send.c revision a63273d4992603979ddb181b6a8f07082839b39f
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/gfp.h> 35#include <net/sock.h> 36#include <linux/in.h> 37#include <linux/list.h> 38 39#include "rds.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 */ 51static int send_batch_count = 64; 52module_param(send_batch_count, int, 0444); 53MODULE_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 */ 58void 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 */ 106int 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 && 167 conn->c_xmit_hdr_off == sizeof(struct rds_header) && 168 conn->c_xmit_sg == rm->data.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 && test_and_clear_bit(0, &conn->c_map_queued)) { 183 if (conn->c_trans->xmit_cong_map) { 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) { 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) { 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->rdma.m_rdma_op.r_active && 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->rdma.m_rdma_op.r_active && !conn->c_xmit_rdma_sent) { 271 ret = conn->c_trans->xmit_rdma(conn, &rm->rdma.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->data.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->data.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->data.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 } 352out: 353 return ret; 354} 355 356static 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 369static 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 */ 382int 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 */ 412void 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 unsigned long flags; 418 419 spin_lock_irqsave(&rm->m_rs_lock, flags); 420 421 ro = &rm->rdma.m_rdma_op; 422 if (test_bit(RDS_MSG_ON_SOCK, &rm->m_flags) && 423 ro->r_active && ro->r_notify && ro->r_notifier) { 424 notifier = ro->r_notifier; 425 rs = rm->m_rs; 426 sock_hold(rds_rs_to_sk(rs)); 427 428 notifier->n_status = status; 429 spin_lock(&rs->rs_lock); 430 list_add_tail(¬ifier->n_list, &rs->rs_notify_queue); 431 spin_unlock(&rs->rs_lock); 432 433 ro->r_notifier = NULL; 434 } 435 436 spin_unlock_irqrestore(&rm->m_rs_lock, flags); 437 438 if (rs) { 439 rds_wake_sk_sleep(rs); 440 sock_put(rds_rs_to_sk(rs)); 441 } 442} 443EXPORT_SYMBOL_GPL(rds_rdma_send_complete); 444 445/* 446 * This is the same as rds_rdma_send_complete except we 447 * don't do any locking - we have all the ingredients (message, 448 * socket, socket lock) and can just move the notifier. 449 */ 450static inline void 451__rds_rdma_send_complete(struct rds_sock *rs, struct rds_message *rm, int status) 452{ 453 struct rds_rdma_op *ro; 454 455 ro = &rm->rdma.m_rdma_op; 456 if (ro->r_active && ro->r_notify && ro->r_notifier) { 457 ro->r_notifier->n_status = status; 458 list_add_tail(&ro->r_notifier->n_list, &rs->rs_notify_queue); 459 ro->r_notifier = NULL; 460 } 461 462 /* No need to wake the app - caller does this */ 463} 464 465/* 466 * This is called from the IB send completion when we detect 467 * a RDMA operation that failed with remote access error. 468 * So speed is not an issue here. 469 */ 470struct rds_message *rds_send_get_message(struct rds_connection *conn, 471 struct rds_rdma_op *op) 472{ 473 struct rds_message *rm, *tmp, *found = NULL; 474 unsigned long flags; 475 476 spin_lock_irqsave(&conn->c_lock, flags); 477 478 list_for_each_entry_safe(rm, tmp, &conn->c_retrans, m_conn_item) { 479 if (&rm->rdma.m_rdma_op == op) { 480 atomic_inc(&rm->m_refcount); 481 found = rm; 482 goto out; 483 } 484 } 485 486 list_for_each_entry_safe(rm, tmp, &conn->c_send_queue, m_conn_item) { 487 if (&rm->rdma.m_rdma_op == op) { 488 atomic_inc(&rm->m_refcount); 489 found = rm; 490 break; 491 } 492 } 493 494out: 495 spin_unlock_irqrestore(&conn->c_lock, flags); 496 497 return found; 498} 499EXPORT_SYMBOL_GPL(rds_send_get_message); 500 501/* 502 * This removes messages from the socket's list if they're on it. The list 503 * argument must be private to the caller, we must be able to modify it 504 * without locks. The messages must have a reference held for their 505 * position on the list. This function will drop that reference after 506 * removing the messages from the 'messages' list regardless of if it found 507 * the messages on the socket list or not. 508 */ 509void rds_send_remove_from_sock(struct list_head *messages, int status) 510{ 511 unsigned long flags; 512 struct rds_sock *rs = NULL; 513 struct rds_message *rm; 514 515 while (!list_empty(messages)) { 516 int was_on_sock = 0; 517 518 rm = list_entry(messages->next, struct rds_message, 519 m_conn_item); 520 list_del_init(&rm->m_conn_item); 521 522 /* 523 * If we see this flag cleared then we're *sure* that someone 524 * else beat us to removing it from the sock. If we race 525 * with their flag update we'll get the lock and then really 526 * see that the flag has been cleared. 527 * 528 * The message spinlock makes sure nobody clears rm->m_rs 529 * while we're messing with it. It does not prevent the 530 * message from being removed from the socket, though. 531 */ 532 spin_lock_irqsave(&rm->m_rs_lock, flags); 533 if (!test_bit(RDS_MSG_ON_SOCK, &rm->m_flags)) 534 goto unlock_and_drop; 535 536 if (rs != rm->m_rs) { 537 if (rs) { 538 rds_wake_sk_sleep(rs); 539 sock_put(rds_rs_to_sk(rs)); 540 } 541 rs = rm->m_rs; 542 sock_hold(rds_rs_to_sk(rs)); 543 } 544 spin_lock(&rs->rs_lock); 545 546 if (test_and_clear_bit(RDS_MSG_ON_SOCK, &rm->m_flags)) { 547 struct rds_rdma_op *ro = &rm->rdma.m_rdma_op; 548 struct rds_notifier *notifier; 549 550 list_del_init(&rm->m_sock_item); 551 rds_send_sndbuf_remove(rs, rm); 552 553 if (ro->r_active && ro->r_notifier && 554 (ro->r_notify || (ro->r_recverr && status))) { 555 notifier = ro->r_notifier; 556 list_add_tail(¬ifier->n_list, 557 &rs->rs_notify_queue); 558 if (!notifier->n_status) 559 notifier->n_status = status; 560 rm->rdma.m_rdma_op.r_notifier = NULL; 561 } 562 was_on_sock = 1; 563 rm->m_rs = NULL; 564 } 565 spin_unlock(&rs->rs_lock); 566 567unlock_and_drop: 568 spin_unlock_irqrestore(&rm->m_rs_lock, flags); 569 rds_message_put(rm); 570 if (was_on_sock) 571 rds_message_put(rm); 572 } 573 574 if (rs) { 575 rds_wake_sk_sleep(rs); 576 sock_put(rds_rs_to_sk(rs)); 577 } 578} 579 580/* 581 * Transports call here when they've determined that the receiver queued 582 * messages up to, and including, the given sequence number. Messages are 583 * moved to the retrans queue when rds_send_xmit picks them off the send 584 * queue. This means that in the TCP case, the message may not have been 585 * assigned the m_ack_seq yet - but that's fine as long as tcp_is_acked 586 * checks the RDS_MSG_HAS_ACK_SEQ bit. 587 * 588 * XXX It's not clear to me how this is safely serialized with socket 589 * destruction. Maybe it should bail if it sees SOCK_DEAD. 590 */ 591void rds_send_drop_acked(struct rds_connection *conn, u64 ack, 592 is_acked_func is_acked) 593{ 594 struct rds_message *rm, *tmp; 595 unsigned long flags; 596 LIST_HEAD(list); 597 598 spin_lock_irqsave(&conn->c_lock, flags); 599 600 list_for_each_entry_safe(rm, tmp, &conn->c_retrans, m_conn_item) { 601 if (!rds_send_is_acked(rm, ack, is_acked)) 602 break; 603 604 list_move(&rm->m_conn_item, &list); 605 clear_bit(RDS_MSG_ON_CONN, &rm->m_flags); 606 } 607 608 /* order flag updates with spin locks */ 609 if (!list_empty(&list)) 610 smp_mb__after_clear_bit(); 611 612 spin_unlock_irqrestore(&conn->c_lock, flags); 613 614 /* now remove the messages from the sock list as needed */ 615 rds_send_remove_from_sock(&list, RDS_RDMA_SUCCESS); 616} 617EXPORT_SYMBOL_GPL(rds_send_drop_acked); 618 619void rds_send_drop_to(struct rds_sock *rs, struct sockaddr_in *dest) 620{ 621 struct rds_message *rm, *tmp; 622 struct rds_connection *conn; 623 unsigned long flags; 624 LIST_HEAD(list); 625 626 /* get all the messages we're dropping under the rs lock */ 627 spin_lock_irqsave(&rs->rs_lock, flags); 628 629 list_for_each_entry_safe(rm, tmp, &rs->rs_send_queue, m_sock_item) { 630 if (dest && (dest->sin_addr.s_addr != rm->m_daddr || 631 dest->sin_port != rm->m_inc.i_hdr.h_dport)) 632 continue; 633 634 list_move(&rm->m_sock_item, &list); 635 rds_send_sndbuf_remove(rs, rm); 636 clear_bit(RDS_MSG_ON_SOCK, &rm->m_flags); 637 } 638 639 /* order flag updates with the rs lock */ 640 smp_mb__after_clear_bit(); 641 642 spin_unlock_irqrestore(&rs->rs_lock, flags); 643 644 if (list_empty(&list)) 645 return; 646 647 /* Remove the messages from the conn */ 648 list_for_each_entry(rm, &list, m_sock_item) { 649 650 conn = rm->m_inc.i_conn; 651 652 spin_lock_irqsave(&conn->c_lock, flags); 653 /* 654 * Maybe someone else beat us to removing rm from the conn. 655 * If we race with their flag update we'll get the lock and 656 * then really see that the flag has been cleared. 657 */ 658 if (!test_and_clear_bit(RDS_MSG_ON_CONN, &rm->m_flags)) { 659 spin_unlock_irqrestore(&conn->c_lock, flags); 660 continue; 661 } 662 list_del_init(&rm->m_conn_item); 663 spin_unlock_irqrestore(&conn->c_lock, flags); 664 665 /* 666 * Couldn't grab m_rs_lock in top loop (lock ordering), 667 * but we can now. 668 */ 669 spin_lock_irqsave(&rm->m_rs_lock, flags); 670 671 spin_lock(&rs->rs_lock); 672 __rds_rdma_send_complete(rs, rm, RDS_RDMA_CANCELED); 673 spin_unlock(&rs->rs_lock); 674 675 rm->m_rs = NULL; 676 spin_unlock_irqrestore(&rm->m_rs_lock, flags); 677 678 rds_message_put(rm); 679 } 680 681 rds_wake_sk_sleep(rs); 682 683 while (!list_empty(&list)) { 684 rm = list_entry(list.next, struct rds_message, m_sock_item); 685 list_del_init(&rm->m_sock_item); 686 687 rds_message_wait(rm); 688 rds_message_put(rm); 689 } 690} 691 692/* 693 * we only want this to fire once so we use the callers 'queued'. It's 694 * possible that another thread can race with us and remove the 695 * message from the flow with RDS_CANCEL_SENT_TO. 696 */ 697static int rds_send_queue_rm(struct rds_sock *rs, struct rds_connection *conn, 698 struct rds_message *rm, __be16 sport, 699 __be16 dport, int *queued) 700{ 701 unsigned long flags; 702 u32 len; 703 704 if (*queued) 705 goto out; 706 707 len = be32_to_cpu(rm->m_inc.i_hdr.h_len); 708 709 /* this is the only place which holds both the socket's rs_lock 710 * and the connection's c_lock */ 711 spin_lock_irqsave(&rs->rs_lock, flags); 712 713 /* 714 * If there is a little space in sndbuf, we don't queue anything, 715 * and userspace gets -EAGAIN. But poll() indicates there's send 716 * room. This can lead to bad behavior (spinning) if snd_bytes isn't 717 * freed up by incoming acks. So we check the *old* value of 718 * rs_snd_bytes here to allow the last msg to exceed the buffer, 719 * and poll() now knows no more data can be sent. 720 */ 721 if (rs->rs_snd_bytes < rds_sk_sndbuf(rs)) { 722 rs->rs_snd_bytes += len; 723 724 /* let recv side know we are close to send space exhaustion. 725 * This is probably not the optimal way to do it, as this 726 * means we set the flag on *all* messages as soon as our 727 * throughput hits a certain threshold. 728 */ 729 if (rs->rs_snd_bytes >= rds_sk_sndbuf(rs) / 2) 730 __set_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags); 731 732 list_add_tail(&rm->m_sock_item, &rs->rs_send_queue); 733 set_bit(RDS_MSG_ON_SOCK, &rm->m_flags); 734 rds_message_addref(rm); 735 rm->m_rs = rs; 736 737 /* The code ordering is a little weird, but we're 738 trying to minimize the time we hold c_lock */ 739 rds_message_populate_header(&rm->m_inc.i_hdr, sport, dport, 0); 740 rm->m_inc.i_conn = conn; 741 rds_message_addref(rm); 742 743 spin_lock(&conn->c_lock); 744 rm->m_inc.i_hdr.h_sequence = cpu_to_be64(conn->c_next_tx_seq++); 745 list_add_tail(&rm->m_conn_item, &conn->c_send_queue); 746 set_bit(RDS_MSG_ON_CONN, &rm->m_flags); 747 spin_unlock(&conn->c_lock); 748 749 rdsdebug("queued msg %p len %d, rs %p bytes %d seq %llu\n", 750 rm, len, rs, rs->rs_snd_bytes, 751 (unsigned long long)be64_to_cpu(rm->m_inc.i_hdr.h_sequence)); 752 753 *queued = 1; 754 } 755 756 spin_unlock_irqrestore(&rs->rs_lock, flags); 757out: 758 return *queued; 759} 760 761/* 762 * rds_message is getting to be quite complicated, and we'd like to allocate 763 * it all in one go. This figures out how big it needs to be up front. 764 */ 765static int rds_rm_size(struct msghdr *msg, int data_len) 766{ 767 struct cmsghdr *cmsg; 768 int size = 0; 769 int retval; 770 771 for (cmsg = CMSG_FIRSTHDR(msg); cmsg; cmsg = CMSG_NXTHDR(msg, cmsg)) { 772 if (!CMSG_OK(msg, cmsg)) 773 return -EINVAL; 774 775 if (cmsg->cmsg_level != SOL_RDS) 776 continue; 777 778 switch (cmsg->cmsg_type) { 779 case RDS_CMSG_RDMA_ARGS: 780 retval = rds_rdma_extra_size(CMSG_DATA(cmsg)); 781 if (retval < 0) 782 return retval; 783 size += retval; 784 break; 785 786 case RDS_CMSG_RDMA_DEST: 787 case RDS_CMSG_RDMA_MAP: 788 /* these are valid but do no add any size */ 789 break; 790 791 default: 792 return -EINVAL; 793 } 794 795 } 796 797 size += ceil(data_len, PAGE_SIZE) * sizeof(struct scatterlist); 798 799 return size; 800} 801 802static int rds_cmsg_send(struct rds_sock *rs, struct rds_message *rm, 803 struct msghdr *msg, int *allocated_mr) 804{ 805 struct cmsghdr *cmsg; 806 int ret = 0; 807 808 for (cmsg = CMSG_FIRSTHDR(msg); cmsg; cmsg = CMSG_NXTHDR(msg, cmsg)) { 809 if (!CMSG_OK(msg, cmsg)) 810 return -EINVAL; 811 812 if (cmsg->cmsg_level != SOL_RDS) 813 continue; 814 815 /* As a side effect, RDMA_DEST and RDMA_MAP will set 816 * rm->m_rdma_cookie and rm->m_rdma_mr. 817 */ 818 switch (cmsg->cmsg_type) { 819 case RDS_CMSG_RDMA_ARGS: 820 ret = rds_cmsg_rdma_args(rs, rm, cmsg); 821 break; 822 823 case RDS_CMSG_RDMA_DEST: 824 ret = rds_cmsg_rdma_dest(rs, rm, cmsg); 825 break; 826 827 case RDS_CMSG_RDMA_MAP: 828 ret = rds_cmsg_rdma_map(rs, rm, cmsg); 829 if (!ret) 830 *allocated_mr = 1; 831 break; 832 833 default: 834 return -EINVAL; 835 } 836 837 if (ret) 838 break; 839 } 840 841 return ret; 842} 843 844int rds_sendmsg(struct kiocb *iocb, struct socket *sock, struct msghdr *msg, 845 size_t payload_len) 846{ 847 struct sock *sk = sock->sk; 848 struct rds_sock *rs = rds_sk_to_rs(sk); 849 struct sockaddr_in *usin = (struct sockaddr_in *)msg->msg_name; 850 __be32 daddr; 851 __be16 dport; 852 struct rds_message *rm = NULL; 853 struct rds_connection *conn; 854 int ret = 0; 855 int queued = 0, allocated_mr = 0; 856 int nonblock = msg->msg_flags & MSG_DONTWAIT; 857 long timeo = sock_sndtimeo(sk, nonblock); 858 859 /* Mirror Linux UDP mirror of BSD error message compatibility */ 860 /* XXX: Perhaps MSG_MORE someday */ 861 if (msg->msg_flags & ~(MSG_DONTWAIT | MSG_CMSG_COMPAT)) { 862 printk(KERN_INFO "msg_flags 0x%08X\n", msg->msg_flags); 863 ret = -EOPNOTSUPP; 864 goto out; 865 } 866 867 if (msg->msg_namelen) { 868 /* XXX fail non-unicast destination IPs? */ 869 if (msg->msg_namelen < sizeof(*usin) || usin->sin_family != AF_INET) { 870 ret = -EINVAL; 871 goto out; 872 } 873 daddr = usin->sin_addr.s_addr; 874 dport = usin->sin_port; 875 } else { 876 /* We only care about consistency with ->connect() */ 877 lock_sock(sk); 878 daddr = rs->rs_conn_addr; 879 dport = rs->rs_conn_port; 880 release_sock(sk); 881 } 882 883 /* racing with another thread binding seems ok here */ 884 if (daddr == 0 || rs->rs_bound_addr == 0) { 885 ret = -ENOTCONN; /* XXX not a great errno */ 886 goto out; 887 } 888 889 /* size of rm including all sgs */ 890 ret = rds_rm_size(msg, payload_len); 891 if (ret < 0) 892 goto out; 893 894 rm = rds_message_alloc(ret, GFP_KERNEL); 895 if (!rm) { 896 ret = -ENOMEM; 897 goto out; 898 } 899 900 rm->data.m_sg = rds_message_alloc_sgs(rm, ceil(payload_len, PAGE_SIZE)); 901 /* XXX fix this to not allocate memory */ 902 ret = rds_message_copy_from_user(rm, msg->msg_iov, payload_len); 903 if (ret) 904 goto out; 905 906 rm->m_daddr = daddr; 907 908 /* rds_conn_create has a spinlock that runs with IRQ off. 909 * Caching the conn in the socket helps a lot. */ 910 if (rs->rs_conn && rs->rs_conn->c_faddr == daddr) 911 conn = rs->rs_conn; 912 else { 913 conn = rds_conn_create_outgoing(rs->rs_bound_addr, daddr, 914 rs->rs_transport, 915 sock->sk->sk_allocation); 916 if (IS_ERR(conn)) { 917 ret = PTR_ERR(conn); 918 goto out; 919 } 920 rs->rs_conn = conn; 921 } 922 923 /* Parse any control messages the user may have included. */ 924 ret = rds_cmsg_send(rs, rm, msg, &allocated_mr); 925 if (ret) 926 goto out; 927 928 if ((rm->m_rdma_cookie || rm->rdma.m_rdma_op.r_active) && 929 !conn->c_trans->xmit_rdma) { 930 if (printk_ratelimit()) 931 printk(KERN_NOTICE "rdma_op %p conn xmit_rdma %p\n", 932 &rm->rdma.m_rdma_op, conn->c_trans->xmit_rdma); 933 ret = -EOPNOTSUPP; 934 goto out; 935 } 936 937 /* If the connection is down, trigger a connect. We may 938 * have scheduled a delayed reconnect however - in this case 939 * we should not interfere. 940 */ 941 if (rds_conn_state(conn) == RDS_CONN_DOWN && 942 !test_and_set_bit(RDS_RECONNECT_PENDING, &conn->c_flags)) 943 queue_delayed_work(rds_wq, &conn->c_conn_w, 0); 944 945 ret = rds_cong_wait(conn->c_fcong, dport, nonblock, rs); 946 if (ret) { 947 rs->rs_seen_congestion = 1; 948 goto out; 949 } 950 951 while (!rds_send_queue_rm(rs, conn, rm, rs->rs_bound_port, 952 dport, &queued)) { 953 rds_stats_inc(s_send_queue_full); 954 /* XXX make sure this is reasonable */ 955 if (payload_len > rds_sk_sndbuf(rs)) { 956 ret = -EMSGSIZE; 957 goto out; 958 } 959 if (nonblock) { 960 ret = -EAGAIN; 961 goto out; 962 } 963 964 timeo = wait_event_interruptible_timeout(*sk_sleep(sk), 965 rds_send_queue_rm(rs, conn, rm, 966 rs->rs_bound_port, 967 dport, 968 &queued), 969 timeo); 970 rdsdebug("sendmsg woke queued %d timeo %ld\n", queued, timeo); 971 if (timeo > 0 || timeo == MAX_SCHEDULE_TIMEOUT) 972 continue; 973 974 ret = timeo; 975 if (ret == 0) 976 ret = -ETIMEDOUT; 977 goto out; 978 } 979 980 /* 981 * By now we've committed to the send. We reuse rds_send_worker() 982 * to retry sends in the rds thread if the transport asks us to. 983 */ 984 rds_stats_inc(s_send_queued); 985 986 if (!test_bit(RDS_LL_SEND_FULL, &conn->c_flags)) 987 rds_send_worker(&conn->c_send_w.work); 988 989 rds_message_put(rm); 990 return payload_len; 991 992out: 993 /* If the user included a RDMA_MAP cmsg, we allocated a MR on the fly. 994 * If the sendmsg goes through, we keep the MR. If it fails with EAGAIN 995 * or in any other way, we need to destroy the MR again */ 996 if (allocated_mr) 997 rds_rdma_unuse(rs, rds_rdma_cookie_key(rm->m_rdma_cookie), 1); 998 999 if (rm) 1000 rds_message_put(rm); 1001 return ret; 1002} 1003 1004/* 1005 * Reply to a ping packet. 1006 */ 1007int 1008rds_send_pong(struct rds_connection *conn, __be16 dport) 1009{ 1010 struct rds_message *rm; 1011 unsigned long flags; 1012 int ret = 0; 1013 1014 rm = rds_message_alloc(0, GFP_ATOMIC); 1015 if (!rm) { 1016 ret = -ENOMEM; 1017 goto out; 1018 } 1019 1020 rm->m_daddr = conn->c_faddr; 1021 1022 /* If the connection is down, trigger a connect. We may 1023 * have scheduled a delayed reconnect however - in this case 1024 * we should not interfere. 1025 */ 1026 if (rds_conn_state(conn) == RDS_CONN_DOWN && 1027 !test_and_set_bit(RDS_RECONNECT_PENDING, &conn->c_flags)) 1028 queue_delayed_work(rds_wq, &conn->c_conn_w, 0); 1029 1030 ret = rds_cong_wait(conn->c_fcong, dport, 1, NULL); 1031 if (ret) 1032 goto out; 1033 1034 spin_lock_irqsave(&conn->c_lock, flags); 1035 list_add_tail(&rm->m_conn_item, &conn->c_send_queue); 1036 set_bit(RDS_MSG_ON_CONN, &rm->m_flags); 1037 rds_message_addref(rm); 1038 rm->m_inc.i_conn = conn; 1039 1040 rds_message_populate_header(&rm->m_inc.i_hdr, 0, dport, 1041 conn->c_next_tx_seq); 1042 conn->c_next_tx_seq++; 1043 spin_unlock_irqrestore(&conn->c_lock, flags); 1044 1045 rds_stats_inc(s_send_queued); 1046 rds_stats_inc(s_send_pong); 1047 1048 queue_delayed_work(rds_wq, &conn->c_send_w, 0); 1049 rds_message_put(rm); 1050 return 0; 1051 1052out: 1053 if (rm) 1054 rds_message_put(rm); 1055 return ret; 1056} 1057