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 "iw.h" 41 42static struct kmem_cache *rds_iw_incoming_slab; 43static struct kmem_cache *rds_iw_frag_slab; 44static atomic_t rds_iw_allocation = ATOMIC_INIT(0); 45 46static void rds_iw_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 53static void rds_iw_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); 57 kmem_cache_free(rds_iw_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 */ 69static void rds_iw_recv_unmap_page(struct rds_iw_connection *ic, 70 struct rds_iw_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 82void rds_iw_recv_init_ring(struct rds_iw_connection *ic) 83{ 84 struct rds_iw_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_iwinc = 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_IW_RECV_SGE; 97 98 sge = rds_iw_data_sge(ic, recv->r_sge); 99 sge->addr = 0; 100 sge->length = RDS_FRAG_SIZE; 101 sge->lkey = 0; 102 103 sge = rds_iw_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 = 0; 107 } 108} 109 110static void rds_iw_recv_clear_one(struct rds_iw_connection *ic, 111 struct rds_iw_recv_work *recv) 112{ 113 if (recv->r_iwinc) { 114 rds_inc_put(&recv->r_iwinc->ii_inc); 115 recv->r_iwinc = NULL; 116 } 117 if (recv->r_frag) { 118 rds_iw_recv_unmap_page(ic, recv); 119 if (recv->r_frag->f_page) 120 rds_iw_frag_drop_page(recv->r_frag); 121 rds_iw_frag_free(recv->r_frag); 122 recv->r_frag = NULL; 123 } 124} 125 126void rds_iw_recv_clear_ring(struct rds_iw_connection *ic) 127{ 128 u32 i; 129 130 for (i = 0; i < ic->i_recv_ring.w_nr; i++) 131 rds_iw_recv_clear_one(ic, &ic->i_recvs[i]); 132 133 if (ic->i_frag.f_page) 134 rds_iw_frag_drop_page(&ic->i_frag); 135} 136 137static int rds_iw_recv_refill_one(struct rds_connection *conn, 138 struct rds_iw_recv_work *recv, 139 gfp_t kptr_gfp, gfp_t page_gfp) 140{ 141 struct rds_iw_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_iwinc) { 147 if (!atomic_add_unless(&rds_iw_allocation, 1, rds_iw_sysctl_max_recv_allocation)) { 148 rds_iw_stats_inc(s_iw_rx_alloc_limit); 149 goto out; 150 } 151 recv->r_iwinc = kmem_cache_alloc(rds_iw_incoming_slab, 152 kptr_gfp); 153 if (!recv->r_iwinc) { 154 atomic_dec(&rds_iw_allocation); 155 goto out; 156 } 157 INIT_LIST_HEAD(&recv->r_iwinc->ii_frags); 158 rds_inc_init(&recv->r_iwinc->ii_inc, conn, conn->c_faddr); 159 } 160 161 if (!recv->r_frag) { 162 recv->r_frag = kmem_cache_alloc(rds_iw_frag_slab, kptr_gfp); 163 if (!recv->r_frag) 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) { 170 ic->i_frag.f_page = alloc_page(page_gfp); 171 if (!ic->i_frag.f_page) 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_iw_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_iw_data_sge(ic, recv->r_sge); 194 sge->addr = dma_addr; 195 sge->length = RDS_FRAG_SIZE; 196 197 sge = rds_iw_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; 212out: 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 */ 224int rds_iw_recv_refill(struct rds_connection *conn, gfp_t kptr_gfp, 225 gfp_t page_gfp, int prefill) 226{ 227 struct rds_iw_connection *ic = conn->c_transport_data; 228 struct rds_iw_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_iw_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_iw_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 iwinc %p page %p addr %lu ret %d\n", recv, 253 recv->r_iwinc, recv->r_frag->f_page, 254 (long) recv->r_frag->f_mapped, ret); 255 if (ret) { 256 rds_iw_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_iw_advertise_credits(conn, posted); 270 271 if (ret) 272 rds_iw_ring_unalloc(&ic->i_recv_ring, 1); 273 return ret; 274} 275 276static void rds_iw_inc_purge(struct rds_incoming *inc) 277{ 278 struct rds_iw_incoming *iwinc; 279 struct rds_page_frag *frag; 280 struct rds_page_frag *pos; 281 282 iwinc = container_of(inc, struct rds_iw_incoming, ii_inc); 283 rdsdebug("purging iwinc %p inc %p\n", iwinc, inc); 284 285 list_for_each_entry_safe(frag, pos, &iwinc->ii_frags, f_item) { 286 list_del_init(&frag->f_item); 287 rds_iw_frag_drop_page(frag); 288 rds_iw_frag_free(frag); 289 } 290} 291 292void rds_iw_inc_free(struct rds_incoming *inc) 293{ 294 struct rds_iw_incoming *iwinc; 295 296 iwinc = container_of(inc, struct rds_iw_incoming, ii_inc); 297 298 rds_iw_inc_purge(inc); 299 rdsdebug("freeing iwinc %p inc %p\n", iwinc, inc); 300 BUG_ON(!list_empty(&iwinc->ii_frags)); 301 kmem_cache_free(rds_iw_incoming_slab, iwinc); 302 atomic_dec(&rds_iw_allocation); 303 BUG_ON(atomic_read(&rds_iw_allocation) < 0); 304} 305 306int rds_iw_inc_copy_to_user(struct rds_incoming *inc, struct iov_iter *to) 307{ 308 struct rds_iw_incoming *iwinc; 309 struct rds_page_frag *frag; 310 unsigned long to_copy; 311 unsigned long frag_off = 0; 312 int copied = 0; 313 int ret; 314 u32 len; 315 316 iwinc = container_of(inc, struct rds_iw_incoming, ii_inc); 317 frag = list_entry(iwinc->ii_frags.next, struct rds_page_frag, f_item); 318 len = be32_to_cpu(inc->i_hdr.h_len); 319 320 while (iov_iter_count(to) && copied < len) { 321 if (frag_off == RDS_FRAG_SIZE) { 322 frag = list_entry(frag->f_item.next, 323 struct rds_page_frag, f_item); 324 frag_off = 0; 325 } 326 to_copy = min_t(unsigned long, iov_iter_count(to), 327 RDS_FRAG_SIZE - frag_off); 328 to_copy = min_t(unsigned long, to_copy, len - copied); 329 330 /* XXX needs + offset for multiple recvs per page */ 331 rds_stats_add(s_copy_to_user, to_copy); 332 ret = copy_page_to_iter(frag->f_page, 333 frag->f_offset + frag_off, 334 to_copy, 335 to); 336 if (ret != to_copy) 337 return -EFAULT; 338 339 frag_off += to_copy; 340 copied += to_copy; 341 } 342 343 return copied; 344} 345 346/* ic starts out kzalloc()ed */ 347void rds_iw_recv_init_ack(struct rds_iw_connection *ic) 348{ 349 struct ib_send_wr *wr = &ic->i_ack_wr; 350 struct ib_sge *sge = &ic->i_ack_sge; 351 352 sge->addr = ic->i_ack_dma; 353 sge->length = sizeof(struct rds_header); 354 sge->lkey = rds_iw_local_dma_lkey(ic); 355 356 wr->sg_list = sge; 357 wr->num_sge = 1; 358 wr->opcode = IB_WR_SEND; 359 wr->wr_id = RDS_IW_ACK_WR_ID; 360 wr->send_flags = IB_SEND_SIGNALED | IB_SEND_SOLICITED; 361} 362 363/* 364 * You'd think that with reliable IB connections you wouldn't need to ack 365 * messages that have been received. The problem is that IB hardware generates 366 * an ack message before it has DMAed the message into memory. This creates a 367 * potential message loss if the HCA is disabled for any reason between when it 368 * sends the ack and before the message is DMAed and processed. This is only a 369 * potential issue if another HCA is available for fail-over. 370 * 371 * When the remote host receives our ack they'll free the sent message from 372 * their send queue. To decrease the latency of this we always send an ack 373 * immediately after we've received messages. 374 * 375 * For simplicity, we only have one ack in flight at a time. This puts 376 * pressure on senders to have deep enough send queues to absorb the latency of 377 * a single ack frame being in flight. This might not be good enough. 378 * 379 * This is implemented by have a long-lived send_wr and sge which point to a 380 * statically allocated ack frame. This ack wr does not fall under the ring 381 * accounting that the tx and rx wrs do. The QP attribute specifically makes 382 * room for it beyond the ring size. Send completion notices its special 383 * wr_id and avoids working with the ring in that case. 384 */ 385#ifndef KERNEL_HAS_ATOMIC64 386static void rds_iw_set_ack(struct rds_iw_connection *ic, u64 seq, 387 int ack_required) 388{ 389 unsigned long flags; 390 391 spin_lock_irqsave(&ic->i_ack_lock, flags); 392 ic->i_ack_next = seq; 393 if (ack_required) 394 set_bit(IB_ACK_REQUESTED, &ic->i_ack_flags); 395 spin_unlock_irqrestore(&ic->i_ack_lock, flags); 396} 397 398static u64 rds_iw_get_ack(struct rds_iw_connection *ic) 399{ 400 unsigned long flags; 401 u64 seq; 402 403 clear_bit(IB_ACK_REQUESTED, &ic->i_ack_flags); 404 405 spin_lock_irqsave(&ic->i_ack_lock, flags); 406 seq = ic->i_ack_next; 407 spin_unlock_irqrestore(&ic->i_ack_lock, flags); 408 409 return seq; 410} 411#else 412static void rds_iw_set_ack(struct rds_iw_connection *ic, u64 seq, 413 int ack_required) 414{ 415 atomic64_set(&ic->i_ack_next, seq); 416 if (ack_required) { 417 smp_mb__before_atomic(); 418 set_bit(IB_ACK_REQUESTED, &ic->i_ack_flags); 419 } 420} 421 422static u64 rds_iw_get_ack(struct rds_iw_connection *ic) 423{ 424 clear_bit(IB_ACK_REQUESTED, &ic->i_ack_flags); 425 smp_mb__after_atomic(); 426 427 return atomic64_read(&ic->i_ack_next); 428} 429#endif 430 431 432static void rds_iw_send_ack(struct rds_iw_connection *ic, unsigned int adv_credits) 433{ 434 struct rds_header *hdr = ic->i_ack; 435 struct ib_send_wr *failed_wr; 436 u64 seq; 437 int ret; 438 439 seq = rds_iw_get_ack(ic); 440 441 rdsdebug("send_ack: ic %p ack %llu\n", ic, (unsigned long long) seq); 442 rds_message_populate_header(hdr, 0, 0, 0); 443 hdr->h_ack = cpu_to_be64(seq); 444 hdr->h_credit = adv_credits; 445 rds_message_make_checksum(hdr); 446 ic->i_ack_queued = jiffies; 447 448 ret = ib_post_send(ic->i_cm_id->qp, &ic->i_ack_wr, &failed_wr); 449 if (unlikely(ret)) { 450 /* Failed to send. Release the WR, and 451 * force another ACK. 452 */ 453 clear_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags); 454 set_bit(IB_ACK_REQUESTED, &ic->i_ack_flags); 455 456 rds_iw_stats_inc(s_iw_ack_send_failure); 457 458 rds_iw_conn_error(ic->conn, "sending ack failed\n"); 459 } else 460 rds_iw_stats_inc(s_iw_ack_sent); 461} 462 463/* 464 * There are 3 ways of getting acknowledgements to the peer: 465 * 1. We call rds_iw_attempt_ack from the recv completion handler 466 * to send an ACK-only frame. 467 * However, there can be only one such frame in the send queue 468 * at any time, so we may have to postpone it. 469 * 2. When another (data) packet is transmitted while there's 470 * an ACK in the queue, we piggyback the ACK sequence number 471 * on the data packet. 472 * 3. If the ACK WR is done sending, we get called from the 473 * send queue completion handler, and check whether there's 474 * another ACK pending (postponed because the WR was on the 475 * queue). If so, we transmit it. 476 * 477 * We maintain 2 variables: 478 * - i_ack_flags, which keeps track of whether the ACK WR 479 * is currently in the send queue or not (IB_ACK_IN_FLIGHT) 480 * - i_ack_next, which is the last sequence number we received 481 * 482 * Potentially, send queue and receive queue handlers can run concurrently. 483 * It would be nice to not have to use a spinlock to synchronize things, 484 * but the one problem that rules this out is that 64bit updates are 485 * not atomic on all platforms. Things would be a lot simpler if 486 * we had atomic64 or maybe cmpxchg64 everywhere. 487 * 488 * Reconnecting complicates this picture just slightly. When we 489 * reconnect, we may be seeing duplicate packets. The peer 490 * is retransmitting them, because it hasn't seen an ACK for 491 * them. It is important that we ACK these. 492 * 493 * ACK mitigation adds a header flag "ACK_REQUIRED"; any packet with 494 * this flag set *MUST* be acknowledged immediately. 495 */ 496 497/* 498 * When we get here, we're called from the recv queue handler. 499 * Check whether we ought to transmit an ACK. 500 */ 501void rds_iw_attempt_ack(struct rds_iw_connection *ic) 502{ 503 unsigned int adv_credits; 504 505 if (!test_bit(IB_ACK_REQUESTED, &ic->i_ack_flags)) 506 return; 507 508 if (test_and_set_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags)) { 509 rds_iw_stats_inc(s_iw_ack_send_delayed); 510 return; 511 } 512 513 /* Can we get a send credit? */ 514 if (!rds_iw_send_grab_credits(ic, 1, &adv_credits, 0, RDS_MAX_ADV_CREDIT)) { 515 rds_iw_stats_inc(s_iw_tx_throttle); 516 clear_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags); 517 return; 518 } 519 520 clear_bit(IB_ACK_REQUESTED, &ic->i_ack_flags); 521 rds_iw_send_ack(ic, adv_credits); 522} 523 524/* 525 * We get here from the send completion handler, when the 526 * adapter tells us the ACK frame was sent. 527 */ 528void rds_iw_ack_send_complete(struct rds_iw_connection *ic) 529{ 530 clear_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags); 531 rds_iw_attempt_ack(ic); 532} 533 534/* 535 * This is called by the regular xmit code when it wants to piggyback 536 * an ACK on an outgoing frame. 537 */ 538u64 rds_iw_piggyb_ack(struct rds_iw_connection *ic) 539{ 540 if (test_and_clear_bit(IB_ACK_REQUESTED, &ic->i_ack_flags)) 541 rds_iw_stats_inc(s_iw_ack_send_piggybacked); 542 return rds_iw_get_ack(ic); 543} 544 545/* 546 * It's kind of lame that we're copying from the posted receive pages into 547 * long-lived bitmaps. We could have posted the bitmaps and rdma written into 548 * them. But receiving new congestion bitmaps should be a *rare* event, so 549 * hopefully we won't need to invest that complexity in making it more 550 * efficient. By copying we can share a simpler core with TCP which has to 551 * copy. 552 */ 553static void rds_iw_cong_recv(struct rds_connection *conn, 554 struct rds_iw_incoming *iwinc) 555{ 556 struct rds_cong_map *map; 557 unsigned int map_off; 558 unsigned int map_page; 559 struct rds_page_frag *frag; 560 unsigned long frag_off; 561 unsigned long to_copy; 562 unsigned long copied; 563 uint64_t uncongested = 0; 564 void *addr; 565 566 /* catch completely corrupt packets */ 567 if (be32_to_cpu(iwinc->ii_inc.i_hdr.h_len) != RDS_CONG_MAP_BYTES) 568 return; 569 570 map = conn->c_fcong; 571 map_page = 0; 572 map_off = 0; 573 574 frag = list_entry(iwinc->ii_frags.next, struct rds_page_frag, f_item); 575 frag_off = 0; 576 577 copied = 0; 578 579 while (copied < RDS_CONG_MAP_BYTES) { 580 uint64_t *src, *dst; 581 unsigned int k; 582 583 to_copy = min(RDS_FRAG_SIZE - frag_off, PAGE_SIZE - map_off); 584 BUG_ON(to_copy & 7); /* Must be 64bit aligned. */ 585 586 addr = kmap_atomic(frag->f_page); 587 588 src = addr + frag_off; 589 dst = (void *)map->m_page_addrs[map_page] + map_off; 590 for (k = 0; k < to_copy; k += 8) { 591 /* Record ports that became uncongested, ie 592 * bits that changed from 0 to 1. */ 593 uncongested |= ~(*src) & *dst; 594 *dst++ = *src++; 595 } 596 kunmap_atomic(addr); 597 598 copied += to_copy; 599 600 map_off += to_copy; 601 if (map_off == PAGE_SIZE) { 602 map_off = 0; 603 map_page++; 604 } 605 606 frag_off += to_copy; 607 if (frag_off == RDS_FRAG_SIZE) { 608 frag = list_entry(frag->f_item.next, 609 struct rds_page_frag, f_item); 610 frag_off = 0; 611 } 612 } 613 614 /* the congestion map is in little endian order */ 615 uncongested = le64_to_cpu(uncongested); 616 617 rds_cong_map_updated(map, uncongested); 618} 619 620/* 621 * Rings are posted with all the allocations they'll need to queue the 622 * incoming message to the receiving socket so this can't fail. 623 * All fragments start with a header, so we can make sure we're not receiving 624 * garbage, and we can tell a small 8 byte fragment from an ACK frame. 625 */ 626struct rds_iw_ack_state { 627 u64 ack_next; 628 u64 ack_recv; 629 unsigned int ack_required:1; 630 unsigned int ack_next_valid:1; 631 unsigned int ack_recv_valid:1; 632}; 633 634static void rds_iw_process_recv(struct rds_connection *conn, 635 struct rds_iw_recv_work *recv, u32 byte_len, 636 struct rds_iw_ack_state *state) 637{ 638 struct rds_iw_connection *ic = conn->c_transport_data; 639 struct rds_iw_incoming *iwinc = ic->i_iwinc; 640 struct rds_header *ihdr, *hdr; 641 642 /* XXX shut down the connection if port 0,0 are seen? */ 643 644 rdsdebug("ic %p iwinc %p recv %p byte len %u\n", ic, iwinc, recv, 645 byte_len); 646 647 if (byte_len < sizeof(struct rds_header)) { 648 rds_iw_conn_error(conn, "incoming message " 649 "from %pI4 didn't include a " 650 "header, disconnecting and " 651 "reconnecting\n", 652 &conn->c_faddr); 653 return; 654 } 655 byte_len -= sizeof(struct rds_header); 656 657 ihdr = &ic->i_recv_hdrs[recv - ic->i_recvs]; 658 659 /* Validate the checksum. */ 660 if (!rds_message_verify_checksum(ihdr)) { 661 rds_iw_conn_error(conn, "incoming message " 662 "from %pI4 has corrupted header - " 663 "forcing a reconnect\n", 664 &conn->c_faddr); 665 rds_stats_inc(s_recv_drop_bad_checksum); 666 return; 667 } 668 669 /* Process the ACK sequence which comes with every packet */ 670 state->ack_recv = be64_to_cpu(ihdr->h_ack); 671 state->ack_recv_valid = 1; 672 673 /* Process the credits update if there was one */ 674 if (ihdr->h_credit) 675 rds_iw_send_add_credits(conn, ihdr->h_credit); 676 677 if (ihdr->h_sport == 0 && ihdr->h_dport == 0 && byte_len == 0) { 678 /* This is an ACK-only packet. The fact that it gets 679 * special treatment here is that historically, ACKs 680 * were rather special beasts. 681 */ 682 rds_iw_stats_inc(s_iw_ack_received); 683 684 /* 685 * Usually the frags make their way on to incs and are then freed as 686 * the inc is freed. We don't go that route, so we have to drop the 687 * page ref ourselves. We can't just leave the page on the recv 688 * because that confuses the dma mapping of pages and each recv's use 689 * of a partial page. We can leave the frag, though, it will be 690 * reused. 691 * 692 * FIXME: Fold this into the code path below. 693 */ 694 rds_iw_frag_drop_page(recv->r_frag); 695 return; 696 } 697 698 /* 699 * If we don't already have an inc on the connection then this 700 * fragment has a header and starts a message.. copy its header 701 * into the inc and save the inc so we can hang upcoming fragments 702 * off its list. 703 */ 704 if (!iwinc) { 705 iwinc = recv->r_iwinc; 706 recv->r_iwinc = NULL; 707 ic->i_iwinc = iwinc; 708 709 hdr = &iwinc->ii_inc.i_hdr; 710 memcpy(hdr, ihdr, sizeof(*hdr)); 711 ic->i_recv_data_rem = be32_to_cpu(hdr->h_len); 712 713 rdsdebug("ic %p iwinc %p rem %u flag 0x%x\n", ic, iwinc, 714 ic->i_recv_data_rem, hdr->h_flags); 715 } else { 716 hdr = &iwinc->ii_inc.i_hdr; 717 /* We can't just use memcmp here; fragments of a 718 * single message may carry different ACKs */ 719 if (hdr->h_sequence != ihdr->h_sequence || 720 hdr->h_len != ihdr->h_len || 721 hdr->h_sport != ihdr->h_sport || 722 hdr->h_dport != ihdr->h_dport) { 723 rds_iw_conn_error(conn, 724 "fragment header mismatch; forcing reconnect\n"); 725 return; 726 } 727 } 728 729 list_add_tail(&recv->r_frag->f_item, &iwinc->ii_frags); 730 recv->r_frag = NULL; 731 732 if (ic->i_recv_data_rem > RDS_FRAG_SIZE) 733 ic->i_recv_data_rem -= RDS_FRAG_SIZE; 734 else { 735 ic->i_recv_data_rem = 0; 736 ic->i_iwinc = NULL; 737 738 if (iwinc->ii_inc.i_hdr.h_flags == RDS_FLAG_CONG_BITMAP) 739 rds_iw_cong_recv(conn, iwinc); 740 else { 741 rds_recv_incoming(conn, conn->c_faddr, conn->c_laddr, 742 &iwinc->ii_inc, GFP_ATOMIC); 743 state->ack_next = be64_to_cpu(hdr->h_sequence); 744 state->ack_next_valid = 1; 745 } 746 747 /* Evaluate the ACK_REQUIRED flag *after* we received 748 * the complete frame, and after bumping the next_rx 749 * sequence. */ 750 if (hdr->h_flags & RDS_FLAG_ACK_REQUIRED) { 751 rds_stats_inc(s_recv_ack_required); 752 state->ack_required = 1; 753 } 754 755 rds_inc_put(&iwinc->ii_inc); 756 } 757} 758 759/* 760 * Plucking the oldest entry from the ring can be done concurrently with 761 * the thread refilling the ring. Each ring operation is protected by 762 * spinlocks and the transient state of refilling doesn't change the 763 * recording of which entry is oldest. 764 * 765 * This relies on IB only calling one cq comp_handler for each cq so that 766 * there will only be one caller of rds_recv_incoming() per RDS connection. 767 */ 768void rds_iw_recv_cq_comp_handler(struct ib_cq *cq, void *context) 769{ 770 struct rds_connection *conn = context; 771 struct rds_iw_connection *ic = conn->c_transport_data; 772 773 rdsdebug("conn %p cq %p\n", conn, cq); 774 775 rds_iw_stats_inc(s_iw_rx_cq_call); 776 777 tasklet_schedule(&ic->i_recv_tasklet); 778} 779 780static inline void rds_poll_cq(struct rds_iw_connection *ic, 781 struct rds_iw_ack_state *state) 782{ 783 struct rds_connection *conn = ic->conn; 784 struct ib_wc wc; 785 struct rds_iw_recv_work *recv; 786 787 while (ib_poll_cq(ic->i_recv_cq, 1, &wc) > 0) { 788 rdsdebug("wc wr_id 0x%llx status %u byte_len %u imm_data %u\n", 789 (unsigned long long)wc.wr_id, wc.status, wc.byte_len, 790 be32_to_cpu(wc.ex.imm_data)); 791 rds_iw_stats_inc(s_iw_rx_cq_event); 792 793 recv = &ic->i_recvs[rds_iw_ring_oldest(&ic->i_recv_ring)]; 794 795 rds_iw_recv_unmap_page(ic, recv); 796 797 /* 798 * Also process recvs in connecting state because it is possible 799 * to get a recv completion _before_ the rdmacm ESTABLISHED 800 * event is processed. 801 */ 802 if (rds_conn_up(conn) || rds_conn_connecting(conn)) { 803 /* We expect errors as the qp is drained during shutdown */ 804 if (wc.status == IB_WC_SUCCESS) { 805 rds_iw_process_recv(conn, recv, wc.byte_len, state); 806 } else { 807 rds_iw_conn_error(conn, "recv completion on " 808 "%pI4 had status %u, disconnecting and " 809 "reconnecting\n", &conn->c_faddr, 810 wc.status); 811 } 812 } 813 814 rds_iw_ring_free(&ic->i_recv_ring, 1); 815 } 816} 817 818void rds_iw_recv_tasklet_fn(unsigned long data) 819{ 820 struct rds_iw_connection *ic = (struct rds_iw_connection *) data; 821 struct rds_connection *conn = ic->conn; 822 struct rds_iw_ack_state state = { 0, }; 823 824 rds_poll_cq(ic, &state); 825 ib_req_notify_cq(ic->i_recv_cq, IB_CQ_SOLICITED); 826 rds_poll_cq(ic, &state); 827 828 if (state.ack_next_valid) 829 rds_iw_set_ack(ic, state.ack_next, state.ack_required); 830 if (state.ack_recv_valid && state.ack_recv > ic->i_ack_recv) { 831 rds_send_drop_acked(conn, state.ack_recv, NULL); 832 ic->i_ack_recv = state.ack_recv; 833 } 834 if (rds_conn_up(conn)) 835 rds_iw_attempt_ack(ic); 836 837 /* If we ever end up with a really empty receive ring, we're 838 * in deep trouble, as the sender will definitely see RNR 839 * timeouts. */ 840 if (rds_iw_ring_empty(&ic->i_recv_ring)) 841 rds_iw_stats_inc(s_iw_rx_ring_empty); 842 843 /* 844 * If the ring is running low, then schedule the thread to refill. 845 */ 846 if (rds_iw_ring_low(&ic->i_recv_ring)) 847 queue_delayed_work(rds_wq, &conn->c_recv_w, 0); 848} 849 850int rds_iw_recv(struct rds_connection *conn) 851{ 852 struct rds_iw_connection *ic = conn->c_transport_data; 853 int ret = 0; 854 855 rdsdebug("conn %p\n", conn); 856 857 /* 858 * If we get a temporary posting failure in this context then 859 * we're really low and we want the caller to back off for a bit. 860 */ 861 mutex_lock(&ic->i_recv_mutex); 862 if (rds_iw_recv_refill(conn, GFP_KERNEL, GFP_HIGHUSER, 0)) 863 ret = -ENOMEM; 864 else 865 rds_iw_stats_inc(s_iw_rx_refill_from_thread); 866 mutex_unlock(&ic->i_recv_mutex); 867 868 if (rds_conn_up(conn)) 869 rds_iw_attempt_ack(ic); 870 871 return ret; 872} 873 874int rds_iw_recv_init(void) 875{ 876 struct sysinfo si; 877 int ret = -ENOMEM; 878 879 /* Default to 30% of all available RAM for recv memory */ 880 si_meminfo(&si); 881 rds_iw_sysctl_max_recv_allocation = si.totalram / 3 * PAGE_SIZE / RDS_FRAG_SIZE; 882 883 rds_iw_incoming_slab = kmem_cache_create("rds_iw_incoming", 884 sizeof(struct rds_iw_incoming), 885 0, 0, NULL); 886 if (!rds_iw_incoming_slab) 887 goto out; 888 889 rds_iw_frag_slab = kmem_cache_create("rds_iw_frag", 890 sizeof(struct rds_page_frag), 891 0, 0, NULL); 892 if (!rds_iw_frag_slab) 893 kmem_cache_destroy(rds_iw_incoming_slab); 894 else 895 ret = 0; 896out: 897 return ret; 898} 899 900void rds_iw_recv_exit(void) 901{ 902 kmem_cache_destroy(rds_iw_incoming_slab); 903 kmem_cache_destroy(rds_iw_frag_slab); 904} 905