net/rds/recv.c

   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
  44 void 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_saddr = *saddr;
  51         inc->i_usercopy.rdma_cookie = 0;
  52         inc->i_usercopy.rx_tstamp = ktime_set(0, 0);
  53
  54         memset(inc->i_rx_lat_trace, 0, sizeof(inc->i_rx_lat_trace));
  55 }
  56 EXPORT_SYMBOL_GPL(rds_inc_init);
  57
  58 void rds_inc_path_init(struct rds_incoming *inc, struct rds_conn_path *cp,
  59                        struct in6_addr  *saddr)
  60 {
  61         refcount_set(&inc->i_refcount, 1);
  62         INIT_LIST_HEAD(&inc->i_item);
  63         inc->i_conn = cp->cp_conn;
  64         inc->i_conn_path = cp;
  65         inc->i_saddr = *saddr;
  66         inc->i_usercopy.rdma_cookie = 0;
  67         inc->i_usercopy.rx_tstamp = ktime_set(0, 0);
  68 }
  69 EXPORT_SYMBOL_GPL(rds_inc_path_init);
  70
  71 static void rds_inc_addref(struct rds_incoming *inc)
  72 {
  73         rdsdebug("addref inc %p ref %d\n", inc, refcount_read(&inc->i_refcount));
  74         refcount_inc(&inc->i_refcount);
  75 }
  76
  77 void rds_inc_put(struct rds_incoming *inc)
  78 {
  79         rdsdebug("put inc %p ref %d\n", inc, refcount_read(&inc->i_refcount));
  80         if (refcount_dec_and_test(&inc->i_refcount)) {
  81                 BUG_ON(!list_empty(&inc->i_item));
  82
  83                 inc->i_conn->c_trans->inc_free(inc);
  84         }
  85 }
  86 EXPORT_SYMBOL_GPL(rds_inc_put);
  87
  88 static void rds_recv_rcvbuf_delta(struct rds_sock *rs, struct sock *sk,
  89                                   struct rds_cong_map *map,
  90                                   int delta, __be16 port)
  91 {
  92         int now_congested;
  93
  94         if (delta == 0)
  95                 return;
  96
  97         rs->rs_rcv_bytes += delta;
  98         if (delta > 0)
  99                 rds_stats_add(s_recv_bytes_added_to_socket, delta);
 100         else
 101                 rds_stats_add(s_recv_bytes_removed_from_socket, -delta);
 102
 103         /* loop transport doesn't send/recv congestion updates */
 104         if (rs->rs_transport->t_type == RDS_TRANS_LOOP)
 105                 return;
 106
 107         now_congested = rs->rs_rcv_bytes > rds_sk_rcvbuf(rs);
 108
 109         rdsdebug("rs %p (%pI6c:%u) recv bytes %d buf %d "
 110           "now_cong %d delta %d\n",
 111           rs, &rs->rs_bound_addr,
 112           ntohs(rs->rs_bound_port), rs->rs_rcv_bytes,
 113           rds_sk_rcvbuf(rs), now_congested, delta);
 114
 115         /* wasn't -> am congested */
 116         if (!rs->rs_congested && now_congested) {
 117                 rs->rs_congested = 1;
 118                 rds_cong_set_bit(map, port);
 119                 rds_cong_queue_updates(map);
 120         }
 121         /* was -> aren't congested */
 122         /* Require more free space before reporting uncongested to prevent
 123            bouncing cong/uncong state too often */
 124         else if (rs->rs_congested && (rs->rs_rcv_bytes < (rds_sk_rcvbuf(rs)/2))) {
 125                 rs->rs_congested = 0;
 126                 rds_cong_clear_bit(map, port);
 127                 rds_cong_queue_updates(map);
 128         }
 129
 130         /* do nothing if no change in cong state */
 131 }
 132
 133 static void rds_conn_peer_gen_update(struct rds_connection *conn,
 134                                      u32 peer_gen_num)
 135 {
 136         int i;
 137         struct rds_message *rm, *tmp;
 138         unsigned long flags;
 139
 140         WARN_ON(conn->c_trans->t_type != RDS_TRANS_TCP);
 141         if (peer_gen_num != 0) {
 142                 if (conn->c_peer_gen_num != 0 &&
 143                     peer_gen_num != conn->c_peer_gen_num) {
 144                         for (i = 0; i < RDS_MPATH_WORKERS; i++) {
 145                                 struct rds_conn_path *cp;
 146
 147                                 cp = &conn->c_path[i];
 148                                 spin_lock_irqsave(&cp->cp_lock, flags);
 149                                 cp->cp_next_tx_seq = 1;
 150                                 cp->cp_next_rx_seq = 0;
 151                                 list_for_each_entry_safe(rm, tmp,
 152                                                          &cp->cp_retrans,
 153                                                          m_conn_item) {
 154                                         set_bit(RDS_MSG_FLUSH, &rm->m_flags);
 155                                 }
 156                                 spin_unlock_irqrestore(&cp->cp_lock, flags);
 157                         }
 158                 }
 159                 conn->c_peer_gen_num = peer_gen_num;
 160         }
 161 }
 162
 163 /*
 164  * Process all extension headers that come with this message.
 165  */
 166 static void rds_recv_incoming_exthdrs(struct rds_incoming *inc, struct rds_sock *rs)
 167 {
 168         struct rds_header *hdr = &inc->i_hdr;
 169         unsigned int pos = 0, type, len;
 170         union {
 171                 struct rds_ext_header_version version;
 172                 struct rds_ext_header_rdma rdma;
 173                 struct rds_ext_header_rdma_dest rdma_dest;
 174         } buffer;
 175
 176         while (1) {
 177                 len = sizeof(buffer);
 178                 type = rds_message_next_extension(hdr, &pos, &buffer, &len);
 179                 if (type == RDS_EXTHDR_NONE)
 180                         break;
 181                 /* Process extension header here */
 182                 switch (type) {
 183                 case RDS_EXTHDR_RDMA:
 184                         rds_rdma_unuse(rs, be32_to_cpu(buffer.rdma.h_rdma_rkey), 0);
 185                         break;
 186
 187                 case RDS_EXTHDR_RDMA_DEST:
 188                         /* We ignore the size for now. We could stash it
 189                          * somewhere and use it for error checking. */
 190                         inc->i_usercopy.rdma_cookie = rds_rdma_make_cookie(
 191                                         be32_to_cpu(buffer.rdma_dest.h_rdma_rkey),
 192                                         be32_to_cpu(buffer.rdma_dest.h_rdma_offset));
 193
 194                         break;
 195                 }
 196         }
 197 }
 198
 199 static void rds_recv_hs_exthdrs(struct rds_header *hdr,
 200                                 struct rds_connection *conn)
 201 {
 202         unsigned int pos = 0, type, len;
 203         union {
 204                 struct rds_ext_header_version version;
 205                 u16 rds_npaths;
 206                 u32 rds_gen_num;
 207         } buffer;
 208         u32 new_peer_gen_num = 0;
 209
 210         while (1) {
 211                 len = sizeof(buffer);
 212                 type = rds_message_next_extension(hdr, &pos, &buffer, &len);
 213                 if (type == RDS_EXTHDR_NONE)
 214                         break;
 215                 /* Process extension header here */
 216                 switch (type) {
 217                 case RDS_EXTHDR_NPATHS:
 218                         conn->c_npaths = min_t(int, RDS_MPATH_WORKERS,
 219                                                be16_to_cpu(buffer.rds_npaths));
 220                         break;
 221                 case RDS_EXTHDR_GEN_NUM:
 222                         new_peer_gen_num = be32_to_cpu(buffer.rds_gen_num);
 223                         break;
 224                 default:
 225                         pr_warn_ratelimited("ignoring unknown exthdr type "
 226                                              "0x%x\n", type);
 227                 }
 228         }
 229         /* if RDS_EXTHDR_NPATHS was not found, default to a single-path */
 230         conn->c_npaths = max_t(int, conn->c_npaths, 1);
 231         conn->c_ping_triggered = 0;
 232         rds_conn_peer_gen_update(conn, new_peer_gen_num);
 233 }
 234
 235 /* rds_start_mprds() will synchronously start multiple paths when appropriate.
 236  * The scheme is based on the following rules:
 237  *
 238  * 1. rds_sendmsg on first connect attempt sends the probe ping, with the
 239  *    sender's npaths (s_npaths)
 240  * 2. rcvr of probe-ping knows the mprds_paths = min(s_npaths, r_npaths). It
 241  *    sends back a probe-pong with r_npaths. After that, if rcvr is the
 242  *    smaller ip addr, it starts rds_conn_path_connect_if_down on all
 243  *    mprds_paths.
 244  * 3. sender gets woken up, and can move to rds_conn_path_connect_if_down.
 245  *    If it is the smaller ipaddr, rds_conn_path_connect_if_down can be
 246  *    called after reception of the probe-pong on all mprds_paths.
 247  *    Otherwise (sender of probe-ping is not the smaller ip addr): just call
 248  *    rds_conn_path_connect_if_down on the hashed path. (see rule 4)
 249  * 4. rds_connect_worker must only trigger a connection if laddr < faddr.
 250  * 5. sender may end up queuing the packet on the cp. will get sent out later.
 251  *    when connection is completed.
 252  */
 253 static void rds_start_mprds(struct rds_connection *conn)
 254 {
 255         int i;
 256         struct rds_conn_path *cp;
 257
 258         if (conn->c_npaths > 1 &&
 259             rds_addr_cmp(&conn->c_laddr, &conn->c_faddr) < 0) {
 260                 for (i = 0; i < conn->c_npaths; i++) {
 261                         cp = &conn->c_path[i];
 262                         rds_conn_path_connect_if_down(cp);
 263                 }
 264         }
 265 }
 266
 267 /*
 268  * The transport must make sure that this is serialized against other
 269  * rx and conn reset on this specific conn.
 270  *
 271  * We currently assert that only one fragmented message will be sent
 272  * down a connection at a time.  This lets us reassemble in the conn
 273  * instead of per-flow which means that we don't have to go digging through
 274  * flows to tear down partial reassembly progress on conn failure and
 275  * we save flow lookup and locking for each frag arrival.  It does mean
 276  * that small messages will wait behind large ones.  Fragmenting at all
 277  * is only to reduce the memory consumption of pre-posted buffers.
 278  *
 279  * The caller passes in saddr and daddr instead of us getting it from the
 280  * conn.  This lets loopback, who only has one conn for both directions,
 281  * tell us which roles the addrs in the conn are playing for this message.
 282  */
 283 void rds_recv_incoming(struct rds_connection *conn, struct in6_addr *saddr,
 284                        struct in6_addr *daddr,
 285                        struct rds_incoming *inc, gfp_t gfp)
 286 {
 287         struct rds_sock *rs = NULL;
 288         struct sock *sk;
 289         unsigned long flags;
 290         struct rds_conn_path *cp;
 291
 292         inc->i_conn = conn;
 293         inc->i_rx_jiffies = jiffies;
 294         if (conn->c_trans->t_mp_capable)
 295                 cp = inc->i_conn_path;
 296         else
 297                 cp = &conn->c_path[0];
 298
 299         rdsdebug("conn %p next %llu inc %p seq %llu len %u sport %u dport %u "
 300                  "flags 0x%x rx_jiffies %lu\n", conn,
 301                  (unsigned long long)cp->cp_next_rx_seq,
 302                  inc,
 303                  (unsigned long long)be64_to_cpu(inc->i_hdr.h_sequence),
 304                  be32_to_cpu(inc->i_hdr.h_len),
 305                  be16_to_cpu(inc->i_hdr.h_sport),
 306                  be16_to_cpu(inc->i_hdr.h_dport),
 307                  inc->i_hdr.h_flags,
 308                  inc->i_rx_jiffies);
 309
 310         /*
 311          * Sequence numbers should only increase.  Messages get their
 312          * sequence number as they're queued in a sending conn.  They
 313          * can be dropped, though, if the sending socket is closed before
 314          * they hit the wire.  So sequence numbers can skip forward
 315          * under normal operation.  They can also drop back in the conn
 316          * failover case as previously sent messages are resent down the
 317          * new instance of a conn.  We drop those, otherwise we have
 318          * to assume that the next valid seq does not come after a
 319          * hole in the fragment stream.
 320          *
 321          * The headers don't give us a way to realize if fragments of
 322          * a message have been dropped.  We assume that frags that arrive
 323          * to a flow are part of the current message on the flow that is
 324          * being reassembled.  This means that senders can't drop messages
 325          * from the sending conn until all their frags are sent.
 326          *
 327          * XXX we could spend more on the wire to get more robust failure
 328          * detection, arguably worth it to avoid data corruption.
 329          */
 330         if (be64_to_cpu(inc->i_hdr.h_sequence) < cp->cp_next_rx_seq &&
 331             (inc->i_hdr.h_flags & RDS_FLAG_RETRANSMITTED)) {
 332                 rds_stats_inc(s_recv_drop_old_seq);
 333                 goto out;
 334         }
 335         cp->cp_next_rx_seq = be64_to_cpu(inc->i_hdr.h_sequence) + 1;
 336
 337         if (rds_sysctl_ping_enable && inc->i_hdr.h_dport == 0) {
 338                 if (inc->i_hdr.h_sport == 0) {
 339                         rdsdebug("ignore ping with 0 sport from %pI6c\n",
 340                                  saddr);
 341                         goto out;
 342                 }
 343                 rds_stats_inc(s_recv_ping);
 344                 rds_send_pong(cp, inc->i_hdr.h_sport);
 345                 /* if this is a handshake ping, start multipath if necessary */
 346                 if (RDS_HS_PROBE(be16_to_cpu(inc->i_hdr.h_sport),
 347                                  be16_to_cpu(inc->i_hdr.h_dport))) {
 348                         rds_recv_hs_exthdrs(&inc->i_hdr, cp->cp_conn);
 349                         rds_start_mprds(cp->cp_conn);
 350                 }
 351                 goto out;
 352         }
 353
 354         if (be16_to_cpu(inc->i_hdr.h_dport) ==  RDS_FLAG_PROBE_PORT &&
 355             inc->i_hdr.h_sport == 0) {
 356                 rds_recv_hs_exthdrs(&inc->i_hdr, cp->cp_conn);
 357                 /* if this is a handshake pong, start multipath if necessary */
 358                 rds_start_mprds(cp->cp_conn);
 359                 wake_up(&cp->cp_conn->c_hs_waitq);
 360                 goto out;
 361         }
 362
 363         rs = rds_find_bound(daddr, inc->i_hdr.h_dport, conn->c_bound_if);
 364         if (!rs) {
 365                 rds_stats_inc(s_recv_drop_no_sock);
 366                 goto out;
 367         }
 368
 369         /* Process extension headers */
 370         rds_recv_incoming_exthdrs(inc, rs);
 371
 372         /* We can be racing with rds_release() which marks the socket dead. */
 373         sk = rds_rs_to_sk(rs);
 374
 375         /* serialize with rds_release -> sock_orphan */
 376         write_lock_irqsave(&rs->rs_recv_lock, flags);
 377         if (!sock_flag(sk, SOCK_DEAD)) {
 378                 rdsdebug("adding inc %p to rs %p's recv queue\n", inc, rs);
 379                 rds_stats_inc(s_recv_queued);
 380                 rds_recv_rcvbuf_delta(rs, sk, inc->i_conn->c_lcong,
 381                                       be32_to_cpu(inc->i_hdr.h_len),
 382                                       inc->i_hdr.h_dport);
 383                 if (sock_flag(sk, SOCK_RCVTSTAMP))
 384                         inc->i_usercopy.rx_tstamp = ktime_get_real();
 385                 rds_inc_addref(inc);
 386                 inc->i_rx_lat_trace[RDS_MSG_RX_END] = local_clock();
 387                 list_add_tail(&inc->i_item, &rs->rs_recv_queue);
 388                 __rds_wake_sk_sleep(sk);
 389         } else {
 390                 rds_stats_inc(s_recv_drop_dead_sock);
 391         }
 392         write_unlock_irqrestore(&rs->rs_recv_lock, flags);
 393
 394 out:
 395         if (rs)
 396                 rds_sock_put(rs);
 397 }
 398 EXPORT_SYMBOL_GPL(rds_recv_incoming);
 399
 400 /*
 401  * be very careful here.  This is being called as the condition in
 402  * wait_event_*() needs to cope with being called many times.
 403  */
 404 static int rds_next_incoming(struct rds_sock *rs, struct rds_incoming **inc)
 405 {
 406         unsigned long flags;
 407
 408         if (!*inc) {
 409                 read_lock_irqsave(&rs->rs_recv_lock, flags);
 410                 if (!list_empty(&rs->rs_recv_queue)) {
 411                         *inc = list_entry(rs->rs_recv_queue.next,
 412                                           struct rds_incoming,
 413                                           i_item);
 414                         rds_inc_addref(*inc);
 415                 }
 416                 read_unlock_irqrestore(&rs->rs_recv_lock, flags);
 417         }
 418
 419         return *inc != NULL;
 420 }
 421
 422 static int rds_still_queued(struct rds_sock *rs, struct rds_incoming *inc,
 423                             int drop)
 424 {
 425         struct sock *sk = rds_rs_to_sk(rs);
 426         int ret = 0;
 427         unsigned long flags;
 428         struct rds_incoming *to_drop = NULL;
 429
 430         write_lock_irqsave(&rs->rs_recv_lock, flags);
 431         if (!list_empty(&inc->i_item)) {
 432                 ret = 1;
 433                 if (drop) {
 434                         /* XXX make sure this i_conn is reliable */
 435                         rds_recv_rcvbuf_delta(rs, sk, inc->i_conn->c_lcong,
 436                                               -be32_to_cpu(inc->i_hdr.h_len),
 437                                               inc->i_hdr.h_dport);
 438                         list_del_init(&inc->i_item);
 439                         to_drop = inc;
 440                 }
 441         }
 442         write_unlock_irqrestore(&rs->rs_recv_lock, flags);
 443
 444         if (to_drop)
 445                 rds_inc_put(to_drop);
 446
 447         rdsdebug("inc %p rs %p still %d dropped %d\n", inc, rs, ret, drop);
 448         return ret;
 449 }
 450
 451 /*
 452  * Pull errors off the error queue.
 453  * If msghdr is NULL, we will just purge the error queue.
 454  */
 455 int rds_notify_queue_get(struct rds_sock *rs, struct msghdr *msghdr)
 456 {
 457         struct rds_notifier *notifier;
 458         struct rds_rdma_notify cmsg;
 459         unsigned int count = 0, max_messages = ~0U;
 460         unsigned long flags;
 461         LIST_HEAD(copy);
 462         int err = 0;
 463
 464         memset(&cmsg, 0, sizeof(cmsg)); /* fill holes with zero */
 465
 466         /* put_cmsg copies to user space and thus may sleep. We can't do this
 467          * with rs_lock held, so first grab as many notifications as we can stuff
 468          * in the user provided cmsg buffer. We don't try to copy more, to avoid
 469          * losing notifications - except when the buffer is so small that it wouldn't
 470          * even hold a single notification. Then we give him as much of this single
 471          * msg as we can squeeze in, and set MSG_CTRUNC.
 472          */
 473         if (msghdr) {
 474                 max_messages = msghdr->msg_controllen / CMSG_SPACE(sizeof(cmsg));
 475                 if (!max_messages)
 476                         max_messages = 1;
 477         }
 478
 479         spin_lock_irqsave(&rs->rs_lock, flags);
 480         while (!list_empty(&rs->rs_notify_queue) && count < max_messages) {
 481                 notifier = list_entry(rs->rs_notify_queue.next,
 482                                 struct rds_notifier, n_list);
 483                 list_move(&notifier->n_list, &copy);
 484                 count++;
 485         }
 486         spin_unlock_irqrestore(&rs->rs_lock, flags);
 487
 488         if (!count)
 489                 return 0;
 490
 491         while (!list_empty(&copy)) {
 492                 notifier = list_entry(copy.next, struct rds_notifier, n_list);
 493
 494                 if (msghdr) {
 495                         cmsg.user_token = notifier->n_user_token;
 496                         cmsg.status = notifier->n_status;
 497
 498                         err = put_cmsg(msghdr, SOL_RDS, RDS_CMSG_RDMA_STATUS,
 499                                        sizeof(cmsg), &cmsg);
 500                         if (err)
 501                                 break;
 502                 }
 503
 504                 list_del_init(&notifier->n_list);
 505                 kfree(notifier);
 506         }
 507
 508         /* If we bailed out because of an error in put_cmsg,
 509          * we may be left with one or more notifications that we
 510          * didn't process. Return them to the head of the list. */
 511         if (!list_empty(&copy)) {
 512                 spin_lock_irqsave(&rs->rs_lock, flags);
 513                 list_splice(&copy, &rs->rs_notify_queue);
 514                 spin_unlock_irqrestore(&rs->rs_lock, flags);
 515         }
 516
 517         return err;
 518 }
 519
 520 /*
 521  * Queue a congestion notification
 522  */
 523 static int rds_notify_cong(struct rds_sock *rs, struct msghdr *msghdr)
 524 {
 525         uint64_t notify = rs->rs_cong_notify;
 526         unsigned long flags;
 527         int err;
 528
 529         err = put_cmsg(msghdr, SOL_RDS, RDS_CMSG_CONG_UPDATE,
 530                         sizeof(notify), &notify);
 531         if (err)
 532                 return err;
 533
 534         spin_lock_irqsave(&rs->rs_lock, flags);
 535         rs->rs_cong_notify &= ~notify;
 536         spin_unlock_irqrestore(&rs->rs_lock, flags);
 537
 538         return 0;
 539 }
 540
 541 /*
 542  * Receive any control messages.
 543  */
 544 static int rds_cmsg_recv(struct rds_incoming *inc, struct msghdr *msg,
 545                          struct rds_sock *rs)
 546 {
 547         int ret = 0;
 548
 549         if (inc->i_usercopy.rdma_cookie) {
 550                 ret = put_cmsg(msg, SOL_RDS, RDS_CMSG_RDMA_DEST,
 551                                 sizeof(inc->i_usercopy.rdma_cookie),
 552                                 &inc->i_usercopy.rdma_cookie);
 553                 if (ret)
 554                         goto out;
 555         }
 556
 557         if ((inc->i_usercopy.rx_tstamp != 0) &&
 558             sock_flag(rds_rs_to_sk(rs), SOCK_RCVTSTAMP)) {
 559                 struct __kernel_old_timeval tv =
 560                         ns_to_kernel_old_timeval(inc->i_usercopy.rx_tstamp);
 561
 562                 if (!sock_flag(rds_rs_to_sk(rs), SOCK_TSTAMP_NEW)) {
 563                         ret = put_cmsg(msg, SOL_SOCKET, SO_TIMESTAMP_OLD,
 564                                        sizeof(tv), &tv);
 565                 } else {
 566                         struct __kernel_sock_timeval sk_tv;
 567
 568                         sk_tv.tv_sec = tv.tv_sec;
 569                         sk_tv.tv_usec = tv.tv_usec;
 570
 571                         ret = put_cmsg(msg, SOL_SOCKET, SO_TIMESTAMP_NEW,
 572                                        sizeof(sk_tv), &sk_tv);
 573                 }
 574
 575                 if (ret)
 576                         goto out;
 577         }
 578
 579         if (rs->rs_rx_traces) {
 580                 struct rds_cmsg_rx_trace t;
 581                 int i, j;
 582
 583                 memset(&t, 0, sizeof(t));
 584                 inc->i_rx_lat_trace[RDS_MSG_RX_CMSG] = local_clock();
 585                 t.rx_traces =  rs->rs_rx_traces;
 586                 for (i = 0; i < rs->rs_rx_traces; i++) {
 587                         j = rs->rs_rx_trace[i];
 588                         t.rx_trace_pos[i] = j;
 589                         t.rx_trace[i] = inc->i_rx_lat_trace[j + 1] -
 590                                           inc->i_rx_lat_trace[j];
 591                 }
 592
 593                 ret = put_cmsg(msg, SOL_RDS, RDS_CMSG_RXPATH_LATENCY,
 594                                sizeof(t), &t);
 595                 if (ret)
 596                         goto out;
 597         }
 598
 599 out:
 600         return ret;
 601 }
 602
 603 static bool rds_recvmsg_zcookie(struct rds_sock *rs, struct msghdr *msg)
 604 {
 605         struct rds_msg_zcopy_queue *q = &rs->rs_zcookie_queue;
 606         struct rds_msg_zcopy_info *info = NULL;
 607         struct rds_zcopy_cookies *done;
 608         unsigned long flags;
 609
 610         if (!msg->msg_control)
 611                 return false;
 612
 613         if (!sock_flag(rds_rs_to_sk(rs), SOCK_ZEROCOPY) ||
 614             msg->msg_controllen < CMSG_SPACE(sizeof(*done)))
 615                 return false;
 616
 617         spin_lock_irqsave(&q->lock, flags);
 618         if (!list_empty(&q->zcookie_head)) {
 619                 info = list_entry(q->zcookie_head.next,
 620                                   struct rds_msg_zcopy_info, rs_zcookie_next);
 621                 list_del(&info->rs_zcookie_next);
 622         }
 623         spin_unlock_irqrestore(&q->lock, flags);
 624         if (!info)
 625                 return false;
 626         done = &info->zcookies;
 627         if (put_cmsg(msg, SOL_RDS, RDS_CMSG_ZCOPY_COMPLETION, sizeof(*done),
 628                      done)) {
 629                 spin_lock_irqsave(&q->lock, flags);
 630                 list_add(&info->rs_zcookie_next, &q->zcookie_head);
 631                 spin_unlock_irqrestore(&q->lock, flags);
 632                 return false;
 633         }
 634         kfree(info);
 635         return true;
 636 }
 637
 638 int rds_recvmsg(struct socket *sock, struct msghdr *msg, size_t size,
 639                 int msg_flags)
 640 {
 641         struct sock *sk = sock->sk;
 642         struct rds_sock *rs = rds_sk_to_rs(sk);
 643         long timeo;
 644         int ret = 0, nonblock = msg_flags & MSG_DONTWAIT;
 645         DECLARE_SOCKADDR(struct sockaddr_in6 *, sin6, msg->msg_name);
 646         DECLARE_SOCKADDR(struct sockaddr_in *, sin, msg->msg_name);
 647         struct rds_incoming *inc = NULL;
 648
 649         /* udp_recvmsg()->sock_recvtimeo() gets away without locking too.. */
 650         timeo = sock_rcvtimeo(sk, nonblock);
 651
 652         rdsdebug("size %zu flags 0x%x timeo %ld\n", size, msg_flags, timeo);
 653
 654         if (msg_flags & MSG_OOB)
 655                 goto out;
 656         if (msg_flags & MSG_ERRQUEUE)
 657                 return sock_recv_errqueue(sk, msg, size, SOL_IP, IP_RECVERR);
 658
 659         while (1) {
 660                 /* If there are pending notifications, do those - and nothing else */
 661                 if (!list_empty(&rs->rs_notify_queue)) {
 662                         ret = rds_notify_queue_get(rs, msg);
 663                         break;
 664                 }
 665
 666                 if (rs->rs_cong_notify) {
 667                         ret = rds_notify_cong(rs, msg);
 668                         break;
 669                 }
 670
 671                 if (!rds_next_incoming(rs, &inc)) {
 672                         if (nonblock) {
 673                                 bool reaped = rds_recvmsg_zcookie(rs, msg);
 674
 675                                 ret = reaped ?  0 : -EAGAIN;
 676                                 break;
 677                         }
 678
 679                         timeo = wait_event_interruptible_timeout(*sk_sleep(sk),
 680                                         (!list_empty(&rs->rs_notify_queue) ||
 681                                          rs->rs_cong_notify ||
 682                                          rds_next_incoming(rs, &inc)), timeo);
 683                         rdsdebug("recvmsg woke inc %p timeo %ld\n", inc,
 684                                  timeo);
 685                         if (timeo > 0 || timeo == MAX_SCHEDULE_TIMEOUT)
 686                                 continue;
 687
 688                         ret = timeo;
 689                         if (ret == 0)
 690                                 ret = -ETIMEDOUT;
 691                         break;
 692                 }
 693
 694                 rdsdebug("copying inc %p from %pI6c:%u to user\n", inc,
 695                          &inc->i_conn->c_faddr,
 696                          ntohs(inc->i_hdr.h_sport));
 697                 ret = inc->i_conn->c_trans->inc_copy_to_user(inc, &msg->msg_iter);
 698                 if (ret < 0)
 699                         break;
 700
 701                 /*
 702                  * if the message we just copied isn't at the head of the
 703                  * recv queue then someone else raced us to return it, try
 704                  * to get the next message.
 705                  */
 706                 if (!rds_still_queued(rs, inc, !(msg_flags & MSG_PEEK))) {
 707                         rds_inc_put(inc);
 708                         inc = NULL;
 709                         rds_stats_inc(s_recv_deliver_raced);
 710                         iov_iter_revert(&msg->msg_iter, ret);
 711                         continue;
 712                 }
 713
 714                 if (ret < be32_to_cpu(inc->i_hdr.h_len)) {
 715                         if (msg_flags & MSG_TRUNC)
 716                                 ret = be32_to_cpu(inc->i_hdr.h_len);
 717                         msg->msg_flags |= MSG_TRUNC;
 718                 }
 719
 720                 if (rds_cmsg_recv(inc, msg, rs)) {
 721                         ret = -EFAULT;
 722                         break;
 723                 }
 724                 rds_recvmsg_zcookie(rs, msg);
 725
 726                 rds_stats_inc(s_recv_delivered);
 727
 728                 if (msg->msg_name) {
 729                         if (ipv6_addr_v4mapped(&inc->i_saddr)) {
 730                                 sin->sin_family = AF_INET;
 731                                 sin->sin_port = inc->i_hdr.h_sport;
 732                                 sin->sin_addr.s_addr =
 733                                     inc->i_saddr.s6_addr32[3];
 734                                 memset(sin->sin_zero, 0, sizeof(sin->sin_zero));
 735                                 msg->msg_namelen = sizeof(*sin);
 736                         } else {
 737                                 sin6->sin6_family = AF_INET6;
 738                                 sin6->sin6_port = inc->i_hdr.h_sport;
 739                                 sin6->sin6_addr = inc->i_saddr;
 740                                 sin6->sin6_flowinfo = 0;
 741                                 sin6->sin6_scope_id = rs->rs_bound_scope_id;
 742                                 msg->msg_namelen = sizeof(*sin6);
 743                         }
 744                 }
 745                 break;
 746         }
 747
 748         if (inc)
 749                 rds_inc_put(inc);
 750
 751 out:
 752         return ret;
 753 }
 754
 755 /*
 756  * The socket is being shut down and we're asked to drop messages that were
 757  * queued for recvmsg.  The caller has unbound the socket so the receive path
 758  * won't queue any more incoming fragments or messages on the socket.
 759  */
 760 void rds_clear_recv_queue(struct rds_sock *rs)
 761 {
 762         struct sock *sk = rds_rs_to_sk(rs);
 763         struct rds_incoming *inc, *tmp;
 764         unsigned long flags;
 765         LIST_HEAD(to_drop);
 766
 767         write_lock_irqsave(&rs->rs_recv_lock, flags);
 768         list_for_each_entry_safe(inc, tmp, &rs->rs_recv_queue, i_item) {
 769                 rds_recv_rcvbuf_delta(rs, sk, inc->i_conn->c_lcong,
 770                                       -be32_to_cpu(inc->i_hdr.h_len),
 771                                       inc->i_hdr.h_dport);
 772                 list_move(&inc->i_item, &to_drop);
 773         }
 774         write_unlock_irqrestore(&rs->rs_recv_lock, flags);
 775
 776         list_for_each_entry_safe(inc, tmp, &to_drop, i_item) {
 777                 list_del_init(&inc->i_item);
 778                 rds_inc_put(inc);
 779         }
 780 }
 781
 782 /*
 783  * inc->i_saddr isn't used here because it is only set in the receive
 784  * path.
 785  */
 786 void rds_inc_info_copy(struct rds_incoming *inc,
 787                        struct rds_info_iterator *iter,
 788                        __be32 saddr, __be32 daddr, int flip)
 789 {
 790         struct rds_info_message minfo;
 791
 792         minfo.seq = be64_to_cpu(inc->i_hdr.h_sequence);
 793         minfo.len = be32_to_cpu(inc->i_hdr.h_len);
 794         minfo.tos = inc->i_conn->c_tos;
 795
 796         if (flip) {
 797                 minfo.laddr = daddr;
 798                 minfo.faddr = saddr;
 799                 minfo.lport = inc->i_hdr.h_dport;
 800                 minfo.fport = inc->i_hdr.h_sport;
 801         } else {
 802                 minfo.laddr = saddr;
 803                 minfo.faddr = daddr;
 804                 minfo.lport = inc->i_hdr.h_sport;
 805                 minfo.fport = inc->i_hdr.h_dport;
 806         }
 807
 808         minfo.flags = 0;
 809
 810         rds_info_copy(iter, &minfo, sizeof(minfo));
 811 }
 812
 813 #if IS_ENABLED(CONFIG_IPV6)
 814 void rds6_inc_info_copy(struct rds_incoming *inc,
 815                         struct rds_info_iterator *iter,
 816                         struct in6_addr *saddr, struct in6_addr *daddr,
 817                         int flip)
 818 {
 819         struct rds6_info_message minfo6;
 820
 821         minfo6.seq = be64_to_cpu(inc->i_hdr.h_sequence);
 822         minfo6.len = be32_to_cpu(inc->i_hdr.h_len);
 823         minfo6.tos = inc->i_conn->c_tos;
 824
 825         if (flip) {
 826                 minfo6.laddr = *daddr;
 827                 minfo6.faddr = *saddr;
 828                 minfo6.lport = inc->i_hdr.h_dport;
 829                 minfo6.fport = inc->i_hdr.h_sport;
 830         } else {
 831                 minfo6.laddr = *saddr;
 832                 minfo6.faddr = *daddr;
 833                 minfo6.lport = inc->i_hdr.h_sport;
 834                 minfo6.fport = inc->i_hdr.h_dport;
 835         }
 836
 837         minfo6.flags = 0;
 838
 839         rds_info_copy(iter, &minfo6, sizeof(minfo6));
 840 }
 841 #endif