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 <net/sock.h> 35 #include <linux/in.h> 36 #include <linux/list.h> 37 38 #include "rds.h" 39 #include "rdma.h" 40 41 /* When transmitting messages in rds_send_xmit, we need to emerge from 42 * time to time and briefly release the CPU. Otherwise the softlock watchdog 43 * will kick our shin. 44 * Also, it seems fairer to not let one busy connection stall all the 45 * others. 46 * 47 * send_batch_count is the number of times we'll loop in send_xmit. Setting 48 * it to 0 will restore the old behavior (where we looped until we had 49 * drained the queue). 50 */ 51 static int send_batch_count = 64; 52 module_param(send_batch_count, int, 0444); 53 MODULE_PARM_DESC(send_batch_count, " batch factor when working the send queue"); 54 55 /* 56 * Reset the send state. Caller must hold c_send_lock when calling here. 57 */ 58 void rds_send_reset(struct rds_connection *conn) 59 { 60 struct rds_message *rm, *tmp; 61 unsigned long flags; 62 63 if (conn->c_xmit_rm) { 64 /* Tell the user the RDMA op is no longer mapped by the 65 * transport. This isn't entirely true (it's flushed out 66 * independently) but as the connection is down, there's 67 * no ongoing RDMA to/from that memory */ 68 rds_message_unmapped(conn->c_xmit_rm); 69 rds_message_put(conn->c_xmit_rm); 70 conn->c_xmit_rm = NULL; 71 } 72 conn->c_xmit_sg = 0; 73 conn->c_xmit_hdr_off = 0; 74 conn->c_xmit_data_off = 0; 75 conn->c_xmit_rdma_sent = 0; 76 77 conn->c_map_queued = 0; 78 79 conn->c_unacked_packets = rds_sysctl_max_unacked_packets; 80 conn->c_unacked_bytes = rds_sysctl_max_unacked_bytes; 81 82 /* Mark messages as retransmissions, and move them to the send q */ 83 spin_lock_irqsave(&conn->c_lock, flags); 84 list_for_each_entry_safe(rm, tmp, &conn->c_retrans, m_conn_item) { 85 set_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags); 86 set_bit(RDS_MSG_RETRANSMITTED, &rm->m_flags); 87 } 88 list_splice_init(&conn->c_retrans, &conn->c_send_queue); 89 spin_unlock_irqrestore(&conn->c_lock, flags); 90 } 91 92 /* 93 * We're making the concious trade-off here to only send one message 94 * down the connection at a time. 95 * Pro: 96 * - tx queueing is a simple fifo list 97 * - reassembly is optional and easily done by transports per conn 98 * - no per flow rx lookup at all, straight to the socket 99 * - less per-frag memory and wire overhead 100 * Con: 101 * - queued acks can be delayed behind large messages 102 * Depends: 103 * - small message latency is higher behind queued large messages 104 * - large message latency isn't starved by intervening small sends 105 */ 106 int rds_send_xmit(struct rds_connection *conn) 107 { 108 struct rds_message *rm; 109 unsigned long flags; 110 unsigned int tmp; 111 unsigned int send_quota = send_batch_count; 112 struct scatterlist *sg; 113 int ret = 0; 114 int was_empty = 0; 115 LIST_HEAD(to_be_dropped); 116 117 /* 118 * sendmsg calls here after having queued its message on the send 119 * queue. We only have one task feeding the connection at a time. If 120 * another thread is already feeding the queue then we back off. This 121 * avoids blocking the caller and trading per-connection data between 122 * caches per message. 123 * 124 * The sem holder will issue a retry if they notice that someone queued 125 * a message after they stopped walking the send queue but before they 126 * dropped the sem. 127 */ 128 if (!mutex_trylock(&conn->c_send_lock)) { 129 rds_stats_inc(s_send_sem_contention); 130 ret = -ENOMEM; 131 goto out; 132 } 133 134 if (conn->c_trans->xmit_prepare) 135 conn->c_trans->xmit_prepare(conn); 136 137 /* 138 * spin trying to push headers and data down the connection until 139 * the connection doens't make forward progress. 140 */ 141 while (--send_quota) { 142 /* 143 * See if need to send a congestion map update if we're 144 * between sending messages. The send_sem protects our sole 145 * use of c_map_offset and _bytes. 146 * Note this is used only by transports that define a special 147 * xmit_cong_map function. For all others, we create allocate 148 * a cong_map message and treat it just like any other send. 149 */ 150 if (conn->c_map_bytes) { 151 ret = conn->c_trans->xmit_cong_map(conn, conn->c_lcong, 152 conn->c_map_offset); 153 if (ret <= 0) 154 break; 155 156 conn->c_map_offset += ret; 157 conn->c_map_bytes -= ret; 158 if (conn->c_map_bytes) 159 continue; 160 } 161 162 /* If we're done sending the current message, clear the 163 * offset and S/G temporaries. 164 */ 165 rm = conn->c_xmit_rm; 166 if (rm != NULL && 167 conn->c_xmit_hdr_off == sizeof(struct rds_header) && 168 conn->c_xmit_sg == rm->m_nents) { 169 conn->c_xmit_rm = NULL; 170 conn->c_xmit_sg = 0; 171 conn->c_xmit_hdr_off = 0; 172 conn->c_xmit_data_off = 0; 173 conn->c_xmit_rdma_sent = 0; 174 175 /* Release the reference to the previous message. */ 176 rds_message_put(rm); 177 rm = NULL; 178 } 179 180 /* If we're asked to send a cong map update, do so. 181 */ 182 if (rm == NULL && test_and_clear_bit(0, &conn->c_map_queued)) { 183 if (conn->c_trans->xmit_cong_map != NULL) { 184 conn->c_map_offset = 0; 185 conn->c_map_bytes = sizeof(struct rds_header) + 186 RDS_CONG_MAP_BYTES; 187 continue; 188 } 189 190 rm = rds_cong_update_alloc(conn); 191 if (IS_ERR(rm)) { 192 ret = PTR_ERR(rm); 193 break; 194 } 195 196 conn->c_xmit_rm = rm; 197 } 198 199 /* 200 * Grab the next message from the send queue, if there is one. 201 * 202 * c_xmit_rm holds a ref while we're sending this message down 203 * the connction. We can use this ref while holding the 204 * send_sem.. rds_send_reset() is serialized with it. 205 */ 206 if (rm == NULL) { 207 unsigned int len; 208 209 spin_lock_irqsave(&conn->c_lock, flags); 210 211 if (!list_empty(&conn->c_send_queue)) { 212 rm = list_entry(conn->c_send_queue.next, 213 struct rds_message, 214 m_conn_item); 215 rds_message_addref(rm); 216 217 /* 218 * Move the message from the send queue to the retransmit 219 * list right away. 220 */ 221 list_move_tail(&rm->m_conn_item, &conn->c_retrans); 222 } 223 224 spin_unlock_irqrestore(&conn->c_lock, flags); 225 226 if (rm == NULL) { 227 was_empty = 1; 228 break; 229 } 230 231 /* Unfortunately, the way Infiniband deals with 232 * RDMA to a bad MR key is by moving the entire 233 * queue pair to error state. We cold possibly 234 * recover from that, but right now we drop the 235 * connection. 236 * Therefore, we never retransmit messages with RDMA ops. 237 */ 238 if (rm->m_rdma_op && 239 test_bit(RDS_MSG_RETRANSMITTED, &rm->m_flags)) { 240 spin_lock_irqsave(&conn->c_lock, flags); 241 if (test_and_clear_bit(RDS_MSG_ON_CONN, &rm->m_flags)) 242 list_move(&rm->m_conn_item, &to_be_dropped); 243 spin_unlock_irqrestore(&conn->c_lock, flags); 244 rds_message_put(rm); 245 continue; 246 } 247 248 /* Require an ACK every once in a while */ 249 len = ntohl(rm->m_inc.i_hdr.h_len); 250 if (conn->c_unacked_packets == 0 || 251 conn->c_unacked_bytes < len) { 252 __set_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags); 253 254 conn->c_unacked_packets = rds_sysctl_max_unacked_packets; 255 conn->c_unacked_bytes = rds_sysctl_max_unacked_bytes; 256 rds_stats_inc(s_send_ack_required); 257 } else { 258 conn->c_unacked_bytes -= len; 259 conn->c_unacked_packets--; 260 } 261 262 conn->c_xmit_rm = rm; 263 } 264 265 /* 266 * Try and send an rdma message. Let's see if we can 267 * keep this simple and require that the transport either 268 * send the whole rdma or none of it. 269 */ 270 if (rm->m_rdma_op && !conn->c_xmit_rdma_sent) { 271 ret = conn->c_trans->xmit_rdma(conn, rm->m_rdma_op); 272 if (ret) 273 break; 274 conn->c_xmit_rdma_sent = 1; 275 /* The transport owns the mapped memory for now. 276 * You can't unmap it while it's on the send queue */ 277 set_bit(RDS_MSG_MAPPED, &rm->m_flags); 278 } 279 280 if (conn->c_xmit_hdr_off < sizeof(struct rds_header) || 281 conn->c_xmit_sg < rm->m_nents) { 282 ret = conn->c_trans->xmit(conn, rm, 283 conn->c_xmit_hdr_off, 284 conn->c_xmit_sg, 285 conn->c_xmit_data_off); 286 if (ret <= 0) 287 break; 288 289 if (conn->c_xmit_hdr_off < sizeof(struct rds_header)) { 290 tmp = min_t(int, ret, 291 sizeof(struct rds_header) - 292 conn->c_xmit_hdr_off); 293 conn->c_xmit_hdr_off += tmp; 294 ret -= tmp; 295 } 296 297 sg = &rm->m_sg[conn->c_xmit_sg]; 298 while (ret) { 299 tmp = min_t(int, ret, sg->length - 300 conn->c_xmit_data_off); 301 conn->c_xmit_data_off += tmp; 302 ret -= tmp; 303 if (conn->c_xmit_data_off == sg->length) { 304 conn->c_xmit_data_off = 0; 305 sg++; 306 conn->c_xmit_sg++; 307 BUG_ON(ret != 0 && 308 conn->c_xmit_sg == rm->m_nents); 309 } 310 } 311 } 312 } 313 314 /* Nuke any messages we decided not to retransmit. */ 315 if (!list_empty(&to_be_dropped)) 316 rds_send_remove_from_sock(&to_be_dropped, RDS_RDMA_DROPPED); 317 318 if (conn->c_trans->xmit_complete) 319 conn->c_trans->xmit_complete(conn); 320 321 /* 322 * We might be racing with another sender who queued a message but 323 * backed off on noticing that we held the c_send_lock. If we check 324 * for queued messages after dropping the sem then either we'll 325 * see the queued message or the queuer will get the sem. If we 326 * notice the queued message then we trigger an immediate retry. 327 * 328 * We need to be careful only to do this when we stopped processing 329 * the send queue because it was empty. It's the only way we 330 * stop processing the loop when the transport hasn't taken 331 * responsibility for forward progress. 332 */ 333 mutex_unlock(&conn->c_send_lock); 334 335 if (conn->c_map_bytes || (send_quota == 0 && !was_empty)) { 336 /* We exhausted the send quota, but there's work left to 337 * do. Return and (re-)schedule the send worker. 338 */ 339 ret = -EAGAIN; 340 } 341 342 if (ret == 0 && was_empty) { 343 /* A simple bit test would be way faster than taking the 344 * spin lock */ 345 spin_lock_irqsave(&conn->c_lock, flags); 346 if (!list_empty(&conn->c_send_queue)) { 347 rds_stats_inc(s_send_sem_queue_raced); 348 ret = -EAGAIN; 349 } 350 spin_unlock_irqrestore(&conn->c_lock, flags); 351 } 352 out: 353 return ret; 354 } 355 356 static void rds_send_sndbuf_remove(struct rds_sock *rs, struct rds_message *rm) 357 { 358 u32 len = be32_to_cpu(rm->m_inc.i_hdr.h_len); 359 360 assert_spin_locked(&rs->rs_lock); 361 362 BUG_ON(rs->rs_snd_bytes < len); 363 rs->rs_snd_bytes -= len; 364 365 if (rs->rs_snd_bytes == 0) 366 rds_stats_inc(s_send_queue_empty); 367 } 368 369 static inline int rds_send_is_acked(struct rds_message *rm, u64 ack, 370 is_acked_func is_acked) 371 { 372 if (is_acked) 373 return is_acked(rm, ack); 374 return be64_to_cpu(rm->m_inc.i_hdr.h_sequence) <= ack; 375 } 376 377 /* 378 * Returns true if there are no messages on the send and retransmit queues 379 * which have a sequence number greater than or equal to the given sequence 380 * number. 381 */ 382 int rds_send_acked_before(struct rds_connection *conn, u64 seq) 383 { 384 struct rds_message *rm, *tmp; 385 int ret = 1; 386 387 spin_lock(&conn->c_lock); 388 389 list_for_each_entry_safe(rm, tmp, &conn->c_retrans, m_conn_item) { 390 if (be64_to_cpu(rm->m_inc.i_hdr.h_sequence) < seq) 391 ret = 0; 392 break; 393 } 394 395 list_for_each_entry_safe(rm, tmp, &conn->c_send_queue, m_conn_item) { 396 if (be64_to_cpu(rm->m_inc.i_hdr.h_sequence) < seq) 397 ret = 0; 398 break; 399 } 400 401 spin_unlock(&conn->c_lock); 402 403 return ret; 404 } 405 406 /* 407 * This is pretty similar to what happens below in the ACK 408 * handling code - except that we call here as soon as we get 409 * the IB send completion on the RDMA op and the accompanying 410 * message. 411 */ 412 void rds_rdma_send_complete(struct rds_message *rm, int status) 413 { 414 struct rds_sock *rs = NULL; 415 struct rds_rdma_op *ro; 416 struct rds_notifier *notifier; 417 418 spin_lock(&rm->m_rs_lock); 419 420 ro = rm->m_rdma_op; 421 if (test_bit(RDS_MSG_ON_SOCK, &rm->m_flags) && 422 ro && ro->r_notify && ro->r_notifier) { 423 notifier = ro->r_notifier; 424 rs = rm->m_rs; 425 sock_hold(rds_rs_to_sk(rs)); 426 427 notifier->n_status = status; 428 spin_lock(&rs->rs_lock); 429 list_add_tail(¬ifier->n_list, &rs->rs_notify_queue); 430 spin_unlock(&rs->rs_lock); 431 432 ro->r_notifier = NULL; 433 } 434 435 spin_unlock(&rm->m_rs_lock); 436 437 if (rs) { 438 rds_wake_sk_sleep(rs); 439 sock_put(rds_rs_to_sk(rs)); 440 } 441 } 442 EXPORT_SYMBOL_GPL(rds_rdma_send_complete); 443 444 /* 445 * This is the same as rds_rdma_send_complete except we 446 * don't do any locking - we have all the ingredients (message, 447 * socket, socket lock) and can just move the notifier. 448 */ 449 static inline void 450 __rds_rdma_send_complete(struct rds_sock *rs, struct rds_message *rm, int status) 451 { 452 struct rds_rdma_op *ro; 453 454 ro = rm->m_rdma_op; 455 if (ro && ro->r_notify && ro->r_notifier) { 456 ro->r_notifier->n_status = status; 457 list_add_tail(&ro->r_notifier->n_list, &rs->rs_notify_queue); 458 ro->r_notifier = NULL; 459 } 460 461 /* No need to wake the app - caller does this */ 462 } 463 464 /* 465 * This is called from the IB send completion when we detect 466 * a RDMA operation that failed with remote access error. 467 * So speed is not an issue here. 468 */ 469 struct rds_message *rds_send_get_message(struct rds_connection *conn, 470 struct rds_rdma_op *op) 471 { 472 struct rds_message *rm, *tmp, *found = NULL; 473 unsigned long flags; 474 475 spin_lock_irqsave(&conn->c_lock, flags); 476 477 list_for_each_entry_safe(rm, tmp, &conn->c_retrans, m_conn_item) { 478 if (rm->m_rdma_op == op) { 479 atomic_inc(&rm->m_refcount); 480 found = rm; 481 goto out; 482 } 483 } 484 485 list_for_each_entry_safe(rm, tmp, &conn->c_send_queue, m_conn_item) { 486 if (rm->m_rdma_op == op) { 487 atomic_inc(&rm->m_refcount); 488 found = rm; 489 break; 490 } 491 } 492 493 out: 494 spin_unlock_irqrestore(&conn->c_lock, flags); 495 496 return found; 497 } 498 EXPORT_SYMBOL_GPL(rds_send_get_message); 499 500 /* 501 * This removes messages from the socket's list if they're on it. The list 502 * argument must be private to the caller, we must be able to modify it 503 * without locks. The messages must have a reference held for their 504 * position on the list. This function will drop that reference after 505 * removing the messages from the 'messages' list regardless of if it found 506 * the messages on the socket list or not. 507 */ 508 void rds_send_remove_from_sock(struct list_head *messages, int status) 509 { 510 unsigned long flags = 0; /* silence gcc :P */ 511 struct rds_sock *rs = NULL; 512 struct rds_message *rm; 513 514 local_irq_save(flags); 515 while (!list_empty(messages)) { 516 rm = list_entry(messages->next, struct rds_message, 517 m_conn_item); 518 list_del_init(&rm->m_conn_item); 519 520 /* 521 * If we see this flag cleared then we're *sure* that someone 522 * else beat us to removing it from the sock. If we race 523 * with their flag update we'll get the lock and then really 524 * see that the flag has been cleared. 525 * 526 * The message spinlock makes sure nobody clears rm->m_rs 527 * while we're messing with it. It does not prevent the 528 * message from being removed from the socket, though. 529 */ 530 spin_lock(&rm->m_rs_lock); 531 if (!test_bit(RDS_MSG_ON_SOCK, &rm->m_flags)) 532 goto unlock_and_drop; 533 534 if (rs != rm->m_rs) { 535 if (rs) { 536 spin_unlock(&rs->rs_lock); 537 rds_wake_sk_sleep(rs); 538 sock_put(rds_rs_to_sk(rs)); 539 } 540 rs = rm->m_rs; 541 spin_lock(&rs->rs_lock); 542 sock_hold(rds_rs_to_sk(rs)); 543 } 544 545 if (test_and_clear_bit(RDS_MSG_ON_SOCK, &rm->m_flags)) { 546 struct rds_rdma_op *ro = rm->m_rdma_op; 547 struct rds_notifier *notifier; 548 549 list_del_init(&rm->m_sock_item); 550 rds_send_sndbuf_remove(rs, rm); 551 552 if (ro && ro->r_notifier && (status || ro->r_notify)) { 553 notifier = ro->r_notifier; 554 list_add_tail(¬ifier->n_list, 555 &rs->rs_notify_queue); 556 if (!notifier->n_status) 557 notifier->n_status = status; 558 rm->m_rdma_op->r_notifier = NULL; 559 } 560 rds_message_put(rm); 561 rm->m_rs = NULL; 562 } 563 564 unlock_and_drop: 565 spin_unlock(&rm->m_rs_lock); 566 rds_message_put(rm); 567 } 568 569 if (rs) { 570 spin_unlock(&rs->rs_lock); 571 rds_wake_sk_sleep(rs); 572 sock_put(rds_rs_to_sk(rs)); 573 } 574 local_irq_restore(flags); 575 } 576 577 /* 578 * Transports call here when they've determined that the receiver queued 579 * messages up to, and including, the given sequence number. Messages are 580 * moved to the retrans queue when rds_send_xmit picks them off the send 581 * queue. This means that in the TCP case, the message may not have been 582 * assigned the m_ack_seq yet - but that's fine as long as tcp_is_acked 583 * checks the RDS_MSG_HAS_ACK_SEQ bit. 584 * 585 * XXX It's not clear to me how this is safely serialized with socket 586 * destruction. Maybe it should bail if it sees SOCK_DEAD. 587 */ 588 void rds_send_drop_acked(struct rds_connection *conn, u64 ack, 589 is_acked_func is_acked) 590 { 591 struct rds_message *rm, *tmp; 592 unsigned long flags; 593 LIST_HEAD(list); 594 595 spin_lock_irqsave(&conn->c_lock, flags); 596 597 list_for_each_entry_safe(rm, tmp, &conn->c_retrans, m_conn_item) { 598 if (!rds_send_is_acked(rm, ack, is_acked)) 599 break; 600 601 list_move(&rm->m_conn_item, &list); 602 clear_bit(RDS_MSG_ON_CONN, &rm->m_flags); 603 } 604 605 /* order flag updates with spin locks */ 606 if (!list_empty(&list)) 607 smp_mb__after_clear_bit(); 608 609 spin_unlock_irqrestore(&conn->c_lock, flags); 610 611 /* now remove the messages from the sock list as needed */ 612 rds_send_remove_from_sock(&list, RDS_RDMA_SUCCESS); 613 } 614 EXPORT_SYMBOL_GPL(rds_send_drop_acked); 615 616 void rds_send_drop_to(struct rds_sock *rs, struct sockaddr_in *dest) 617 { 618 struct rds_message *rm, *tmp; 619 struct rds_connection *conn; 620 unsigned long flags, flags2; 621 LIST_HEAD(list); 622 int wake = 0; 623 624 /* get all the messages we're dropping under the rs lock */ 625 spin_lock_irqsave(&rs->rs_lock, flags); 626 627 list_for_each_entry_safe(rm, tmp, &rs->rs_send_queue, m_sock_item) { 628 if (dest && (dest->sin_addr.s_addr != rm->m_daddr || 629 dest->sin_port != rm->m_inc.i_hdr.h_dport)) 630 continue; 631 632 wake = 1; 633 list_move(&rm->m_sock_item, &list); 634 rds_send_sndbuf_remove(rs, rm); 635 clear_bit(RDS_MSG_ON_SOCK, &rm->m_flags); 636 637 /* If this is a RDMA operation, notify the app. */ 638 __rds_rdma_send_complete(rs, rm, RDS_RDMA_CANCELED); 639 } 640 641 /* order flag updates with the rs lock */ 642 if (wake) 643 smp_mb__after_clear_bit(); 644 645 spin_unlock_irqrestore(&rs->rs_lock, flags); 646 647 if (wake) 648 rds_wake_sk_sleep(rs); 649 650 conn = NULL; 651 652 /* now remove the messages from the conn list as needed */ 653 list_for_each_entry(rm, &list, m_sock_item) { 654 /* We do this here rather than in the loop above, so that 655 * we don't have to nest m_rs_lock under rs->rs_lock */ 656 spin_lock_irqsave(&rm->m_rs_lock, flags2); 657 rm->m_rs = NULL; 658 spin_unlock_irqrestore(&rm->m_rs_lock, flags2); 659 660 /* 661 * If we see this flag cleared then we're *sure* that someone 662 * else beat us to removing it from the conn. If we race 663 * with their flag update we'll get the lock and then really 664 * see that the flag has been cleared. 665 */ 666 if (!test_bit(RDS_MSG_ON_CONN, &rm->m_flags)) 667 continue; 668 669 if (conn != rm->m_inc.i_conn) { 670 if (conn) 671 spin_unlock_irqrestore(&conn->c_lock, flags); 672 conn = rm->m_inc.i_conn; 673 spin_lock_irqsave(&conn->c_lock, flags); 674 } 675 676 if (test_and_clear_bit(RDS_MSG_ON_CONN, &rm->m_flags)) { 677 list_del_init(&rm->m_conn_item); 678 rds_message_put(rm); 679 } 680 } 681 682 if (conn) 683 spin_unlock_irqrestore(&conn->c_lock, flags); 684 685 while (!list_empty(&list)) { 686 rm = list_entry(list.next, struct rds_message, m_sock_item); 687 list_del_init(&rm->m_sock_item); 688 689 rds_message_wait(rm); 690 rds_message_put(rm); 691 } 692 } 693 694 /* 695 * we only want this to fire once so we use the callers 'queued'. It's 696 * possible that another thread can race with us and remove the 697 * message from the flow with RDS_CANCEL_SENT_TO. 698 */ 699 static int rds_send_queue_rm(struct rds_sock *rs, struct rds_connection *conn, 700 struct rds_message *rm, __be16 sport, 701 __be16 dport, int *queued) 702 { 703 unsigned long flags; 704 u32 len; 705 706 if (*queued) 707 goto out; 708 709 len = be32_to_cpu(rm->m_inc.i_hdr.h_len); 710 711 /* this is the only place which holds both the socket's rs_lock 712 * and the connection's c_lock */ 713 spin_lock_irqsave(&rs->rs_lock, flags); 714 715 /* 716 * If there is a little space in sndbuf, we don't queue anything, 717 * and userspace gets -EAGAIN. But poll() indicates there's send 718 * room. This can lead to bad behavior (spinning) if snd_bytes isn't 719 * freed up by incoming acks. So we check the *old* value of 720 * rs_snd_bytes here to allow the last msg to exceed the buffer, 721 * and poll() now knows no more data can be sent. 722 */ 723 if (rs->rs_snd_bytes < rds_sk_sndbuf(rs)) { 724 rs->rs_snd_bytes += len; 725 726 /* let recv side know we are close to send space exhaustion. 727 * This is probably not the optimal way to do it, as this 728 * means we set the flag on *all* messages as soon as our 729 * throughput hits a certain threshold. 730 */ 731 if (rs->rs_snd_bytes >= rds_sk_sndbuf(rs) / 2) 732 __set_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags); 733 734 list_add_tail(&rm->m_sock_item, &rs->rs_send_queue); 735 set_bit(RDS_MSG_ON_SOCK, &rm->m_flags); 736 rds_message_addref(rm); 737 rm->m_rs = rs; 738 739 /* The code ordering is a little weird, but we're 740 trying to minimize the time we hold c_lock */ 741 rds_message_populate_header(&rm->m_inc.i_hdr, sport, dport, 0); 742 rm->m_inc.i_conn = conn; 743 rds_message_addref(rm); 744 745 spin_lock(&conn->c_lock); 746 rm->m_inc.i_hdr.h_sequence = cpu_to_be64(conn->c_next_tx_seq++); 747 list_add_tail(&rm->m_conn_item, &conn->c_send_queue); 748 set_bit(RDS_MSG_ON_CONN, &rm->m_flags); 749 spin_unlock(&conn->c_lock); 750 751 rdsdebug("queued msg %p len %d, rs %p bytes %d seq %llu\n", 752 rm, len, rs, rs->rs_snd_bytes, 753 (unsigned long long)be64_to_cpu(rm->m_inc.i_hdr.h_sequence)); 754 755 *queued = 1; 756 } 757 758 spin_unlock_irqrestore(&rs->rs_lock, flags); 759 out: 760 return *queued; 761 } 762 763 static int rds_cmsg_send(struct rds_sock *rs, struct rds_message *rm, 764 struct msghdr *msg, int *allocated_mr) 765 { 766 struct cmsghdr *cmsg; 767 int ret = 0; 768 769 for (cmsg = CMSG_FIRSTHDR(msg); cmsg; cmsg = CMSG_NXTHDR(msg, cmsg)) { 770 if (!CMSG_OK(msg, cmsg)) 771 return -EINVAL; 772 773 if (cmsg->cmsg_level != SOL_RDS) 774 continue; 775 776 /* As a side effect, RDMA_DEST and RDMA_MAP will set 777 * rm->m_rdma_cookie and rm->m_rdma_mr. 778 */ 779 switch (cmsg->cmsg_type) { 780 case RDS_CMSG_RDMA_ARGS: 781 ret = rds_cmsg_rdma_args(rs, rm, cmsg); 782 break; 783 784 case RDS_CMSG_RDMA_DEST: 785 ret = rds_cmsg_rdma_dest(rs, rm, cmsg); 786 break; 787 788 case RDS_CMSG_RDMA_MAP: 789 ret = rds_cmsg_rdma_map(rs, rm, cmsg); 790 if (!ret) 791 *allocated_mr = 1; 792 break; 793 794 default: 795 return -EINVAL; 796 } 797 798 if (ret) 799 break; 800 } 801 802 return ret; 803 } 804 805 int rds_sendmsg(struct kiocb *iocb, struct socket *sock, struct msghdr *msg, 806 size_t payload_len) 807 { 808 struct sock *sk = sock->sk; 809 struct rds_sock *rs = rds_sk_to_rs(sk); 810 struct sockaddr_in *usin = (struct sockaddr_in *)msg->msg_name; 811 __be32 daddr; 812 __be16 dport; 813 struct rds_message *rm = NULL; 814 struct rds_connection *conn; 815 int ret = 0; 816 int queued = 0, allocated_mr = 0; 817 int nonblock = msg->msg_flags & MSG_DONTWAIT; 818 long timeo = sock_rcvtimeo(sk, nonblock); 819 820 /* Mirror Linux UDP mirror of BSD error message compatibility */ 821 /* XXX: Perhaps MSG_MORE someday */ 822 if (msg->msg_flags & ~(MSG_DONTWAIT | MSG_CMSG_COMPAT)) { 823 printk(KERN_INFO "msg_flags 0x%08X\n", msg->msg_flags); 824 ret = -EOPNOTSUPP; 825 goto out; 826 } 827 828 if (msg->msg_namelen) { 829 /* XXX fail non-unicast destination IPs? */ 830 if (msg->msg_namelen < sizeof(*usin) || usin->sin_family != AF_INET) { 831 ret = -EINVAL; 832 goto out; 833 } 834 daddr = usin->sin_addr.s_addr; 835 dport = usin->sin_port; 836 } else { 837 /* We only care about consistency with ->connect() */ 838 lock_sock(sk); 839 daddr = rs->rs_conn_addr; 840 dport = rs->rs_conn_port; 841 release_sock(sk); 842 } 843 844 /* racing with another thread binding seems ok here */ 845 if (daddr == 0 || rs->rs_bound_addr == 0) { 846 ret = -ENOTCONN; /* XXX not a great errno */ 847 goto out; 848 } 849 850 rm = rds_message_copy_from_user(msg->msg_iov, payload_len); 851 if (IS_ERR(rm)) { 852 ret = PTR_ERR(rm); 853 rm = NULL; 854 goto out; 855 } 856 857 rm->m_daddr = daddr; 858 859 /* rds_conn_create has a spinlock that runs with IRQ off. 860 * Caching the conn in the socket helps a lot. */ 861 if (rs->rs_conn && rs->rs_conn->c_faddr == daddr) 862 conn = rs->rs_conn; 863 else { 864 conn = rds_conn_create_outgoing(rs->rs_bound_addr, daddr, 865 rs->rs_transport, 866 sock->sk->sk_allocation); 867 if (IS_ERR(conn)) { 868 ret = PTR_ERR(conn); 869 goto out; 870 } 871 rs->rs_conn = conn; 872 } 873 874 /* Parse any control messages the user may have included. */ 875 ret = rds_cmsg_send(rs, rm, msg, &allocated_mr); 876 if (ret) 877 goto out; 878 879 if ((rm->m_rdma_cookie || rm->m_rdma_op) && 880 conn->c_trans->xmit_rdma == NULL) { 881 if (printk_ratelimit()) 882 printk(KERN_NOTICE "rdma_op %p conn xmit_rdma %p\n", 883 rm->m_rdma_op, conn->c_trans->xmit_rdma); 884 ret = -EOPNOTSUPP; 885 goto out; 886 } 887 888 /* If the connection is down, trigger a connect. We may 889 * have scheduled a delayed reconnect however - in this case 890 * we should not interfere. 891 */ 892 if (rds_conn_state(conn) == RDS_CONN_DOWN && 893 !test_and_set_bit(RDS_RECONNECT_PENDING, &conn->c_flags)) 894 queue_delayed_work(rds_wq, &conn->c_conn_w, 0); 895 896 ret = rds_cong_wait(conn->c_fcong, dport, nonblock, rs); 897 if (ret) 898 goto out; 899 900 while (!rds_send_queue_rm(rs, conn, rm, rs->rs_bound_port, 901 dport, &queued)) { 902 rds_stats_inc(s_send_queue_full); 903 /* XXX make sure this is reasonable */ 904 if (payload_len > rds_sk_sndbuf(rs)) { 905 ret = -EMSGSIZE; 906 goto out; 907 } 908 if (nonblock) { 909 ret = -EAGAIN; 910 goto out; 911 } 912 913 timeo = wait_event_interruptible_timeout(*sk->sk_sleep, 914 rds_send_queue_rm(rs, conn, rm, 915 rs->rs_bound_port, 916 dport, 917 &queued), 918 timeo); 919 rdsdebug("sendmsg woke queued %d timeo %ld\n", queued, timeo); 920 if (timeo > 0 || timeo == MAX_SCHEDULE_TIMEOUT) 921 continue; 922 923 ret = timeo; 924 if (ret == 0) 925 ret = -ETIMEDOUT; 926 goto out; 927 } 928 929 /* 930 * By now we've committed to the send. We reuse rds_send_worker() 931 * to retry sends in the rds thread if the transport asks us to. 932 */ 933 rds_stats_inc(s_send_queued); 934 935 if (!test_bit(RDS_LL_SEND_FULL, &conn->c_flags)) 936 rds_send_worker(&conn->c_send_w.work); 937 938 rds_message_put(rm); 939 return payload_len; 940 941 out: 942 /* If the user included a RDMA_MAP cmsg, we allocated a MR on the fly. 943 * If the sendmsg goes through, we keep the MR. If it fails with EAGAIN 944 * or in any other way, we need to destroy the MR again */ 945 if (allocated_mr) 946 rds_rdma_unuse(rs, rds_rdma_cookie_key(rm->m_rdma_cookie), 1); 947 948 if (rm) 949 rds_message_put(rm); 950 return ret; 951 } 952 953 /* 954 * Reply to a ping packet. 955 */ 956 int 957 rds_send_pong(struct rds_connection *conn, __be16 dport) 958 { 959 struct rds_message *rm; 960 unsigned long flags; 961 int ret = 0; 962 963 rm = rds_message_alloc(0, GFP_ATOMIC); 964 if (rm == NULL) { 965 ret = -ENOMEM; 966 goto out; 967 } 968 969 rm->m_daddr = conn->c_faddr; 970 971 /* If the connection is down, trigger a connect. We may 972 * have scheduled a delayed reconnect however - in this case 973 * we should not interfere. 974 */ 975 if (rds_conn_state(conn) == RDS_CONN_DOWN && 976 !test_and_set_bit(RDS_RECONNECT_PENDING, &conn->c_flags)) 977 queue_delayed_work(rds_wq, &conn->c_conn_w, 0); 978 979 ret = rds_cong_wait(conn->c_fcong, dport, 1, NULL); 980 if (ret) 981 goto out; 982 983 spin_lock_irqsave(&conn->c_lock, flags); 984 list_add_tail(&rm->m_conn_item, &conn->c_send_queue); 985 set_bit(RDS_MSG_ON_CONN, &rm->m_flags); 986 rds_message_addref(rm); 987 rm->m_inc.i_conn = conn; 988 989 rds_message_populate_header(&rm->m_inc.i_hdr, 0, dport, 990 conn->c_next_tx_seq); 991 conn->c_next_tx_seq++; 992 spin_unlock_irqrestore(&conn->c_lock, flags); 993 994 rds_stats_inc(s_send_queued); 995 rds_stats_inc(s_send_pong); 996 997 queue_delayed_work(rds_wq, &conn->c_send_w, 0); 998 rds_message_put(rm); 999 return 0; 1000 1001 out: 1002 if (rm) 1003 rds_message_put(rm); 1004 return ret; 1005 } 1006