xref: /linux/net/rds/send.c (revision 1422f288)
1 /*
2  * Copyright (c) 2006, 2018 Oracle and/or its affiliates. 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>
34 #include <linux/moduleparam.h>
35 #include <linux/gfp.h>
36 #include <net/sock.h>
37 #include <linux/in.h>
38 #include <linux/list.h>
39 #include <linux/ratelimit.h>
40 #include <linux/export.h>
41 #include <linux/sizes.h>
42 
43 #include "rds.h"
44 
45 /* When transmitting messages in rds_send_xmit, we need to emerge from
46  * time to time and briefly release the CPU. Otherwise the softlock watchdog
47  * will kick our shin.
48  * Also, it seems fairer to not let one busy connection stall all the
49  * others.
50  *
51  * send_batch_count is the number of times we'll loop in send_xmit. Setting
52  * it to 0 will restore the old behavior (where we looped until we had
53  * drained the queue).
54  */
55 static int send_batch_count = SZ_1K;
56 module_param(send_batch_count, int, 0444);
57 MODULE_PARM_DESC(send_batch_count, " batch factor when working the send queue");
58 
59 static void rds_send_remove_from_sock(struct list_head *messages, int status);
60 
61 /*
62  * Reset the send state.  Callers must ensure that this doesn't race with
63  * rds_send_xmit().
64  */
rds_send_path_reset(struct rds_conn_path * cp)65 void rds_send_path_reset(struct rds_conn_path *cp)
66 {
67 	struct rds_message *rm, *tmp;
68 	unsigned long flags;
69 
70 	if (cp->cp_xmit_rm) {
71 		rm = cp->cp_xmit_rm;
72 		cp->cp_xmit_rm = NULL;
73 		/* Tell the user the RDMA op is no longer mapped by the
74 		 * transport. This isn't entirely true (it's flushed out
75 		 * independently) but as the connection is down, there's
76 		 * no ongoing RDMA to/from that memory */
77 		rds_message_unmapped(rm);
78 		rds_message_put(rm);
79 	}
80 
81 	cp->cp_xmit_sg = 0;
82 	cp->cp_xmit_hdr_off = 0;
83 	cp->cp_xmit_data_off = 0;
84 	cp->cp_xmit_atomic_sent = 0;
85 	cp->cp_xmit_rdma_sent = 0;
86 	cp->cp_xmit_data_sent = 0;
87 
88 	cp->cp_conn->c_map_queued = 0;
89 
90 	cp->cp_unacked_packets = rds_sysctl_max_unacked_packets;
91 	cp->cp_unacked_bytes = rds_sysctl_max_unacked_bytes;
92 
93 	/* Mark messages as retransmissions, and move them to the send q */
94 	spin_lock_irqsave(&cp->cp_lock, flags);
95 	list_for_each_entry_safe(rm, tmp, &cp->cp_retrans, m_conn_item) {
96 		set_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags);
97 		set_bit(RDS_MSG_RETRANSMITTED, &rm->m_flags);
98 	}
99 	list_splice_init(&cp->cp_retrans, &cp->cp_send_queue);
100 	spin_unlock_irqrestore(&cp->cp_lock, flags);
101 }
102 EXPORT_SYMBOL_GPL(rds_send_path_reset);
103 
acquire_in_xmit(struct rds_conn_path * cp)104 static int acquire_in_xmit(struct rds_conn_path *cp)
105 {
106 	return test_and_set_bit_lock(RDS_IN_XMIT, &cp->cp_flags) == 0;
107 }
108 
release_in_xmit(struct rds_conn_path * cp)109 static void release_in_xmit(struct rds_conn_path *cp)
110 {
111 	clear_bit_unlock(RDS_IN_XMIT, &cp->cp_flags);
112 	/*
113 	 * We don't use wait_on_bit()/wake_up_bit() because our waking is in a
114 	 * hot path and finding waiters is very rare.  We don't want to walk
115 	 * the system-wide hashed waitqueue buckets in the fast path only to
116 	 * almost never find waiters.
117 	 */
118 	if (waitqueue_active(&cp->cp_waitq))
119 		wake_up_all(&cp->cp_waitq);
120 }
121 
122 /*
123  * We're making the conscious trade-off here to only send one message
124  * down the connection at a time.
125  *   Pro:
126  *      - tx queueing is a simple fifo list
127  *   	- reassembly is optional and easily done by transports per conn
128  *      - no per flow rx lookup at all, straight to the socket
129  *   	- less per-frag memory and wire overhead
130  *   Con:
131  *      - queued acks can be delayed behind large messages
132  *   Depends:
133  *      - small message latency is higher behind queued large messages
134  *      - large message latency isn't starved by intervening small sends
135  */
rds_send_xmit(struct rds_conn_path * cp)136 int rds_send_xmit(struct rds_conn_path *cp)
137 {
138 	struct rds_connection *conn = cp->cp_conn;
139 	struct rds_message *rm;
140 	unsigned long flags;
141 	unsigned int tmp;
142 	struct scatterlist *sg;
143 	int ret = 0;
144 	LIST_HEAD(to_be_dropped);
145 	int batch_count;
146 	unsigned long send_gen = 0;
147 	int same_rm = 0;
148 
149 restart:
150 	batch_count = 0;
151 
152 	/*
153 	 * sendmsg calls here after having queued its message on the send
154 	 * queue.  We only have one task feeding the connection at a time.  If
155 	 * another thread is already feeding the queue then we back off.  This
156 	 * avoids blocking the caller and trading per-connection data between
157 	 * caches per message.
158 	 */
159 	if (!acquire_in_xmit(cp)) {
160 		rds_stats_inc(s_send_lock_contention);
161 		ret = -ENOMEM;
162 		goto out;
163 	}
164 
165 	if (rds_destroy_pending(cp->cp_conn)) {
166 		release_in_xmit(cp);
167 		ret = -ENETUNREACH; /* dont requeue send work */
168 		goto out;
169 	}
170 
171 	/*
172 	 * we record the send generation after doing the xmit acquire.
173 	 * if someone else manages to jump in and do some work, we'll use
174 	 * this to avoid a goto restart farther down.
175 	 *
176 	 * The acquire_in_xmit() check above ensures that only one
177 	 * caller can increment c_send_gen at any time.
178 	 */
179 	send_gen = READ_ONCE(cp->cp_send_gen) + 1;
180 	WRITE_ONCE(cp->cp_send_gen, send_gen);
181 
182 	/*
183 	 * rds_conn_shutdown() sets the conn state and then tests RDS_IN_XMIT,
184 	 * we do the opposite to avoid races.
185 	 */
186 	if (!rds_conn_path_up(cp)) {
187 		release_in_xmit(cp);
188 		ret = 0;
189 		goto out;
190 	}
191 
192 	if (conn->c_trans->xmit_path_prepare)
193 		conn->c_trans->xmit_path_prepare(cp);
194 
195 	/*
196 	 * spin trying to push headers and data down the connection until
197 	 * the connection doesn't make forward progress.
198 	 */
199 	while (1) {
200 
201 		rm = cp->cp_xmit_rm;
202 
203 		if (!rm) {
204 			same_rm = 0;
205 		} else {
206 			same_rm++;
207 			if (same_rm >= 4096) {
208 				rds_stats_inc(s_send_stuck_rm);
209 				ret = -EAGAIN;
210 				break;
211 			}
212 		}
213 
214 		/*
215 		 * If between sending messages, we can send a pending congestion
216 		 * map update.
217 		 */
218 		if (!rm && test_and_clear_bit(0, &conn->c_map_queued)) {
219 			rm = rds_cong_update_alloc(conn);
220 			if (IS_ERR(rm)) {
221 				ret = PTR_ERR(rm);
222 				break;
223 			}
224 			rm->data.op_active = 1;
225 			rm->m_inc.i_conn_path = cp;
226 			rm->m_inc.i_conn = cp->cp_conn;
227 
228 			cp->cp_xmit_rm = rm;
229 		}
230 
231 		/*
232 		 * If not already working on one, grab the next message.
233 		 *
234 		 * cp_xmit_rm holds a ref while we're sending this message down
235 		 * the connction.  We can use this ref while holding the
236 		 * send_sem.. rds_send_reset() is serialized with it.
237 		 */
238 		if (!rm) {
239 			unsigned int len;
240 
241 			batch_count++;
242 
243 			/* we want to process as big a batch as we can, but
244 			 * we also want to avoid softlockups.  If we've been
245 			 * through a lot of messages, lets back off and see
246 			 * if anyone else jumps in
247 			 */
248 			if (batch_count >= send_batch_count)
249 				goto over_batch;
250 
251 			spin_lock_irqsave(&cp->cp_lock, flags);
252 
253 			if (!list_empty(&cp->cp_send_queue)) {
254 				rm = list_entry(cp->cp_send_queue.next,
255 						struct rds_message,
256 						m_conn_item);
257 				rds_message_addref(rm);
258 
259 				/*
260 				 * Move the message from the send queue to the retransmit
261 				 * list right away.
262 				 */
263 				list_move_tail(&rm->m_conn_item,
264 					       &cp->cp_retrans);
265 			}
266 
267 			spin_unlock_irqrestore(&cp->cp_lock, flags);
268 
269 			if (!rm)
270 				break;
271 
272 			/* Unfortunately, the way Infiniband deals with
273 			 * RDMA to a bad MR key is by moving the entire
274 			 * queue pair to error state. We could possibly
275 			 * recover from that, but right now we drop the
276 			 * connection.
277 			 * Therefore, we never retransmit messages with RDMA ops.
278 			 */
279 			if (test_bit(RDS_MSG_FLUSH, &rm->m_flags) ||
280 			    (rm->rdma.op_active &&
281 			    test_bit(RDS_MSG_RETRANSMITTED, &rm->m_flags))) {
282 				spin_lock_irqsave(&cp->cp_lock, flags);
283 				if (test_and_clear_bit(RDS_MSG_ON_CONN, &rm->m_flags))
284 					list_move(&rm->m_conn_item, &to_be_dropped);
285 				spin_unlock_irqrestore(&cp->cp_lock, flags);
286 				continue;
287 			}
288 
289 			/* Require an ACK every once in a while */
290 			len = ntohl(rm->m_inc.i_hdr.h_len);
291 			if (cp->cp_unacked_packets == 0 ||
292 			    cp->cp_unacked_bytes < len) {
293 				set_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags);
294 
295 				cp->cp_unacked_packets =
296 					rds_sysctl_max_unacked_packets;
297 				cp->cp_unacked_bytes =
298 					rds_sysctl_max_unacked_bytes;
299 				rds_stats_inc(s_send_ack_required);
300 			} else {
301 				cp->cp_unacked_bytes -= len;
302 				cp->cp_unacked_packets--;
303 			}
304 
305 			cp->cp_xmit_rm = rm;
306 		}
307 
308 		/* The transport either sends the whole rdma or none of it */
309 		if (rm->rdma.op_active && !cp->cp_xmit_rdma_sent) {
310 			rm->m_final_op = &rm->rdma;
311 			/* The transport owns the mapped memory for now.
312 			 * You can't unmap it while it's on the send queue
313 			 */
314 			set_bit(RDS_MSG_MAPPED, &rm->m_flags);
315 			ret = conn->c_trans->xmit_rdma(conn, &rm->rdma);
316 			if (ret) {
317 				clear_bit(RDS_MSG_MAPPED, &rm->m_flags);
318 				wake_up_interruptible(&rm->m_flush_wait);
319 				break;
320 			}
321 			cp->cp_xmit_rdma_sent = 1;
322 
323 		}
324 
325 		if (rm->atomic.op_active && !cp->cp_xmit_atomic_sent) {
326 			rm->m_final_op = &rm->atomic;
327 			/* The transport owns the mapped memory for now.
328 			 * You can't unmap it while it's on the send queue
329 			 */
330 			set_bit(RDS_MSG_MAPPED, &rm->m_flags);
331 			ret = conn->c_trans->xmit_atomic(conn, &rm->atomic);
332 			if (ret) {
333 				clear_bit(RDS_MSG_MAPPED, &rm->m_flags);
334 				wake_up_interruptible(&rm->m_flush_wait);
335 				break;
336 			}
337 			cp->cp_xmit_atomic_sent = 1;
338 
339 		}
340 
341 		/*
342 		 * A number of cases require an RDS header to be sent
343 		 * even if there is no data.
344 		 * We permit 0-byte sends; rds-ping depends on this.
345 		 * However, if there are exclusively attached silent ops,
346 		 * we skip the hdr/data send, to enable silent operation.
347 		 */
348 		if (rm->data.op_nents == 0) {
349 			int ops_present;
350 			int all_ops_are_silent = 1;
351 
352 			ops_present = (rm->atomic.op_active || rm->rdma.op_active);
353 			if (rm->atomic.op_active && !rm->atomic.op_silent)
354 				all_ops_are_silent = 0;
355 			if (rm->rdma.op_active && !rm->rdma.op_silent)
356 				all_ops_are_silent = 0;
357 
358 			if (ops_present && all_ops_are_silent
359 			    && !rm->m_rdma_cookie)
360 				rm->data.op_active = 0;
361 		}
362 
363 		if (rm->data.op_active && !cp->cp_xmit_data_sent) {
364 			rm->m_final_op = &rm->data;
365 
366 			ret = conn->c_trans->xmit(conn, rm,
367 						  cp->cp_xmit_hdr_off,
368 						  cp->cp_xmit_sg,
369 						  cp->cp_xmit_data_off);
370 			if (ret <= 0)
371 				break;
372 
373 			if (cp->cp_xmit_hdr_off < sizeof(struct rds_header)) {
374 				tmp = min_t(int, ret,
375 					    sizeof(struct rds_header) -
376 					    cp->cp_xmit_hdr_off);
377 				cp->cp_xmit_hdr_off += tmp;
378 				ret -= tmp;
379 			}
380 
381 			sg = &rm->data.op_sg[cp->cp_xmit_sg];
382 			while (ret) {
383 				tmp = min_t(int, ret, sg->length -
384 						      cp->cp_xmit_data_off);
385 				cp->cp_xmit_data_off += tmp;
386 				ret -= tmp;
387 				if (cp->cp_xmit_data_off == sg->length) {
388 					cp->cp_xmit_data_off = 0;
389 					sg++;
390 					cp->cp_xmit_sg++;
391 					BUG_ON(ret != 0 && cp->cp_xmit_sg ==
392 					       rm->data.op_nents);
393 				}
394 			}
395 
396 			if (cp->cp_xmit_hdr_off == sizeof(struct rds_header) &&
397 			    (cp->cp_xmit_sg == rm->data.op_nents))
398 				cp->cp_xmit_data_sent = 1;
399 		}
400 
401 		/*
402 		 * A rm will only take multiple times through this loop
403 		 * if there is a data op. Thus, if the data is sent (or there was
404 		 * none), then we're done with the rm.
405 		 */
406 		if (!rm->data.op_active || cp->cp_xmit_data_sent) {
407 			cp->cp_xmit_rm = NULL;
408 			cp->cp_xmit_sg = 0;
409 			cp->cp_xmit_hdr_off = 0;
410 			cp->cp_xmit_data_off = 0;
411 			cp->cp_xmit_rdma_sent = 0;
412 			cp->cp_xmit_atomic_sent = 0;
413 			cp->cp_xmit_data_sent = 0;
414 
415 			rds_message_put(rm);
416 		}
417 	}
418 
419 over_batch:
420 	if (conn->c_trans->xmit_path_complete)
421 		conn->c_trans->xmit_path_complete(cp);
422 	release_in_xmit(cp);
423 
424 	/* Nuke any messages we decided not to retransmit. */
425 	if (!list_empty(&to_be_dropped)) {
426 		/* irqs on here, so we can put(), unlike above */
427 		list_for_each_entry(rm, &to_be_dropped, m_conn_item)
428 			rds_message_put(rm);
429 		rds_send_remove_from_sock(&to_be_dropped, RDS_RDMA_DROPPED);
430 	}
431 
432 	/*
433 	 * Other senders can queue a message after we last test the send queue
434 	 * but before we clear RDS_IN_XMIT.  In that case they'd back off and
435 	 * not try and send their newly queued message.  We need to check the
436 	 * send queue after having cleared RDS_IN_XMIT so that their message
437 	 * doesn't get stuck on the send queue.
438 	 *
439 	 * If the transport cannot continue (i.e ret != 0), then it must
440 	 * call us when more room is available, such as from the tx
441 	 * completion handler.
442 	 *
443 	 * We have an extra generation check here so that if someone manages
444 	 * to jump in after our release_in_xmit, we'll see that they have done
445 	 * some work and we will skip our goto
446 	 */
447 	if (ret == 0) {
448 		bool raced;
449 
450 		smp_mb();
451 		raced = send_gen != READ_ONCE(cp->cp_send_gen);
452 
453 		if ((test_bit(0, &conn->c_map_queued) ||
454 		    !list_empty(&cp->cp_send_queue)) && !raced) {
455 			if (batch_count < send_batch_count)
456 				goto restart;
457 			rcu_read_lock();
458 			if (rds_destroy_pending(cp->cp_conn))
459 				ret = -ENETUNREACH;
460 			else
461 				queue_delayed_work(rds_wq, &cp->cp_send_w, 1);
462 			rcu_read_unlock();
463 		} else if (raced) {
464 			rds_stats_inc(s_send_lock_queue_raced);
465 		}
466 	}
467 out:
468 	return ret;
469 }
470 EXPORT_SYMBOL_GPL(rds_send_xmit);
471 
rds_send_sndbuf_remove(struct rds_sock * rs,struct rds_message * rm)472 static void rds_send_sndbuf_remove(struct rds_sock *rs, struct rds_message *rm)
473 {
474 	u32 len = be32_to_cpu(rm->m_inc.i_hdr.h_len);
475 
476 	assert_spin_locked(&rs->rs_lock);
477 
478 	BUG_ON(rs->rs_snd_bytes < len);
479 	rs->rs_snd_bytes -= len;
480 
481 	if (rs->rs_snd_bytes == 0)
482 		rds_stats_inc(s_send_queue_empty);
483 }
484 
rds_send_is_acked(struct rds_message * rm,u64 ack,is_acked_func is_acked)485 static inline int rds_send_is_acked(struct rds_message *rm, u64 ack,
486 				    is_acked_func is_acked)
487 {
488 	if (is_acked)
489 		return is_acked(rm, ack);
490 	return be64_to_cpu(rm->m_inc.i_hdr.h_sequence) <= ack;
491 }
492 
493 /*
494  * This is pretty similar to what happens below in the ACK
495  * handling code - except that we call here as soon as we get
496  * the IB send completion on the RDMA op and the accompanying
497  * message.
498  */
rds_rdma_send_complete(struct rds_message * rm,int status)499 void rds_rdma_send_complete(struct rds_message *rm, int status)
500 {
501 	struct rds_sock *rs = NULL;
502 	struct rm_rdma_op *ro;
503 	struct rds_notifier *notifier;
504 	unsigned long flags;
505 
506 	spin_lock_irqsave(&rm->m_rs_lock, flags);
507 
508 	ro = &rm->rdma;
509 	if (test_bit(RDS_MSG_ON_SOCK, &rm->m_flags) &&
510 	    ro->op_active && ro->op_notify && ro->op_notifier) {
511 		notifier = ro->op_notifier;
512 		rs = rm->m_rs;
513 		sock_hold(rds_rs_to_sk(rs));
514 
515 		notifier->n_status = status;
516 		spin_lock(&rs->rs_lock);
517 		list_add_tail(&notifier->n_list, &rs->rs_notify_queue);
518 		spin_unlock(&rs->rs_lock);
519 
520 		ro->op_notifier = NULL;
521 	}
522 
523 	spin_unlock_irqrestore(&rm->m_rs_lock, flags);
524 
525 	if (rs) {
526 		rds_wake_sk_sleep(rs);
527 		sock_put(rds_rs_to_sk(rs));
528 	}
529 }
530 EXPORT_SYMBOL_GPL(rds_rdma_send_complete);
531 
532 /*
533  * Just like above, except looks at atomic op
534  */
rds_atomic_send_complete(struct rds_message * rm,int status)535 void rds_atomic_send_complete(struct rds_message *rm, int status)
536 {
537 	struct rds_sock *rs = NULL;
538 	struct rm_atomic_op *ao;
539 	struct rds_notifier *notifier;
540 	unsigned long flags;
541 
542 	spin_lock_irqsave(&rm->m_rs_lock, flags);
543 
544 	ao = &rm->atomic;
545 	if (test_bit(RDS_MSG_ON_SOCK, &rm->m_flags)
546 	    && ao->op_active && ao->op_notify && ao->op_notifier) {
547 		notifier = ao->op_notifier;
548 		rs = rm->m_rs;
549 		sock_hold(rds_rs_to_sk(rs));
550 
551 		notifier->n_status = status;
552 		spin_lock(&rs->rs_lock);
553 		list_add_tail(&notifier->n_list, &rs->rs_notify_queue);
554 		spin_unlock(&rs->rs_lock);
555 
556 		ao->op_notifier = NULL;
557 	}
558 
559 	spin_unlock_irqrestore(&rm->m_rs_lock, flags);
560 
561 	if (rs) {
562 		rds_wake_sk_sleep(rs);
563 		sock_put(rds_rs_to_sk(rs));
564 	}
565 }
566 EXPORT_SYMBOL_GPL(rds_atomic_send_complete);
567 
568 /*
569  * This is the same as rds_rdma_send_complete except we
570  * don't do any locking - we have all the ingredients (message,
571  * socket, socket lock) and can just move the notifier.
572  */
573 static inline void
__rds_send_complete(struct rds_sock * rs,struct rds_message * rm,int status)574 __rds_send_complete(struct rds_sock *rs, struct rds_message *rm, int status)
575 {
576 	struct rm_rdma_op *ro;
577 	struct rm_atomic_op *ao;
578 
579 	ro = &rm->rdma;
580 	if (ro->op_active && ro->op_notify && ro->op_notifier) {
581 		ro->op_notifier->n_status = status;
582 		list_add_tail(&ro->op_notifier->n_list, &rs->rs_notify_queue);
583 		ro->op_notifier = NULL;
584 	}
585 
586 	ao = &rm->atomic;
587 	if (ao->op_active && ao->op_notify && ao->op_notifier) {
588 		ao->op_notifier->n_status = status;
589 		list_add_tail(&ao->op_notifier->n_list, &rs->rs_notify_queue);
590 		ao->op_notifier = NULL;
591 	}
592 
593 	/* No need to wake the app - caller does this */
594 }
595 
596 /*
597  * This removes messages from the socket's list if they're on it.  The list
598  * argument must be private to the caller, we must be able to modify it
599  * without locks.  The messages must have a reference held for their
600  * position on the list.  This function will drop that reference after
601  * removing the messages from the 'messages' list regardless of if it found
602  * the messages on the socket list or not.
603  */
rds_send_remove_from_sock(struct list_head * messages,int status)604 static void rds_send_remove_from_sock(struct list_head *messages, int status)
605 {
606 	unsigned long flags;
607 	struct rds_sock *rs = NULL;
608 	struct rds_message *rm;
609 
610 	while (!list_empty(messages)) {
611 		int was_on_sock = 0;
612 
613 		rm = list_entry(messages->next, struct rds_message,
614 				m_conn_item);
615 		list_del_init(&rm->m_conn_item);
616 
617 		/*
618 		 * If we see this flag cleared then we're *sure* that someone
619 		 * else beat us to removing it from the sock.  If we race
620 		 * with their flag update we'll get the lock and then really
621 		 * see that the flag has been cleared.
622 		 *
623 		 * The message spinlock makes sure nobody clears rm->m_rs
624 		 * while we're messing with it. It does not prevent the
625 		 * message from being removed from the socket, though.
626 		 */
627 		spin_lock_irqsave(&rm->m_rs_lock, flags);
628 		if (!test_bit(RDS_MSG_ON_SOCK, &rm->m_flags))
629 			goto unlock_and_drop;
630 
631 		if (rs != rm->m_rs) {
632 			if (rs) {
633 				rds_wake_sk_sleep(rs);
634 				sock_put(rds_rs_to_sk(rs));
635 			}
636 			rs = rm->m_rs;
637 			if (rs)
638 				sock_hold(rds_rs_to_sk(rs));
639 		}
640 		if (!rs)
641 			goto unlock_and_drop;
642 		spin_lock(&rs->rs_lock);
643 
644 		if (test_and_clear_bit(RDS_MSG_ON_SOCK, &rm->m_flags)) {
645 			struct rm_rdma_op *ro = &rm->rdma;
646 			struct rds_notifier *notifier;
647 
648 			list_del_init(&rm->m_sock_item);
649 			rds_send_sndbuf_remove(rs, rm);
650 
651 			if (ro->op_active && ro->op_notifier &&
652 			       (ro->op_notify || (ro->op_recverr && status))) {
653 				notifier = ro->op_notifier;
654 				list_add_tail(&notifier->n_list,
655 						&rs->rs_notify_queue);
656 				if (!notifier->n_status)
657 					notifier->n_status = status;
658 				rm->rdma.op_notifier = NULL;
659 			}
660 			was_on_sock = 1;
661 		}
662 		spin_unlock(&rs->rs_lock);
663 
664 unlock_and_drop:
665 		spin_unlock_irqrestore(&rm->m_rs_lock, flags);
666 		rds_message_put(rm);
667 		if (was_on_sock)
668 			rds_message_put(rm);
669 	}
670 
671 	if (rs) {
672 		rds_wake_sk_sleep(rs);
673 		sock_put(rds_rs_to_sk(rs));
674 	}
675 }
676 
677 /*
678  * Transports call here when they've determined that the receiver queued
679  * messages up to, and including, the given sequence number.  Messages are
680  * moved to the retrans queue when rds_send_xmit picks them off the send
681  * queue. This means that in the TCP case, the message may not have been
682  * assigned the m_ack_seq yet - but that's fine as long as tcp_is_acked
683  * checks the RDS_MSG_HAS_ACK_SEQ bit.
684  */
rds_send_path_drop_acked(struct rds_conn_path * cp,u64 ack,is_acked_func is_acked)685 void rds_send_path_drop_acked(struct rds_conn_path *cp, u64 ack,
686 			      is_acked_func is_acked)
687 {
688 	struct rds_message *rm, *tmp;
689 	unsigned long flags;
690 	LIST_HEAD(list);
691 
692 	spin_lock_irqsave(&cp->cp_lock, flags);
693 
694 	list_for_each_entry_safe(rm, tmp, &cp->cp_retrans, m_conn_item) {
695 		if (!rds_send_is_acked(rm, ack, is_acked))
696 			break;
697 
698 		list_move(&rm->m_conn_item, &list);
699 		clear_bit(RDS_MSG_ON_CONN, &rm->m_flags);
700 	}
701 
702 	/* order flag updates with spin locks */
703 	if (!list_empty(&list))
704 		smp_mb__after_atomic();
705 
706 	spin_unlock_irqrestore(&cp->cp_lock, flags);
707 
708 	/* now remove the messages from the sock list as needed */
709 	rds_send_remove_from_sock(&list, RDS_RDMA_SUCCESS);
710 }
711 EXPORT_SYMBOL_GPL(rds_send_path_drop_acked);
712 
rds_send_drop_acked(struct rds_connection * conn,u64 ack,is_acked_func is_acked)713 void rds_send_drop_acked(struct rds_connection *conn, u64 ack,
714 			 is_acked_func is_acked)
715 {
716 	WARN_ON(conn->c_trans->t_mp_capable);
717 	rds_send_path_drop_acked(&conn->c_path[0], ack, is_acked);
718 }
719 EXPORT_SYMBOL_GPL(rds_send_drop_acked);
720 
rds_send_drop_to(struct rds_sock * rs,struct sockaddr_in6 * dest)721 void rds_send_drop_to(struct rds_sock *rs, struct sockaddr_in6 *dest)
722 {
723 	struct rds_message *rm, *tmp;
724 	struct rds_connection *conn;
725 	struct rds_conn_path *cp;
726 	unsigned long flags;
727 	LIST_HEAD(list);
728 
729 	/* get all the messages we're dropping under the rs lock */
730 	spin_lock_irqsave(&rs->rs_lock, flags);
731 
732 	list_for_each_entry_safe(rm, tmp, &rs->rs_send_queue, m_sock_item) {
733 		if (dest &&
734 		    (!ipv6_addr_equal(&dest->sin6_addr, &rm->m_daddr) ||
735 		     dest->sin6_port != rm->m_inc.i_hdr.h_dport))
736 			continue;
737 
738 		list_move(&rm->m_sock_item, &list);
739 		rds_send_sndbuf_remove(rs, rm);
740 		clear_bit(RDS_MSG_ON_SOCK, &rm->m_flags);
741 	}
742 
743 	/* order flag updates with the rs lock */
744 	smp_mb__after_atomic();
745 
746 	spin_unlock_irqrestore(&rs->rs_lock, flags);
747 
748 	if (list_empty(&list))
749 		return;
750 
751 	/* Remove the messages from the conn */
752 	list_for_each_entry(rm, &list, m_sock_item) {
753 
754 		conn = rm->m_inc.i_conn;
755 		if (conn->c_trans->t_mp_capable)
756 			cp = rm->m_inc.i_conn_path;
757 		else
758 			cp = &conn->c_path[0];
759 
760 		spin_lock_irqsave(&cp->cp_lock, flags);
761 		/*
762 		 * Maybe someone else beat us to removing rm from the conn.
763 		 * If we race with their flag update we'll get the lock and
764 		 * then really see that the flag has been cleared.
765 		 */
766 		if (!test_and_clear_bit(RDS_MSG_ON_CONN, &rm->m_flags)) {
767 			spin_unlock_irqrestore(&cp->cp_lock, flags);
768 			continue;
769 		}
770 		list_del_init(&rm->m_conn_item);
771 		spin_unlock_irqrestore(&cp->cp_lock, flags);
772 
773 		/*
774 		 * Couldn't grab m_rs_lock in top loop (lock ordering),
775 		 * but we can now.
776 		 */
777 		spin_lock_irqsave(&rm->m_rs_lock, flags);
778 
779 		spin_lock(&rs->rs_lock);
780 		__rds_send_complete(rs, rm, RDS_RDMA_CANCELED);
781 		spin_unlock(&rs->rs_lock);
782 
783 		spin_unlock_irqrestore(&rm->m_rs_lock, flags);
784 
785 		rds_message_put(rm);
786 	}
787 
788 	rds_wake_sk_sleep(rs);
789 
790 	while (!list_empty(&list)) {
791 		rm = list_entry(list.next, struct rds_message, m_sock_item);
792 		list_del_init(&rm->m_sock_item);
793 		rds_message_wait(rm);
794 
795 		/* just in case the code above skipped this message
796 		 * because RDS_MSG_ON_CONN wasn't set, run it again here
797 		 * taking m_rs_lock is the only thing that keeps us
798 		 * from racing with ack processing.
799 		 */
800 		spin_lock_irqsave(&rm->m_rs_lock, flags);
801 
802 		spin_lock(&rs->rs_lock);
803 		__rds_send_complete(rs, rm, RDS_RDMA_CANCELED);
804 		spin_unlock(&rs->rs_lock);
805 
806 		spin_unlock_irqrestore(&rm->m_rs_lock, flags);
807 
808 		rds_message_put(rm);
809 	}
810 }
811 
812 /*
813  * we only want this to fire once so we use the callers 'queued'.  It's
814  * possible that another thread can race with us and remove the
815  * message from the flow with RDS_CANCEL_SENT_TO.
816  */
rds_send_queue_rm(struct rds_sock * rs,struct rds_connection * conn,struct rds_conn_path * cp,struct rds_message * rm,__be16 sport,__be16 dport,int * queued)817 static int rds_send_queue_rm(struct rds_sock *rs, struct rds_connection *conn,
818 			     struct rds_conn_path *cp,
819 			     struct rds_message *rm, __be16 sport,
820 			     __be16 dport, int *queued)
821 {
822 	unsigned long flags;
823 	u32 len;
824 
825 	if (*queued)
826 		goto out;
827 
828 	len = be32_to_cpu(rm->m_inc.i_hdr.h_len);
829 
830 	/* this is the only place which holds both the socket's rs_lock
831 	 * and the connection's c_lock */
832 	spin_lock_irqsave(&rs->rs_lock, flags);
833 
834 	/*
835 	 * If there is a little space in sndbuf, we don't queue anything,
836 	 * and userspace gets -EAGAIN. But poll() indicates there's send
837 	 * room. This can lead to bad behavior (spinning) if snd_bytes isn't
838 	 * freed up by incoming acks. So we check the *old* value of
839 	 * rs_snd_bytes here to allow the last msg to exceed the buffer,
840 	 * and poll() now knows no more data can be sent.
841 	 */
842 	if (rs->rs_snd_bytes < rds_sk_sndbuf(rs)) {
843 		rs->rs_snd_bytes += len;
844 
845 		/* let recv side know we are close to send space exhaustion.
846 		 * This is probably not the optimal way to do it, as this
847 		 * means we set the flag on *all* messages as soon as our
848 		 * throughput hits a certain threshold.
849 		 */
850 		if (rs->rs_snd_bytes >= rds_sk_sndbuf(rs) / 2)
851 			set_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags);
852 
853 		list_add_tail(&rm->m_sock_item, &rs->rs_send_queue);
854 		set_bit(RDS_MSG_ON_SOCK, &rm->m_flags);
855 		rds_message_addref(rm);
856 		sock_hold(rds_rs_to_sk(rs));
857 		rm->m_rs = rs;
858 
859 		/* The code ordering is a little weird, but we're
860 		   trying to minimize the time we hold c_lock */
861 		rds_message_populate_header(&rm->m_inc.i_hdr, sport, dport, 0);
862 		rm->m_inc.i_conn = conn;
863 		rm->m_inc.i_conn_path = cp;
864 		rds_message_addref(rm);
865 
866 		spin_lock(&cp->cp_lock);
867 		rm->m_inc.i_hdr.h_sequence = cpu_to_be64(cp->cp_next_tx_seq++);
868 		list_add_tail(&rm->m_conn_item, &cp->cp_send_queue);
869 		set_bit(RDS_MSG_ON_CONN, &rm->m_flags);
870 		spin_unlock(&cp->cp_lock);
871 
872 		rdsdebug("queued msg %p len %d, rs %p bytes %d seq %llu\n",
873 			 rm, len, rs, rs->rs_snd_bytes,
874 			 (unsigned long long)be64_to_cpu(rm->m_inc.i_hdr.h_sequence));
875 
876 		*queued = 1;
877 	}
878 
879 	spin_unlock_irqrestore(&rs->rs_lock, flags);
880 out:
881 	return *queued;
882 }
883 
884 /*
885  * rds_message is getting to be quite complicated, and we'd like to allocate
886  * it all in one go. This figures out how big it needs to be up front.
887  */
rds_rm_size(struct msghdr * msg,int num_sgs,struct rds_iov_vector_arr * vct)888 static int rds_rm_size(struct msghdr *msg, int num_sgs,
889 		       struct rds_iov_vector_arr *vct)
890 {
891 	struct cmsghdr *cmsg;
892 	int size = 0;
893 	int cmsg_groups = 0;
894 	int retval;
895 	bool zcopy_cookie = false;
896 	struct rds_iov_vector *iov, *tmp_iov;
897 
898 	if (num_sgs < 0)
899 		return -EINVAL;
900 
901 	for_each_cmsghdr(cmsg, msg) {
902 		if (!CMSG_OK(msg, cmsg))
903 			return -EINVAL;
904 
905 		if (cmsg->cmsg_level != SOL_RDS)
906 			continue;
907 
908 		switch (cmsg->cmsg_type) {
909 		case RDS_CMSG_RDMA_ARGS:
910 			if (vct->indx >= vct->len) {
911 				vct->len += vct->incr;
912 				tmp_iov =
913 					krealloc(vct->vec,
914 						 vct->len *
915 						 sizeof(struct rds_iov_vector),
916 						 GFP_KERNEL);
917 				if (!tmp_iov) {
918 					vct->len -= vct->incr;
919 					return -ENOMEM;
920 				}
921 				vct->vec = tmp_iov;
922 			}
923 			iov = &vct->vec[vct->indx];
924 			memset(iov, 0, sizeof(struct rds_iov_vector));
925 			vct->indx++;
926 			cmsg_groups |= 1;
927 			retval = rds_rdma_extra_size(CMSG_DATA(cmsg), iov);
928 			if (retval < 0)
929 				return retval;
930 			size += retval;
931 
932 			break;
933 
934 		case RDS_CMSG_ZCOPY_COOKIE:
935 			zcopy_cookie = true;
936 			fallthrough;
937 
938 		case RDS_CMSG_RDMA_DEST:
939 		case RDS_CMSG_RDMA_MAP:
940 			cmsg_groups |= 2;
941 			/* these are valid but do no add any size */
942 			break;
943 
944 		case RDS_CMSG_ATOMIC_CSWP:
945 		case RDS_CMSG_ATOMIC_FADD:
946 		case RDS_CMSG_MASKED_ATOMIC_CSWP:
947 		case RDS_CMSG_MASKED_ATOMIC_FADD:
948 			cmsg_groups |= 1;
949 			size += sizeof(struct scatterlist);
950 			break;
951 
952 		default:
953 			return -EINVAL;
954 		}
955 
956 	}
957 
958 	if ((msg->msg_flags & MSG_ZEROCOPY) && !zcopy_cookie)
959 		return -EINVAL;
960 
961 	size += num_sgs * sizeof(struct scatterlist);
962 
963 	/* Ensure (DEST, MAP) are never used with (ARGS, ATOMIC) */
964 	if (cmsg_groups == 3)
965 		return -EINVAL;
966 
967 	return size;
968 }
969 
rds_cmsg_zcopy(struct rds_sock * rs,struct rds_message * rm,struct cmsghdr * cmsg)970 static int rds_cmsg_zcopy(struct rds_sock *rs, struct rds_message *rm,
971 			  struct cmsghdr *cmsg)
972 {
973 	u32 *cookie;
974 
975 	if (cmsg->cmsg_len < CMSG_LEN(sizeof(*cookie)) ||
976 	    !rm->data.op_mmp_znotifier)
977 		return -EINVAL;
978 	cookie = CMSG_DATA(cmsg);
979 	rm->data.op_mmp_znotifier->z_cookie = *cookie;
980 	return 0;
981 }
982 
rds_cmsg_send(struct rds_sock * rs,struct rds_message * rm,struct msghdr * msg,int * allocated_mr,struct rds_iov_vector_arr * vct)983 static int rds_cmsg_send(struct rds_sock *rs, struct rds_message *rm,
984 			 struct msghdr *msg, int *allocated_mr,
985 			 struct rds_iov_vector_arr *vct)
986 {
987 	struct cmsghdr *cmsg;
988 	int ret = 0, ind = 0;
989 
990 	for_each_cmsghdr(cmsg, msg) {
991 		if (!CMSG_OK(msg, cmsg))
992 			return -EINVAL;
993 
994 		if (cmsg->cmsg_level != SOL_RDS)
995 			continue;
996 
997 		/* As a side effect, RDMA_DEST and RDMA_MAP will set
998 		 * rm->rdma.m_rdma_cookie and rm->rdma.m_rdma_mr.
999 		 */
1000 		switch (cmsg->cmsg_type) {
1001 		case RDS_CMSG_RDMA_ARGS:
1002 			if (ind >= vct->indx)
1003 				return -ENOMEM;
1004 			ret = rds_cmsg_rdma_args(rs, rm, cmsg, &vct->vec[ind]);
1005 			ind++;
1006 			break;
1007 
1008 		case RDS_CMSG_RDMA_DEST:
1009 			ret = rds_cmsg_rdma_dest(rs, rm, cmsg);
1010 			break;
1011 
1012 		case RDS_CMSG_RDMA_MAP:
1013 			ret = rds_cmsg_rdma_map(rs, rm, cmsg);
1014 			if (!ret)
1015 				*allocated_mr = 1;
1016 			else if (ret == -ENODEV)
1017 				/* Accommodate the get_mr() case which can fail
1018 				 * if connection isn't established yet.
1019 				 */
1020 				ret = -EAGAIN;
1021 			break;
1022 		case RDS_CMSG_ATOMIC_CSWP:
1023 		case RDS_CMSG_ATOMIC_FADD:
1024 		case RDS_CMSG_MASKED_ATOMIC_CSWP:
1025 		case RDS_CMSG_MASKED_ATOMIC_FADD:
1026 			ret = rds_cmsg_atomic(rs, rm, cmsg);
1027 			break;
1028 
1029 		case RDS_CMSG_ZCOPY_COOKIE:
1030 			ret = rds_cmsg_zcopy(rs, rm, cmsg);
1031 			break;
1032 
1033 		default:
1034 			return -EINVAL;
1035 		}
1036 
1037 		if (ret)
1038 			break;
1039 	}
1040 
1041 	return ret;
1042 }
1043 
rds_send_mprds_hash(struct rds_sock * rs,struct rds_connection * conn,int nonblock)1044 static int rds_send_mprds_hash(struct rds_sock *rs,
1045 			       struct rds_connection *conn, int nonblock)
1046 {
1047 	int hash;
1048 
1049 	if (conn->c_npaths == 0)
1050 		hash = RDS_MPATH_HASH(rs, RDS_MPATH_WORKERS);
1051 	else
1052 		hash = RDS_MPATH_HASH(rs, conn->c_npaths);
1053 	if (conn->c_npaths == 0 && hash != 0) {
1054 		rds_send_ping(conn, 0);
1055 
1056 		/* The underlying connection is not up yet.  Need to wait
1057 		 * until it is up to be sure that the non-zero c_path can be
1058 		 * used.  But if we are interrupted, we have to use the zero
1059 		 * c_path in case the connection ends up being non-MP capable.
1060 		 */
1061 		if (conn->c_npaths == 0) {
1062 			/* Cannot wait for the connection be made, so just use
1063 			 * the base c_path.
1064 			 */
1065 			if (nonblock)
1066 				return 0;
1067 			if (wait_event_interruptible(conn->c_hs_waitq,
1068 						     conn->c_npaths != 0))
1069 				hash = 0;
1070 		}
1071 		if (conn->c_npaths == 1)
1072 			hash = 0;
1073 	}
1074 	return hash;
1075 }
1076 
rds_rdma_bytes(struct msghdr * msg,size_t * rdma_bytes)1077 static int rds_rdma_bytes(struct msghdr *msg, size_t *rdma_bytes)
1078 {
1079 	struct rds_rdma_args *args;
1080 	struct cmsghdr *cmsg;
1081 
1082 	for_each_cmsghdr(cmsg, msg) {
1083 		if (!CMSG_OK(msg, cmsg))
1084 			return -EINVAL;
1085 
1086 		if (cmsg->cmsg_level != SOL_RDS)
1087 			continue;
1088 
1089 		if (cmsg->cmsg_type == RDS_CMSG_RDMA_ARGS) {
1090 			if (cmsg->cmsg_len <
1091 			    CMSG_LEN(sizeof(struct rds_rdma_args)))
1092 				return -EINVAL;
1093 			args = CMSG_DATA(cmsg);
1094 			*rdma_bytes += args->remote_vec.bytes;
1095 		}
1096 	}
1097 	return 0;
1098 }
1099 
rds_sendmsg(struct socket * sock,struct msghdr * msg,size_t payload_len)1100 int rds_sendmsg(struct socket *sock, struct msghdr *msg, size_t payload_len)
1101 {
1102 	struct sock *sk = sock->sk;
1103 	struct rds_sock *rs = rds_sk_to_rs(sk);
1104 	DECLARE_SOCKADDR(struct sockaddr_in6 *, sin6, msg->msg_name);
1105 	DECLARE_SOCKADDR(struct sockaddr_in *, usin, msg->msg_name);
1106 	__be16 dport;
1107 	struct rds_message *rm = NULL;
1108 	struct rds_connection *conn;
1109 	int ret = 0;
1110 	int queued = 0, allocated_mr = 0;
1111 	int nonblock = msg->msg_flags & MSG_DONTWAIT;
1112 	long timeo = sock_sndtimeo(sk, nonblock);
1113 	struct rds_conn_path *cpath;
1114 	struct in6_addr daddr;
1115 	__u32 scope_id = 0;
1116 	size_t rdma_payload_len = 0;
1117 	bool zcopy = ((msg->msg_flags & MSG_ZEROCOPY) &&
1118 		      sock_flag(rds_rs_to_sk(rs), SOCK_ZEROCOPY));
1119 	int num_sgs = DIV_ROUND_UP(payload_len, PAGE_SIZE);
1120 	int namelen;
1121 	struct rds_iov_vector_arr vct;
1122 	int ind;
1123 
1124 	memset(&vct, 0, sizeof(vct));
1125 
1126 	/* expect 1 RDMA CMSG per rds_sendmsg. can still grow if more needed. */
1127 	vct.incr = 1;
1128 
1129 	/* Mirror Linux UDP mirror of BSD error message compatibility */
1130 	/* XXX: Perhaps MSG_MORE someday */
1131 	if (msg->msg_flags & ~(MSG_DONTWAIT | MSG_CMSG_COMPAT | MSG_ZEROCOPY)) {
1132 		ret = -EOPNOTSUPP;
1133 		goto out;
1134 	}
1135 
1136 	namelen = msg->msg_namelen;
1137 	if (namelen != 0) {
1138 		if (namelen < sizeof(*usin)) {
1139 			ret = -EINVAL;
1140 			goto out;
1141 		}
1142 		switch (usin->sin_family) {
1143 		case AF_INET:
1144 			if (usin->sin_addr.s_addr == htonl(INADDR_ANY) ||
1145 			    usin->sin_addr.s_addr == htonl(INADDR_BROADCAST) ||
1146 			    ipv4_is_multicast(usin->sin_addr.s_addr)) {
1147 				ret = -EINVAL;
1148 				goto out;
1149 			}
1150 			ipv6_addr_set_v4mapped(usin->sin_addr.s_addr, &daddr);
1151 			dport = usin->sin_port;
1152 			break;
1153 
1154 #if IS_ENABLED(CONFIG_IPV6)
1155 		case AF_INET6: {
1156 			int addr_type;
1157 
1158 			if (namelen < sizeof(*sin6)) {
1159 				ret = -EINVAL;
1160 				goto out;
1161 			}
1162 			addr_type = ipv6_addr_type(&sin6->sin6_addr);
1163 			if (!(addr_type & IPV6_ADDR_UNICAST)) {
1164 				__be32 addr4;
1165 
1166 				if (!(addr_type & IPV6_ADDR_MAPPED)) {
1167 					ret = -EINVAL;
1168 					goto out;
1169 				}
1170 
1171 				/* It is a mapped address.  Need to do some
1172 				 * sanity checks.
1173 				 */
1174 				addr4 = sin6->sin6_addr.s6_addr32[3];
1175 				if (addr4 == htonl(INADDR_ANY) ||
1176 				    addr4 == htonl(INADDR_BROADCAST) ||
1177 				    ipv4_is_multicast(addr4)) {
1178 					ret = -EINVAL;
1179 					goto out;
1180 				}
1181 			}
1182 			if (addr_type & IPV6_ADDR_LINKLOCAL) {
1183 				if (sin6->sin6_scope_id == 0) {
1184 					ret = -EINVAL;
1185 					goto out;
1186 				}
1187 				scope_id = sin6->sin6_scope_id;
1188 			}
1189 
1190 			daddr = sin6->sin6_addr;
1191 			dport = sin6->sin6_port;
1192 			break;
1193 		}
1194 #endif
1195 
1196 		default:
1197 			ret = -EINVAL;
1198 			goto out;
1199 		}
1200 	} else {
1201 		/* We only care about consistency with ->connect() */
1202 		lock_sock(sk);
1203 		daddr = rs->rs_conn_addr;
1204 		dport = rs->rs_conn_port;
1205 		scope_id = rs->rs_bound_scope_id;
1206 		release_sock(sk);
1207 	}
1208 
1209 	lock_sock(sk);
1210 	if (ipv6_addr_any(&rs->rs_bound_addr) || ipv6_addr_any(&daddr)) {
1211 		release_sock(sk);
1212 		ret = -ENOTCONN;
1213 		goto out;
1214 	} else if (namelen != 0) {
1215 		/* Cannot send to an IPv4 address using an IPv6 source
1216 		 * address and cannot send to an IPv6 address using an
1217 		 * IPv4 source address.
1218 		 */
1219 		if (ipv6_addr_v4mapped(&daddr) ^
1220 		    ipv6_addr_v4mapped(&rs->rs_bound_addr)) {
1221 			release_sock(sk);
1222 			ret = -EOPNOTSUPP;
1223 			goto out;
1224 		}
1225 		/* If the socket is already bound to a link local address,
1226 		 * it can only send to peers on the same link.  But allow
1227 		 * communicating between link local and non-link local address.
1228 		 */
1229 		if (scope_id != rs->rs_bound_scope_id) {
1230 			if (!scope_id) {
1231 				scope_id = rs->rs_bound_scope_id;
1232 			} else if (rs->rs_bound_scope_id) {
1233 				release_sock(sk);
1234 				ret = -EINVAL;
1235 				goto out;
1236 			}
1237 		}
1238 	}
1239 	release_sock(sk);
1240 
1241 	ret = rds_rdma_bytes(msg, &rdma_payload_len);
1242 	if (ret)
1243 		goto out;
1244 
1245 	if (max_t(size_t, payload_len, rdma_payload_len) > RDS_MAX_MSG_SIZE) {
1246 		ret = -EMSGSIZE;
1247 		goto out;
1248 	}
1249 
1250 	if (payload_len > rds_sk_sndbuf(rs)) {
1251 		ret = -EMSGSIZE;
1252 		goto out;
1253 	}
1254 
1255 	if (zcopy) {
1256 		if (rs->rs_transport->t_type != RDS_TRANS_TCP) {
1257 			ret = -EOPNOTSUPP;
1258 			goto out;
1259 		}
1260 		num_sgs = iov_iter_npages(&msg->msg_iter, INT_MAX);
1261 	}
1262 	/* size of rm including all sgs */
1263 	ret = rds_rm_size(msg, num_sgs, &vct);
1264 	if (ret < 0)
1265 		goto out;
1266 
1267 	rm = rds_message_alloc(ret, GFP_KERNEL);
1268 	if (!rm) {
1269 		ret = -ENOMEM;
1270 		goto out;
1271 	}
1272 
1273 	/* Attach data to the rm */
1274 	if (payload_len) {
1275 		rm->data.op_sg = rds_message_alloc_sgs(rm, num_sgs);
1276 		if (IS_ERR(rm->data.op_sg)) {
1277 			ret = PTR_ERR(rm->data.op_sg);
1278 			goto out;
1279 		}
1280 		ret = rds_message_copy_from_user(rm, &msg->msg_iter, zcopy);
1281 		if (ret)
1282 			goto out;
1283 	}
1284 	rm->data.op_active = 1;
1285 
1286 	rm->m_daddr = daddr;
1287 
1288 	/* rds_conn_create has a spinlock that runs with IRQ off.
1289 	 * Caching the conn in the socket helps a lot. */
1290 	if (rs->rs_conn && ipv6_addr_equal(&rs->rs_conn->c_faddr, &daddr) &&
1291 	    rs->rs_tos == rs->rs_conn->c_tos) {
1292 		conn = rs->rs_conn;
1293 	} else {
1294 		conn = rds_conn_create_outgoing(sock_net(sock->sk),
1295 						&rs->rs_bound_addr, &daddr,
1296 						rs->rs_transport, rs->rs_tos,
1297 						sock->sk->sk_allocation,
1298 						scope_id);
1299 		if (IS_ERR(conn)) {
1300 			ret = PTR_ERR(conn);
1301 			goto out;
1302 		}
1303 		rs->rs_conn = conn;
1304 	}
1305 
1306 	if (conn->c_trans->t_mp_capable)
1307 		cpath = &conn->c_path[rds_send_mprds_hash(rs, conn, nonblock)];
1308 	else
1309 		cpath = &conn->c_path[0];
1310 
1311 	rm->m_conn_path = cpath;
1312 
1313 	/* Parse any control messages the user may have included. */
1314 	ret = rds_cmsg_send(rs, rm, msg, &allocated_mr, &vct);
1315 	if (ret)
1316 		goto out;
1317 
1318 	if (rm->rdma.op_active && !conn->c_trans->xmit_rdma) {
1319 		printk_ratelimited(KERN_NOTICE "rdma_op %p conn xmit_rdma %p\n",
1320 			       &rm->rdma, conn->c_trans->xmit_rdma);
1321 		ret = -EOPNOTSUPP;
1322 		goto out;
1323 	}
1324 
1325 	if (rm->atomic.op_active && !conn->c_trans->xmit_atomic) {
1326 		printk_ratelimited(KERN_NOTICE "atomic_op %p conn xmit_atomic %p\n",
1327 			       &rm->atomic, conn->c_trans->xmit_atomic);
1328 		ret = -EOPNOTSUPP;
1329 		goto out;
1330 	}
1331 
1332 	if (rds_destroy_pending(conn)) {
1333 		ret = -EAGAIN;
1334 		goto out;
1335 	}
1336 
1337 	if (rds_conn_path_down(cpath))
1338 		rds_check_all_paths(conn);
1339 
1340 	ret = rds_cong_wait(conn->c_fcong, dport, nonblock, rs);
1341 	if (ret) {
1342 		rs->rs_seen_congestion = 1;
1343 		goto out;
1344 	}
1345 	while (!rds_send_queue_rm(rs, conn, cpath, rm, rs->rs_bound_port,
1346 				  dport, &queued)) {
1347 		rds_stats_inc(s_send_queue_full);
1348 
1349 		if (nonblock) {
1350 			ret = -EAGAIN;
1351 			goto out;
1352 		}
1353 
1354 		timeo = wait_event_interruptible_timeout(*sk_sleep(sk),
1355 					rds_send_queue_rm(rs, conn, cpath, rm,
1356 							  rs->rs_bound_port,
1357 							  dport,
1358 							  &queued),
1359 					timeo);
1360 		rdsdebug("sendmsg woke queued %d timeo %ld\n", queued, timeo);
1361 		if (timeo > 0 || timeo == MAX_SCHEDULE_TIMEOUT)
1362 			continue;
1363 
1364 		ret = timeo;
1365 		if (ret == 0)
1366 			ret = -ETIMEDOUT;
1367 		goto out;
1368 	}
1369 
1370 	/*
1371 	 * By now we've committed to the send.  We reuse rds_send_worker()
1372 	 * to retry sends in the rds thread if the transport asks us to.
1373 	 */
1374 	rds_stats_inc(s_send_queued);
1375 
1376 	ret = rds_send_xmit(cpath);
1377 	if (ret == -ENOMEM || ret == -EAGAIN) {
1378 		ret = 0;
1379 		rcu_read_lock();
1380 		if (rds_destroy_pending(cpath->cp_conn))
1381 			ret = -ENETUNREACH;
1382 		else
1383 			queue_delayed_work(rds_wq, &cpath->cp_send_w, 1);
1384 		rcu_read_unlock();
1385 	}
1386 	if (ret)
1387 		goto out;
1388 	rds_message_put(rm);
1389 
1390 	for (ind = 0; ind < vct.indx; ind++)
1391 		kfree(vct.vec[ind].iov);
1392 	kfree(vct.vec);
1393 
1394 	return payload_len;
1395 
1396 out:
1397 	for (ind = 0; ind < vct.indx; ind++)
1398 		kfree(vct.vec[ind].iov);
1399 	kfree(vct.vec);
1400 
1401 	/* If the user included a RDMA_MAP cmsg, we allocated a MR on the fly.
1402 	 * If the sendmsg goes through, we keep the MR. If it fails with EAGAIN
1403 	 * or in any other way, we need to destroy the MR again */
1404 	if (allocated_mr)
1405 		rds_rdma_unuse(rs, rds_rdma_cookie_key(rm->m_rdma_cookie), 1);
1406 
1407 	if (rm)
1408 		rds_message_put(rm);
1409 	return ret;
1410 }
1411 
1412 /*
1413  * send out a probe. Can be shared by rds_send_ping,
1414  * rds_send_pong, rds_send_hb.
1415  * rds_send_hb should use h_flags
1416  *   RDS_FLAG_HB_PING|RDS_FLAG_ACK_REQUIRED
1417  * or
1418  *   RDS_FLAG_HB_PONG|RDS_FLAG_ACK_REQUIRED
1419  */
1420 static int
rds_send_probe(struct rds_conn_path * cp,__be16 sport,__be16 dport,u8 h_flags)1421 rds_send_probe(struct rds_conn_path *cp, __be16 sport,
1422 	       __be16 dport, u8 h_flags)
1423 {
1424 	struct rds_message *rm;
1425 	unsigned long flags;
1426 	int ret = 0;
1427 
1428 	rm = rds_message_alloc(0, GFP_ATOMIC);
1429 	if (!rm) {
1430 		ret = -ENOMEM;
1431 		goto out;
1432 	}
1433 
1434 	rm->m_daddr = cp->cp_conn->c_faddr;
1435 	rm->data.op_active = 1;
1436 
1437 	rds_conn_path_connect_if_down(cp);
1438 
1439 	ret = rds_cong_wait(cp->cp_conn->c_fcong, dport, 1, NULL);
1440 	if (ret)
1441 		goto out;
1442 
1443 	spin_lock_irqsave(&cp->cp_lock, flags);
1444 	list_add_tail(&rm->m_conn_item, &cp->cp_send_queue);
1445 	set_bit(RDS_MSG_ON_CONN, &rm->m_flags);
1446 	rds_message_addref(rm);
1447 	rm->m_inc.i_conn = cp->cp_conn;
1448 	rm->m_inc.i_conn_path = cp;
1449 
1450 	rds_message_populate_header(&rm->m_inc.i_hdr, sport, dport,
1451 				    cp->cp_next_tx_seq);
1452 	rm->m_inc.i_hdr.h_flags |= h_flags;
1453 	cp->cp_next_tx_seq++;
1454 
1455 	if (RDS_HS_PROBE(be16_to_cpu(sport), be16_to_cpu(dport)) &&
1456 	    cp->cp_conn->c_trans->t_mp_capable) {
1457 		u16 npaths = cpu_to_be16(RDS_MPATH_WORKERS);
1458 		u32 my_gen_num = cpu_to_be32(cp->cp_conn->c_my_gen_num);
1459 
1460 		rds_message_add_extension(&rm->m_inc.i_hdr,
1461 					  RDS_EXTHDR_NPATHS, &npaths,
1462 					  sizeof(npaths));
1463 		rds_message_add_extension(&rm->m_inc.i_hdr,
1464 					  RDS_EXTHDR_GEN_NUM,
1465 					  &my_gen_num,
1466 					  sizeof(u32));
1467 	}
1468 	spin_unlock_irqrestore(&cp->cp_lock, flags);
1469 
1470 	rds_stats_inc(s_send_queued);
1471 	rds_stats_inc(s_send_pong);
1472 
1473 	/* schedule the send work on rds_wq */
1474 	rcu_read_lock();
1475 	if (!rds_destroy_pending(cp->cp_conn))
1476 		queue_delayed_work(rds_wq, &cp->cp_send_w, 1);
1477 	rcu_read_unlock();
1478 
1479 	rds_message_put(rm);
1480 	return 0;
1481 
1482 out:
1483 	if (rm)
1484 		rds_message_put(rm);
1485 	return ret;
1486 }
1487 
1488 int
rds_send_pong(struct rds_conn_path * cp,__be16 dport)1489 rds_send_pong(struct rds_conn_path *cp, __be16 dport)
1490 {
1491 	return rds_send_probe(cp, 0, dport, 0);
1492 }
1493 
1494 void
rds_send_ping(struct rds_connection * conn,int cp_index)1495 rds_send_ping(struct rds_connection *conn, int cp_index)
1496 {
1497 	unsigned long flags;
1498 	struct rds_conn_path *cp = &conn->c_path[cp_index];
1499 
1500 	spin_lock_irqsave(&cp->cp_lock, flags);
1501 	if (conn->c_ping_triggered) {
1502 		spin_unlock_irqrestore(&cp->cp_lock, flags);
1503 		return;
1504 	}
1505 	conn->c_ping_triggered = 1;
1506 	spin_unlock_irqrestore(&cp->cp_lock, flags);
1507 	rds_send_probe(cp, cpu_to_be16(RDS_FLAG_PROBE_PORT), 0, 0);
1508 }
1509 EXPORT_SYMBOL_GPL(rds_send_ping);
1510