]>
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> |
4bebdd7a | 41 | #include <linux/sizes.h> |
5c115590 | 42 | |
0cb43965 | 43 | #include "rds_single_path.h" |
5c115590 | 44 | #include "rds.h" |
5c115590 AG |
45 | |
46 | /* When transmitting messages in rds_send_xmit, we need to emerge from | |
47 | * time to time and briefly release the CPU. Otherwise the softlock watchdog | |
48 | * will kick our shin. | |
49 | * Also, it seems fairer to not let one busy connection stall all the | |
50 | * others. | |
51 | * | |
52 | * send_batch_count is the number of times we'll loop in send_xmit. Setting | |
53 | * it to 0 will restore the old behavior (where we looped until we had | |
54 | * drained the queue). | |
55 | */ | |
4bebdd7a | 56 | static int send_batch_count = SZ_1K; |
5c115590 AG |
57 | module_param(send_batch_count, int, 0444); |
58 | MODULE_PARM_DESC(send_batch_count, " batch factor when working the send queue"); | |
59 | ||
ff51bf84 | 60 | static void rds_send_remove_from_sock(struct list_head *messages, int status); |
61 | ||
5c115590 | 62 | /* |
0f4b1c7e ZB |
63 | * Reset the send state. Callers must ensure that this doesn't race with |
64 | * rds_send_xmit(). | |
5c115590 | 65 | */ |
4e9b551c | 66 | static void rds_send_path_reset(struct rds_conn_path *cp) |
5c115590 AG |
67 | { |
68 | struct rds_message *rm, *tmp; | |
69 | unsigned long flags; | |
70 | ||
4e9b551c SV |
71 | if (cp->cp_xmit_rm) { |
72 | rm = cp->cp_xmit_rm; | |
73 | cp->cp_xmit_rm = NULL; | |
5c115590 AG |
74 | /* Tell the user the RDMA op is no longer mapped by the |
75 | * transport. This isn't entirely true (it's flushed out | |
76 | * independently) but as the connection is down, there's | |
77 | * no ongoing RDMA to/from that memory */ | |
7e3f2952 | 78 | rds_message_unmapped(rm); |
7e3f2952 | 79 | rds_message_put(rm); |
5c115590 | 80 | } |
7e3f2952 | 81 | |
4e9b551c SV |
82 | cp->cp_xmit_sg = 0; |
83 | cp->cp_xmit_hdr_off = 0; | |
84 | cp->cp_xmit_data_off = 0; | |
85 | cp->cp_xmit_atomic_sent = 0; | |
86 | cp->cp_xmit_rdma_sent = 0; | |
87 | cp->cp_xmit_data_sent = 0; | |
5c115590 | 88 | |
4e9b551c | 89 | cp->cp_conn->c_map_queued = 0; |
5c115590 | 90 | |
4e9b551c SV |
91 | cp->cp_unacked_packets = rds_sysctl_max_unacked_packets; |
92 | cp->cp_unacked_bytes = rds_sysctl_max_unacked_bytes; | |
5c115590 AG |
93 | |
94 | /* Mark messages as retransmissions, and move them to the send q */ | |
4e9b551c SV |
95 | spin_lock_irqsave(&cp->cp_lock, flags); |
96 | list_for_each_entry_safe(rm, tmp, &cp->cp_retrans, m_conn_item) { | |
5c115590 AG |
97 | set_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags); |
98 | set_bit(RDS_MSG_RETRANSMITTED, &rm->m_flags); | |
99 | } | |
4e9b551c SV |
100 | list_splice_init(&cp->cp_retrans, &cp->cp_send_queue); |
101 | spin_unlock_irqrestore(&cp->cp_lock, flags); | |
102 | } | |
103 | ||
104 | void rds_send_reset(struct rds_connection *conn) | |
105 | { | |
106 | rds_send_path_reset(&conn->c_path[0]); | |
5c115590 | 107 | } |
0b6f760c | 108 | EXPORT_SYMBOL_GPL(rds_send_reset); |
5c115590 | 109 | |
0f4b1c7e ZB |
110 | static int acquire_in_xmit(struct rds_connection *conn) |
111 | { | |
112 | return test_and_set_bit(RDS_IN_XMIT, &conn->c_flags) == 0; | |
113 | } | |
114 | ||
115 | static void release_in_xmit(struct rds_connection *conn) | |
116 | { | |
117 | clear_bit(RDS_IN_XMIT, &conn->c_flags); | |
4e857c58 | 118 | smp_mb__after_atomic(); |
0f4b1c7e ZB |
119 | /* |
120 | * We don't use wait_on_bit()/wake_up_bit() because our waking is in a | |
121 | * hot path and finding waiters is very rare. We don't want to walk | |
122 | * the system-wide hashed waitqueue buckets in the fast path only to | |
123 | * almost never find waiters. | |
124 | */ | |
125 | if (waitqueue_active(&conn->c_waitq)) | |
126 | wake_up_all(&conn->c_waitq); | |
127 | } | |
128 | ||
5c115590 | 129 | /* |
25985edc | 130 | * We're making the conscious trade-off here to only send one message |
5c115590 AG |
131 | * down the connection at a time. |
132 | * Pro: | |
133 | * - tx queueing is a simple fifo list | |
134 | * - reassembly is optional and easily done by transports per conn | |
135 | * - no per flow rx lookup at all, straight to the socket | |
136 | * - less per-frag memory and wire overhead | |
137 | * Con: | |
138 | * - queued acks can be delayed behind large messages | |
139 | * Depends: | |
140 | * - small message latency is higher behind queued large messages | |
141 | * - large message latency isn't starved by intervening small sends | |
142 | */ | |
143 | int rds_send_xmit(struct rds_connection *conn) | |
144 | { | |
145 | struct rds_message *rm; | |
146 | unsigned long flags; | |
147 | unsigned int tmp; | |
5c115590 AG |
148 | struct scatterlist *sg; |
149 | int ret = 0; | |
5c115590 | 150 | LIST_HEAD(to_be_dropped); |
443be0e5 SV |
151 | int batch_count; |
152 | unsigned long send_gen = 0; | |
5c115590 | 153 | |
fcc5450c | 154 | restart: |
443be0e5 | 155 | batch_count = 0; |
049ee3f5 | 156 | |
5c115590 AG |
157 | /* |
158 | * sendmsg calls here after having queued its message on the send | |
159 | * queue. We only have one task feeding the connection at a time. If | |
160 | * another thread is already feeding the queue then we back off. This | |
161 | * avoids blocking the caller and trading per-connection data between | |
162 | * caches per message. | |
5c115590 | 163 | */ |
0f4b1c7e | 164 | if (!acquire_in_xmit(conn)) { |
049ee3f5 | 165 | rds_stats_inc(s_send_lock_contention); |
5c115590 AG |
166 | ret = -ENOMEM; |
167 | goto out; | |
168 | } | |
0f4b1c7e | 169 | |
443be0e5 SV |
170 | /* |
171 | * we record the send generation after doing the xmit acquire. | |
172 | * if someone else manages to jump in and do some work, we'll use | |
173 | * this to avoid a goto restart farther down. | |
174 | * | |
175 | * The acquire_in_xmit() check above ensures that only one | |
176 | * caller can increment c_send_gen at any time. | |
177 | */ | |
178 | conn->c_send_gen++; | |
179 | send_gen = conn->c_send_gen; | |
180 | ||
0f4b1c7e ZB |
181 | /* |
182 | * rds_conn_shutdown() sets the conn state and then tests RDS_IN_XMIT, | |
183 | * we do the opposite to avoid races. | |
184 | */ | |
185 | if (!rds_conn_up(conn)) { | |
186 | release_in_xmit(conn); | |
187 | ret = 0; | |
188 | goto out; | |
189 | } | |
5c115590 AG |
190 | |
191 | if (conn->c_trans->xmit_prepare) | |
192 | conn->c_trans->xmit_prepare(conn); | |
193 | ||
194 | /* | |
195 | * spin trying to push headers and data down the connection until | |
5b2366bd | 196 | * the connection doesn't make forward progress. |
5c115590 | 197 | */ |
fcc5450c | 198 | while (1) { |
5c115590 | 199 | |
5c115590 | 200 | rm = conn->c_xmit_rm; |
5c115590 | 201 | |
5b2366bd AG |
202 | /* |
203 | * If between sending messages, we can send a pending congestion | |
204 | * map update. | |
5c115590 | 205 | */ |
8690bfa1 | 206 | if (!rm && test_and_clear_bit(0, &conn->c_map_queued)) { |
77dd550e AG |
207 | rm = rds_cong_update_alloc(conn); |
208 | if (IS_ERR(rm)) { | |
209 | ret = PTR_ERR(rm); | |
210 | break; | |
5b2366bd | 211 | } |
77dd550e AG |
212 | rm->data.op_active = 1; |
213 | ||
214 | conn->c_xmit_rm = rm; | |
5c115590 AG |
215 | } |
216 | ||
217 | /* | |
5b2366bd | 218 | * If not already working on one, grab the next message. |
5c115590 AG |
219 | * |
220 | * c_xmit_rm holds a ref while we're sending this message down | |
221 | * the connction. We can use this ref while holding the | |
222 | * send_sem.. rds_send_reset() is serialized with it. | |
223 | */ | |
8690bfa1 | 224 | if (!rm) { |
5c115590 AG |
225 | unsigned int len; |
226 | ||
443be0e5 SV |
227 | batch_count++; |
228 | ||
229 | /* we want to process as big a batch as we can, but | |
230 | * we also want to avoid softlockups. If we've been | |
231 | * through a lot of messages, lets back off and see | |
232 | * if anyone else jumps in | |
233 | */ | |
4bebdd7a | 234 | if (batch_count >= send_batch_count) |
443be0e5 SV |
235 | goto over_batch; |
236 | ||
0f4b1c7e | 237 | spin_lock_irqsave(&conn->c_lock, flags); |
5c115590 AG |
238 | |
239 | if (!list_empty(&conn->c_send_queue)) { | |
240 | rm = list_entry(conn->c_send_queue.next, | |
241 | struct rds_message, | |
242 | m_conn_item); | |
243 | rds_message_addref(rm); | |
244 | ||
245 | /* | |
246 | * Move the message from the send queue to the retransmit | |
247 | * list right away. | |
248 | */ | |
249 | list_move_tail(&rm->m_conn_item, &conn->c_retrans); | |
250 | } | |
251 | ||
0f4b1c7e | 252 | spin_unlock_irqrestore(&conn->c_lock, flags); |
5c115590 | 253 | |
fcc5450c | 254 | if (!rm) |
5c115590 | 255 | break; |
5c115590 AG |
256 | |
257 | /* Unfortunately, the way Infiniband deals with | |
258 | * RDMA to a bad MR key is by moving the entire | |
259 | * queue pair to error state. We cold possibly | |
260 | * recover from that, but right now we drop the | |
261 | * connection. | |
262 | * Therefore, we never retransmit messages with RDMA ops. | |
263 | */ | |
f8b3aaf2 | 264 | if (rm->rdma.op_active && |
f64f9e71 | 265 | test_bit(RDS_MSG_RETRANSMITTED, &rm->m_flags)) { |
0f4b1c7e | 266 | spin_lock_irqsave(&conn->c_lock, flags); |
5c115590 AG |
267 | if (test_and_clear_bit(RDS_MSG_ON_CONN, &rm->m_flags)) |
268 | list_move(&rm->m_conn_item, &to_be_dropped); | |
0f4b1c7e | 269 | spin_unlock_irqrestore(&conn->c_lock, flags); |
5c115590 AG |
270 | continue; |
271 | } | |
272 | ||
273 | /* Require an ACK every once in a while */ | |
274 | len = ntohl(rm->m_inc.i_hdr.h_len); | |
f64f9e71 JP |
275 | if (conn->c_unacked_packets == 0 || |
276 | conn->c_unacked_bytes < len) { | |
5c115590 AG |
277 | __set_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags); |
278 | ||
279 | conn->c_unacked_packets = rds_sysctl_max_unacked_packets; | |
280 | conn->c_unacked_bytes = rds_sysctl_max_unacked_bytes; | |
281 | rds_stats_inc(s_send_ack_required); | |
282 | } else { | |
283 | conn->c_unacked_bytes -= len; | |
284 | conn->c_unacked_packets--; | |
285 | } | |
286 | ||
287 | conn->c_xmit_rm = rm; | |
288 | } | |
289 | ||
2c3a5f9a AG |
290 | /* The transport either sends the whole rdma or none of it */ |
291 | if (rm->rdma.op_active && !conn->c_xmit_rdma_sent) { | |
ff3d7d36 | 292 | rm->m_final_op = &rm->rdma; |
4f73113c | 293 | /* The transport owns the mapped memory for now. |
294 | * You can't unmap it while it's on the send queue | |
295 | */ | |
296 | set_bit(RDS_MSG_MAPPED, &rm->m_flags); | |
2c3a5f9a | 297 | ret = conn->c_trans->xmit_rdma(conn, &rm->rdma); |
4f73113c | 298 | if (ret) { |
299 | clear_bit(RDS_MSG_MAPPED, &rm->m_flags); | |
300 | wake_up_interruptible(&rm->m_flush_wait); | |
15133f6e | 301 | break; |
4f73113c | 302 | } |
2c3a5f9a AG |
303 | conn->c_xmit_rdma_sent = 1; |
304 | ||
15133f6e AG |
305 | } |
306 | ||
2c3a5f9a | 307 | if (rm->atomic.op_active && !conn->c_xmit_atomic_sent) { |
ff3d7d36 | 308 | rm->m_final_op = &rm->atomic; |
4f73113c | 309 | /* The transport owns the mapped memory for now. |
310 | * You can't unmap it while it's on the send queue | |
311 | */ | |
312 | set_bit(RDS_MSG_MAPPED, &rm->m_flags); | |
ff3d7d36 | 313 | ret = conn->c_trans->xmit_atomic(conn, &rm->atomic); |
4f73113c | 314 | if (ret) { |
315 | clear_bit(RDS_MSG_MAPPED, &rm->m_flags); | |
316 | wake_up_interruptible(&rm->m_flush_wait); | |
5c115590 | 317 | break; |
4f73113c | 318 | } |
2c3a5f9a | 319 | conn->c_xmit_atomic_sent = 1; |
ff3d7d36 | 320 | |
5c115590 AG |
321 | } |
322 | ||
2c3a5f9a AG |
323 | /* |
324 | * A number of cases require an RDS header to be sent | |
325 | * even if there is no data. | |
326 | * We permit 0-byte sends; rds-ping depends on this. | |
327 | * However, if there are exclusively attached silent ops, | |
328 | * we skip the hdr/data send, to enable silent operation. | |
329 | */ | |
330 | if (rm->data.op_nents == 0) { | |
331 | int ops_present; | |
332 | int all_ops_are_silent = 1; | |
333 | ||
334 | ops_present = (rm->atomic.op_active || rm->rdma.op_active); | |
335 | if (rm->atomic.op_active && !rm->atomic.op_silent) | |
336 | all_ops_are_silent = 0; | |
337 | if (rm->rdma.op_active && !rm->rdma.op_silent) | |
338 | all_ops_are_silent = 0; | |
339 | ||
340 | if (ops_present && all_ops_are_silent | |
341 | && !rm->m_rdma_cookie) | |
342 | rm->data.op_active = 0; | |
343 | } | |
344 | ||
5b2366bd | 345 | if (rm->data.op_active && !conn->c_xmit_data_sent) { |
ff3d7d36 | 346 | rm->m_final_op = &rm->data; |
5c115590 AG |
347 | ret = conn->c_trans->xmit(conn, rm, |
348 | conn->c_xmit_hdr_off, | |
349 | conn->c_xmit_sg, | |
350 | conn->c_xmit_data_off); | |
351 | if (ret <= 0) | |
352 | break; | |
353 | ||
354 | if (conn->c_xmit_hdr_off < sizeof(struct rds_header)) { | |
355 | tmp = min_t(int, ret, | |
356 | sizeof(struct rds_header) - | |
357 | conn->c_xmit_hdr_off); | |
358 | conn->c_xmit_hdr_off += tmp; | |
359 | ret -= tmp; | |
360 | } | |
361 | ||
6c7cc6e4 | 362 | sg = &rm->data.op_sg[conn->c_xmit_sg]; |
5c115590 AG |
363 | while (ret) { |
364 | tmp = min_t(int, ret, sg->length - | |
365 | conn->c_xmit_data_off); | |
366 | conn->c_xmit_data_off += tmp; | |
367 | ret -= tmp; | |
368 | if (conn->c_xmit_data_off == sg->length) { | |
369 | conn->c_xmit_data_off = 0; | |
370 | sg++; | |
371 | conn->c_xmit_sg++; | |
372 | BUG_ON(ret != 0 && | |
6c7cc6e4 | 373 | conn->c_xmit_sg == rm->data.op_nents); |
5c115590 AG |
374 | } |
375 | } | |
5b2366bd AG |
376 | |
377 | if (conn->c_xmit_hdr_off == sizeof(struct rds_header) && | |
378 | (conn->c_xmit_sg == rm->data.op_nents)) | |
379 | conn->c_xmit_data_sent = 1; | |
380 | } | |
381 | ||
382 | /* | |
383 | * A rm will only take multiple times through this loop | |
384 | * if there is a data op. Thus, if the data is sent (or there was | |
385 | * none), then we're done with the rm. | |
386 | */ | |
387 | if (!rm->data.op_active || conn->c_xmit_data_sent) { | |
388 | conn->c_xmit_rm = NULL; | |
389 | conn->c_xmit_sg = 0; | |
390 | conn->c_xmit_hdr_off = 0; | |
391 | conn->c_xmit_data_off = 0; | |
392 | conn->c_xmit_rdma_sent = 0; | |
393 | conn->c_xmit_atomic_sent = 0; | |
394 | conn->c_xmit_data_sent = 0; | |
395 | ||
396 | rds_message_put(rm); | |
5c115590 AG |
397 | } |
398 | } | |
399 | ||
443be0e5 | 400 | over_batch: |
5c115590 AG |
401 | if (conn->c_trans->xmit_complete) |
402 | conn->c_trans->xmit_complete(conn); | |
0f4b1c7e | 403 | release_in_xmit(conn); |
5c115590 | 404 | |
2ad8099b AG |
405 | /* Nuke any messages we decided not to retransmit. */ |
406 | if (!list_empty(&to_be_dropped)) { | |
407 | /* irqs on here, so we can put(), unlike above */ | |
408 | list_for_each_entry(rm, &to_be_dropped, m_conn_item) | |
409 | rds_message_put(rm); | |
410 | rds_send_remove_from_sock(&to_be_dropped, RDS_RDMA_DROPPED); | |
411 | } | |
412 | ||
fcc5450c | 413 | /* |
0f4b1c7e ZB |
414 | * Other senders can queue a message after we last test the send queue |
415 | * but before we clear RDS_IN_XMIT. In that case they'd back off and | |
416 | * not try and send their newly queued message. We need to check the | |
417 | * send queue after having cleared RDS_IN_XMIT so that their message | |
418 | * doesn't get stuck on the send queue. | |
fcc5450c AG |
419 | * |
420 | * If the transport cannot continue (i.e ret != 0), then it must | |
421 | * call us when more room is available, such as from the tx | |
422 | * completion handler. | |
443be0e5 SV |
423 | * |
424 | * We have an extra generation check here so that if someone manages | |
425 | * to jump in after our release_in_xmit, we'll see that they have done | |
426 | * some work and we will skip our goto | |
fcc5450c AG |
427 | */ |
428 | if (ret == 0) { | |
9e29db0e | 429 | smp_mb(); |
0c484240 | 430 | if ((test_bit(0, &conn->c_map_queued) || |
431 | !list_empty(&conn->c_send_queue)) && | |
443be0e5 | 432 | send_gen == conn->c_send_gen) { |
049ee3f5 | 433 | rds_stats_inc(s_send_lock_queue_raced); |
4bebdd7a SS |
434 | if (batch_count < send_batch_count) |
435 | goto restart; | |
436 | queue_delayed_work(rds_wq, &conn->c_send_w, 1); | |
5c115590 | 437 | } |
5c115590 AG |
438 | } |
439 | out: | |
440 | return ret; | |
441 | } | |
0c28c045 | 442 | EXPORT_SYMBOL_GPL(rds_send_xmit); |
5c115590 AG |
443 | |
444 | static void rds_send_sndbuf_remove(struct rds_sock *rs, struct rds_message *rm) | |
445 | { | |
446 | u32 len = be32_to_cpu(rm->m_inc.i_hdr.h_len); | |
447 | ||
448 | assert_spin_locked(&rs->rs_lock); | |
449 | ||
450 | BUG_ON(rs->rs_snd_bytes < len); | |
451 | rs->rs_snd_bytes -= len; | |
452 | ||
453 | if (rs->rs_snd_bytes == 0) | |
454 | rds_stats_inc(s_send_queue_empty); | |
455 | } | |
456 | ||
457 | static inline int rds_send_is_acked(struct rds_message *rm, u64 ack, | |
458 | is_acked_func is_acked) | |
459 | { | |
460 | if (is_acked) | |
461 | return is_acked(rm, ack); | |
462 | return be64_to_cpu(rm->m_inc.i_hdr.h_sequence) <= ack; | |
463 | } | |
464 | ||
5c115590 AG |
465 | /* |
466 | * This is pretty similar to what happens below in the ACK | |
467 | * handling code - except that we call here as soon as we get | |
468 | * the IB send completion on the RDMA op and the accompanying | |
469 | * message. | |
470 | */ | |
471 | void rds_rdma_send_complete(struct rds_message *rm, int status) | |
472 | { | |
473 | struct rds_sock *rs = NULL; | |
f8b3aaf2 | 474 | struct rm_rdma_op *ro; |
5c115590 | 475 | struct rds_notifier *notifier; |
9de0864c | 476 | unsigned long flags; |
5c115590 | 477 | |
9de0864c | 478 | spin_lock_irqsave(&rm->m_rs_lock, flags); |
5c115590 | 479 | |
f8b3aaf2 | 480 | ro = &rm->rdma; |
f64f9e71 | 481 | if (test_bit(RDS_MSG_ON_SOCK, &rm->m_flags) && |
f8b3aaf2 AG |
482 | ro->op_active && ro->op_notify && ro->op_notifier) { |
483 | notifier = ro->op_notifier; | |
5c115590 AG |
484 | rs = rm->m_rs; |
485 | sock_hold(rds_rs_to_sk(rs)); | |
486 | ||
487 | notifier->n_status = status; | |
488 | spin_lock(&rs->rs_lock); | |
489 | list_add_tail(¬ifier->n_list, &rs->rs_notify_queue); | |
490 | spin_unlock(&rs->rs_lock); | |
491 | ||
f8b3aaf2 | 492 | ro->op_notifier = NULL; |
5c115590 AG |
493 | } |
494 | ||
9de0864c | 495 | spin_unlock_irqrestore(&rm->m_rs_lock, flags); |
5c115590 AG |
496 | |
497 | if (rs) { | |
498 | rds_wake_sk_sleep(rs); | |
499 | sock_put(rds_rs_to_sk(rs)); | |
500 | } | |
501 | } | |
616b757a | 502 | EXPORT_SYMBOL_GPL(rds_rdma_send_complete); |
5c115590 | 503 | |
15133f6e AG |
504 | /* |
505 | * Just like above, except looks at atomic op | |
506 | */ | |
507 | void rds_atomic_send_complete(struct rds_message *rm, int status) | |
508 | { | |
509 | struct rds_sock *rs = NULL; | |
510 | struct rm_atomic_op *ao; | |
511 | struct rds_notifier *notifier; | |
cf4b7389 | 512 | unsigned long flags; |
15133f6e | 513 | |
cf4b7389 | 514 | spin_lock_irqsave(&rm->m_rs_lock, flags); |
15133f6e AG |
515 | |
516 | ao = &rm->atomic; | |
517 | if (test_bit(RDS_MSG_ON_SOCK, &rm->m_flags) | |
518 | && ao->op_active && ao->op_notify && ao->op_notifier) { | |
519 | notifier = ao->op_notifier; | |
520 | rs = rm->m_rs; | |
521 | sock_hold(rds_rs_to_sk(rs)); | |
522 | ||
523 | notifier->n_status = status; | |
524 | spin_lock(&rs->rs_lock); | |
525 | list_add_tail(¬ifier->n_list, &rs->rs_notify_queue); | |
526 | spin_unlock(&rs->rs_lock); | |
527 | ||
528 | ao->op_notifier = NULL; | |
529 | } | |
530 | ||
cf4b7389 | 531 | spin_unlock_irqrestore(&rm->m_rs_lock, flags); |
15133f6e AG |
532 | |
533 | if (rs) { | |
534 | rds_wake_sk_sleep(rs); | |
535 | sock_put(rds_rs_to_sk(rs)); | |
536 | } | |
537 | } | |
538 | EXPORT_SYMBOL_GPL(rds_atomic_send_complete); | |
539 | ||
5c115590 AG |
540 | /* |
541 | * This is the same as rds_rdma_send_complete except we | |
542 | * don't do any locking - we have all the ingredients (message, | |
543 | * socket, socket lock) and can just move the notifier. | |
544 | */ | |
545 | static inline void | |
940786eb | 546 | __rds_send_complete(struct rds_sock *rs, struct rds_message *rm, int status) |
5c115590 | 547 | { |
f8b3aaf2 | 548 | struct rm_rdma_op *ro; |
940786eb | 549 | struct rm_atomic_op *ao; |
5c115590 | 550 | |
f8b3aaf2 AG |
551 | ro = &rm->rdma; |
552 | if (ro->op_active && ro->op_notify && ro->op_notifier) { | |
553 | ro->op_notifier->n_status = status; | |
554 | list_add_tail(&ro->op_notifier->n_list, &rs->rs_notify_queue); | |
555 | ro->op_notifier = NULL; | |
5c115590 AG |
556 | } |
557 | ||
940786eb AG |
558 | ao = &rm->atomic; |
559 | if (ao->op_active && ao->op_notify && ao->op_notifier) { | |
560 | ao->op_notifier->n_status = status; | |
561 | list_add_tail(&ao->op_notifier->n_list, &rs->rs_notify_queue); | |
562 | ao->op_notifier = NULL; | |
563 | } | |
564 | ||
5c115590 AG |
565 | /* No need to wake the app - caller does this */ |
566 | } | |
567 | ||
5c115590 AG |
568 | /* |
569 | * This removes messages from the socket's list if they're on it. The list | |
570 | * argument must be private to the caller, we must be able to modify it | |
571 | * without locks. The messages must have a reference held for their | |
572 | * position on the list. This function will drop that reference after | |
573 | * removing the messages from the 'messages' list regardless of if it found | |
574 | * the messages on the socket list or not. | |
575 | */ | |
ff51bf84 | 576 | static void rds_send_remove_from_sock(struct list_head *messages, int status) |
5c115590 | 577 | { |
561c7df6 | 578 | unsigned long flags; |
5c115590 AG |
579 | struct rds_sock *rs = NULL; |
580 | struct rds_message *rm; | |
581 | ||
5c115590 | 582 | while (!list_empty(messages)) { |
561c7df6 AG |
583 | int was_on_sock = 0; |
584 | ||
5c115590 AG |
585 | rm = list_entry(messages->next, struct rds_message, |
586 | m_conn_item); | |
587 | list_del_init(&rm->m_conn_item); | |
588 | ||
589 | /* | |
590 | * If we see this flag cleared then we're *sure* that someone | |
591 | * else beat us to removing it from the sock. If we race | |
592 | * with their flag update we'll get the lock and then really | |
593 | * see that the flag has been cleared. | |
594 | * | |
595 | * The message spinlock makes sure nobody clears rm->m_rs | |
596 | * while we're messing with it. It does not prevent the | |
597 | * message from being removed from the socket, though. | |
598 | */ | |
561c7df6 | 599 | spin_lock_irqsave(&rm->m_rs_lock, flags); |
5c115590 AG |
600 | if (!test_bit(RDS_MSG_ON_SOCK, &rm->m_flags)) |
601 | goto unlock_and_drop; | |
602 | ||
603 | if (rs != rm->m_rs) { | |
604 | if (rs) { | |
5c115590 AG |
605 | rds_wake_sk_sleep(rs); |
606 | sock_put(rds_rs_to_sk(rs)); | |
607 | } | |
608 | rs = rm->m_rs; | |
593cbb3e HK |
609 | if (rs) |
610 | sock_hold(rds_rs_to_sk(rs)); | |
5c115590 | 611 | } |
593cbb3e HK |
612 | if (!rs) |
613 | goto unlock_and_drop; | |
048c15e6 | 614 | spin_lock(&rs->rs_lock); |
5c115590 AG |
615 | |
616 | if (test_and_clear_bit(RDS_MSG_ON_SOCK, &rm->m_flags)) { | |
f8b3aaf2 | 617 | struct rm_rdma_op *ro = &rm->rdma; |
5c115590 AG |
618 | struct rds_notifier *notifier; |
619 | ||
620 | list_del_init(&rm->m_sock_item); | |
621 | rds_send_sndbuf_remove(rs, rm); | |
622 | ||
f8b3aaf2 AG |
623 | if (ro->op_active && ro->op_notifier && |
624 | (ro->op_notify || (ro->op_recverr && status))) { | |
625 | notifier = ro->op_notifier; | |
5c115590 AG |
626 | list_add_tail(¬ifier->n_list, |
627 | &rs->rs_notify_queue); | |
628 | if (!notifier->n_status) | |
629 | notifier->n_status = status; | |
f8b3aaf2 | 630 | rm->rdma.op_notifier = NULL; |
5c115590 | 631 | } |
561c7df6 | 632 | was_on_sock = 1; |
5c115590 AG |
633 | rm->m_rs = NULL; |
634 | } | |
048c15e6 | 635 | spin_unlock(&rs->rs_lock); |
5c115590 AG |
636 | |
637 | unlock_and_drop: | |
561c7df6 | 638 | spin_unlock_irqrestore(&rm->m_rs_lock, flags); |
5c115590 | 639 | rds_message_put(rm); |
561c7df6 AG |
640 | if (was_on_sock) |
641 | rds_message_put(rm); | |
5c115590 AG |
642 | } |
643 | ||
644 | if (rs) { | |
5c115590 AG |
645 | rds_wake_sk_sleep(rs); |
646 | sock_put(rds_rs_to_sk(rs)); | |
647 | } | |
5c115590 AG |
648 | } |
649 | ||
650 | /* | |
651 | * Transports call here when they've determined that the receiver queued | |
652 | * messages up to, and including, the given sequence number. Messages are | |
653 | * moved to the retrans queue when rds_send_xmit picks them off the send | |
654 | * queue. This means that in the TCP case, the message may not have been | |
655 | * assigned the m_ack_seq yet - but that's fine as long as tcp_is_acked | |
656 | * checks the RDS_MSG_HAS_ACK_SEQ bit. | |
5c115590 | 657 | */ |
5c3d274c SV |
658 | void rds_send_path_drop_acked(struct rds_conn_path *cp, u64 ack, |
659 | is_acked_func is_acked) | |
5c115590 AG |
660 | { |
661 | struct rds_message *rm, *tmp; | |
662 | unsigned long flags; | |
663 | LIST_HEAD(list); | |
664 | ||
5c3d274c | 665 | spin_lock_irqsave(&cp->cp_lock, flags); |
5c115590 | 666 | |
5c3d274c | 667 | list_for_each_entry_safe(rm, tmp, &cp->cp_retrans, m_conn_item) { |
5c115590 AG |
668 | if (!rds_send_is_acked(rm, ack, is_acked)) |
669 | break; | |
670 | ||
671 | list_move(&rm->m_conn_item, &list); | |
672 | clear_bit(RDS_MSG_ON_CONN, &rm->m_flags); | |
673 | } | |
674 | ||
675 | /* order flag updates with spin locks */ | |
676 | if (!list_empty(&list)) | |
4e857c58 | 677 | smp_mb__after_atomic(); |
5c115590 | 678 | |
5c3d274c | 679 | spin_unlock_irqrestore(&cp->cp_lock, flags); |
5c115590 AG |
680 | |
681 | /* now remove the messages from the sock list as needed */ | |
682 | rds_send_remove_from_sock(&list, RDS_RDMA_SUCCESS); | |
683 | } | |
5c3d274c SV |
684 | EXPORT_SYMBOL_GPL(rds_send_path_drop_acked); |
685 | ||
686 | void rds_send_drop_acked(struct rds_connection *conn, u64 ack, | |
687 | is_acked_func is_acked) | |
688 | { | |
689 | WARN_ON(conn->c_trans->t_mp_capable); | |
690 | rds_send_path_drop_acked(&conn->c_path[0], ack, is_acked); | |
691 | } | |
616b757a | 692 | EXPORT_SYMBOL_GPL(rds_send_drop_acked); |
5c115590 AG |
693 | |
694 | void rds_send_drop_to(struct rds_sock *rs, struct sockaddr_in *dest) | |
695 | { | |
696 | struct rds_message *rm, *tmp; | |
697 | struct rds_connection *conn; | |
7c82eaf0 | 698 | unsigned long flags; |
5c115590 | 699 | LIST_HEAD(list); |
5c115590 AG |
700 | |
701 | /* get all the messages we're dropping under the rs lock */ | |
702 | spin_lock_irqsave(&rs->rs_lock, flags); | |
703 | ||
704 | list_for_each_entry_safe(rm, tmp, &rs->rs_send_queue, m_sock_item) { | |
705 | if (dest && (dest->sin_addr.s_addr != rm->m_daddr || | |
706 | dest->sin_port != rm->m_inc.i_hdr.h_dport)) | |
707 | continue; | |
708 | ||
5c115590 AG |
709 | list_move(&rm->m_sock_item, &list); |
710 | rds_send_sndbuf_remove(rs, rm); | |
711 | clear_bit(RDS_MSG_ON_SOCK, &rm->m_flags); | |
5c115590 AG |
712 | } |
713 | ||
714 | /* order flag updates with the rs lock */ | |
4e857c58 | 715 | smp_mb__after_atomic(); |
5c115590 AG |
716 | |
717 | spin_unlock_irqrestore(&rs->rs_lock, flags); | |
718 | ||
7c82eaf0 AG |
719 | if (list_empty(&list)) |
720 | return; | |
5c115590 | 721 | |
7c82eaf0 | 722 | /* Remove the messages from the conn */ |
5c115590 | 723 | list_for_each_entry(rm, &list, m_sock_item) { |
7c82eaf0 AG |
724 | |
725 | conn = rm->m_inc.i_conn; | |
5c115590 | 726 | |
9de0864c | 727 | spin_lock_irqsave(&conn->c_lock, flags); |
5c115590 | 728 | /* |
7c82eaf0 AG |
729 | * Maybe someone else beat us to removing rm from the conn. |
730 | * If we race with their flag update we'll get the lock and | |
731 | * then really see that the flag has been cleared. | |
5c115590 | 732 | */ |
7c82eaf0 AG |
733 | if (!test_and_clear_bit(RDS_MSG_ON_CONN, &rm->m_flags)) { |
734 | spin_unlock_irqrestore(&conn->c_lock, flags); | |
593cbb3e HK |
735 | spin_lock_irqsave(&rm->m_rs_lock, flags); |
736 | rm->m_rs = NULL; | |
737 | spin_unlock_irqrestore(&rm->m_rs_lock, flags); | |
5c115590 | 738 | continue; |
5c115590 | 739 | } |
9de0864c AG |
740 | list_del_init(&rm->m_conn_item); |
741 | spin_unlock_irqrestore(&conn->c_lock, flags); | |
5c115590 | 742 | |
7c82eaf0 AG |
743 | /* |
744 | * Couldn't grab m_rs_lock in top loop (lock ordering), | |
745 | * but we can now. | |
746 | */ | |
9de0864c | 747 | spin_lock_irqsave(&rm->m_rs_lock, flags); |
5c115590 | 748 | |
7c82eaf0 | 749 | spin_lock(&rs->rs_lock); |
940786eb | 750 | __rds_send_complete(rs, rm, RDS_RDMA_CANCELED); |
7c82eaf0 AG |
751 | spin_unlock(&rs->rs_lock); |
752 | ||
753 | rm->m_rs = NULL; | |
9de0864c | 754 | spin_unlock_irqrestore(&rm->m_rs_lock, flags); |
7c82eaf0 | 755 | |
7c82eaf0 | 756 | rds_message_put(rm); |
7c82eaf0 | 757 | } |
5c115590 | 758 | |
7c82eaf0 | 759 | rds_wake_sk_sleep(rs); |
550a8002 | 760 | |
5c115590 AG |
761 | while (!list_empty(&list)) { |
762 | rm = list_entry(list.next, struct rds_message, m_sock_item); | |
763 | list_del_init(&rm->m_sock_item); | |
5c115590 | 764 | rds_message_wait(rm); |
dfcec251 | 765 | |
766 | /* just in case the code above skipped this message | |
767 | * because RDS_MSG_ON_CONN wasn't set, run it again here | |
768 | * taking m_rs_lock is the only thing that keeps us | |
769 | * from racing with ack processing. | |
770 | */ | |
771 | spin_lock_irqsave(&rm->m_rs_lock, flags); | |
772 | ||
773 | spin_lock(&rs->rs_lock); | |
774 | __rds_send_complete(rs, rm, RDS_RDMA_CANCELED); | |
775 | spin_unlock(&rs->rs_lock); | |
776 | ||
777 | rm->m_rs = NULL; | |
778 | spin_unlock_irqrestore(&rm->m_rs_lock, flags); | |
779 | ||
5c115590 AG |
780 | rds_message_put(rm); |
781 | } | |
782 | } | |
783 | ||
784 | /* | |
785 | * we only want this to fire once so we use the callers 'queued'. It's | |
786 | * possible that another thread can race with us and remove the | |
787 | * message from the flow with RDS_CANCEL_SENT_TO. | |
788 | */ | |
789 | static int rds_send_queue_rm(struct rds_sock *rs, struct rds_connection *conn, | |
780a6d9e | 790 | struct rds_conn_path *cp, |
5c115590 AG |
791 | struct rds_message *rm, __be16 sport, |
792 | __be16 dport, int *queued) | |
793 | { | |
794 | unsigned long flags; | |
795 | u32 len; | |
796 | ||
797 | if (*queued) | |
798 | goto out; | |
799 | ||
800 | len = be32_to_cpu(rm->m_inc.i_hdr.h_len); | |
801 | ||
802 | /* this is the only place which holds both the socket's rs_lock | |
803 | * and the connection's c_lock */ | |
804 | spin_lock_irqsave(&rs->rs_lock, flags); | |
805 | ||
806 | /* | |
807 | * If there is a little space in sndbuf, we don't queue anything, | |
808 | * and userspace gets -EAGAIN. But poll() indicates there's send | |
809 | * room. This can lead to bad behavior (spinning) if snd_bytes isn't | |
810 | * freed up by incoming acks. So we check the *old* value of | |
811 | * rs_snd_bytes here to allow the last msg to exceed the buffer, | |
812 | * and poll() now knows no more data can be sent. | |
813 | */ | |
814 | if (rs->rs_snd_bytes < rds_sk_sndbuf(rs)) { | |
815 | rs->rs_snd_bytes += len; | |
816 | ||
817 | /* let recv side know we are close to send space exhaustion. | |
818 | * This is probably not the optimal way to do it, as this | |
819 | * means we set the flag on *all* messages as soon as our | |
820 | * throughput hits a certain threshold. | |
821 | */ | |
822 | if (rs->rs_snd_bytes >= rds_sk_sndbuf(rs) / 2) | |
823 | __set_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags); | |
824 | ||
825 | list_add_tail(&rm->m_sock_item, &rs->rs_send_queue); | |
826 | set_bit(RDS_MSG_ON_SOCK, &rm->m_flags); | |
827 | rds_message_addref(rm); | |
828 | rm->m_rs = rs; | |
829 | ||
830 | /* The code ordering is a little weird, but we're | |
831 | trying to minimize the time we hold c_lock */ | |
832 | rds_message_populate_header(&rm->m_inc.i_hdr, sport, dport, 0); | |
833 | rm->m_inc.i_conn = conn; | |
780a6d9e | 834 | rm->m_inc.i_conn_path = cp; |
5c115590 AG |
835 | rds_message_addref(rm); |
836 | ||
780a6d9e SV |
837 | spin_lock(&cp->cp_lock); |
838 | rm->m_inc.i_hdr.h_sequence = cpu_to_be64(cp->cp_next_tx_seq++); | |
839 | list_add_tail(&rm->m_conn_item, &cp->cp_send_queue); | |
5c115590 | 840 | set_bit(RDS_MSG_ON_CONN, &rm->m_flags); |
780a6d9e | 841 | spin_unlock(&cp->cp_lock); |
5c115590 AG |
842 | |
843 | rdsdebug("queued msg %p len %d, rs %p bytes %d seq %llu\n", | |
844 | rm, len, rs, rs->rs_snd_bytes, | |
845 | (unsigned long long)be64_to_cpu(rm->m_inc.i_hdr.h_sequence)); | |
846 | ||
847 | *queued = 1; | |
848 | } | |
849 | ||
850 | spin_unlock_irqrestore(&rs->rs_lock, flags); | |
851 | out: | |
852 | return *queued; | |
853 | } | |
854 | ||
fc445084 AG |
855 | /* |
856 | * rds_message is getting to be quite complicated, and we'd like to allocate | |
857 | * it all in one go. This figures out how big it needs to be up front. | |
858 | */ | |
859 | static int rds_rm_size(struct msghdr *msg, int data_len) | |
860 | { | |
ff87e97a | 861 | struct cmsghdr *cmsg; |
fc445084 | 862 | int size = 0; |
aa0a4ef4 | 863 | int cmsg_groups = 0; |
ff87e97a AG |
864 | int retval; |
865 | ||
f95b414e | 866 | for_each_cmsghdr(cmsg, msg) { |
ff87e97a AG |
867 | if (!CMSG_OK(msg, cmsg)) |
868 | return -EINVAL; | |
869 | ||
870 | if (cmsg->cmsg_level != SOL_RDS) | |
871 | continue; | |
872 | ||
873 | switch (cmsg->cmsg_type) { | |
874 | case RDS_CMSG_RDMA_ARGS: | |
aa0a4ef4 | 875 | cmsg_groups |= 1; |
ff87e97a AG |
876 | retval = rds_rdma_extra_size(CMSG_DATA(cmsg)); |
877 | if (retval < 0) | |
878 | return retval; | |
879 | size += retval; | |
aa0a4ef4 | 880 | |
ff87e97a AG |
881 | break; |
882 | ||
883 | case RDS_CMSG_RDMA_DEST: | |
884 | case RDS_CMSG_RDMA_MAP: | |
aa0a4ef4 | 885 | cmsg_groups |= 2; |
ff87e97a AG |
886 | /* these are valid but do no add any size */ |
887 | break; | |
888 | ||
15133f6e AG |
889 | case RDS_CMSG_ATOMIC_CSWP: |
890 | case RDS_CMSG_ATOMIC_FADD: | |
20c72bd5 AG |
891 | case RDS_CMSG_MASKED_ATOMIC_CSWP: |
892 | case RDS_CMSG_MASKED_ATOMIC_FADD: | |
aa0a4ef4 | 893 | cmsg_groups |= 1; |
15133f6e AG |
894 | size += sizeof(struct scatterlist); |
895 | break; | |
896 | ||
ff87e97a AG |
897 | default: |
898 | return -EINVAL; | |
899 | } | |
900 | ||
901 | } | |
fc445084 | 902 | |
ff87e97a | 903 | size += ceil(data_len, PAGE_SIZE) * sizeof(struct scatterlist); |
fc445084 | 904 | |
aa0a4ef4 AG |
905 | /* Ensure (DEST, MAP) are never used with (ARGS, ATOMIC) */ |
906 | if (cmsg_groups == 3) | |
907 | return -EINVAL; | |
908 | ||
fc445084 AG |
909 | return size; |
910 | } | |
911 | ||
5c115590 AG |
912 | static int rds_cmsg_send(struct rds_sock *rs, struct rds_message *rm, |
913 | struct msghdr *msg, int *allocated_mr) | |
914 | { | |
915 | struct cmsghdr *cmsg; | |
916 | int ret = 0; | |
917 | ||
f95b414e | 918 | for_each_cmsghdr(cmsg, msg) { |
5c115590 AG |
919 | if (!CMSG_OK(msg, cmsg)) |
920 | return -EINVAL; | |
921 | ||
922 | if (cmsg->cmsg_level != SOL_RDS) | |
923 | continue; | |
924 | ||
925 | /* As a side effect, RDMA_DEST and RDMA_MAP will set | |
15133f6e | 926 | * rm->rdma.m_rdma_cookie and rm->rdma.m_rdma_mr. |
5c115590 AG |
927 | */ |
928 | switch (cmsg->cmsg_type) { | |
929 | case RDS_CMSG_RDMA_ARGS: | |
930 | ret = rds_cmsg_rdma_args(rs, rm, cmsg); | |
931 | break; | |
932 | ||
933 | case RDS_CMSG_RDMA_DEST: | |
934 | ret = rds_cmsg_rdma_dest(rs, rm, cmsg); | |
935 | break; | |
936 | ||
937 | case RDS_CMSG_RDMA_MAP: | |
938 | ret = rds_cmsg_rdma_map(rs, rm, cmsg); | |
939 | if (!ret) | |
940 | *allocated_mr = 1; | |
941 | break; | |
15133f6e AG |
942 | case RDS_CMSG_ATOMIC_CSWP: |
943 | case RDS_CMSG_ATOMIC_FADD: | |
20c72bd5 AG |
944 | case RDS_CMSG_MASKED_ATOMIC_CSWP: |
945 | case RDS_CMSG_MASKED_ATOMIC_FADD: | |
15133f6e AG |
946 | ret = rds_cmsg_atomic(rs, rm, cmsg); |
947 | break; | |
5c115590 AG |
948 | |
949 | default: | |
950 | return -EINVAL; | |
951 | } | |
952 | ||
953 | if (ret) | |
954 | break; | |
955 | } | |
956 | ||
957 | return ret; | |
958 | } | |
959 | ||
1b784140 | 960 | int rds_sendmsg(struct socket *sock, struct msghdr *msg, size_t payload_len) |
5c115590 AG |
961 | { |
962 | struct sock *sk = sock->sk; | |
963 | struct rds_sock *rs = rds_sk_to_rs(sk); | |
342dfc30 | 964 | DECLARE_SOCKADDR(struct sockaddr_in *, usin, msg->msg_name); |
5c115590 AG |
965 | __be32 daddr; |
966 | __be16 dport; | |
967 | struct rds_message *rm = NULL; | |
968 | struct rds_connection *conn; | |
969 | int ret = 0; | |
970 | int queued = 0, allocated_mr = 0; | |
971 | int nonblock = msg->msg_flags & MSG_DONTWAIT; | |
1123fd73 | 972 | long timeo = sock_sndtimeo(sk, nonblock); |
780a6d9e | 973 | struct rds_conn_path *cpath; |
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 | ||
8c7188b2 | 998 | lock_sock(sk); |
5c115590 | 999 | if (daddr == 0 || rs->rs_bound_addr == 0) { |
8c7188b2 | 1000 | release_sock(sk); |
5c115590 AG |
1001 | ret = -ENOTCONN; /* XXX not a great errno */ |
1002 | goto out; | |
1003 | } | |
8c7188b2 | 1004 | release_sock(sk); |
5c115590 | 1005 | |
06e8941e MK |
1006 | if (payload_len > rds_sk_sndbuf(rs)) { |
1007 | ret = -EMSGSIZE; | |
1008 | goto out; | |
1009 | } | |
1010 | ||
fc445084 AG |
1011 | /* size of rm including all sgs */ |
1012 | ret = rds_rm_size(msg, payload_len); | |
1013 | if (ret < 0) | |
1014 | goto out; | |
1015 | ||
1016 | rm = rds_message_alloc(ret, GFP_KERNEL); | |
1017 | if (!rm) { | |
1018 | ret = -ENOMEM; | |
5c115590 AG |
1019 | goto out; |
1020 | } | |
1021 | ||
372cd7de AG |
1022 | /* Attach data to the rm */ |
1023 | if (payload_len) { | |
1024 | rm->data.op_sg = rds_message_alloc_sgs(rm, ceil(payload_len, PAGE_SIZE)); | |
d139ff09 AG |
1025 | if (!rm->data.op_sg) { |
1026 | ret = -ENOMEM; | |
1027 | goto out; | |
1028 | } | |
c0371da6 | 1029 | ret = rds_message_copy_from_user(rm, &msg->msg_iter); |
372cd7de AG |
1030 | if (ret) |
1031 | goto out; | |
1032 | } | |
1033 | rm->data.op_active = 1; | |
fc445084 | 1034 | |
5c115590 AG |
1035 | rm->m_daddr = daddr; |
1036 | ||
5c115590 AG |
1037 | /* rds_conn_create has a spinlock that runs with IRQ off. |
1038 | * Caching the conn in the socket helps a lot. */ | |
1039 | if (rs->rs_conn && rs->rs_conn->c_faddr == daddr) | |
1040 | conn = rs->rs_conn; | |
1041 | else { | |
d5a8ac28 SV |
1042 | conn = rds_conn_create_outgoing(sock_net(sock->sk), |
1043 | rs->rs_bound_addr, daddr, | |
5c115590 AG |
1044 | rs->rs_transport, |
1045 | sock->sk->sk_allocation); | |
1046 | if (IS_ERR(conn)) { | |
1047 | ret = PTR_ERR(conn); | |
1048 | goto out; | |
1049 | } | |
1050 | rs->rs_conn = conn; | |
1051 | } | |
1052 | ||
49f69691 AG |
1053 | /* Parse any control messages the user may have included. */ |
1054 | ret = rds_cmsg_send(rs, rm, msg, &allocated_mr); | |
1055 | if (ret) | |
1056 | goto out; | |
1057 | ||
2c3a5f9a | 1058 | if (rm->rdma.op_active && !conn->c_trans->xmit_rdma) { |
cb0a6056 | 1059 | printk_ratelimited(KERN_NOTICE "rdma_op %p conn xmit_rdma %p\n", |
f8b3aaf2 | 1060 | &rm->rdma, conn->c_trans->xmit_rdma); |
15133f6e AG |
1061 | ret = -EOPNOTSUPP; |
1062 | goto out; | |
1063 | } | |
1064 | ||
1065 | if (rm->atomic.op_active && !conn->c_trans->xmit_atomic) { | |
cb0a6056 | 1066 | printk_ratelimited(KERN_NOTICE "atomic_op %p conn xmit_atomic %p\n", |
15133f6e | 1067 | &rm->atomic, conn->c_trans->xmit_atomic); |
5c115590 AG |
1068 | ret = -EOPNOTSUPP; |
1069 | goto out; | |
1070 | } | |
1071 | ||
f3c6808d | 1072 | rds_conn_connect_if_down(conn); |
5c115590 AG |
1073 | |
1074 | ret = rds_cong_wait(conn->c_fcong, dport, nonblock, rs); | |
b98ba52f AG |
1075 | if (ret) { |
1076 | rs->rs_seen_congestion = 1; | |
5c115590 | 1077 | goto out; |
b98ba52f | 1078 | } |
5c115590 | 1079 | |
780a6d9e SV |
1080 | cpath = &conn->c_path[0]; |
1081 | ||
1082 | while (!rds_send_queue_rm(rs, conn, cpath, rm, rs->rs_bound_port, | |
5c115590 AG |
1083 | dport, &queued)) { |
1084 | rds_stats_inc(s_send_queue_full); | |
06e8941e | 1085 | |
5c115590 AG |
1086 | if (nonblock) { |
1087 | ret = -EAGAIN; | |
1088 | goto out; | |
1089 | } | |
1090 | ||
aa395145 | 1091 | timeo = wait_event_interruptible_timeout(*sk_sleep(sk), |
780a6d9e | 1092 | rds_send_queue_rm(rs, conn, cpath, rm, |
5c115590 AG |
1093 | rs->rs_bound_port, |
1094 | dport, | |
1095 | &queued), | |
1096 | timeo); | |
1097 | rdsdebug("sendmsg woke queued %d timeo %ld\n", queued, timeo); | |
1098 | if (timeo > 0 || timeo == MAX_SCHEDULE_TIMEOUT) | |
1099 | continue; | |
1100 | ||
1101 | ret = timeo; | |
1102 | if (ret == 0) | |
1103 | ret = -ETIMEDOUT; | |
1104 | goto out; | |
1105 | } | |
1106 | ||
1107 | /* | |
1108 | * By now we've committed to the send. We reuse rds_send_worker() | |
1109 | * to retry sends in the rds thread if the transport asks us to. | |
1110 | */ | |
1111 | rds_stats_inc(s_send_queued); | |
1112 | ||
db6526dc SS |
1113 | ret = rds_send_xmit(conn); |
1114 | if (ret == -ENOMEM || ret == -EAGAIN) | |
1115 | queue_delayed_work(rds_wq, &conn->c_send_w, 1); | |
5c115590 AG |
1116 | |
1117 | rds_message_put(rm); | |
1118 | return payload_len; | |
1119 | ||
1120 | out: | |
1121 | /* If the user included a RDMA_MAP cmsg, we allocated a MR on the fly. | |
1122 | * If the sendmsg goes through, we keep the MR. If it fails with EAGAIN | |
1123 | * or in any other way, we need to destroy the MR again */ | |
1124 | if (allocated_mr) | |
1125 | rds_rdma_unuse(rs, rds_rdma_cookie_key(rm->m_rdma_cookie), 1); | |
1126 | ||
1127 | if (rm) | |
1128 | rds_message_put(rm); | |
1129 | return ret; | |
1130 | } | |
1131 | ||
1132 | /* | |
1133 | * Reply to a ping packet. | |
1134 | */ | |
1135 | int | |
1136 | rds_send_pong(struct rds_connection *conn, __be16 dport) | |
1137 | { | |
1138 | struct rds_message *rm; | |
1139 | unsigned long flags; | |
1140 | int ret = 0; | |
1141 | ||
1142 | rm = rds_message_alloc(0, GFP_ATOMIC); | |
8690bfa1 | 1143 | if (!rm) { |
5c115590 AG |
1144 | ret = -ENOMEM; |
1145 | goto out; | |
1146 | } | |
1147 | ||
1148 | rm->m_daddr = conn->c_faddr; | |
acfcd4d4 | 1149 | rm->data.op_active = 1; |
5c115590 | 1150 | |
f3c6808d | 1151 | rds_conn_connect_if_down(conn); |
5c115590 AG |
1152 | |
1153 | ret = rds_cong_wait(conn->c_fcong, dport, 1, NULL); | |
1154 | if (ret) | |
1155 | goto out; | |
1156 | ||
1157 | spin_lock_irqsave(&conn->c_lock, flags); | |
1158 | list_add_tail(&rm->m_conn_item, &conn->c_send_queue); | |
1159 | set_bit(RDS_MSG_ON_CONN, &rm->m_flags); | |
1160 | rds_message_addref(rm); | |
1161 | rm->m_inc.i_conn = conn; | |
1162 | ||
1163 | rds_message_populate_header(&rm->m_inc.i_hdr, 0, dport, | |
1164 | conn->c_next_tx_seq); | |
1165 | conn->c_next_tx_seq++; | |
1166 | spin_unlock_irqrestore(&conn->c_lock, flags); | |
1167 | ||
1168 | rds_stats_inc(s_send_queued); | |
1169 | rds_stats_inc(s_send_pong); | |
1170 | ||
7b4b0009 | 1171 | /* schedule the send work on rds_wq */ |
1172 | queue_delayed_work(rds_wq, &conn->c_send_w, 1); | |
acfcd4d4 | 1173 | |
5c115590 AG |
1174 | rds_message_put(rm); |
1175 | return 0; | |
1176 | ||
1177 | out: | |
1178 | if (rm) | |
1179 | rds_message_put(rm); | |
1180 | return ret; | |
1181 | } |