kernel/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                         ret = conn->c_trans->xmit_rdma(conn, &rm->rdma);
 286                         if (ret)
 287                                 break;
 288                         conn->c_xmit_rdma_sent = 1;
 289
 290                         /* The transport owns the mapped memory for now.
 291                          * You can't unmap it while it's on the send queue */
 292                         set_bit(RDS_MSG_MAPPED, &rm->m_flags);
 293                 }
 294
 295                 if (rm->atomic.op_active && !conn->c_xmit_atomic_sent) {
 296                         rm->m_final_op = &rm->atomic;
 297                         ret = conn->c_trans->xmit_atomic(conn, &rm->atomic);
 298                         if (ret)
 299                                 break;
 300                         conn->c_xmit_atomic_sent = 1;
 301
 302                         /* The transport owns the mapped memory for now.
 303                          * You can't unmap it while it's on the send queue */
 304                         set_bit(RDS_MSG_MAPPED, &rm->m_flags);
 305                 }
 306
 307                 /*
 308                  * A number of cases require an RDS header to be sent
 309                  * even if there is no data.
 310                  * We permit 0-byte sends; rds-ping depends on this.
 311                  * However, if there are exclusively attached silent ops,
 312                  * we skip the hdr/data send, to enable silent operation.
 313                  */
 314                 if (rm->data.op_nents == 0) {
 315                         int ops_present;
 316                         int all_ops_are_silent = 1;
 317
 318                         ops_present = (rm->atomic.op_active || rm->rdma.op_active);
 319                         if (rm->atomic.op_active && !rm->atomic.op_silent)
 320                                 all_ops_are_silent = 0;
 321                         if (rm->rdma.op_active && !rm->rdma.op_silent)
 322                                 all_ops_are_silent = 0;
 323
 324                         if (ops_present && all_ops_are_silent
 325                             && !rm->m_rdma_cookie)
 326                                 rm->data.op_active = 0;
 327                 }
 328
 329                 if (rm->data.op_active && !conn->c_xmit_data_sent) {
 330                         rm->m_final_op = &rm->data;
 331                         ret = conn->c_trans->xmit(conn, rm,
 332                                                   conn->c_xmit_hdr_off,
 333                                                   conn->c_xmit_sg,
 334                                                   conn->c_xmit_data_off);
 335                         if (ret <= 0)
 336                                 break;
 337
 338                         if (conn->c_xmit_hdr_off < sizeof(struct rds_header)) {
 339                                 tmp = min_t(int, ret,
 340                                             sizeof(struct rds_header) -
 341                                             conn->c_xmit_hdr_off);
 342                                 conn->c_xmit_hdr_off += tmp;
 343                                 ret -= tmp;
 344                         }
 345
 346                         sg = &rm->data.op_sg[conn->c_xmit_sg];
 347                         while (ret) {
 348                                 tmp = min_t(int, ret, sg->length -
 349                                                       conn->c_xmit_data_off);
 350                                 conn->c_xmit_data_off += tmp;
 351                                 ret -= tmp;
 352                                 if (conn->c_xmit_data_off == sg->length) {
 353                                         conn->c_xmit_data_off = 0;
 354                                         sg++;
 355                                         conn->c_xmit_sg++;
 356                                         BUG_ON(ret != 0 &&
 357                                                conn->c_xmit_sg == rm->data.op_nents);
 358                                 }
 359                         }
 360
 361                         if (conn->c_xmit_hdr_off == sizeof(struct rds_header) &&
 362                             (conn->c_xmit_sg == rm->data.op_nents))
 363                                 conn->c_xmit_data_sent = 1;
 364                 }
 365
 366                 /*
 367                  * A rm will only take multiple times through this loop
 368                  * if there is a data op. Thus, if the data is sent (or there was
 369                  * none), then we're done with the rm.
 370                  */
 371                 if (!rm->data.op_active || conn->c_xmit_data_sent) {
 372                         conn->c_xmit_rm = NULL;
 373                         conn->c_xmit_sg = 0;
 374                         conn->c_xmit_hdr_off = 0;
 375                         conn->c_xmit_data_off = 0;
 376                         conn->c_xmit_rdma_sent = 0;
 377                         conn->c_xmit_atomic_sent = 0;
 378                         conn->c_xmit_data_sent = 0;
 379
 380                         rds_message_put(rm);
 381                 }
 382         }
 383
 384 over_batch:
 385         if (conn->c_trans->xmit_complete)
 386                 conn->c_trans->xmit_complete(conn);
 387         release_in_xmit(conn);
 388
 389         /* Nuke any messages we decided not to retransmit. */
 390         if (!list_empty(&to_be_dropped)) {
 391                 /* irqs on here, so we can put(), unlike above */
 392                 list_for_each_entry(rm, &to_be_dropped, m_conn_item)
 393                         rds_message_put(rm);
 394                 rds_send_remove_from_sock(&to_be_dropped, RDS_RDMA_DROPPED);
 395         }
 396
 397         /*
 398          * Other senders can queue a message after we last test the send queue
 399          * but before we clear RDS_IN_XMIT.  In that case they'd back off and
 400          * not try and send their newly queued message.  We need to check the
 401          * send queue after having cleared RDS_IN_XMIT so that their message
 402          * doesn't get stuck on the send queue.
 403          *
 404          * If the transport cannot continue (i.e ret != 0), then it must
 405          * call us when more room is available, such as from the tx
 406          * completion handler.
 407          *
 408          * We have an extra generation check here so that if someone manages
 409          * to jump in after our release_in_xmit, we'll see that they have done
 410          * some work and we will skip our goto
 411          */
 412         if (ret == 0) {
 413                 smp_mb();
 414                 if (!list_empty(&conn->c_send_queue) &&
 415                     send_gen == conn->c_send_gen) {
 416                         rds_stats_inc(s_send_lock_queue_raced);
 417                         goto restart;
 418                 }
 419         }
 420 out:
 421         return ret;
 422 }
 423
 424 static void rds_send_sndbuf_remove(struct rds_sock *rs, struct rds_message *rm)
 425 {
 426         u32 len = be32_to_cpu(rm->m_inc.i_hdr.h_len);
 427
 428         assert_spin_locked(&rs->rs_lock);
 429
 430         BUG_ON(rs->rs_snd_bytes < len);
 431         rs->rs_snd_bytes -= len;
 432
 433         if (rs->rs_snd_bytes == 0)
 434                 rds_stats_inc(s_send_queue_empty);
 435 }
 436
 437 static inline int rds_send_is_acked(struct rds_message *rm, u64 ack,
 438                                     is_acked_func is_acked)
 439 {
 440         if (is_acked)
 441                 return is_acked(rm, ack);
 442         return be64_to_cpu(rm->m_inc.i_hdr.h_sequence) <= ack;
 443 }
 444
 445 /*
 446  * This is pretty similar to what happens below in the ACK
 447  * handling code - except that we call here as soon as we get
 448  * the IB send completion on the RDMA op and the accompanying
 449  * message.
 450  */
 451 void rds_rdma_send_complete(struct rds_message *rm, int status)
 452 {
 453         struct rds_sock *rs = NULL;
 454         struct rm_rdma_op *ro;
 455         struct rds_notifier *notifier;
 456         unsigned long flags;
 457
 458         spin_lock_irqsave(&rm->m_rs_lock, flags);
 459
 460         ro = &rm->rdma;
 461         if (test_bit(RDS_MSG_ON_SOCK, &rm->m_flags) &&
 462             ro->op_active && ro->op_notify && ro->op_notifier) {
 463                 notifier = ro->op_notifier;
 464                 rs = rm->m_rs;
 465                 sock_hold(rds_rs_to_sk(rs));
 466
 467                 notifier->n_status = status;
 468                 spin_lock(&rs->rs_lock);
 469                 list_add_tail(&notifier->n_list, &rs->rs_notify_queue);
 470                 spin_unlock(&rs->rs_lock);
 471
 472                 ro->op_notifier = NULL;
 473         }
 474
 475         spin_unlock_irqrestore(&rm->m_rs_lock, flags);
 476
 477         if (rs) {
 478                 rds_wake_sk_sleep(rs);
 479                 sock_put(rds_rs_to_sk(rs));
 480         }
 481 }
 482 EXPORT_SYMBOL_GPL(rds_rdma_send_complete);
 483
 484 /*
 485  * Just like above, except looks at atomic op
 486  */
 487 void rds_atomic_send_complete(struct rds_message *rm, int status)
 488 {
 489         struct rds_sock *rs = NULL;
 490         struct rm_atomic_op *ao;
 491         struct rds_notifier *notifier;
 492         unsigned long flags;
 493
 494         spin_lock_irqsave(&rm->m_rs_lock, flags);
 495
 496         ao = &rm->atomic;
 497         if (test_bit(RDS_MSG_ON_SOCK, &rm->m_flags)
 498             && ao->op_active && ao->op_notify && ao->op_notifier) {
 499                 notifier = ao->op_notifier;
 500                 rs = rm->m_rs;
 501                 sock_hold(rds_rs_to_sk(rs));
 502
 503                 notifier->n_status = status;
 504                 spin_lock(&rs->rs_lock);
 505                 list_add_tail(&notifier->n_list, &rs->rs_notify_queue);
 506                 spin_unlock(&rs->rs_lock);
 507
 508                 ao->op_notifier = NULL;
 509         }
 510
 511         spin_unlock_irqrestore(&rm->m_rs_lock, flags);
 512
 513         if (rs) {
 514                 rds_wake_sk_sleep(rs);
 515                 sock_put(rds_rs_to_sk(rs));
 516         }
 517 }
 518 EXPORT_SYMBOL_GPL(rds_atomic_send_complete);
 519
 520 /*
 521  * This is the same as rds_rdma_send_complete except we
 522  * don't do any locking - we have all the ingredients (message,
 523  * socket, socket lock) and can just move the notifier.
 524  */
 525 static inline void
 526 __rds_send_complete(struct rds_sock *rs, struct rds_message *rm, int status)
 527 {
 528         struct rm_rdma_op *ro;
 529         struct rm_atomic_op *ao;
 530
 531         ro = &rm->rdma;
 532         if (ro->op_active && ro->op_notify && ro->op_notifier) {
 533                 ro->op_notifier->n_status = status;
 534                 list_add_tail(&ro->op_notifier->n_list, &rs->rs_notify_queue);
 535                 ro->op_notifier = NULL;
 536         }
 537
 538         ao = &rm->atomic;
 539         if (ao->op_active && ao->op_notify && ao->op_notifier) {
 540                 ao->op_notifier->n_status = status;
 541                 list_add_tail(&ao->op_notifier->n_list, &rs->rs_notify_queue);
 542                 ao->op_notifier = NULL;
 543         }
 544
 545         /* No need to wake the app - caller does this */
 546 }
 547
 548 /*
 549  * This is called from the IB send completion when we detect
 550  * a RDMA operation that failed with remote access error.
 551  * So speed is not an issue here.
 552  */
 553 struct rds_message *rds_send_get_message(struct rds_connection *conn,
 554                                          struct rm_rdma_op *op)
 555 {
 556         struct rds_message *rm, *tmp, *found = NULL;
 557         unsigned long flags;
 558
 559         spin_lock_irqsave(&conn->c_lock, flags);
 560
 561         list_for_each_entry_safe(rm, tmp, &conn->c_retrans, m_conn_item) {
 562                 if (&rm->rdma == op) {
 563                         atomic_inc(&rm->m_refcount);
 564                         found = rm;
 565                         goto out;
 566                 }
 567         }
 568
 569         list_for_each_entry_safe(rm, tmp, &conn->c_send_queue, m_conn_item) {
 570                 if (&rm->rdma == op) {
 571                         atomic_inc(&rm->m_refcount);
 572                         found = rm;
 573                         break;
 574                 }
 575         }
 576
 577 out:
 578         spin_unlock_irqrestore(&conn->c_lock, flags);
 579
 580         return found;
 581 }
 582 EXPORT_SYMBOL_GPL(rds_send_get_message);
 583
 584 /*
 585  * This removes messages from the socket's list if they're on it.  The list
 586  * argument must be private to the caller, we must be able to modify it
 587  * without locks.  The messages must have a reference held for their
 588  * position on the list.  This function will drop that reference after
 589  * removing the messages from the 'messages' list regardless of if it found
 590  * the messages on the socket list or not.
 591  */
 592 static void rds_send_remove_from_sock(struct list_head *messages, int status)
 593 {
 594         unsigned long flags;
 595         struct rds_sock *rs = NULL;
 596         struct rds_message *rm;
 597
 598         while (!list_empty(messages)) {
 599                 int was_on_sock = 0;
 600
 601                 rm = list_entry(messages->next, struct rds_message,
 602                                 m_conn_item);
 603                 list_del_init(&rm->m_conn_item);
 604
 605                 /*
 606                  * If we see this flag cleared then we're *sure* that someone
 607                  * else beat us to removing it from the sock.  If we race
 608                  * with their flag update we'll get the lock and then really
 609                  * see that the flag has been cleared.
 610                  *
 611                  * The message spinlock makes sure nobody clears rm->m_rs
 612                  * while we're messing with it. It does not prevent the
 613                  * message from being removed from the socket, though.
 614                  */
 615                 spin_lock_irqsave(&rm->m_rs_lock, flags);
 616                 if (!test_bit(RDS_MSG_ON_SOCK, &rm->m_flags))
 617                         goto unlock_and_drop;
 618
 619                 if (rs != rm->m_rs) {
 620                         if (rs) {
 621                                 rds_wake_sk_sleep(rs);
 622                                 sock_put(rds_rs_to_sk(rs));
 623                         }
 624                         rs = rm->m_rs;
 625                         if (rs)
 626                                 sock_hold(rds_rs_to_sk(rs));
 627                 }
 628                 if (!rs)
 629                         goto unlock_and_drop;
 630                 spin_lock(&rs->rs_lock);
 631
 632                 if (test_and_clear_bit(RDS_MSG_ON_SOCK, &rm->m_flags)) {
 633                         struct rm_rdma_op *ro = &rm->rdma;
 634                         struct rds_notifier *notifier;
 635
 636                         list_del_init(&rm->m_sock_item);
 637                         rds_send_sndbuf_remove(rs, rm);
 638
 639                         if (ro->op_active && ro->op_notifier &&
 640                                (ro->op_notify || (ro->op_recverr && status))) {
 641                                 notifier = ro->op_notifier;
 642                                 list_add_tail(&notifier->n_list,
 643                                                 &rs->rs_notify_queue);
 644                                 if (!notifier->n_status)
 645                                         notifier->n_status = status;
 646                                 rm->rdma.op_notifier = NULL;
 647                         }
 648                         was_on_sock = 1;
 649                         rm->m_rs = NULL;
 650                 }
 651                 spin_unlock(&rs->rs_lock);
 652
 653 unlock_and_drop:
 654                 spin_unlock_irqrestore(&rm->m_rs_lock, flags);
 655                 rds_message_put(rm);
 656                 if (was_on_sock)
 657                         rds_message_put(rm);
 658         }
 659
 660         if (rs) {
 661                 rds_wake_sk_sleep(rs);
 662                 sock_put(rds_rs_to_sk(rs));
 663         }
 664 }
 665
 666 /*
 667  * Transports call here when they've determined that the receiver queued
 668  * messages up to, and including, the given sequence number.  Messages are
 669  * moved to the retrans queue when rds_send_xmit picks them off the send
 670  * queue. This means that in the TCP case, the message may not have been
 671  * assigned the m_ack_seq yet - but that's fine as long as tcp_is_acked
 672  * checks the RDS_MSG_HAS_ACK_SEQ bit.
 673  */
 674 void rds_send_drop_acked(struct rds_connection *conn, u64 ack,
 675                          is_acked_func is_acked)
 676 {
 677         struct rds_message *rm, *tmp;
 678         unsigned long flags;
 679         LIST_HEAD(list);
 680
 681         spin_lock_irqsave(&conn->c_lock, flags);
 682
 683         list_for_each_entry_safe(rm, tmp, &conn->c_retrans, m_conn_item) {
 684                 if (!rds_send_is_acked(rm, ack, is_acked))
 685                         break;
 686
 687                 list_move(&rm->m_conn_item, &list);
 688                 clear_bit(RDS_MSG_ON_CONN, &rm->m_flags);
 689         }
 690
 691         /* order flag updates with spin locks */
 692         if (!list_empty(&list))
 693                 smp_mb__after_atomic();
 694
 695         spin_unlock_irqrestore(&conn->c_lock, flags);
 696
 697         /* now remove the messages from the sock list as needed */
 698         rds_send_remove_from_sock(&list, RDS_RDMA_SUCCESS);
 699 }
 700 EXPORT_SYMBOL_GPL(rds_send_drop_acked);
 701
 702 void rds_send_drop_to(struct rds_sock *rs, struct sockaddr_in *dest)
 703 {
 704         struct rds_message *rm, *tmp;
 705         struct rds_connection *conn;
 706         unsigned long flags;
 707         LIST_HEAD(list);
 708
 709         /* get all the messages we're dropping under the rs lock */
 710         spin_lock_irqsave(&rs->rs_lock, flags);
 711
 712         list_for_each_entry_safe(rm, tmp, &rs->rs_send_queue, m_sock_item) {
 713                 if (dest && (dest->sin_addr.s_addr != rm->m_daddr ||
 714                              dest->sin_port != rm->m_inc.i_hdr.h_dport))
 715                         continue;
 716
 717                 list_move(&rm->m_sock_item, &list);
 718                 rds_send_sndbuf_remove(rs, rm);
 719                 clear_bit(RDS_MSG_ON_SOCK, &rm->m_flags);
 720         }
 721
 722         /* order flag updates with the rs lock */
 723         smp_mb__after_atomic();
 724
 725         spin_unlock_irqrestore(&rs->rs_lock, flags);
 726
 727         if (list_empty(&list))
 728                 return;
 729
 730         /* Remove the messages from the conn */
 731         list_for_each_entry(rm, &list, m_sock_item) {
 732
 733                 conn = rm->m_inc.i_conn;
 734
 735                 spin_lock_irqsave(&conn->c_lock, flags);
 736                 /*
 737                  * Maybe someone else beat us to removing rm from the conn.
 738                  * If we race with their flag update we'll get the lock and
 739                  * then really see that the flag has been cleared.
 740                  */
 741                 if (!test_and_clear_bit(RDS_MSG_ON_CONN, &rm->m_flags)) {
 742                         spin_unlock_irqrestore(&conn->c_lock, flags);
 743                         spin_lock_irqsave(&rm->m_rs_lock, flags);
 744                         rm->m_rs = NULL;
 745                         spin_unlock_irqrestore(&rm->m_rs_lock, flags);
 746                         continue;
 747                 }
 748                 list_del_init(&rm->m_conn_item);
 749                 spin_unlock_irqrestore(&conn->c_lock, flags);
 750
 751                 /*
 752                  * Couldn't grab m_rs_lock in top loop (lock ordering),
 753                  * but we can now.
 754                  */
 755                 spin_lock_irqsave(&rm->m_rs_lock, flags);
 756
 757                 spin_lock(&rs->rs_lock);
 758                 __rds_send_complete(rs, rm, RDS_RDMA_CANCELED);
 759                 spin_unlock(&rs->rs_lock);
 760
 761                 rm->m_rs = NULL;
 762                 spin_unlock_irqrestore(&rm->m_rs_lock, flags);
 763
 764                 rds_message_put(rm);
 765         }
 766
 767         rds_wake_sk_sleep(rs);
 768
 769         while (!list_empty(&list)) {
 770                 rm = list_entry(list.next, struct rds_message, m_sock_item);
 771                 list_del_init(&rm->m_sock_item);
 772
 773                 rds_message_wait(rm);
 774                 rds_message_put(rm);
 775         }
 776 }
 777
 778 /*
 779  * we only want this to fire once so we use the callers 'queued'.  It's
 780  * possible that another thread can race with us and remove the
 781  * message from the flow with RDS_CANCEL_SENT_TO.
 782  */
 783 static int rds_send_queue_rm(struct rds_sock *rs, struct rds_connection *conn,
 784                              struct rds_message *rm, __be16 sport,
 785                              __be16 dport, int *queued)
 786 {
 787         unsigned long flags;
 788         u32 len;
 789
 790         if (*queued)
 791                 goto out;
 792
 793         len = be32_to_cpu(rm->m_inc.i_hdr.h_len);
 794
 795         /* this is the only place which holds both the socket's rs_lock
 796          * and the connection's c_lock */
 797         spin_lock_irqsave(&rs->rs_lock, flags);
 798
 799         /*
 800          * If there is a little space in sndbuf, we don't queue anything,
 801          * and userspace gets -EAGAIN. But poll() indicates there's send
 802          * room. This can lead to bad behavior (spinning) if snd_bytes isn't
 803          * freed up by incoming acks. So we check the *old* value of
 804          * rs_snd_bytes here to allow the last msg to exceed the buffer,
 805          * and poll() now knows no more data can be sent.
 806          */
 807         if (rs->rs_snd_bytes < rds_sk_sndbuf(rs)) {
 808                 rs->rs_snd_bytes += len;
 809
 810                 /* let recv side know we are close to send space exhaustion.
 811                  * This is probably not the optimal way to do it, as this
 812                  * means we set the flag on *all* messages as soon as our
 813                  * throughput hits a certain threshold.
 814                  */
 815                 if (rs->rs_snd_bytes >= rds_sk_sndbuf(rs) / 2)
 816                         __set_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags);
 817
 818                 list_add_tail(&rm->m_sock_item, &rs->rs_send_queue);
 819                 set_bit(RDS_MSG_ON_SOCK, &rm->m_flags);
 820                 rds_message_addref(rm);
 821                 rm->m_rs = rs;
 822
 823                 /* The code ordering is a little weird, but we're
 824                    trying to minimize the time we hold c_lock */
 825                 rds_message_populate_header(&rm->m_inc.i_hdr, sport, dport, 0);
 826                 rm->m_inc.i_conn = conn;
 827                 rds_message_addref(rm);
 828
 829                 spin_lock(&conn->c_lock);
 830                 rm->m_inc.i_hdr.h_sequence = cpu_to_be64(conn->c_next_tx_seq++);
 831                 list_add_tail(&rm->m_conn_item, &conn->c_send_queue);
 832                 set_bit(RDS_MSG_ON_CONN, &rm->m_flags);
 833                 spin_unlock(&conn->c_lock);
 834
 835                 rdsdebug("queued msg %p len %d, rs %p bytes %d seq %llu\n",
 836                          rm, len, rs, rs->rs_snd_bytes,
 837                          (unsigned long long)be64_to_cpu(rm->m_inc.i_hdr.h_sequence));
 838
 839                 *queued = 1;
 840         }
 841
 842         spin_unlock_irqrestore(&rs->rs_lock, flags);
 843 out:
 844         return *queued;
 845 }
 846
 847 /*
 848  * rds_message is getting to be quite complicated, and we'd like to allocate
 849  * it all in one go. This figures out how big it needs to be up front.
 850  */
 851 static int rds_rm_size(struct msghdr *msg, int data_len)
 852 {
 853         struct cmsghdr *cmsg;
 854         int size = 0;
 855         int cmsg_groups = 0;
 856         int retval;
 857
 858         for_each_cmsghdr(cmsg, msg) {
 859                 if (!CMSG_OK(msg, cmsg))
 860                         return -EINVAL;
 861
 862                 if (cmsg->cmsg_level != SOL_RDS)
 863                         continue;
 864
 865                 switch (cmsg->cmsg_type) {
 866                 case RDS_CMSG_RDMA_ARGS:
 867                         cmsg_groups |= 1;
 868                         retval = rds_rdma_extra_size(CMSG_DATA(cmsg));
 869                         if (retval < 0)
 870                                 return retval;
 871                         size += retval;
 872
 873                         break;
 874
 875                 case RDS_CMSG_RDMA_DEST:
 876                 case RDS_CMSG_RDMA_MAP:
 877                         cmsg_groups |= 2;
 878                         /* these are valid but do no add any size */
 879                         break;
 880
 881                 case RDS_CMSG_ATOMIC_CSWP:
 882                 case RDS_CMSG_ATOMIC_FADD:
 883                 case RDS_CMSG_MASKED_ATOMIC_CSWP:
 884                 case RDS_CMSG_MASKED_ATOMIC_FADD:
 885                         cmsg_groups |= 1;
 886                         size += sizeof(struct scatterlist);
 887                         break;
 888
 889                 default:
 890                         return -EINVAL;
 891                 }
 892
 893         }
 894
 895         size += ceil(data_len, PAGE_SIZE) * sizeof(struct scatterlist);
 896
 897         /* Ensure (DEST, MAP) are never used with (ARGS, ATOMIC) */
 898         if (cmsg_groups == 3)
 899                 return -EINVAL;
 900
 901         return size;
 902 }
 903
 904 static int rds_cmsg_send(struct rds_sock *rs, struct rds_message *rm,
 905                          struct msghdr *msg, int *allocated_mr)
 906 {
 907         struct cmsghdr *cmsg;
 908         int ret = 0;
 909
 910         for_each_cmsghdr(cmsg, msg) {
 911                 if (!CMSG_OK(msg, cmsg))
 912                         return -EINVAL;
 913
 914                 if (cmsg->cmsg_level != SOL_RDS)
 915                         continue;
 916
 917                 /* As a side effect, RDMA_DEST and RDMA_MAP will set
 918                  * rm->rdma.m_rdma_cookie and rm->rdma.m_rdma_mr.
 919                  */
 920                 switch (cmsg->cmsg_type) {
 921                 case RDS_CMSG_RDMA_ARGS:
 922                         ret = rds_cmsg_rdma_args(rs, rm, cmsg);
 923                         break;
 924
 925                 case RDS_CMSG_RDMA_DEST:
 926                         ret = rds_cmsg_rdma_dest(rs, rm, cmsg);
 927                         break;
 928
 929                 case RDS_CMSG_RDMA_MAP:
 930                         ret = rds_cmsg_rdma_map(rs, rm, cmsg);
 931                         if (!ret)
 932                                 *allocated_mr = 1;
 933                         break;
 934                 case RDS_CMSG_ATOMIC_CSWP:
 935                 case RDS_CMSG_ATOMIC_FADD:
 936                 case RDS_CMSG_MASKED_ATOMIC_CSWP:
 937                 case RDS_CMSG_MASKED_ATOMIC_FADD:
 938                         ret = rds_cmsg_atomic(rs, rm, cmsg);
 939                         break;
 940
 941                 default:
 942                         return -EINVAL;
 943                 }
 944
 945                 if (ret)
 946                         break;
 947         }
 948
 949         return ret;
 950 }
 951
 952 int rds_sendmsg(struct socket *sock, struct msghdr *msg, size_t payload_len)
 953 {
 954         struct sock *sk = sock->sk;
 955         struct rds_sock *rs = rds_sk_to_rs(sk);
 956         DECLARE_SOCKADDR(struct sockaddr_in *, usin, msg->msg_name);
 957         __be32 daddr;
 958         __be16 dport;
 959         struct rds_message *rm = NULL;
 960         struct rds_connection *conn;
 961         int ret = 0;
 962         int queued = 0, allocated_mr = 0;
 963         int nonblock = msg->msg_flags & MSG_DONTWAIT;
 964         long timeo = sock_sndtimeo(sk, nonblock);
 965
 966         /* Mirror Linux UDP mirror of BSD error message compatibility */
 967         /* XXX: Perhaps MSG_MORE someday */
 968         if (msg->msg_flags & ~(MSG_DONTWAIT | MSG_CMSG_COMPAT)) {
 969                 ret = -EOPNOTSUPP;
 970                 goto out;
 971         }
 972
 973         if (msg->msg_namelen) {
 974                 /* XXX fail non-unicast destination IPs? */
 975                 if (msg->msg_namelen < sizeof(*usin) || usin->sin_family != AF_INET) {
 976                         ret = -EINVAL;
 977                         goto out;
 978                 }
 979                 daddr = usin->sin_addr.s_addr;
 980                 dport = usin->sin_port;
 981         } else {
 982                 /* We only care about consistency with ->connect() */
 983                 lock_sock(sk);
 984                 daddr = rs->rs_conn_addr;
 985                 dport = rs->rs_conn_port;
 986                 release_sock(sk);
 987         }
 988
 989         /* racing with another thread binding seems ok here */
 990         if (daddr == 0 || rs->rs_bound_addr == 0) {
 991                 ret = -ENOTCONN; /* XXX not a great errno */
 992                 goto out;
 993         }
 994
 995         /* size of rm including all sgs */
 996         ret = rds_rm_size(msg, payload_len);
 997         if (ret < 0)
 998                 goto out;
 999
1000         rm = rds_message_alloc(ret, GFP_KERNEL);
1001         if (!rm) {
1002                 ret = -ENOMEM;
1003                 goto out;
1004         }
1005
1006         /* Attach data to the rm */
1007         if (payload_len) {
1008                 rm->data.op_sg = rds_message_alloc_sgs(rm, ceil(payload_len, PAGE_SIZE));
1009                 if (!rm->data.op_sg) {
1010                         ret = -ENOMEM;
1011                         goto out;
1012                 }
1013                 ret = rds_message_copy_from_user(rm, &msg->msg_iter);
1014                 if (ret)
1015                         goto out;
1016         }
1017         rm->data.op_active = 1;
1018
1019         rm->m_daddr = daddr;
1020
1021         /* rds_conn_create has a spinlock that runs with IRQ off.
1022          * Caching the conn in the socket helps a lot. */
1023         if (rs->rs_conn && rs->rs_conn->c_faddr == daddr)
1024                 conn = rs->rs_conn;
1025         else {
1026                 conn = rds_conn_create_outgoing(rs->rs_bound_addr, daddr,
1027                                         rs->rs_transport,
1028                                         sock->sk->sk_allocation);
1029                 if (IS_ERR(conn)) {
1030                         ret = PTR_ERR(conn);
1031                         goto out;
1032                 }
1033                 rs->rs_conn = conn;
1034         }
1035
1036         /* Parse any control messages the user may have included. */
1037         ret = rds_cmsg_send(rs, rm, msg, &allocated_mr);
1038         if (ret)
1039                 goto out;
1040
1041         if (rm->rdma.op_active && !conn->c_trans->xmit_rdma) {
1042                 printk_ratelimited(KERN_NOTICE "rdma_op %p conn xmit_rdma %p\n",
1043                                &rm->rdma, conn->c_trans->xmit_rdma);
1044                 ret = -EOPNOTSUPP;
1045                 goto out;
1046         }
1047
1048         if (rm->atomic.op_active && !conn->c_trans->xmit_atomic) {
1049                 printk_ratelimited(KERN_NOTICE "atomic_op %p conn xmit_atomic %p\n",
1050                                &rm->atomic, conn->c_trans->xmit_atomic);
1051                 ret = -EOPNOTSUPP;
1052                 goto out;
1053         }
1054
1055         rds_conn_connect_if_down(conn);
1056
1057         ret = rds_cong_wait(conn->c_fcong, dport, nonblock, rs);
1058         if (ret) {
1059                 rs->rs_seen_congestion = 1;
1060                 goto out;
1061         }
1062
1063         while (!rds_send_queue_rm(rs, conn, rm, rs->rs_bound_port,
1064                                   dport, &queued)) {
1065                 rds_stats_inc(s_send_queue_full);
1066                 /* XXX make sure this is reasonable */
1067                 if (payload_len > rds_sk_sndbuf(rs)) {
1068                         ret = -EMSGSIZE;
1069                         goto out;
1070                 }
1071                 if (nonblock) {
1072                         ret = -EAGAIN;
1073                         goto out;
1074                 }
1075
1076                 timeo = wait_event_interruptible_timeout(*sk_sleep(sk),
1077                                         rds_send_queue_rm(rs, conn, rm,
1078                                                           rs->rs_bound_port,
1079                                                           dport,
1080                                                           &queued),
1081                                         timeo);
1082                 rdsdebug("sendmsg woke queued %d timeo %ld\n", queued, timeo);
1083                 if (timeo > 0 || timeo == MAX_SCHEDULE_TIMEOUT)
1084                         continue;
1085
1086                 ret = timeo;
1087                 if (ret == 0)
1088                         ret = -ETIMEDOUT;
1089                 goto out;
1090         }
1091
1092         /*
1093          * By now we've committed to the send.  We reuse rds_send_worker()
1094          * to retry sends in the rds thread if the transport asks us to.
1095          */
1096         rds_stats_inc(s_send_queued);
1097
1098         if (!test_bit(RDS_LL_SEND_FULL, &conn->c_flags))
1099                 rds_send_xmit(conn);
1100
1101         rds_message_put(rm);
1102         return payload_len;
1103
1104 out:
1105         /* If the user included a RDMA_MAP cmsg, we allocated a MR on the fly.
1106          * If the sendmsg goes through, we keep the MR. If it fails with EAGAIN
1107          * or in any other way, we need to destroy the MR again */
1108         if (allocated_mr)
1109                 rds_rdma_unuse(rs, rds_rdma_cookie_key(rm->m_rdma_cookie), 1);
1110
1111         if (rm)
1112                 rds_message_put(rm);
1113         return ret;
1114 }
1115
1116 /*
1117  * Reply to a ping packet.
1118  */
1119 int
1120 rds_send_pong(struct rds_connection *conn, __be16 dport)
1121 {
1122         struct rds_message *rm;
1123         unsigned long flags;
1124         int ret = 0;
1125
1126         rm = rds_message_alloc(0, GFP_ATOMIC);
1127         if (!rm) {
1128                 ret = -ENOMEM;
1129                 goto out;
1130         }
1131
1132         rm->m_daddr = conn->c_faddr;
1133         rm->data.op_active = 1;
1134
1135         rds_conn_connect_if_down(conn);
1136
1137         ret = rds_cong_wait(conn->c_fcong, dport, 1, NULL);
1138         if (ret)
1139                 goto out;
1140
1141         spin_lock_irqsave(&conn->c_lock, flags);
1142         list_add_tail(&rm->m_conn_item, &conn->c_send_queue);
1143         set_bit(RDS_MSG_ON_CONN, &rm->m_flags);
1144         rds_message_addref(rm);
1145         rm->m_inc.i_conn = conn;
1146
1147         rds_message_populate_header(&rm->m_inc.i_hdr, 0, dport,
1148                                     conn->c_next_tx_seq);
1149         conn->c_next_tx_seq++;
1150         spin_unlock_irqrestore(&conn->c_lock, flags);
1151
1152         rds_stats_inc(s_send_queued);
1153         rds_stats_inc(s_send_pong);
1154
1155         if (!test_bit(RDS_LL_SEND_FULL, &conn->c_flags))
1156                 queue_delayed_work(rds_wq, &conn->c_send_w, 0);
1157
1158         rds_message_put(rm);
1159         return 0;
1160
1161 out:
1162         if (rm)
1163                 rds_message_put(rm);
1164         return ret;
1165 }