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 <linux/pci.h> 36 #include <linux/dma-mapping.h> 37 #include <rdma/rdma_cm.h> 38 39 #include "rds.h" 40 #include "ib.h" 41 42 static struct kmem_cache *rds_ib_incoming_slab; 43 static struct kmem_cache *rds_ib_frag_slab; 44 static atomic_t rds_ib_allocation = ATOMIC_INIT(0); 45 46 static void rds_ib_frag_drop_page(struct rds_page_frag *frag) 47 { 48 rdsdebug("frag %p page %p\n", frag, frag->f_page); 49 __free_page(frag->f_page); 50 frag->f_page = NULL; 51 } 52 53 static void rds_ib_frag_free(struct rds_page_frag *frag) 54 { 55 rdsdebug("frag %p page %p\n", frag, frag->f_page); 56 BUG_ON(frag->f_page != NULL); 57 kmem_cache_free(rds_ib_frag_slab, frag); 58 } 59 60 /* 61 * We map a page at a time. Its fragments are posted in order. This 62 * is called in fragment order as the fragments get send completion events. 63 * Only the last frag in the page performs the unmapping. 64 * 65 * It's OK for ring cleanup to call this in whatever order it likes because 66 * DMA is not in flight and so we can unmap while other ring entries still 67 * hold page references in their frags. 68 */ 69 static void rds_ib_recv_unmap_page(struct rds_ib_connection *ic, 70 struct rds_ib_recv_work *recv) 71 { 72 struct rds_page_frag *frag = recv->r_frag; 73 74 rdsdebug("recv %p frag %p page %p\n", recv, frag, frag->f_page); 75 if (frag->f_mapped) 76 ib_dma_unmap_page(ic->i_cm_id->device, 77 frag->f_mapped, 78 RDS_FRAG_SIZE, DMA_FROM_DEVICE); 79 frag->f_mapped = 0; 80 } 81 82 void rds_ib_recv_init_ring(struct rds_ib_connection *ic) 83 { 84 struct rds_ib_recv_work *recv; 85 u32 i; 86 87 for (i = 0, recv = ic->i_recvs; i < ic->i_recv_ring.w_nr; i++, recv++) { 88 struct ib_sge *sge; 89 90 recv->r_ibinc = NULL; 91 recv->r_frag = NULL; 92 93 recv->r_wr.next = NULL; 94 recv->r_wr.wr_id = i; 95 recv->r_wr.sg_list = recv->r_sge; 96 recv->r_wr.num_sge = RDS_IB_RECV_SGE; 97 98 sge = rds_ib_data_sge(ic, recv->r_sge); 99 sge->addr = 0; 100 sge->length = RDS_FRAG_SIZE; 101 sge->lkey = ic->i_mr->lkey; 102 103 sge = rds_ib_header_sge(ic, recv->r_sge); 104 sge->addr = ic->i_recv_hdrs_dma + (i * sizeof(struct rds_header)); 105 sge->length = sizeof(struct rds_header); 106 sge->lkey = ic->i_mr->lkey; 107 } 108 } 109 110 static void rds_ib_recv_clear_one(struct rds_ib_connection *ic, 111 struct rds_ib_recv_work *recv) 112 { 113 if (recv->r_ibinc) { 114 rds_inc_put(&recv->r_ibinc->ii_inc); 115 recv->r_ibinc = NULL; 116 } 117 if (recv->r_frag) { 118 rds_ib_recv_unmap_page(ic, recv); 119 if (recv->r_frag->f_page) 120 rds_ib_frag_drop_page(recv->r_frag); 121 rds_ib_frag_free(recv->r_frag); 122 recv->r_frag = NULL; 123 } 124 } 125 126 void rds_ib_recv_clear_ring(struct rds_ib_connection *ic) 127 { 128 u32 i; 129 130 for (i = 0; i < ic->i_recv_ring.w_nr; i++) 131 rds_ib_recv_clear_one(ic, &ic->i_recvs[i]); 132 133 if (ic->i_frag.f_page) 134 rds_ib_frag_drop_page(&ic->i_frag); 135 } 136 137 static int rds_ib_recv_refill_one(struct rds_connection *conn, 138 struct rds_ib_recv_work *recv, 139 gfp_t kptr_gfp, gfp_t page_gfp) 140 { 141 struct rds_ib_connection *ic = conn->c_transport_data; 142 dma_addr_t dma_addr; 143 struct ib_sge *sge; 144 int ret = -ENOMEM; 145 146 if (recv->r_ibinc == NULL) { 147 if (!atomic_add_unless(&rds_ib_allocation, 1, rds_ib_sysctl_max_recv_allocation)) { 148 rds_ib_stats_inc(s_ib_rx_alloc_limit); 149 goto out; 150 } 151 recv->r_ibinc = kmem_cache_alloc(rds_ib_incoming_slab, 152 kptr_gfp); 153 if (recv->r_ibinc == NULL) { 154 atomic_dec(&rds_ib_allocation); 155 goto out; 156 } 157 INIT_LIST_HEAD(&recv->r_ibinc->ii_frags); 158 rds_inc_init(&recv->r_ibinc->ii_inc, conn, conn->c_faddr); 159 } 160 161 if (recv->r_frag == NULL) { 162 recv->r_frag = kmem_cache_alloc(rds_ib_frag_slab, kptr_gfp); 163 if (recv->r_frag == NULL) 164 goto out; 165 INIT_LIST_HEAD(&recv->r_frag->f_item); 166 recv->r_frag->f_page = NULL; 167 } 168 169 if (ic->i_frag.f_page == NULL) { 170 ic->i_frag.f_page = alloc_page(page_gfp); 171 if (ic->i_frag.f_page == NULL) 172 goto out; 173 ic->i_frag.f_offset = 0; 174 } 175 176 dma_addr = ib_dma_map_page(ic->i_cm_id->device, 177 ic->i_frag.f_page, 178 ic->i_frag.f_offset, 179 RDS_FRAG_SIZE, 180 DMA_FROM_DEVICE); 181 if (ib_dma_mapping_error(ic->i_cm_id->device, dma_addr)) 182 goto out; 183 184 /* 185 * Once we get the RDS_PAGE_LAST_OFF frag then rds_ib_frag_unmap() 186 * must be called on this recv. This happens as completions hit 187 * in order or on connection shutdown. 188 */ 189 recv->r_frag->f_page = ic->i_frag.f_page; 190 recv->r_frag->f_offset = ic->i_frag.f_offset; 191 recv->r_frag->f_mapped = dma_addr; 192 193 sge = rds_ib_data_sge(ic, recv->r_sge); 194 sge->addr = dma_addr; 195 sge->length = RDS_FRAG_SIZE; 196 197 sge = rds_ib_header_sge(ic, recv->r_sge); 198 sge->addr = ic->i_recv_hdrs_dma + (recv - ic->i_recvs) * sizeof(struct rds_header); 199 sge->length = sizeof(struct rds_header); 200 201 get_page(recv->r_frag->f_page); 202 203 if (ic->i_frag.f_offset < RDS_PAGE_LAST_OFF) { 204 ic->i_frag.f_offset += RDS_FRAG_SIZE; 205 } else { 206 put_page(ic->i_frag.f_page); 207 ic->i_frag.f_page = NULL; 208 ic->i_frag.f_offset = 0; 209 } 210 211 ret = 0; 212 out: 213 return ret; 214 } 215 216 /* 217 * This tries to allocate and post unused work requests after making sure that 218 * they have all the allocations they need to queue received fragments into 219 * sockets. The i_recv_mutex is held here so that ring_alloc and _unalloc 220 * pairs don't go unmatched. 221 * 222 * -1 is returned if posting fails due to temporary resource exhaustion. 223 */ 224 int rds_ib_recv_refill(struct rds_connection *conn, gfp_t kptr_gfp, 225 gfp_t page_gfp, int prefill) 226 { 227 struct rds_ib_connection *ic = conn->c_transport_data; 228 struct rds_ib_recv_work *recv; 229 struct ib_recv_wr *failed_wr; 230 unsigned int posted = 0; 231 int ret = 0; 232 u32 pos; 233 234 while ((prefill || rds_conn_up(conn)) && 235 rds_ib_ring_alloc(&ic->i_recv_ring, 1, &pos)) { 236 if (pos >= ic->i_recv_ring.w_nr) { 237 printk(KERN_NOTICE "Argh - ring alloc returned pos=%u\n", 238 pos); 239 ret = -EINVAL; 240 break; 241 } 242 243 recv = &ic->i_recvs[pos]; 244 ret = rds_ib_recv_refill_one(conn, recv, kptr_gfp, page_gfp); 245 if (ret) { 246 ret = -1; 247 break; 248 } 249 250 /* XXX when can this fail? */ 251 ret = ib_post_recv(ic->i_cm_id->qp, &recv->r_wr, &failed_wr); 252 rdsdebug("recv %p ibinc %p page %p addr %lu ret %d\n", recv, 253 recv->r_ibinc, recv->r_frag->f_page, 254 (long) recv->r_frag->f_mapped, ret); 255 if (ret) { 256 rds_ib_conn_error(conn, "recv post on " 257 "%pI4 returned %d, disconnecting and " 258 "reconnecting\n", &conn->c_faddr, 259 ret); 260 ret = -1; 261 break; 262 } 263 264 posted++; 265 } 266 267 /* We're doing flow control - update the window. */ 268 if (ic->i_flowctl && posted) 269 rds_ib_advertise_credits(conn, posted); 270 271 if (ret) 272 rds_ib_ring_unalloc(&ic->i_recv_ring, 1); 273 return ret; 274 } 275 276 void rds_ib_inc_purge(struct rds_incoming *inc) 277 { 278 struct rds_ib_incoming *ibinc; 279 struct rds_page_frag *frag; 280 struct rds_page_frag *pos; 281 282 ibinc = container_of(inc, struct rds_ib_incoming, ii_inc); 283 rdsdebug("purging ibinc %p inc %p\n", ibinc, inc); 284 285 list_for_each_entry_safe(frag, pos, &ibinc->ii_frags, f_item) { 286 list_del_init(&frag->f_item); 287 rds_ib_frag_drop_page(frag); 288 rds_ib_frag_free(frag); 289 } 290 } 291 292 void rds_ib_inc_free(struct rds_incoming *inc) 293 { 294 struct rds_ib_incoming *ibinc; 295 296 ibinc = container_of(inc, struct rds_ib_incoming, ii_inc); 297 298 rds_ib_inc_purge(inc); 299 rdsdebug("freeing ibinc %p inc %p\n", ibinc, inc); 300 BUG_ON(!list_empty(&ibinc->ii_frags)); 301 kmem_cache_free(rds_ib_incoming_slab, ibinc); 302 atomic_dec(&rds_ib_allocation); 303 BUG_ON(atomic_read(&rds_ib_allocation) < 0); 304 } 305 306 int rds_ib_inc_copy_to_user(struct rds_incoming *inc, struct iovec *first_iov, 307 size_t size) 308 { 309 struct rds_ib_incoming *ibinc; 310 struct rds_page_frag *frag; 311 struct iovec *iov = first_iov; 312 unsigned long to_copy; 313 unsigned long frag_off = 0; 314 unsigned long iov_off = 0; 315 int copied = 0; 316 int ret; 317 u32 len; 318 319 ibinc = container_of(inc, struct rds_ib_incoming, ii_inc); 320 frag = list_entry(ibinc->ii_frags.next, struct rds_page_frag, f_item); 321 len = be32_to_cpu(inc->i_hdr.h_len); 322 323 while (copied < size && copied < len) { 324 if (frag_off == RDS_FRAG_SIZE) { 325 frag = list_entry(frag->f_item.next, 326 struct rds_page_frag, f_item); 327 frag_off = 0; 328 } 329 while (iov_off == iov->iov_len) { 330 iov_off = 0; 331 iov++; 332 } 333 334 to_copy = min(iov->iov_len - iov_off, RDS_FRAG_SIZE - frag_off); 335 to_copy = min_t(size_t, to_copy, size - copied); 336 to_copy = min_t(unsigned long, to_copy, len - copied); 337 338 rdsdebug("%lu bytes to user [%p, %zu] + %lu from frag " 339 "[%p, %lu] + %lu\n", 340 to_copy, iov->iov_base, iov->iov_len, iov_off, 341 frag->f_page, frag->f_offset, frag_off); 342 343 /* XXX needs + offset for multiple recvs per page */ 344 ret = rds_page_copy_to_user(frag->f_page, 345 frag->f_offset + frag_off, 346 iov->iov_base + iov_off, 347 to_copy); 348 if (ret) { 349 copied = ret; 350 break; 351 } 352 353 iov_off += to_copy; 354 frag_off += to_copy; 355 copied += to_copy; 356 } 357 358 return copied; 359 } 360 361 /* ic starts out kzalloc()ed */ 362 void rds_ib_recv_init_ack(struct rds_ib_connection *ic) 363 { 364 struct ib_send_wr *wr = &ic->i_ack_wr; 365 struct ib_sge *sge = &ic->i_ack_sge; 366 367 sge->addr = ic->i_ack_dma; 368 sge->length = sizeof(struct rds_header); 369 sge->lkey = ic->i_mr->lkey; 370 371 wr->sg_list = sge; 372 wr->num_sge = 1; 373 wr->opcode = IB_WR_SEND; 374 wr->wr_id = RDS_IB_ACK_WR_ID; 375 wr->send_flags = IB_SEND_SIGNALED | IB_SEND_SOLICITED; 376 } 377 378 /* 379 * You'd think that with reliable IB connections you wouldn't need to ack 380 * messages that have been received. The problem is that IB hardware generates 381 * an ack message before it has DMAed the message into memory. This creates a 382 * potential message loss if the HCA is disabled for any reason between when it 383 * sends the ack and before the message is DMAed and processed. This is only a 384 * potential issue if another HCA is available for fail-over. 385 * 386 * When the remote host receives our ack they'll free the sent message from 387 * their send queue. To decrease the latency of this we always send an ack 388 * immediately after we've received messages. 389 * 390 * For simplicity, we only have one ack in flight at a time. This puts 391 * pressure on senders to have deep enough send queues to absorb the latency of 392 * a single ack frame being in flight. This might not be good enough. 393 * 394 * This is implemented by have a long-lived send_wr and sge which point to a 395 * statically allocated ack frame. This ack wr does not fall under the ring 396 * accounting that the tx and rx wrs do. The QP attribute specifically makes 397 * room for it beyond the ring size. Send completion notices its special 398 * wr_id and avoids working with the ring in that case. 399 */ 400 #ifndef KERNEL_HAS_ATOMIC64 401 static void rds_ib_set_ack(struct rds_ib_connection *ic, u64 seq, 402 int ack_required) 403 { 404 unsigned long flags; 405 406 spin_lock_irqsave(&ic->i_ack_lock, flags); 407 ic->i_ack_next = seq; 408 if (ack_required) 409 set_bit(IB_ACK_REQUESTED, &ic->i_ack_flags); 410 spin_unlock_irqrestore(&ic->i_ack_lock, flags); 411 } 412 413 static u64 rds_ib_get_ack(struct rds_ib_connection *ic) 414 { 415 unsigned long flags; 416 u64 seq; 417 418 clear_bit(IB_ACK_REQUESTED, &ic->i_ack_flags); 419 420 spin_lock_irqsave(&ic->i_ack_lock, flags); 421 seq = ic->i_ack_next; 422 spin_unlock_irqrestore(&ic->i_ack_lock, flags); 423 424 return seq; 425 } 426 #else 427 static void rds_ib_set_ack(struct rds_ib_connection *ic, u64 seq, 428 int ack_required) 429 { 430 atomic64_set(&ic->i_ack_next, seq); 431 if (ack_required) { 432 smp_mb__before_clear_bit(); 433 set_bit(IB_ACK_REQUESTED, &ic->i_ack_flags); 434 } 435 } 436 437 static u64 rds_ib_get_ack(struct rds_ib_connection *ic) 438 { 439 clear_bit(IB_ACK_REQUESTED, &ic->i_ack_flags); 440 smp_mb__after_clear_bit(); 441 442 return atomic64_read(&ic->i_ack_next); 443 } 444 #endif 445 446 447 static void rds_ib_send_ack(struct rds_ib_connection *ic, unsigned int adv_credits) 448 { 449 struct rds_header *hdr = ic->i_ack; 450 struct ib_send_wr *failed_wr; 451 u64 seq; 452 int ret; 453 454 seq = rds_ib_get_ack(ic); 455 456 rdsdebug("send_ack: ic %p ack %llu\n", ic, (unsigned long long) seq); 457 rds_message_populate_header(hdr, 0, 0, 0); 458 hdr->h_ack = cpu_to_be64(seq); 459 hdr->h_credit = adv_credits; 460 rds_message_make_checksum(hdr); 461 ic->i_ack_queued = jiffies; 462 463 ret = ib_post_send(ic->i_cm_id->qp, &ic->i_ack_wr, &failed_wr); 464 if (unlikely(ret)) { 465 /* Failed to send. Release the WR, and 466 * force another ACK. 467 */ 468 clear_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags); 469 set_bit(IB_ACK_REQUESTED, &ic->i_ack_flags); 470 471 rds_ib_stats_inc(s_ib_ack_send_failure); 472 473 rds_ib_conn_error(ic->conn, "sending ack failed\n"); 474 } else 475 rds_ib_stats_inc(s_ib_ack_sent); 476 } 477 478 /* 479 * There are 3 ways of getting acknowledgements to the peer: 480 * 1. We call rds_ib_attempt_ack from the recv completion handler 481 * to send an ACK-only frame. 482 * However, there can be only one such frame in the send queue 483 * at any time, so we may have to postpone it. 484 * 2. When another (data) packet is transmitted while there's 485 * an ACK in the queue, we piggyback the ACK sequence number 486 * on the data packet. 487 * 3. If the ACK WR is done sending, we get called from the 488 * send queue completion handler, and check whether there's 489 * another ACK pending (postponed because the WR was on the 490 * queue). If so, we transmit it. 491 * 492 * We maintain 2 variables: 493 * - i_ack_flags, which keeps track of whether the ACK WR 494 * is currently in the send queue or not (IB_ACK_IN_FLIGHT) 495 * - i_ack_next, which is the last sequence number we received 496 * 497 * Potentially, send queue and receive queue handlers can run concurrently. 498 * It would be nice to not have to use a spinlock to synchronize things, 499 * but the one problem that rules this out is that 64bit updates are 500 * not atomic on all platforms. Things would be a lot simpler if 501 * we had atomic64 or maybe cmpxchg64 everywhere. 502 * 503 * Reconnecting complicates this picture just slightly. When we 504 * reconnect, we may be seeing duplicate packets. The peer 505 * is retransmitting them, because it hasn't seen an ACK for 506 * them. It is important that we ACK these. 507 * 508 * ACK mitigation adds a header flag "ACK_REQUIRED"; any packet with 509 * this flag set *MUST* be acknowledged immediately. 510 */ 511 512 /* 513 * When we get here, we're called from the recv queue handler. 514 * Check whether we ought to transmit an ACK. 515 */ 516 void rds_ib_attempt_ack(struct rds_ib_connection *ic) 517 { 518 unsigned int adv_credits; 519 520 if (!test_bit(IB_ACK_REQUESTED, &ic->i_ack_flags)) 521 return; 522 523 if (test_and_set_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags)) { 524 rds_ib_stats_inc(s_ib_ack_send_delayed); 525 return; 526 } 527 528 /* Can we get a send credit? */ 529 if (!rds_ib_send_grab_credits(ic, 1, &adv_credits, 0, RDS_MAX_ADV_CREDIT)) { 530 rds_ib_stats_inc(s_ib_tx_throttle); 531 clear_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags); 532 return; 533 } 534 535 clear_bit(IB_ACK_REQUESTED, &ic->i_ack_flags); 536 rds_ib_send_ack(ic, adv_credits); 537 } 538 539 /* 540 * We get here from the send completion handler, when the 541 * adapter tells us the ACK frame was sent. 542 */ 543 void rds_ib_ack_send_complete(struct rds_ib_connection *ic) 544 { 545 clear_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags); 546 rds_ib_attempt_ack(ic); 547 } 548 549 /* 550 * This is called by the regular xmit code when it wants to piggyback 551 * an ACK on an outgoing frame. 552 */ 553 u64 rds_ib_piggyb_ack(struct rds_ib_connection *ic) 554 { 555 if (test_and_clear_bit(IB_ACK_REQUESTED, &ic->i_ack_flags)) 556 rds_ib_stats_inc(s_ib_ack_send_piggybacked); 557 return rds_ib_get_ack(ic); 558 } 559 560 static struct rds_header *rds_ib_get_header(struct rds_connection *conn, 561 struct rds_ib_recv_work *recv, 562 u32 data_len) 563 { 564 struct rds_ib_connection *ic = conn->c_transport_data; 565 void *hdr_buff = &ic->i_recv_hdrs[recv - ic->i_recvs]; 566 void *addr; 567 u32 misplaced_hdr_bytes; 568 569 /* 570 * Support header at the front (RDS 3.1+) as well as header-at-end. 571 * 572 * Cases: 573 * 1) header all in header buff (great!) 574 * 2) header all in data page (copy all to header buff) 575 * 3) header split across hdr buf + data page 576 * (move bit in hdr buff to end before copying other bit from data page) 577 */ 578 if (conn->c_version > RDS_PROTOCOL_3_0 || data_len == RDS_FRAG_SIZE) 579 return hdr_buff; 580 581 if (data_len <= (RDS_FRAG_SIZE - sizeof(struct rds_header))) { 582 addr = kmap_atomic(recv->r_frag->f_page, KM_SOFTIRQ0); 583 memcpy(hdr_buff, 584 addr + recv->r_frag->f_offset + data_len, 585 sizeof(struct rds_header)); 586 kunmap_atomic(addr, KM_SOFTIRQ0); 587 return hdr_buff; 588 } 589 590 misplaced_hdr_bytes = (sizeof(struct rds_header) - (RDS_FRAG_SIZE - data_len)); 591 592 memmove(hdr_buff + misplaced_hdr_bytes, hdr_buff, misplaced_hdr_bytes); 593 594 addr = kmap_atomic(recv->r_frag->f_page, KM_SOFTIRQ0); 595 memcpy(hdr_buff, addr + recv->r_frag->f_offset + data_len, 596 sizeof(struct rds_header) - misplaced_hdr_bytes); 597 kunmap_atomic(addr, KM_SOFTIRQ0); 598 return hdr_buff; 599 } 600 601 /* 602 * It's kind of lame that we're copying from the posted receive pages into 603 * long-lived bitmaps. We could have posted the bitmaps and rdma written into 604 * them. But receiving new congestion bitmaps should be a *rare* event, so 605 * hopefully we won't need to invest that complexity in making it more 606 * efficient. By copying we can share a simpler core with TCP which has to 607 * copy. 608 */ 609 static void rds_ib_cong_recv(struct rds_connection *conn, 610 struct rds_ib_incoming *ibinc) 611 { 612 struct rds_cong_map *map; 613 unsigned int map_off; 614 unsigned int map_page; 615 struct rds_page_frag *frag; 616 unsigned long frag_off; 617 unsigned long to_copy; 618 unsigned long copied; 619 uint64_t uncongested = 0; 620 void *addr; 621 622 /* catch completely corrupt packets */ 623 if (be32_to_cpu(ibinc->ii_inc.i_hdr.h_len) != RDS_CONG_MAP_BYTES) 624 return; 625 626 map = conn->c_fcong; 627 map_page = 0; 628 map_off = 0; 629 630 frag = list_entry(ibinc->ii_frags.next, struct rds_page_frag, f_item); 631 frag_off = 0; 632 633 copied = 0; 634 635 while (copied < RDS_CONG_MAP_BYTES) { 636 uint64_t *src, *dst; 637 unsigned int k; 638 639 to_copy = min(RDS_FRAG_SIZE - frag_off, PAGE_SIZE - map_off); 640 BUG_ON(to_copy & 7); /* Must be 64bit aligned. */ 641 642 addr = kmap_atomic(frag->f_page, KM_SOFTIRQ0); 643 644 src = addr + frag_off; 645 dst = (void *)map->m_page_addrs[map_page] + map_off; 646 for (k = 0; k < to_copy; k += 8) { 647 /* Record ports that became uncongested, ie 648 * bits that changed from 0 to 1. */ 649 uncongested |= ~(*src) & *dst; 650 *dst++ = *src++; 651 } 652 kunmap_atomic(addr, KM_SOFTIRQ0); 653 654 copied += to_copy; 655 656 map_off += to_copy; 657 if (map_off == PAGE_SIZE) { 658 map_off = 0; 659 map_page++; 660 } 661 662 frag_off += to_copy; 663 if (frag_off == RDS_FRAG_SIZE) { 664 frag = list_entry(frag->f_item.next, 665 struct rds_page_frag, f_item); 666 frag_off = 0; 667 } 668 } 669 670 /* the congestion map is in little endian order */ 671 uncongested = le64_to_cpu(uncongested); 672 673 rds_cong_map_updated(map, uncongested); 674 } 675 676 /* 677 * Rings are posted with all the allocations they'll need to queue the 678 * incoming message to the receiving socket so this can't fail. 679 * All fragments start with a header, so we can make sure we're not receiving 680 * garbage, and we can tell a small 8 byte fragment from an ACK frame. 681 */ 682 struct rds_ib_ack_state { 683 u64 ack_next; 684 u64 ack_recv; 685 unsigned int ack_required:1; 686 unsigned int ack_next_valid:1; 687 unsigned int ack_recv_valid:1; 688 }; 689 690 static void rds_ib_process_recv(struct rds_connection *conn, 691 struct rds_ib_recv_work *recv, u32 data_len, 692 struct rds_ib_ack_state *state) 693 { 694 struct rds_ib_connection *ic = conn->c_transport_data; 695 struct rds_ib_incoming *ibinc = ic->i_ibinc; 696 struct rds_header *ihdr, *hdr; 697 698 /* XXX shut down the connection if port 0,0 are seen? */ 699 700 rdsdebug("ic %p ibinc %p recv %p byte len %u\n", ic, ibinc, recv, 701 data_len); 702 703 if (data_len < sizeof(struct rds_header)) { 704 rds_ib_conn_error(conn, "incoming message " 705 "from %pI4 didn't inclue a " 706 "header, disconnecting and " 707 "reconnecting\n", 708 &conn->c_faddr); 709 return; 710 } 711 data_len -= sizeof(struct rds_header); 712 713 ihdr = rds_ib_get_header(conn, recv, data_len); 714 715 /* Validate the checksum. */ 716 if (!rds_message_verify_checksum(ihdr)) { 717 rds_ib_conn_error(conn, "incoming message " 718 "from %pI4 has corrupted header - " 719 "forcing a reconnect\n", 720 &conn->c_faddr); 721 rds_stats_inc(s_recv_drop_bad_checksum); 722 return; 723 } 724 725 /* Process the ACK sequence which comes with every packet */ 726 state->ack_recv = be64_to_cpu(ihdr->h_ack); 727 state->ack_recv_valid = 1; 728 729 /* Process the credits update if there was one */ 730 if (ihdr->h_credit) 731 rds_ib_send_add_credits(conn, ihdr->h_credit); 732 733 if (ihdr->h_sport == 0 && ihdr->h_dport == 0 && data_len == 0) { 734 /* This is an ACK-only packet. The fact that it gets 735 * special treatment here is that historically, ACKs 736 * were rather special beasts. 737 */ 738 rds_ib_stats_inc(s_ib_ack_received); 739 740 /* 741 * Usually the frags make their way on to incs and are then freed as 742 * the inc is freed. We don't go that route, so we have to drop the 743 * page ref ourselves. We can't just leave the page on the recv 744 * because that confuses the dma mapping of pages and each recv's use 745 * of a partial page. We can leave the frag, though, it will be 746 * reused. 747 * 748 * FIXME: Fold this into the code path below. 749 */ 750 rds_ib_frag_drop_page(recv->r_frag); 751 return; 752 } 753 754 /* 755 * If we don't already have an inc on the connection then this 756 * fragment has a header and starts a message.. copy its header 757 * into the inc and save the inc so we can hang upcoming fragments 758 * off its list. 759 */ 760 if (ibinc == NULL) { 761 ibinc = recv->r_ibinc; 762 recv->r_ibinc = NULL; 763 ic->i_ibinc = ibinc; 764 765 hdr = &ibinc->ii_inc.i_hdr; 766 memcpy(hdr, ihdr, sizeof(*hdr)); 767 ic->i_recv_data_rem = be32_to_cpu(hdr->h_len); 768 769 rdsdebug("ic %p ibinc %p rem %u flag 0x%x\n", ic, ibinc, 770 ic->i_recv_data_rem, hdr->h_flags); 771 } else { 772 hdr = &ibinc->ii_inc.i_hdr; 773 /* We can't just use memcmp here; fragments of a 774 * single message may carry different ACKs */ 775 if (hdr->h_sequence != ihdr->h_sequence || 776 hdr->h_len != ihdr->h_len || 777 hdr->h_sport != ihdr->h_sport || 778 hdr->h_dport != ihdr->h_dport) { 779 rds_ib_conn_error(conn, 780 "fragment header mismatch; forcing reconnect\n"); 781 return; 782 } 783 } 784 785 list_add_tail(&recv->r_frag->f_item, &ibinc->ii_frags); 786 recv->r_frag = NULL; 787 788 if (ic->i_recv_data_rem > RDS_FRAG_SIZE) 789 ic->i_recv_data_rem -= RDS_FRAG_SIZE; 790 else { 791 ic->i_recv_data_rem = 0; 792 ic->i_ibinc = NULL; 793 794 if (ibinc->ii_inc.i_hdr.h_flags == RDS_FLAG_CONG_BITMAP) 795 rds_ib_cong_recv(conn, ibinc); 796 else { 797 rds_recv_incoming(conn, conn->c_faddr, conn->c_laddr, 798 &ibinc->ii_inc, GFP_ATOMIC, 799 KM_SOFTIRQ0); 800 state->ack_next = be64_to_cpu(hdr->h_sequence); 801 state->ack_next_valid = 1; 802 } 803 804 /* Evaluate the ACK_REQUIRED flag *after* we received 805 * the complete frame, and after bumping the next_rx 806 * sequence. */ 807 if (hdr->h_flags & RDS_FLAG_ACK_REQUIRED) { 808 rds_stats_inc(s_recv_ack_required); 809 state->ack_required = 1; 810 } 811 812 rds_inc_put(&ibinc->ii_inc); 813 } 814 } 815 816 /* 817 * Plucking the oldest entry from the ring can be done concurrently with 818 * the thread refilling the ring. Each ring operation is protected by 819 * spinlocks and the transient state of refilling doesn't change the 820 * recording of which entry is oldest. 821 * 822 * This relies on IB only calling one cq comp_handler for each cq so that 823 * there will only be one caller of rds_recv_incoming() per RDS connection. 824 */ 825 void rds_ib_recv_cq_comp_handler(struct ib_cq *cq, void *context) 826 { 827 struct rds_connection *conn = context; 828 struct rds_ib_connection *ic = conn->c_transport_data; 829 830 rdsdebug("conn %p cq %p\n", conn, cq); 831 832 rds_ib_stats_inc(s_ib_rx_cq_call); 833 834 tasklet_schedule(&ic->i_recv_tasklet); 835 } 836 837 static inline void rds_poll_cq(struct rds_ib_connection *ic, 838 struct rds_ib_ack_state *state) 839 { 840 struct rds_connection *conn = ic->conn; 841 struct ib_wc wc; 842 struct rds_ib_recv_work *recv; 843 844 while (ib_poll_cq(ic->i_recv_cq, 1, &wc) > 0) { 845 rdsdebug("wc wr_id 0x%llx status %u byte_len %u imm_data %u\n", 846 (unsigned long long)wc.wr_id, wc.status, wc.byte_len, 847 be32_to_cpu(wc.ex.imm_data)); 848 rds_ib_stats_inc(s_ib_rx_cq_event); 849 850 recv = &ic->i_recvs[rds_ib_ring_oldest(&ic->i_recv_ring)]; 851 852 rds_ib_recv_unmap_page(ic, recv); 853 854 /* 855 * Also process recvs in connecting state because it is possible 856 * to get a recv completion _before_ the rdmacm ESTABLISHED 857 * event is processed. 858 */ 859 if (rds_conn_up(conn) || rds_conn_connecting(conn)) { 860 /* We expect errors as the qp is drained during shutdown */ 861 if (wc.status == IB_WC_SUCCESS) { 862 rds_ib_process_recv(conn, recv, wc.byte_len, state); 863 } else { 864 rds_ib_conn_error(conn, "recv completion on " 865 "%pI4 had status %u, disconnecting and " 866 "reconnecting\n", &conn->c_faddr, 867 wc.status); 868 } 869 } 870 871 rds_ib_ring_free(&ic->i_recv_ring, 1); 872 } 873 } 874 875 void rds_ib_recv_tasklet_fn(unsigned long data) 876 { 877 struct rds_ib_connection *ic = (struct rds_ib_connection *) data; 878 struct rds_connection *conn = ic->conn; 879 struct rds_ib_ack_state state = { 0, }; 880 881 rds_poll_cq(ic, &state); 882 ib_req_notify_cq(ic->i_recv_cq, IB_CQ_SOLICITED); 883 rds_poll_cq(ic, &state); 884 885 if (state.ack_next_valid) 886 rds_ib_set_ack(ic, state.ack_next, state.ack_required); 887 if (state.ack_recv_valid && state.ack_recv > ic->i_ack_recv) { 888 rds_send_drop_acked(conn, state.ack_recv, NULL); 889 ic->i_ack_recv = state.ack_recv; 890 } 891 if (rds_conn_up(conn)) 892 rds_ib_attempt_ack(ic); 893 894 /* If we ever end up with a really empty receive ring, we're 895 * in deep trouble, as the sender will definitely see RNR 896 * timeouts. */ 897 if (rds_ib_ring_empty(&ic->i_recv_ring)) 898 rds_ib_stats_inc(s_ib_rx_ring_empty); 899 900 /* 901 * If the ring is running low, then schedule the thread to refill. 902 */ 903 if (rds_ib_ring_low(&ic->i_recv_ring)) 904 queue_delayed_work(rds_wq, &conn->c_recv_w, 0); 905 } 906 907 int rds_ib_recv(struct rds_connection *conn) 908 { 909 struct rds_ib_connection *ic = conn->c_transport_data; 910 int ret = 0; 911 912 rdsdebug("conn %p\n", conn); 913 914 /* 915 * If we get a temporary posting failure in this context then 916 * we're really low and we want the caller to back off for a bit. 917 */ 918 mutex_lock(&ic->i_recv_mutex); 919 if (rds_ib_recv_refill(conn, GFP_KERNEL, GFP_HIGHUSER, 0)) 920 ret = -ENOMEM; 921 else 922 rds_ib_stats_inc(s_ib_rx_refill_from_thread); 923 mutex_unlock(&ic->i_recv_mutex); 924 925 if (rds_conn_up(conn)) 926 rds_ib_attempt_ack(ic); 927 928 return ret; 929 } 930 931 int __init rds_ib_recv_init(void) 932 { 933 struct sysinfo si; 934 int ret = -ENOMEM; 935 936 /* Default to 30% of all available RAM for recv memory */ 937 si_meminfo(&si); 938 rds_ib_sysctl_max_recv_allocation = si.totalram / 3 * PAGE_SIZE / RDS_FRAG_SIZE; 939 940 rds_ib_incoming_slab = kmem_cache_create("rds_ib_incoming", 941 sizeof(struct rds_ib_incoming), 942 0, 0, NULL); 943 if (rds_ib_incoming_slab == NULL) 944 goto out; 945 946 rds_ib_frag_slab = kmem_cache_create("rds_ib_frag", 947 sizeof(struct rds_page_frag), 948 0, 0, NULL); 949 if (rds_ib_frag_slab == NULL) 950 kmem_cache_destroy(rds_ib_incoming_slab); 951 else 952 ret = 0; 953 out: 954 return ret; 955 } 956 957 void rds_ib_recv_exit(void) 958 { 959 kmem_cache_destroy(rds_ib_incoming_slab); 960 kmem_cache_destroy(rds_ib_frag_slab); 961 } 962