]>
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> | |
d9b93842 | 34 | #include <linux/moduleparam.h> |
5a0e3ad6 | 35 | #include <linux/gfp.h> |
5c115590 AG |
36 | #include <net/sock.h> |
37 | #include <linux/in.h> | |
38 | #include <linux/list.h> | |
cb0a6056 | 39 | #include <linux/ratelimit.h> |
bc3b2d7f | 40 | #include <linux/export.h> |
5c115590 AG |
41 | |
42 | #include "rds.h" | |
5c115590 AG |
43 | |
44 | /* When transmitting messages in rds_send_xmit, we need to emerge from | |
45 | * time to time and briefly release the CPU. Otherwise the softlock watchdog | |
46 | * will kick our shin. | |
47 | * Also, it seems fairer to not let one busy connection stall all the | |
48 | * others. | |
49 | * | |
50 | * send_batch_count is the number of times we'll loop in send_xmit. Setting | |
51 | * it to 0 will restore the old behavior (where we looped until we had | |
52 | * drained the queue). | |
53 | */ | |
54 | static int send_batch_count = 64; | |
55 | module_param(send_batch_count, int, 0444); | |
56 | MODULE_PARM_DESC(send_batch_count, " batch factor when working the send queue"); | |
57 | ||
ff51bf84 | 58 | static void rds_send_remove_from_sock(struct list_head *messages, int status); |
59 | ||
5c115590 | 60 | /* |
0f4b1c7e ZB |
61 | * Reset the send state. Callers must ensure that this doesn't race with |
62 | * rds_send_xmit(). | |
5c115590 AG |
63 | */ |
64 | void 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) { | |
7e3f2952 CM |
70 | rm = conn->c_xmit_rm; |
71 | conn->c_xmit_rm = NULL; | |
5c115590 AG |
72 | /* Tell the user the RDMA op is no longer mapped by the |
73 | * transport. This isn't entirely true (it's flushed out | |
74 | * independently) but as the connection is down, there's | |
75 | * no ongoing RDMA to/from that memory */ | |
7e3f2952 | 76 | rds_message_unmapped(rm); |
7e3f2952 | 77 | rds_message_put(rm); |
5c115590 | 78 | } |
7e3f2952 | 79 | |
5c115590 AG |
80 | conn->c_xmit_sg = 0; |
81 | conn->c_xmit_hdr_off = 0; | |
82 | conn->c_xmit_data_off = 0; | |
15133f6e | 83 | conn->c_xmit_atomic_sent = 0; |
5b2366bd AG |
84 | conn->c_xmit_rdma_sent = 0; |
85 | conn->c_xmit_data_sent = 0; | |
5c115590 AG |
86 | |
87 | conn->c_map_queued = 0; | |
88 | ||
89 | conn->c_unacked_packets = rds_sysctl_max_unacked_packets; | |
90 | conn->c_unacked_bytes = rds_sysctl_max_unacked_bytes; | |
91 | ||
92 | /* Mark messages as retransmissions, and move them to the send q */ | |
93 | spin_lock_irqsave(&conn->c_lock, flags); | |
94 | list_for_each_entry_safe(rm, tmp, &conn->c_retrans, m_conn_item) { | |
95 | set_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags); | |
96 | set_bit(RDS_MSG_RETRANSMITTED, &rm->m_flags); | |
97 | } | |
98 | list_splice_init(&conn->c_retrans, &conn->c_send_queue); | |
99 | spin_unlock_irqrestore(&conn->c_lock, flags); | |
100 | } | |
101 | ||
0f4b1c7e ZB |
102 | static int acquire_in_xmit(struct rds_connection *conn) |
103 | { | |
104 | return test_and_set_bit(RDS_IN_XMIT, &conn->c_flags) == 0; | |
105 | } | |
106 | ||
107 | static void release_in_xmit(struct rds_connection *conn) | |
108 | { | |
109 | clear_bit(RDS_IN_XMIT, &conn->c_flags); | |
4e857c58 | 110 | smp_mb__after_atomic(); |
0f4b1c7e ZB |
111 | /* |
112 | * We don't use wait_on_bit()/wake_up_bit() because our waking is in a | |
113 | * hot path and finding waiters is very rare. We don't want to walk | |
114 | * the system-wide hashed waitqueue buckets in the fast path only to | |
115 | * almost never find waiters. | |
116 | */ | |
117 | if (waitqueue_active(&conn->c_waitq)) | |
118 | wake_up_all(&conn->c_waitq); | |
119 | } | |
120 | ||
5c115590 | 121 | /* |
25985edc | 122 | * We're making the conscious trade-off here to only send one message |
5c115590 AG |
123 | * down the connection at a time. |
124 | * Pro: | |
125 | * - tx queueing is a simple fifo list | |
126 | * - reassembly is optional and easily done by transports per conn | |
127 | * - no per flow rx lookup at all, straight to the socket | |
128 | * - less per-frag memory and wire overhead | |
129 | * Con: | |
130 | * - queued acks can be delayed behind large messages | |
131 | * Depends: | |
132 | * - small message latency is higher behind queued large messages | |
133 | * - large message latency isn't starved by intervening small sends | |
134 | */ | |
135 | int rds_send_xmit(struct rds_connection *conn) | |
136 | { | |
137 | struct rds_message *rm; | |
138 | unsigned long flags; | |
139 | unsigned int tmp; | |
5c115590 AG |
140 | struct scatterlist *sg; |
141 | int ret = 0; | |
5c115590 | 142 | LIST_HEAD(to_be_dropped); |
443be0e5 SV |
143 | int batch_count; |
144 | unsigned long send_gen = 0; | |
5c115590 | 145 | |
fcc5450c | 146 | restart: |
443be0e5 | 147 | batch_count = 0; |
049ee3f5 | 148 | |
5c115590 AG |
149 | /* |
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. | |
5c115590 | 155 | */ |
0f4b1c7e | 156 | if (!acquire_in_xmit(conn)) { |
049ee3f5 | 157 | rds_stats_inc(s_send_lock_contention); |
5c115590 AG |
158 | ret = -ENOMEM; |
159 | goto out; | |
160 | } | |
0f4b1c7e | 161 | |
443be0e5 SV |
162 | /* |
163 | * 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 | ||
0f4b1c7e ZB |
173 | /* |
174 | * 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 | } | |
5c115590 AG |
182 | |
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 | |
5b2366bd | 188 | * the connection doesn't make forward progress. |
5c115590 | 189 | */ |
fcc5450c | 190 | while (1) { |
5c115590 | 191 | |
5c115590 | 192 | rm = conn->c_xmit_rm; |
5c115590 | 193 | |
5b2366bd AG |
194 | /* |
195 | * If between sending messages, we can send a pending congestion | |
196 | * map update. | |
5c115590 | 197 | */ |
8690bfa1 | 198 | if (!rm && test_and_clear_bit(0, &conn->c_map_queued)) { |
77dd550e AG |
199 | rm = rds_cong_update_alloc(conn); |
200 | if (IS_ERR(rm)) { | |
201 | ret = PTR_ERR(rm); | |
202 | break; | |
5b2366bd | 203 | } |
77dd550e AG |
204 | rm->data.op_active = 1; |
205 | ||
206 | conn->c_xmit_rm = rm; | |
5c115590 AG |
207 | } |
208 | ||
209 | /* | |
5b2366bd | 210 | * If not already working on one, grab the next message. |
5c115590 AG |
211 | * |
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 | */ | |
8690bfa1 | 216 | if (!rm) { |
5c115590 AG |
217 | unsigned int len; |
218 | ||
443be0e5 SV |
219 | 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 | ||
0f4b1c7e | 229 | spin_lock_irqsave(&conn->c_lock, flags); |
5c115590 AG |
230 | |
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 | ||
0f4b1c7e | 244 | spin_unlock_irqrestore(&conn->c_lock, flags); |
5c115590 | 245 | |
fcc5450c | 246 | if (!rm) |
5c115590 | 247 | break; |
5c115590 AG |
248 | |
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 | */ | |
f8b3aaf2 | 256 | if (rm->rdma.op_active && |
f64f9e71 | 257 | test_bit(RDS_MSG_RETRANSMITTED, &rm->m_flags)) { |
0f4b1c7e | 258 | spin_lock_irqsave(&conn->c_lock, flags); |
5c115590 AG |
259 | if (test_and_clear_bit(RDS_MSG_ON_CONN, &rm->m_flags)) |
260 | list_move(&rm->m_conn_item, &to_be_dropped); | |
0f4b1c7e | 261 | spin_unlock_irqrestore(&conn->c_lock, flags); |
5c115590 AG |
262 | continue; |
263 | } | |
264 | ||
265 | /* Require an ACK every once in a while */ | |
266 | len = ntohl(rm->m_inc.i_hdr.h_len); | |
f64f9e71 JP |
267 | if (conn->c_unacked_packets == 0 || |
268 | conn->c_unacked_bytes < len) { | |
5c115590 AG |
269 | __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 | ||
2c3a5f9a AG |
282 | /* The transport either sends the whole rdma or none of it */ |
283 | if (rm->rdma.op_active && !conn->c_xmit_rdma_sent) { | |
ff3d7d36 | 284 | rm->m_final_op = &rm->rdma; |
4f73113c | 285 | /* The transport owns the mapped memory for now. |
286 | * You can't unmap it while it's on the send queue | |
287 | */ | |
288 | set_bit(RDS_MSG_MAPPED, &rm->m_flags); | |
2c3a5f9a | 289 | ret = conn->c_trans->xmit_rdma(conn, &rm->rdma); |
4f73113c | 290 | if (ret) { |
291 | clear_bit(RDS_MSG_MAPPED, &rm->m_flags); | |
292 | wake_up_interruptible(&rm->m_flush_wait); | |
15133f6e | 293 | break; |
4f73113c | 294 | } |
2c3a5f9a AG |
295 | conn->c_xmit_rdma_sent = 1; |
296 | ||
15133f6e AG |
297 | } |
298 | ||
2c3a5f9a | 299 | if (rm->atomic.op_active && !conn->c_xmit_atomic_sent) { |
ff3d7d36 | 300 | rm->m_final_op = &rm->atomic; |
4f73113c | 301 | /* The transport owns the mapped memory for now. |
302 | * You can't unmap it while it's on the send queue | |
303 | */ | |
304 | set_bit(RDS_MSG_MAPPED, &rm->m_flags); | |
ff3d7d36 | 305 | ret = conn->c_trans->xmit_atomic(conn, &rm->atomic); |
4f73113c | 306 | if (ret) { |
307 | clear_bit(RDS_MSG_MAPPED, &rm->m_flags); | |
308 | wake_up_interruptible(&rm->m_flush_wait); | |
5c115590 | 309 | break; |
4f73113c | 310 | } |
2c3a5f9a | 311 | conn->c_xmit_atomic_sent = 1; |
ff3d7d36 | 312 | |
5c115590 AG |
313 | } |
314 | ||
2c3a5f9a AG |
315 | /* |
316 | * A number of cases require an RDS header to be sent | |
317 | * even if there is no data. | |
318 | * We permit 0-byte sends; rds-ping depends on this. | |
319 | * However, if there are exclusively attached silent ops, | |
320 | * we skip the hdr/data send, to enable silent operation. | |
321 | */ | |
322 | if (rm->data.op_nents == 0) { | |
323 | int ops_present; | |
324 | int all_ops_are_silent = 1; | |
325 | ||
326 | ops_present = (rm->atomic.op_active || rm->rdma.op_active); | |
327 | if (rm->atomic.op_active && !rm->atomic.op_silent) | |
328 | all_ops_are_silent = 0; | |
329 | if (rm->rdma.op_active && !rm->rdma.op_silent) | |
330 | all_ops_are_silent = 0; | |
331 | ||
332 | if (ops_present && all_ops_are_silent | |
333 | && !rm->m_rdma_cookie) | |
334 | rm->data.op_active = 0; | |
335 | } | |
336 | ||
5b2366bd | 337 | if (rm->data.op_active && !conn->c_xmit_data_sent) { |
ff3d7d36 | 338 | rm->m_final_op = &rm->data; |
5c115590 AG |
339 | ret = conn->c_trans->xmit(conn, rm, |
340 | conn->c_xmit_hdr_off, | |
341 | conn->c_xmit_sg, | |
342 | conn->c_xmit_data_off); | |
343 | if (ret <= 0) | |
344 | break; | |
345 | ||
346 | if (conn->c_xmit_hdr_off < sizeof(struct rds_header)) { | |
347 | tmp = min_t(int, ret, | |
348 | sizeof(struct rds_header) - | |
349 | conn->c_xmit_hdr_off); | |
350 | conn->c_xmit_hdr_off += tmp; | |
351 | ret -= tmp; | |
352 | } | |
353 | ||
6c7cc6e4 | 354 | sg = &rm->data.op_sg[conn->c_xmit_sg]; |
5c115590 AG |
355 | while (ret) { |
356 | tmp = min_t(int, ret, sg->length - | |
357 | conn->c_xmit_data_off); | |
358 | conn->c_xmit_data_off += tmp; | |
359 | ret -= tmp; | |
360 | if (conn->c_xmit_data_off == sg->length) { | |
361 | conn->c_xmit_data_off = 0; | |
362 | sg++; | |
363 | conn->c_xmit_sg++; | |
364 | BUG_ON(ret != 0 && | |
6c7cc6e4 | 365 | conn->c_xmit_sg == rm->data.op_nents); |
5c115590 AG |
366 | } |
367 | } | |
5b2366bd AG |
368 | |
369 | if (conn->c_xmit_hdr_off == sizeof(struct rds_header) && | |
370 | (conn->c_xmit_sg == rm->data.op_nents)) | |
371 | conn->c_xmit_data_sent = 1; | |
372 | } | |
373 | ||
374 | /* | |
375 | * A rm will only take multiple times through this loop | |
376 | * if there is a data op. Thus, if the data is sent (or there was | |
377 | * none), then we're done with the rm. | |
378 | */ | |
379 | if (!rm->data.op_active || conn->c_xmit_data_sent) { | |
380 | conn->c_xmit_rm = NULL; | |
381 | conn->c_xmit_sg = 0; | |
382 | conn->c_xmit_hdr_off = 0; | |
383 | conn->c_xmit_data_off = 0; | |
384 | conn->c_xmit_rdma_sent = 0; | |
385 | conn->c_xmit_atomic_sent = 0; | |
386 | conn->c_xmit_data_sent = 0; | |
387 | ||
388 | rds_message_put(rm); | |
5c115590 AG |
389 | } |
390 | } | |
391 | ||
443be0e5 | 392 | over_batch: |
5c115590 AG |
393 | if (conn->c_trans->xmit_complete) |
394 | conn->c_trans->xmit_complete(conn); | |
0f4b1c7e | 395 | release_in_xmit(conn); |
5c115590 | 396 | |
2ad8099b AG |
397 | /* Nuke any messages we decided not to retransmit. */ |
398 | if (!list_empty(&to_be_dropped)) { | |
399 | /* irqs on here, so we can put(), unlike above */ | |
400 | list_for_each_entry(rm, &to_be_dropped, m_conn_item) | |
401 | rds_message_put(rm); | |
402 | rds_send_remove_from_sock(&to_be_dropped, RDS_RDMA_DROPPED); | |
403 | } | |
404 | ||
fcc5450c | 405 | /* |
0f4b1c7e ZB |
406 | * Other senders can queue a message after we last test the send queue |
407 | * but before we clear RDS_IN_XMIT. In that case they'd back off and | |
408 | * not try and send their newly queued message. We need to check the | |
409 | * send queue after having cleared RDS_IN_XMIT so that their message | |
410 | * doesn't get stuck on the send queue. | |
fcc5450c AG |
411 | * |
412 | * If the transport cannot continue (i.e ret != 0), then it must | |
413 | * call us when more room is available, such as from the tx | |
414 | * completion handler. | |
443be0e5 SV |
415 | * |
416 | * We have an extra generation check here so that if someone manages | |
417 | * to jump in after our release_in_xmit, we'll see that they have done | |
418 | * some work and we will skip our goto | |
fcc5450c AG |
419 | */ |
420 | if (ret == 0) { | |
9e29db0e | 421 | smp_mb(); |
0c484240 | 422 | if ((test_bit(0, &conn->c_map_queued) || |
423 | !list_empty(&conn->c_send_queue)) && | |
443be0e5 | 424 | send_gen == conn->c_send_gen) { |
049ee3f5 | 425 | rds_stats_inc(s_send_lock_queue_raced); |
0f4b1c7e | 426 | goto restart; |
5c115590 | 427 | } |
5c115590 AG |
428 | } |
429 | out: | |
430 | return ret; | |
431 | } | |
432 | ||
433 | static void rds_send_sndbuf_remove(struct rds_sock *rs, struct rds_message *rm) | |
434 | { | |
435 | u32 len = be32_to_cpu(rm->m_inc.i_hdr.h_len); | |
436 | ||
437 | assert_spin_locked(&rs->rs_lock); | |
438 | ||
439 | BUG_ON(rs->rs_snd_bytes < len); | |
440 | rs->rs_snd_bytes -= len; | |
441 | ||
442 | if (rs->rs_snd_bytes == 0) | |
443 | rds_stats_inc(s_send_queue_empty); | |
444 | } | |
445 | ||
446 | static inline int rds_send_is_acked(struct rds_message *rm, u64 ack, | |
447 | is_acked_func is_acked) | |
448 | { | |
449 | if (is_acked) | |
450 | return is_acked(rm, ack); | |
451 | return be64_to_cpu(rm->m_inc.i_hdr.h_sequence) <= ack; | |
452 | } | |
453 | ||
5c115590 AG |
454 | /* |
455 | * This is pretty similar to what happens below in the ACK | |
456 | * handling code - except that we call here as soon as we get | |
457 | * the IB send completion on the RDMA op and the accompanying | |
458 | * message. | |
459 | */ | |
460 | void rds_rdma_send_complete(struct rds_message *rm, int status) | |
461 | { | |
462 | struct rds_sock *rs = NULL; | |
f8b3aaf2 | 463 | struct rm_rdma_op *ro; |
5c115590 | 464 | struct rds_notifier *notifier; |
9de0864c | 465 | unsigned long flags; |
5c115590 | 466 | |
9de0864c | 467 | spin_lock_irqsave(&rm->m_rs_lock, flags); |
5c115590 | 468 | |
f8b3aaf2 | 469 | ro = &rm->rdma; |
f64f9e71 | 470 | if (test_bit(RDS_MSG_ON_SOCK, &rm->m_flags) && |
f8b3aaf2 AG |
471 | ro->op_active && ro->op_notify && ro->op_notifier) { |
472 | notifier = ro->op_notifier; | |
5c115590 AG |
473 | rs = rm->m_rs; |
474 | sock_hold(rds_rs_to_sk(rs)); | |
475 | ||
476 | notifier->n_status = status; | |
477 | spin_lock(&rs->rs_lock); | |
478 | list_add_tail(¬ifier->n_list, &rs->rs_notify_queue); | |
479 | spin_unlock(&rs->rs_lock); | |
480 | ||
f8b3aaf2 | 481 | ro->op_notifier = NULL; |
5c115590 AG |
482 | } |
483 | ||
9de0864c | 484 | spin_unlock_irqrestore(&rm->m_rs_lock, flags); |
5c115590 AG |
485 | |
486 | if (rs) { | |
487 | rds_wake_sk_sleep(rs); | |
488 | sock_put(rds_rs_to_sk(rs)); | |
489 | } | |
490 | } | |
616b757a | 491 | EXPORT_SYMBOL_GPL(rds_rdma_send_complete); |
5c115590 | 492 | |
15133f6e AG |
493 | /* |
494 | * Just like above, except looks at atomic op | |
495 | */ | |
496 | void rds_atomic_send_complete(struct rds_message *rm, int status) | |
497 | { | |
498 | struct rds_sock *rs = NULL; | |
499 | struct rm_atomic_op *ao; | |
500 | struct rds_notifier *notifier; | |
cf4b7389 | 501 | unsigned long flags; |
15133f6e | 502 | |
cf4b7389 | 503 | spin_lock_irqsave(&rm->m_rs_lock, flags); |
15133f6e AG |
504 | |
505 | ao = &rm->atomic; | |
506 | if (test_bit(RDS_MSG_ON_SOCK, &rm->m_flags) | |
507 | && ao->op_active && ao->op_notify && ao->op_notifier) { | |
508 | notifier = ao->op_notifier; | |
509 | rs = rm->m_rs; | |
510 | sock_hold(rds_rs_to_sk(rs)); | |
511 | ||
512 | notifier->n_status = status; | |
513 | spin_lock(&rs->rs_lock); | |
514 | list_add_tail(¬ifier->n_list, &rs->rs_notify_queue); | |
515 | spin_unlock(&rs->rs_lock); | |
516 | ||
517 | ao->op_notifier = NULL; | |
518 | } | |
519 | ||
cf4b7389 | 520 | spin_unlock_irqrestore(&rm->m_rs_lock, flags); |
15133f6e AG |
521 | |
522 | if (rs) { | |
523 | rds_wake_sk_sleep(rs); | |
524 | sock_put(rds_rs_to_sk(rs)); | |
525 | } | |
526 | } | |
527 | EXPORT_SYMBOL_GPL(rds_atomic_send_complete); | |
528 | ||
5c115590 AG |
529 | /* |
530 | * This is the same as rds_rdma_send_complete except we | |
531 | * don't do any locking - we have all the ingredients (message, | |
532 | * socket, socket lock) and can just move the notifier. | |
533 | */ | |
534 | static inline void | |
940786eb | 535 | __rds_send_complete(struct rds_sock *rs, struct rds_message *rm, int status) |
5c115590 | 536 | { |
f8b3aaf2 | 537 | struct rm_rdma_op *ro; |
940786eb | 538 | struct rm_atomic_op *ao; |
5c115590 | 539 | |
f8b3aaf2 AG |
540 | ro = &rm->rdma; |
541 | if (ro->op_active && ro->op_notify && ro->op_notifier) { | |
542 | ro->op_notifier->n_status = status; | |
543 | list_add_tail(&ro->op_notifier->n_list, &rs->rs_notify_queue); | |
544 | ro->op_notifier = NULL; | |
5c115590 AG |
545 | } |
546 | ||
940786eb AG |
547 | ao = &rm->atomic; |
548 | if (ao->op_active && ao->op_notify && ao->op_notifier) { | |
549 | ao->op_notifier->n_status = status; | |
550 | list_add_tail(&ao->op_notifier->n_list, &rs->rs_notify_queue); | |
551 | ao->op_notifier = NULL; | |
552 | } | |
553 | ||
5c115590 AG |
554 | /* No need to wake the app - caller does this */ |
555 | } | |
556 | ||
557 | /* | |
558 | * This is called from the IB send completion when we detect | |
559 | * a RDMA operation that failed with remote access error. | |
560 | * So speed is not an issue here. | |
561 | */ | |
562 | struct rds_message *rds_send_get_message(struct rds_connection *conn, | |
f8b3aaf2 | 563 | struct rm_rdma_op *op) |
5c115590 AG |
564 | { |
565 | struct rds_message *rm, *tmp, *found = NULL; | |
566 | unsigned long flags; | |
567 | ||
568 | spin_lock_irqsave(&conn->c_lock, flags); | |
569 | ||
570 | list_for_each_entry_safe(rm, tmp, &conn->c_retrans, m_conn_item) { | |
f8b3aaf2 | 571 | if (&rm->rdma == op) { |
5c115590 AG |
572 | atomic_inc(&rm->m_refcount); |
573 | found = rm; | |
574 | goto out; | |
575 | } | |
576 | } | |
577 | ||
578 | list_for_each_entry_safe(rm, tmp, &conn->c_send_queue, m_conn_item) { | |
f8b3aaf2 | 579 | if (&rm->rdma == op) { |
5c115590 AG |
580 | atomic_inc(&rm->m_refcount); |
581 | found = rm; | |
582 | break; | |
583 | } | |
584 | } | |
585 | ||
586 | out: | |
587 | spin_unlock_irqrestore(&conn->c_lock, flags); | |
588 | ||
589 | return found; | |
590 | } | |
616b757a | 591 | EXPORT_SYMBOL_GPL(rds_send_get_message); |
5c115590 AG |
592 | |
593 | /* | |
594 | * This removes messages from the socket's list if they're on it. The list | |
595 | * argument must be private to the caller, we must be able to modify it | |
596 | * without locks. The messages must have a reference held for their | |
597 | * position on the list. This function will drop that reference after | |
598 | * removing the messages from the 'messages' list regardless of if it found | |
599 | * the messages on the socket list or not. | |
600 | */ | |
ff51bf84 | 601 | static void rds_send_remove_from_sock(struct list_head *messages, int status) |
5c115590 | 602 | { |
561c7df6 | 603 | unsigned long flags; |
5c115590 AG |
604 | struct rds_sock *rs = NULL; |
605 | struct rds_message *rm; | |
606 | ||
5c115590 | 607 | while (!list_empty(messages)) { |
561c7df6 AG |
608 | int was_on_sock = 0; |
609 | ||
5c115590 AG |
610 | rm = list_entry(messages->next, struct rds_message, |
611 | m_conn_item); | |
612 | list_del_init(&rm->m_conn_item); | |
613 | ||
614 | /* | |
615 | * If we see this flag cleared then we're *sure* that someone | |
616 | * else beat us to removing it from the sock. If we race | |
617 | * with their flag update we'll get the lock and then really | |
618 | * see that the flag has been cleared. | |
619 | * | |
620 | * The message spinlock makes sure nobody clears rm->m_rs | |
621 | * while we're messing with it. It does not prevent the | |
622 | * message from being removed from the socket, though. | |
623 | */ | |
561c7df6 | 624 | spin_lock_irqsave(&rm->m_rs_lock, flags); |
5c115590 AG |
625 | if (!test_bit(RDS_MSG_ON_SOCK, &rm->m_flags)) |
626 | goto unlock_and_drop; | |
627 | ||
628 | if (rs != rm->m_rs) { | |
629 | if (rs) { | |
5c115590 AG |
630 | rds_wake_sk_sleep(rs); |
631 | sock_put(rds_rs_to_sk(rs)); | |
632 | } | |
633 | rs = rm->m_rs; | |
593cbb3e HK |
634 | if (rs) |
635 | sock_hold(rds_rs_to_sk(rs)); | |
5c115590 | 636 | } |
593cbb3e HK |
637 | if (!rs) |
638 | goto unlock_and_drop; | |
048c15e6 | 639 | spin_lock(&rs->rs_lock); |
5c115590 AG |
640 | |
641 | if (test_and_clear_bit(RDS_MSG_ON_SOCK, &rm->m_flags)) { | |
f8b3aaf2 | 642 | struct rm_rdma_op *ro = &rm->rdma; |
5c115590 AG |
643 | struct rds_notifier *notifier; |
644 | ||
645 | list_del_init(&rm->m_sock_item); | |
646 | rds_send_sndbuf_remove(rs, rm); | |
647 | ||
f8b3aaf2 AG |
648 | if (ro->op_active && ro->op_notifier && |
649 | (ro->op_notify || (ro->op_recverr && status))) { | |
650 | notifier = ro->op_notifier; | |
5c115590 AG |
651 | list_add_tail(¬ifier->n_list, |
652 | &rs->rs_notify_queue); | |
653 | if (!notifier->n_status) | |
654 | notifier->n_status = status; | |
f8b3aaf2 | 655 | rm->rdma.op_notifier = NULL; |
5c115590 | 656 | } |
561c7df6 | 657 | was_on_sock = 1; |
5c115590 AG |
658 | rm->m_rs = NULL; |
659 | } | |
048c15e6 | 660 | spin_unlock(&rs->rs_lock); |
5c115590 AG |
661 | |
662 | unlock_and_drop: | |
561c7df6 | 663 | spin_unlock_irqrestore(&rm->m_rs_lock, flags); |
5c115590 | 664 | rds_message_put(rm); |
561c7df6 AG |
665 | if (was_on_sock) |
666 | rds_message_put(rm); | |
5c115590 AG |
667 | } |
668 | ||
669 | if (rs) { | |
5c115590 AG |
670 | rds_wake_sk_sleep(rs); |
671 | sock_put(rds_rs_to_sk(rs)); | |
672 | } | |
5c115590 AG |
673 | } |
674 | ||
675 | /* | |
676 | * Transports call here when they've determined that the receiver queued | |
677 | * messages up to, and including, the given sequence number. Messages are | |
678 | * moved to the retrans queue when rds_send_xmit picks them off the send | |
679 | * queue. This means that in the TCP case, the message may not have been | |
680 | * assigned the m_ack_seq yet - but that's fine as long as tcp_is_acked | |
681 | * checks the RDS_MSG_HAS_ACK_SEQ bit. | |
5c115590 AG |
682 | */ |
683 | void rds_send_drop_acked(struct rds_connection *conn, u64 ack, | |
684 | is_acked_func is_acked) | |
685 | { | |
686 | struct rds_message *rm, *tmp; | |
687 | unsigned long flags; | |
688 | LIST_HEAD(list); | |
689 | ||
690 | spin_lock_irqsave(&conn->c_lock, flags); | |
691 | ||
692 | list_for_each_entry_safe(rm, tmp, &conn->c_retrans, m_conn_item) { | |
693 | if (!rds_send_is_acked(rm, ack, is_acked)) | |
694 | break; | |
695 | ||
696 | list_move(&rm->m_conn_item, &list); | |
697 | clear_bit(RDS_MSG_ON_CONN, &rm->m_flags); | |
698 | } | |
699 | ||
700 | /* order flag updates with spin locks */ | |
701 | if (!list_empty(&list)) | |
4e857c58 | 702 | smp_mb__after_atomic(); |
5c115590 AG |
703 | |
704 | spin_unlock_irqrestore(&conn->c_lock, flags); | |
705 | ||
706 | /* now remove the messages from the sock list as needed */ | |
707 | rds_send_remove_from_sock(&list, RDS_RDMA_SUCCESS); | |
708 | } | |
616b757a | 709 | EXPORT_SYMBOL_GPL(rds_send_drop_acked); |
5c115590 AG |
710 | |
711 | void rds_send_drop_to(struct rds_sock *rs, struct sockaddr_in *dest) | |
712 | { | |
713 | struct rds_message *rm, *tmp; | |
714 | struct rds_connection *conn; | |
7c82eaf0 | 715 | unsigned long flags; |
5c115590 | 716 | LIST_HEAD(list); |
5c115590 AG |
717 | |
718 | /* get all the messages we're dropping under the rs lock */ | |
719 | spin_lock_irqsave(&rs->rs_lock, flags); | |
720 | ||
721 | list_for_each_entry_safe(rm, tmp, &rs->rs_send_queue, m_sock_item) { | |
722 | if (dest && (dest->sin_addr.s_addr != rm->m_daddr || | |
723 | dest->sin_port != rm->m_inc.i_hdr.h_dport)) | |
724 | continue; | |
725 | ||
5c115590 AG |
726 | list_move(&rm->m_sock_item, &list); |
727 | rds_send_sndbuf_remove(rs, rm); | |
728 | clear_bit(RDS_MSG_ON_SOCK, &rm->m_flags); | |
5c115590 AG |
729 | } |
730 | ||
731 | /* order flag updates with the rs lock */ | |
4e857c58 | 732 | smp_mb__after_atomic(); |
5c115590 AG |
733 | |
734 | spin_unlock_irqrestore(&rs->rs_lock, flags); | |
735 | ||
7c82eaf0 AG |
736 | if (list_empty(&list)) |
737 | return; | |
5c115590 | 738 | |
7c82eaf0 | 739 | /* Remove the messages from the conn */ |
5c115590 | 740 | list_for_each_entry(rm, &list, m_sock_item) { |
7c82eaf0 AG |
741 | |
742 | conn = rm->m_inc.i_conn; | |
5c115590 | 743 | |
9de0864c | 744 | spin_lock_irqsave(&conn->c_lock, flags); |
5c115590 | 745 | /* |
7c82eaf0 AG |
746 | * Maybe someone else beat us to removing rm from the conn. |
747 | * If we race with their flag update we'll get the lock and | |
748 | * then really see that the flag has been cleared. | |
5c115590 | 749 | */ |
7c82eaf0 AG |
750 | if (!test_and_clear_bit(RDS_MSG_ON_CONN, &rm->m_flags)) { |
751 | spin_unlock_irqrestore(&conn->c_lock, flags); | |
593cbb3e HK |
752 | spin_lock_irqsave(&rm->m_rs_lock, flags); |
753 | rm->m_rs = NULL; | |
754 | spin_unlock_irqrestore(&rm->m_rs_lock, flags); | |
5c115590 | 755 | continue; |
5c115590 | 756 | } |
9de0864c AG |
757 | list_del_init(&rm->m_conn_item); |
758 | spin_unlock_irqrestore(&conn->c_lock, flags); | |
5c115590 | 759 | |
7c82eaf0 AG |
760 | /* |
761 | * Couldn't grab m_rs_lock in top loop (lock ordering), | |
762 | * but we can now. | |
763 | */ | |
9de0864c | 764 | spin_lock_irqsave(&rm->m_rs_lock, flags); |
5c115590 | 765 | |
7c82eaf0 | 766 | spin_lock(&rs->rs_lock); |
940786eb | 767 | __rds_send_complete(rs, rm, RDS_RDMA_CANCELED); |
7c82eaf0 AG |
768 | spin_unlock(&rs->rs_lock); |
769 | ||
770 | rm->m_rs = NULL; | |
9de0864c | 771 | spin_unlock_irqrestore(&rm->m_rs_lock, flags); |
7c82eaf0 | 772 | |
7c82eaf0 | 773 | rds_message_put(rm); |
7c82eaf0 | 774 | } |
5c115590 | 775 | |
7c82eaf0 | 776 | rds_wake_sk_sleep(rs); |
550a8002 | 777 | |
5c115590 AG |
778 | while (!list_empty(&list)) { |
779 | rm = list_entry(list.next, struct rds_message, m_sock_item); | |
780 | list_del_init(&rm->m_sock_item); | |
781 | ||
782 | rds_message_wait(rm); | |
783 | rds_message_put(rm); | |
784 | } | |
785 | } | |
786 | ||
787 | /* | |
788 | * we only want this to fire once so we use the callers 'queued'. It's | |
789 | * possible that another thread can race with us and remove the | |
790 | * message from the flow with RDS_CANCEL_SENT_TO. | |
791 | */ | |
792 | static int rds_send_queue_rm(struct rds_sock *rs, struct rds_connection *conn, | |
793 | struct rds_message *rm, __be16 sport, | |
794 | __be16 dport, int *queued) | |
795 | { | |
796 | unsigned long flags; | |
797 | u32 len; | |
798 | ||
799 | if (*queued) | |
800 | goto out; | |
801 | ||
802 | len = be32_to_cpu(rm->m_inc.i_hdr.h_len); | |
803 | ||
804 | /* this is the only place which holds both the socket's rs_lock | |
805 | * and the connection's c_lock */ | |
806 | spin_lock_irqsave(&rs->rs_lock, flags); | |
807 | ||
808 | /* | |
809 | * If there is a little space in sndbuf, we don't queue anything, | |
810 | * and userspace gets -EAGAIN. But poll() indicates there's send | |
811 | * room. This can lead to bad behavior (spinning) if snd_bytes isn't | |
812 | * freed up by incoming acks. So we check the *old* value of | |
813 | * rs_snd_bytes here to allow the last msg to exceed the buffer, | |
814 | * and poll() now knows no more data can be sent. | |
815 | */ | |
816 | if (rs->rs_snd_bytes < rds_sk_sndbuf(rs)) { | |
817 | rs->rs_snd_bytes += len; | |
818 | ||
819 | /* let recv side know we are close to send space exhaustion. | |
820 | * This is probably not the optimal way to do it, as this | |
821 | * means we set the flag on *all* messages as soon as our | |
822 | * throughput hits a certain threshold. | |
823 | */ | |
824 | if (rs->rs_snd_bytes >= rds_sk_sndbuf(rs) / 2) | |
825 | __set_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags); | |
826 | ||
827 | list_add_tail(&rm->m_sock_item, &rs->rs_send_queue); | |
828 | set_bit(RDS_MSG_ON_SOCK, &rm->m_flags); | |
829 | rds_message_addref(rm); | |
830 | rm->m_rs = rs; | |
831 | ||
832 | /* The code ordering is a little weird, but we're | |
833 | trying to minimize the time we hold c_lock */ | |
834 | rds_message_populate_header(&rm->m_inc.i_hdr, sport, dport, 0); | |
835 | rm->m_inc.i_conn = conn; | |
836 | rds_message_addref(rm); | |
837 | ||
838 | spin_lock(&conn->c_lock); | |
839 | rm->m_inc.i_hdr.h_sequence = cpu_to_be64(conn->c_next_tx_seq++); | |
840 | list_add_tail(&rm->m_conn_item, &conn->c_send_queue); | |
841 | set_bit(RDS_MSG_ON_CONN, &rm->m_flags); | |
842 | spin_unlock(&conn->c_lock); | |
843 | ||
844 | rdsdebug("queued msg %p len %d, rs %p bytes %d seq %llu\n", | |
845 | rm, len, rs, rs->rs_snd_bytes, | |
846 | (unsigned long long)be64_to_cpu(rm->m_inc.i_hdr.h_sequence)); | |
847 | ||
848 | *queued = 1; | |
849 | } | |
850 | ||
851 | spin_unlock_irqrestore(&rs->rs_lock, flags); | |
852 | out: | |
853 | return *queued; | |
854 | } | |
855 | ||
fc445084 AG |
856 | /* |
857 | * rds_message is getting to be quite complicated, and we'd like to allocate | |
858 | * it all in one go. This figures out how big it needs to be up front. | |
859 | */ | |
860 | static int rds_rm_size(struct msghdr *msg, int data_len) | |
861 | { | |
ff87e97a | 862 | struct cmsghdr *cmsg; |
fc445084 | 863 | int size = 0; |
aa0a4ef4 | 864 | int cmsg_groups = 0; |
ff87e97a AG |
865 | int retval; |
866 | ||
f95b414e | 867 | for_each_cmsghdr(cmsg, msg) { |
ff87e97a AG |
868 | if (!CMSG_OK(msg, cmsg)) |
869 | return -EINVAL; | |
870 | ||
871 | if (cmsg->cmsg_level != SOL_RDS) | |
872 | continue; | |
873 | ||
874 | switch (cmsg->cmsg_type) { | |
875 | case RDS_CMSG_RDMA_ARGS: | |
aa0a4ef4 | 876 | cmsg_groups |= 1; |
ff87e97a AG |
877 | retval = rds_rdma_extra_size(CMSG_DATA(cmsg)); |
878 | if (retval < 0) | |
879 | return retval; | |
880 | size += retval; | |
aa0a4ef4 | 881 | |
ff87e97a AG |
882 | break; |
883 | ||
884 | case RDS_CMSG_RDMA_DEST: | |
885 | case RDS_CMSG_RDMA_MAP: | |
aa0a4ef4 | 886 | cmsg_groups |= 2; |
ff87e97a AG |
887 | /* these are valid but do no add any size */ |
888 | break; | |
889 | ||
15133f6e AG |
890 | case RDS_CMSG_ATOMIC_CSWP: |
891 | case RDS_CMSG_ATOMIC_FADD: | |
20c72bd5 AG |
892 | case RDS_CMSG_MASKED_ATOMIC_CSWP: |
893 | case RDS_CMSG_MASKED_ATOMIC_FADD: | |
aa0a4ef4 | 894 | cmsg_groups |= 1; |
15133f6e AG |
895 | size += sizeof(struct scatterlist); |
896 | break; | |
897 | ||
ff87e97a AG |
898 | default: |
899 | return -EINVAL; | |
900 | } | |
901 | ||
902 | } | |
fc445084 | 903 | |
ff87e97a | 904 | size += ceil(data_len, PAGE_SIZE) * sizeof(struct scatterlist); |
fc445084 | 905 | |
aa0a4ef4 AG |
906 | /* Ensure (DEST, MAP) are never used with (ARGS, ATOMIC) */ |
907 | if (cmsg_groups == 3) | |
908 | return -EINVAL; | |
909 | ||
fc445084 AG |
910 | return size; |
911 | } | |
912 | ||
5c115590 AG |
913 | static int rds_cmsg_send(struct rds_sock *rs, struct rds_message *rm, |
914 | struct msghdr *msg, int *allocated_mr) | |
915 | { | |
916 | struct cmsghdr *cmsg; | |
917 | int ret = 0; | |
918 | ||
f95b414e | 919 | for_each_cmsghdr(cmsg, msg) { |
5c115590 AG |
920 | if (!CMSG_OK(msg, cmsg)) |
921 | return -EINVAL; | |
922 | ||
923 | if (cmsg->cmsg_level != SOL_RDS) | |
924 | continue; | |
925 | ||
926 | /* As a side effect, RDMA_DEST and RDMA_MAP will set | |
15133f6e | 927 | * rm->rdma.m_rdma_cookie and rm->rdma.m_rdma_mr. |
5c115590 AG |
928 | */ |
929 | switch (cmsg->cmsg_type) { | |
930 | case RDS_CMSG_RDMA_ARGS: | |
931 | ret = rds_cmsg_rdma_args(rs, rm, cmsg); | |
932 | break; | |
933 | ||
934 | case RDS_CMSG_RDMA_DEST: | |
935 | ret = rds_cmsg_rdma_dest(rs, rm, cmsg); | |
936 | break; | |
937 | ||
938 | case RDS_CMSG_RDMA_MAP: | |
939 | ret = rds_cmsg_rdma_map(rs, rm, cmsg); | |
940 | if (!ret) | |
941 | *allocated_mr = 1; | |
942 | break; | |
15133f6e AG |
943 | case RDS_CMSG_ATOMIC_CSWP: |
944 | case RDS_CMSG_ATOMIC_FADD: | |
20c72bd5 AG |
945 | case RDS_CMSG_MASKED_ATOMIC_CSWP: |
946 | case RDS_CMSG_MASKED_ATOMIC_FADD: | |
15133f6e AG |
947 | ret = rds_cmsg_atomic(rs, rm, cmsg); |
948 | break; | |
5c115590 AG |
949 | |
950 | default: | |
951 | return -EINVAL; | |
952 | } | |
953 | ||
954 | if (ret) | |
955 | break; | |
956 | } | |
957 | ||
958 | return ret; | |
959 | } | |
960 | ||
1b784140 | 961 | int rds_sendmsg(struct socket *sock, struct msghdr *msg, size_t payload_len) |
5c115590 AG |
962 | { |
963 | struct sock *sk = sock->sk; | |
964 | struct rds_sock *rs = rds_sk_to_rs(sk); | |
342dfc30 | 965 | DECLARE_SOCKADDR(struct sockaddr_in *, usin, msg->msg_name); |
5c115590 AG |
966 | __be32 daddr; |
967 | __be16 dport; | |
968 | struct rds_message *rm = NULL; | |
969 | struct rds_connection *conn; | |
970 | int ret = 0; | |
971 | int queued = 0, allocated_mr = 0; | |
972 | int nonblock = msg->msg_flags & MSG_DONTWAIT; | |
1123fd73 | 973 | long timeo = sock_sndtimeo(sk, nonblock); |
5c115590 AG |
974 | |
975 | /* Mirror Linux UDP mirror of BSD error message compatibility */ | |
976 | /* XXX: Perhaps MSG_MORE someday */ | |
977 | if (msg->msg_flags & ~(MSG_DONTWAIT | MSG_CMSG_COMPAT)) { | |
5c115590 AG |
978 | ret = -EOPNOTSUPP; |
979 | goto out; | |
980 | } | |
981 | ||
982 | if (msg->msg_namelen) { | |
983 | /* XXX fail non-unicast destination IPs? */ | |
984 | if (msg->msg_namelen < sizeof(*usin) || usin->sin_family != AF_INET) { | |
985 | ret = -EINVAL; | |
986 | goto out; | |
987 | } | |
988 | daddr = usin->sin_addr.s_addr; | |
989 | dport = usin->sin_port; | |
990 | } else { | |
991 | /* We only care about consistency with ->connect() */ | |
992 | lock_sock(sk); | |
993 | daddr = rs->rs_conn_addr; | |
994 | dport = rs->rs_conn_port; | |
995 | release_sock(sk); | |
996 | } | |
997 | ||
998 | /* racing with another thread binding seems ok here */ | |
999 | if (daddr == 0 || rs->rs_bound_addr == 0) { | |
1000 | ret = -ENOTCONN; /* XXX not a great errno */ | |
1001 | goto out; | |
1002 | } | |
1003 | ||
fc445084 AG |
1004 | /* size of rm including all sgs */ |
1005 | ret = rds_rm_size(msg, payload_len); | |
1006 | if (ret < 0) | |
1007 | goto out; | |
1008 | ||
1009 | rm = rds_message_alloc(ret, GFP_KERNEL); | |
1010 | if (!rm) { | |
1011 | ret = -ENOMEM; | |
5c115590 AG |
1012 | goto out; |
1013 | } | |
1014 | ||
372cd7de AG |
1015 | /* Attach data to the rm */ |
1016 | if (payload_len) { | |
1017 | rm->data.op_sg = rds_message_alloc_sgs(rm, ceil(payload_len, PAGE_SIZE)); | |
d139ff09 AG |
1018 | if (!rm->data.op_sg) { |
1019 | ret = -ENOMEM; | |
1020 | goto out; | |
1021 | } | |
c0371da6 | 1022 | ret = rds_message_copy_from_user(rm, &msg->msg_iter); |
372cd7de AG |
1023 | if (ret) |
1024 | goto out; | |
1025 | } | |
1026 | rm->data.op_active = 1; | |
fc445084 | 1027 | |
5c115590 AG |
1028 | rm->m_daddr = daddr; |
1029 | ||
5c115590 AG |
1030 | /* rds_conn_create has a spinlock that runs with IRQ off. |
1031 | * Caching the conn in the socket helps a lot. */ | |
1032 | if (rs->rs_conn && rs->rs_conn->c_faddr == daddr) | |
1033 | conn = rs->rs_conn; | |
1034 | else { | |
d5a8ac28 SV |
1035 | conn = rds_conn_create_outgoing(sock_net(sock->sk), |
1036 | rs->rs_bound_addr, daddr, | |
5c115590 AG |
1037 | rs->rs_transport, |
1038 | sock->sk->sk_allocation); | |
1039 | if (IS_ERR(conn)) { | |
1040 | ret = PTR_ERR(conn); | |
1041 | goto out; | |
1042 | } | |
1043 | rs->rs_conn = conn; | |
1044 | } | |
1045 | ||
49f69691 AG |
1046 | /* Parse any control messages the user may have included. */ |
1047 | ret = rds_cmsg_send(rs, rm, msg, &allocated_mr); | |
1048 | if (ret) | |
1049 | goto out; | |
1050 | ||
2c3a5f9a | 1051 | if (rm->rdma.op_active && !conn->c_trans->xmit_rdma) { |
cb0a6056 | 1052 | printk_ratelimited(KERN_NOTICE "rdma_op %p conn xmit_rdma %p\n", |
f8b3aaf2 | 1053 | &rm->rdma, conn->c_trans->xmit_rdma); |
15133f6e AG |
1054 | ret = -EOPNOTSUPP; |
1055 | goto out; | |
1056 | } | |
1057 | ||
1058 | if (rm->atomic.op_active && !conn->c_trans->xmit_atomic) { | |
cb0a6056 | 1059 | printk_ratelimited(KERN_NOTICE "atomic_op %p conn xmit_atomic %p\n", |
15133f6e | 1060 | &rm->atomic, conn->c_trans->xmit_atomic); |
5c115590 AG |
1061 | ret = -EOPNOTSUPP; |
1062 | goto out; | |
1063 | } | |
1064 | ||
f3c6808d | 1065 | rds_conn_connect_if_down(conn); |
5c115590 AG |
1066 | |
1067 | ret = rds_cong_wait(conn->c_fcong, dport, nonblock, rs); | |
b98ba52f AG |
1068 | if (ret) { |
1069 | rs->rs_seen_congestion = 1; | |
5c115590 | 1070 | goto out; |
b98ba52f | 1071 | } |
5c115590 AG |
1072 | |
1073 | while (!rds_send_queue_rm(rs, conn, rm, rs->rs_bound_port, | |
1074 | dport, &queued)) { | |
1075 | rds_stats_inc(s_send_queue_full); | |
1076 | /* XXX make sure this is reasonable */ | |
1077 | if (payload_len > rds_sk_sndbuf(rs)) { | |
1078 | ret = -EMSGSIZE; | |
1079 | goto out; | |
1080 | } | |
1081 | if (nonblock) { | |
1082 | ret = -EAGAIN; | |
1083 | goto out; | |
1084 | } | |
1085 | ||
aa395145 | 1086 | timeo = wait_event_interruptible_timeout(*sk_sleep(sk), |
5c115590 AG |
1087 | rds_send_queue_rm(rs, conn, rm, |
1088 | rs->rs_bound_port, | |
1089 | dport, | |
1090 | &queued), | |
1091 | timeo); | |
1092 | rdsdebug("sendmsg woke queued %d timeo %ld\n", queued, timeo); | |
1093 | if (timeo > 0 || timeo == MAX_SCHEDULE_TIMEOUT) | |
1094 | continue; | |
1095 | ||
1096 | ret = timeo; | |
1097 | if (ret == 0) | |
1098 | ret = -ETIMEDOUT; | |
1099 | goto out; | |
1100 | } | |
1101 | ||
1102 | /* | |
1103 | * By now we've committed to the send. We reuse rds_send_worker() | |
1104 | * to retry sends in the rds thread if the transport asks us to. | |
1105 | */ | |
1106 | rds_stats_inc(s_send_queued); | |
1107 | ||
1108 | if (!test_bit(RDS_LL_SEND_FULL, &conn->c_flags)) | |
a7d3a281 | 1109 | rds_send_xmit(conn); |
5c115590 AG |
1110 | |
1111 | rds_message_put(rm); | |
1112 | return payload_len; | |
1113 | ||
1114 | out: | |
1115 | /* If the user included a RDMA_MAP cmsg, we allocated a MR on the fly. | |
1116 | * If the sendmsg goes through, we keep the MR. If it fails with EAGAIN | |
1117 | * or in any other way, we need to destroy the MR again */ | |
1118 | if (allocated_mr) | |
1119 | rds_rdma_unuse(rs, rds_rdma_cookie_key(rm->m_rdma_cookie), 1); | |
1120 | ||
1121 | if (rm) | |
1122 | rds_message_put(rm); | |
1123 | return ret; | |
1124 | } | |
1125 | ||
1126 | /* | |
1127 | * Reply to a ping packet. | |
1128 | */ | |
1129 | int | |
1130 | rds_send_pong(struct rds_connection *conn, __be16 dport) | |
1131 | { | |
1132 | struct rds_message *rm; | |
1133 | unsigned long flags; | |
1134 | int ret = 0; | |
1135 | ||
1136 | rm = rds_message_alloc(0, GFP_ATOMIC); | |
8690bfa1 | 1137 | if (!rm) { |
5c115590 AG |
1138 | ret = -ENOMEM; |
1139 | goto out; | |
1140 | } | |
1141 | ||
1142 | rm->m_daddr = conn->c_faddr; | |
acfcd4d4 | 1143 | rm->data.op_active = 1; |
5c115590 | 1144 | |
f3c6808d | 1145 | rds_conn_connect_if_down(conn); |
5c115590 AG |
1146 | |
1147 | ret = rds_cong_wait(conn->c_fcong, dport, 1, NULL); | |
1148 | if (ret) | |
1149 | goto out; | |
1150 | ||
1151 | spin_lock_irqsave(&conn->c_lock, flags); | |
1152 | list_add_tail(&rm->m_conn_item, &conn->c_send_queue); | |
1153 | set_bit(RDS_MSG_ON_CONN, &rm->m_flags); | |
1154 | rds_message_addref(rm); | |
1155 | rm->m_inc.i_conn = conn; | |
1156 | ||
1157 | rds_message_populate_header(&rm->m_inc.i_hdr, 0, dport, | |
1158 | conn->c_next_tx_seq); | |
1159 | conn->c_next_tx_seq++; | |
1160 | spin_unlock_irqrestore(&conn->c_lock, flags); | |
1161 | ||
1162 | rds_stats_inc(s_send_queued); | |
1163 | rds_stats_inc(s_send_pong); | |
1164 | ||
acfcd4d4 | 1165 | if (!test_bit(RDS_LL_SEND_FULL, &conn->c_flags)) |
5175a5e7 | 1166 | queue_delayed_work(rds_wq, &conn->c_send_w, 0); |
acfcd4d4 | 1167 | |
5c115590 AG |
1168 | rds_message_put(rm); |
1169 | return 0; | |
1170 | ||
1171 | out: | |
1172 | if (rm) | |
1173 | rds_message_put(rm); | |
1174 | return ret; | |
1175 | } |