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/slab.h> 35 #include <net/sock.h> 36 #include <linux/in.h> 37 38 #include "rds.h" 39 #include "rdma.h" 40 41 void rds_inc_init(struct rds_incoming *inc, struct rds_connection *conn, 42 __be32 saddr) 43 { 44 atomic_set(&inc->i_refcount, 1); 45 INIT_LIST_HEAD(&inc->i_item); 46 inc->i_conn = conn; 47 inc->i_saddr = saddr; 48 inc->i_rdma_cookie = 0; 49 } 50 EXPORT_SYMBOL_GPL(rds_inc_init); 51 52 void rds_inc_addref(struct rds_incoming *inc) 53 { 54 rdsdebug("addref inc %p ref %d\n", inc, atomic_read(&inc->i_refcount)); 55 atomic_inc(&inc->i_refcount); 56 } 57 EXPORT_SYMBOL_GPL(rds_inc_addref); 58 59 void rds_inc_put(struct rds_incoming *inc) 60 { 61 rdsdebug("put inc %p ref %d\n", inc, atomic_read(&inc->i_refcount)); 62 if (atomic_dec_and_test(&inc->i_refcount)) { 63 BUG_ON(!list_empty(&inc->i_item)); 64 65 inc->i_conn->c_trans->inc_free(inc); 66 } 67 } 68 EXPORT_SYMBOL_GPL(rds_inc_put); 69 70 static void rds_recv_rcvbuf_delta(struct rds_sock *rs, struct sock *sk, 71 struct rds_cong_map *map, 72 int delta, __be16 port) 73 { 74 int now_congested; 75 76 if (delta == 0) 77 return; 78 79 rs->rs_rcv_bytes += delta; 80 now_congested = rs->rs_rcv_bytes > rds_sk_rcvbuf(rs); 81 82 rdsdebug("rs %p (%pI4:%u) recv bytes %d buf %d " 83 "now_cong %d delta %d\n", 84 rs, &rs->rs_bound_addr, 85 ntohs(rs->rs_bound_port), rs->rs_rcv_bytes, 86 rds_sk_rcvbuf(rs), now_congested, delta); 87 88 /* wasn't -> am congested */ 89 if (!rs->rs_congested && now_congested) { 90 rs->rs_congested = 1; 91 rds_cong_set_bit(map, port); 92 rds_cong_queue_updates(map); 93 } 94 /* was -> aren't congested */ 95 /* Require more free space before reporting uncongested to prevent 96 bouncing cong/uncong state too often */ 97 else if (rs->rs_congested && (rs->rs_rcv_bytes < (rds_sk_rcvbuf(rs)/2))) { 98 rs->rs_congested = 0; 99 rds_cong_clear_bit(map, port); 100 rds_cong_queue_updates(map); 101 } 102 103 /* do nothing if no change in cong state */ 104 } 105 106 /* 107 * Process all extension headers that come with this message. 108 */ 109 static void rds_recv_incoming_exthdrs(struct rds_incoming *inc, struct rds_sock *rs) 110 { 111 struct rds_header *hdr = &inc->i_hdr; 112 unsigned int pos = 0, type, len; 113 union { 114 struct rds_ext_header_version version; 115 struct rds_ext_header_rdma rdma; 116 struct rds_ext_header_rdma_dest rdma_dest; 117 } buffer; 118 119 while (1) { 120 len = sizeof(buffer); 121 type = rds_message_next_extension(hdr, &pos, &buffer, &len); 122 if (type == RDS_EXTHDR_NONE) 123 break; 124 /* Process extension header here */ 125 switch (type) { 126 case RDS_EXTHDR_RDMA: 127 rds_rdma_unuse(rs, be32_to_cpu(buffer.rdma.h_rdma_rkey), 0); 128 break; 129 130 case RDS_EXTHDR_RDMA_DEST: 131 /* We ignore the size for now. We could stash it 132 * somewhere and use it for error checking. */ 133 inc->i_rdma_cookie = rds_rdma_make_cookie( 134 be32_to_cpu(buffer.rdma_dest.h_rdma_rkey), 135 be32_to_cpu(buffer.rdma_dest.h_rdma_offset)); 136 137 break; 138 } 139 } 140 } 141 142 /* 143 * The transport must make sure that this is serialized against other 144 * rx and conn reset on this specific conn. 145 * 146 * We currently assert that only one fragmented message will be sent 147 * down a connection at a time. This lets us reassemble in the conn 148 * instead of per-flow which means that we don't have to go digging through 149 * flows to tear down partial reassembly progress on conn failure and 150 * we save flow lookup and locking for each frag arrival. It does mean 151 * that small messages will wait behind large ones. Fragmenting at all 152 * is only to reduce the memory consumption of pre-posted buffers. 153 * 154 * The caller passes in saddr and daddr instead of us getting it from the 155 * conn. This lets loopback, who only has one conn for both directions, 156 * tell us which roles the addrs in the conn are playing for this message. 157 */ 158 void rds_recv_incoming(struct rds_connection *conn, __be32 saddr, __be32 daddr, 159 struct rds_incoming *inc, gfp_t gfp, enum km_type km) 160 { 161 struct rds_sock *rs = NULL; 162 struct sock *sk; 163 unsigned long flags; 164 165 inc->i_conn = conn; 166 inc->i_rx_jiffies = jiffies; 167 168 rdsdebug("conn %p next %llu inc %p seq %llu len %u sport %u dport %u " 169 "flags 0x%x rx_jiffies %lu\n", conn, 170 (unsigned long long)conn->c_next_rx_seq, 171 inc, 172 (unsigned long long)be64_to_cpu(inc->i_hdr.h_sequence), 173 be32_to_cpu(inc->i_hdr.h_len), 174 be16_to_cpu(inc->i_hdr.h_sport), 175 be16_to_cpu(inc->i_hdr.h_dport), 176 inc->i_hdr.h_flags, 177 inc->i_rx_jiffies); 178 179 /* 180 * Sequence numbers should only increase. Messages get their 181 * sequence number as they're queued in a sending conn. They 182 * can be dropped, though, if the sending socket is closed before 183 * they hit the wire. So sequence numbers can skip forward 184 * under normal operation. They can also drop back in the conn 185 * failover case as previously sent messages are resent down the 186 * new instance of a conn. We drop those, otherwise we have 187 * to assume that the next valid seq does not come after a 188 * hole in the fragment stream. 189 * 190 * The headers don't give us a way to realize if fragments of 191 * a message have been dropped. We assume that frags that arrive 192 * to a flow are part of the current message on the flow that is 193 * being reassembled. This means that senders can't drop messages 194 * from the sending conn until all their frags are sent. 195 * 196 * XXX we could spend more on the wire to get more robust failure 197 * detection, arguably worth it to avoid data corruption. 198 */ 199 if (be64_to_cpu(inc->i_hdr.h_sequence) < conn->c_next_rx_seq && 200 (inc->i_hdr.h_flags & RDS_FLAG_RETRANSMITTED)) { 201 rds_stats_inc(s_recv_drop_old_seq); 202 goto out; 203 } 204 conn->c_next_rx_seq = be64_to_cpu(inc->i_hdr.h_sequence) + 1; 205 206 if (rds_sysctl_ping_enable && inc->i_hdr.h_dport == 0) { 207 rds_stats_inc(s_recv_ping); 208 rds_send_pong(conn, inc->i_hdr.h_sport); 209 goto out; 210 } 211 212 rs = rds_find_bound(daddr, inc->i_hdr.h_dport); 213 if (rs == NULL) { 214 rds_stats_inc(s_recv_drop_no_sock); 215 goto out; 216 } 217 218 /* Process extension headers */ 219 rds_recv_incoming_exthdrs(inc, rs); 220 221 /* We can be racing with rds_release() which marks the socket dead. */ 222 sk = rds_rs_to_sk(rs); 223 224 /* serialize with rds_release -> sock_orphan */ 225 write_lock_irqsave(&rs->rs_recv_lock, flags); 226 if (!sock_flag(sk, SOCK_DEAD)) { 227 rdsdebug("adding inc %p to rs %p's recv queue\n", inc, rs); 228 rds_stats_inc(s_recv_queued); 229 rds_recv_rcvbuf_delta(rs, sk, inc->i_conn->c_lcong, 230 be32_to_cpu(inc->i_hdr.h_len), 231 inc->i_hdr.h_dport); 232 rds_inc_addref(inc); 233 list_add_tail(&inc->i_item, &rs->rs_recv_queue); 234 __rds_wake_sk_sleep(sk); 235 } else { 236 rds_stats_inc(s_recv_drop_dead_sock); 237 } 238 write_unlock_irqrestore(&rs->rs_recv_lock, flags); 239 240 out: 241 if (rs) 242 rds_sock_put(rs); 243 } 244 EXPORT_SYMBOL_GPL(rds_recv_incoming); 245 246 /* 247 * be very careful here. This is being called as the condition in 248 * wait_event_*() needs to cope with being called many times. 249 */ 250 static int rds_next_incoming(struct rds_sock *rs, struct rds_incoming **inc) 251 { 252 unsigned long flags; 253 254 if (*inc == NULL) { 255 read_lock_irqsave(&rs->rs_recv_lock, flags); 256 if (!list_empty(&rs->rs_recv_queue)) { 257 *inc = list_entry(rs->rs_recv_queue.next, 258 struct rds_incoming, 259 i_item); 260 rds_inc_addref(*inc); 261 } 262 read_unlock_irqrestore(&rs->rs_recv_lock, flags); 263 } 264 265 return *inc != NULL; 266 } 267 268 static int rds_still_queued(struct rds_sock *rs, struct rds_incoming *inc, 269 int drop) 270 { 271 struct sock *sk = rds_rs_to_sk(rs); 272 int ret = 0; 273 unsigned long flags; 274 275 write_lock_irqsave(&rs->rs_recv_lock, flags); 276 if (!list_empty(&inc->i_item)) { 277 ret = 1; 278 if (drop) { 279 /* XXX make sure this i_conn is reliable */ 280 rds_recv_rcvbuf_delta(rs, sk, inc->i_conn->c_lcong, 281 -be32_to_cpu(inc->i_hdr.h_len), 282 inc->i_hdr.h_dport); 283 list_del_init(&inc->i_item); 284 rds_inc_put(inc); 285 } 286 } 287 write_unlock_irqrestore(&rs->rs_recv_lock, flags); 288 289 rdsdebug("inc %p rs %p still %d dropped %d\n", inc, rs, ret, drop); 290 return ret; 291 } 292 293 /* 294 * Pull errors off the error queue. 295 * If msghdr is NULL, we will just purge the error queue. 296 */ 297 int rds_notify_queue_get(struct rds_sock *rs, struct msghdr *msghdr) 298 { 299 struct rds_notifier *notifier; 300 struct rds_rdma_notify cmsg; 301 unsigned int count = 0, max_messages = ~0U; 302 unsigned long flags; 303 LIST_HEAD(copy); 304 int err = 0; 305 306 307 /* put_cmsg copies to user space and thus may sleep. We can't do this 308 * with rs_lock held, so first grab as many notifications as we can stuff 309 * in the user provided cmsg buffer. We don't try to copy more, to avoid 310 * losing notifications - except when the buffer is so small that it wouldn't 311 * even hold a single notification. Then we give him as much of this single 312 * msg as we can squeeze in, and set MSG_CTRUNC. 313 */ 314 if (msghdr) { 315 max_messages = msghdr->msg_controllen / CMSG_SPACE(sizeof(cmsg)); 316 if (!max_messages) 317 max_messages = 1; 318 } 319 320 spin_lock_irqsave(&rs->rs_lock, flags); 321 while (!list_empty(&rs->rs_notify_queue) && count < max_messages) { 322 notifier = list_entry(rs->rs_notify_queue.next, 323 struct rds_notifier, n_list); 324 list_move(¬ifier->n_list, ©); 325 count++; 326 } 327 spin_unlock_irqrestore(&rs->rs_lock, flags); 328 329 if (!count) 330 return 0; 331 332 while (!list_empty(©)) { 333 notifier = list_entry(copy.next, struct rds_notifier, n_list); 334 335 if (msghdr) { 336 cmsg.user_token = notifier->n_user_token; 337 cmsg.status = notifier->n_status; 338 339 err = put_cmsg(msghdr, SOL_RDS, RDS_CMSG_RDMA_STATUS, 340 sizeof(cmsg), &cmsg); 341 if (err) 342 break; 343 } 344 345 list_del_init(¬ifier->n_list); 346 kfree(notifier); 347 } 348 349 /* If we bailed out because of an error in put_cmsg, 350 * we may be left with one or more notifications that we 351 * didn't process. Return them to the head of the list. */ 352 if (!list_empty(©)) { 353 spin_lock_irqsave(&rs->rs_lock, flags); 354 list_splice(©, &rs->rs_notify_queue); 355 spin_unlock_irqrestore(&rs->rs_lock, flags); 356 } 357 358 return err; 359 } 360 361 /* 362 * Queue a congestion notification 363 */ 364 static int rds_notify_cong(struct rds_sock *rs, struct msghdr *msghdr) 365 { 366 uint64_t notify = rs->rs_cong_notify; 367 unsigned long flags; 368 int err; 369 370 err = put_cmsg(msghdr, SOL_RDS, RDS_CMSG_CONG_UPDATE, 371 sizeof(notify), ¬ify); 372 if (err) 373 return err; 374 375 spin_lock_irqsave(&rs->rs_lock, flags); 376 rs->rs_cong_notify &= ~notify; 377 spin_unlock_irqrestore(&rs->rs_lock, flags); 378 379 return 0; 380 } 381 382 /* 383 * Receive any control messages. 384 */ 385 static int rds_cmsg_recv(struct rds_incoming *inc, struct msghdr *msg) 386 { 387 int ret = 0; 388 389 if (inc->i_rdma_cookie) { 390 ret = put_cmsg(msg, SOL_RDS, RDS_CMSG_RDMA_DEST, 391 sizeof(inc->i_rdma_cookie), &inc->i_rdma_cookie); 392 if (ret) 393 return ret; 394 } 395 396 return 0; 397 } 398 399 int rds_recvmsg(struct kiocb *iocb, struct socket *sock, struct msghdr *msg, 400 size_t size, int msg_flags) 401 { 402 struct sock *sk = sock->sk; 403 struct rds_sock *rs = rds_sk_to_rs(sk); 404 long timeo; 405 int ret = 0, nonblock = msg_flags & MSG_DONTWAIT; 406 struct sockaddr_in *sin; 407 struct rds_incoming *inc = NULL; 408 409 /* udp_recvmsg()->sock_recvtimeo() gets away without locking too.. */ 410 timeo = sock_rcvtimeo(sk, nonblock); 411 412 rdsdebug("size %zu flags 0x%x timeo %ld\n", size, msg_flags, timeo); 413 414 if (msg_flags & MSG_OOB) 415 goto out; 416 417 while (1) { 418 /* If there are pending notifications, do those - and nothing else */ 419 if (!list_empty(&rs->rs_notify_queue)) { 420 ret = rds_notify_queue_get(rs, msg); 421 break; 422 } 423 424 if (rs->rs_cong_notify) { 425 ret = rds_notify_cong(rs, msg); 426 break; 427 } 428 429 if (!rds_next_incoming(rs, &inc)) { 430 if (nonblock) { 431 ret = -EAGAIN; 432 break; 433 } 434 435 timeo = wait_event_interruptible_timeout(*sk_sleep(sk), 436 (!list_empty(&rs->rs_notify_queue) || 437 rs->rs_cong_notify || 438 rds_next_incoming(rs, &inc)), timeo); 439 rdsdebug("recvmsg woke inc %p timeo %ld\n", inc, 440 timeo); 441 if (timeo > 0 || timeo == MAX_SCHEDULE_TIMEOUT) 442 continue; 443 444 ret = timeo; 445 if (ret == 0) 446 ret = -ETIMEDOUT; 447 break; 448 } 449 450 rdsdebug("copying inc %p from %pI4:%u to user\n", inc, 451 &inc->i_conn->c_faddr, 452 ntohs(inc->i_hdr.h_sport)); 453 ret = inc->i_conn->c_trans->inc_copy_to_user(inc, msg->msg_iov, 454 size); 455 if (ret < 0) 456 break; 457 458 /* 459 * if the message we just copied isn't at the head of the 460 * recv queue then someone else raced us to return it, try 461 * to get the next message. 462 */ 463 if (!rds_still_queued(rs, inc, !(msg_flags & MSG_PEEK))) { 464 rds_inc_put(inc); 465 inc = NULL; 466 rds_stats_inc(s_recv_deliver_raced); 467 continue; 468 } 469 470 if (ret < be32_to_cpu(inc->i_hdr.h_len)) { 471 if (msg_flags & MSG_TRUNC) 472 ret = be32_to_cpu(inc->i_hdr.h_len); 473 msg->msg_flags |= MSG_TRUNC; 474 } 475 476 if (rds_cmsg_recv(inc, msg)) { 477 ret = -EFAULT; 478 goto out; 479 } 480 481 rds_stats_inc(s_recv_delivered); 482 483 sin = (struct sockaddr_in *)msg->msg_name; 484 if (sin) { 485 sin->sin_family = AF_INET; 486 sin->sin_port = inc->i_hdr.h_sport; 487 sin->sin_addr.s_addr = inc->i_saddr; 488 memset(sin->sin_zero, 0, sizeof(sin->sin_zero)); 489 } 490 break; 491 } 492 493 if (inc) 494 rds_inc_put(inc); 495 496 out: 497 return ret; 498 } 499 500 /* 501 * The socket is being shut down and we're asked to drop messages that were 502 * queued for recvmsg. The caller has unbound the socket so the receive path 503 * won't queue any more incoming fragments or messages on the socket. 504 */ 505 void rds_clear_recv_queue(struct rds_sock *rs) 506 { 507 struct sock *sk = rds_rs_to_sk(rs); 508 struct rds_incoming *inc, *tmp; 509 unsigned long flags; 510 511 write_lock_irqsave(&rs->rs_recv_lock, flags); 512 list_for_each_entry_safe(inc, tmp, &rs->rs_recv_queue, i_item) { 513 rds_recv_rcvbuf_delta(rs, sk, inc->i_conn->c_lcong, 514 -be32_to_cpu(inc->i_hdr.h_len), 515 inc->i_hdr.h_dport); 516 list_del_init(&inc->i_item); 517 rds_inc_put(inc); 518 } 519 write_unlock_irqrestore(&rs->rs_recv_lock, flags); 520 } 521 522 /* 523 * inc->i_saddr isn't used here because it is only set in the receive 524 * path. 525 */ 526 void rds_inc_info_copy(struct rds_incoming *inc, 527 struct rds_info_iterator *iter, 528 __be32 saddr, __be32 daddr, int flip) 529 { 530 struct rds_info_message minfo; 531 532 minfo.seq = be64_to_cpu(inc->i_hdr.h_sequence); 533 minfo.len = be32_to_cpu(inc->i_hdr.h_len); 534 535 if (flip) { 536 minfo.laddr = daddr; 537 minfo.faddr = saddr; 538 minfo.lport = inc->i_hdr.h_dport; 539 minfo.fport = inc->i_hdr.h_sport; 540 } else { 541 minfo.laddr = saddr; 542 minfo.faddr = daddr; 543 minfo.lport = inc->i_hdr.h_sport; 544 minfo.fport = inc->i_hdr.h_dport; 545 } 546 547 rds_info_copy(iter, &minfo, sizeof(minfo)); 548 } 549