:
Sowmini Varadhan443be0e2015-04-08 12:33:47 -0400147 batch_count = 0;
Andy Grover049ee3f2010-03-23 17:39:07 -0700148
Andy Grover5c115592009-02-24 15:30:27 +0000149 /*
150 * sendmsg calls here after having queued its message on the send
151 * queue. We only have one task feeding the connection at a time. If
152 * another thread is already feeding the queue then we back off. This
153 * avoids blocking the caller and trading per-connection data between
154 * caches per message.
Andy Grover5c115592009-02-24 15:30:27 +0000155 */
Zach Brown0f4b1c72010-06-04 14:41:41 -0700156 if (!acquire_in_xmit(conn)) {
Andy Grover049ee3f2010-03-23 17:39:07 -0700157 rds_stats_inc(s_send_lock_contention);
Andy Grover5c115592009-02-24 15:30:27 +0000158 ret = -ENOMEM;
159 goto out;
160 }
Zach Brown0f4b1c72010-06-04 14:41:41 -0700161
162 /*
Sowmini Varadhan443be0e2015-04-08 12:33:47 -0400163 * we record the send generation after doing the xmit acquire.
164 * if someone else manages to jump in and do some work, we'll use
165 * this to avoid a goto restart farther down.
166 *
167 * The acquire_in_xmit() check above ensures that only one
168 * caller can increment c_send_gen at any time.
169 */
170 conn->c_send_gen++;
171 send_gen = conn->c_send_gen;
172
173 /*
Zach Brown0f4b1c72010-06-04 14:41:41 -0700174 * rds_conn_shutdown() sets the conn state and then tests RDS_IN_XMIT,
175 * we do the opposite to avoid races.
176 */
177 if (!rds_conn_up(conn)) {
178 release_in_xmit(conn);
179 ret = 0;
180 goto out;
181 }
Andy Grover5c115592009-02-24 15:30:27 +0000182
183 if (conn->c_trans->xmit_prepare)
184 conn->c_trans->xmit_prepare(conn);
185
186 /*
187 * spin trying to push headers and data down the connection until
Andy Grover5b2366b2010-02-03 19:36:44 -0800188 * the connection doesn't make forward progress.
Andy Grover5c115592009-02-24 15:30:27 +0000189 */
Andy Groverfcc54502010-03-29 17:08:49 -0700190 while (1) {
Andy Grover5c115592009-02-24 15:30:27 +0000191
Andy Grover5c115592009-02-24 15:30:27 +0000192 rm = conn->c_xmit_rm;
Andy Grover5c115592009-02-24 15:30:27 +0000193
Andy Grover5b2366b2010-02-03 19:36:44 -0800194 /*
195 * If between sending messages, we can send a pending congestion
196 * map update.
Andy Grover5c115592009-02-24 15:30:27 +0000197 */
Andy Grover8690bfa2010-01-12 11:56:44 -0800198 if (!rm && test_and_clear_bit(0, &conn->c_map_queued)) {
Andy Grover77dd5502010-03-22 15:22:04 -0700199 rm = rds_cong_update_alloc(conn);
200 if (IS_ERR(rm)) {
201 ret = PTR_ERR(rm);
202 break;
Andy Grover5b2366b2010-02-03 19:36:44 -0800203 }
Andy Grover77dd5502010-03-22 15:22:04 -0700204 rm->data.op_active = 1;
205
206 conn->c_xmit_rm = rm;
Andy Grover5c115592009-02-24 15:30:27 +0000207 }
208
209 /*
Andy Grover5b2366b2010-02-03 19:36:44 -0800210 * If not already working on one, grab the next message.
Andy Grover5c115592009-02-24 15:30:27 +0000211 *
212 * c_xmit_rm holds a ref while we're sending this message down
213 * the connction. We can use this ref while holding the
214 * send_sem.. rds_send_reset() is serialized with it.
215 */
Andy Grover8690bfa2010-01-12 11:56:44 -0800216 if (!rm) {
Andy Grover5c115592009-02-24 15:30:27 +0000217 unsigned int len;
218
Sowmini Varadhan443be0e2015-04-08 12:33:47 -0400219 batch_count++;
220
221 /* we want to process as big a batch as we can, but
222 * we also want to avoid softlockups. If we've been
223 * through a lot of messages, lets back off and see
224 * if anyone else jumps in
225 */
226 if (batch_count >= 1024)
227 goto over_batch;
228
Zach Brown0f4b1c72010-06-04 14:41:41 -0700229 spin_lock_irqsave(&conn->c_lock, flags);
Andy Grover5c115592009-02-24 15:30:27 +0000230
231 if (!list_empty(&conn->c_send_queue)) {
232 rm = list_entry(conn->c_send_queue.next,
233 struct rds_message,
234 m_conn_item);
235 rds_message_addref(rm);
236
237 /*
238 * Move the message from the send queue to the retransmit
239 * list right away.
240 */
241 list_move_tail(&rm->m_conn_item, &conn->c_retrans);
242 }
243
Zach Brown0f4b1c72010-06-04 14:41:41 -0700244 spin_unlock_irqrestore(&conn->c_lock, flags);
Andy Grover5c115592009-02-24 15:30:27 +0000245
Andy Groverfcc54502010-03-29 17:08:49 -0700246 if (!rm)
Andy Grover5c115592009-02-24 15:30:27 +0000247 break;
Andy Grover5c115592009-02-24 15:30:27 +0000248
249 /* Unfortunately, the way Infiniband deals with
250 * RDMA to a bad MR key is by moving the entire
251 * queue pair to error state. We cold possibly
252 * recover from that, but right now we drop the
253 * connection.
254 * Therefore, we never retransmit messages with RDMA ops.
255 */
Andy Groverf8b3aaf2010-03-01 14:11:53 -0800256 if (rm->rdma.op_active &&
Joe Perchesf64f9e72009-11-29 16:55:45 -0800257 test_bit(RDS_MSG_RETRANSMITTED, &rm->m_flags)) {
Zach Brown0f4b1c72010-06-04 14:41:41 -0700258 spin_lock_irqsave(&conn->c_lock, flags);
Andy Grover5c115592009-02-24 15:30:27 +0000259 if (test_and_clear_bit(RDS_MSG_ON_CONN, &rm->m_flags))
260 list_move(&rm->m_conn_item, &to_be_dropped);
Zach Brown0f4b1c72010-06-04 14:41:41 -0700261 spin_unlock_irqrestore(&conn->c_lock, flags);
Andy Grover5c115592009-02-24 15:30:27 +0000262 continue;
263 }
264
265 /* Require an ACK every once in a while */
266 len = ntohl(rm->m_inc.i_hdr.h_len);
Joe Perchesf64f9e72009-11-29 16:55:45 -0800267 if (conn->c_unacked_packets == 0 ||
268 conn->c_unacked_bytes < len) {
Andy Grover5c115592009-02-24 15:30:27 +0000269 __set_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags);
270
271 conn->c_unacked_packets = rds_sysctl_max_unacked_packets;
272 conn->c_unacked_bytes = rds_sysctl_max_unacked_bytes;
273 rds_stats_inc(s_send_ack_required);
274 } else {
275 conn->c_unacked_bytes -= len;
276 conn->c_unacked_packets--;
277 }
278
279 conn->c_xmit_rm = rm;
280 }
281
Andy Grover2c3a5f92010-03-01 16:10:40 -0800282 /* The transport either sends the whole rdma or none of it */
283 if (rm->rdma.op_active && !conn->c_xmit_rdma_sent) {
Andy Groverff3d7d32010-03-01 14:03:09 -0800284 rm->m_final_op = &rm->rdma;
Andy Grover2c3a5f92010-03-01 16:10:40 -0800285 ret = conn->c_trans->xmit_rdma(conn, &rm->rdma);
Chris Mason1cc22282010-05-11 16:15:35 -0700286 if (ret)
Andy Grover2c3a5f92010-03-01 16:10:40 -0800287 break;
288 conn->c_xmit_rdma_sent = 1;
289
290 /* The transport owns the mapped memory for now.
291 * You can't unmap it while it's on the send queue */
292 set_bit(RDS_MSG_MAPPED, &rm->m_flags);
293 }
294
Andy Grover15133f62010-01-12 14:33:38 -0800295 if (rm->atomic.op_active && !conn->c_xmit_atomic_sent) {
Andy Groverff3d7d32010-03-01 14:03:09 -0800296 rm->m_final_op = &rm->atomic;
297 ret = conn->c_trans->xmit_atomic(conn, &rm->atomic);
Chris Mason1cc22282010-05-11 16:15:35 -0700298 if (ret)
Andy Grover15133f62010-01-12 14:33:38 -0800299 break;
300 conn->c_xmit_atomic_sent = 1;
Andy Groverff3d7d32010-03-01 14:03:09 -0800301
Andy Grover15133f62010-01-12 14:33:38 -0800302 /* The transport owns the mapped memory for now.
303 * You can't unmap it while it's on the send queue */
304 set_bit(RDS_MSG_MAPPED, &rm->m_flags);
305 }
306
Andy Grover2c3a5f92010-03-01 16:10:40 -0800307 /*
308 * A number of cases require an RDS header to be sent
309 * even if there is no data.
310 * We permit 0-byte sends; rds-ping depends on this.
311 * However, if there are exclusively attached silent ops,
312 * we skip the hdr/data send, to enable silent operation.
313 */
314 if (rm->data.op_nents == 0) {
315 int ops_present;
316 int all_ops_are_silent = 1;
Andy Grover241eef32010-01-19 21:25:26 -0800317
Andy Grover2c3a5f92010-03-01 16:10:40 -0800318 ops_present = (rm->atomic.op_active || rm->rdma.op_active);
319 if (rm->atomic.op_active && !rm->atomic.op_silent)
320 all_ops_are_silent = 0;
321 if (rm->rdma.op_active && !rm->rdma.op_silent)
322 all_ops_are_silent = 0;
Andy Grover241eef32010-01-19 21:25:26 -0800323
Andy Grover2c3a5f92010-03-01 16:10:40 -0800324 if (ops_present && all_ops_are_silent
325 && !rm->m_rdma_cookie)
326 rm->data.op_active = 0;
Andy Grover5c115592009-02-24 15:30:27 +0000327 }
328
Andy Grover5b2366b2010-02-03 19:36:44 -0800329 if (rm->data.op_active && !conn->c_xmit_data_sent) {
Andy Groverff3d7d32010-03-01 14:03:09 -0800330 rm->m_final_op = &rm->data;
Andy Grover5c115592009-02-24 15:30:27 +0000331 ret = conn->c_trans->xmit(conn, rm,
332 conn->c_xmit_hdr_off,
333 conn->c_xmit_sg,
334 conn->c_xmit_data_off);
335 if (ret <= 0)
336 break;
337
338 if (conn->c_xmit_hdr_off < sizeof(struct rds_header)) {
339 tmp = min_t(int, ret,
340 sizeof(struct rds_header) -
341 conn->c_xmit_hdr_off);
342 conn->c_xmit_hdr_off += tmp;
343 ret -= tmp;
344 }
345
Andy Grover6c7cc6e2010-01-27 18:04:18 -0800346 sg = &rm->data.op_sg[conn->c_xmit_sg];
Andy Grover5c115592009-02-24 15:30:27 +0000347 while (ret) {
348 tmp = min_t(int, ret, sg->length -
349 conn->c_xmit_data_off);
350 conn->c_xmit_data_off += tmp;
351 ret -= tmp;
352 if (conn->c_xmit_data_off == sg->length) {
353 conn->c_xmit_data_off = 0;
354 sg++;
355 conn->c_xmit_sg++;
356 BUG_ON(ret != 0 &&
Andy Grover6c7cc6e2010-01-27 18:04:18 -0800357 conn->c_xmit_sg == rm->data.op_nents);
Andy Grover5c115592009-02-24 15:30:27 +0000358 }
359 }
Andy Grover5b2366b2010-02-03 19:36:44 -0800360
361 if (conn->c_xmit_hdr_off == sizeof(struct rds_header) &&
362 (conn->c_xmit_sg == rm->data.op_nents))
363 conn->c_xmit_data_sent = 1;
364 }
365
366 /*
367 * A rm will only take multiple times through this loop
368 * if there is a data op. Thus, if the data is sent (or there was
369 * none), then we're done with the rm.
370 */
371 if (!rm->data.op_active || conn->c_xmit_data_sent) {
372 conn->c_xmit_rm = NULL;
373 conn->c_xmit_sg = 0;
374 conn->c_xmit_hdr_off = 0;
375 conn->c_xmit_data_off = 0;
376 conn->c_xmit_rdma_sent = 0;
377 conn->c_xmit_atomic_sent = 0;
378 conn->c_xmit_data_sent = 0;
379
380 rds_message_put(rm);
Andy Grover5c115592009-02-24 15:30:27 +0000381 }
382 }
383
Sowmini Varadhan443be0e2015-04-08 12:33:47 -0400384over_batch:
Andy Grover5c115592009-02-24 15:30:27 +0000385 if (conn->c_trans->xmit_complete)
386 conn->c_trans->xmit_complete(conn);
Zach Brown0f4b1c72010-06-04 14:41:41 -0700387 release_in_xmit(conn);
Andy Grover5c115592009-02-24 15:30:27 +0000388
Andy Grover2ad80992010-03-23 17:48:04 -0700389 /* Nuke any messages we decided not to retransmit. */
390 if (!list_empty(&to_be_dropped)) {
391 /* irqs on here, so we can put(), unlike above */
392 list_for_each_entry(rm, &to_be_dropped, m_conn_item)
393 rds_message_put(rm);
394 rds_send_remove_from_sock(&to_be_dropped, RDS_RDMA_DROPPED);
395 }
396
Andy Groverfcc54502010-03-29 17:08:49 -0700397 /*
Zach Brown0f4b1c72010-06-04 14:41:41 -0700398 * Other senders can queue a message after we last test the send queue
399 * but before we clear RDS_IN_XMIT. In that case they'd back off and
400 * not try and send their newly queued message. We need to check the
401 * send queue after having cleared RDS_IN_XMIT so that their message
402 * doesn't get stuck on the send queue.
Andy Groverfcc54502010-03-29 17:08:49 -0700403 *
404 * If the transport cannot continue (i.e ret != 0), then it must
405 * call us when more room is available, such as from the tx
406 * completion handler.
Sowmini Varadhan443be0e2015-04-08 12:33:47 -0400407 *
408 * We have an extra generation check here so that if someone manages
409 * to jump in after our release_in_xmit, we'll see that they have done
410 * some work and we will skip our goto
Andy Groverfcc54502010-03-29 17:08:49 -0700411 */
412 if (ret == 0) {
Chris Mason9e29db02010-04-15 16:38:14 -0400413 smp_mb();
Sowmini Varadhan443be0e2015-04-08 12:33:47 -0400414 if (!list_empty(&conn->c_send_queue) &&
415 send_gen == conn->c_send_gen) {
Andy Grover049ee3f2010-03-23 17:39:07 -0700416 rds_stats_inc(s_send_lock_queue_raced);
Zach Brown0f4b1c72010-06-04 14:41:41 -0700417 goto restart;
Andy Grover5c115592009-02-24 15:30:27 +0000418 }
Andy Grover5c115592009-02-24 15:30:27 +0000419 }
420out:
421 return ret;
422}
423
424static void rds_send_sndbuf_remove(struct rds_sock *rs, struct rds_message *rm)
425{
426 u32 len = be32_to_cpu(rm->m_inc.i_hdr.h_len);
427
428 assert_spin_locked(&rs->rs_lock);
429
430 BUG_ON(rs->rs_snd_bytes < len);
431 rs->rs_snd_bytes -= len;
432
433 if (rs->rs_snd_bytes == 0)
434 rds_stats_inc(s_send_queue_empty);
435}
436
437static inline int rds_send_is_acked(struct rds_message *rm, u64 ack,
438 is_acked_func is_acked)
439{
440 if (is_acked)
441 return is_acked(rm, ack);
442 return be64_to_cpu(rm->m_inc.i_hdr.h_sequence) <= ack;
443}
444
445/*
Andy Grover5c115592009-02-24 15:30:27 +0000446 * This is pretty similar to what happens below in the ACK
447 * handling code - except that we call here as soon as we get
448 * the IB send completion on the RDMA op and the accompanying
449 * message.
450 */
451void rds_rdma_send_complete(struct rds_message *rm, int status)
452{
453 struct rds_sock *rs = NULL;
Andy Groverf8b3aaf2010-03-01 14:11:53 -0800454 struct rm_rdma_op *ro;
Andy Grover5c115592009-02-24 15:30:27 +0000455 struct rds_notifier *notifier;
Andy Grover9de08642010-03-29 16:50:54 -0700456 unsigned long flags;
Andy Grover5c115592009-02-24 15:30:27 +0000457
Andy Grover9de08642010-03-29 16:50:54 -0700458 spin_lock_irqsave(&rm->m_rs_lock, flags);
Andy Grover5c115592009-02-24 15:30:27 +0000459
Andy Groverf8b3aaf2010-03-01 14:11:53 -0800460 ro = &rm->rdma;
Joe Perchesf64f9e72009-11-29 16:55:45 -0800461 if (test_bit(RDS_MSG_ON_SOCK, &rm->m_flags) &&
Andy Groverf8b3aaf2010-03-01 14:11:53 -0800462 ro->op_active && ro->op_notify && ro->op_notifier) {
463 notifier = ro->op_notifier;
Andy Grover5c115592009-02-24 15:30:27 +0000464 rs = rm->m_rs;
465 sock_hold(rds_rs_to_sk(rs));
466
467 notifier->n_status = status;
468 spin_lock(&rs->rs_lock);
469 list_add_tail(&notifier->n_list, &rs->rs_notify_queue);
470 spin_unlock(&rs->rs_lock);
471
Andy Groverf8b3aaf2010-03-01 14:11:53 -0800472 ro->op_notifier = NULL;
Andy Grover5c115592009-02-24 15:30:27 +0000473 }
474
Andy Grover9de08642010-03-29 16:50:54 -0700475 spin_unlock_irqrestore(&rm->m_rs_lock, flags);
Andy Grover5c115592009-02-24 15:30:27 +0000476
477 if (rs) {
478 rds_wake_sk_sleep(rs);
479 sock_put(rds_rs_to_sk(rs));
480 }
481}
Andy Grover616b7572009-08-21 12:28:32 +0000482EXPORT_SYMBOL_GPL(rds_rdma_send_complete);
Andy Grover5c115592009-02-24 15:30:27 +0000483
484/*
Andy Grover15133f62010-01-12 14:33:38 -0800485 * Just like above, except looks at atomic op
486 */
487void rds_atomic_send_complete(struct rds_message *rm, int status)
488{
489 struct rds_sock *rs = NULL;
490 struct rm_atomic_op *ao;
491 struct rds_notifier *notifier;
Andy Grovercf4b7382010-03-29 16:50:54 -0700492 unsigned long flags;
Andy Grover15133f62010-01-12 14:33:38 -0800493
Andy Grovercf4b7382010-03-29 16:50:54 -0700494 spin_lock_irqsave(&rm->m_rs_lock, flags);
Andy Grover15133f62010-01-12 14:33:38 -0800495
496 ao = &rm->atomic;
497 if (test_bit(RDS_MSG_ON_SOCK, &rm->m_flags)
498 && ao->op_active && ao->op_notify && ao->op_notifier) {
499 notifier = ao->op_notifier;
500 rs = rm->m_rs;
501 sock_hold(rds_rs_to_sk(rs));
502
503 notifier->n_status = status;
504 spin_lock(&rs->rs_lock);
505 list_add_tail(&notifier->n_list, &rs->rs_notify_queue);
506 spin_unlock(&rs->rs_lock);
507
508 ao->op_notifier = NULL;
509 }
510
Andy Grovercf4b7382010-03-29 16:50:54 -0700511 spin_unlock_irqrestore(&rm->m_rs_lock, flags);
Andy Grover15133f62010-01-12 14:33:38 -0800512
513 if (rs) {
514 rds_wake_sk_sleep(rs);
515 sock_put(rds_rs_to_sk(rs));
516 }
517}
518EXPORT_SYMBOL_GPL(rds_atomic_send_complete);
519
520/*
Andy Grover5c115592009-02-24 15:30:27 +0000521 * This is the same as rds_rdma_send_complete except we
522 * don't do any locking - we have all the ingredients (message,
523 * socket, socket lock) and can just move the notifier.
524 */
525static inline void
Andy Grover940786e2010-02-19 18:04:58 -0800526__rds_send_complete(struct rds_sock *rs, struct rds_message *rm, int status)
Andy Grover5c115592009-02-24 15:30:27 +0000527{
Andy Groverf8b3aaf2010-03-01 14:11:53 -0800528 struct rm_rdma_op *ro;
Andy Grover940786e2010-02-19 18:04:58 -0800529 struct rm_atomic_op *ao;
Andy Grover5c115592009-02-24 15:30:27 +0000530
Andy Groverf8b3aaf2010-03-01 14:11:53 -0800531 ro = &rm->rdma;
532 if (ro->op_active && ro->op_notify && ro->op_notifier) {
533 ro->op_notifier->n_status = status;
534 list_add_tail(&ro->op_notifier->n_list, &rs->rs_notify_queue);
535 ro->op_notifier = NULL;
Andy Grover5c115592009-02-24 15:30:27 +0000536 }
537
Andy Grover940786e2010-02-19 18:04:58 -0800538 ao = &rm->atomic;
539 if (ao->op_active && ao->op_notify && ao->op_notifier) {
540 ao->op_notifier->n_status = status;
541 list_add_tail(&ao->op_notifier->n_list, &rs->rs_notify_queue);
542 ao->op_notifier = NULL;
543 }
544
Andy Grover5c115592009-02-24 15:30:27 +0000545 /* No need to wake the app - caller does this */
546}
547
548/*
549 * This is called from the IB send completion when we detect
550 * a RDMA operation that failed with remote access error.
551 * So speed is not an issue here.
552 */
553struct rds_message *rds_send_get_message(struct rds_connection *conn,
Andy Groverf8b3aaf2010-03-01 14:11:53 -0800554 struct rm_rdma_op *op)
Andy Grover5c115592009-02-24 15:30:27 +0000555{
556 struct rds_message *rm, *tmp, *found = NULL;
557 unsigned long flags;
558
559 spin_lock_irqsave(&conn->c_lock, flags);
560
561 list_for_each_entry_safe(rm, tmp, &conn->c_retrans, m_conn_item) {
Andy Groverf8b3aaf2010-03-01 14:11:53 -0800562 if (&rm->rdma == op) {
Andy Grover5c115592009-02-24 15:30:27 +0000563 atomic_inc(&rm->m_refcount);
564 found = rm;
565 goto out;
566 }
567 }
568
569 list_for_each_entry_safe(rm, tmp, &conn->c_send_queue, m_conn_item) {
Andy Groverf8b3aaf2010-03-01 14:11:53 -0800570 if (&rm->rdma == op) {
Andy Grover5c115592009-02-24 15:30:27 +0000571 atomic_inc(&rm->m_refcount);
572 found = rm;
573 break;
574 }
575 }
576
577out:
578 spin_unlock_irqrestore(&conn->c_lock, flags);
579
580 return found;
581}
Andy Grover616b7572009-08-21 12:28:32 +0000582EXPORT_SYMBOL_GPL(rds_send_get_message);
Andy Grover5c115592009-02-24 15:30:27 +0000583
584/*
585 * This removes messages from the socket's list if they're on it. The list
586 * argument must be private to the caller, we must be able to modify it
587 * without locks. The messages must have a reference held for their
588 * position on the list. This function will drop that reference after
589 * removing the messages from the 'messages' list regardless of if it found
590 * the messages on the socket list or not.
591 */
stephen hemmingerff51bf82010-10-19 08:08:33 +0000592static void rds_send_remove_from_sock(struct list_head *messages, int status)
Andy Grover5c115592009-02-24 15:30:27 +0000593{
Andy Grover561c7df2010-03-11 13:50:06 +0000594 unsigned long flags;
Andy Grover5c115592009-02-24 15:30:27 +0000595 struct rds_sock *rs = NULL;
596 struct rds_message *rm;
597
Andy Grover5c115592009-02-24 15:30:27 +0000598 while (!list_empty(messages)) {
Andy Grover561c7df2010-03-11 13:50:06 +0000599 int was_on_sock = 0;
600
Andy Grover5c115592009-02-24 15:30:27 +0000601 rm = list_entry(messages->next, struct rds_message,
602 m_conn_item);
603 list_del_init(&rm->m_conn_item);
604
605 /*
606 * If we see this flag cleared then we're *sure* that someone
607 * else beat us to removing it from the sock. If we race
608 * with their flag update we'll get the lock and then really
609 * see that the flag has been cleared.
610 *
611 * The message spinlock makes sure nobody clears rm->m_rs
612 * while we're messing with it. It does not prevent the
613 * message from being removed from the socket, though.
614 */
Andy Grover561c7df2010-03-11 13:50:06 +0000615 spin_lock_irqsave(&rm->m_rs_lock, flags);
Andy Grover5c115592009-02-24 15:30:27 +0000616 if (!test_bit(RDS_MSG_ON_SOCK, &rm->m_flags))
617 goto unlock_and_drop;
618
619 if (rs != rm->m_rs) {
620 if (rs) {
Andy Grover5c115592009-02-24 15:30:27 +0000621 rds_wake_sk_sleep(rs);
622 sock_put(rds_rs_to_sk(rs));
623 }
624 rs = rm->m_rs;
Herton R. Krzesinski593cbb32014-10-01 18:49:54 -0300625 if (rs)
626 sock_hold(rds_rs_to_sk(rs));
Andy Grover5c115592009-02-24 15:30:27 +0000627 }
Herton R. Krzesinski593cbb32014-10-01 18:49:54 -0300628 if (!rs)
629 goto unlock_and_drop;
Tina Yang048c15e2010-03-11 13:50:00 +0000630 spin_lock(&rs->rs_lock);
Andy Grover5c115592009-02-24 15:30:27 +0000631
632 if (test_and_clear_bit(RDS_MSG_ON_SOCK, &rm->m_flags)) {
Andy Groverf8b3aaf2010-03-01 14:11:53 -0800633 struct rm_rdma_op *ro = &rm->rdma;
Andy Grover5c115592009-02-24 15:30:27 +0000634 struct rds_notifier *notifier;
635
636 list_del_init(&rm->m_sock_item);
637 rds_send_sndbuf_remove(rs, rm);
638
Andy Groverf8b3aaf2010-03-01 14:11:53 -0800639 if (ro->op_active && ro->op_notifier &&
640 (ro->op_notify || (ro->op_recverr && status))) {
641 notifier = ro->op_notifier;
Andy Grover5c115592009-02-24 15:30:27 +0000642 list_add_tail(&notifier->n_list,
643 &rs->rs_notify_queue);
644 if (!notifier->n_status)
645 notifier->n_status = status;
Andy Groverf8b3aaf2010-03-01 14:11:53 -0800646 rm->rdma.op_notifier = NULL;
Andy Grover5c115592009-02-24 15:30:27 +0000647 }
Andy Grover561c7df2010-03-11 13:50:06 +0000648 was_on_sock = 1;
Andy Grover5c115592009-02-24 15:30:27 +0000649 rm->m_rs = NULL;
650 }
Tina Yang048c15e2010-03-11 13:50:00 +0000651 spin_unlock(&rs->rs_lock);
Andy Grover5c115592009-02-24 15:30:27 +0000652
653unlock_and_drop:
Andy Grover561c7df2010-03-11 13:50:06 +0000654 spin_unlock_irqrestore(&rm->m_rs_lock, flags);
Andy Grover5c115592009-02-24 15:30:27 +0000655 rds_message_put(rm);
Andy Grover561c7df2010-03-11 13:50:06 +0000656 if (was_on_sock)
657 rds_message_put(rm);
Andy Grover5c115592009-02-24 15:30:27 +0000658 }
659
660 if (rs) {
Andy Grover5c115592009-02-24 15:30:27 +0000661 rds_wake_sk_sleep(rs);
662 sock_put(rds_rs_to_sk(rs));
663 }
Andy Grover5c115592009-02-24 15:30:27 +0000664}
665
666/*
667 * Transports call here when they've determined that the receiver queued
668 * messages up to, and including, the given sequence number. Messages are
669 * moved to the retrans queue when rds_send_xmit picks them off the send
670 * queue. This means that in the TCP case, the message may not have been
671 * assigned the m_ack_seq yet - but that's fine as long as tcp_is_acked
672 * checks the RDS_MSG_HAS_ACK_SEQ bit.
Andy Grover5c115592009-02-24 15:30:27 +0000673 */
674void rds_send_drop_acked(struct rds_connection *conn, u64 ack,
675 is_acked_func is_acked)
676{
677 struct rds_message *rm, *tmp;
678 unsigned long flags;
679 LIST_HEAD(list);
680
681 spin_lock_irqsave(&conn->c_lock, flags);
682
683 list_for_each_entry_safe(rm, tmp, &conn->c_retrans, m_conn_item) {
684 if (!rds_send_is_acked(rm, ack, is_acked))
685 break;
686
687 list_move(&rm->m_conn_item, &list);
688 clear_bit(RDS_MSG_ON_CONN, &rm->m_flags);
689 }
690
691 /* order flag updates with spin locks */
692 if (!list_empty(&list))
Peter Zijlstra4e857c52014-03-17 18:06:10 +0100693 smp_mb__after_atomic();
Andy Grover5c115592009-02-24 15:30:27 +0000694
695 spin_unlock_irqrestore(&conn->c_lock, flags);
696
697 /* now remove the messages from the sock list as needed */
698 rds_send_remove_from_sock(&list, RDS_RDMA_SUCCESS);
699}
Andy Grover616b7572009-08-21 12:28:32 +0000700EXPORT_SYMBOL_GPL(rds_send_drop_acked);
Andy Grover5c115592009-02-24 15:30:27 +0000701
702void rds_send_drop_to(struct rds_sock *rs, struct sockaddr_in *dest)
703{
704 struct rds_message *rm, *tmp;
705 struct rds_connection *conn;
Andy Grover7c82eaf2010-02-19 18:01:41 -0800706 unsigned long flags;
Andy Grover5c115592009-02-24 15:30:27 +0000707 LIST_HEAD(list);
Andy Grover5c115592009-02-24 15:30:27 +0000708
709 /* get all the messages we're dropping under the rs lock */
710 spin_lock_irqsave(&rs->rs_lock, flags);
711
712 list_for_each_entry_safe(rm, tmp, &rs->rs_send_queue, m_sock_item) {
713 if (dest && (dest->sin_addr.s_addr != rm->m_daddr ||
714 dest->sin_port != rm->m_inc.i_hdr.h_dport))
715 continue;
716
Andy Grover5c115592009-02-24 15:30:27 +0000717 list_move(&rm->m_sock_item, &list);
718 rds_send_sndbuf_remove(rs, rm);
719 clear_bit(RDS_MSG_ON_SOCK, &rm->m_flags);
Andy Grover5c115592009-02-24 15:30:27 +0000720 }
721
722 /* order flag updates with the rs lock */
Peter Zijlstra4e857c52014-03-17 18:06:10 +0100723 smp_mb__after_atomic();
Andy Grover5c115592009-02-24 15:30:27 +0000724
725 spin_unlock_irqrestore(&rs->rs_lock, flags);
726
Andy Grover7c82eaf2010-02-19 18:01:41 -0800727 if (list_empty(&list))
728 return;
Andy Grover5c115592009-02-24 15:30:27 +0000729
Andy Grover7c82eaf2010-02-19 18:01:41 -0800730 /* Remove the messages from the conn */
Andy Grover5c115592009-02-24 15:30:27 +0000731 list_for_each_entry(rm, &list, m_sock_item) {
Andy Grover7c82eaf2010-02-19 18:01:41 -0800732
733 conn = rm->m_inc.i_conn;
Andy Grover7c82eaf2010-02-19 18:01:41 -0800734
Andy Grover9de08642010-03-29 16:50:54 -0700735 spin_lock_irqsave(&conn->c_lock, flags);
Andy Grover7c82eaf2010-02-19 18:01:41 -0800736 /*
737 * Maybe someone else beat us to removing rm from the conn.
738 * If we race with their flag update we'll get the lock and
739 * then really see that the flag has been cleared.
740 */
741 if (!test_and_clear_bit(RDS_MSG_ON_CONN, &rm->m_flags)) {
742 spin_unlock_irqrestore(&conn->c_lock, flags);
Herton R. Krzesinski593cbb32014-10-01 18:49:54 -0300743 spin_lock_irqsave(&rm->m_rs_lock, flags);
744 rm->m_rs = NULL;
745 spin_unlock_irqrestore(&rm->m_rs_lock, flags);
Andy Grover7c82eaf2010-02-19 18:01:41 -0800746 continue;
747 }
Andy Grover9de08642010-03-29 16:50:54 -0700748 list_del_init(&rm->m_conn_item);
749 spin_unlock_irqrestore(&conn->c_lock, flags);
Andy Grover7c82eaf2010-02-19 18:01:41 -0800750
751 /*
752 * Couldn't grab m_rs_lock in top loop (lock ordering),
753 * but we can now.
754 */
Andy Grover9de08642010-03-29 16:50:54 -0700755 spin_lock_irqsave(&rm->m_rs_lock, flags);
Andy Grover7c82eaf2010-02-19 18:01:41 -0800756
Tina Yang550a8002010-03-11 13:50:03 +0000757 spin_lock(&rs->rs_lock);
Andy Grover940786e2010-02-19 18:04:58 -0800758 __rds_send_complete(rs, rm, RDS_RDMA_CANCELED);
Tina Yang550a8002010-03-11 13:50:03 +0000759 spin_unlock(&rs->rs_lock);
Andy Grover7c82eaf2010-02-19 18:01:41 -0800760
Andy Grover5c115592009-02-24 15:30:27 +0000761 rm->m_rs = NULL;
Andy Grover9de08642010-03-29 16:50:54 -0700762 spin_unlock_irqrestore(&rm->m_rs_lock, flags);
Andy Grover5c115592009-02-24 15:30:27 +0000763
Andy Grover7c82eaf2010-02-19 18:01:41 -0800764 rds_message_put(rm);
Andy Grover5c115592009-02-24 15:30:27 +0000765 }
766
Andy Grover7c82eaf2010-02-19 18:01:41 -0800767 rds_wake_sk_sleep(rs);
Tina Yang550a8002010-03-11 13:50:03 +0000768
Andy Grover5c115592009-02-24 15:30:27 +0000769 while (!list_empty(&list)) {
770 rm = list_entry(list.next, struct rds_message, m_sock_item);
771 list_del_init(&rm->m_sock_item);
772
773 rds_message_wait(rm);
774 rds_message_put(rm);
775 }
776}
777
778/*
779 * we only want this to fire once so we use the callers 'queued'. It's
780 * possible that another thread can race with us and remove the
781 * message from the flow with RDS_CANCEL_SENT_TO.
782 */
783static int rds_send_queue_rm(struct rds_sock *rs, struct rds_connection *conn,
784 struct rds_message *rm, __be16 sport,
785 __be16 dport, int *queued)
786{
787 unsigned long flags;
788 u32 len;
789
790 if (*queued)
791 goto out;
792
793 len = be32_to_cpu(rm->m_inc.i_hdr.h_len);
794
795 /* this is the only place which holds both the socket's rs_lock
796 * and the connection's c_lock */
797 spin_lock_irqsave(&rs->rs_lock, flags);
798
799 /*
800 * If there is a little space in sndbuf, we don't queue anything,
801 * and userspace gets -EAGAIN. But poll() indicates there's send
802 * room. This can lead to bad behavior (spinning) if snd_bytes isn't
803 * freed up by incoming acks. So we check the *old* value of
804 * rs_snd_bytes here to allow the last msg to exceed the buffer,
805 * and poll() now knows no more data can be sent.
806 */
807 if (rs->rs_snd_bytes < rds_sk_sndbuf(rs)) {
808 rs->rs_snd_bytes += len;
809
810 /* let recv side know we are close to send space exhaustion.
811 * This is probably not the optimal way to do it, as this
812 * means we set the flag on *all* messages as soon as our
813 * throughput hits a certain threshold.
814 */
815 if (rs->rs_snd_bytes >= rds_sk_sndbuf(rs) / 2)
816 __set_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags);
817
818 list_add_tail(&rm->m_sock_item, &rs->rs_send_queue);
819 set_bit(RDS_MSG_ON_SOCK, &rm->m_flags);
820 rds_message_addref(rm);
821 rm->m_rs = rs;
822
823 /* The code ordering is a little weird, but we're
824 trying to minimize the time we hold c_lock */
825 rds_message_populate_header(&rm->m_inc.i_hdr, sport, dport, 0);
826 rm->m_inc.i_conn = conn;
827 rds_message_addref(rm);
828
829 spin_lock(&conn->c_lock);
830 rm->m_inc.i_hdr.h_sequence = cpu_to_be64(conn->c_next_tx_seq++);
831 list_add_tail(&rm->m_conn_item, &conn->c_send_queue);
832 set_bit(RDS_MSG_ON_CONN, &rm->m_flags);
833 spin_unlock(&conn->c_lock);
834
835 rdsdebug("queued msg %p len %d, rs %p bytes %d seq %llu\n",
836 rm, len, rs, rs->rs_snd_bytes,
837 (unsigned long long)be64_to_cpu(rm->m_inc.i_hdr.h_sequence));
838
839 *queued = 1;
840 }
841
842 spin_unlock_irqrestore(&rs->rs_lock, flags);
843out:
844 return *queued;
845}
846
Andy Groverfc445082010-01-12 12:56:06 -0800847/*
848 * rds_message is getting to be quite complicated, and we'd like to allocate
849 * it all in one go. This figures out how big it needs to be up front.
850 */
851static int rds_rm_size(struct msghdr *msg, int data_len)
852{
Andy Groverff87e972010-01-12 14:13:15 -0800853 struct cmsghdr *cmsg;
Andy Groverfc445082010-01-12 12:56:06 -0800854 int size = 0;
Andy Groveraa0a4ef2010-04-13 12:00:35 -0700855 int cmsg_groups = 0;
Andy Groverff87e972010-01-12 14:13:15 -0800856 int retval;
Andy Groverfc445082010-01-12 12:56:06 -0800857
Gu Zhengf95b4142014-12-11 11:22:04 +0800858 for_each_cmsghdr(cmsg, msg) {
Andy Groverff87e972010-01-12 14:13:15 -0800859 if (!CMSG_OK(msg, cmsg))
860 return -EINVAL;
861
862 if (cmsg->cmsg_level != SOL_RDS)
863 continue;
864
865 switch (cmsg->cmsg_type) {
866 case RDS_CMSG_RDMA_ARGS:
Andy Groveraa0a4ef2010-04-13 12:00:35 -0700867 cmsg_groups |= 1;
Andy Groverff87e972010-01-12 14:13:15 -0800868 retval = rds_rdma_extra_size(CMSG_DATA(cmsg));
869 if (retval < 0)
870 return retval;
871 size += retval;
Andy Groveraa0a4ef2010-04-13 12:00:35 -0700872
Andy Groverff87e972010-01-12 14:13:15 -0800873 break;
874
875 case RDS_CMSG_RDMA_DEST:
876 case RDS_CMSG_RDMA_MAP:
Andy Groveraa0a4ef2010-04-13 12:00:35 -0700877 cmsg_groups |= 2;
Andy Groverff87e972010-01-12 14:13:15 -0800878 /* these are valid but do no add any size */
879 break;
880
Andy Grover15133f62010-01-12 14:33:38 -0800881 case RDS_CMSG_ATOMIC_CSWP:
882 case RDS_CMSG_ATOMIC_FADD:
Andy Grover20c72bd2010-08-25 05:51:28 -0700883 case RDS_CMSG_MASKED_ATOMIC_CSWP:
884 case RDS_CMSG_MASKED_ATOMIC_FADD:
Andy Groveraa0a4ef2010-04-13 12:00:35 -0700885 cmsg_groups |= 1;
Andy Grover15133f62010-01-12 14:33:38 -0800886 size += sizeof(struct scatterlist);
887 break;
888
Andy Groverff87e972010-01-12 14:13:15 -0800889 default:
890 return -EINVAL;
891 }
892
893 }
894
895 size += ceil(data_len, PAGE_SIZE) * sizeof(struct scatterlist);
Andy Groverfc445082010-01-12 12:56:06 -0800896
Andy Groveraa0a4ef2010-04-13 12:00:35 -0700897 /* Ensure (DEST, MAP) are never used with (ARGS, ATOMIC) */
898 if (cmsg_groups == 3)
899 return -EINVAL;
900
Andy Groverfc445082010-01-12 12:56:06 -0800901 return size;
902}
903
Andy Grover5c115592009-02-24 15:30:27 +0000904static int rds_cmsg_send(struct rds_sock *rs, struct rds_message *rm,
905 struct msghdr *msg, int *allocated_mr)
906{
907 struct cmsghdr *cmsg;
908 int ret = 0;
909
Gu Zhengf95b4142014-12-11 11:22:04 +0800910 for_each_cmsghdr(cmsg, msg) {
Andy Grover5c115592009-02-24 15:30:27 +0000911 if (!CMSG_OK(msg, cmsg))
912 return -EINVAL;
913
914 if (cmsg->cmsg_level != SOL_RDS)
915 continue;
916
917 /* As a side effect, RDMA_DEST and RDMA_MAP will set
Andy Grover15133f62010-01-12 14:33:38 -0800918 * rm->rdma.m_rdma_cookie and rm->rdma.m_rdma_mr.
Andy Grover5c115592009-02-24 15:30:27 +0000919 */
920 switch (cmsg->cmsg_type) {
921 case RDS_CMSG_RDMA_ARGS:
922 ret = rds_cmsg_rdma_args(rs, rm, cmsg);
923 break;
924
925 case RDS_CMSG_RDMA_DEST:
926 ret = rds_cmsg_rdma_dest(rs, rm, cmsg);
927 break;
928
929 case RDS_CMSG_RDMA_MAP:
930 ret = rds_cmsg_rdma_map(rs, rm, cmsg);
931 if (!ret)
932 *allocated_mr = 1;
933 break;
Andy Grover15133f62010-01-12 14:33:38 -0800934 case RDS_CMSG_ATOMIC_CSWP:
935 case RDS_CMSG_ATOMIC_FADD:
Andy Grover20c72bd2010-08-25 05:51:28 -0700936 case RDS_CMSG_MASKED_ATOMIC_CSWP:
937 case RDS_CMSG_MASKED_ATOMIC_FADD:
Andy Grover15133f62010-01-12 14:33:38 -0800938 ret = rds_cmsg_atomic(rs, rm, cmsg);
939 break;
Andy Grover5c115592009-02-24 15:30:27 +0000940
941 default:
942 return -EINVAL;
943 }
944
945 if (ret)
946 break;
947 }
948
949 return ret;
950}
951
Ying Xue1b784142015-03-02 15:37:48 +0800952int rds_sendmsg(struct socket *sock, struct msghdr *msg, size_t payload_len)
Andy Grover5c115592009-02-24 15:30:27 +0000953{
954 struct sock *sk = sock->sk;
955 struct rds_sock *rs = rds_sk_to_rs(sk);
Steffen Hurrle342dfc32014-01-17 22:53:15 +0100956 DECLARE_SOCKADDR(struct sockaddr_in *, usin, msg->msg_name);
Andy Grover5c115592009-02-24 15:30:27 +0000957 __be32 daddr;
958 __be16 dport;
959 struct rds_message *rm = NULL;
960 struct rds_connection *conn;
961 int ret = 0;
962 int queued = 0, allocated_mr = 0;
963 int nonblock = msg->msg_flags & MSG_DONTWAIT;
Andy Grover1123fd72010-03-11 13:49:56 +0000964 long timeo = sock_sndtimeo(sk, nonblock);
Andy Grover5c115592009-02-24 15:30:27 +0000965
966 /* Mirror Linux UDP mirror of BSD error message compatibility */
967 /* XXX: Perhaps MSG_MORE someday */
968 if (msg->msg_flags & ~(MSG_DONTWAIT | MSG_CMSG_COMPAT)) {
Andy Grover5c115592009-02-24 15:30:27 +0000969 ret = -EOPNOTSUPP;
970 goto out;
971 }
972
973 if (msg->msg_namelen) {
974 /* XXX fail non-unicast destination IPs? */
975 if (msg->msg_namelen < sizeof(*usin) || usin->sin_family != AF_INET) {
976 ret = -EINVAL;
977 goto out;
978 }
979 daddr = usin->sin_addr.s_addr;
980 dport = usin->sin_port;
981 } else {
982 /* We only care about consistency with ->connect() */
983 lock_sock(sk);
984 daddr = rs->rs_conn_addr;
985 dport = rs->rs_conn_port;
986 release_sock(sk);
987 }
988
989 /* racing with another thread binding seems ok here */
990 if (daddr == 0 || rs->rs_bound_addr == 0) {
991 ret = -ENOTCONN; /* XXX not a great errno */
992 goto out;
993 }
994
Andy Groverfc445082010-01-12 12:56:06 -0800995 /* size of rm including all sgs */
996 ret = rds_rm_size(msg, payload_len);
997 if (ret < 0)
998 goto out;
999
1000 rm = rds_message_alloc(ret, GFP_KERNEL);
1001 if (!rm) {
1002 ret = -ENOMEM;
Andy Grover5c115592009-02-24 15:30:27 +00001003 goto out;
1004 }
1005
Andy Grover372cd7d2010-02-03 19:40:32 -08001006 /* Attach data to the rm */
1007 if (payload_len) {
1008 rm->data.op_sg = rds_message_alloc_sgs(rm, ceil(payload_len, PAGE_SIZE));
Andy Groverd139ff02010-10-28 15:40:59 +00001009 if (!rm->data.op_sg) {
1010 ret = -ENOMEM;
1011 goto out;
1012 }
Al Viroc0371da2014-11-24 10:42:55 -05001013 ret = rds_message_copy_from_user(rm, &msg->msg_iter);
Andy Grover372cd7d2010-02-03 19:40:32 -08001014 if (ret)
1015 goto out;
1016 }
1017 rm->data.op_active = 1;
Andy Groverfc445082010-01-12 12:56:06 -08001018
Andy Grover5c115592009-02-24 15:30:27 +00001019 rm->m_daddr = daddr;
1020
Andy Grover5c115592009-02-24 15:30:27 +00001021 /* rds_conn_create has a spinlock that runs with IRQ off.
1022 * Caching the conn in the socket helps a lot. */
1023 if (rs->rs_conn && rs->rs_conn->c_faddr == daddr)
1024 conn = rs->rs_conn;
1025 else {
1026 conn = rds_conn_create_outgoing(rs->rs_bound_addr, daddr,
1027 rs->rs_transport,
1028 sock->sk->sk_allocation);
1029 if (IS_ERR(conn)) {
1030 ret = PTR_ERR(conn);
1031 goto out;
1032 }
1033 rs->rs_conn = conn;
1034 }
1035
Andy Grover49f69692009-04-09 14:09:41 +00001036 /* Parse any control messages the user may have included. */
1037 ret = rds_cmsg_send(rs, rm, msg, &allocated_mr);
1038 if (ret)
1039 goto out;
1040
Andy Grover2c3a5f92010-03-01 16:10:40 -08001041 if (rm->rdma.op_active && !conn->c_trans->xmit_rdma) {
Manuel Zerpiescb0a6052011-06-16 02:09:57 +00001042 printk_ratelimited(KERN_NOTICE "rdma_op %p conn xmit_rdma %p\n",
Andy Groverf8b3aaf2010-03-01 14:11:53 -08001043 &rm->rdma, conn->c_trans->xmit_rdma);
Andy Grover15133f62010-01-12 14:33:38 -08001044 ret = -EOPNOTSUPP;
1045 goto out;
1046 }
1047
1048 if (rm->atomic.op_active && !conn->c_trans->xmit_atomic) {
Manuel Zerpiescb0a6052011-06-16 02:09:57 +00001049 printk_ratelimited(KERN_NOTICE "atomic_op %p conn xmit_atomic %p\n",
Andy Grover15133f62010-01-12 14:33:38 -08001050 &rm->atomic, conn->c_trans->xmit_atomic);
Andy Grover5c115592009-02-24 15:30:27 +00001051 ret = -EOPNOTSUPP;
1052 goto out;
1053 }
1054
Zach Brownf3c68082010-05-24 13:14:36 -07001055 rds_conn_connect_if_down(conn);
Andy Grover5c115592009-02-24 15:30:27 +00001056
1057 ret = rds_cong_wait(conn->c_fcong, dport, nonblock, rs);
Andy Groverb98ba522010-03-11 13:50:04 +00001058 if (ret) {
1059 rs->rs_seen_congestion = 1;
Andy Grover5c115592009-02-24 15:30:27 +00001060 goto out;
Andy Groverb98ba522010-03-11 13:50:04 +00001061 }
Andy Grover5c115592009-02-24 15:30:27 +00001062
1063 while (!rds_send_queue_rm(rs, conn, rm, rs->rs_bound_port,
1064 dport, &queued)) {
1065 rds_stats_inc(s_send_queue_full);
1066 /* XXX make sure this is reasonable */
1067 if (payload_len > rds_sk_sndbuf(rs)) {
1068 ret = -EMSGSIZE;
1069 goto out;
1070 }
1071 if (nonblock) {
1072 ret = -EAGAIN;
1073 goto out;
1074 }
1075
Eric Dumazetaa395142010-04-20 13:03:51 +00001076 timeo = wait_event_interruptible_timeout(*sk_sleep(sk),
Andy Grover5c115592009-02-24 15:30:27 +00001077 rds_send_queue_rm(rs, conn, rm,
1078 rs->rs_bound_port,
1079 dport,
1080 &queued),
1081 timeo);
1082 rdsdebug("sendmsg woke queued %d timeo %ld\n", queued, timeo);
1083 if (timeo > 0 || timeo == MAX_SCHEDULE_TIMEOUT)
1084 continue;
1085
1086 ret = timeo;
1087 if (ret == 0)
1088 ret = -ETIMEDOUT;
1089 goto out;
1090 }
1091
1092 /*
1093 * By now we've committed to the send. We reuse rds_send_worker()
1094 * to retry sends in the rds thread if the transport asks us to.
1095 */
1096 rds_stats_inc(s_send_queued);
1097
1098 if (!test_bit(RDS_LL_SEND_FULL, &conn->c_flags))
Andy Grovera7d3a282010-03-29 16:20:18 -07001099 rds_send_xmit(conn);
Andy Grover5c115592009-02-24 15:30:27 +00001100
1101 rds_message_put(rm);
1102 return payload_len;
1103
1104out:
1105 /* If the user included a RDMA_MAP cmsg, we allocated a MR on the fly.
1106 * If the sendmsg goes through, we keep the MR. If it fails with EAGAIN
1107 * or in any other way, we need to destroy the MR again */
1108 if (allocated_mr)
1109 rds_rdma_unuse(rs, rds_rdma_cookie_key(rm->m_rdma_cookie), 1);
1110
1111 if (rm)
1112 rds_message_put(rm);
1113 return ret;
1114}
1115
1116/*
1117 * Reply to a ping packet.
1118 */
1119int
1120rds_send_pong(struct rds_connection *conn, __be16 dport)
1121{
1122 struct rds_message *rm;
1123 unsigned long flags;
1124 int ret = 0;
1125
1126 rm = rds_message_alloc(0, GFP_ATOMIC);
Andy Grover8690bfa2010-01-12 11:56:44 -08001127 if (!rm) {
Andy Grover5c115592009-02-24 15:30:27 +00001128 ret = -ENOMEM;
1129 goto out;
1130 }
1131
1132 rm->m_daddr = conn->c_faddr;
Andy Groveracfcd4d2010-03-31 18:56:25 -07001133 rm->data.op_active = 1;
Andy Grover5c115592009-02-24 15:30:27 +00001134
Zach Brownf3c68082010-05-24 13:14:36 -07001135 rds_conn_connect_if_down(conn);
Andy Grover5c115592009-02-24 15:30:27 +00001136
1137 ret = rds_cong_wait(conn->c_fcong, dport, 1, NULL);
1138 if (ret)
1139 goto out;
1140
1141 spin_lock_irqsave(&conn->c_lock, flags);
1142 list_add_tail(&rm->m_conn_item, &conn->c_send_queue);
1143 set_bit(RDS_MSG_ON_CONN, &rm->m_flags);
1144 rds_message_addref(rm);
1145 rm->m_inc.i_conn = conn;
1146
1147 rds_message_populate_header(&rm->m_inc.i_hdr, 0, dport,
1148 conn->c_next_tx_seq);
1149 conn->c_next_tx_seq++;
1150 spin_unlock_irqrestore(&conn->c_lock, flags);
1151
1152 rds_stats_inc(s_send_queued);
1153 rds_stats_inc(s_send_pong);
1154
Andy Groveracfcd4d2010-03-31 18:56:25 -07001155 if (!test_bit(RDS_LL_SEND_FULL, &conn->c_flags))
jeff.liu5175a5e2012-10-08 18:57:27 +00001156 queue_delayed_work(rds_wq, &conn->c_send_w, 0);
Andy Groveracfcd4d2010-03-31 18:56:25 -07001157
Andy Grover5c115592009-02-24 15:30:27 +00001158 rds_message_put(rm);
1159 return 0;
1160
1161out:
1162 if (rm)
1163 rds_message_put(rm);
1164 return ret;
1165}