projects/linux·blob master · refs

892 lines · 25.0 KiB

1/*
2 * Copyright (c) 2006, 2019 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/slab.h>
35#include <net/sock.h>
36#include <linux/in.h>
37#include <linux/export.h>
38#include <linux/sched/clock.h>
39#include <linux/time.h>
40#include <linux/rds.h>
41
42#include "rds.h"
43
44void rds_inc_init(struct rds_incoming *inc, struct rds_connection *conn,
45 struct in6_addr *saddr)
46{
47 refcount_set(&inc->i_refcount, 1);
48 INIT_LIST_HEAD(&inc->i_item);
49 inc->i_conn = conn;
50 inc->i_conn_path = NULL;
51 inc->i_saddr = *saddr;
52 inc->i_usercopy.rdma_cookie = 0;
53 inc->i_usercopy.rx_tstamp = ktime_set(0, 0);
54
55 memset(inc->i_rx_lat_trace, 0, sizeof(inc->i_rx_lat_trace));
56}
57EXPORT_SYMBOL_GPL(rds_inc_init);
58
59void rds_inc_path_init(struct rds_incoming *inc, struct rds_conn_path *cp,
60 struct in6_addr *saddr)
61{
62 refcount_set(&inc->i_refcount, 1);
63 INIT_LIST_HEAD(&inc->i_item);
64 inc->i_conn = cp->cp_conn;
65 inc->i_conn_path = cp;
66 inc->i_saddr = *saddr;
67 inc->i_usercopy.rdma_cookie = 0;
68 inc->i_usercopy.rx_tstamp = ktime_set(0, 0);
69
70 memset(inc->i_rx_lat_trace, 0, sizeof(inc->i_rx_lat_trace));
71}
72EXPORT_SYMBOL_GPL(rds_inc_path_init);
73
74static void rds_inc_addref(struct rds_incoming *inc)
75{
76 rdsdebug("addref inc %p ref %d\n", inc, refcount_read(&inc->i_refcount));
77 refcount_inc(&inc->i_refcount);
78}
79
80void rds_inc_put(struct rds_incoming *inc)
81{
82 rdsdebug("put inc %p ref %d\n", inc, refcount_read(&inc->i_refcount));
83 if (refcount_dec_and_test(&inc->i_refcount)) {
84 BUG_ON(!list_empty(&inc->i_item));
85
86 inc->i_conn->c_trans->inc_free(inc);
87 }
88}
89EXPORT_SYMBOL_GPL(rds_inc_put);
90
91static void rds_recv_rcvbuf_delta(struct rds_sock *rs, struct sock *sk,
92 struct rds_cong_map *map,
93 int delta, __be16 port)
94{
95 int now_congested;
96
97 if (delta == 0)
98 return;
99
100 rs->rs_rcv_bytes += delta;
101 if (delta > 0)
102 rds_stats_add(s_recv_bytes_added_to_socket, delta);
103 else
104 rds_stats_add(s_recv_bytes_removed_from_socket, -delta);
105
106 /* loop transport doesn't send/recv congestion updates */
107 if (rs->rs_transport->t_type == RDS_TRANS_LOOP)
108 return;
109
110 now_congested = rs->rs_rcv_bytes > rds_sk_rcvbuf(rs);
111
112 rdsdebug("rs %p (%pI6c:%u) recv bytes %d buf %d "
113 "now_cong %d delta %d\n",
114 rs, &rs->rs_bound_addr,
115 ntohs(rs->rs_bound_port), rs->rs_rcv_bytes,
116 rds_sk_rcvbuf(rs), now_congested, delta);
117
118 /* wasn't -> am congested */
119 if (!rs->rs_congested && now_congested) {
120 rs->rs_congested = 1;
121 rds_cong_set_bit(map, port);
122 rds_cong_queue_updates(map);
123 }
124 /* was -> aren't congested */
125 /* Require more free space before reporting uncongested to prevent
126 bouncing cong/uncong state too often */
127 else if (rs->rs_congested && (rs->rs_rcv_bytes < (rds_sk_rcvbuf(rs)/2))) {
128 rs->rs_congested = 0;
129 rds_cong_clear_bit(map, port);
130 rds_cong_queue_updates(map);
131 }
132
133 /* do nothing if no change in cong state */
134}
135
136static void rds_conn_peer_gen_update(struct rds_connection *conn,
137 u32 peer_gen_num)
138{
139 int i;
140 struct rds_message *rm, *tmp;
141 unsigned long flags;
142
143 WARN_ON(conn->c_trans->t_type != RDS_TRANS_TCP);
144 if (peer_gen_num != 0) {
145 if (conn->c_peer_gen_num != 0 &&
146 peer_gen_num != conn->c_peer_gen_num) {
147 for (i = 0; i < RDS_MPATH_WORKERS; i++) {
148 struct rds_conn_path *cp;
149
150 cp = &conn->c_path[i];
151 spin_lock_irqsave(&cp->cp_lock, flags);
152 cp->cp_next_tx_seq = 1;
153 cp->cp_next_rx_seq = 0;
154 list_for_each_entry_safe(rm, tmp,
155 &cp->cp_retrans,
156 m_conn_item) {
157 set_bit(RDS_MSG_FLUSH, &rm->m_flags);
158 }
159 spin_unlock_irqrestore(&cp->cp_lock, flags);
160 }
161 }
162 conn->c_peer_gen_num = peer_gen_num;
163 }
164}
165
166/*
167 * Process all extension headers that come with this message.
168 */
169static void rds_recv_incoming_exthdrs(struct rds_incoming *inc, struct rds_sock *rs)
170{
171 struct rds_header *hdr = &inc->i_hdr;
172 unsigned int pos = 0, type, len;
173 union {
174 struct rds_ext_header_version version;
175 struct rds_ext_header_rdma rdma;
176 struct rds_ext_header_rdma_dest rdma_dest;
177 } buffer;
178
179 while (1) {
180 len = sizeof(buffer);
181 type = rds_message_next_extension(hdr, &pos, &buffer, &len);
182 if (type == RDS_EXTHDR_NONE)
183 break;
184 /* Process extension header here */
185 switch (type) {
186 case RDS_EXTHDR_RDMA:
187 rds_rdma_unuse(rs, be32_to_cpu(buffer.rdma.h_rdma_rkey), 0);
188 break;
189
190 case RDS_EXTHDR_RDMA_DEST:
191 /* We ignore the size for now. We could stash it
192 * somewhere and use it for error checking. */
193 inc->i_usercopy.rdma_cookie = rds_rdma_make_cookie(
194 be32_to_cpu(buffer.rdma_dest.h_rdma_rkey),
195 be32_to_cpu(buffer.rdma_dest.h_rdma_offset));
196
197 break;
198 }
199 }
200}
201
202static void rds_recv_hs_exthdrs(struct rds_header *hdr,
203 struct rds_connection *conn)
204{
205 unsigned int pos = 0, type, len;
206 union {
207 struct rds_ext_header_version version;
208 __be16 rds_npaths;
209 __be32 rds_gen_num;
210 u8 dummy;
211 } buffer;
212 bool new_with_sport_idx = false;
213 u32 new_peer_gen_num = 0;
214 int new_npaths;
215 bool fan_out;
216
217 new_npaths = conn->c_npaths;
218
219 while (1) {
220 len = sizeof(buffer);
221 type = rds_message_next_extension(hdr, &pos, &buffer, &len);
222 if (type == RDS_EXTHDR_NONE)
223 break;
224 /* Process extension header here */
225 switch (type) {
226 case RDS_EXTHDR_NPATHS:
227 new_npaths = min_t(int, RDS_MPATH_WORKERS,
228 be16_to_cpu(buffer.rds_npaths));
229 break;
230 case RDS_EXTHDR_GEN_NUM:
231 new_peer_gen_num = be32_to_cpu(buffer.rds_gen_num);
232 break;
233 case RDS_EXTHDR_SPORT_IDX:
234 new_with_sport_idx = true;
235 break;
236 default:
237 pr_warn_ratelimited("ignoring unknown exthdr type "
238 "0x%x\n", type);
239 }
240 }
241
242 conn->c_with_sport_idx = new_with_sport_idx;
243
244 if (new_npaths > 1 && new_npaths != conn->c_npaths) {
245 /* We're about to fan-out.
246 * Make sure that messages from cp_index#0
247 * are sent prior to handling other lanes.
248 */
249 struct rds_conn_path *cp0 = conn->c_path;
250 unsigned long flags;
251
252 spin_lock_irqsave(&cp0->cp_lock, flags);
253 conn->c_cp0_mprds_catchup_tx_seq = cp0->cp_next_tx_seq;
254 spin_unlock_irqrestore(&cp0->cp_lock, flags);
255 fan_out = true;
256 } else {
257 fan_out = false;
258 }
259
260 /* if RDS_EXTHDR_NPATHS was not found, default to a single-path */
261 conn->c_npaths = max_t(int, new_npaths, 1);
262
263 conn->c_ping_triggered = 0;
264 rds_conn_peer_gen_update(conn, new_peer_gen_num);
265
266 if (conn->c_npaths > 1 &&
267 conn->c_trans->conn_slots_available)
268 conn->c_trans->conn_slots_available(conn, fan_out);
269}
270
271/* rds_start_mprds() will synchronously start multiple paths when appropriate.
272 * The scheme is based on the following rules:
273 *
274 * 1. rds_sendmsg on first connect attempt sends the probe ping, with the
275 * sender's npaths (s_npaths)
276 * 2. rcvr of probe-ping knows the mprds_paths = min(s_npaths, r_npaths). It
277 * sends back a probe-pong with r_npaths. After that, if rcvr is the
278 * smaller ip addr, it starts rds_conn_path_connect_if_down on all
279 * mprds_paths.
280 * 3. sender gets woken up, and can move to rds_conn_path_connect_if_down.
281 * If it is the smaller ipaddr, rds_conn_path_connect_if_down can be
282 * called after reception of the probe-pong on all mprds_paths.
283 * Otherwise (sender of probe-ping is not the smaller ip addr): just call
284 * rds_conn_path_connect_if_down on the hashed path. (see rule 4)
285 * 4. rds_connect_worker must only trigger a connection if laddr < faddr.
286 * 5. sender may end up queuing the packet on the cp. will get sent out later.
287 * when connection is completed.
288 */
289static void rds_start_mprds(struct rds_connection *conn)
290{
291 int i;
292 struct rds_conn_path *cp;
293
294 if (conn->c_npaths > 1 &&
295 rds_addr_cmp(&conn->c_laddr, &conn->c_faddr) < 0) {
296 for (i = 0; i < conn->c_npaths; i++) {
297 cp = &conn->c_path[i];
298 rds_conn_path_connect_if_down(cp);
299 }
300 }
301}
302
303/*
304 * The transport must make sure that this is serialized against other
305 * rx and conn reset on this specific conn.
306 *
307 * We currently assert that only one fragmented message will be sent
308 * down a connection at a time. This lets us reassemble in the conn
309 * instead of per-flow which means that we don't have to go digging through
310 * flows to tear down partial reassembly progress on conn failure and
311 * we save flow lookup and locking for each frag arrival. It does mean
312 * that small messages will wait behind large ones. Fragmenting at all
313 * is only to reduce the memory consumption of pre-posted buffers.
314 *
315 * The caller passes in saddr and daddr instead of us getting it from the
316 * conn. This lets loopback, who only has one conn for both directions,
317 * tell us which roles the addrs in the conn are playing for this message.
318 */
319void rds_recv_incoming(struct rds_connection *conn, struct in6_addr *saddr,
320 struct in6_addr *daddr,
321 struct rds_incoming *inc, gfp_t gfp)
322{
323 struct rds_sock *rs = NULL;
324 struct sock *sk;
325 unsigned long flags;
326 struct rds_conn_path *cp;
327
328 inc->i_conn = conn;
329 inc->i_rx_jiffies = jiffies;
330 if (conn->c_trans->t_mp_capable)
331 cp = inc->i_conn_path;
332 else
333 cp = &conn->c_path[0];
334
335 rdsdebug("conn %p next %llu inc %p seq %llu len %u sport %u dport %u "
336 "flags 0x%x rx_jiffies %lu\n", conn,
337 (unsigned long long)cp->cp_next_rx_seq,
338 inc,
339 (unsigned long long)be64_to_cpu(inc->i_hdr.h_sequence),
340 be32_to_cpu(inc->i_hdr.h_len),
341 be16_to_cpu(inc->i_hdr.h_sport),
342 be16_to_cpu(inc->i_hdr.h_dport),
343 inc->i_hdr.h_flags,
344 inc->i_rx_jiffies);
345
346 /*
347 * Sequence numbers should only increase. Messages get their
348 * sequence number as they're queued in a sending conn. They
349 * can be dropped, though, if the sending socket is closed before
350 * they hit the wire. So sequence numbers can skip forward
351 * under normal operation. They can also drop back in the conn
352 * failover case as previously sent messages are resent down the
353 * new instance of a conn. We drop those, otherwise we have
354 * to assume that the next valid seq does not come after a
355 * hole in the fragment stream.
356 *
357 * The headers don't give us a way to realize if fragments of
358 * a message have been dropped. We assume that frags that arrive
359 * to a flow are part of the current message on the flow that is
360 * being reassembled. This means that senders can't drop messages
361 * from the sending conn until all their frags are sent.
362 *
363 * XXX we could spend more on the wire to get more robust failure
364 * detection, arguably worth it to avoid data corruption.
365 */
366 if (be64_to_cpu(inc->i_hdr.h_sequence) < cp->cp_next_rx_seq &&
367 (inc->i_hdr.h_flags & RDS_FLAG_RETRANSMITTED)) {
368 rds_stats_inc(s_recv_drop_old_seq);
369 goto out;
370 }
371 cp->cp_next_rx_seq = be64_to_cpu(inc->i_hdr.h_sequence) + 1;
372
373 if (rds_sysctl_ping_enable && inc->i_hdr.h_dport == 0) {
374 if (inc->i_hdr.h_sport == 0) {
375 rdsdebug("ignore ping with 0 sport from %pI6c\n",
376 saddr);
377 goto out;
378 }
379 rds_stats_inc(s_recv_ping);
380 rds_send_pong(cp, inc->i_hdr.h_sport);
381 /* if this is a handshake ping, start multipath if necessary */
382 if (RDS_HS_PROBE(be16_to_cpu(inc->i_hdr.h_sport),
383 be16_to_cpu(inc->i_hdr.h_dport))) {
384 rds_recv_hs_exthdrs(&inc->i_hdr, cp->cp_conn);
385 rds_start_mprds(cp->cp_conn);
386 }
387 goto out;
388 }
389
390 if (be16_to_cpu(inc->i_hdr.h_dport) == RDS_FLAG_PROBE_PORT &&
391 inc->i_hdr.h_sport == 0) {
392 rds_recv_hs_exthdrs(&inc->i_hdr, cp->cp_conn);
393 /* if this is a handshake pong, start multipath if necessary */
394 rds_start_mprds(cp->cp_conn);
395 wake_up(&cp->cp_conn->c_hs_waitq);
396 goto out;
397 }
398
399 rs = rds_find_bound(daddr, inc->i_hdr.h_dport, conn->c_bound_if);
400 if (!rs) {
401 rds_stats_inc(s_recv_drop_no_sock);
402 goto out;
403 }
404
405 /*
406 * rds_find_bound() uses a global (netns-agnostic) hash table.
407 * An RDS connection created in netns A can match a socket bound
408 * in the init netns, delivering inc cross-netns with inc->i_conn
409 * pointing into netns A. When cleanup_net() then frees that conn,
410 * any subsequent dereference of inc->i_conn is a use-after-free.
411 * Drop the inc if the receiving socket lives in a different netns.
412 */
413 if (!net_eq(sock_net(rds_rs_to_sk(rs)), rds_conn_net(conn))) {
414 rds_stats_inc(s_recv_drop_no_sock);
415 rds_sock_put(rs);
416 rs = NULL;
417 goto out;
418 }
419
420 /* Process extension headers */
421 rds_recv_incoming_exthdrs(inc, rs);
422
423 /* We can be racing with rds_release() which marks the socket dead. */
424 sk = rds_rs_to_sk(rs);
425
426 /* serialize with rds_release -> sock_orphan */
427 write_lock_irqsave(&rs->rs_recv_lock, flags);
428 if (!sock_flag(sk, SOCK_DEAD)) {
429 rdsdebug("adding inc %p to rs %p's recv queue\n", inc, rs);
430 rds_stats_inc(s_recv_queued);
431 rds_recv_rcvbuf_delta(rs, sk, inc->i_conn->c_lcong,
432 be32_to_cpu(inc->i_hdr.h_len),
433 inc->i_hdr.h_dport);
434 if (sock_flag(sk, SOCK_RCVTSTAMP))
435 inc->i_usercopy.rx_tstamp = ktime_get_real();
436 rds_inc_addref(inc);
437 inc->i_rx_lat_trace[RDS_MSG_RX_END] = local_clock();
438 list_add_tail(&inc->i_item, &rs->rs_recv_queue);
439 __rds_wake_sk_sleep(sk);
440 } else {
441 rds_stats_inc(s_recv_drop_dead_sock);
442 }
443 write_unlock_irqrestore(&rs->rs_recv_lock, flags);
444
445out:
446 if (rs)
447 rds_sock_put(rs);
448}
449EXPORT_SYMBOL_GPL(rds_recv_incoming);
450
451/*
452 * be very careful here. This is being called as the condition in
453 * wait_event_*() needs to cope with being called many times.
454 */
455static int rds_next_incoming(struct rds_sock *rs, struct rds_incoming **inc)
456{
457 unsigned long flags;
458
459 if (!*inc) {
460 read_lock_irqsave(&rs->rs_recv_lock, flags);
461 if (!list_empty(&rs->rs_recv_queue)) {
462 *inc = list_entry(rs->rs_recv_queue.next,
463 struct rds_incoming,
464 i_item);
465 rds_inc_addref(*inc);
466 }
467 read_unlock_irqrestore(&rs->rs_recv_lock, flags);
468 }
469
470 return *inc != NULL;
471}
472
473static int rds_still_queued(struct rds_sock *rs, struct rds_incoming *inc,
474 int drop)
475{
476 struct sock *sk = rds_rs_to_sk(rs);
477 int ret = 0;
478 unsigned long flags;
479 struct rds_incoming *to_drop = NULL;
480
481 write_lock_irqsave(&rs->rs_recv_lock, flags);
482 if (!list_empty(&inc->i_item)) {
483 ret = 1;
484 if (drop) {
485 /* XXX make sure this i_conn is reliable */
486 rds_recv_rcvbuf_delta(rs, sk, inc->i_conn->c_lcong,
487 -be32_to_cpu(inc->i_hdr.h_len),
488 inc->i_hdr.h_dport);
489 list_del_init(&inc->i_item);
490 to_drop = inc;
491 }
492 }
493 write_unlock_irqrestore(&rs->rs_recv_lock, flags);
494
495 if (to_drop)
496 rds_inc_put(to_drop);
497
498 rdsdebug("inc %p rs %p still %d dropped %d\n", inc, rs, ret, drop);
499 return ret;
500}
501
502/*
503 * Pull errors off the error queue.
504 * If msghdr is NULL, we will just purge the error queue.
505 */
506int rds_notify_queue_get(struct rds_sock *rs, struct msghdr *msghdr)
507{
508 struct rds_notifier *notifier;
509 struct rds_rdma_notify cmsg;
510 unsigned int count = 0, max_messages = ~0U;
511 unsigned long flags;
512 LIST_HEAD(copy);
513 int err = 0;
514
515 memset(&cmsg, 0, sizeof(cmsg)); /* fill holes with zero */
516
517 /* put_cmsg copies to user space and thus may sleep. We can't do this
518 * with rs_lock held, so first grab as many notifications as we can stuff
519 * in the user provided cmsg buffer. We don't try to copy more, to avoid
520 * losing notifications - except when the buffer is so small that it wouldn't
521 * even hold a single notification. Then we give him as much of this single
522 * msg as we can squeeze in, and set MSG_CTRUNC.
523 */
524 if (msghdr) {
525 max_messages = msghdr->msg_controllen / CMSG_SPACE(sizeof(cmsg));
526 if (!max_messages)
527 max_messages = 1;
528 }
529
530 spin_lock_irqsave(&rs->rs_lock, flags);
531 while (!list_empty(&rs->rs_notify_queue) && count < max_messages) {
532 notifier = list_entry(rs->rs_notify_queue.next,
533 struct rds_notifier, n_list);
534 list_move(&notifier->n_list, &copy);
535 count++;
536 }
537 spin_unlock_irqrestore(&rs->rs_lock, flags);
538
539 if (!count)
540 return 0;
541
542 while (!list_empty(&copy)) {
543 notifier = list_entry(copy.next, struct rds_notifier, n_list);
544
545 if (msghdr) {
546 cmsg.user_token = notifier->n_user_token;
547 cmsg.status = notifier->n_status;
548
549 err = put_cmsg(msghdr, SOL_RDS, RDS_CMSG_RDMA_STATUS,
550 sizeof(cmsg), &cmsg);
551 if (err)
552 break;
553 }
554
555 list_del_init(&notifier->n_list);
556 kfree(notifier);
557 }
558
559 /* If we bailed out because of an error in put_cmsg,
560 * we may be left with one or more notifications that we
561 * didn't process. Return them to the head of the list. */
562 if (!list_empty(&copy)) {
563 spin_lock_irqsave(&rs->rs_lock, flags);
564 list_splice(&copy, &rs->rs_notify_queue);
565 spin_unlock_irqrestore(&rs->rs_lock, flags);
566 }
567
568 return err;
569}
570
571/*
572 * Queue a congestion notification
573 */
574static int rds_notify_cong(struct rds_sock *rs, struct msghdr *msghdr)
575{
576 uint64_t notify = rs->rs_cong_notify;
577 unsigned long flags;
578 int err;
579
580 err = put_cmsg(msghdr, SOL_RDS, RDS_CMSG_CONG_UPDATE,
581 sizeof(notify), &notify);
582 if (err)
583 return err;
584
585 spin_lock_irqsave(&rs->rs_lock, flags);
586 rs->rs_cong_notify &= ~notify;
587 spin_unlock_irqrestore(&rs->rs_lock, flags);
588
589 return 0;
590}
591
592/*
593 * Receive any control messages.
594 */
595static int rds_cmsg_recv(struct rds_incoming *inc, struct msghdr *msg,
596 struct rds_sock *rs)
597{
598 int ret = 0;
599
600 if (inc->i_usercopy.rdma_cookie) {
601 ret = put_cmsg(msg, SOL_RDS, RDS_CMSG_RDMA_DEST,
602 sizeof(inc->i_usercopy.rdma_cookie),
603 &inc->i_usercopy.rdma_cookie);
604 if (ret)
605 goto out;
606 }
607
608 if ((inc->i_usercopy.rx_tstamp != 0) &&
609 sock_flag(rds_rs_to_sk(rs), SOCK_RCVTSTAMP)) {
610 struct __kernel_old_timeval tv =
611 ns_to_kernel_old_timeval(inc->i_usercopy.rx_tstamp);
612
613 if (!sock_flag(rds_rs_to_sk(rs), SOCK_TSTAMP_NEW)) {
614 ret = put_cmsg(msg, SOL_SOCKET, SO_TIMESTAMP_OLD,
615 sizeof(tv), &tv);
616 } else {
617 struct __kernel_sock_timeval sk_tv;
618
619 sk_tv.tv_sec = tv.tv_sec;
620 sk_tv.tv_usec = tv.tv_usec;
621
622 ret = put_cmsg(msg, SOL_SOCKET, SO_TIMESTAMP_NEW,
623 sizeof(sk_tv), &sk_tv);
624 }
625
626 if (ret)
627 goto out;
628 }
629
630 if (rs->rs_rx_traces) {
631 struct rds_cmsg_rx_trace t;
632 int i, j;
633
634 memset(&t, 0, sizeof(t));
635 inc->i_rx_lat_trace[RDS_MSG_RX_CMSG] = local_clock();
636 t.rx_traces = rs->rs_rx_traces;
637 for (i = 0; i < rs->rs_rx_traces; i++) {
638 j = rs->rs_rx_trace[i];
639 t.rx_trace_pos[i] = j;
640 t.rx_trace[i] = inc->i_rx_lat_trace[j + 1] -
641 inc->i_rx_lat_trace[j];
642 }
643
644 ret = put_cmsg(msg, SOL_RDS, RDS_CMSG_RXPATH_LATENCY,
645 sizeof(t), &t);
646 if (ret)
647 goto out;
648 }
649
650out:
651 return ret;
652}
653
654static bool rds_recvmsg_zcookie(struct rds_sock *rs, struct msghdr *msg)
655{
656 struct rds_msg_zcopy_queue *q = &rs->rs_zcookie_queue;
657 struct rds_msg_zcopy_info *info = NULL;
658 struct rds_zcopy_cookies *done;
659 unsigned long flags;
660
661 if (!msg->msg_control)
662 return false;
663
664 if (!sock_flag(rds_rs_to_sk(rs), SOCK_ZEROCOPY) ||
665 msg->msg_controllen < CMSG_SPACE(sizeof(*done)))
666 return false;
667
668 spin_lock_irqsave(&q->lock, flags);
669 if (!list_empty(&q->zcookie_head)) {
670 info = list_entry(q->zcookie_head.next,
671 struct rds_msg_zcopy_info, rs_zcookie_next);
672 list_del(&info->rs_zcookie_next);
673 }
674 spin_unlock_irqrestore(&q->lock, flags);
675 if (!info)
676 return false;
677 done = &info->zcookies;
678 if (put_cmsg(msg, SOL_RDS, RDS_CMSG_ZCOPY_COMPLETION, sizeof(*done),
679 done)) {
680 spin_lock_irqsave(&q->lock, flags);
681 list_add(&info->rs_zcookie_next, &q->zcookie_head);
682 spin_unlock_irqrestore(&q->lock, flags);
683 return false;
684 }
685 kfree(info);
686 return true;
687}
688
689int rds_recvmsg(struct socket *sock, struct msghdr *msg, size_t size,
690 int msg_flags)
691{
692 struct sock *sk = sock->sk;
693 struct rds_sock *rs = rds_sk_to_rs(sk);
694 long timeo;
695 int ret = 0, nonblock = msg_flags & MSG_DONTWAIT;
696 DECLARE_SOCKADDR(struct sockaddr_in6 *, sin6, msg->msg_name);
697 DECLARE_SOCKADDR(struct sockaddr_in *, sin, msg->msg_name);
698 struct rds_incoming *inc = NULL;
699
700 /* udp_recvmsg()->sock_recvtimeo() gets away without locking too.. */
701 timeo = sock_rcvtimeo(sk, nonblock);
702
703 rdsdebug("size %zu flags 0x%x timeo %ld\n", size, msg_flags, timeo);
704
705 if (msg_flags & MSG_OOB)
706 goto out;
707 if (msg_flags & MSG_ERRQUEUE)
708 return sock_recv_errqueue(sk, msg, size, SOL_IP, IP_RECVERR);
709
710 while (1) {
711 /* If there are pending notifications, do those - and nothing else */
712 if (!list_empty(&rs->rs_notify_queue)) {
713 ret = rds_notify_queue_get(rs, msg);
714 break;
715 }
716
717 if (rs->rs_cong_notify) {
718 ret = rds_notify_cong(rs, msg);
719 break;
720 }
721
722 if (!rds_next_incoming(rs, &inc)) {
723 if (nonblock) {
724 bool reaped = rds_recvmsg_zcookie(rs, msg);
725
726 ret = reaped ? 0 : -EAGAIN;
727 break;
728 }
729
730 timeo = wait_event_interruptible_timeout(*sk_sleep(sk),
731 (!list_empty(&rs->rs_notify_queue) ||
732 rs->rs_cong_notify ||
733 rds_next_incoming(rs, &inc)), timeo);
734 rdsdebug("recvmsg woke inc %p timeo %ld\n", inc,
735 timeo);
736 if (timeo > 0 || timeo == MAX_SCHEDULE_TIMEOUT)
737 continue;
738
739 ret = timeo;
740 if (ret == 0)
741 ret = -ETIMEDOUT;
742 break;
743 }
744
745 rdsdebug("copying inc %p from %pI6c:%u to user\n", inc,
746 &inc->i_conn->c_faddr,
747 ntohs(inc->i_hdr.h_sport));
748 ret = inc->i_conn->c_trans->inc_copy_to_user(inc, &msg->msg_iter);
749 if (ret < 0)
750 break;
751
752 /*
753 * if the message we just copied isn't at the head of the
754 * recv queue then someone else raced us to return it, try
755 * to get the next message.
756 */
757 if (!rds_still_queued(rs, inc, !(msg_flags & MSG_PEEK))) {
758 rds_inc_put(inc);
759 inc = NULL;
760 rds_stats_inc(s_recv_deliver_raced);
761 iov_iter_revert(&msg->msg_iter, ret);
762 continue;
763 }
764
765 if (ret < be32_to_cpu(inc->i_hdr.h_len)) {
766 if (msg_flags & MSG_TRUNC)
767 ret = be32_to_cpu(inc->i_hdr.h_len);
768 msg->msg_flags |= MSG_TRUNC;
769 }
770
771 if (rds_cmsg_recv(inc, msg, rs)) {
772 ret = -EFAULT;
773 break;
774 }
775 rds_recvmsg_zcookie(rs, msg);
776
777 rds_stats_inc(s_recv_delivered);
778
779 if (msg->msg_name) {
780 if (ipv6_addr_v4mapped(&inc->i_saddr)) {
781 sin->sin_family = AF_INET;
782 sin->sin_port = inc->i_hdr.h_sport;
783 sin->sin_addr.s_addr =
784 inc->i_saddr.s6_addr32[3];
785 memset(sin->sin_zero, 0, sizeof(sin->sin_zero));
786 msg->msg_namelen = sizeof(*sin);
787 } else {
788 sin6->sin6_family = AF_INET6;
789 sin6->sin6_port = inc->i_hdr.h_sport;
790 sin6->sin6_addr = inc->i_saddr;
791 sin6->sin6_flowinfo = 0;
792 sin6->sin6_scope_id = rs->rs_bound_scope_id;
793 msg->msg_namelen = sizeof(*sin6);
794 }
795 }
796 break;
797 }
798
799 if (inc)
800 rds_inc_put(inc);
801
802out:
803 return ret;
804}
805
806/*
807 * The socket is being shut down and we're asked to drop messages that were
808 * queued for recvmsg. The caller has unbound the socket so the receive path
809 * won't queue any more incoming fragments or messages on the socket.
810 */
811void rds_clear_recv_queue(struct rds_sock *rs)
812{
813 struct sock *sk = rds_rs_to_sk(rs);
814 struct rds_incoming *inc, *tmp;
815 unsigned long flags;
816 LIST_HEAD(to_drop);
817
818 write_lock_irqsave(&rs->rs_recv_lock, flags);
819 list_for_each_entry_safe(inc, tmp, &rs->rs_recv_queue, i_item) {
820 rds_recv_rcvbuf_delta(rs, sk, inc->i_conn->c_lcong,
821 -be32_to_cpu(inc->i_hdr.h_len),
822 inc->i_hdr.h_dport);
823 list_move(&inc->i_item, &to_drop);
824 }
825 write_unlock_irqrestore(&rs->rs_recv_lock, flags);
826
827 list_for_each_entry_safe(inc, tmp, &to_drop, i_item) {
828 list_del_init(&inc->i_item);
829 rds_inc_put(inc);
830 }
831}
832
833/*
834 * inc->i_saddr isn't used here because it is only set in the receive
835 * path.
836 */
837void rds_inc_info_copy(struct rds_incoming *inc,
838 struct rds_info_iterator *iter,
839 __be32 saddr, __be32 daddr, int flip)
840{
841 struct rds_info_message minfo;
842
843 minfo.seq = be64_to_cpu(inc->i_hdr.h_sequence);
844 minfo.len = be32_to_cpu(inc->i_hdr.h_len);
845 minfo.tos = inc->i_conn->c_tos;
846
847 if (flip) {
848 minfo.laddr = daddr;
849 minfo.faddr = saddr;
850 minfo.lport = inc->i_hdr.h_dport;
851 minfo.fport = inc->i_hdr.h_sport;
852 } else {
853 minfo.laddr = saddr;
854 minfo.faddr = daddr;
855 minfo.lport = inc->i_hdr.h_sport;
856 minfo.fport = inc->i_hdr.h_dport;
857 }
858
859 minfo.flags = 0;
860
861 rds_info_copy(iter, &minfo, sizeof(minfo));
862}
863
864#if IS_ENABLED(CONFIG_IPV6)
865void rds6_inc_info_copy(struct rds_incoming *inc,
866 struct rds_info_iterator *iter,
867 struct in6_addr *saddr, struct in6_addr *daddr,
868 int flip)
869{
870 struct rds6_info_message minfo6;
871
872 minfo6.seq = be64_to_cpu(inc->i_hdr.h_sequence);
873 minfo6.len = be32_to_cpu(inc->i_hdr.h_len);
874 minfo6.tos = inc->i_conn->c_tos;
875
876 if (flip) {
877 minfo6.laddr = *daddr;
878 minfo6.faddr = *saddr;
879 minfo6.lport = inc->i_hdr.h_dport;
880 minfo6.fport = inc->i_hdr.h_sport;
881 } else {
882 minfo6.laddr = *saddr;
883 minfo6.faddr = *daddr;
884 minfo6.lport = inc->i_hdr.h_sport;
885 minfo6.fport = inc->i_hdr.h_dport;
886 }
887
888 minfo6.flags = 0;
889
890 rds_info_copy(iter, &minfo6, sizeof(minfo6));
891}
892#endif