net/rds/send.c

   1 /*
   2  * Copyright (c) 2006 Oracle.  All rights reserved.
   3  *
   4  * This software is available to you under a choice of one of two
   5  * licenses.  You may choose to be licensed under the terms of the GNU
   6  * General Public License (GPL) Version 2, available from the file
   7  * COPYING in the main directory of this source tree, or the
   8  * OpenIB.org BSD license below:
   9  *
  10  *     Redistribution and use in source and binary forms, with or
  11  *     without modification, are permitted provided that the following
  12  *     conditions are met:
  13  *
  14  *      - Redistributions of source code must retain the above
  15  *        copyright notice, this list of conditions and the following
  16  *        disclaimer.
  17  *
  18  *      - Redistributions in binary form must reproduce the above
  19  *        copyright notice, this list of conditions and the following
  20  *        disclaimer in the documentation and/or other materials
  21  *        provided with the distribution.
  22  *
  23  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  24  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  25  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  26  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  27  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  28  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  29  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  30  * SOFTWARE.
  31  *
  32  */
  33 #include <linux/kernel.h>
  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
  42 #include "rds.h"
  43
  44 /* When transmitting messages in rds_send_xmit, we need to emerge from
  45  * time to time and briefly release the CPU. Otherwise the softlock watchdog
  46  * will kick our shin.
  47  * Also, it seems fairer to not let one busy connection stall all the
  48  * others.
  49  *
  50  * send_batch_count is the number of times we'll loop in send_xmit. Setting
  51  * it to 0 will restore the old behavior (where we looped until we had
  52  * drained the queue).
  53  */
  54 static int send_batch_count = 64;
  55 module_param(send_batch_count, int, 0444);
  56 MODULE_PARM_DESC(send_batch_count, " batch factor when working the send queue");
  57
  58 static void rds_send_remove_from_sock(struct list_head *messages, int status);
  59
  60 /*
  61  * Reset the send state.  Callers must ensure that this doesn't race with
  62  * rds_send_xmit().
  63  */
  64 void rds_send_reset(struct rds_connection *conn)
  65 {
  66         struct rds_message *rm, *tmp;
  67         unsigned long flags;
  68
  69         if (conn->c_xmit_rm) {
  70                 rm = conn->c_xmit_rm;
  71                 conn->c_xmit_rm = NULL;
  72                 /* Tell the user the RDMA op is no longer mapped by the
  73                  * transport. This isn't entirely true (it's flushed out
  74                  * independently) but as the connection is down, there's
  75                  * no ongoing RDMA to/from that memory */
  76                 rds_message_unmapped(rm);
  77                 rds_message_put(rm);
  78         }
  79
  80         conn->c_xmit_sg = 0;
  81         conn->c_xmit_hdr_off = 0;
  82         conn->c_xmit_data_off = 0;
  83         conn->c_xmit_atomic_sent = 0;
  84         conn->c_xmit_rdma_sent = 0;
  85         conn->c_xmit_data_sent = 0;
  86
  87         conn->c_map_queued = 0;
  88
  89         conn->c_unacked_packets = rds_sysctl_max_unacked_packets;
  90         conn->c_unacked_bytes = rds_sysctl_max_unacked_bytes;
  91
  92         /* Mark messages as retransmissions, and move them to the send q */
  93         spin_lock_irqsave(&conn->c_lock, flags);
  94         list_for_each_entry_safe(rm, tmp, &conn->c_retrans, m_conn_item) {
  95                 set_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags);
  96                 set_bit(RDS_MSG_RETRANSMITTED, &rm->m_flags);
  97         }
  98         list_splice_init(&conn->c_retrans, &conn->c_send_queue);
  99         spin_unlock_irqrestore(&conn->c_lock, flags);
 100 }
 101
 102 static int acquire_in_xmit(struct rds_connection *conn)
 103 {
 104         return test_and_set_bit(RDS_IN_XMIT, &conn->c_flags) == 0;
 105 }
 106
 107 static void release_in_xmit(struct rds_connection *conn)
 108 {
 109         clear_bit(RDS_IN_XMIT, &conn->c_flags);
 110         smp_mb__after_atomic();
 111         /*
 112          * We don't use wait_on_bit()/wake_up_bit() because our waking is in a
 113          * hot path and finding waiters is very rare.  We don't want to walk
 114          * the system-wide hashed waitqueue buckets in the fast path only to
 115          * almost never find waiters.
 116          */
 117         if (waitqueue_active(&conn->c_waitq))
 118                 wake_up_all(&conn->c_waitq);
 119 }
 120
 121 /*
 122  * We're making the conscious trade-off here to only send one message
 123  * down the connection at a time.
 124  *   Pro:
 125  *      - tx queueing is a simple fifo list
 126  *      - reassembly is optional and easily done by transports per conn
 127  *      - no per flow rx lookup at all, straight to the socket
 128  *      - less per-frag memory and wire overhead
 129  *   Con:
 130  *      - queued acks can be delayed behind large messages
 131  *   Depends:
 132  *      - small message latency is higher behind queued large messages
 133  *      - large message latency isn't starved by intervening small sends
 134  */
 135 int rds_send_xmit(struct rds_connection *conn)
 136 {
 137         struct rds_message *rm;
 138         unsigned long flags;
 139         unsigned int tmp;
 140         struct scatterlist *sg;
 141         int ret = 0;
 142         LIST_HEAD(to_be_dropped);
 143         int batch_count;
 144         unsigned long send_gen = 0;
 145
 146 restart:
 147         batch_count = 0;
 148
 149         /*
 150          * sendmsg calls here after having queued its message on the send
 151          * queue.  We only have one task feeding the connection at a time.  If
 152          * another thread is already feeding the queue then we back off.  This
 153          * avoids blocking the caller and trading per-connection data between
 154          * caches per message.
 155          */
 156         if (!acquire_in_xmit(conn)) {
 157                 rds_stats_inc(s_send_lock_contention);
 158                 ret = -ENOMEM;
 159                 goto out;
 160         }
 161
 162         /*
 163          * we record the send generation after doing the xmit acquire.
 164          * if someone else manages to jump in and do some work, we'll use
 165          * this to avoid a goto restart farther down.
 166          *
 167          * The acquire_in_xmit() check above ensures that only one
 168          * caller can increment c_send_gen at any time.
 169          */
 170         conn->c_send_gen++;
 171         send_gen = conn->c_send_gen;
 172
 173         /*
 174          * rds_conn_shutdown() sets the conn state and then tests RDS_IN_XMIT,
 175          * we do the opposite to avoid races.
 176          */
 177         if (!rds_conn_up(conn)) {
 178                 release_in_xmit(conn);
 179                 ret = 0;
 180                 goto out;
 181         }
 182
 183         if (conn->c_trans->xmit_prepare)
 184                 conn->c_trans->xmit_prepare(conn);
 185
 186         /*
 187          * spin trying to push headers and data down the connection until
 188          * the connection doesn't make forward progress.
 189          */
 190         while (1) {
 191
 192                 rm = conn->c_xmit_rm;
 193
 194                 /*
 195                  * If between sending messages, we can send a pending congestion
 196                  * map update.
 197                  */
 198                 if (!rm && test_and_clear_bit(0, &conn->c_map_queued)) {
 199                         rm = rds_cong_update_alloc(conn);
 200                         if (IS_ERR(rm)) {
 201                                 ret = PTR_ERR(rm);
 202                                 break;
 203                         }
 204                         rm->data.op_active = 1;
 205
 206                         conn->c_xmit_rm = rm;
 207                 }
 208
 209                 /*
 210                  * If not already working on one, grab the next message.
 211                  *
 212                  * c_xmit_rm holds a ref while we're sending this message down
 213                  * the connction.  We can use this ref while holding the
 214                  * send_sem.. rds_send_reset() is serialized with it.
 215                  */
 216                 if (!rm) {
 217                         unsigned int len;
 218
 219                         batch_count++;
 220
 221                         /* we want to process as big a batch as we can, but
 222                          * we also want to avoid softlockups.  If we've been
 223                          * through a lot of messages, lets back off and see
 224                          * if anyone else jumps in
 225                          */
 226                         if (batch_count >= 1024)
 227                                 goto over_batch;
 228
 229                         spin_lock_irqsave(&conn->c_lock, flags);
 230
 231                         if (!list_empty(&conn->c_send_queue)) {
 232                                 rm = list_entry(conn->c_send_queue.next,
 233                                                 struct rds_message,
 234                                                 m_conn_item);
 235                                 rds_message_addref(rm);
 236
 237                                 /*
 238                                  * Move the message from the send queue to the retransmit
 239                                  * list right away.
 240                                  */
 241                                 list_move_tail(&rm->m_conn_item, &conn->c_retrans);
 242                         }
 243
 244                         spin_unlock_irqrestore(&conn->c_lock, flags);
 245
 246                         if (!rm)
 247                                 break;
 248
 249                         /* Unfortunately, the way Infiniband deals with
 250                          * RDMA to a bad MR key is by moving the entire
 251                          * queue pair to error state. We cold possibly
 252                          * recover from that, but right now we drop the
 253                          * connection.
 254                          * Therefore, we never retransmit messages with RDMA ops.
 255                          */
 256                         if (rm->rdma.op_active &&
 257                             test_bit(RDS_MSG_RETRANSMITTED, &rm->m_flags)) {
 258                                 spin_lock_irqsave(&conn->c_lock, flags);
 259                                 if (test_and_clear_bit(RDS_MSG_ON_CONN, &rm->m_flags))
 260                                         list_move(&rm->m_conn_item, &to_be_dropped);
 261                                 spin_unlock_irqrestore(&conn->c_lock, flags);
 262                                 continue;
 263                         }
 264
 265                         /* Require an ACK every once in a while */
 266                         len = ntohl(rm->m_inc.i_hdr.h_len);
 267                         if (conn->c_unacked_packets == 0 ||
 268                             conn->c_unacked_bytes < len) {
 269                                 __set_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags);
 270
 271                                 conn->c_unacked_packets = rds_sysctl_max_unacked_packets;
 272                                 conn->c_unacked_bytes = rds_sysctl_max_unacked_bytes;
 273                                 rds_stats_inc(s_send_ack_required);
 274                         } else {
 275                                 conn->c_unacked_bytes -= len;
 276                                 conn->c_unacked_packets--;
 277                         }
 278
 279                         conn->c_xmit_rm = rm;
 280                 }
 281
 282                 /* The transport either sends the whole rdma or none of it */
 283                 if (rm->rdma.op_active && !conn->c_xmit_rdma_sent) {
 284                         rm->m_final_op = &rm->rdma;
 285                         /* The transport owns the mapped memory for now.
 286                          * You can't unmap it while it's on the send queue
 287                          */
 288                         set_bit(RDS_MSG_MAPPED, &rm->m_flags);
 289                         ret = conn->c_trans->xmit_rdma(conn, &rm->rdma);
 290                         if (ret) {
 291                                 clear_bit(RDS_MSG_MAPPED, &rm->m_flags);
 292                                 wake_up_interruptible(&rm->m_flush_wait);
 293                                 break;
 294                         }
 295                         conn->c_xmit_rdma_sent = 1;
 296
 297                 }
 298
 299                 if (rm->atomic.op_active && !conn->c_xmit_atomic_sent) {
 300                         rm->m_final_op = &rm->atomic;
 301                         /* The transport owns the mapped memory for now.
 302                          * You can't unmap it while it's on the send queue
 303                          */
 304                         set_bit(RDS_MSG_MAPPED, &rm->m_flags);
 305                         ret = conn->c_trans->xmit_atomic(conn, &rm->atomic);
 306                         if (ret) {
 307                                 clear_bit(RDS_MSG_MAPPED, &rm->m_flags);
 308                                 wake_up_interruptible(&rm->m_flush_wait);
 309                                 break;
 310                         }
 311                         conn->c_xmit_atomic_sent = 1;
 312
 313                 }
 314
 315                 /*
 316                  * A number of cases require an RDS header to be sent
 317                  * even if there is no data.
 318                  * We permit 0-byte sends; rds-ping depends on this.
 319                  * However, if there are exclusively attached silent ops,
 320                  * we skip the hdr/data send, to enable silent operation.
 321                  */
 322                 if (rm->data.op_nents == 0) {
 323                         int ops_present;
 324                         int all_ops_are_silent = 1;
 325
 326                         ops_present = (rm->atomic.op_active || rm->rdma.op_active);
 327                         if (rm->atomic.op_active && !rm->atomic.op_silent)
 328                                 all_ops_are_silent = 0;
 329                         if (rm->rdma.op_active && !rm->rdma.op_silent)
 330                                 all_ops_are_silent = 0;
 331
 332                         if (ops_present && all_ops_are_silent
 333                             && !rm->m_rdma_cookie)
 334                                 rm->data.op_active = 0;
 335                 }
 336
 337                 if (rm->data.op_active && !conn->c_xmit_data_sent) {
 338                         rm->m_final_op = &rm->data;
 339                         ret = conn->c_trans->xmit(conn, rm,
 340                                                   conn->c_xmit_hdr_off,
 341                                                   conn->c_xmit_sg,
 342                                                   conn->c_xmit_data_off);
 343                         if (ret <= 0)
 344                                 break;
 345
 346                         if (conn->c_xmit_hdr_off < sizeof(struct rds_header)) {
 347                                 tmp = min_t(int, ret,
 348                                             sizeof(struct rds_header) -
 349                                             conn->c_xmit_hdr_off);
 350                                 conn->c_xmit_hdr_off += tmp;
 351                                 ret -= tmp;
 352                         }
 353
 354                         sg = &rm->data.op_sg[conn->c_xmit_sg];
 355                         while (ret) {
 356                                 tmp = min_t(int, ret, sg->length -
 357                                                       conn->c_xmit_data_off);
 358                                 conn->c_xmit_data_off += tmp;
 359                                 ret -= tmp;
 360                                 if (conn->c_xmit_data_off == sg->length) {
 361                                         conn->c_xmit_data_off = 0;
 362                                         sg++;
 363                                         conn->c_xmit_sg++;
 364                                         BUG_ON(ret != 0 &&
 365                                                conn->c_xmit_sg == rm->data.op_nents);
 366                                 }
 367                         }
 368
 369                         if (conn->c_xmit_hdr_off == sizeof(struct rds_header) &&
 370                             (conn->c_xmit_sg == rm->data.op_nents))
 371                                 conn->c_xmit_data_sent = 1;
 372                 }
 373
 374                 /*
 375                  * A rm will only take multiple times through this loop
 376                  * if there is a data op. Thus, if the data is sent (or there was
 377                  * none), then we're done with the rm.
 378                  */
 379                 if (!rm->data.op_active || conn->c_xmit_data_sent) {
 380                         conn->c_xmit_rm = NULL;
 381                         conn->c_xmit_sg = 0;
 382                         conn->c_xmit_hdr_off = 0;
 383                         conn->c_xmit_data_off = 0;
 384                         conn->c_xmit_rdma_sent = 0;
 385                         conn->c_xmit_atomic_sent = 0;
 386                         conn->c_xmit_data_sent = 0;
 387
 388                         rds_message_put(rm);
 389                 }
 390         }
 391
 392 over_batch:
 393         if (conn->c_trans->xmit_complete)
 394                 conn->c_trans->xmit_complete(conn);
 395         release_in_xmit(conn);
 396
 397         /* Nuke any messages we decided not to retransmit. */
 398         if (!list_empty(&to_be_dropped)) {
 399                 /* irqs on here, so we can put(), unlike above */
 400                 list_for_each_entry(rm, &to_be_dropped, m_conn_item)
 401                         rds_message_put(rm);
 402                 rds_send_remove_from_sock(&to_be_dropped, RDS_RDMA_DROPPED);
 403         }
 404
 405         /*
 406          * Other senders can queue a message after we last test the send queue
 407          * but before we clear RDS_IN_XMIT.  In that case they'd back off and
 408          * not try and send their newly queued message.  We need to check the
 409          * send queue after having cleared RDS_IN_XMIT so that their message
 410          * doesn't get stuck on the send queue.
 411          *
 412          * If the transport cannot continue (i.e ret != 0), then it must
 413          * call us when more room is available, such as from the tx
 414          * completion handler.
 415          *
 416          * We have an extra generation check here so that if someone manages
 417          * to jump in after our release_in_xmit, we'll see that they have done
 418          * some work and we will skip our goto
 419          */
 420         if (ret == 0) {
 421                 smp_mb();
 422                 if ((test_bit(0, &conn->c_map_queued) ||
 423                      !list_empty(&conn->c_send_queue)) &&
 424                     send_gen == conn->c_send_gen) {
 425                         rds_stats_inc(s_send_lock_queue_raced);
 426                         goto restart;
 427                 }
 428         }
 429 out:
 430         return ret;
 431 }
 432
 433 static void rds_send_sndbuf_remove(struct rds_sock *rs, struct rds_message *rm)
 434 {
 435         u32 len = be32_to_cpu(rm->m_inc.i_hdr.h_len);
 436
 437         assert_spin_locked(&rs->rs_lock);
 438
 439         BUG_ON(rs->rs_snd_bytes < len);
 440         rs->rs_snd_bytes -= len;
 441
 442         if (rs->rs_snd_bytes == 0)
 443                 rds_stats_inc(s_send_queue_empty);
 444 }
 445
 446 static inline int rds_send_is_acked(struct rds_message *rm, u64 ack,
 447                                     is_acked_func is_acked)
 448 {
 449         if (is_acked)
 450                 return is_acked(rm, ack);
 451         return be64_to_cpu(rm->m_inc.i_hdr.h_sequence) <= ack;
 452 }
 453
 454 /*
 455  * This is pretty similar to what happens below in the ACK
 456  * handling code - except that we call here as soon as we get
 457  * the IB send completion on the RDMA op and the accompanying
 458  * message.
 459  */
 460 void rds_rdma_send_complete(struct rds_message *rm, int status)
 461 {
 462         struct rds_sock *rs = NULL;
 463         struct rm_rdma_op *ro;
 464         struct rds_notifier *notifier;
 465         unsigned long flags;
 466
 467         spin_lock_irqsave(&rm->m_rs_lock, flags);
 468
 469         ro = &rm->rdma;
 470         if (test_bit(RDS_MSG_ON_SOCK, &rm->m_flags) &&
 471             ro->op_active && ro->op_notify && ro->op_notifier) {
 472                 notifier = ro->op_notifier;
 473                 rs = rm->m_rs;
 474                 sock_hold(rds_rs_to_sk(rs));
 475
 476                 notifier->n_status = status;
 477                 spin_lock(&rs->rs_lock);
 478                 list_add_tail(&notifier->n_list, &rs->rs_notify_queue);
 479                 spin_unlock(&rs->rs_lock);
 480
 481                 ro->op_notifier = NULL;
 482         }
 483
 484         spin_unlock_irqrestore(&rm->m_rs_lock, flags);
 485
 486         if (rs) {
 487                 rds_wake_sk_sleep(rs);
 488                 sock_put(rds_rs_to_sk(rs));
 489         }
 490 }
 491 EXPORT_SYMBOL_GPL(rds_rdma_send_complete);
 492
 493 /*
 494  * Just like above, except looks at atomic op
 495  */
 496 void rds_atomic_send_complete(struct rds_message *rm, int status)
 497 {
 498         struct rds_sock *rs = NULL;
 499         struct rm_atomic_op *ao;
 500         struct rds_notifier *notifier;
 501         unsigned long flags;
 502
 503         spin_lock_irqsave(&rm->m_rs_lock, flags);
 504
 505         ao = &rm->atomic;
 506         if (test_bit(RDS_MSG_ON_SOCK, &rm->m_flags)
 507             && ao->op_active && ao->op_notify && ao->op_notifier) {
 508                 notifier = ao->op_notifier;
 509                 rs = rm->m_rs;
 510                 sock_hold(rds_rs_to_sk(rs));
 511
 512                 notifier->n_status = status;
 513                 spin_lock(&rs->rs_lock);
 514                 list_add_tail(&notifier->n_list, &rs->rs_notify_queue);
 515                 spin_unlock(&rs->rs_lock);
 516
 517                 ao->op_notifier = NULL;
 518         }
 519
 520         spin_unlock_irqrestore(&rm->m_rs_lock, flags);
 521
 522         if (rs) {
 523                 rds_wake_sk_sleep(rs);
 524                 sock_put(rds_rs_to_sk(rs));
 525         }
 526 }
 527 EXPORT_SYMBOL_GPL(rds_atomic_send_complete);
 528
 529 /*
 530  * This is the same as rds_rdma_send_complete except we
 531  * don't do any locking - we have all the ingredients (message,
 532  * socket, socket lock) and can just move the notifier.
 533  */
 534 static inline void
 535 __rds_send_complete(struct rds_sock *rs, struct rds_message *rm, int status)
 536 {
 537         struct rm_rdma_op *ro;
 538         struct rm_atomic_op *ao;
 539
 540         ro = &rm->rdma;
 541         if (ro->op_active && ro->op_notify && ro->op_notifier) {
 542                 ro->op_notifier->n_status = status;
 543                 list_add_tail(&ro->op_notifier->n_list, &rs->rs_notify_queue);
 544                 ro->op_notifier = NULL;
 545         }
 546
 547         ao = &rm->atomic;
 548         if (ao->op_active && ao->op_notify && ao->op_notifier) {
 549                 ao->op_notifier->n_status = status;
 550                 list_add_tail(&ao->op_notifier->n_list, &rs->rs_notify_queue);
 551                 ao->op_notifier = NULL;
 552         }
 553
 554         /* No need to wake the app - caller does this */
 555 }
 556
 557 /*
 558  * This is called from the IB send completion when we detect
 559  * a RDMA operation that failed with remote access error.
 560  * So speed is not an issue here.
 561  */
 562 struct rds_message *rds_send_get_message(struct rds_connection *conn,
 563                                          struct rm_rdma_op *op)
 564 {
 565         struct rds_message *rm, *tmp, *found = NULL;
 566         unsigned long flags;
 567
 568         spin_lock_irqsave(&conn->c_lock, flags);
 569
 570         list_for_each_entry_safe(rm, tmp, &conn->c_retrans, m_conn_item) {
 571                 if (&rm->rdma == op) {
 572                         atomic_inc(&rm->m_refcount);
 573                         found = rm;
 574                         goto out;
 575                 }
 576         }
 577
 578         list_for_each_entry_safe(rm, tmp, &conn->c_send_queue, m_conn_item) {
 579                 if (&rm->rdma == op) {
 580                         atomic_inc(&rm->m_refcount);
 581                         found = rm;
 582                         break;
 583                 }
 584         }
 585
 586 out:
 587         spin_unlock_irqrestore(&conn->c_lock, flags);
 588
 589         return found;
 590 }
 591 EXPORT_SYMBOL_GPL(rds_send_get_message);
 592
 593 /*
 594  * This removes messages from the socket's list if they're on it.  The list
 595  * argument must be private to the caller, we must be able to modify it
 596  * without locks.  The messages must have a reference held for their
 597  * position on the list.  This function will drop that reference after
 598  * removing the messages from the 'messages' list regardless of if it found
 599  * the messages on the socket list or not.
 600  */
 601 static void rds_send_remove_from_sock(struct list_head *messages, int status)
 602 {
 603         unsigned long flags;
 604         struct rds_sock *rs = NULL;
 605         struct rds_message *rm;
 606
 607         while (!list_empty(messages)) {
 608                 int was_on_sock = 0;
 609
 610                 rm = list_entry(messages->next, struct rds_message,
 611                                 m_conn_item);
 612                 list_del_init(&rm->m_conn_item);
 613
 614                 /*
 615                  * If we see this flag cleared then we're *sure* that someone
 616                  * else beat us to removing it from the sock.  If we race
 617                  * with their flag update we'll get the lock and then really
 618                  * see that the flag has been cleared.
 619                  *
 620                  * The message spinlock makes sure nobody clears rm->m_rs
 621                  * while we're messing with it. It does not prevent the
 622                  * message from being removed from the socket, though.
 623                  */
 624                 spin_lock_irqsave(&rm->m_rs_lock, flags);
 625                 if (!test_bit(RDS_MSG_ON_SOCK, &rm->m_flags))
 626                         goto unlock_and_drop;
 627
 628                 if (rs != rm->m_rs) {
 629                         if (rs) {
 630                                 rds_wake_sk_sleep(rs);
 631                                 sock_put(rds_rs_to_sk(rs));
 632                         }
 633                         rs = rm->m_rs;
 634                         if (rs)
 635                                 sock_hold(rds_rs_to_sk(rs));
 636                 }
 637                 if (!rs)
 638                         goto unlock_and_drop;
 639                 spin_lock(&rs->rs_lock);
 640
 641                 if (test_and_clear_bit(RDS_MSG_ON_SOCK, &rm->m_flags)) {
 642                         struct rm_rdma_op *ro = &rm->rdma;
 643                         struct rds_notifier *notifier;
 644
 645                         list_del_init(&rm->m_sock_item);
 646                         rds_send_sndbuf_remove(rs, rm);
 647
 648                         if (ro->op_active && ro->op_notifier &&
 649                                (ro->op_notify || (ro->op_recverr && status))) {
 650                                 notifier = ro->op_notifier;
 651                                 list_add_tail(&notifier->n_list,
 652                                                 &rs->rs_notify_queue);
 653                                 if (!notifier->n_status)
 654                                         notifier->n_status = status;
 655                                 rm->rdma.op_notifier = NULL;
 656                         }
 657                         was_on_sock = 1;
 658                         rm->m_rs = NULL;
 659                 }
 660                 spin_unlock(&rs->rs_lock);
 661
 662 unlock_and_drop:
 663                 spin_unlock_irqrestore(&rm->m_rs_lock, flags);
 664                 rds_message_put(rm);
 665                 if (was_on_sock)
 666                         rds_message_put(rm);
 667         }
 668
 669         if (rs) {
 670                 rds_wake_sk_sleep(rs);
 671                 sock_put(rds_rs_to_sk(rs));
 672         }
 673 }
 674
 675 /*
 676  * Transports call here when they've determined that the receiver queued
 677  * messages up to, and including, the given sequence number.  Messages are
 678  * moved to the retrans queue when rds_send_xmit picks them off the send
 679  * queue. This means that in the TCP case, the message may not have been
 680  * assigned the m_ack_seq yet - but that's fine as long as tcp_is_acked
 681  * checks the RDS_MSG_HAS_ACK_SEQ bit.
 682  */
 683 void rds_send_drop_acked(struct rds_connection *conn, u64 ack,
 684                          is_acked_func is_acked)
 685 {
 686         struct rds_message *rm, *tmp;
 687         unsigned long flags;
 688         LIST_HEAD(list);
 689
 690         spin_lock_irqsave(&conn->c_lock, flags);
 691
 692         list_for_each_entry_safe(rm, tmp, &conn->c_retrans, m_conn_item) {
 693                 if (!rds_send_is_acked(rm, ack, is_acked))
 694                         break;
 695
 696                 list_move(&rm->m_conn_item, &list);
 697                 clear_bit(RDS_MSG_ON_CONN, &rm->m_flags);
 698         }
 699
 700         /* order flag updates with spin locks */
 701         if (!list_empty(&list))
 702                 smp_mb__after_atomic();
 703
 704         spin_unlock_irqrestore(&conn->c_lock, flags);
 705
 706         /* now remove the messages from the sock list as needed */
 707         rds_send_remove_from_sock(&list, RDS_RDMA_SUCCESS);
 708 }
 709 EXPORT_SYMBOL_GPL(rds_send_drop_acked);
 710
 711 void rds_send_drop_to(struct rds_sock *rs, struct sockaddr_in *dest)
 712 {
 713         struct rds_message *rm, *tmp;
 714         struct rds_connection *conn;
 715         unsigned long flags;
 716         LIST_HEAD(list);
 717
 718         /* get all the messages we're dropping under the rs lock */
 719         spin_lock_irqsave(&rs->rs_lock, flags);
 720
 721         list_for_each_entry_safe(rm, tmp, &rs->rs_send_queue, m_sock_item) {
 722                 if (dest && (dest->sin_addr.s_addr != rm->m_daddr ||
 723                              dest->sin_port != rm->m_inc.i_hdr.h_dport))
 724                         continue;
 725
 726                 list_move(&rm->m_sock_item, &list);
 727                 rds_send_sndbuf_remove(rs, rm);
 728                 clear_bit(RDS_MSG_ON_SOCK, &rm->m_flags);
 729         }
 730
 731         /* order flag updates with the rs lock */
 732         smp_mb__after_atomic();
 733
 734         spin_unlock_irqrestore(&rs->rs_lock, flags);
 735
 736         if (list_empty(&list))
 737                 return;
 738
 739         /* Remove the messages from the conn */
 740         list_for_each_entry(rm, &list, m_sock_item) {
 741
 742                 conn = rm->m_inc.i_conn;
 743
 744                 spin_lock_irqsave(&conn->c_lock, flags);
 745                 /*
 746                  * Maybe someone else beat us to removing rm from the conn.
 747                  * If we race with their flag update we'll get the lock and
 748                  * then really see that the flag has been cleared.
 749                  */
 750                 if (!test_and_clear_bit(RDS_MSG_ON_CONN, &rm->m_flags)) {
 751                         spin_unlock_irqrestore(&conn->c_lock, flags);
 752                         spin_lock_irqsave(&rm->m_rs_lock, flags);
 753                         rm->m_rs = NULL;
 754                         spin_unlock_irqrestore(&rm->m_rs_lock, flags);
 755                         continue;
 756                 }
 757                 list_del_init(&rm->m_conn_item);
 758                 spin_unlock_irqrestore(&conn->c_lock, flags);
 759
 760                 /*
 761                  * Couldn't grab m_rs_lock in top loop (lock ordering),
 762                  * but we can now.
 763                  */
 764                 spin_lock_irqsave(&rm->m_rs_lock, flags);
 765
 766                 spin_lock(&rs->rs_lock);
 767                 __rds_send_complete(rs, rm, RDS_RDMA_CANCELED);
 768                 spin_unlock(&rs->rs_lock);
 769
 770                 rm->m_rs = NULL;
 771                 spin_unlock_irqrestore(&rm->m_rs_lock, flags);
 772
 773                 rds_message_put(rm);
 774         }
 775
 776         rds_wake_sk_sleep(rs);
 777
 778         while (!list_empty(&list)) {
 779                 rm = list_entry(list.next, struct rds_message, m_sock_item);
 780                 list_del_init(&rm->m_sock_item);
 781                 rds_message_wait(rm);
 782
 783                 /* just in case the code above skipped this message
 784                  * because RDS_MSG_ON_CONN wasn't set, run it again here
 785                  * taking m_rs_lock is the only thing that keeps us
 786                  * from racing with ack processing.
 787                  */
 788                 spin_lock_irqsave(&rm->m_rs_lock, flags);
 789
 790                 spin_lock(&rs->rs_lock);
 791                 __rds_send_complete(rs, rm, RDS_RDMA_CANCELED);
 792                 spin_unlock(&rs->rs_lock);
 793
 794                 rm->m_rs = NULL;
 795                 spin_unlock_irqrestore(&rm->m_rs_lock, flags);
 796
 797                 rds_message_put(rm);
 798         }
 799 }
 800
 801 /*
 802  * we only want this to fire once so we use the callers 'queued'.  It's
 803  * possible that another thread can race with us and remove the
 804  * message from the flow with RDS_CANCEL_SENT_TO.
 805  */
 806 static int rds_send_queue_rm(struct rds_sock *rs, struct rds_connection *conn,
 807                              struct rds_message *rm, __be16 sport,
 808                              __be16 dport, int *queued)
 809 {
 810         unsigned long flags;
 811         u32 len;
 812
 813         if (*queued)
 814                 goto out;
 815
 816         len = be32_to_cpu(rm->m_inc.i_hdr.h_len);
 817
 818         /* this is the only place which holds both the socket's rs_lock
 819          * and the connection's c_lock */
 820         spin_lock_irqsave(&rs->rs_lock, flags);
 821
 822         /*
 823          * If there is a little space in sndbuf, we don't queue anything,
 824          * and userspace gets -EAGAIN. But poll() indicates there's send
 825          * room. This can lead to bad behavior (spinning) if snd_bytes isn't
 826          * freed up by incoming acks. So we check the *old* value of
 827          * rs_snd_bytes here to allow the last msg to exceed the buffer,
 828          * and poll() now knows no more data can be sent.
 829          */
 830         if (rs->rs_snd_bytes < rds_sk_sndbuf(rs)) {
 831                 rs->rs_snd_bytes += len;
 832
 833                 /* let recv side know we are close to send space exhaustion.
 834                  * This is probably not the optimal way to do it, as this
 835                  * means we set the flag on *all* messages as soon as our
 836                  * throughput hits a certain threshold.
 837                  */
 838                 if (rs->rs_snd_bytes >= rds_sk_sndbuf(rs) / 2)
 839                         __set_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags);
 840
 841                 list_add_tail(&rm->m_sock_item, &rs->rs_send_queue);
 842                 set_bit(RDS_MSG_ON_SOCK, &rm->m_flags);
 843                 rds_message_addref(rm);
 844                 rm->m_rs = rs;
 845
 846                 /* The code ordering is a little weird, but we're
 847                    trying to minimize the time we hold c_lock */
 848                 rds_message_populate_header(&rm->m_inc.i_hdr, sport, dport, 0);
 849                 rm->m_inc.i_conn = conn;
 850                 rds_message_addref(rm);
 851
 852                 spin_lock(&conn->c_lock);
 853                 rm->m_inc.i_hdr.h_sequence = cpu_to_be64(conn->c_next_tx_seq++);
 854                 list_add_tail(&rm->m_conn_item, &conn->c_send_queue);
 855                 set_bit(RDS_MSG_ON_CONN, &rm->m_flags);
 856                 spin_unlock(&conn->c_lock);
 857
 858                 rdsdebug("queued msg %p len %d, rs %p bytes %d seq %llu\n",
 859                          rm, len, rs, rs->rs_snd_bytes,
 860                          (unsigned long long)be64_to_cpu(rm->m_inc.i_hdr.h_sequence));
 861
 862                 *queued = 1;
 863         }
 864
 865         spin_unlock_irqrestore(&rs->rs_lock, flags);
 866 out:
 867         return *queued;
 868 }
 869
 870 /*
 871  * rds_message is getting to be quite complicated, and we'd like to allocate
 872  * it all in one go. This figures out how big it needs to be up front.
 873  */
 874 static int rds_rm_size(struct msghdr *msg, int data_len)
 875 {
 876         struct cmsghdr *cmsg;
 877         int size = 0;
 878         int cmsg_groups = 0;
 879         int retval;
 880
 881         for_each_cmsghdr(cmsg, msg) {
 882                 if (!CMSG_OK(msg, cmsg))
 883                         return -EINVAL;
 884
 885                 if (cmsg->cmsg_level != SOL_RDS)
 886                         continue;
 887
 888                 switch (cmsg->cmsg_type) {
 889                 case RDS_CMSG_RDMA_ARGS:
 890                         cmsg_groups |= 1;
 891                         retval = rds_rdma_extra_size(CMSG_DATA(cmsg));
 892                         if (retval < 0)
 893                                 return retval;
 894                         size += retval;
 895
 896                         break;
 897
 898                 case RDS_CMSG_RDMA_DEST:
 899                 case RDS_CMSG_RDMA_MAP:
 900                         cmsg_groups |= 2;
 901                         /* these are valid but do no add any size */
 902                         break;
 903
 904                 case RDS_CMSG_ATOMIC_CSWP:
 905                 case RDS_CMSG_ATOMIC_FADD:
 906                 case RDS_CMSG_MASKED_ATOMIC_CSWP:
 907                 case RDS_CMSG_MASKED_ATOMIC_FADD:
 908                         cmsg_groups |= 1;
 909                         size += sizeof(struct scatterlist);
 910                         break;
 911
 912                 default:
 913                         return -EINVAL;
 914                 }
 915
 916         }
 917
 918         size += ceil(data_len, PAGE_SIZE) * sizeof(struct scatterlist);
 919
 920         /* Ensure (DEST, MAP) are never used with (ARGS, ATOMIC) */
 921         if (cmsg_groups == 3)
 922                 return -EINVAL;
 923
 924         return size;
 925 }
 926
 927 static int rds_cmsg_send(struct rds_sock *rs, struct rds_message *rm,
 928                          struct msghdr *msg, int *allocated_mr)
 929 {
 930         struct cmsghdr *cmsg;
 931         int ret = 0;
 932
 933         for_each_cmsghdr(cmsg, msg) {
 934                 if (!CMSG_OK(msg, cmsg))
 935                         return -EINVAL;
 936
 937                 if (cmsg->cmsg_level != SOL_RDS)
 938                         continue;
 939
 940                 /* As a side effect, RDMA_DEST and RDMA_MAP will set
 941                  * rm->rdma.m_rdma_cookie and rm->rdma.m_rdma_mr.
 942                  */
 943                 switch (cmsg->cmsg_type) {
 944                 case RDS_CMSG_RDMA_ARGS:
 945                         ret = rds_cmsg_rdma_args(rs, rm, cmsg);
 946                         break;
 947
 948                 case RDS_CMSG_RDMA_DEST:
 949                         ret = rds_cmsg_rdma_dest(rs, rm, cmsg);
 950                         break;
 951
 952                 case RDS_CMSG_RDMA_MAP:
 953                         ret = rds_cmsg_rdma_map(rs, rm, cmsg);
 954                         if (!ret)
 955                                 *allocated_mr = 1;
 956                         break;
 957                 case RDS_CMSG_ATOMIC_CSWP:
 958                 case RDS_CMSG_ATOMIC_FADD:
 959                 case RDS_CMSG_MASKED_ATOMIC_CSWP:
 960                 case RDS_CMSG_MASKED_ATOMIC_FADD:
 961                         ret = rds_cmsg_atomic(rs, rm, cmsg);
 962                         break;
 963
 964                 default:
 965                         return -EINVAL;
 966                 }
 967
 968                 if (ret)
 969                         break;
 970         }
 971
 972         return ret;
 973 }
 974
 975 int rds_sendmsg(struct socket *sock, struct msghdr *msg, size_t payload_len)
 976 {
 977         struct sock *sk = sock->sk;
 978         struct rds_sock *rs = rds_sk_to_rs(sk);
 979         DECLARE_SOCKADDR(struct sockaddr_in *, usin, msg->msg_name);
 980         __be32 daddr;
 981         __be16 dport;
 982         struct rds_message *rm = NULL;
 983         struct rds_connection *conn;
 984         int ret = 0;
 985         int queued = 0, allocated_mr = 0;
 986         int nonblock = msg->msg_flags & MSG_DONTWAIT;
 987         long timeo = sock_sndtimeo(sk, nonblock);
 988
 989         /* Mirror Linux UDP mirror of BSD error message compatibility */
 990         /* XXX: Perhaps MSG_MORE someday */
 991         if (msg->msg_flags & ~(MSG_DONTWAIT | MSG_CMSG_COMPAT)) {
 992                 ret = -EOPNOTSUPP;
 993                 goto out;
 994         }
 995
 996         if (msg->msg_namelen) {
 997                 /* XXX fail non-unicast destination IPs? */
 998                 if (msg->msg_namelen < sizeof(*usin) || usin->sin_family != AF_INET) {
 999                         ret = -EINVAL;
1000                         goto out;
1001                 }
1002                 daddr = usin->sin_addr.s_addr;
1003                 dport = usin->sin_port;
1004         } else {
1005                 /* We only care about consistency with ->connect() */
1006                 lock_sock(sk);
1007                 daddr = rs->rs_conn_addr;
1008                 dport = rs->rs_conn_port;
1009                 release_sock(sk);
1010         }
1011
1012         /* racing with another thread binding seems ok here */
1013         if (daddr == 0 || rs->rs_bound_addr == 0) {
1014                 ret = -ENOTCONN; /* XXX not a great errno */
1015                 goto out;
1016         }
1017
1018         if (payload_len > rds_sk_sndbuf(rs)) {
1019                 ret = -EMSGSIZE;
1020                 goto out;
1021         }
1022
1023         /* size of rm including all sgs */
1024         ret = rds_rm_size(msg, payload_len);
1025         if (ret < 0)
1026                 goto out;
1027
1028         rm = rds_message_alloc(ret, GFP_KERNEL);
1029         if (!rm) {
1030                 ret = -ENOMEM;
1031                 goto out;
1032         }
1033
1034         /* Attach data to the rm */
1035         if (payload_len) {
1036                 rm->data.op_sg = rds_message_alloc_sgs(rm, ceil(payload_len, PAGE_SIZE));
1037                 if (!rm->data.op_sg) {
1038                         ret = -ENOMEM;
1039                         goto out;
1040                 }
1041                 ret = rds_message_copy_from_user(rm, &msg->msg_iter);
1042                 if (ret)
1043                         goto out;
1044         }
1045         rm->data.op_active = 1;
1046
1047         rm->m_daddr = daddr;
1048
1049         /* rds_conn_create has a spinlock that runs with IRQ off.
1050          * Caching the conn in the socket helps a lot. */
1051         if (rs->rs_conn && rs->rs_conn->c_faddr == daddr)
1052                 conn = rs->rs_conn;
1053         else {
1054                 conn = rds_conn_create_outgoing(sock_net(sock->sk),
1055                                                 rs->rs_bound_addr, daddr,
1056                                         rs->rs_transport,
1057                                         sock->sk->sk_allocation);
1058                 if (IS_ERR(conn)) {
1059                         ret = PTR_ERR(conn);
1060                         goto out;
1061                 }
1062                 rs->rs_conn = conn;
1063         }
1064
1065         /* Parse any control messages the user may have included. */
1066         ret = rds_cmsg_send(rs, rm, msg, &allocated_mr);
1067         if (ret)
1068                 goto out;
1069
1070         if (rm->rdma.op_active && !conn->c_trans->xmit_rdma) {
1071                 printk_ratelimited(KERN_NOTICE "rdma_op %p conn xmit_rdma %p\n",
1072                                &rm->rdma, conn->c_trans->xmit_rdma);
1073                 ret = -EOPNOTSUPP;
1074                 goto out;
1075         }
1076
1077         if (rm->atomic.op_active && !conn->c_trans->xmit_atomic) {
1078                 printk_ratelimited(KERN_NOTICE "atomic_op %p conn xmit_atomic %p\n",
1079                                &rm->atomic, conn->c_trans->xmit_atomic);
1080                 ret = -EOPNOTSUPP;
1081                 goto out;
1082         }
1083
1084         rds_conn_connect_if_down(conn);
1085
1086         ret = rds_cong_wait(conn->c_fcong, dport, nonblock, rs);
1087         if (ret) {
1088                 rs->rs_seen_congestion = 1;
1089                 goto out;
1090         }
1091
1092         while (!rds_send_queue_rm(rs, conn, rm, rs->rs_bound_port,
1093                                   dport, &queued)) {
1094                 rds_stats_inc(s_send_queue_full);
1095
1096                 if (nonblock) {
1097                         ret = -EAGAIN;
1098                         goto out;
1099                 }
1100
1101                 timeo = wait_event_interruptible_timeout(*sk_sleep(sk),
1102                                         rds_send_queue_rm(rs, conn, rm,
1103                                                           rs->rs_bound_port,
1104                                                           dport,
1105                                                           &queued),
1106                                         timeo);
1107                 rdsdebug("sendmsg woke queued %d timeo %ld\n", queued, timeo);
1108                 if (timeo > 0 || timeo == MAX_SCHEDULE_TIMEOUT)
1109                         continue;
1110
1111                 ret = timeo;
1112                 if (ret == 0)
1113                         ret = -ETIMEDOUT;
1114                 goto out;
1115         }
1116
1117         /*
1118          * By now we've committed to the send.  We reuse rds_send_worker()
1119          * to retry sends in the rds thread if the transport asks us to.
1120          */
1121         rds_stats_inc(s_send_queued);
1122
1123         if (!test_bit(RDS_LL_SEND_FULL, &conn->c_flags))
1124                 rds_send_xmit(conn);
1125
1126         rds_message_put(rm);
1127         return payload_len;
1128
1129 out:
1130         /* If the user included a RDMA_MAP cmsg, we allocated a MR on the fly.
1131          * If the sendmsg goes through, we keep the MR. If it fails with EAGAIN
1132          * or in any other way, we need to destroy the MR again */
1133         if (allocated_mr)
1134                 rds_rdma_unuse(rs, rds_rdma_cookie_key(rm->m_rdma_cookie), 1);
1135
1136         if (rm)
1137                 rds_message_put(rm);
1138         return ret;
1139 }
1140
1141 /*
1142  * Reply to a ping packet.
1143  */
1144 int
1145 rds_send_pong(struct rds_connection *conn, __be16 dport)
1146 {
1147         struct rds_message *rm;
1148         unsigned long flags;
1149         int ret = 0;
1150
1151         rm = rds_message_alloc(0, GFP_ATOMIC);
1152         if (!rm) {
1153                 ret = -ENOMEM;
1154                 goto out;
1155         }
1156
1157         rm->m_daddr = conn->c_faddr;
1158         rm->data.op_active = 1;
1159
1160         rds_conn_connect_if_down(conn);
1161
1162         ret = rds_cong_wait(conn->c_fcong, dport, 1, NULL);
1163         if (ret)
1164                 goto out;
1165
1166         spin_lock_irqsave(&conn->c_lock, flags);
1167         list_add_tail(&rm->m_conn_item, &conn->c_send_queue);
1168         set_bit(RDS_MSG_ON_CONN, &rm->m_flags);
1169         rds_message_addref(rm);
1170         rm->m_inc.i_conn = conn;
1171
1172         rds_message_populate_header(&rm->m_inc.i_hdr, 0, dport,
1173                                     conn->c_next_tx_seq);
1174         conn->c_next_tx_seq++;
1175         spin_unlock_irqrestore(&conn->c_lock, flags);
1176
1177         rds_stats_inc(s_send_queued);
1178         rds_stats_inc(s_send_pong);
1179
1180         if (!test_bit(RDS_LL_SEND_FULL, &conn->c_flags))
1181                 queue_delayed_work(rds_wq, &conn->c_send_w, 0);
1182
1183         rds_message_put(rm);
1184         return 0;
1185
1186 out:
1187         if (rm)
1188                 rds_message_put(rm);
1189         return ret;
1190 }