blob: b0acd45a9d42ae7abc86f3a028aabffe9d6576f4 [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>
Santosh Shilimkar4bebdd72015-09-10 11:57:14 -070041#include <linux/sizes.h>
Andy Grover5c115592009-02-24 15:30:27 +000042
43#include "rds.h"
Andy Grover5c115592009-02-24 15:30:27 +000044
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 */
Santosh Shilimkar4bebdd72015-09-10 11:57:14 -070055static int send_batch_count = SZ_1K;
Andy Grover5c115592009-02-24 15:30:27 +000056module_param(send_batch_count, int, 0444);
57MODULE_PARM_DESC(send_batch_count, " batch factor when working the send queue");
58
stephen hemmingerff51bf82010-10-19 08:08:33 +000059static void rds_send_remove_from_sock(struct list_head *messages, int status);
60
Andy Grover5c115592009-02-24 15:30:27 +000061/*
Zach Brown0f4b1c72010-06-04 14:41:41 -070062 * Reset the send state. Callers must ensure that this doesn't race with
63 * rds_send_xmit().
Andy Grover5c115592009-02-24 15:30:27 +000064 */
65void rds_send_reset(struct rds_connection *conn)
66{
67 struct rds_message *rm, *tmp;
68 unsigned long flags;
69
70 if (conn->c_xmit_rm) {
Chris Mason7e3f2952010-05-11 15:11:11 -070071 rm = conn->c_xmit_rm;
72 conn->c_xmit_rm = NULL;
Andy Grover5c115592009-02-24 15:30:27 +000073 /* 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 */
Chris Mason7e3f2952010-05-11 15:11:11 -070077 rds_message_unmapped(rm);
Chris Mason7e3f2952010-05-11 15:11:11 -070078 rds_message_put(rm);
Andy Grover5c115592009-02-24 15:30:27 +000079 }
Chris Mason7e3f2952010-05-11 15:11:11 -070080
Andy Grover5c115592009-02-24 15:30:27 +000081 conn->c_xmit_sg = 0;
82 conn->c_xmit_hdr_off = 0;
83 conn->c_xmit_data_off = 0;
Andy Grover15133f62010-01-12 14:33:38 -080084 conn->c_xmit_atomic_sent = 0;
Andy Grover5b2366b2010-02-03 19:36:44 -080085 conn->c_xmit_rdma_sent = 0;
86 conn->c_xmit_data_sent = 0;
Andy Grover5c115592009-02-24 15:30:27 +000087
88 conn->c_map_queued = 0;
89
90 conn->c_unacked_packets = rds_sysctl_max_unacked_packets;
91 conn->c_unacked_bytes = rds_sysctl_max_unacked_bytes;
92
93 /* Mark messages as retransmissions, and move them to the send q */
94 spin_lock_irqsave(&conn->c_lock, flags);
95 list_for_each_entry_safe(rm, tmp, &conn->c_retrans, m_conn_item) {
96 set_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags);
97 set_bit(RDS_MSG_RETRANSMITTED, &rm->m_flags);
98 }
99 list_splice_init(&conn->c_retrans, &conn->c_send_queue);
100 spin_unlock_irqrestore(&conn->c_lock, flags);
101}
102
Zach Brown0f4b1c72010-06-04 14:41:41 -0700103static int acquire_in_xmit(struct rds_connection *conn)
104{
105 return test_and_set_bit(RDS_IN_XMIT, &conn->c_flags) == 0;
106}
107
108static void release_in_xmit(struct rds_connection *conn)
109{
110 clear_bit(RDS_IN_XMIT, &conn->c_flags);
Peter Zijlstra4e857c52014-03-17 18:06:10 +0100111 smp_mb__after_atomic();
Zach Brown0f4b1c72010-06-04 14:41:41 -0700112 /*
113 * We don't use wait_on_bit()/wake_up_bit() because our waking is in a
114 * hot path and finding waiters is very rare. We don't want to walk
115 * the system-wide hashed waitqueue buckets in the fast path only to
116 * almost never find waiters.
117 */
118 if (waitqueue_active(&conn->c_waitq))
119 wake_up_all(&conn->c_waitq);
120}
121
Andy Grover5c115592009-02-24 15:30:27 +0000122/*
Lucas De Marchi25985ed2011-03-30 22:57:33 -0300123 * We're making the conscious trade-off here to only send one message
Andy Grover5c115592009-02-24 15:30:27 +0000124 * down the connection at a time.
125 * Pro:
126 * - tx queueing is a simple fifo list
127 * - reassembly is optional and easily done by transports per conn
128 * - no per flow rx lookup at all, straight to the socket
129 * - less per-frag memory and wire overhead
130 * Con:
131 * - queued acks can be delayed behind large messages
132 * Depends:
133 * - small message latency is higher behind queued large messages
134 * - large message latency isn't starved by intervening small sends
135 */
136int rds_send_xmit(struct rds_connection *conn)
137{
138 struct rds_message *rm;
139 unsigned long flags;
140 unsigned int tmp;
Andy Grover5c115592009-02-24 15:30:27 +0000141 struct scatterlist *sg;
142 int ret = 0;
Andy Grover5c115592009-02-24 15:30:27 +0000143 LIST_HEAD(to_be_dropped);
Sowmini Varadhan443be0e2015-04-08 12:33:47 -0400144 int batch_count;
145 unsigned long send_gen = 0;
Andy Grover5c115592009-02-24 15:30:27 +0000146
Andy Groverfcc54502010-03-29 17:08:49 -0700147restart:
Sowmini Varadhan443be0e2015-04-08 12:33:47 -0400148 batch_count = 0;
Andy Grover049ee3f2010-03-23 17:39:07 -0700149
Andy Grover5c115592009-02-24 15:30:27 +0000150 /*
151 * sendmsg calls here after having queued its message on the send
152 * queue. We only have one task feeding the connection at a time. If
153 * another thread is already feeding the queue then we back off. This
154 * avoids blocking the caller and trading per-connection data between
155 * caches per message.
Andy Grover5c115592009-02-24 15:30:27 +0000156 */
Zach Brown0f4b1c72010-06-04 14:41:41 -0700157 if (!acquire_in_xmit(conn)) {
Andy Grover049ee3f2010-03-23 17:39:07 -0700158 rds_stats_inc(s_send_lock_contention);
Andy Grover5c115592009-02-24 15:30:27 +0000159 ret = -ENOMEM;
160 goto out;
161 }
Zach Brown0f4b1c72010-06-04 14:41:41 -0700162
163 /*
Sowmini Varadhan443be0e2015-04-08 12:33:47 -0400164 * we record the send generation after doing the xmit acquire.
165 * if someone else manages to jump in and do some work, we'll use
166 * this to avoid a goto restart farther down.
167 *
168 * The acquire_in_xmit() check above ensures that only one
169 * caller can increment c_send_gen at any time.
170 */
171 conn->c_send_gen++;
172 send_gen = conn->c_send_gen;
173
174 /*
Zach Brown0f4b1c72010-06-04 14:41:41 -0700175 * rds_conn_shutdown() sets the conn state and then tests RDS_IN_XMIT,
176 * we do the opposite to avoid races.
177 */
178 if (!rds_conn_up(conn)) {
179 release_in_xmit(conn);
180 ret = 0;
181 goto out;
182 }
Andy Grover5c115592009-02-24 15:30:27 +0000183
184 if (conn->c_trans->xmit_prepare)
185 conn->c_trans->xmit_prepare(conn);
186
187 /*
188 * spin trying to push headers and data down the connection until
Andy Grover5b2366b2010-02-03 19:36:44 -0800189 * the connection doesn't make forward progress.
Andy Grover5c115592009-02-24 15:30:27 +0000190 */
Andy Groverfcc54502010-03-29 17:08:49 -0700191 while (1) {
Andy Grover5c115592009-02-24 15:30:27 +0000192
Andy Grover5c115592009-02-24 15:30:27 +0000193 rm = conn->c_xmit_rm;
Andy Grover5c115592009-02-24 15:30:27 +0000194
Andy Grover5b2366b2010-02-03 19:36:44 -0800195 /*
196 * If between sending messages, we can send a pending congestion
197 * map update.
Andy Grover5c115592009-02-24 15:30:27 +0000198 */
Andy Grover8690bfa2010-01-12 11:56:44 -0800199 if (!rm && test_and_clear_bit(0, &conn->c_map_queued)) {
Andy Grover77dd5502010-03-22 15:22:04 -0700200 rm = rds_cong_update_alloc(conn);
201 if (IS_ERR(rm)) {
202 ret = PTR_ERR(rm);
203 break;
Andy Grover5b2366b2010-02-03 19:36:44 -0800204 }
Andy Grover77dd5502010-03-22 15:22:04 -0700205 rm->data.op_active = 1;
206
207 conn->c_xmit_rm = rm;
Andy Grover5c115592009-02-24 15:30:27 +0000208 }
209
210 /*
Andy Grover5b2366b2010-02-03 19:36:44 -0800211 * If not already working on one, grab the next message.
Andy Grover5c115592009-02-24 15:30:27 +0000212 *
213 * c_xmit_rm holds a ref while we're sending this message down
214 * the connction. We can use this ref while holding the
215 * send_sem.. rds_send_reset() is serialized with it.
216 */
Andy Grover8690bfa2010-01-12 11:56:44 -0800217 if (!rm) {
Andy Grover5c115592009-02-24 15:30:27 +0000218 unsigned int len;
219
Sowmini Varadhan443be0e2015-04-08 12:33:47 -0400220 batch_count++;
221
222 /* we want to process as big a batch as we can, but
223 * we also want to avoid softlockups. If we've been
224 * through a lot of messages, lets back off and see
225 * if anyone else jumps in
226 */
Santosh Shilimkar4bebdd72015-09-10 11:57:14 -0700227 if (batch_count >= send_batch_count)
Sowmini Varadhan443be0e2015-04-08 12:33:47 -0400228 goto over_batch;
229
Zach Brown0f4b1c72010-06-04 14:41:41 -0700230 spin_lock_irqsave(&conn->c_lock, flags);
Andy Grover5c115592009-02-24 15:30:27 +0000231
232 if (!list_empty(&conn->c_send_queue)) {
233 rm = list_entry(conn->c_send_queue.next,
234 struct rds_message,
235 m_conn_item);
236 rds_message_addref(rm);
237
238 /*
239 * Move the message from the send queue to the retransmit
240 * list right away.
241 */
242 list_move_tail(&rm->m_conn_item, &conn->c_retrans);
243 }
244
Zach Brown0f4b1c72010-06-04 14:41:41 -0700245 spin_unlock_irqrestore(&conn->c_lock, flags);
Andy Grover5c115592009-02-24 15:30:27 +0000246
Andy Groverfcc54502010-03-29 17:08:49 -0700247 if (!rm)
Andy Grover5c115592009-02-24 15:30:27 +0000248 break;
Andy Grover5c115592009-02-24 15:30:27 +0000249
250 /* Unfortunately, the way Infiniband deals with
251 * RDMA to a bad MR key is by moving the entire
252 * queue pair to error state. We cold possibly
253 * recover from that, but right now we drop the
254 * connection.
255 * Therefore, we never retransmit messages with RDMA ops.
256 */
Andy Groverf8b3aaf2010-03-01 14:11:53 -0800257 if (rm->rdma.op_active &&
Joe Perchesf64f9e72009-11-29 16:55:45 -0800258 test_bit(RDS_MSG_RETRANSMITTED, &rm->m_flags)) {
Zach Brown0f4b1c72010-06-04 14:41:41 -0700259 spin_lock_irqsave(&conn->c_lock, flags);
Andy Grover5c115592009-02-24 15:30:27 +0000260 if (test_and_clear_bit(RDS_MSG_ON_CONN, &rm->m_flags))
261 list_move(&rm->m_conn_item, &to_be_dropped);
Zach Brown0f4b1c72010-06-04 14:41:41 -0700262 spin_unlock_irqrestore(&conn->c_lock, flags);
Andy Grover5c115592009-02-24 15:30:27 +0000263 continue;
264 }
265
266 /* Require an ACK every once in a while */
267 len = ntohl(rm->m_inc.i_hdr.h_len);
Joe Perchesf64f9e72009-11-29 16:55:45 -0800268 if (conn->c_unacked_packets == 0 ||
269 conn->c_unacked_bytes < len) {
Andy Grover5c115592009-02-24 15:30:27 +0000270 __set_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags);
271
272 conn->c_unacked_packets = rds_sysctl_max_unacked_packets;
273 conn->c_unacked_bytes = rds_sysctl_max_unacked_bytes;
274 rds_stats_inc(s_send_ack_required);
275 } else {
276 conn->c_unacked_bytes -= len;
277 conn->c_unacked_packets--;
278 }
279
280 conn->c_xmit_rm = rm;
281 }
282
Andy Grover2c3a5f92010-03-01 16:10:40 -0800283 /* The transport either sends the whole rdma or none of it */
284 if (rm->rdma.op_active && !conn->c_xmit_rdma_sent) {
Andy Groverff3d7d32010-03-01 14:03:09 -0800285 rm->m_final_op = &rm->rdma;
santosh.shilimkar@oracle.com4f731132015-08-22 15:45:29 -0700286 /* The transport owns the mapped memory for now.
287 * You can't unmap it while it's on the send queue
288 */
289 set_bit(RDS_MSG_MAPPED, &rm->m_flags);
Andy Grover2c3a5f92010-03-01 16:10:40 -0800290 ret = conn->c_trans->xmit_rdma(conn, &rm->rdma);
santosh.shilimkar@oracle.com4f731132015-08-22 15:45:29 -0700291 if (ret) {
292 clear_bit(RDS_MSG_MAPPED, &rm->m_flags);
293 wake_up_interruptible(&rm->m_flush_wait);
Andy Grover2c3a5f92010-03-01 16:10:40 -0800294 break;
santosh.shilimkar@oracle.com4f731132015-08-22 15:45:29 -0700295 }
Andy Grover2c3a5f92010-03-01 16:10:40 -0800296 conn->c_xmit_rdma_sent = 1;
297
Andy Grover2c3a5f92010-03-01 16:10:40 -0800298 }
299
Andy Grover15133f62010-01-12 14:33:38 -0800300 if (rm->atomic.op_active && !conn->c_xmit_atomic_sent) {
Andy Groverff3d7d32010-03-01 14:03:09 -0800301 rm->m_final_op = &rm->atomic;
santosh.shilimkar@oracle.com4f731132015-08-22 15:45:29 -0700302 /* The transport owns the mapped memory for now.
303 * You can't unmap it while it's on the send queue
304 */
305 set_bit(RDS_MSG_MAPPED, &rm->m_flags);
Andy Groverff3d7d32010-03-01 14:03:09 -0800306 ret = conn->c_trans->xmit_atomic(conn, &rm->atomic);
santosh.shilimkar@oracle.com4f731132015-08-22 15:45:29 -0700307 if (ret) {
308 clear_bit(RDS_MSG_MAPPED, &rm->m_flags);
309 wake_up_interruptible(&rm->m_flush_wait);
Andy Grover15133f62010-01-12 14:33:38 -0800310 break;
santosh.shilimkar@oracle.com4f731132015-08-22 15:45:29 -0700311 }
Andy Grover15133f62010-01-12 14:33:38 -0800312 conn->c_xmit_atomic_sent = 1;
Andy Groverff3d7d32010-03-01 14:03:09 -0800313
Andy Grover15133f62010-01-12 14:33:38 -0800314 }
315
Andy Grover2c3a5f92010-03-01 16:10:40 -0800316 /*
317 * A number of cases require an RDS header to be sent
318 * even if there is no data.
319 * We permit 0-byte sends; rds-ping depends on this.
320 * However, if there are exclusively attached silent ops,
321 * we skip the hdr/data send, to enable silent operation.
322 */
323 if (rm->data.op_nents == 0) {
324 int ops_present;
325 int all_ops_are_silent = 1;
Andy Grover241eef32010-01-19 21:25:26 -0800326
Andy Grover2c3a5f92010-03-01 16:10:40 -0800327 ops_present = (rm->atomic.op_active || rm->rdma.op_active);
328 if (rm->atomic.op_active && !rm->atomic.op_silent)
329 all_ops_are_silent = 0;
330 if (rm->rdma.op_active && !rm->rdma.op_silent)
331 all_ops_are_silent = 0;
Andy Grover241eef32010-01-19 21:25:26 -0800332
Andy Grover2c3a5f92010-03-01 16:10:40 -0800333 if (ops_present && all_ops_are_silent
334 && !rm->m_rdma_cookie)
335 rm->data.op_active = 0;
Andy Grover5c115592009-02-24 15:30:27 +0000336 }
337
Andy Grover5b2366b2010-02-03 19:36:44 -0800338 if (rm->data.op_active && !conn->c_xmit_data_sent) {
Andy Groverff3d7d32010-03-01 14:03:09 -0800339 rm->m_final_op = &rm->data;
Andy Grover5c115592009-02-24 15:30:27 +0000340 ret = conn->c_trans->xmit(conn, rm,
341 conn->c_xmit_hdr_off,
342 conn->c_xmit_sg,
343 conn->c_xmit_data_off);
344 if (ret <= 0)
345 break;
346
347 if (conn->c_xmit_hdr_off < sizeof(struct rds_header)) {
348 tmp = min_t(int, ret,
349 sizeof(struct rds_header) -
350 conn->c_xmit_hdr_off);
351 conn->c_xmit_hdr_off += tmp;
352 ret -= tmp;
353 }
354
Andy Grover6c7cc6e2010-01-27 18:04:18 -0800355 sg = &rm->data.op_sg[conn->c_xmit_sg];
Andy Grover5c115592009-02-24 15:30:27 +0000356 while (ret) {
357 tmp = min_t(int, ret, sg->length -
358 conn->c_xmit_data_off);
359 conn->c_xmit_data_off += tmp;
360 ret -= tmp;
361 if (conn->c_xmit_data_off == sg->length) {
362 conn->c_xmit_data_off = 0;
363 sg++;
364 conn->c_xmit_sg++;
365 BUG_ON(ret != 0 &&
Andy Grover6c7cc6e2010-01-27 18:04:18 -0800366 conn->c_xmit_sg == rm->data.op_nents);
Andy Grover5c115592009-02-24 15:30:27 +0000367 }
368 }
Andy Grover5b2366b2010-02-03 19:36:44 -0800369
370 if (conn->c_xmit_hdr_off == sizeof(struct rds_header) &&
371 (conn->c_xmit_sg == rm->data.op_nents))
372 conn->c_xmit_data_sent = 1;
373 }
374
375 /*
376 * A rm will only take multiple times through this loop
377 * if there is a data op. Thus, if the data is sent (or there was
378 * none), then we're done with the rm.
379 */
380 if (!rm->data.op_active || conn->c_xmit_data_sent) {
381 conn->c_xmit_rm = NULL;
382 conn->c_xmit_sg = 0;
383 conn->c_xmit_hdr_off = 0;
384 conn->c_xmit_data_off = 0;
385 conn->c_xmit_rdma_sent = 0;
386 conn->c_xmit_atomic_sent = 0;
387 conn->c_xmit_data_sent = 0;
388
389 rds_message_put(rm);
Andy Grover5c115592009-02-24 15:30:27 +0000390 }
391 }
392
Sowmini Varadhan443be0e2015-04-08 12:33:47 -0400393over_batch:
Andy Grover5c115592009-02-24 15:30:27 +0000394 if (conn->c_trans->xmit_complete)
395 conn->c_trans->xmit_complete(conn);
Zach Brown0f4b1c72010-06-04 14:41:41 -0700396 release_in_xmit(conn);
Andy Grover5c115592009-02-24 15:30:27 +0000397
Andy Grover2ad80992010-03-23 17:48:04 -0700398 /* Nuke any messages we decided not to retransmit. */
399 if (!list_empty(&to_be_dropped)) {
400 /* irqs on here, so we can put(), unlike above */
401 list_for_each_entry(rm, &to_be_dropped, m_conn_item)
402 rds_message_put(rm);
403 rds_send_remove_from_sock(&to_be_dropped, RDS_RDMA_DROPPED);
404 }
405
Andy Groverfcc54502010-03-29 17:08:49 -0700406 /*
Zach Brown0f4b1c72010-06-04 14:41:41 -0700407 * Other senders can queue a message after we last test the send queue
408 * but before we clear RDS_IN_XMIT. In that case they'd back off and
409 * not try and send their newly queued message. We need to check the
410 * send queue after having cleared RDS_IN_XMIT so that their message
411 * doesn't get stuck on the send queue.
Andy Groverfcc54502010-03-29 17:08:49 -0700412 *
413 * If the transport cannot continue (i.e ret != 0), then it must
414 * call us when more room is available, such as from the tx
415 * completion handler.
Sowmini Varadhan443be0e2015-04-08 12:33:47 -0400416 *
417 * We have an extra generation check here so that if someone manages
418 * to jump in after our release_in_xmit, we'll see that they have done
419 * some work and we will skip our goto
Andy Groverfcc54502010-03-29 17:08:49 -0700420 */
421 if (ret == 0) {
Chris Mason9e29db02010-04-15 16:38:14 -0400422 smp_mb();
santosh.shilimkar@oracle.com0c484242015-08-22 15:45:27 -0700423 if ((test_bit(0, &conn->c_map_queued) ||
424 !list_empty(&conn->c_send_queue)) &&
Sowmini Varadhan443be0e2015-04-08 12:33:47 -0400425 send_gen == conn->c_send_gen) {
Andy Grover049ee3f2010-03-23 17:39:07 -0700426 rds_stats_inc(s_send_lock_queue_raced);
Santosh Shilimkar4bebdd72015-09-10 11:57:14 -0700427 if (batch_count < send_batch_count)
428 goto restart;
429 queue_delayed_work(rds_wq, &conn->c_send_w, 1);
Andy Grover5c115592009-02-24 15:30:27 +0000430 }
Andy Grover5c115592009-02-24 15:30:27 +0000431 }
432out:
433 return ret;
434}
435
436static void rds_send_sndbuf_remove(struct rds_sock *rs, struct rds_message *rm)
437{
438 u32 len = be32_to_cpu(rm->m_inc.i_hdr.h_len);
439
440 assert_spin_locked(&rs->rs_lock);
441
442 BUG_ON(rs->rs_snd_bytes < len);
443 rs->rs_snd_bytes -= len;
444
445 if (rs->rs_snd_bytes == 0)
446 rds_stats_inc(s_send_queue_empty);
447}
448
449static inline int rds_send_is_acked(struct rds_message *rm, u64 ack,
450 is_acked_func is_acked)
451{
452 if (is_acked)
453 return is_acked(rm, ack);
454 return be64_to_cpu(rm->m_inc.i_hdr.h_sequence) <= ack;
455}
456
457/*
Andy Grover5c115592009-02-24 15:30:27 +0000458 * This is pretty similar to what happens below in the ACK
459 * handling code - except that we call here as soon as we get
460 * the IB send completion on the RDMA op and the accompanying
461 * message.
462 */
463void rds_rdma_send_complete(struct rds_message *rm, int status)
464{
465 struct rds_sock *rs = NULL;
Andy Groverf8b3aaf2010-03-01 14:11:53 -0800466 struct rm_rdma_op *ro;
Andy Grover5c115592009-02-24 15:30:27 +0000467 struct rds_notifier *notifier;
Andy Grover9de08642010-03-29 16:50:54 -0700468 unsigned long flags;
Andy Grover5c115592009-02-24 15:30:27 +0000469
Andy Grover9de08642010-03-29 16:50:54 -0700470 spin_lock_irqsave(&rm->m_rs_lock, flags);
Andy Grover5c115592009-02-24 15:30:27 +0000471
Andy Groverf8b3aaf2010-03-01 14:11:53 -0800472 ro = &rm->rdma;
Joe Perchesf64f9e72009-11-29 16:55:45 -0800473 if (test_bit(RDS_MSG_ON_SOCK, &rm->m_flags) &&
Andy Groverf8b3aaf2010-03-01 14:11:53 -0800474 ro->op_active && ro->op_notify && ro->op_notifier) {
475 notifier = ro->op_notifier;
Andy Grover5c115592009-02-24 15:30:27 +0000476 rs = rm->m_rs;
477 sock_hold(rds_rs_to_sk(rs));
478
479 notifier->n_status = status;
480 spin_lock(&rs->rs_lock);
481 list_add_tail(&notifier->n_list, &rs->rs_notify_queue);
482 spin_unlock(&rs->rs_lock);
483
Andy Groverf8b3aaf2010-03-01 14:11:53 -0800484 ro->op_notifier = NULL;
Andy Grover5c115592009-02-24 15:30:27 +0000485 }
486
Andy Grover9de08642010-03-29 16:50:54 -0700487 spin_unlock_irqrestore(&rm->m_rs_lock, flags);
Andy Grover5c115592009-02-24 15:30:27 +0000488
489 if (rs) {
490 rds_wake_sk_sleep(rs);
491 sock_put(rds_rs_to_sk(rs));
492 }
493}
Andy Grover616b7572009-08-21 12:28:32 +0000494EXPORT_SYMBOL_GPL(rds_rdma_send_complete);
Andy Grover5c115592009-02-24 15:30:27 +0000495
496/*
Andy Grover15133f62010-01-12 14:33:38 -0800497 * Just like above, except looks at atomic op
498 */
499void rds_atomic_send_complete(struct rds_message *rm, int status)
500{
501 struct rds_sock *rs = NULL;
502 struct rm_atomic_op *ao;
503 struct rds_notifier *notifier;
Andy Grovercf4b7382010-03-29 16:50:54 -0700504 unsigned long flags;
Andy Grover15133f62010-01-12 14:33:38 -0800505
Andy Grovercf4b7382010-03-29 16:50:54 -0700506 spin_lock_irqsave(&rm->m_rs_lock, flags);
Andy Grover15133f62010-01-12 14:33:38 -0800507
508 ao = &rm->atomic;
509 if (test_bit(RDS_MSG_ON_SOCK, &rm->m_flags)
510 && ao->op_active && ao->op_notify && ao->op_notifier) {
511 notifier = ao->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(&notifier->n_list, &rs->rs_notify_queue);
518 spin_unlock(&rs->rs_lock);
519
520 ao->op_notifier = NULL;
521 }
522
Andy Grovercf4b7382010-03-29 16:50:54 -0700523 spin_unlock_irqrestore(&rm->m_rs_lock, flags);
Andy Grover15133f62010-01-12 14:33:38 -0800524
525 if (rs) {
526 rds_wake_sk_sleep(rs);
527 sock_put(rds_rs_to_sk(rs));
528 }
529}
530EXPORT_SYMBOL_GPL(rds_atomic_send_complete);
531
532/*
Andy Grover5c115592009-02-24 15:30:27 +0000533 * This is the same as rds_rdma_send_complete except we
534 * don't do any locking - we have all the ingredients (message,
535 * socket, socket lock) and can just move the notifier.
536 */
537static inline void
Andy Grover940786e2010-02-19 18:04:58 -0800538__rds_send_complete(struct rds_sock *rs, struct rds_message *rm, int status)
Andy Grover5c115592009-02-24 15:30:27 +0000539{
Andy Groverf8b3aaf2010-03-01 14:11:53 -0800540 struct rm_rdma_op *ro;
Andy Grover940786e2010-02-19 18:04:58 -0800541 struct rm_atomic_op *ao;
Andy Grover5c115592009-02-24 15:30:27 +0000542
Andy Groverf8b3aaf2010-03-01 14:11:53 -0800543 ro = &rm->rdma;
544 if (ro->op_active && ro->op_notify && ro->op_notifier) {
545 ro->op_notifier->n_status = status;
546 list_add_tail(&ro->op_notifier->n_list, &rs->rs_notify_queue);
547 ro->op_notifier = NULL;
Andy Grover5c115592009-02-24 15:30:27 +0000548 }
549
Andy Grover940786e2010-02-19 18:04:58 -0800550 ao = &rm->atomic;
551 if (ao->op_active && ao->op_notify && ao->op_notifier) {
552 ao->op_notifier->n_status = status;
553 list_add_tail(&ao->op_notifier->n_list, &rs->rs_notify_queue);
554 ao->op_notifier = NULL;
555 }
556
Andy Grover5c115592009-02-24 15:30:27 +0000557 /* No need to wake the app - caller does this */
558}
559
560/*
561 * This is called from the IB send completion when we detect
562 * a RDMA operation that failed with remote access error.
563 * So speed is not an issue here.
564 */
565struct rds_message *rds_send_get_message(struct rds_connection *conn,
Andy Groverf8b3aaf2010-03-01 14:11:53 -0800566 struct rm_rdma_op *op)
Andy Grover5c115592009-02-24 15:30:27 +0000567{
568 struct rds_message *rm, *tmp, *found = NULL;
569 unsigned long flags;
570
571 spin_lock_irqsave(&conn->c_lock, flags);
572
573 list_for_each_entry_safe(rm, tmp, &conn->c_retrans, m_conn_item) {
Andy Groverf8b3aaf2010-03-01 14:11:53 -0800574 if (&rm->rdma == op) {
Andy Grover5c115592009-02-24 15:30:27 +0000575 atomic_inc(&rm->m_refcount);
576 found = rm;
577 goto out;
578 }
579 }
580
581 list_for_each_entry_safe(rm, tmp, &conn->c_send_queue, m_conn_item) {
Andy Groverf8b3aaf2010-03-01 14:11:53 -0800582 if (&rm->rdma == op) {
Andy Grover5c115592009-02-24 15:30:27 +0000583 atomic_inc(&rm->m_refcount);
584 found = rm;
585 break;
586 }
587 }
588
589out:
590 spin_unlock_irqrestore(&conn->c_lock, flags);
591
592 return found;
593}
Andy Grover616b7572009-08-21 12:28:32 +0000594EXPORT_SYMBOL_GPL(rds_send_get_message);
Andy Grover5c115592009-02-24 15:30:27 +0000595
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 */
stephen hemmingerff51bf82010-10-19 08:08:33 +0000604static void rds_send_remove_from_sock(struct list_head *messages, int status)
Andy Grover5c115592009-02-24 15:30:27 +0000605{
Andy Grover561c7df2010-03-11 13:50:06 +0000606 unsigned long flags;
Andy Grover5c115592009-02-24 15:30:27 +0000607 struct rds_sock *rs = NULL;
608 struct rds_message *rm;
609
Andy Grover5c115592009-02-24 15:30:27 +0000610 while (!list_empty(messages)) {
Andy Grover561c7df2010-03-11 13:50:06 +0000611 int was_on_sock = 0;
612
Andy Grover5c115592009-02-24 15:30:27 +0000613 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 */
Andy Grover561c7df2010-03-11 13:50:06 +0000627 spin_lock_irqsave(&rm->m_rs_lock, flags);
Andy Grover5c115592009-02-24 15:30:27 +0000628 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) {
Andy Grover5c115592009-02-24 15:30:27 +0000633 rds_wake_sk_sleep(rs);
634 sock_put(rds_rs_to_sk(rs));
635 }
636 rs = rm->m_rs;
Herton R. Krzesinski593cbb32014-10-01 18:49:54 -0300637 if (rs)
638 sock_hold(rds_rs_to_sk(rs));
Andy Grover5c115592009-02-24 15:30:27 +0000639 }
Herton R. Krzesinski593cbb32014-10-01 18:49:54 -0300640 if (!rs)
641 goto unlock_and_drop;
Tina Yang048c15e2010-03-11 13:50:00 +0000642 spin_lock(&rs->rs_lock);
Andy Grover5c115592009-02-24 15:30:27 +0000643
644 if (test_and_clear_bit(RDS_MSG_ON_SOCK, &rm->m_flags)) {
Andy Groverf8b3aaf2010-03-01 14:11:53 -0800645 struct rm_rdma_op *ro = &rm->rdma;
Andy Grover5c115592009-02-24 15:30:27 +0000646 struct rds_notifier *notifier;
647
648 list_del_init(&rm->m_sock_item);
649 rds_send_sndbuf_remove(rs, rm);
650
Andy Groverf8b3aaf2010-03-01 14:11:53 -0800651 if (ro->op_active && ro->op_notifier &&
652 (ro->op_notify || (ro->op_recverr && status))) {
653 notifier = ro->op_notifier;
Andy Grover5c115592009-02-24 15:30:27 +0000654 list_add_tail(&notifier->n_list,
655 &rs->rs_notify_queue);
656 if (!notifier->n_status)
657 notifier->n_status = status;
Andy Groverf8b3aaf2010-03-01 14:11:53 -0800658 rm->rdma.op_notifier = NULL;
Andy Grover5c115592009-02-24 15:30:27 +0000659 }
Andy Grover561c7df2010-03-11 13:50:06 +0000660 was_on_sock = 1;
Andy Grover5c115592009-02-24 15:30:27 +0000661 rm->m_rs = NULL;
662 }
Tina Yang048c15e2010-03-11 13:50:00 +0000663 spin_unlock(&rs->rs_lock);
Andy Grover5c115592009-02-24 15:30:27 +0000664
665unlock_and_drop:
Andy Grover561c7df2010-03-11 13:50:06 +0000666 spin_unlock_irqrestore(&rm->m_rs_lock, flags);
Andy Grover5c115592009-02-24 15:30:27 +0000667 rds_message_put(rm);
Andy Grover561c7df2010-03-11 13:50:06 +0000668 if (was_on_sock)
669 rds_message_put(rm);
Andy Grover5c115592009-02-24 15:30:27 +0000670 }
671
672 if (rs) {
Andy Grover5c115592009-02-24 15:30:27 +0000673 rds_wake_sk_sleep(rs);
674 sock_put(rds_rs_to_sk(rs));
675 }
Andy Grover5c115592009-02-24 15:30:27 +0000676}
677
678/*
679 * Transports call here when they've determined that the receiver queued
680 * messages up to, and including, the given sequence number. Messages are
681 * moved to the retrans queue when rds_send_xmit picks them off the send
682 * queue. This means that in the TCP case, the message may not have been
683 * assigned the m_ack_seq yet - but that's fine as long as tcp_is_acked
684 * checks the RDS_MSG_HAS_ACK_SEQ bit.
Andy Grover5c115592009-02-24 15:30:27 +0000685 */
686void rds_send_drop_acked(struct rds_connection *conn, u64 ack,
687 is_acked_func is_acked)
688{
689 struct rds_message *rm, *tmp;
690 unsigned long flags;
691 LIST_HEAD(list);
692
693 spin_lock_irqsave(&conn->c_lock, flags);
694
695 list_for_each_entry_safe(rm, tmp, &conn->c_retrans, m_conn_item) {
696 if (!rds_send_is_acked(rm, ack, is_acked))
697 break;
698
699 list_move(&rm->m_conn_item, &list);
700 clear_bit(RDS_MSG_ON_CONN, &rm->m_flags);
701 }
702
703 /* order flag updates with spin locks */
704 if (!list_empty(&list))
Peter Zijlstra4e857c52014-03-17 18:06:10 +0100705 smp_mb__after_atomic();
Andy Grover5c115592009-02-24 15:30:27 +0000706
707 spin_unlock_irqrestore(&conn->c_lock, flags);
708
709 /* now remove the messages from the sock list as needed */
710 rds_send_remove_from_sock(&list, RDS_RDMA_SUCCESS);
711}
Andy Grover616b7572009-08-21 12:28:32 +0000712EXPORT_SYMBOL_GPL(rds_send_drop_acked);
Andy Grover5c115592009-02-24 15:30:27 +0000713
714void rds_send_drop_to(struct rds_sock *rs, struct sockaddr_in *dest)
715{
716 struct rds_message *rm, *tmp;
717 struct rds_connection *conn;
Andy Grover7c82eaf2010-02-19 18:01:41 -0800718 unsigned long flags;
Andy Grover5c115592009-02-24 15:30:27 +0000719 LIST_HEAD(list);
Andy Grover5c115592009-02-24 15:30:27 +0000720
721 /* get all the messages we're dropping under the rs lock */
722 spin_lock_irqsave(&rs->rs_lock, flags);
723
724 list_for_each_entry_safe(rm, tmp, &rs->rs_send_queue, m_sock_item) {
725 if (dest && (dest->sin_addr.s_addr != rm->m_daddr ||
726 dest->sin_port != rm->m_inc.i_hdr.h_dport))
727 continue;
728
Andy Grover5c115592009-02-24 15:30:27 +0000729 list_move(&rm->m_sock_item, &list);
730 rds_send_sndbuf_remove(rs, rm);
731 clear_bit(RDS_MSG_ON_SOCK, &rm->m_flags);
Andy Grover5c115592009-02-24 15:30:27 +0000732 }
733
734 /* order flag updates with the rs lock */
Peter Zijlstra4e857c52014-03-17 18:06:10 +0100735 smp_mb__after_atomic();
Andy Grover5c115592009-02-24 15:30:27 +0000736
737 spin_unlock_irqrestore(&rs->rs_lock, flags);
738
Andy Grover7c82eaf2010-02-19 18:01:41 -0800739 if (list_empty(&list))
740 return;
Andy Grover5c115592009-02-24 15:30:27 +0000741
Andy Grover7c82eaf2010-02-19 18:01:41 -0800742 /* Remove the messages from the conn */
Andy Grover5c115592009-02-24 15:30:27 +0000743 list_for_each_entry(rm, &list, m_sock_item) {
Andy Grover7c82eaf2010-02-19 18:01:41 -0800744
745 conn = rm->m_inc.i_conn;
Andy Grover7c82eaf2010-02-19 18:01:41 -0800746
Andy Grover9de08642010-03-29 16:50:54 -0700747 spin_lock_irqsave(&conn->c_lock, flags);
Andy Grover7c82eaf2010-02-19 18:01:41 -0800748 /*
749 * Maybe someone else beat us to removing rm from the conn.
750 * If we race with their flag update we'll get the lock and
751 * then really see that the flag has been cleared.
752 */
753 if (!test_and_clear_bit(RDS_MSG_ON_CONN, &rm->m_flags)) {
754 spin_unlock_irqrestore(&conn->c_lock, flags);
Herton R. Krzesinski593cbb32014-10-01 18:49:54 -0300755 spin_lock_irqsave(&rm->m_rs_lock, flags);
756 rm->m_rs = NULL;
757 spin_unlock_irqrestore(&rm->m_rs_lock, flags);
Andy Grover7c82eaf2010-02-19 18:01:41 -0800758 continue;
759 }
Andy Grover9de08642010-03-29 16:50:54 -0700760 list_del_init(&rm->m_conn_item);
761 spin_unlock_irqrestore(&conn->c_lock, flags);
Andy Grover7c82eaf2010-02-19 18:01:41 -0800762
763 /*
764 * Couldn't grab m_rs_lock in top loop (lock ordering),
765 * but we can now.
766 */
Andy Grover9de08642010-03-29 16:50:54 -0700767 spin_lock_irqsave(&rm->m_rs_lock, flags);
Andy Grover7c82eaf2010-02-19 18:01:41 -0800768
Tina Yang550a8002010-03-11 13:50:03 +0000769 spin_lock(&rs->rs_lock);
Andy Grover940786e2010-02-19 18:04:58 -0800770 __rds_send_complete(rs, rm, RDS_RDMA_CANCELED);
Tina Yang550a8002010-03-11 13:50:03 +0000771 spin_unlock(&rs->rs_lock);
Andy Grover7c82eaf2010-02-19 18:01:41 -0800772
Andy Grover5c115592009-02-24 15:30:27 +0000773 rm->m_rs = NULL;
Andy Grover9de08642010-03-29 16:50:54 -0700774 spin_unlock_irqrestore(&rm->m_rs_lock, flags);
Andy Grover5c115592009-02-24 15:30:27 +0000775
Andy Grover7c82eaf2010-02-19 18:01:41 -0800776 rds_message_put(rm);
Andy Grover5c115592009-02-24 15:30:27 +0000777 }
778
Andy Grover7c82eaf2010-02-19 18:01:41 -0800779 rds_wake_sk_sleep(rs);
Tina Yang550a8002010-03-11 13:50:03 +0000780
Andy Grover5c115592009-02-24 15:30:27 +0000781 while (!list_empty(&list)) {
782 rm = list_entry(list.next, struct rds_message, m_sock_item);
783 list_del_init(&rm->m_sock_item);
Andy Grover5c115592009-02-24 15:30:27 +0000784 rds_message_wait(rm);
santosh.shilimkar@oracle.comdfcec252015-08-22 15:45:33 -0700785
786 /* just in case the code above skipped this message
787 * because RDS_MSG_ON_CONN wasn't set, run it again here
788 * taking m_rs_lock is the only thing that keeps us
789 * from racing with ack processing.
790 */
791 spin_lock_irqsave(&rm->m_rs_lock, flags);
792
793 spin_lock(&rs->rs_lock);
794 __rds_send_complete(rs, rm, RDS_RDMA_CANCELED);
795 spin_unlock(&rs->rs_lock);
796
797 rm->m_rs = NULL;
798 spin_unlock_irqrestore(&rm->m_rs_lock, flags);
799
Andy Grover5c115592009-02-24 15:30:27 +0000800 rds_message_put(rm);
801 }
802}
803
804/*
805 * we only want this to fire once so we use the callers 'queued'. It's
806 * possible that another thread can race with us and remove the
807 * message from the flow with RDS_CANCEL_SENT_TO.
808 */
809static int rds_send_queue_rm(struct rds_sock *rs, struct rds_connection *conn,
810 struct rds_message *rm, __be16 sport,
811 __be16 dport, int *queued)
812{
813 unsigned long flags;
814 u32 len;
815
816 if (*queued)
817 goto out;
818
819 len = be32_to_cpu(rm->m_inc.i_hdr.h_len);
820
821 /* this is the only place which holds both the socket's rs_lock
822 * and the connection's c_lock */
823 spin_lock_irqsave(&rs->rs_lock, flags);
824
825 /*
826 * If there is a little space in sndbuf, we don't queue anything,
827 * and userspace gets -EAGAIN. But poll() indicates there's send
828 * room. This can lead to bad behavior (spinning) if snd_bytes isn't
829 * freed up by incoming acks. So we check the *old* value of
830 * rs_snd_bytes here to allow the last msg to exceed the buffer,
831 * and poll() now knows no more data can be sent.
832 */
833 if (rs->rs_snd_bytes < rds_sk_sndbuf(rs)) {
834 rs->rs_snd_bytes += len;
835
836 /* let recv side know we are close to send space exhaustion.
837 * This is probably not the optimal way to do it, as this
838 * means we set the flag on *all* messages as soon as our
839 * throughput hits a certain threshold.
840 */
841 if (rs->rs_snd_bytes >= rds_sk_sndbuf(rs) / 2)
842 __set_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags);
843
844 list_add_tail(&rm->m_sock_item, &rs->rs_send_queue);
845 set_bit(RDS_MSG_ON_SOCK, &rm->m_flags);
846 rds_message_addref(rm);
847 rm->m_rs = rs;
848
849 /* The code ordering is a little weird, but we're
850 trying to minimize the time we hold c_lock */
851 rds_message_populate_header(&rm->m_inc.i_hdr, sport, dport, 0);
852 rm->m_inc.i_conn = conn;
853 rds_message_addref(rm);
854
855 spin_lock(&conn->c_lock);
856 rm->m_inc.i_hdr.h_sequence = cpu_to_be64(conn->c_next_tx_seq++);
857 list_add_tail(&rm->m_conn_item, &conn->c_send_queue);
858 set_bit(RDS_MSG_ON_CONN, &rm->m_flags);
859 spin_unlock(&conn->c_lock);
860
861 rdsdebug("queued msg %p len %d, rs %p bytes %d seq %llu\n",
862 rm, len, rs, rs->rs_snd_bytes,
863 (unsigned long long)be64_to_cpu(rm->m_inc.i_hdr.h_sequence));
864
865 *queued = 1;
866 }
867
868 spin_unlock_irqrestore(&rs->rs_lock, flags);
869out:
870 return *queued;
871}
872
Andy Groverfc445082010-01-12 12:56:06 -0800873/*
874 * rds_message is getting to be quite complicated, and we'd like to allocate
875 * it all in one go. This figures out how big it needs to be up front.
876 */
877static int rds_rm_size(struct msghdr *msg, int data_len)
878{
Andy Groverff87e972010-01-12 14:13:15 -0800879 struct cmsghdr *cmsg;
Andy Groverfc445082010-01-12 12:56:06 -0800880 int size = 0;
Andy Groveraa0a4ef2010-04-13 12:00:35 -0700881 int cmsg_groups = 0;
Andy Groverff87e972010-01-12 14:13:15 -0800882 int retval;
Andy Groverfc445082010-01-12 12:56:06 -0800883
Gu Zhengf95b4142014-12-11 11:22:04 +0800884 for_each_cmsghdr(cmsg, msg) {
Andy Groverff87e972010-01-12 14:13:15 -0800885 if (!CMSG_OK(msg, cmsg))
886 return -EINVAL;
887
888 if (cmsg->cmsg_level != SOL_RDS)
889 continue;
890
891 switch (cmsg->cmsg_type) {
892 case RDS_CMSG_RDMA_ARGS:
Andy Groveraa0a4ef2010-04-13 12:00:35 -0700893 cmsg_groups |= 1;
Andy Groverff87e972010-01-12 14:13:15 -0800894 retval = rds_rdma_extra_size(CMSG_DATA(cmsg));
895 if (retval < 0)
896 return retval;
897 size += retval;
Andy Groveraa0a4ef2010-04-13 12:00:35 -0700898
Andy Groverff87e972010-01-12 14:13:15 -0800899 break;
900
901 case RDS_CMSG_RDMA_DEST:
902 case RDS_CMSG_RDMA_MAP:
Andy Groveraa0a4ef2010-04-13 12:00:35 -0700903 cmsg_groups |= 2;
Andy Groverff87e972010-01-12 14:13:15 -0800904 /* these are valid but do no add any size */
905 break;
906
Andy Grover15133f62010-01-12 14:33:38 -0800907 case RDS_CMSG_ATOMIC_CSWP:
908 case RDS_CMSG_ATOMIC_FADD:
Andy Grover20c72bd2010-08-25 05:51:28 -0700909 case RDS_CMSG_MASKED_ATOMIC_CSWP:
910 case RDS_CMSG_MASKED_ATOMIC_FADD:
Andy Groveraa0a4ef2010-04-13 12:00:35 -0700911 cmsg_groups |= 1;
Andy Grover15133f62010-01-12 14:33:38 -0800912 size += sizeof(struct scatterlist);
913 break;
914
Andy Groverff87e972010-01-12 14:13:15 -0800915 default:
916 return -EINVAL;
917 }
918
919 }
920
921 size += ceil(data_len, PAGE_SIZE) * sizeof(struct scatterlist);
Andy Groverfc445082010-01-12 12:56:06 -0800922
Andy Groveraa0a4ef2010-04-13 12:00:35 -0700923 /* Ensure (DEST, MAP) are never used with (ARGS, ATOMIC) */
924 if (cmsg_groups == 3)
925 return -EINVAL;
926
Andy Groverfc445082010-01-12 12:56:06 -0800927 return size;
928}
929
Andy Grover5c115592009-02-24 15:30:27 +0000930static int rds_cmsg_send(struct rds_sock *rs, struct rds_message *rm,
931 struct msghdr *msg, int *allocated_mr)
932{
933 struct cmsghdr *cmsg;
934 int ret = 0;
935
Gu Zhengf95b4142014-12-11 11:22:04 +0800936 for_each_cmsghdr(cmsg, msg) {
Andy Grover5c115592009-02-24 15:30:27 +0000937 if (!CMSG_OK(msg, cmsg))
938 return -EINVAL;
939
940 if (cmsg->cmsg_level != SOL_RDS)
941 continue;
942
943 /* As a side effect, RDMA_DEST and RDMA_MAP will set
Andy Grover15133f62010-01-12 14:33:38 -0800944 * rm->rdma.m_rdma_cookie and rm->rdma.m_rdma_mr.
Andy Grover5c115592009-02-24 15:30:27 +0000945 */
946 switch (cmsg->cmsg_type) {
947 case RDS_CMSG_RDMA_ARGS:
948 ret = rds_cmsg_rdma_args(rs, rm, cmsg);
949 break;
950
951 case RDS_CMSG_RDMA_DEST:
952 ret = rds_cmsg_rdma_dest(rs, rm, cmsg);
953 break;
954
955 case RDS_CMSG_RDMA_MAP:
956 ret = rds_cmsg_rdma_map(rs, rm, cmsg);
957 if (!ret)
958 *allocated_mr = 1;
959 break;
Andy Grover15133f62010-01-12 14:33:38 -0800960 case RDS_CMSG_ATOMIC_CSWP:
961 case RDS_CMSG_ATOMIC_FADD:
Andy Grover20c72bd2010-08-25 05:51:28 -0700962 case RDS_CMSG_MASKED_ATOMIC_CSWP:
963 case RDS_CMSG_MASKED_ATOMIC_FADD:
Andy Grover15133f62010-01-12 14:33:38 -0800964 ret = rds_cmsg_atomic(rs, rm, cmsg);
965 break;
Andy Grover5c115592009-02-24 15:30:27 +0000966
967 default:
968 return -EINVAL;
969 }
970
971 if (ret)
972 break;
973 }
974
975 return ret;
976}
977
Ying Xue1b784142015-03-02 15:37:48 +0800978int rds_sendmsg(struct socket *sock, struct msghdr *msg, size_t payload_len)
Andy Grover5c115592009-02-24 15:30:27 +0000979{
980 struct sock *sk = sock->sk;
981 struct rds_sock *rs = rds_sk_to_rs(sk);
Steffen Hurrle342dfc32014-01-17 22:53:15 +0100982 DECLARE_SOCKADDR(struct sockaddr_in *, usin, msg->msg_name);
Andy Grover5c115592009-02-24 15:30:27 +0000983 __be32 daddr;
984 __be16 dport;
985 struct rds_message *rm = NULL;
986 struct rds_connection *conn;
987 int ret = 0;
988 int queued = 0, allocated_mr = 0;
989 int nonblock = msg->msg_flags & MSG_DONTWAIT;
Andy Grover1123fd72010-03-11 13:49:56 +0000990 long timeo = sock_sndtimeo(sk, nonblock);
Andy Grover5c115592009-02-24 15:30:27 +0000991
992 /* Mirror Linux UDP mirror of BSD error message compatibility */
993 /* XXX: Perhaps MSG_MORE someday */
994 if (msg->msg_flags & ~(MSG_DONTWAIT | MSG_CMSG_COMPAT)) {
Andy Grover5c115592009-02-24 15:30:27 +0000995 ret = -EOPNOTSUPP;
996 goto out;
997 }
998
999 if (msg->msg_namelen) {
1000 /* XXX fail non-unicast destination IPs? */
1001 if (msg->msg_namelen < sizeof(*usin) || usin->sin_family != AF_INET) {
1002 ret = -EINVAL;
1003 goto out;
1004 }
1005 daddr = usin->sin_addr.s_addr;
1006 dport = usin->sin_port;
1007 } else {
1008 /* We only care about consistency with ->connect() */
1009 lock_sock(sk);
1010 daddr = rs->rs_conn_addr;
1011 dport = rs->rs_conn_port;
1012 release_sock(sk);
1013 }
1014
1015 /* racing with another thread binding seems ok here */
1016 if (daddr == 0 || rs->rs_bound_addr == 0) {
1017 ret = -ENOTCONN; /* XXX not a great errno */
1018 goto out;
1019 }
1020
Mukesh Kacker06e89412015-08-22 15:45:34 -07001021 if (payload_len > rds_sk_sndbuf(rs)) {
1022 ret = -EMSGSIZE;
1023 goto out;
1024 }
1025
Andy Groverfc445082010-01-12 12:56:06 -08001026 /* size of rm including all sgs */
1027 ret = rds_rm_size(msg, payload_len);
1028 if (ret < 0)
1029 goto out;
1030
1031 rm = rds_message_alloc(ret, GFP_KERNEL);
1032 if (!rm) {
1033 ret = -ENOMEM;
Andy Grover5c115592009-02-24 15:30:27 +00001034 goto out;
1035 }
1036
Andy Grover372cd7d2010-02-03 19:40:32 -08001037 /* Attach data to the rm */
1038 if (payload_len) {
1039 rm->data.op_sg = rds_message_alloc_sgs(rm, ceil(payload_len, PAGE_SIZE));
Andy Groverd139ff02010-10-28 15:40:59 +00001040 if (!rm->data.op_sg) {
1041 ret = -ENOMEM;
1042 goto out;
1043 }
Al Viroc0371da2014-11-24 10:42:55 -05001044 ret = rds_message_copy_from_user(rm, &msg->msg_iter);
Andy Grover372cd7d2010-02-03 19:40:32 -08001045 if (ret)
1046 goto out;
1047 }
1048 rm->data.op_active = 1;
Andy Groverfc445082010-01-12 12:56:06 -08001049
Andy Grover5c115592009-02-24 15:30:27 +00001050 rm->m_daddr = daddr;
1051
Andy Grover5c115592009-02-24 15:30:27 +00001052 /* rds_conn_create has a spinlock that runs with IRQ off.
1053 * Caching the conn in the socket helps a lot. */
1054 if (rs->rs_conn && rs->rs_conn->c_faddr == daddr)
1055 conn = rs->rs_conn;
1056 else {
Sowmini Varadhand5a8ac22015-08-05 01:43:25 -04001057 conn = rds_conn_create_outgoing(sock_net(sock->sk),
1058 rs->rs_bound_addr, daddr,
Andy Grover5c115592009-02-24 15:30:27 +00001059 rs->rs_transport,
1060 sock->sk->sk_allocation);
1061 if (IS_ERR(conn)) {
1062 ret = PTR_ERR(conn);
1063 goto out;
1064 }
1065 rs->rs_conn = conn;
1066 }
1067
Andy Grover49f69692009-04-09 14:09:41 +00001068 /* Parse any control messages the user may have included. */
1069 ret = rds_cmsg_send(rs, rm, msg, &allocated_mr);
1070 if (ret)
1071 goto out;
1072
Andy Grover2c3a5f92010-03-01 16:10:40 -08001073 if (rm->rdma.op_active && !conn->c_trans->xmit_rdma) {
Manuel Zerpiescb0a6052011-06-16 02:09:57 +00001074 printk_ratelimited(KERN_NOTICE "rdma_op %p conn xmit_rdma %p\n",
Andy Groverf8b3aaf2010-03-01 14:11:53 -08001075 &rm->rdma, conn->c_trans->xmit_rdma);
Andy Grover15133f62010-01-12 14:33:38 -08001076 ret = -EOPNOTSUPP;
1077 goto out;
1078 }
1079
1080 if (rm->atomic.op_active && !conn->c_trans->xmit_atomic) {
Manuel Zerpiescb0a6052011-06-16 02:09:57 +00001081 printk_ratelimited(KERN_NOTICE "atomic_op %p conn xmit_atomic %p\n",
Andy Grover15133f62010-01-12 14:33:38 -08001082 &rm->atomic, conn->c_trans->xmit_atomic);
Andy Grover5c115592009-02-24 15:30:27 +00001083 ret = -EOPNOTSUPP;
1084 goto out;
1085 }
1086
Zach Brownf3c68082010-05-24 13:14:36 -07001087 rds_conn_connect_if_down(conn);
Andy Grover5c115592009-02-24 15:30:27 +00001088
1089 ret = rds_cong_wait(conn->c_fcong, dport, nonblock, rs);
Andy Groverb98ba522010-03-11 13:50:04 +00001090 if (ret) {
1091 rs->rs_seen_congestion = 1;
Andy Grover5c115592009-02-24 15:30:27 +00001092 goto out;
Andy Groverb98ba522010-03-11 13:50:04 +00001093 }
Andy Grover5c115592009-02-24 15:30:27 +00001094
1095 while (!rds_send_queue_rm(rs, conn, rm, rs->rs_bound_port,
1096 dport, &queued)) {
1097 rds_stats_inc(s_send_queue_full);
Mukesh Kacker06e89412015-08-22 15:45:34 -07001098
Andy Grover5c115592009-02-24 15:30:27 +00001099 if (nonblock) {
1100 ret = -EAGAIN;
1101 goto out;
1102 }
1103
Eric Dumazetaa395142010-04-20 13:03:51 +00001104 timeo = wait_event_interruptible_timeout(*sk_sleep(sk),
Andy Grover5c115592009-02-24 15:30:27 +00001105 rds_send_queue_rm(rs, conn, rm,
1106 rs->rs_bound_port,
1107 dport,
1108 &queued),
1109 timeo);
1110 rdsdebug("sendmsg woke queued %d timeo %ld\n", queued, timeo);
1111 if (timeo > 0 || timeo == MAX_SCHEDULE_TIMEOUT)
1112 continue;
1113
1114 ret = timeo;
1115 if (ret == 0)
1116 ret = -ETIMEDOUT;
1117 goto out;
1118 }
1119
1120 /*
1121 * By now we've committed to the send. We reuse rds_send_worker()
1122 * to retry sends in the rds thread if the transport asks us to.
1123 */
1124 rds_stats_inc(s_send_queued);
1125
1126 if (!test_bit(RDS_LL_SEND_FULL, &conn->c_flags))
Andy Grovera7d3a282010-03-29 16:20:18 -07001127 rds_send_xmit(conn);
Andy Grover5c115592009-02-24 15:30:27 +00001128
1129 rds_message_put(rm);
1130 return payload_len;
1131
1132out:
1133 /* If the user included a RDMA_MAP cmsg, we allocated a MR on the fly.
1134 * If the sendmsg goes through, we keep the MR. If it fails with EAGAIN
1135 * or in any other way, we need to destroy the MR again */
1136 if (allocated_mr)
1137 rds_rdma_unuse(rs, rds_rdma_cookie_key(rm->m_rdma_cookie), 1);
1138
1139 if (rm)
1140 rds_message_put(rm);
1141 return ret;
1142}
1143
1144/*
1145 * Reply to a ping packet.
1146 */
1147int
1148rds_send_pong(struct rds_connection *conn, __be16 dport)
1149{
1150 struct rds_message *rm;
1151 unsigned long flags;
1152 int ret = 0;
1153
1154 rm = rds_message_alloc(0, GFP_ATOMIC);
Andy Grover8690bfa2010-01-12 11:56:44 -08001155 if (!rm) {
Andy Grover5c115592009-02-24 15:30:27 +00001156 ret = -ENOMEM;
1157 goto out;
1158 }
1159
1160 rm->m_daddr = conn->c_faddr;
Andy Groveracfcd4d2010-03-31 18:56:25 -07001161 rm->data.op_active = 1;
Andy Grover5c115592009-02-24 15:30:27 +00001162
Zach Brownf3c68082010-05-24 13:14:36 -07001163 rds_conn_connect_if_down(conn);
Andy Grover5c115592009-02-24 15:30:27 +00001164
1165 ret = rds_cong_wait(conn->c_fcong, dport, 1, NULL);
1166 if (ret)
1167 goto out;
1168
1169 spin_lock_irqsave(&conn->c_lock, flags);
1170 list_add_tail(&rm->m_conn_item, &conn->c_send_queue);
1171 set_bit(RDS_MSG_ON_CONN, &rm->m_flags);
1172 rds_message_addref(rm);
1173 rm->m_inc.i_conn = conn;
1174
1175 rds_message_populate_header(&rm->m_inc.i_hdr, 0, dport,
1176 conn->c_next_tx_seq);
1177 conn->c_next_tx_seq++;
1178 spin_unlock_irqrestore(&conn->c_lock, flags);
1179
1180 rds_stats_inc(s_send_queued);
1181 rds_stats_inc(s_send_pong);
1182
Andy Groveracfcd4d2010-03-31 18:56:25 -07001183 if (!test_bit(RDS_LL_SEND_FULL, &conn->c_flags))
jeff.liu5175a5e2012-10-08 18:57:27 +00001184 queue_delayed_work(rds_wq, &conn->c_send_w, 0);
Andy Groveracfcd4d2010-03-31 18:56:25 -07001185
Andy Grover5c115592009-02-24 15:30:27 +00001186 rds_message_put(rm);
1187 return 0;
1188
1189out:
1190 if (rm)
1191 rds_message_put(rm);
1192 return ret;
1193}