Commit | Line | Data |
---|---|---|
5c115590 AG |
1 | /* |
2 | * Copyright (c) 2006 Oracle. All rights reserved. | |
3 | * | |
4 | * This software is available to you under a choice of one of two | |
5 | * licenses. You may choose to be licensed under the terms of the GNU | |
6 | * General Public License (GPL) Version 2, available from the file | |
7 | * COPYING in the main directory of this source tree, or the | |
8 | * OpenIB.org BSD license below: | |
9 | * | |
10 | * Redistribution and use in source and binary forms, with or | |
11 | * without modification, are permitted provided that the following | |
12 | * conditions are met: | |
13 | * | |
14 | * - Redistributions of source code must retain the above | |
15 | * copyright notice, this list of conditions and the following | |
16 | * disclaimer. | |
17 | * | |
18 | * - Redistributions in binary form must reproduce the above | |
19 | * copyright notice, this list of conditions and the following | |
20 | * disclaimer in the documentation and/or other materials | |
21 | * provided with the distribution. | |
22 | * | |
23 | * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, | |
24 | * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF | |
25 | * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND | |
26 | * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS | |
27 | * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN | |
28 | * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN | |
29 | * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE | |
30 | * SOFTWARE. | |
31 | * | |
32 | */ | |
33 | #include <linux/kernel.h> | |
5a0e3ad6 | 34 | #include <linux/gfp.h> |
5c115590 AG |
35 | #include <net/sock.h> |
36 | #include <linux/in.h> | |
37 | #include <linux/list.h> | |
38 | ||
39 | #include "rds.h" | |
5c115590 AG |
40 | |
41 | /* When transmitting messages in rds_send_xmit, we need to emerge from | |
42 | * time to time and briefly release the CPU. Otherwise the softlock watchdog | |
43 | * will kick our shin. | |
44 | * Also, it seems fairer to not let one busy connection stall all the | |
45 | * others. | |
46 | * | |
47 | * send_batch_count is the number of times we'll loop in send_xmit. Setting | |
48 | * it to 0 will restore the old behavior (where we looped until we had | |
49 | * drained the queue). | |
50 | */ | |
51 | static int send_batch_count = 64; | |
52 | module_param(send_batch_count, int, 0444); | |
53 | MODULE_PARM_DESC(send_batch_count, " batch factor when working the send queue"); | |
54 | ||
55 | /* | |
56 | * Reset the send state. Caller must hold c_send_lock when calling here. | |
57 | */ | |
58 | void rds_send_reset(struct rds_connection *conn) | |
59 | { | |
60 | struct rds_message *rm, *tmp; | |
61 | unsigned long flags; | |
62 | ||
63 | if (conn->c_xmit_rm) { | |
64 | /* Tell the user the RDMA op is no longer mapped by the | |
65 | * transport. This isn't entirely true (it's flushed out | |
66 | * independently) but as the connection is down, there's | |
67 | * no ongoing RDMA to/from that memory */ | |
68 | rds_message_unmapped(conn->c_xmit_rm); | |
69 | rds_message_put(conn->c_xmit_rm); | |
70 | conn->c_xmit_rm = NULL; | |
71 | } | |
72 | conn->c_xmit_sg = 0; | |
73 | conn->c_xmit_hdr_off = 0; | |
74 | conn->c_xmit_data_off = 0; | |
15133f6e | 75 | conn->c_xmit_atomic_sent = 0; |
5b2366bd AG |
76 | conn->c_xmit_rdma_sent = 0; |
77 | conn->c_xmit_data_sent = 0; | |
5c115590 AG |
78 | |
79 | conn->c_map_queued = 0; | |
80 | ||
81 | conn->c_unacked_packets = rds_sysctl_max_unacked_packets; | |
82 | conn->c_unacked_bytes = rds_sysctl_max_unacked_bytes; | |
83 | ||
84 | /* Mark messages as retransmissions, and move them to the send q */ | |
85 | spin_lock_irqsave(&conn->c_lock, flags); | |
86 | list_for_each_entry_safe(rm, tmp, &conn->c_retrans, m_conn_item) { | |
87 | set_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags); | |
88 | set_bit(RDS_MSG_RETRANSMITTED, &rm->m_flags); | |
89 | } | |
90 | list_splice_init(&conn->c_retrans, &conn->c_send_queue); | |
91 | spin_unlock_irqrestore(&conn->c_lock, flags); | |
92 | } | |
93 | ||
94 | /* | |
95 | * We're making the concious trade-off here to only send one message | |
96 | * down the connection at a time. | |
97 | * Pro: | |
98 | * - tx queueing is a simple fifo list | |
99 | * - reassembly is optional and easily done by transports per conn | |
100 | * - no per flow rx lookup at all, straight to the socket | |
101 | * - less per-frag memory and wire overhead | |
102 | * Con: | |
103 | * - queued acks can be delayed behind large messages | |
104 | * Depends: | |
105 | * - small message latency is higher behind queued large messages | |
106 | * - large message latency isn't starved by intervening small sends | |
107 | */ | |
108 | int rds_send_xmit(struct rds_connection *conn) | |
109 | { | |
110 | struct rds_message *rm; | |
111 | unsigned long flags; | |
112 | unsigned int tmp; | |
113 | unsigned int send_quota = send_batch_count; | |
114 | struct scatterlist *sg; | |
115 | int ret = 0; | |
116 | int was_empty = 0; | |
117 | LIST_HEAD(to_be_dropped); | |
118 | ||
049ee3f5 AG |
119 | if (!rds_conn_up(conn)) |
120 | goto out; | |
121 | ||
5c115590 AG |
122 | /* |
123 | * sendmsg calls here after having queued its message on the send | |
124 | * queue. We only have one task feeding the connection at a time. If | |
125 | * another thread is already feeding the queue then we back off. This | |
126 | * avoids blocking the caller and trading per-connection data between | |
127 | * caches per message. | |
5c115590 | 128 | */ |
049ee3f5 AG |
129 | if (!spin_trylock_irqsave(&conn->c_send_lock, flags)) { |
130 | rds_stats_inc(s_send_lock_contention); | |
5c115590 AG |
131 | ret = -ENOMEM; |
132 | goto out; | |
133 | } | |
134 | ||
135 | if (conn->c_trans->xmit_prepare) | |
136 | conn->c_trans->xmit_prepare(conn); | |
137 | ||
138 | /* | |
139 | * spin trying to push headers and data down the connection until | |
5b2366bd | 140 | * the connection doesn't make forward progress. |
5c115590 AG |
141 | */ |
142 | while (--send_quota) { | |
5c115590 | 143 | |
5c115590 | 144 | rm = conn->c_xmit_rm; |
5c115590 | 145 | |
5b2366bd AG |
146 | /* |
147 | * If between sending messages, we can send a pending congestion | |
148 | * map update. | |
5c115590 | 149 | */ |
8690bfa1 | 150 | if (!rm && test_and_clear_bit(0, &conn->c_map_queued)) { |
77dd550e AG |
151 | rm = rds_cong_update_alloc(conn); |
152 | if (IS_ERR(rm)) { | |
153 | ret = PTR_ERR(rm); | |
154 | break; | |
5b2366bd | 155 | } |
77dd550e AG |
156 | rm->data.op_active = 1; |
157 | ||
158 | conn->c_xmit_rm = rm; | |
5c115590 AG |
159 | } |
160 | ||
161 | /* | |
5b2366bd | 162 | * If not already working on one, grab the next message. |
5c115590 AG |
163 | * |
164 | * c_xmit_rm holds a ref while we're sending this message down | |
165 | * the connction. We can use this ref while holding the | |
166 | * send_sem.. rds_send_reset() is serialized with it. | |
167 | */ | |
8690bfa1 | 168 | if (!rm) { |
5c115590 AG |
169 | unsigned int len; |
170 | ||
2ad8099b | 171 | spin_lock(&conn->c_lock); |
5c115590 AG |
172 | |
173 | if (!list_empty(&conn->c_send_queue)) { | |
174 | rm = list_entry(conn->c_send_queue.next, | |
175 | struct rds_message, | |
176 | m_conn_item); | |
177 | rds_message_addref(rm); | |
178 | ||
179 | /* | |
180 | * Move the message from the send queue to the retransmit | |
181 | * list right away. | |
182 | */ | |
183 | list_move_tail(&rm->m_conn_item, &conn->c_retrans); | |
184 | } | |
185 | ||
2ad8099b | 186 | spin_unlock(&conn->c_lock); |
5c115590 | 187 | |
8690bfa1 | 188 | if (!rm) { |
5c115590 AG |
189 | was_empty = 1; |
190 | break; | |
191 | } | |
192 | ||
193 | /* Unfortunately, the way Infiniband deals with | |
194 | * RDMA to a bad MR key is by moving the entire | |
195 | * queue pair to error state. We cold possibly | |
196 | * recover from that, but right now we drop the | |
197 | * connection. | |
198 | * Therefore, we never retransmit messages with RDMA ops. | |
199 | */ | |
f8b3aaf2 | 200 | if (rm->rdma.op_active && |
f64f9e71 | 201 | test_bit(RDS_MSG_RETRANSMITTED, &rm->m_flags)) { |
2ad8099b | 202 | spin_lock(&conn->c_lock); |
5c115590 AG |
203 | if (test_and_clear_bit(RDS_MSG_ON_CONN, &rm->m_flags)) |
204 | list_move(&rm->m_conn_item, &to_be_dropped); | |
2ad8099b | 205 | spin_unlock(&conn->c_lock); |
5c115590 AG |
206 | continue; |
207 | } | |
208 | ||
209 | /* Require an ACK every once in a while */ | |
210 | len = ntohl(rm->m_inc.i_hdr.h_len); | |
f64f9e71 JP |
211 | if (conn->c_unacked_packets == 0 || |
212 | conn->c_unacked_bytes < len) { | |
5c115590 AG |
213 | __set_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags); |
214 | ||
215 | conn->c_unacked_packets = rds_sysctl_max_unacked_packets; | |
216 | conn->c_unacked_bytes = rds_sysctl_max_unacked_bytes; | |
217 | rds_stats_inc(s_send_ack_required); | |
218 | } else { | |
219 | conn->c_unacked_bytes -= len; | |
220 | conn->c_unacked_packets--; | |
221 | } | |
222 | ||
223 | conn->c_xmit_rm = rm; | |
224 | } | |
225 | ||
2c3a5f9a AG |
226 | /* The transport either sends the whole rdma or none of it */ |
227 | if (rm->rdma.op_active && !conn->c_xmit_rdma_sent) { | |
ff3d7d36 | 228 | rm->m_final_op = &rm->rdma; |
2c3a5f9a | 229 | ret = conn->c_trans->xmit_rdma(conn, &rm->rdma); |
15133f6e AG |
230 | if (ret) |
231 | break; | |
2c3a5f9a AG |
232 | conn->c_xmit_rdma_sent = 1; |
233 | ||
15133f6e AG |
234 | /* The transport owns the mapped memory for now. |
235 | * You can't unmap it while it's on the send queue */ | |
236 | set_bit(RDS_MSG_MAPPED, &rm->m_flags); | |
237 | } | |
238 | ||
2c3a5f9a | 239 | if (rm->atomic.op_active && !conn->c_xmit_atomic_sent) { |
ff3d7d36 AG |
240 | rm->m_final_op = &rm->atomic; |
241 | ret = conn->c_trans->xmit_atomic(conn, &rm->atomic); | |
5c115590 AG |
242 | if (ret) |
243 | break; | |
2c3a5f9a | 244 | conn->c_xmit_atomic_sent = 1; |
ff3d7d36 | 245 | |
5c115590 AG |
246 | /* The transport owns the mapped memory for now. |
247 | * You can't unmap it while it's on the send queue */ | |
248 | set_bit(RDS_MSG_MAPPED, &rm->m_flags); | |
249 | } | |
250 | ||
2c3a5f9a AG |
251 | /* |
252 | * A number of cases require an RDS header to be sent | |
253 | * even if there is no data. | |
254 | * We permit 0-byte sends; rds-ping depends on this. | |
255 | * However, if there are exclusively attached silent ops, | |
256 | * we skip the hdr/data send, to enable silent operation. | |
257 | */ | |
258 | if (rm->data.op_nents == 0) { | |
259 | int ops_present; | |
260 | int all_ops_are_silent = 1; | |
261 | ||
262 | ops_present = (rm->atomic.op_active || rm->rdma.op_active); | |
263 | if (rm->atomic.op_active && !rm->atomic.op_silent) | |
264 | all_ops_are_silent = 0; | |
265 | if (rm->rdma.op_active && !rm->rdma.op_silent) | |
266 | all_ops_are_silent = 0; | |
267 | ||
268 | if (ops_present && all_ops_are_silent | |
269 | && !rm->m_rdma_cookie) | |
270 | rm->data.op_active = 0; | |
271 | } | |
272 | ||
5b2366bd | 273 | if (rm->data.op_active && !conn->c_xmit_data_sent) { |
ff3d7d36 | 274 | rm->m_final_op = &rm->data; |
5c115590 AG |
275 | ret = conn->c_trans->xmit(conn, rm, |
276 | conn->c_xmit_hdr_off, | |
277 | conn->c_xmit_sg, | |
278 | conn->c_xmit_data_off); | |
279 | if (ret <= 0) | |
280 | break; | |
281 | ||
282 | if (conn->c_xmit_hdr_off < sizeof(struct rds_header)) { | |
283 | tmp = min_t(int, ret, | |
284 | sizeof(struct rds_header) - | |
285 | conn->c_xmit_hdr_off); | |
286 | conn->c_xmit_hdr_off += tmp; | |
287 | ret -= tmp; | |
288 | } | |
289 | ||
6c7cc6e4 | 290 | sg = &rm->data.op_sg[conn->c_xmit_sg]; |
5c115590 AG |
291 | while (ret) { |
292 | tmp = min_t(int, ret, sg->length - | |
293 | conn->c_xmit_data_off); | |
294 | conn->c_xmit_data_off += tmp; | |
295 | ret -= tmp; | |
296 | if (conn->c_xmit_data_off == sg->length) { | |
297 | conn->c_xmit_data_off = 0; | |
298 | sg++; | |
299 | conn->c_xmit_sg++; | |
300 | BUG_ON(ret != 0 && | |
6c7cc6e4 | 301 | conn->c_xmit_sg == rm->data.op_nents); |
5c115590 AG |
302 | } |
303 | } | |
5b2366bd AG |
304 | |
305 | if (conn->c_xmit_hdr_off == sizeof(struct rds_header) && | |
306 | (conn->c_xmit_sg == rm->data.op_nents)) | |
307 | conn->c_xmit_data_sent = 1; | |
308 | } | |
309 | ||
310 | /* | |
311 | * A rm will only take multiple times through this loop | |
312 | * if there is a data op. Thus, if the data is sent (or there was | |
313 | * none), then we're done with the rm. | |
314 | */ | |
315 | if (!rm->data.op_active || conn->c_xmit_data_sent) { | |
316 | conn->c_xmit_rm = NULL; | |
317 | conn->c_xmit_sg = 0; | |
318 | conn->c_xmit_hdr_off = 0; | |
319 | conn->c_xmit_data_off = 0; | |
320 | conn->c_xmit_rdma_sent = 0; | |
321 | conn->c_xmit_atomic_sent = 0; | |
322 | conn->c_xmit_data_sent = 0; | |
323 | ||
324 | rds_message_put(rm); | |
5c115590 AG |
325 | } |
326 | } | |
327 | ||
5c115590 AG |
328 | if (conn->c_trans->xmit_complete) |
329 | conn->c_trans->xmit_complete(conn); | |
330 | ||
331 | /* | |
332 | * We might be racing with another sender who queued a message but | |
333 | * backed off on noticing that we held the c_send_lock. If we check | |
334 | * for queued messages after dropping the sem then either we'll | |
335 | * see the queued message or the queuer will get the sem. If we | |
336 | * notice the queued message then we trigger an immediate retry. | |
337 | * | |
338 | * We need to be careful only to do this when we stopped processing | |
339 | * the send queue because it was empty. It's the only way we | |
340 | * stop processing the loop when the transport hasn't taken | |
341 | * responsibility for forward progress. | |
342 | */ | |
049ee3f5 | 343 | spin_unlock_irqrestore(&conn->c_send_lock, flags); |
5c115590 | 344 | |
2ad8099b AG |
345 | /* Nuke any messages we decided not to retransmit. */ |
346 | if (!list_empty(&to_be_dropped)) { | |
347 | /* irqs on here, so we can put(), unlike above */ | |
348 | list_for_each_entry(rm, &to_be_dropped, m_conn_item) | |
349 | rds_message_put(rm); | |
350 | rds_send_remove_from_sock(&to_be_dropped, RDS_RDMA_DROPPED); | |
351 | } | |
352 | ||
5b2366bd | 353 | if (send_quota == 0 && !was_empty) { |
5c115590 AG |
354 | /* We exhausted the send quota, but there's work left to |
355 | * do. Return and (re-)schedule the send worker. | |
356 | */ | |
357 | ret = -EAGAIN; | |
358 | } | |
359 | ||
360 | if (ret == 0 && was_empty) { | |
361 | /* A simple bit test would be way faster than taking the | |
362 | * spin lock */ | |
363 | spin_lock_irqsave(&conn->c_lock, flags); | |
364 | if (!list_empty(&conn->c_send_queue)) { | |
049ee3f5 | 365 | rds_stats_inc(s_send_lock_queue_raced); |
5c115590 AG |
366 | ret = -EAGAIN; |
367 | } | |
368 | spin_unlock_irqrestore(&conn->c_lock, flags); | |
369 | } | |
370 | out: | |
371 | return ret; | |
372 | } | |
373 | ||
374 | static void rds_send_sndbuf_remove(struct rds_sock *rs, struct rds_message *rm) | |
375 | { | |
376 | u32 len = be32_to_cpu(rm->m_inc.i_hdr.h_len); | |
377 | ||
378 | assert_spin_locked(&rs->rs_lock); | |
379 | ||
380 | BUG_ON(rs->rs_snd_bytes < len); | |
381 | rs->rs_snd_bytes -= len; | |
382 | ||
383 | if (rs->rs_snd_bytes == 0) | |
384 | rds_stats_inc(s_send_queue_empty); | |
385 | } | |
386 | ||
387 | static inline int rds_send_is_acked(struct rds_message *rm, u64 ack, | |
388 | is_acked_func is_acked) | |
389 | { | |
390 | if (is_acked) | |
391 | return is_acked(rm, ack); | |
392 | return be64_to_cpu(rm->m_inc.i_hdr.h_sequence) <= ack; | |
393 | } | |
394 | ||
395 | /* | |
396 | * Returns true if there are no messages on the send and retransmit queues | |
397 | * which have a sequence number greater than or equal to the given sequence | |
398 | * number. | |
399 | */ | |
400 | int rds_send_acked_before(struct rds_connection *conn, u64 seq) | |
401 | { | |
402 | struct rds_message *rm, *tmp; | |
403 | int ret = 1; | |
404 | ||
405 | spin_lock(&conn->c_lock); | |
406 | ||
407 | list_for_each_entry_safe(rm, tmp, &conn->c_retrans, m_conn_item) { | |
408 | if (be64_to_cpu(rm->m_inc.i_hdr.h_sequence) < seq) | |
409 | ret = 0; | |
410 | break; | |
411 | } | |
412 | ||
413 | list_for_each_entry_safe(rm, tmp, &conn->c_send_queue, m_conn_item) { | |
414 | if (be64_to_cpu(rm->m_inc.i_hdr.h_sequence) < seq) | |
415 | ret = 0; | |
416 | break; | |
417 | } | |
418 | ||
419 | spin_unlock(&conn->c_lock); | |
420 | ||
421 | return ret; | |
422 | } | |
423 | ||
424 | /* | |
425 | * This is pretty similar to what happens below in the ACK | |
426 | * handling code - except that we call here as soon as we get | |
427 | * the IB send completion on the RDMA op and the accompanying | |
428 | * message. | |
429 | */ | |
430 | void rds_rdma_send_complete(struct rds_message *rm, int status) | |
431 | { | |
432 | struct rds_sock *rs = NULL; | |
f8b3aaf2 | 433 | struct rm_rdma_op *ro; |
5c115590 | 434 | struct rds_notifier *notifier; |
9de0864c | 435 | unsigned long flags; |
5c115590 | 436 | |
9de0864c | 437 | spin_lock_irqsave(&rm->m_rs_lock, flags); |
5c115590 | 438 | |
f8b3aaf2 | 439 | ro = &rm->rdma; |
f64f9e71 | 440 | if (test_bit(RDS_MSG_ON_SOCK, &rm->m_flags) && |
f8b3aaf2 AG |
441 | ro->op_active && ro->op_notify && ro->op_notifier) { |
442 | notifier = ro->op_notifier; | |
5c115590 AG |
443 | rs = rm->m_rs; |
444 | sock_hold(rds_rs_to_sk(rs)); | |
445 | ||
446 | notifier->n_status = status; | |
447 | spin_lock(&rs->rs_lock); | |
448 | list_add_tail(¬ifier->n_list, &rs->rs_notify_queue); | |
449 | spin_unlock(&rs->rs_lock); | |
450 | ||
f8b3aaf2 | 451 | ro->op_notifier = NULL; |
5c115590 AG |
452 | } |
453 | ||
9de0864c | 454 | spin_unlock_irqrestore(&rm->m_rs_lock, flags); |
5c115590 AG |
455 | |
456 | if (rs) { | |
457 | rds_wake_sk_sleep(rs); | |
458 | sock_put(rds_rs_to_sk(rs)); | |
459 | } | |
460 | } | |
616b757a | 461 | EXPORT_SYMBOL_GPL(rds_rdma_send_complete); |
5c115590 | 462 | |
15133f6e AG |
463 | /* |
464 | * Just like above, except looks at atomic op | |
465 | */ | |
466 | void rds_atomic_send_complete(struct rds_message *rm, int status) | |
467 | { | |
468 | struct rds_sock *rs = NULL; | |
469 | struct rm_atomic_op *ao; | |
470 | struct rds_notifier *notifier; | |
471 | ||
472 | spin_lock(&rm->m_rs_lock); | |
473 | ||
474 | ao = &rm->atomic; | |
475 | if (test_bit(RDS_MSG_ON_SOCK, &rm->m_flags) | |
476 | && ao->op_active && ao->op_notify && ao->op_notifier) { | |
477 | notifier = ao->op_notifier; | |
478 | rs = rm->m_rs; | |
479 | sock_hold(rds_rs_to_sk(rs)); | |
480 | ||
481 | notifier->n_status = status; | |
482 | spin_lock(&rs->rs_lock); | |
483 | list_add_tail(¬ifier->n_list, &rs->rs_notify_queue); | |
484 | spin_unlock(&rs->rs_lock); | |
485 | ||
486 | ao->op_notifier = NULL; | |
487 | } | |
488 | ||
489 | spin_unlock(&rm->m_rs_lock); | |
490 | ||
491 | if (rs) { | |
492 | rds_wake_sk_sleep(rs); | |
493 | sock_put(rds_rs_to_sk(rs)); | |
494 | } | |
495 | } | |
496 | EXPORT_SYMBOL_GPL(rds_atomic_send_complete); | |
497 | ||
5c115590 AG |
498 | /* |
499 | * This is the same as rds_rdma_send_complete except we | |
500 | * don't do any locking - we have all the ingredients (message, | |
501 | * socket, socket lock) and can just move the notifier. | |
502 | */ | |
503 | static inline void | |
940786eb | 504 | __rds_send_complete(struct rds_sock *rs, struct rds_message *rm, int status) |
5c115590 | 505 | { |
f8b3aaf2 | 506 | struct rm_rdma_op *ro; |
940786eb | 507 | struct rm_atomic_op *ao; |
5c115590 | 508 | |
f8b3aaf2 AG |
509 | ro = &rm->rdma; |
510 | if (ro->op_active && ro->op_notify && ro->op_notifier) { | |
511 | ro->op_notifier->n_status = status; | |
512 | list_add_tail(&ro->op_notifier->n_list, &rs->rs_notify_queue); | |
513 | ro->op_notifier = NULL; | |
5c115590 AG |
514 | } |
515 | ||
940786eb AG |
516 | ao = &rm->atomic; |
517 | if (ao->op_active && ao->op_notify && ao->op_notifier) { | |
518 | ao->op_notifier->n_status = status; | |
519 | list_add_tail(&ao->op_notifier->n_list, &rs->rs_notify_queue); | |
520 | ao->op_notifier = NULL; | |
521 | } | |
522 | ||
5c115590 AG |
523 | /* No need to wake the app - caller does this */ |
524 | } | |
525 | ||
526 | /* | |
527 | * This is called from the IB send completion when we detect | |
528 | * a RDMA operation that failed with remote access error. | |
529 | * So speed is not an issue here. | |
530 | */ | |
531 | struct rds_message *rds_send_get_message(struct rds_connection *conn, | |
f8b3aaf2 | 532 | struct rm_rdma_op *op) |
5c115590 AG |
533 | { |
534 | struct rds_message *rm, *tmp, *found = NULL; | |
535 | unsigned long flags; | |
536 | ||
537 | spin_lock_irqsave(&conn->c_lock, flags); | |
538 | ||
539 | list_for_each_entry_safe(rm, tmp, &conn->c_retrans, m_conn_item) { | |
f8b3aaf2 | 540 | if (&rm->rdma == op) { |
5c115590 AG |
541 | atomic_inc(&rm->m_refcount); |
542 | found = rm; | |
543 | goto out; | |
544 | } | |
545 | } | |
546 | ||
547 | list_for_each_entry_safe(rm, tmp, &conn->c_send_queue, m_conn_item) { | |
f8b3aaf2 | 548 | if (&rm->rdma == op) { |
5c115590 AG |
549 | atomic_inc(&rm->m_refcount); |
550 | found = rm; | |
551 | break; | |
552 | } | |
553 | } | |
554 | ||
555 | out: | |
556 | spin_unlock_irqrestore(&conn->c_lock, flags); | |
557 | ||
558 | return found; | |
559 | } | |
616b757a | 560 | EXPORT_SYMBOL_GPL(rds_send_get_message); |
5c115590 AG |
561 | |
562 | /* | |
563 | * This removes messages from the socket's list if they're on it. The list | |
564 | * argument must be private to the caller, we must be able to modify it | |
565 | * without locks. The messages must have a reference held for their | |
566 | * position on the list. This function will drop that reference after | |
567 | * removing the messages from the 'messages' list regardless of if it found | |
568 | * the messages on the socket list or not. | |
569 | */ | |
570 | void rds_send_remove_from_sock(struct list_head *messages, int status) | |
571 | { | |
561c7df6 | 572 | unsigned long flags; |
5c115590 AG |
573 | struct rds_sock *rs = NULL; |
574 | struct rds_message *rm; | |
575 | ||
5c115590 | 576 | while (!list_empty(messages)) { |
561c7df6 AG |
577 | int was_on_sock = 0; |
578 | ||
5c115590 AG |
579 | rm = list_entry(messages->next, struct rds_message, |
580 | m_conn_item); | |
581 | list_del_init(&rm->m_conn_item); | |
582 | ||
583 | /* | |
584 | * If we see this flag cleared then we're *sure* that someone | |
585 | * else beat us to removing it from the sock. If we race | |
586 | * with their flag update we'll get the lock and then really | |
587 | * see that the flag has been cleared. | |
588 | * | |
589 | * The message spinlock makes sure nobody clears rm->m_rs | |
590 | * while we're messing with it. It does not prevent the | |
591 | * message from being removed from the socket, though. | |
592 | */ | |
561c7df6 | 593 | spin_lock_irqsave(&rm->m_rs_lock, flags); |
5c115590 AG |
594 | if (!test_bit(RDS_MSG_ON_SOCK, &rm->m_flags)) |
595 | goto unlock_and_drop; | |
596 | ||
597 | if (rs != rm->m_rs) { | |
598 | if (rs) { | |
5c115590 AG |
599 | rds_wake_sk_sleep(rs); |
600 | sock_put(rds_rs_to_sk(rs)); | |
601 | } | |
602 | rs = rm->m_rs; | |
5c115590 AG |
603 | sock_hold(rds_rs_to_sk(rs)); |
604 | } | |
048c15e6 | 605 | spin_lock(&rs->rs_lock); |
5c115590 AG |
606 | |
607 | if (test_and_clear_bit(RDS_MSG_ON_SOCK, &rm->m_flags)) { | |
f8b3aaf2 | 608 | struct rm_rdma_op *ro = &rm->rdma; |
5c115590 AG |
609 | struct rds_notifier *notifier; |
610 | ||
611 | list_del_init(&rm->m_sock_item); | |
612 | rds_send_sndbuf_remove(rs, rm); | |
613 | ||
f8b3aaf2 AG |
614 | if (ro->op_active && ro->op_notifier && |
615 | (ro->op_notify || (ro->op_recverr && status))) { | |
616 | notifier = ro->op_notifier; | |
5c115590 AG |
617 | list_add_tail(¬ifier->n_list, |
618 | &rs->rs_notify_queue); | |
619 | if (!notifier->n_status) | |
620 | notifier->n_status = status; | |
f8b3aaf2 | 621 | rm->rdma.op_notifier = NULL; |
5c115590 | 622 | } |
561c7df6 | 623 | was_on_sock = 1; |
5c115590 AG |
624 | rm->m_rs = NULL; |
625 | } | |
048c15e6 | 626 | spin_unlock(&rs->rs_lock); |
5c115590 AG |
627 | |
628 | unlock_and_drop: | |
561c7df6 | 629 | spin_unlock_irqrestore(&rm->m_rs_lock, flags); |
5c115590 | 630 | rds_message_put(rm); |
561c7df6 AG |
631 | if (was_on_sock) |
632 | rds_message_put(rm); | |
5c115590 AG |
633 | } |
634 | ||
635 | if (rs) { | |
5c115590 AG |
636 | rds_wake_sk_sleep(rs); |
637 | sock_put(rds_rs_to_sk(rs)); | |
638 | } | |
5c115590 AG |
639 | } |
640 | ||
641 | /* | |
642 | * Transports call here when they've determined that the receiver queued | |
643 | * messages up to, and including, the given sequence number. Messages are | |
644 | * moved to the retrans queue when rds_send_xmit picks them off the send | |
645 | * queue. This means that in the TCP case, the message may not have been | |
646 | * assigned the m_ack_seq yet - but that's fine as long as tcp_is_acked | |
647 | * checks the RDS_MSG_HAS_ACK_SEQ bit. | |
648 | * | |
649 | * XXX It's not clear to me how this is safely serialized with socket | |
650 | * destruction. Maybe it should bail if it sees SOCK_DEAD. | |
651 | */ | |
652 | void rds_send_drop_acked(struct rds_connection *conn, u64 ack, | |
653 | is_acked_func is_acked) | |
654 | { | |
655 | struct rds_message *rm, *tmp; | |
656 | unsigned long flags; | |
657 | LIST_HEAD(list); | |
658 | ||
659 | spin_lock_irqsave(&conn->c_lock, flags); | |
660 | ||
661 | list_for_each_entry_safe(rm, tmp, &conn->c_retrans, m_conn_item) { | |
662 | if (!rds_send_is_acked(rm, ack, is_acked)) | |
663 | break; | |
664 | ||
665 | list_move(&rm->m_conn_item, &list); | |
666 | clear_bit(RDS_MSG_ON_CONN, &rm->m_flags); | |
667 | } | |
668 | ||
669 | /* order flag updates with spin locks */ | |
670 | if (!list_empty(&list)) | |
671 | smp_mb__after_clear_bit(); | |
672 | ||
673 | spin_unlock_irqrestore(&conn->c_lock, flags); | |
674 | ||
675 | /* now remove the messages from the sock list as needed */ | |
676 | rds_send_remove_from_sock(&list, RDS_RDMA_SUCCESS); | |
677 | } | |
616b757a | 678 | EXPORT_SYMBOL_GPL(rds_send_drop_acked); |
5c115590 AG |
679 | |
680 | void rds_send_drop_to(struct rds_sock *rs, struct sockaddr_in *dest) | |
681 | { | |
682 | struct rds_message *rm, *tmp; | |
683 | struct rds_connection *conn; | |
7c82eaf0 | 684 | unsigned long flags; |
5c115590 | 685 | LIST_HEAD(list); |
5c115590 AG |
686 | |
687 | /* get all the messages we're dropping under the rs lock */ | |
688 | spin_lock_irqsave(&rs->rs_lock, flags); | |
689 | ||
690 | list_for_each_entry_safe(rm, tmp, &rs->rs_send_queue, m_sock_item) { | |
691 | if (dest && (dest->sin_addr.s_addr != rm->m_daddr || | |
692 | dest->sin_port != rm->m_inc.i_hdr.h_dport)) | |
693 | continue; | |
694 | ||
5c115590 AG |
695 | list_move(&rm->m_sock_item, &list); |
696 | rds_send_sndbuf_remove(rs, rm); | |
697 | clear_bit(RDS_MSG_ON_SOCK, &rm->m_flags); | |
5c115590 AG |
698 | } |
699 | ||
700 | /* order flag updates with the rs lock */ | |
7c82eaf0 | 701 | smp_mb__after_clear_bit(); |
5c115590 AG |
702 | |
703 | spin_unlock_irqrestore(&rs->rs_lock, flags); | |
704 | ||
7c82eaf0 AG |
705 | if (list_empty(&list)) |
706 | return; | |
5c115590 | 707 | |
7c82eaf0 | 708 | /* Remove the messages from the conn */ |
5c115590 | 709 | list_for_each_entry(rm, &list, m_sock_item) { |
7c82eaf0 AG |
710 | |
711 | conn = rm->m_inc.i_conn; | |
5c115590 | 712 | |
9de0864c | 713 | spin_lock_irqsave(&conn->c_lock, flags); |
5c115590 | 714 | /* |
7c82eaf0 AG |
715 | * Maybe someone else beat us to removing rm from the conn. |
716 | * If we race with their flag update we'll get the lock and | |
717 | * then really see that the flag has been cleared. | |
5c115590 | 718 | */ |
7c82eaf0 AG |
719 | if (!test_and_clear_bit(RDS_MSG_ON_CONN, &rm->m_flags)) { |
720 | spin_unlock_irqrestore(&conn->c_lock, flags); | |
5c115590 | 721 | continue; |
5c115590 | 722 | } |
9de0864c AG |
723 | list_del_init(&rm->m_conn_item); |
724 | spin_unlock_irqrestore(&conn->c_lock, flags); | |
5c115590 | 725 | |
7c82eaf0 AG |
726 | /* |
727 | * Couldn't grab m_rs_lock in top loop (lock ordering), | |
728 | * but we can now. | |
729 | */ | |
9de0864c | 730 | spin_lock_irqsave(&rm->m_rs_lock, flags); |
5c115590 | 731 | |
7c82eaf0 | 732 | spin_lock(&rs->rs_lock); |
940786eb | 733 | __rds_send_complete(rs, rm, RDS_RDMA_CANCELED); |
7c82eaf0 AG |
734 | spin_unlock(&rs->rs_lock); |
735 | ||
736 | rm->m_rs = NULL; | |
9de0864c | 737 | spin_unlock_irqrestore(&rm->m_rs_lock, flags); |
7c82eaf0 | 738 | |
7c82eaf0 | 739 | rds_message_put(rm); |
7c82eaf0 | 740 | } |
5c115590 | 741 | |
7c82eaf0 | 742 | rds_wake_sk_sleep(rs); |
550a8002 | 743 | |
5c115590 AG |
744 | while (!list_empty(&list)) { |
745 | rm = list_entry(list.next, struct rds_message, m_sock_item); | |
746 | list_del_init(&rm->m_sock_item); | |
747 | ||
748 | rds_message_wait(rm); | |
749 | rds_message_put(rm); | |
750 | } | |
751 | } | |
752 | ||
753 | /* | |
754 | * we only want this to fire once so we use the callers 'queued'. It's | |
755 | * possible that another thread can race with us and remove the | |
756 | * message from the flow with RDS_CANCEL_SENT_TO. | |
757 | */ | |
758 | static int rds_send_queue_rm(struct rds_sock *rs, struct rds_connection *conn, | |
759 | struct rds_message *rm, __be16 sport, | |
760 | __be16 dport, int *queued) | |
761 | { | |
762 | unsigned long flags; | |
763 | u32 len; | |
764 | ||
765 | if (*queued) | |
766 | goto out; | |
767 | ||
768 | len = be32_to_cpu(rm->m_inc.i_hdr.h_len); | |
769 | ||
770 | /* this is the only place which holds both the socket's rs_lock | |
771 | * and the connection's c_lock */ | |
772 | spin_lock_irqsave(&rs->rs_lock, flags); | |
773 | ||
774 | /* | |
775 | * If there is a little space in sndbuf, we don't queue anything, | |
776 | * and userspace gets -EAGAIN. But poll() indicates there's send | |
777 | * room. This can lead to bad behavior (spinning) if snd_bytes isn't | |
778 | * freed up by incoming acks. So we check the *old* value of | |
779 | * rs_snd_bytes here to allow the last msg to exceed the buffer, | |
780 | * and poll() now knows no more data can be sent. | |
781 | */ | |
782 | if (rs->rs_snd_bytes < rds_sk_sndbuf(rs)) { | |
783 | rs->rs_snd_bytes += len; | |
784 | ||
785 | /* let recv side know we are close to send space exhaustion. | |
786 | * This is probably not the optimal way to do it, as this | |
787 | * means we set the flag on *all* messages as soon as our | |
788 | * throughput hits a certain threshold. | |
789 | */ | |
790 | if (rs->rs_snd_bytes >= rds_sk_sndbuf(rs) / 2) | |
791 | __set_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags); | |
792 | ||
793 | list_add_tail(&rm->m_sock_item, &rs->rs_send_queue); | |
794 | set_bit(RDS_MSG_ON_SOCK, &rm->m_flags); | |
795 | rds_message_addref(rm); | |
796 | rm->m_rs = rs; | |
797 | ||
798 | /* The code ordering is a little weird, but we're | |
799 | trying to minimize the time we hold c_lock */ | |
800 | rds_message_populate_header(&rm->m_inc.i_hdr, sport, dport, 0); | |
801 | rm->m_inc.i_conn = conn; | |
802 | rds_message_addref(rm); | |
803 | ||
804 | spin_lock(&conn->c_lock); | |
805 | rm->m_inc.i_hdr.h_sequence = cpu_to_be64(conn->c_next_tx_seq++); | |
806 | list_add_tail(&rm->m_conn_item, &conn->c_send_queue); | |
807 | set_bit(RDS_MSG_ON_CONN, &rm->m_flags); | |
808 | spin_unlock(&conn->c_lock); | |
809 | ||
810 | rdsdebug("queued msg %p len %d, rs %p bytes %d seq %llu\n", | |
811 | rm, len, rs, rs->rs_snd_bytes, | |
812 | (unsigned long long)be64_to_cpu(rm->m_inc.i_hdr.h_sequence)); | |
813 | ||
814 | *queued = 1; | |
815 | } | |
816 | ||
817 | spin_unlock_irqrestore(&rs->rs_lock, flags); | |
818 | out: | |
819 | return *queued; | |
820 | } | |
821 | ||
fc445084 AG |
822 | /* |
823 | * rds_message is getting to be quite complicated, and we'd like to allocate | |
824 | * it all in one go. This figures out how big it needs to be up front. | |
825 | */ | |
826 | static int rds_rm_size(struct msghdr *msg, int data_len) | |
827 | { | |
ff87e97a | 828 | struct cmsghdr *cmsg; |
fc445084 | 829 | int size = 0; |
aa0a4ef4 | 830 | int cmsg_groups = 0; |
ff87e97a AG |
831 | int retval; |
832 | ||
833 | for (cmsg = CMSG_FIRSTHDR(msg); cmsg; cmsg = CMSG_NXTHDR(msg, cmsg)) { | |
834 | if (!CMSG_OK(msg, cmsg)) | |
835 | return -EINVAL; | |
836 | ||
837 | if (cmsg->cmsg_level != SOL_RDS) | |
838 | continue; | |
839 | ||
840 | switch (cmsg->cmsg_type) { | |
841 | case RDS_CMSG_RDMA_ARGS: | |
aa0a4ef4 | 842 | cmsg_groups |= 1; |
ff87e97a AG |
843 | retval = rds_rdma_extra_size(CMSG_DATA(cmsg)); |
844 | if (retval < 0) | |
845 | return retval; | |
846 | size += retval; | |
aa0a4ef4 | 847 | |
ff87e97a AG |
848 | break; |
849 | ||
850 | case RDS_CMSG_RDMA_DEST: | |
851 | case RDS_CMSG_RDMA_MAP: | |
aa0a4ef4 | 852 | cmsg_groups |= 2; |
ff87e97a AG |
853 | /* these are valid but do no add any size */ |
854 | break; | |
855 | ||
15133f6e AG |
856 | case RDS_CMSG_ATOMIC_CSWP: |
857 | case RDS_CMSG_ATOMIC_FADD: | |
aa0a4ef4 | 858 | cmsg_groups |= 1; |
15133f6e AG |
859 | size += sizeof(struct scatterlist); |
860 | break; | |
861 | ||
ff87e97a AG |
862 | default: |
863 | return -EINVAL; | |
864 | } | |
865 | ||
866 | } | |
fc445084 | 867 | |
ff87e97a | 868 | size += ceil(data_len, PAGE_SIZE) * sizeof(struct scatterlist); |
fc445084 | 869 | |
aa0a4ef4 AG |
870 | /* Ensure (DEST, MAP) are never used with (ARGS, ATOMIC) */ |
871 | if (cmsg_groups == 3) | |
872 | return -EINVAL; | |
873 | ||
fc445084 AG |
874 | return size; |
875 | } | |
876 | ||
5c115590 AG |
877 | static int rds_cmsg_send(struct rds_sock *rs, struct rds_message *rm, |
878 | struct msghdr *msg, int *allocated_mr) | |
879 | { | |
880 | struct cmsghdr *cmsg; | |
881 | int ret = 0; | |
882 | ||
883 | for (cmsg = CMSG_FIRSTHDR(msg); cmsg; cmsg = CMSG_NXTHDR(msg, cmsg)) { | |
884 | if (!CMSG_OK(msg, cmsg)) | |
885 | return -EINVAL; | |
886 | ||
887 | if (cmsg->cmsg_level != SOL_RDS) | |
888 | continue; | |
889 | ||
890 | /* As a side effect, RDMA_DEST and RDMA_MAP will set | |
15133f6e | 891 | * rm->rdma.m_rdma_cookie and rm->rdma.m_rdma_mr. |
5c115590 AG |
892 | */ |
893 | switch (cmsg->cmsg_type) { | |
894 | case RDS_CMSG_RDMA_ARGS: | |
895 | ret = rds_cmsg_rdma_args(rs, rm, cmsg); | |
896 | break; | |
897 | ||
898 | case RDS_CMSG_RDMA_DEST: | |
899 | ret = rds_cmsg_rdma_dest(rs, rm, cmsg); | |
900 | break; | |
901 | ||
902 | case RDS_CMSG_RDMA_MAP: | |
903 | ret = rds_cmsg_rdma_map(rs, rm, cmsg); | |
904 | if (!ret) | |
905 | *allocated_mr = 1; | |
906 | break; | |
15133f6e AG |
907 | case RDS_CMSG_ATOMIC_CSWP: |
908 | case RDS_CMSG_ATOMIC_FADD: | |
909 | ret = rds_cmsg_atomic(rs, rm, cmsg); | |
910 | break; | |
5c115590 AG |
911 | |
912 | default: | |
913 | return -EINVAL; | |
914 | } | |
915 | ||
916 | if (ret) | |
917 | break; | |
918 | } | |
919 | ||
920 | return ret; | |
921 | } | |
922 | ||
923 | int rds_sendmsg(struct kiocb *iocb, struct socket *sock, struct msghdr *msg, | |
924 | size_t payload_len) | |
925 | { | |
926 | struct sock *sk = sock->sk; | |
927 | struct rds_sock *rs = rds_sk_to_rs(sk); | |
928 | struct sockaddr_in *usin = (struct sockaddr_in *)msg->msg_name; | |
929 | __be32 daddr; | |
930 | __be16 dport; | |
931 | struct rds_message *rm = NULL; | |
932 | struct rds_connection *conn; | |
933 | int ret = 0; | |
934 | int queued = 0, allocated_mr = 0; | |
935 | int nonblock = msg->msg_flags & MSG_DONTWAIT; | |
1123fd73 | 936 | long timeo = sock_sndtimeo(sk, nonblock); |
5c115590 AG |
937 | |
938 | /* Mirror Linux UDP mirror of BSD error message compatibility */ | |
939 | /* XXX: Perhaps MSG_MORE someday */ | |
940 | if (msg->msg_flags & ~(MSG_DONTWAIT | MSG_CMSG_COMPAT)) { | |
941 | printk(KERN_INFO "msg_flags 0x%08X\n", msg->msg_flags); | |
942 | ret = -EOPNOTSUPP; | |
943 | goto out; | |
944 | } | |
945 | ||
946 | if (msg->msg_namelen) { | |
947 | /* XXX fail non-unicast destination IPs? */ | |
948 | if (msg->msg_namelen < sizeof(*usin) || usin->sin_family != AF_INET) { | |
949 | ret = -EINVAL; | |
950 | goto out; | |
951 | } | |
952 | daddr = usin->sin_addr.s_addr; | |
953 | dport = usin->sin_port; | |
954 | } else { | |
955 | /* We only care about consistency with ->connect() */ | |
956 | lock_sock(sk); | |
957 | daddr = rs->rs_conn_addr; | |
958 | dport = rs->rs_conn_port; | |
959 | release_sock(sk); | |
960 | } | |
961 | ||
962 | /* racing with another thread binding seems ok here */ | |
963 | if (daddr == 0 || rs->rs_bound_addr == 0) { | |
964 | ret = -ENOTCONN; /* XXX not a great errno */ | |
965 | goto out; | |
966 | } | |
967 | ||
fc445084 AG |
968 | /* size of rm including all sgs */ |
969 | ret = rds_rm_size(msg, payload_len); | |
970 | if (ret < 0) | |
971 | goto out; | |
972 | ||
973 | rm = rds_message_alloc(ret, GFP_KERNEL); | |
974 | if (!rm) { | |
975 | ret = -ENOMEM; | |
5c115590 AG |
976 | goto out; |
977 | } | |
978 | ||
372cd7de AG |
979 | /* Attach data to the rm */ |
980 | if (payload_len) { | |
981 | rm->data.op_sg = rds_message_alloc_sgs(rm, ceil(payload_len, PAGE_SIZE)); | |
982 | ret = rds_message_copy_from_user(rm, msg->msg_iov, payload_len); | |
983 | if (ret) | |
984 | goto out; | |
985 | } | |
986 | rm->data.op_active = 1; | |
fc445084 | 987 | |
5c115590 AG |
988 | rm->m_daddr = daddr; |
989 | ||
5c115590 AG |
990 | /* rds_conn_create has a spinlock that runs with IRQ off. |
991 | * Caching the conn in the socket helps a lot. */ | |
992 | if (rs->rs_conn && rs->rs_conn->c_faddr == daddr) | |
993 | conn = rs->rs_conn; | |
994 | else { | |
995 | conn = rds_conn_create_outgoing(rs->rs_bound_addr, daddr, | |
996 | rs->rs_transport, | |
997 | sock->sk->sk_allocation); | |
998 | if (IS_ERR(conn)) { | |
999 | ret = PTR_ERR(conn); | |
1000 | goto out; | |
1001 | } | |
1002 | rs->rs_conn = conn; | |
1003 | } | |
1004 | ||
49f69691 AG |
1005 | /* Parse any control messages the user may have included. */ |
1006 | ret = rds_cmsg_send(rs, rm, msg, &allocated_mr); | |
1007 | if (ret) | |
1008 | goto out; | |
1009 | ||
2c3a5f9a | 1010 | if (rm->rdma.op_active && !conn->c_trans->xmit_rdma) { |
5c115590 AG |
1011 | if (printk_ratelimit()) |
1012 | printk(KERN_NOTICE "rdma_op %p conn xmit_rdma %p\n", | |
f8b3aaf2 | 1013 | &rm->rdma, conn->c_trans->xmit_rdma); |
15133f6e AG |
1014 | ret = -EOPNOTSUPP; |
1015 | goto out; | |
1016 | } | |
1017 | ||
1018 | if (rm->atomic.op_active && !conn->c_trans->xmit_atomic) { | |
1019 | if (printk_ratelimit()) | |
1020 | printk(KERN_NOTICE "atomic_op %p conn xmit_atomic %p\n", | |
1021 | &rm->atomic, conn->c_trans->xmit_atomic); | |
5c115590 AG |
1022 | ret = -EOPNOTSUPP; |
1023 | goto out; | |
1024 | } | |
1025 | ||
1026 | /* If the connection is down, trigger a connect. We may | |
1027 | * have scheduled a delayed reconnect however - in this case | |
1028 | * we should not interfere. | |
1029 | */ | |
f64f9e71 JP |
1030 | if (rds_conn_state(conn) == RDS_CONN_DOWN && |
1031 | !test_and_set_bit(RDS_RECONNECT_PENDING, &conn->c_flags)) | |
5c115590 AG |
1032 | queue_delayed_work(rds_wq, &conn->c_conn_w, 0); |
1033 | ||
1034 | ret = rds_cong_wait(conn->c_fcong, dport, nonblock, rs); | |
b98ba52f AG |
1035 | if (ret) { |
1036 | rs->rs_seen_congestion = 1; | |
5c115590 | 1037 | goto out; |
b98ba52f | 1038 | } |
5c115590 AG |
1039 | |
1040 | while (!rds_send_queue_rm(rs, conn, rm, rs->rs_bound_port, | |
1041 | dport, &queued)) { | |
1042 | rds_stats_inc(s_send_queue_full); | |
1043 | /* XXX make sure this is reasonable */ | |
1044 | if (payload_len > rds_sk_sndbuf(rs)) { | |
1045 | ret = -EMSGSIZE; | |
1046 | goto out; | |
1047 | } | |
1048 | if (nonblock) { | |
1049 | ret = -EAGAIN; | |
1050 | goto out; | |
1051 | } | |
1052 | ||
aa395145 | 1053 | timeo = wait_event_interruptible_timeout(*sk_sleep(sk), |
5c115590 AG |
1054 | rds_send_queue_rm(rs, conn, rm, |
1055 | rs->rs_bound_port, | |
1056 | dport, | |
1057 | &queued), | |
1058 | timeo); | |
1059 | rdsdebug("sendmsg woke queued %d timeo %ld\n", queued, timeo); | |
1060 | if (timeo > 0 || timeo == MAX_SCHEDULE_TIMEOUT) | |
1061 | continue; | |
1062 | ||
1063 | ret = timeo; | |
1064 | if (ret == 0) | |
1065 | ret = -ETIMEDOUT; | |
1066 | goto out; | |
1067 | } | |
1068 | ||
1069 | /* | |
1070 | * By now we've committed to the send. We reuse rds_send_worker() | |
1071 | * to retry sends in the rds thread if the transport asks us to. | |
1072 | */ | |
1073 | rds_stats_inc(s_send_queued); | |
1074 | ||
1075 | if (!test_bit(RDS_LL_SEND_FULL, &conn->c_flags)) | |
a7d3a281 | 1076 | rds_send_xmit(conn); |
5c115590 AG |
1077 | |
1078 | rds_message_put(rm); | |
1079 | return payload_len; | |
1080 | ||
1081 | out: | |
1082 | /* If the user included a RDMA_MAP cmsg, we allocated a MR on the fly. | |
1083 | * If the sendmsg goes through, we keep the MR. If it fails with EAGAIN | |
1084 | * or in any other way, we need to destroy the MR again */ | |
1085 | if (allocated_mr) | |
1086 | rds_rdma_unuse(rs, rds_rdma_cookie_key(rm->m_rdma_cookie), 1); | |
1087 | ||
1088 | if (rm) | |
1089 | rds_message_put(rm); | |
1090 | return ret; | |
1091 | } | |
1092 | ||
1093 | /* | |
1094 | * Reply to a ping packet. | |
1095 | */ | |
1096 | int | |
1097 | rds_send_pong(struct rds_connection *conn, __be16 dport) | |
1098 | { | |
1099 | struct rds_message *rm; | |
1100 | unsigned long flags; | |
1101 | int ret = 0; | |
1102 | ||
1103 | rm = rds_message_alloc(0, GFP_ATOMIC); | |
8690bfa1 | 1104 | if (!rm) { |
5c115590 AG |
1105 | ret = -ENOMEM; |
1106 | goto out; | |
1107 | } | |
1108 | ||
1109 | rm->m_daddr = conn->c_faddr; | |
1110 | ||
1111 | /* If the connection is down, trigger a connect. We may | |
1112 | * have scheduled a delayed reconnect however - in this case | |
1113 | * we should not interfere. | |
1114 | */ | |
f64f9e71 JP |
1115 | if (rds_conn_state(conn) == RDS_CONN_DOWN && |
1116 | !test_and_set_bit(RDS_RECONNECT_PENDING, &conn->c_flags)) | |
5c115590 AG |
1117 | queue_delayed_work(rds_wq, &conn->c_conn_w, 0); |
1118 | ||
1119 | ret = rds_cong_wait(conn->c_fcong, dport, 1, NULL); | |
1120 | if (ret) | |
1121 | goto out; | |
1122 | ||
1123 | spin_lock_irqsave(&conn->c_lock, flags); | |
1124 | list_add_tail(&rm->m_conn_item, &conn->c_send_queue); | |
1125 | set_bit(RDS_MSG_ON_CONN, &rm->m_flags); | |
1126 | rds_message_addref(rm); | |
1127 | rm->m_inc.i_conn = conn; | |
1128 | ||
1129 | rds_message_populate_header(&rm->m_inc.i_hdr, 0, dport, | |
1130 | conn->c_next_tx_seq); | |
1131 | conn->c_next_tx_seq++; | |
1132 | spin_unlock_irqrestore(&conn->c_lock, flags); | |
1133 | ||
1134 | rds_stats_inc(s_send_queued); | |
1135 | rds_stats_inc(s_send_pong); | |
1136 | ||
1137 | queue_delayed_work(rds_wq, &conn->c_send_w, 0); | |
1138 | rds_message_put(rm); | |
1139 | return 0; | |
1140 | ||
1141 | out: | |
1142 | if (rm) | |
1143 | rds_message_put(rm); | |
1144 | return ret; | |
1145 | } |