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
  41 #include "rds.h"
  42
  43 /* When transmitting messages in rds_send_xmit, we need to emerge from
  44  * time to time and briefly release the CPU. Otherwise the softlock watchdog
  45  * will kick our shin.
  46  * Also, it seems fairer to not let one busy connection stall all the
  47  * others.
  48  *
  49  * send_batch_count is the number of times we'll loop in send_xmit. Setting
  50  * it to 0 will restore the old behavior (where we looped until we had
  51  * drained the queue).
  52  */
  53 static int send_batch_count = 64;
  54 module_param(send_batch_count, int, 0444);
  55 MODULE_PARM_DESC(send_batch_count, " batch factor when working the send queue");
  56
  57 static void rds_send_remove_from_sock(struct list_head *messages, int status);
  58
  59 /*
  60  * Reset the send state.  Callers must ensure that this doesn't race with
  61  * rds_send_xmit().
  62  */
  63 void rds_send_reset(struct rds_connection *conn)
  64 {
  65         struct rds_message *rm, *tmp;
  66         unsigned long flags;
  67
  68         if (conn->c_xmit_rm) {
  69                 rm = conn->c_xmit_rm;
  70                 conn->c_xmit_rm = NULL;
  71                 /* Tell the user the RDMA op is no longer mapped by the
  72                  * transport. This isn't entirely true (it's flushed out
  73                  * independently) but as the connection is down, there's
  74                  * no ongoing RDMA to/from that memory */
  75                 rds_message_unmapped(rm);
  76                 rds_message_put(rm);
  77         }
  78
  79         conn->c_xmit_sg = 0;
  80         conn->c_xmit_hdr_off = 0;
  81         conn->c_xmit_data_off = 0;
  82         conn->c_xmit_atomic_sent = 0;
  83         conn->c_xmit_rdma_sent = 0;
  84         conn->c_xmit_data_sent = 0;
  85
  86         conn->c_map_queued = 0;
  87
  88         conn->c_unacked_packets = rds_sysctl_max_unacked_packets;
  89         conn->c_unacked_bytes = rds_sysctl_max_unacked_bytes;
  90
  91         /* Mark messages as retransmissions, and move them to the send q */
  92         spin_lock_irqsave(&conn->c_lock, flags);
  93         list_for_each_entry_safe(rm, tmp, &conn->c_retrans, m_conn_item) {
  94                 set_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags);
  95                 set_bit(RDS_MSG_RETRANSMITTED, &rm->m_flags);
  96         }
  97         list_splice_init(&conn->c_retrans, &conn->c_send_queue);
  98         spin_unlock_irqrestore(&conn->c_lock, flags);
  99 }
 100
 101 static int acquire_in_xmit(struct rds_connection *conn)
 102 {
 103         return test_and_set_bit(RDS_IN_XMIT, &conn->c_flags) == 0;
 104 }
 105
 106 static void release_in_xmit(struct rds_connection *conn)
 107 {
 108         clear_bit(RDS_IN_XMIT, &conn->c_flags);
 109         smp_mb__after_clear_bit();
 110         /*
 111          * We don't use wait_on_bit()/wake_up_bit() because our waking is in a
 112          * hot path and finding waiters is very rare.  We don't want to walk
 113          * the system-wide hashed waitqueue buckets in the fast path only to
 114          * almost never find waiters.
 115          */
 116         if (waitqueue_active(&conn->c_waitq))
 117                 wake_up_all(&conn->c_waitq);
 118 }
 119
 120 /*
 121  * We're making the conscious trade-off here to only send one message
 122  * down the connection at a time.
 123  *   Pro:
 124  *      - tx queueing is a simple fifo list
 125  *      - reassembly is optional and easily done by transports per conn
 126  *      - no per flow rx lookup at all, straight to the socket
 127  *      - less per-frag memory and wire overhead
 128  *   Con:
 129  *      - queued acks can be delayed behind large messages
 130  *   Depends:
 131  *      - small message latency is higher behind queued large messages
 132  *      - large message latency isn't starved by intervening small sends
 133  */
 134 int rds_send_xmit(struct rds_connection *conn)
 135 {
 136         struct rds_message *rm;
 137         unsigned long flags;
 138         unsigned int tmp;
 139         struct scatterlist *sg;
 140         int ret = 0;
 141         LIST_HEAD(to_be_dropped);
 142
 143 restart:
 144
 145         /*
 146          * sendmsg calls here after having queued its message on the send
 147          * queue.  We only have one task feeding the connection at a time.  If
 148          * another thread is already feeding the queue then we back off.  This
 149          * avoids blocking the caller and trading per-connection data between
 150          * caches per message.
 151          */
 152         if (!acquire_in_xmit(conn)) {
 153                 rds_stats_inc(s_send_lock_contention);
 154                 ret = -ENOMEM;
 155                 goto out;
 156         }
 157
 158         /*
 159          * rds_conn_shutdown() sets the conn state and then tests RDS_IN_XMIT,
 160          * we do the opposite to avoid races.
 161          */
 162         if (!rds_conn_up(conn)) {
 163                 release_in_xmit(conn);
 164                 ret = 0;
 165                 goto out;
 166         }
 167
 168         if (conn->c_trans->xmit_prepare)
 169                 conn->c_trans->xmit_prepare(conn);
 170
 171         /*
 172          * spin trying to push headers and data down the connection until
 173          * the connection doesn't make forward progress.
 174          */
 175         while (1) {
 176
 177                 rm = conn->c_xmit_rm;
 178
 179                 /*
 180                  * If between sending messages, we can send a pending congestion
 181                  * map update.
 182                  */
 183                 if (!rm && test_and_clear_bit(0, &conn->c_map_queued)) {
 184                         rm = rds_cong_update_alloc(conn);
 185                         if (IS_ERR(rm)) {
 186                                 ret = PTR_ERR(rm);
 187                                 break;
 188                         }
 189                         rm->data.op_active = 1;
 190
 191                         conn->c_xmit_rm = rm;
 192                 }
 193
 194                 /*
 195                  * If not already working on one, grab the next message.
 196                  *
 197                  * c_xmit_rm holds a ref while we're sending this message down
 198                  * the connction.  We can use this ref while holding the
 199                  * send_sem.. rds_send_reset() is serialized with it.
 200                  */
 201                 if (!rm) {
 202                         unsigned int len;
 203
 204                         spin_lock_irqsave(&conn->c_lock, flags);
 205
 206                         if (!list_empty(&conn->c_send_queue)) {
 207                                 rm = list_entry(conn->c_send_queue.next,
 208                                                 struct rds_message,
 209                                                 m_conn_item);
 210                                 rds_message_addref(rm);
 211
 212                                 /*
 213                                  * Move the message from the send queue to the retransmit
 214                                  * list right away.
 215                                  */
 216                                 list_move_tail(&rm->m_conn_item, &conn->c_retrans);
 217                         }
 218
 219                         spin_unlock_irqrestore(&conn->c_lock, flags);
 220
 221                         if (!rm)
 222                                 break;
 223
 224                         /* Unfortunately, the way Infiniband deals with
 225                          * RDMA to a bad MR key is by moving the entire
 226                          * queue pair to error state. We cold possibly
 227                          * recover from that, but right now we drop the
 228                          * connection.
 229                          * Therefore, we never retransmit messages with RDMA ops.
 230                          */
 231                         if (rm->rdma.op_active &&
 232                             test_bit(RDS_MSG_RETRANSMITTED, &rm->m_flags)) {
 233                                 spin_lock_irqsave(&conn->c_lock, flags);
 234                                 if (test_and_clear_bit(RDS_MSG_ON_CONN, &rm->m_flags))
 235                                         list_move(&rm->m_conn_item, &to_be_dropped);
 236                                 spin_unlock_irqrestore(&conn->c_lock, flags);
 237                                 continue;
 238                         }
 239
 240                         /* Require an ACK every once in a while */
 241                         len = ntohl(rm->m_inc.i_hdr.h_len);
 242                         if (conn->c_unacked_packets == 0 ||
 243                             conn->c_unacked_bytes < len) {
 244                                 __set_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags);
 245
 246                                 conn->c_unacked_packets = rds_sysctl_max_unacked_packets;
 247                                 conn->c_unacked_bytes = rds_sysctl_max_unacked_bytes;
 248                                 rds_stats_inc(s_send_ack_required);
 249                         } else {
 250                                 conn->c_unacked_bytes -= len;
 251                                 conn->c_unacked_packets--;
 252                         }
 253
 254                         conn->c_xmit_rm = rm;
 255                 }
 256
 257                 /* The transport either sends the whole rdma or none of it */
 258                 if (rm->rdma.op_active && !conn->c_xmit_rdma_sent) {
 259                         rm->m_final_op = &rm->rdma;
 260                         ret = conn->c_trans->xmit_rdma(conn, &rm->rdma);
 261                         if (ret)
 262                                 break;
 263                         conn->c_xmit_rdma_sent = 1;
 264
 265                         /* The transport owns the mapped memory for now.
 266                          * You can't unmap it while it's on the send queue */
 267                         set_bit(RDS_MSG_MAPPED, &rm->m_flags);
 268                 }
 269
 270                 if (rm->atomic.op_active && !conn->c_xmit_atomic_sent) {
 271                         rm->m_final_op = &rm->atomic;
 272                         ret = conn->c_trans->xmit_atomic(conn, &rm->atomic);
 273                         if (ret)
 274                                 break;
 275                         conn->c_xmit_atomic_sent = 1;
 276
 277                         /* The transport owns the mapped memory for now.
 278                          * You can't unmap it while it's on the send queue */
 279                         set_bit(RDS_MSG_MAPPED, &rm->m_flags);
 280                 }
 281
 282                 /*
 283                  * A number of cases require an RDS header to be sent
 284                  * even if there is no data.
 285                  * We permit 0-byte sends; rds-ping depends on this.
 286                  * However, if there are exclusively attached silent ops,
 287                  * we skip the hdr/data send, to enable silent operation.
 288                  */
 289                 if (rm->data.op_nents == 0) {
 290                         int ops_present;
 291                         int all_ops_are_silent = 1;
 292
 293                         ops_present = (rm->atomic.op_active || rm->rdma.op_active);
 294                         if (rm->atomic.op_active && !rm->atomic.op_silent)
 295                                 all_ops_are_silent = 0;
 296                         if (rm->rdma.op_active && !rm->rdma.op_silent)
 297                                 all_ops_are_silent = 0;
 298
 299                         if (ops_present && all_ops_are_silent
 300                             && !rm->m_rdma_cookie)
 301                                 rm->data.op_active = 0;
 302                 }
 303
 304                 if (rm->data.op_active && !conn->c_xmit_data_sent) {
 305                         rm->m_final_op = &rm->data;
 306                         ret = conn->c_trans->xmit(conn, rm,
 307                                                   conn->c_xmit_hdr_off,
 308                                                   conn->c_xmit_sg,
 309                                                   conn->c_xmit_data_off);
 310                         if (ret <= 0)
 311                                 break;
 312
 313                         if (conn->c_xmit_hdr_off < sizeof(struct rds_header)) {
 314                                 tmp = min_t(int, ret,
 315                                             sizeof(struct rds_header) -
 316                                             conn->c_xmit_hdr_off);
 317                                 conn->c_xmit_hdr_off += tmp;
 318                                 ret -= tmp;
 319                         }
 320
 321                         sg = &rm->data.op_sg[conn->c_xmit_sg];
 322                         while (ret) {
 323                                 tmp = min_t(int, ret, sg->length -
 324                                                       conn->c_xmit_data_off);
 325                                 conn->c_xmit_data_off += tmp;
 326                                 ret -= tmp;
 327                                 if (conn->c_xmit_data_off == sg->length) {
 328                                         conn->c_xmit_data_off = 0;
 329                                         sg++;
 330                                         conn->c_xmit_sg++;
 331                                         BUG_ON(ret != 0 &&
 332                                                conn->c_xmit_sg == rm->data.op_nents);
 333                                 }
 334                         }
 335
 336                         if (conn->c_xmit_hdr_off == sizeof(struct rds_header) &&
 337                             (conn->c_xmit_sg == rm->data.op_nents))
 338                                 conn->c_xmit_data_sent = 1;
 339                 }
 340
 341                 /*
 342                  * A rm will only take multiple times through this loop
 343                  * if there is a data op. Thus, if the data is sent (or there was
 344                  * none), then we're done with the rm.
 345                  */
 346                 if (!rm->data.op_active || conn->c_xmit_data_sent) {
 347                         conn->c_xmit_rm = NULL;
 348                         conn->c_xmit_sg = 0;
 349                         conn->c_xmit_hdr_off = 0;
 350                         conn->c_xmit_data_off = 0;
 351                         conn->c_xmit_rdma_sent = 0;
 352                         conn->c_xmit_atomic_sent = 0;
 353                         conn->c_xmit_data_sent = 0;
 354
 355                         rds_message_put(rm);
 356                 }
 357         }
 358
 359         if (conn->c_trans->xmit_complete)
 360                 conn->c_trans->xmit_complete(conn);
 361
 362         release_in_xmit(conn);
 363
 364         /* Nuke any messages we decided not to retransmit. */
 365         if (!list_empty(&to_be_dropped)) {
 366                 /* irqs on here, so we can put(), unlike above */
 367                 list_for_each_entry(rm, &to_be_dropped, m_conn_item)
 368                         rds_message_put(rm);
 369                 rds_send_remove_from_sock(&to_be_dropped, RDS_RDMA_DROPPED);
 370         }
 371
 372         /*
 373          * Other senders can queue a message after we last test the send queue
 374          * but before we clear RDS_IN_XMIT.  In that case they'd back off and
 375          * not try and send their newly queued message.  We need to check the
 376          * send queue after having cleared RDS_IN_XMIT so that their message
 377          * doesn't get stuck on the send queue.
 378          *
 379          * If the transport cannot continue (i.e ret != 0), then it must
 380          * call us when more room is available, such as from the tx
 381          * completion handler.
 382          */
 383         if (ret == 0) {
 384                 smp_mb();
 385                 if (!list_empty(&conn->c_send_queue)) {
 386                         rds_stats_inc(s_send_lock_queue_raced);
 387                         goto restart;
 388                 }
 389         }
 390 out:
 391         return ret;
 392 }
 393
 394 static void rds_send_sndbuf_remove(struct rds_sock *rs, struct rds_message *rm)
 395 {
 396         u32 len = be32_to_cpu(rm->m_inc.i_hdr.h_len);
 397
 398         assert_spin_locked(&rs->rs_lock);
 399
 400         BUG_ON(rs->rs_snd_bytes < len);
 401         rs->rs_snd_bytes -= len;
 402
 403         if (rs->rs_snd_bytes == 0)
 404                 rds_stats_inc(s_send_queue_empty);
 405 }
 406
 407 static inline int rds_send_is_acked(struct rds_message *rm, u64 ack,
 408                                     is_acked_func is_acked)
 409 {
 410         if (is_acked)
 411                 return is_acked(rm, ack);
 412         return be64_to_cpu(rm->m_inc.i_hdr.h_sequence) <= ack;
 413 }
 414
 415 /*
 416  * This is pretty similar to what happens below in the ACK
 417  * handling code - except that we call here as soon as we get
 418  * the IB send completion on the RDMA op and the accompanying
 419  * message.
 420  */
 421 void rds_rdma_send_complete(struct rds_message *rm, int status)
 422 {
 423         struct rds_sock *rs = NULL;
 424         struct rm_rdma_op *ro;
 425         struct rds_notifier *notifier;
 426         unsigned long flags;
 427
 428         spin_lock_irqsave(&rm->m_rs_lock, flags);
 429
 430         ro = &rm->rdma;
 431         if (test_bit(RDS_MSG_ON_SOCK, &rm->m_flags) &&
 432             ro->op_active && ro->op_notify && ro->op_notifier) {
 433                 notifier = ro->op_notifier;
 434                 rs = rm->m_rs;
 435                 sock_hold(rds_rs_to_sk(rs));
 436
 437                 notifier->n_status = status;
 438                 spin_lock(&rs->rs_lock);
 439                 list_add_tail(&notifier->n_list, &rs->rs_notify_queue);
 440                 spin_unlock(&rs->rs_lock);
 441
 442                 ro->op_notifier = NULL;
 443         }
 444
 445         spin_unlock_irqrestore(&rm->m_rs_lock, flags);
 446
 447         if (rs) {
 448                 rds_wake_sk_sleep(rs);
 449                 sock_put(rds_rs_to_sk(rs));
 450         }
 451 }
 452 EXPORT_SYMBOL_GPL(rds_rdma_send_complete);
 453
 454 /*
 455  * Just like above, except looks at atomic op
 456  */
 457 void rds_atomic_send_complete(struct rds_message *rm, int status)
 458 {
 459         struct rds_sock *rs = NULL;
 460         struct rm_atomic_op *ao;
 461         struct rds_notifier *notifier;
 462         unsigned long flags;
 463
 464         spin_lock_irqsave(&rm->m_rs_lock, flags);
 465
 466         ao = &rm->atomic;
 467         if (test_bit(RDS_MSG_ON_SOCK, &rm->m_flags)
 468             && ao->op_active && ao->op_notify && ao->op_notifier) {
 469                 notifier = ao->op_notifier;
 470                 rs = rm->m_rs;
 471                 sock_hold(rds_rs_to_sk(rs));
 472
 473                 notifier->n_status = status;
 474                 spin_lock(&rs->rs_lock);
 475                 list_add_tail(&notifier->n_list, &rs->rs_notify_queue);
 476                 spin_unlock(&rs->rs_lock);
 477
 478                 ao->op_notifier = NULL;
 479         }
 480
 481         spin_unlock_irqrestore(&rm->m_rs_lock, flags);
 482
 483         if (rs) {
 484                 rds_wake_sk_sleep(rs);
 485                 sock_put(rds_rs_to_sk(rs));
 486         }
 487 }
 488 EXPORT_SYMBOL_GPL(rds_atomic_send_complete);
 489
 490 /*
 491  * This is the same as rds_rdma_send_complete except we
 492  * don't do any locking - we have all the ingredients (message,
 493  * socket, socket lock) and can just move the notifier.
 494  */
 495 static inline void
 496 __rds_send_complete(struct rds_sock *rs, struct rds_message *rm, int status)
 497 {
 498         struct rm_rdma_op *ro;
 499         struct rm_atomic_op *ao;
 500
 501         ro = &rm->rdma;
 502         if (ro->op_active && ro->op_notify && ro->op_notifier) {
 503                 ro->op_notifier->n_status = status;
 504                 list_add_tail(&ro->op_notifier->n_list, &rs->rs_notify_queue);
 505                 ro->op_notifier = NULL;
 506         }
 507
 508         ao = &rm->atomic;
 509         if (ao->op_active && ao->op_notify && ao->op_notifier) {
 510                 ao->op_notifier->n_status = status;
 511                 list_add_tail(&ao->op_notifier->n_list, &rs->rs_notify_queue);
 512                 ao->op_notifier = NULL;
 513         }
 514
 515         /* No need to wake the app - caller does this */
 516 }
 517
 518 /*
 519  * This is called from the IB send completion when we detect
 520  * a RDMA operation that failed with remote access error.
 521  * So speed is not an issue here.
 522  */
 523 struct rds_message *rds_send_get_message(struct rds_connection *conn,
 524                                          struct rm_rdma_op *op)
 525 {
 526         struct rds_message *rm, *tmp, *found = NULL;
 527         unsigned long flags;
 528
 529         spin_lock_irqsave(&conn->c_lock, flags);
 530
 531         list_for_each_entry_safe(rm, tmp, &conn->c_retrans, m_conn_item) {
 532                 if (&rm->rdma == op) {
 533                         atomic_inc(&rm->m_refcount);
 534                         found = rm;
 535                         goto out;
 536                 }
 537         }
 538
 539         list_for_each_entry_safe(rm, tmp, &conn->c_send_queue, m_conn_item) {
 540                 if (&rm->rdma == op) {
 541                         atomic_inc(&rm->m_refcount);
 542                         found = rm;
 543                         break;
 544                 }
 545         }
 546
 547 out:
 548         spin_unlock_irqrestore(&conn->c_lock, flags);
 549
 550         return found;
 551 }
 552 EXPORT_SYMBOL_GPL(rds_send_get_message);
 553
 554 /*
 555  * This removes messages from the socket's list if they're on it.  The list
 556  * argument must be private to the caller, we must be able to modify it
 557  * without locks.  The messages must have a reference held for their
 558  * position on the list.  This function will drop that reference after
 559  * removing the messages from the 'messages' list regardless of if it found
 560  * the messages on the socket list or not.
 561  */
 562 static void rds_send_remove_from_sock(struct list_head *messages, int status)
 563 {
 564         unsigned long flags;
 565         struct rds_sock *rs = NULL;
 566         struct rds_message *rm;
 567
 568         while (!list_empty(messages)) {
 569                 int was_on_sock = 0;
 570
 571                 rm = list_entry(messages->next, struct rds_message,
 572                                 m_conn_item);
 573                 list_del_init(&rm->m_conn_item);
 574
 575                 /*
 576                  * If we see this flag cleared then we're *sure* that someone
 577                  * else beat us to removing it from the sock.  If we race
 578                  * with their flag update we'll get the lock and then really
 579                  * see that the flag has been cleared.
 580                  *
 581                  * The message spinlock makes sure nobody clears rm->m_rs
 582                  * while we're messing with it. It does not prevent the
 583                  * message from being removed from the socket, though.
 584                  */
 585                 spin_lock_irqsave(&rm->m_rs_lock, flags);
 586                 if (!test_bit(RDS_MSG_ON_SOCK, &rm->m_flags))
 587                         goto unlock_and_drop;
 588
 589                 if (rs != rm->m_rs) {
 590                         if (rs) {
 591                                 rds_wake_sk_sleep(rs);
 592                                 sock_put(rds_rs_to_sk(rs));
 593                         }
 594                         rs = rm->m_rs;
 595                         sock_hold(rds_rs_to_sk(rs));
 596                 }
 597                 spin_lock(&rs->rs_lock);
 598
 599                 if (test_and_clear_bit(RDS_MSG_ON_SOCK, &rm->m_flags)) {
 600                         struct rm_rdma_op *ro = &rm->rdma;
 601                         struct rds_notifier *notifier;
 602
 603                         list_del_init(&rm->m_sock_item);
 604                         rds_send_sndbuf_remove(rs, rm);
 605
 606                         if (ro->op_active && ro->op_notifier &&
 607                                (ro->op_notify || (ro->op_recverr && status))) {
 608                                 notifier = ro->op_notifier;
 609                                 list_add_tail(&notifier->n_list,
 610                                                 &rs->rs_notify_queue);
 611                                 if (!notifier->n_status)
 612                                         notifier->n_status = status;
 613                                 rm->rdma.op_notifier = NULL;
 614                         }
 615                         was_on_sock = 1;
 616                         rm->m_rs = NULL;
 617                 }
 618                 spin_unlock(&rs->rs_lock);
 619
 620 unlock_and_drop:
 621                 spin_unlock_irqrestore(&rm->m_rs_lock, flags);
 622                 rds_message_put(rm);
 623                 if (was_on_sock)
 624                         rds_message_put(rm);
 625         }
 626
 627         if (rs) {
 628                 rds_wake_sk_sleep(rs);
 629                 sock_put(rds_rs_to_sk(rs));
 630         }
 631 }
 632
 633 /*
 634  * Transports call here when they've determined that the receiver queued
 635  * messages up to, and including, the given sequence number.  Messages are
 636  * moved to the retrans queue when rds_send_xmit picks them off the send
 637  * queue. This means that in the TCP case, the message may not have been
 638  * assigned the m_ack_seq yet - but that's fine as long as tcp_is_acked
 639  * checks the RDS_MSG_HAS_ACK_SEQ bit.
 640  *
 641  * XXX It's not clear to me how this is safely serialized with socket
 642  * destruction.  Maybe it should bail if it sees SOCK_DEAD.
 643  */
 644 void rds_send_drop_acked(struct rds_connection *conn, u64 ack,
 645                          is_acked_func is_acked)
 646 {
 647         struct rds_message *rm, *tmp;
 648         unsigned long flags;
 649         LIST_HEAD(list);
 650
 651         spin_lock_irqsave(&conn->c_lock, flags);
 652
 653         list_for_each_entry_safe(rm, tmp, &conn->c_retrans, m_conn_item) {
 654                 if (!rds_send_is_acked(rm, ack, is_acked))
 655                         break;
 656
 657                 list_move(&rm->m_conn_item, &list);
 658                 clear_bit(RDS_MSG_ON_CONN, &rm->m_flags);
 659         }
 660
 661         /* order flag updates with spin locks */
 662         if (!list_empty(&list))
 663                 smp_mb__after_clear_bit();
 664
 665         spin_unlock_irqrestore(&conn->c_lock, flags);
 666
 667         /* now remove the messages from the sock list as needed */
 668         rds_send_remove_from_sock(&list, RDS_RDMA_SUCCESS);
 669 }
 670 EXPORT_SYMBOL_GPL(rds_send_drop_acked);
 671
 672 void rds_send_drop_to(struct rds_sock *rs, struct sockaddr_in *dest)
 673 {
 674         struct rds_message *rm, *tmp;
 675         struct rds_connection *conn;
 676         unsigned long flags;
 677         LIST_HEAD(list);
 678
 679         /* get all the messages we're dropping under the rs lock */
 680         spin_lock_irqsave(&rs->rs_lock, flags);
 681
 682         list_for_each_entry_safe(rm, tmp, &rs->rs_send_queue, m_sock_item) {
 683                 if (dest && (dest->sin_addr.s_addr != rm->m_daddr ||
 684                              dest->sin_port != rm->m_inc.i_hdr.h_dport))
 685                         continue;
 686
 687                 list_move(&rm->m_sock_item, &list);
 688                 rds_send_sndbuf_remove(rs, rm);
 689                 clear_bit(RDS_MSG_ON_SOCK, &rm->m_flags);
 690         }
 691
 692         /* order flag updates with the rs lock */
 693         smp_mb__after_clear_bit();
 694
 695         spin_unlock_irqrestore(&rs->rs_lock, flags);
 696
 697         if (list_empty(&list))
 698                 return;
 699
 700         /* Remove the messages from the conn */
 701         list_for_each_entry(rm, &list, m_sock_item) {
 702
 703                 conn = rm->m_inc.i_conn;
 704
 705                 spin_lock_irqsave(&conn->c_lock, flags);
 706                 /*
 707                  * Maybe someone else beat us to removing rm from the conn.
 708                  * If we race with their flag update we'll get the lock and
 709                  * then really see that the flag has been cleared.
 710                  */
 711                 if (!test_and_clear_bit(RDS_MSG_ON_CONN, &rm->m_flags)) {
 712                         spin_unlock_irqrestore(&conn->c_lock, flags);
 713                         continue;
 714                 }
 715                 list_del_init(&rm->m_conn_item);
 716                 spin_unlock_irqrestore(&conn->c_lock, flags);
 717
 718                 /*
 719                  * Couldn't grab m_rs_lock in top loop (lock ordering),
 720                  * but we can now.
 721                  */
 722                 spin_lock_irqsave(&rm->m_rs_lock, flags);
 723
 724                 spin_lock(&rs->rs_lock);
 725                 __rds_send_complete(rs, rm, RDS_RDMA_CANCELED);
 726                 spin_unlock(&rs->rs_lock);
 727
 728                 rm->m_rs = NULL;
 729                 spin_unlock_irqrestore(&rm->m_rs_lock, flags);
 730
 731                 rds_message_put(rm);
 732         }
 733
 734         rds_wake_sk_sleep(rs);
 735
 736         while (!list_empty(&list)) {
 737                 rm = list_entry(list.next, struct rds_message, m_sock_item);
 738                 list_del_init(&rm->m_sock_item);
 739
 740                 rds_message_wait(rm);
 741                 rds_message_put(rm);
 742         }
 743 }
 744
 745 /*
 746  * we only want this to fire once so we use the callers 'queued'.  It's
 747  * possible that another thread can race with us and remove the
 748  * message from the flow with RDS_CANCEL_SENT_TO.
 749  */
 750 static int rds_send_queue_rm(struct rds_sock *rs, struct rds_connection *conn,
 751                              struct rds_message *rm, __be16 sport,
 752                              __be16 dport, int *queued)
 753 {
 754         unsigned long flags;
 755         u32 len;
 756
 757         if (*queued)
 758                 goto out;
 759
 760         len = be32_to_cpu(rm->m_inc.i_hdr.h_len);
 761
 762         /* this is the only place which holds both the socket's rs_lock
 763          * and the connection's c_lock */
 764         spin_lock_irqsave(&rs->rs_lock, flags);
 765
 766         /*
 767          * If there is a little space in sndbuf, we don't queue anything,
 768          * and userspace gets -EAGAIN. But poll() indicates there's send
 769          * room. This can lead to bad behavior (spinning) if snd_bytes isn't
 770          * freed up by incoming acks. So we check the *old* value of
 771          * rs_snd_bytes here to allow the last msg to exceed the buffer,
 772          * and poll() now knows no more data can be sent.
 773          */
 774         if (rs->rs_snd_bytes < rds_sk_sndbuf(rs)) {
 775                 rs->rs_snd_bytes += len;
 776
 777                 /* let recv side know we are close to send space exhaustion.
 778                  * This is probably not the optimal way to do it, as this
 779                  * means we set the flag on *all* messages as soon as our
 780                  * throughput hits a certain threshold.
 781                  */
 782                 if (rs->rs_snd_bytes >= rds_sk_sndbuf(rs) / 2)
 783                         __set_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags);
 784
 785                 list_add_tail(&rm->m_sock_item, &rs->rs_send_queue);
 786                 set_bit(RDS_MSG_ON_SOCK, &rm->m_flags);
 787                 rds_message_addref(rm);
 788                 rm->m_rs = rs;
 789
 790                 /* The code ordering is a little weird, but we're
 791                    trying to minimize the time we hold c_lock */
 792                 rds_message_populate_header(&rm->m_inc.i_hdr, sport, dport, 0);
 793                 rm->m_inc.i_conn = conn;
 794                 rds_message_addref(rm);
 795
 796                 spin_lock(&conn->c_lock);
 797                 rm->m_inc.i_hdr.h_sequence = cpu_to_be64(conn->c_next_tx_seq++);
 798                 list_add_tail(&rm->m_conn_item, &conn->c_send_queue);
 799                 set_bit(RDS_MSG_ON_CONN, &rm->m_flags);
 800                 spin_unlock(&conn->c_lock);
 801
 802                 rdsdebug("queued msg %p len %d, rs %p bytes %d seq %llu\n",
 803                          rm, len, rs, rs->rs_snd_bytes,
 804                          (unsigned long long)be64_to_cpu(rm->m_inc.i_hdr.h_sequence));
 805
 806                 *queued = 1;
 807         }
 808
 809         spin_unlock_irqrestore(&rs->rs_lock, flags);
 810 out:
 811         return *queued;
 812 }
 813
 814 /*
 815  * rds_message is getting to be quite complicated, and we'd like to allocate
 816  * it all in one go. This figures out how big it needs to be up front.
 817  */
 818 static int rds_rm_size(struct msghdr *msg, int data_len)
 819 {
 820         struct cmsghdr *cmsg;
 821         int size = 0;
 822         int cmsg_groups = 0;
 823         int retval;
 824
 825         for (cmsg = CMSG_FIRSTHDR(msg); cmsg; cmsg = CMSG_NXTHDR(msg, cmsg)) {
 826                 if (!CMSG_OK(msg, cmsg))
 827                         return -EINVAL;
 828
 829                 if (cmsg->cmsg_level != SOL_RDS)
 830                         continue;
 831
 832                 switch (cmsg->cmsg_type) {
 833                 case RDS_CMSG_RDMA_ARGS:
 834                         cmsg_groups |= 1;
 835                         retval = rds_rdma_extra_size(CMSG_DATA(cmsg));
 836                         if (retval < 0)
 837                                 return retval;
 838                         size += retval;
 839
 840                         break;
 841
 842                 case RDS_CMSG_RDMA_DEST:
 843                 case RDS_CMSG_RDMA_MAP:
 844                         cmsg_groups |= 2;
 845                         /* these are valid but do no add any size */
 846                         break;
 847
 848                 case RDS_CMSG_ATOMIC_CSWP:
 849                 case RDS_CMSG_ATOMIC_FADD:
 850                 case RDS_CMSG_MASKED_ATOMIC_CSWP:
 851                 case RDS_CMSG_MASKED_ATOMIC_FADD:
 852                         cmsg_groups |= 1;
 853                         size += sizeof(struct scatterlist);
 854                         break;
 855
 856                 default:
 857                         return -EINVAL;
 858                 }
 859
 860         }
 861
 862         size += ceil(data_len, PAGE_SIZE) * sizeof(struct scatterlist);
 863
 864         /* Ensure (DEST, MAP) are never used with (ARGS, ATOMIC) */
 865         if (cmsg_groups == 3)
 866                 return -EINVAL;
 867
 868         return size;
 869 }
 870
 871 static int rds_cmsg_send(struct rds_sock *rs, struct rds_message *rm,
 872                          struct msghdr *msg, int *allocated_mr)
 873 {
 874         struct cmsghdr *cmsg;
 875         int ret = 0;
 876
 877         for (cmsg = CMSG_FIRSTHDR(msg); cmsg; cmsg = CMSG_NXTHDR(msg, cmsg)) {
 878                 if (!CMSG_OK(msg, cmsg))
 879                         return -EINVAL;
 880
 881                 if (cmsg->cmsg_level != SOL_RDS)
 882                         continue;
 883
 884                 /* As a side effect, RDMA_DEST and RDMA_MAP will set
 885                  * rm->rdma.m_rdma_cookie and rm->rdma.m_rdma_mr.
 886                  */
 887                 switch (cmsg->cmsg_type) {
 888                 case RDS_CMSG_RDMA_ARGS:
 889                         ret = rds_cmsg_rdma_args(rs, rm, cmsg);
 890                         break;
 891
 892                 case RDS_CMSG_RDMA_DEST:
 893                         ret = rds_cmsg_rdma_dest(rs, rm, cmsg);
 894                         break;
 895
 896                 case RDS_CMSG_RDMA_MAP:
 897                         ret = rds_cmsg_rdma_map(rs, rm, cmsg);
 898                         if (!ret)
 899                                 *allocated_mr = 1;
 900                         break;
 901                 case RDS_CMSG_ATOMIC_CSWP:
 902                 case RDS_CMSG_ATOMIC_FADD:
 903                 case RDS_CMSG_MASKED_ATOMIC_CSWP:
 904                 case RDS_CMSG_MASKED_ATOMIC_FADD:
 905                         ret = rds_cmsg_atomic(rs, rm, cmsg);
 906                         break;
 907
 908                 default:
 909                         return -EINVAL;
 910                 }
 911
 912                 if (ret)
 913                         break;
 914         }
 915
 916         return ret;
 917 }
 918
 919 int rds_sendmsg(struct kiocb *iocb, struct socket *sock, struct msghdr *msg,
 920                 size_t payload_len)
 921 {
 922         struct sock *sk = sock->sk;
 923         struct rds_sock *rs = rds_sk_to_rs(sk);
 924         struct sockaddr_in *usin = (struct sockaddr_in *)msg->msg_name;
 925         __be32 daddr;
 926         __be16 dport;
 927         struct rds_message *rm = NULL;
 928         struct rds_connection *conn;
 929         int ret = 0;
 930         int queued = 0, allocated_mr = 0;
 931         int nonblock = msg->msg_flags & MSG_DONTWAIT;
 932         long timeo = sock_sndtimeo(sk, nonblock);
 933
 934         /* Mirror Linux UDP mirror of BSD error message compatibility */
 935         /* XXX: Perhaps MSG_MORE someday */
 936         if (msg->msg_flags & ~(MSG_DONTWAIT | MSG_CMSG_COMPAT)) {
 937                 printk(KERN_INFO "msg_flags 0x%08X\n", msg->msg_flags);
 938                 ret = -EOPNOTSUPP;
 939                 goto out;
 940         }
 941
 942         if (msg->msg_namelen) {
 943                 /* XXX fail non-unicast destination IPs? */
 944                 if (msg->msg_namelen < sizeof(*usin) || usin->sin_family != AF_INET) {
 945                         ret = -EINVAL;
 946                         goto out;
 947                 }
 948                 daddr = usin->sin_addr.s_addr;
 949                 dport = usin->sin_port;
 950         } else {
 951                 /* We only care about consistency with ->connect() */
 952                 lock_sock(sk);
 953                 daddr = rs->rs_conn_addr;
 954                 dport = rs->rs_conn_port;
 955                 release_sock(sk);
 956         }
 957
 958         /* racing with another thread binding seems ok here */
 959         if (daddr == 0 || rs->rs_bound_addr == 0) {
 960                 ret = -ENOTCONN; /* XXX not a great errno */
 961                 goto out;
 962         }
 963
 964         /* size of rm including all sgs */
 965         ret = rds_rm_size(msg, payload_len);
 966         if (ret < 0)
 967                 goto out;
 968
 969         rm = rds_message_alloc(ret, GFP_KERNEL);
 970         if (!rm) {
 971                 ret = -ENOMEM;
 972                 goto out;
 973         }
 974
 975         /* Attach data to the rm */
 976         if (payload_len) {
 977                 rm->data.op_sg = rds_message_alloc_sgs(rm, ceil(payload_len, PAGE_SIZE));
 978                 if (!rm->data.op_sg) {
 979                         ret = -ENOMEM;
 980                         goto out;
 981                 }
 982                 ret = rds_message_copy_from_user(rm, msg->msg_iov, payload_len);
 983                 if (ret)
 984                         goto out;
 985         }
 986         rm->data.op_active = 1;
 987
 988         rm->m_daddr = daddr;
 989
 990         /* rds_conn_create has a spinlock that runs with IRQ off.
 991          * Caching the conn in the socket helps a lot. */
 992         if (rs->rs_conn && rs->rs_conn->c_faddr == daddr)
 993                 conn = rs->rs_conn;
 994         else {
 995                 conn = rds_conn_create_outgoing(rs->rs_bound_addr, daddr,
 996                                         rs->rs_transport,
 997                                         sock->sk->sk_allocation);
 998                 if (IS_ERR(conn)) {
 999                         ret = PTR_ERR(conn);
1000                         goto out;
1001                 }
1002                 rs->rs_conn = conn;
1003         }
1004
1005         /* Parse any control messages the user may have included. */
1006         ret = rds_cmsg_send(rs, rm, msg, &allocated_mr);
1007         if (ret)
1008                 goto out;
1009
1010         if (rm->rdma.op_active && !conn->c_trans->xmit_rdma) {
1011                 printk_ratelimited(KERN_NOTICE "rdma_op %p conn xmit_rdma %p\n",
1012                                &rm->rdma, conn->c_trans->xmit_rdma);
1013                 ret = -EOPNOTSUPP;
1014                 goto out;
1015         }
1016
1017         if (rm->atomic.op_active && !conn->c_trans->xmit_atomic) {
1018                 printk_ratelimited(KERN_NOTICE "atomic_op %p conn xmit_atomic %p\n",
1019                                &rm->atomic, conn->c_trans->xmit_atomic);
1020                 ret = -EOPNOTSUPP;
1021                 goto out;
1022         }
1023
1024         rds_conn_connect_if_down(conn);
1025
1026         ret = rds_cong_wait(conn->c_fcong, dport, nonblock, rs);
1027         if (ret) {
1028                 rs->rs_seen_congestion = 1;
1029                 goto out;
1030         }
1031
1032         while (!rds_send_queue_rm(rs, conn, rm, rs->rs_bound_port,
1033                                   dport, &queued)) {
1034                 rds_stats_inc(s_send_queue_full);
1035                 /* XXX make sure this is reasonable */
1036                 if (payload_len > rds_sk_sndbuf(rs)) {
1037                         ret = -EMSGSIZE;
1038                         goto out;
1039                 }
1040                 if (nonblock) {
1041                         ret = -EAGAIN;
1042                         goto out;
1043                 }
1044
1045                 timeo = wait_event_interruptible_timeout(*sk_sleep(sk),
1046                                         rds_send_queue_rm(rs, conn, rm,
1047                                                           rs->rs_bound_port,
1048                                                           dport,
1049                                                           &queued),
1050                                         timeo);
1051                 rdsdebug("sendmsg woke queued %d timeo %ld\n", queued, timeo);
1052                 if (timeo > 0 || timeo == MAX_SCHEDULE_TIMEOUT)
1053                         continue;
1054
1055                 ret = timeo;
1056                 if (ret == 0)
1057                         ret = -ETIMEDOUT;
1058                 goto out;
1059         }
1060
1061         /*
1062          * By now we've committed to the send.  We reuse rds_send_worker()
1063          * to retry sends in the rds thread if the transport asks us to.
1064          */
1065         rds_stats_inc(s_send_queued);
1066
1067         if (!test_bit(RDS_LL_SEND_FULL, &conn->c_flags))
1068                 rds_send_xmit(conn);
1069
1070         rds_message_put(rm);
1071         return payload_len;
1072
1073 out:
1074         /* If the user included a RDMA_MAP cmsg, we allocated a MR on the fly.
1075          * If the sendmsg goes through, we keep the MR. If it fails with EAGAIN
1076          * or in any other way, we need to destroy the MR again */
1077         if (allocated_mr)
1078                 rds_rdma_unuse(rs, rds_rdma_cookie_key(rm->m_rdma_cookie), 1);
1079
1080         if (rm)
1081                 rds_message_put(rm);
1082         return ret;
1083 }
1084
1085 /*
1086  * Reply to a ping packet.
1087  */
1088 int
1089 rds_send_pong(struct rds_connection *conn, __be16 dport)
1090 {
1091         struct rds_message *rm;
1092         unsigned long flags;
1093         int ret = 0;
1094
1095         rm = rds_message_alloc(0, GFP_ATOMIC);
1096         if (!rm) {
1097                 ret = -ENOMEM;
1098                 goto out;
1099         }
1100
1101         rm->m_daddr = conn->c_faddr;
1102         rm->data.op_active = 1;
1103
1104         rds_conn_connect_if_down(conn);
1105
1106         ret = rds_cong_wait(conn->c_fcong, dport, 1, NULL);
1107         if (ret)
1108                 goto out;
1109
1110         spin_lock_irqsave(&conn->c_lock, flags);
1111         list_add_tail(&rm->m_conn_item, &conn->c_send_queue);
1112         set_bit(RDS_MSG_ON_CONN, &rm->m_flags);
1113         rds_message_addref(rm);
1114         rm->m_inc.i_conn = conn;
1115
1116         rds_message_populate_header(&rm->m_inc.i_hdr, 0, dport,
1117                                     conn->c_next_tx_seq);
1118         conn->c_next_tx_seq++;
1119         spin_unlock_irqrestore(&conn->c_lock, flags);
1120
1121         rds_stats_inc(s_send_queued);
1122         rds_stats_inc(s_send_pong);
1123
1124         if (!test_bit(RDS_LL_SEND_FULL, &conn->c_flags))
1125                 rds_send_xmit(conn);
1126
1127         rds_message_put(rm);
1128         return 0;
1129
1130 out:
1131         if (rm)
1132                 rds_message_put(rm);
1133         return ret;
1134 }