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 42static struct kmem_cache *rds_ib_incoming_slab; 43static struct kmem_cache *rds_ib_frag_slab; 44static atomic_t rds_ib_allocation = ATOMIC_INIT(0); 45 46void rds_ib_recv_init_ring(struct rds_ib_connection *ic) 47{ 48 struct rds_ib_recv_work *recv; 49 u32 i; 50 51 for (i = 0, recv = ic->i_recvs; i < ic->i_recv_ring.w_nr; i++, recv++) { 52 struct ib_sge *sge; 53 54 recv->r_ibinc = NULL; 55 recv->r_frag = NULL; 56 57 recv->r_wr.next = NULL; 58 recv->r_wr.wr_id = i; 59 recv->r_wr.sg_list = recv->r_sge; 60 recv->r_wr.num_sge = RDS_IB_RECV_SGE; 61 62 sge = &recv->r_sge[0]; 63 sge->addr = ic->i_recv_hdrs_dma + (i * sizeof(struct rds_header)); 64 sge->length = sizeof(struct rds_header); 65 sge->lkey = ic->i_mr->lkey; 66 67 sge = &recv->r_sge[1]; 68 sge->addr = 0; 69 sge->length = RDS_FRAG_SIZE; 70 sge->lkey = ic->i_mr->lkey; 71 } 72} 73 74/* 75 * The entire 'from' list, including the from element itself, is put on 76 * to the tail of the 'to' list. 77 */ 78static void list_splice_entire_tail(struct list_head *from, 79 struct list_head *to) 80{ 81 struct list_head *from_last = from->prev; 82 83 list_splice_tail(from_last, to); 84 list_add_tail(from_last, to); 85} 86 87static void rds_ib_cache_xfer_to_ready(struct rds_ib_refill_cache *cache) 88{ 89 struct list_head *tmp; 90 91 tmp = xchg(&cache->xfer, NULL); 92 if (tmp) { 93 if (cache->ready) 94 list_splice_entire_tail(tmp, cache->ready); 95 else 96 cache->ready = tmp; 97 } 98} 99 100static int rds_ib_recv_alloc_cache(struct rds_ib_refill_cache *cache) 101{ 102 struct rds_ib_cache_head *head; 103 int cpu; 104 105 cache->percpu = alloc_percpu(struct rds_ib_cache_head); 106 if (!cache->percpu) 107 return -ENOMEM; 108 109 for_each_possible_cpu(cpu) { 110 head = per_cpu_ptr(cache->percpu, cpu); 111 head->first = NULL; 112 head->count = 0; 113 } 114 cache->xfer = NULL; 115 cache->ready = NULL; 116 117 return 0; 118} 119 120int rds_ib_recv_alloc_caches(struct rds_ib_connection *ic) 121{ 122 int ret; 123 124 ret = rds_ib_recv_alloc_cache(&ic->i_cache_incs); 125 if (!ret) { 126 ret = rds_ib_recv_alloc_cache(&ic->i_cache_frags); 127 if (ret) 128 free_percpu(ic->i_cache_incs.percpu); 129 } 130 131 return ret; 132} 133 134static void rds_ib_cache_splice_all_lists(struct rds_ib_refill_cache *cache, 135 struct list_head *caller_list) 136{ 137 struct rds_ib_cache_head *head; 138 int cpu; 139 140 for_each_possible_cpu(cpu) { 141 head = per_cpu_ptr(cache->percpu, cpu); 142 if (head->first) { 143 list_splice_entire_tail(head->first, caller_list); 144 head->first = NULL; 145 } 146 } 147 148 if (cache->ready) { 149 list_splice_entire_tail(cache->ready, caller_list); 150 cache->ready = NULL; 151 } 152} 153 154void rds_ib_recv_free_caches(struct rds_ib_connection *ic) 155{ 156 struct rds_ib_incoming *inc; 157 struct rds_ib_incoming *inc_tmp; 158 struct rds_page_frag *frag; 159 struct rds_page_frag *frag_tmp; 160 LIST_HEAD(list); 161 162 rds_ib_cache_xfer_to_ready(&ic->i_cache_incs); 163 rds_ib_cache_splice_all_lists(&ic->i_cache_incs, &list); 164 free_percpu(ic->i_cache_incs.percpu); 165 166 list_for_each_entry_safe(inc, inc_tmp, &list, ii_cache_entry) { 167 list_del(&inc->ii_cache_entry); 168 WARN_ON(!list_empty(&inc->ii_frags)); 169 kmem_cache_free(rds_ib_incoming_slab, inc); 170 } 171 172 rds_ib_cache_xfer_to_ready(&ic->i_cache_frags); 173 rds_ib_cache_splice_all_lists(&ic->i_cache_frags, &list); 174 free_percpu(ic->i_cache_frags.percpu); 175 176 list_for_each_entry_safe(frag, frag_tmp, &list, f_cache_entry) { 177 list_del(&frag->f_cache_entry); 178 WARN_ON(!list_empty(&frag->f_item)); 179 kmem_cache_free(rds_ib_frag_slab, frag); 180 } 181} 182 183/* fwd decl */ 184static void rds_ib_recv_cache_put(struct list_head *new_item, 185 struct rds_ib_refill_cache *cache); 186static struct list_head *rds_ib_recv_cache_get(struct rds_ib_refill_cache *cache); 187 188 189/* Recycle frag and attached recv buffer f_sg */ 190static void rds_ib_frag_free(struct rds_ib_connection *ic, 191 struct rds_page_frag *frag) 192{ 193 rdsdebug("frag %p page %p\n", frag, sg_page(&frag->f_sg)); 194 195 rds_ib_recv_cache_put(&frag->f_cache_entry, &ic->i_cache_frags); 196} 197 198/* Recycle inc after freeing attached frags */ 199void rds_ib_inc_free(struct rds_incoming *inc) 200{ 201 struct rds_ib_incoming *ibinc; 202 struct rds_page_frag *frag; 203 struct rds_page_frag *pos; 204 struct rds_ib_connection *ic = inc->i_conn->c_transport_data; 205 206 ibinc = container_of(inc, struct rds_ib_incoming, ii_inc); 207 208 /* Free attached frags */ 209 list_for_each_entry_safe(frag, pos, &ibinc->ii_frags, f_item) { 210 list_del_init(&frag->f_item); 211 rds_ib_frag_free(ic, frag); 212 } 213 BUG_ON(!list_empty(&ibinc->ii_frags)); 214 215 rdsdebug("freeing ibinc %p inc %p\n", ibinc, inc); 216 rds_ib_recv_cache_put(&ibinc->ii_cache_entry, &ic->i_cache_incs); 217} 218 219static void rds_ib_recv_clear_one(struct rds_ib_connection *ic, 220 struct rds_ib_recv_work *recv) 221{ 222 if (recv->r_ibinc) { 223 rds_inc_put(&recv->r_ibinc->ii_inc); 224 recv->r_ibinc = NULL; 225 } 226 if (recv->r_frag) { 227 ib_dma_unmap_sg(ic->i_cm_id->device, &recv->r_frag->f_sg, 1, DMA_FROM_DEVICE); 228 rds_ib_frag_free(ic, recv->r_frag); 229 recv->r_frag = NULL; 230 } 231} 232 233void rds_ib_recv_clear_ring(struct rds_ib_connection *ic) 234{ 235 u32 i; 236 237 for (i = 0; i < ic->i_recv_ring.w_nr; i++) 238 rds_ib_recv_clear_one(ic, &ic->i_recvs[i]); 239} 240 241static struct rds_ib_incoming *rds_ib_refill_one_inc(struct rds_ib_connection *ic, 242 gfp_t slab_mask) 243{ 244 struct rds_ib_incoming *ibinc; 245 struct list_head *cache_item; 246 int avail_allocs; 247 248 cache_item = rds_ib_recv_cache_get(&ic->i_cache_incs); 249 if (cache_item) { 250 ibinc = container_of(cache_item, struct rds_ib_incoming, ii_cache_entry); 251 } else { 252 avail_allocs = atomic_add_unless(&rds_ib_allocation, 253 1, rds_ib_sysctl_max_recv_allocation); 254 if (!avail_allocs) { 255 rds_ib_stats_inc(s_ib_rx_alloc_limit); 256 return NULL; 257 } 258 ibinc = kmem_cache_alloc(rds_ib_incoming_slab, slab_mask); 259 if (!ibinc) { 260 atomic_dec(&rds_ib_allocation); 261 return NULL; 262 } 263 } 264 INIT_LIST_HEAD(&ibinc->ii_frags); 265 rds_inc_init(&ibinc->ii_inc, ic->conn, ic->conn->c_faddr); 266 267 return ibinc; 268} 269 270static struct rds_page_frag *rds_ib_refill_one_frag(struct rds_ib_connection *ic, 271 gfp_t slab_mask, gfp_t page_mask) 272{ 273 struct rds_page_frag *frag; 274 struct list_head *cache_item; 275 int ret; 276 277 cache_item = rds_ib_recv_cache_get(&ic->i_cache_frags); 278 if (cache_item) { 279 frag = container_of(cache_item, struct rds_page_frag, f_cache_entry); 280 } else { 281 frag = kmem_cache_alloc(rds_ib_frag_slab, slab_mask); 282 if (!frag) 283 return NULL; 284 285 sg_init_table(&frag->f_sg, 1); 286 ret = rds_page_remainder_alloc(&frag->f_sg, 287 RDS_FRAG_SIZE, page_mask); 288 if (ret) { 289 kmem_cache_free(rds_ib_frag_slab, frag); 290 return NULL; 291 } 292 } 293 294 INIT_LIST_HEAD(&frag->f_item); 295 296 return frag; 297} 298 299static int rds_ib_recv_refill_one(struct rds_connection *conn, 300 struct rds_ib_recv_work *recv, int prefill) 301{ 302 struct rds_ib_connection *ic = conn->c_transport_data; 303 struct ib_sge *sge; 304 int ret = -ENOMEM; 305 gfp_t slab_mask = GFP_NOWAIT; 306 gfp_t page_mask = GFP_NOWAIT; 307 308 if (prefill) { 309 slab_mask = GFP_KERNEL; 310 page_mask = GFP_HIGHUSER; 311 } 312 313 if (!ic->i_cache_incs.ready) 314 rds_ib_cache_xfer_to_ready(&ic->i_cache_incs); 315 if (!ic->i_cache_frags.ready) 316 rds_ib_cache_xfer_to_ready(&ic->i_cache_frags); 317 318 /* 319 * ibinc was taken from recv if recv contained the start of a message. 320 * recvs that were continuations will still have this allocated. 321 */ 322 if (!recv->r_ibinc) { 323 recv->r_ibinc = rds_ib_refill_one_inc(ic, slab_mask); 324 if (!recv->r_ibinc) 325 goto out; 326 } 327 328 WARN_ON(recv->r_frag); /* leak! */ 329 recv->r_frag = rds_ib_refill_one_frag(ic, slab_mask, page_mask); 330 if (!recv->r_frag) 331 goto out; 332 333 ret = ib_dma_map_sg(ic->i_cm_id->device, &recv->r_frag->f_sg, 334 1, DMA_FROM_DEVICE); 335 WARN_ON(ret != 1); 336 337 sge = &recv->r_sge[0]; 338 sge->addr = ic->i_recv_hdrs_dma + (recv - ic->i_recvs) * sizeof(struct rds_header); 339 sge->length = sizeof(struct rds_header); 340 341 sge = &recv->r_sge[1]; 342 sge->addr = ib_sg_dma_address(ic->i_cm_id->device, &recv->r_frag->f_sg); 343 sge->length = ib_sg_dma_len(ic->i_cm_id->device, &recv->r_frag->f_sg); 344 345 ret = 0; 346out: 347 return ret; 348} 349 350/* 351 * This tries to allocate and post unused work requests after making sure that 352 * they have all the allocations they need to queue received fragments into 353 * sockets. 354 * 355 * -1 is returned if posting fails due to temporary resource exhaustion. 356 */ 357void rds_ib_recv_refill(struct rds_connection *conn, int prefill) 358{ 359 struct rds_ib_connection *ic = conn->c_transport_data; 360 struct rds_ib_recv_work *recv; 361 struct ib_recv_wr *failed_wr; 362 unsigned int posted = 0; 363 int ret = 0; 364 u32 pos; 365 366 while ((prefill || rds_conn_up(conn)) && 367 rds_ib_ring_alloc(&ic->i_recv_ring, 1, &pos)) { 368 if (pos >= ic->i_recv_ring.w_nr) { 369 printk(KERN_NOTICE "Argh - ring alloc returned pos=%u\n", 370 pos); 371 break; 372 } 373 374 recv = &ic->i_recvs[pos]; 375 ret = rds_ib_recv_refill_one(conn, recv, prefill); 376 if (ret) { 377 break; 378 } 379 380 /* XXX when can this fail? */ 381 ret = ib_post_recv(ic->i_cm_id->qp, &recv->r_wr, &failed_wr); 382 rdsdebug("recv %p ibinc %p page %p addr %lu ret %d\n", recv, 383 recv->r_ibinc, sg_page(&recv->r_frag->f_sg), 384 (long) ib_sg_dma_address( 385 ic->i_cm_id->device, 386 &recv->r_frag->f_sg), 387 ret); 388 if (ret) { 389 rds_ib_conn_error(conn, "recv post on " 390 "%pI4 returned %d, disconnecting and " 391 "reconnecting\n", &conn->c_faddr, 392 ret); 393 break; 394 } 395 396 posted++; 397 } 398 399 /* We're doing flow control - update the window. */ 400 if (ic->i_flowctl && posted) 401 rds_ib_advertise_credits(conn, posted); 402 403 if (ret) 404 rds_ib_ring_unalloc(&ic->i_recv_ring, 1); 405} 406 407/* 408 * We want to recycle several types of recv allocations, like incs and frags. 409 * To use this, the *_free() function passes in the ptr to a list_head within 410 * the recyclee, as well as the cache to put it on. 411 * 412 * First, we put the memory on a percpu list. When this reaches a certain size, 413 * We move it to an intermediate non-percpu list in a lockless manner, with some 414 * xchg/compxchg wizardry. 415 * 416 * N.B. Instead of a list_head as the anchor, we use a single pointer, which can 417 * be NULL and xchg'd. The list is actually empty when the pointer is NULL, and 418 * list_empty() will return true with one element is actually present. 419 */ 420static void rds_ib_recv_cache_put(struct list_head *new_item, 421 struct rds_ib_refill_cache *cache) 422{ 423 unsigned long flags; 424 struct list_head *old, *chpfirst; 425 426 local_irq_save(flags); 427 428 chpfirst = __this_cpu_read(cache->percpu->first); 429 if (!chpfirst) 430 INIT_LIST_HEAD(new_item); 431 else /* put on front */ 432 list_add_tail(new_item, chpfirst); 433 434 __this_cpu_write(cache->percpu->first, new_item); 435 __this_cpu_inc(cache->percpu->count); 436 437 if (__this_cpu_read(cache->percpu->count) < RDS_IB_RECYCLE_BATCH_COUNT) 438 goto end; 439 440 /* 441 * Return our per-cpu first list to the cache's xfer by atomically 442 * grabbing the current xfer list, appending it to our per-cpu list, 443 * and then atomically returning that entire list back to the 444 * cache's xfer list as long as it's still empty. 445 */ 446 do { 447 old = xchg(&cache->xfer, NULL); 448 if (old) 449 list_splice_entire_tail(old, chpfirst); 450 old = cmpxchg(&cache->xfer, NULL, chpfirst); 451 } while (old); 452 453 454 __this_cpu_write(cache->percpu->first, NULL); 455 __this_cpu_write(cache->percpu->count, 0); 456end: 457 local_irq_restore(flags); 458} 459 460static struct list_head *rds_ib_recv_cache_get(struct rds_ib_refill_cache *cache) 461{ 462 struct list_head *head = cache->ready; 463 464 if (head) { 465 if (!list_empty(head)) { 466 cache->ready = head->next; 467 list_del_init(head); 468 } else 469 cache->ready = NULL; 470 } 471 472 return head; 473} 474 475int rds_ib_inc_copy_to_user(struct rds_incoming *inc, struct iovec *first_iov, 476 size_t size) 477{ 478 struct rds_ib_incoming *ibinc; 479 struct rds_page_frag *frag; 480 struct iovec *iov = first_iov; 481 unsigned long to_copy; 482 unsigned long frag_off = 0; 483 unsigned long iov_off = 0; 484 int copied = 0; 485 int ret; 486 u32 len; 487 488 ibinc = container_of(inc, struct rds_ib_incoming, ii_inc); 489 frag = list_entry(ibinc->ii_frags.next, struct rds_page_frag, f_item); 490 len = be32_to_cpu(inc->i_hdr.h_len); 491 492 while (copied < size && copied < len) { 493 if (frag_off == RDS_FRAG_SIZE) { 494 frag = list_entry(frag->f_item.next, 495 struct rds_page_frag, f_item); 496 frag_off = 0; 497 } 498 while (iov_off == iov->iov_len) { 499 iov_off = 0; 500 iov++; 501 } 502 503 to_copy = min(iov->iov_len - iov_off, RDS_FRAG_SIZE - frag_off); 504 to_copy = min_t(size_t, to_copy, size - copied); 505 to_copy = min_t(unsigned long, to_copy, len - copied); 506 507 rdsdebug("%lu bytes to user [%p, %zu] + %lu from frag " 508 "[%p, %u] + %lu\n", 509 to_copy, iov->iov_base, iov->iov_len, iov_off, 510 sg_page(&frag->f_sg), frag->f_sg.offset, frag_off); 511 512 /* XXX needs + offset for multiple recvs per page */ 513 ret = rds_page_copy_to_user(sg_page(&frag->f_sg), 514 frag->f_sg.offset + frag_off, 515 iov->iov_base + iov_off, 516 to_copy); 517 if (ret) { 518 copied = ret; 519 break; 520 } 521 522 iov_off += to_copy; 523 frag_off += to_copy; 524 copied += to_copy; 525 } 526 527 return copied; 528} 529 530/* ic starts out kzalloc()ed */ 531void rds_ib_recv_init_ack(struct rds_ib_connection *ic) 532{ 533 struct ib_send_wr *wr = &ic->i_ack_wr; 534 struct ib_sge *sge = &ic->i_ack_sge; 535 536 sge->addr = ic->i_ack_dma; 537 sge->length = sizeof(struct rds_header); 538 sge->lkey = ic->i_mr->lkey; 539 540 wr->sg_list = sge; 541 wr->num_sge = 1; 542 wr->opcode = IB_WR_SEND; 543 wr->wr_id = RDS_IB_ACK_WR_ID; 544 wr->send_flags = IB_SEND_SIGNALED | IB_SEND_SOLICITED; 545} 546 547/* 548 * You'd think that with reliable IB connections you wouldn't need to ack 549 * messages that have been received. The problem is that IB hardware generates 550 * an ack message before it has DMAed the message into memory. This creates a 551 * potential message loss if the HCA is disabled for any reason between when it 552 * sends the ack and before the message is DMAed and processed. This is only a 553 * potential issue if another HCA is available for fail-over. 554 * 555 * When the remote host receives our ack they'll free the sent message from 556 * their send queue. To decrease the latency of this we always send an ack 557 * immediately after we've received messages. 558 * 559 * For simplicity, we only have one ack in flight at a time. This puts 560 * pressure on senders to have deep enough send queues to absorb the latency of 561 * a single ack frame being in flight. This might not be good enough. 562 * 563 * This is implemented by have a long-lived send_wr and sge which point to a 564 * statically allocated ack frame. This ack wr does not fall under the ring 565 * accounting that the tx and rx wrs do. The QP attribute specifically makes 566 * room for it beyond the ring size. Send completion notices its special 567 * wr_id and avoids working with the ring in that case. 568 */ 569#ifndef KERNEL_HAS_ATOMIC64 570static void rds_ib_set_ack(struct rds_ib_connection *ic, u64 seq, 571 int ack_required) 572{ 573 unsigned long flags; 574 575 spin_lock_irqsave(&ic->i_ack_lock, flags); 576 ic->i_ack_next = seq; 577 if (ack_required) 578 set_bit(IB_ACK_REQUESTED, &ic->i_ack_flags); 579 spin_unlock_irqrestore(&ic->i_ack_lock, flags); 580} 581 582static u64 rds_ib_get_ack(struct rds_ib_connection *ic) 583{ 584 unsigned long flags; 585 u64 seq; 586 587 clear_bit(IB_ACK_REQUESTED, &ic->i_ack_flags); 588 589 spin_lock_irqsave(&ic->i_ack_lock, flags); 590 seq = ic->i_ack_next; 591 spin_unlock_irqrestore(&ic->i_ack_lock, flags); 592 593 return seq; 594} 595#else 596static void rds_ib_set_ack(struct rds_ib_connection *ic, u64 seq, 597 int ack_required) 598{ 599 atomic64_set(&ic->i_ack_next, seq); 600 if (ack_required) { 601 smp_mb__before_atomic(); 602 set_bit(IB_ACK_REQUESTED, &ic->i_ack_flags); 603 } 604} 605 606static u64 rds_ib_get_ack(struct rds_ib_connection *ic) 607{ 608 clear_bit(IB_ACK_REQUESTED, &ic->i_ack_flags); 609 smp_mb__after_atomic(); 610 611 return atomic64_read(&ic->i_ack_next); 612} 613#endif 614 615 616static void rds_ib_send_ack(struct rds_ib_connection *ic, unsigned int adv_credits) 617{ 618 struct rds_header *hdr = ic->i_ack; 619 struct ib_send_wr *failed_wr; 620 u64 seq; 621 int ret; 622 623 seq = rds_ib_get_ack(ic); 624 625 rdsdebug("send_ack: ic %p ack %llu\n", ic, (unsigned long long) seq); 626 rds_message_populate_header(hdr, 0, 0, 0); 627 hdr->h_ack = cpu_to_be64(seq); 628 hdr->h_credit = adv_credits; 629 rds_message_make_checksum(hdr); 630 ic->i_ack_queued = jiffies; 631 632 ret = ib_post_send(ic->i_cm_id->qp, &ic->i_ack_wr, &failed_wr); 633 if (unlikely(ret)) { 634 /* Failed to send. Release the WR, and 635 * force another ACK. 636 */ 637 clear_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags); 638 set_bit(IB_ACK_REQUESTED, &ic->i_ack_flags); 639 640 rds_ib_stats_inc(s_ib_ack_send_failure); 641 642 rds_ib_conn_error(ic->conn, "sending ack failed\n"); 643 } else 644 rds_ib_stats_inc(s_ib_ack_sent); 645} 646 647/* 648 * There are 3 ways of getting acknowledgements to the peer: 649 * 1. We call rds_ib_attempt_ack from the recv completion handler 650 * to send an ACK-only frame. 651 * However, there can be only one such frame in the send queue 652 * at any time, so we may have to postpone it. 653 * 2. When another (data) packet is transmitted while there's 654 * an ACK in the queue, we piggyback the ACK sequence number 655 * on the data packet. 656 * 3. If the ACK WR is done sending, we get called from the 657 * send queue completion handler, and check whether there's 658 * another ACK pending (postponed because the WR was on the 659 * queue). If so, we transmit it. 660 * 661 * We maintain 2 variables: 662 * - i_ack_flags, which keeps track of whether the ACK WR 663 * is currently in the send queue or not (IB_ACK_IN_FLIGHT) 664 * - i_ack_next, which is the last sequence number we received 665 * 666 * Potentially, send queue and receive queue handlers can run concurrently. 667 * It would be nice to not have to use a spinlock to synchronize things, 668 * but the one problem that rules this out is that 64bit updates are 669 * not atomic on all platforms. Things would be a lot simpler if 670 * we had atomic64 or maybe cmpxchg64 everywhere. 671 * 672 * Reconnecting complicates this picture just slightly. When we 673 * reconnect, we may be seeing duplicate packets. The peer 674 * is retransmitting them, because it hasn't seen an ACK for 675 * them. It is important that we ACK these. 676 * 677 * ACK mitigation adds a header flag "ACK_REQUIRED"; any packet with 678 * this flag set *MUST* be acknowledged immediately. 679 */ 680 681/* 682 * When we get here, we're called from the recv queue handler. 683 * Check whether we ought to transmit an ACK. 684 */ 685void rds_ib_attempt_ack(struct rds_ib_connection *ic) 686{ 687 unsigned int adv_credits; 688 689 if (!test_bit(IB_ACK_REQUESTED, &ic->i_ack_flags)) 690 return; 691 692 if (test_and_set_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags)) { 693 rds_ib_stats_inc(s_ib_ack_send_delayed); 694 return; 695 } 696 697 /* Can we get a send credit? */ 698 if (!rds_ib_send_grab_credits(ic, 1, &adv_credits, 0, RDS_MAX_ADV_CREDIT)) { 699 rds_ib_stats_inc(s_ib_tx_throttle); 700 clear_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags); 701 return; 702 } 703 704 clear_bit(IB_ACK_REQUESTED, &ic->i_ack_flags); 705 rds_ib_send_ack(ic, adv_credits); 706} 707 708/* 709 * We get here from the send completion handler, when the 710 * adapter tells us the ACK frame was sent. 711 */ 712void rds_ib_ack_send_complete(struct rds_ib_connection *ic) 713{ 714 clear_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags); 715 rds_ib_attempt_ack(ic); 716} 717 718/* 719 * This is called by the regular xmit code when it wants to piggyback 720 * an ACK on an outgoing frame. 721 */ 722u64 rds_ib_piggyb_ack(struct rds_ib_connection *ic) 723{ 724 if (test_and_clear_bit(IB_ACK_REQUESTED, &ic->i_ack_flags)) 725 rds_ib_stats_inc(s_ib_ack_send_piggybacked); 726 return rds_ib_get_ack(ic); 727} 728 729/* 730 * It's kind of lame that we're copying from the posted receive pages into 731 * long-lived bitmaps. We could have posted the bitmaps and rdma written into 732 * them. But receiving new congestion bitmaps should be a *rare* event, so 733 * hopefully we won't need to invest that complexity in making it more 734 * efficient. By copying we can share a simpler core with TCP which has to 735 * copy. 736 */ 737static void rds_ib_cong_recv(struct rds_connection *conn, 738 struct rds_ib_incoming *ibinc) 739{ 740 struct rds_cong_map *map; 741 unsigned int map_off; 742 unsigned int map_page; 743 struct rds_page_frag *frag; 744 unsigned long frag_off; 745 unsigned long to_copy; 746 unsigned long copied; 747 uint64_t uncongested = 0; 748 void *addr; 749 750 /* catch completely corrupt packets */ 751 if (be32_to_cpu(ibinc->ii_inc.i_hdr.h_len) != RDS_CONG_MAP_BYTES) 752 return; 753 754 map = conn->c_fcong; 755 map_page = 0; 756 map_off = 0; 757 758 frag = list_entry(ibinc->ii_frags.next, struct rds_page_frag, f_item); 759 frag_off = 0; 760 761 copied = 0; 762 763 while (copied < RDS_CONG_MAP_BYTES) { 764 uint64_t *src, *dst; 765 unsigned int k; 766 767 to_copy = min(RDS_FRAG_SIZE - frag_off, PAGE_SIZE - map_off); 768 BUG_ON(to_copy & 7); /* Must be 64bit aligned. */ 769 770 addr = kmap_atomic(sg_page(&frag->f_sg)); 771 772 src = addr + frag_off; 773 dst = (void *)map->m_page_addrs[map_page] + map_off; 774 for (k = 0; k < to_copy; k += 8) { 775 /* Record ports that became uncongested, ie 776 * bits that changed from 0 to 1. */ 777 uncongested |= ~(*src) & *dst; 778 *dst++ = *src++; 779 } 780 kunmap_atomic(addr); 781 782 copied += to_copy; 783 784 map_off += to_copy; 785 if (map_off == PAGE_SIZE) { 786 map_off = 0; 787 map_page++; 788 } 789 790 frag_off += to_copy; 791 if (frag_off == RDS_FRAG_SIZE) { 792 frag = list_entry(frag->f_item.next, 793 struct rds_page_frag, f_item); 794 frag_off = 0; 795 } 796 } 797 798 /* the congestion map is in little endian order */ 799 uncongested = le64_to_cpu(uncongested); 800 801 rds_cong_map_updated(map, uncongested); 802} 803 804/* 805 * Rings are posted with all the allocations they'll need to queue the 806 * incoming message to the receiving socket so this can't fail. 807 * All fragments start with a header, so we can make sure we're not receiving 808 * garbage, and we can tell a small 8 byte fragment from an ACK frame. 809 */ 810struct rds_ib_ack_state { 811 u64 ack_next; 812 u64 ack_recv; 813 unsigned int ack_required:1; 814 unsigned int ack_next_valid:1; 815 unsigned int ack_recv_valid:1; 816}; 817 818static void rds_ib_process_recv(struct rds_connection *conn, 819 struct rds_ib_recv_work *recv, u32 data_len, 820 struct rds_ib_ack_state *state) 821{ 822 struct rds_ib_connection *ic = conn->c_transport_data; 823 struct rds_ib_incoming *ibinc = ic->i_ibinc; 824 struct rds_header *ihdr, *hdr; 825 826 /* XXX shut down the connection if port 0,0 are seen? */ 827 828 rdsdebug("ic %p ibinc %p recv %p byte len %u\n", ic, ibinc, recv, 829 data_len); 830 831 if (data_len < sizeof(struct rds_header)) { 832 rds_ib_conn_error(conn, "incoming message " 833 "from %pI4 didn't include a " 834 "header, disconnecting and " 835 "reconnecting\n", 836 &conn->c_faddr); 837 return; 838 } 839 data_len -= sizeof(struct rds_header); 840 841 ihdr = &ic->i_recv_hdrs[recv - ic->i_recvs]; 842 843 /* Validate the checksum. */ 844 if (!rds_message_verify_checksum(ihdr)) { 845 rds_ib_conn_error(conn, "incoming message " 846 "from %pI4 has corrupted header - " 847 "forcing a reconnect\n", 848 &conn->c_faddr); 849 rds_stats_inc(s_recv_drop_bad_checksum); 850 return; 851 } 852 853 /* Process the ACK sequence which comes with every packet */ 854 state->ack_recv = be64_to_cpu(ihdr->h_ack); 855 state->ack_recv_valid = 1; 856 857 /* Process the credits update if there was one */ 858 if (ihdr->h_credit) 859 rds_ib_send_add_credits(conn, ihdr->h_credit); 860 861 if (ihdr->h_sport == 0 && ihdr->h_dport == 0 && data_len == 0) { 862 /* This is an ACK-only packet. The fact that it gets 863 * special treatment here is that historically, ACKs 864 * were rather special beasts. 865 */ 866 rds_ib_stats_inc(s_ib_ack_received); 867 868 /* 869 * Usually the frags make their way on to incs and are then freed as 870 * the inc is freed. We don't go that route, so we have to drop the 871 * page ref ourselves. We can't just leave the page on the recv 872 * because that confuses the dma mapping of pages and each recv's use 873 * of a partial page. 874 * 875 * FIXME: Fold this into the code path below. 876 */ 877 rds_ib_frag_free(ic, recv->r_frag); 878 recv->r_frag = NULL; 879 return; 880 } 881 882 /* 883 * If we don't already have an inc on the connection then this 884 * fragment has a header and starts a message.. copy its header 885 * into the inc and save the inc so we can hang upcoming fragments 886 * off its list. 887 */ 888 if (!ibinc) { 889 ibinc = recv->r_ibinc; 890 recv->r_ibinc = NULL; 891 ic->i_ibinc = ibinc; 892 893 hdr = &ibinc->ii_inc.i_hdr; 894 memcpy(hdr, ihdr, sizeof(*hdr)); 895 ic->i_recv_data_rem = be32_to_cpu(hdr->h_len); 896 897 rdsdebug("ic %p ibinc %p rem %u flag 0x%x\n", ic, ibinc, 898 ic->i_recv_data_rem, hdr->h_flags); 899 } else { 900 hdr = &ibinc->ii_inc.i_hdr; 901 /* We can't just use memcmp here; fragments of a 902 * single message may carry different ACKs */ 903 if (hdr->h_sequence != ihdr->h_sequence || 904 hdr->h_len != ihdr->h_len || 905 hdr->h_sport != ihdr->h_sport || 906 hdr->h_dport != ihdr->h_dport) { 907 rds_ib_conn_error(conn, 908 "fragment header mismatch; forcing reconnect\n"); 909 return; 910 } 911 } 912 913 list_add_tail(&recv->r_frag->f_item, &ibinc->ii_frags); 914 recv->r_frag = NULL; 915 916 if (ic->i_recv_data_rem > RDS_FRAG_SIZE) 917 ic->i_recv_data_rem -= RDS_FRAG_SIZE; 918 else { 919 ic->i_recv_data_rem = 0; 920 ic->i_ibinc = NULL; 921 922 if (ibinc->ii_inc.i_hdr.h_flags == RDS_FLAG_CONG_BITMAP) 923 rds_ib_cong_recv(conn, ibinc); 924 else { 925 rds_recv_incoming(conn, conn->c_faddr, conn->c_laddr, 926 &ibinc->ii_inc, GFP_ATOMIC); 927 state->ack_next = be64_to_cpu(hdr->h_sequence); 928 state->ack_next_valid = 1; 929 } 930 931 /* Evaluate the ACK_REQUIRED flag *after* we received 932 * the complete frame, and after bumping the next_rx 933 * sequence. */ 934 if (hdr->h_flags & RDS_FLAG_ACK_REQUIRED) { 935 rds_stats_inc(s_recv_ack_required); 936 state->ack_required = 1; 937 } 938 939 rds_inc_put(&ibinc->ii_inc); 940 } 941} 942 943/* 944 * Plucking the oldest entry from the ring can be done concurrently with 945 * the thread refilling the ring. Each ring operation is protected by 946 * spinlocks and the transient state of refilling doesn't change the 947 * recording of which entry is oldest. 948 * 949 * This relies on IB only calling one cq comp_handler for each cq so that 950 * there will only be one caller of rds_recv_incoming() per RDS connection. 951 */ 952void rds_ib_recv_cq_comp_handler(struct ib_cq *cq, void *context) 953{ 954 struct rds_connection *conn = context; 955 struct rds_ib_connection *ic = conn->c_transport_data; 956 957 rdsdebug("conn %p cq %p\n", conn, cq); 958 959 rds_ib_stats_inc(s_ib_rx_cq_call); 960 961 tasklet_schedule(&ic->i_recv_tasklet); 962} 963 964static inline void rds_poll_cq(struct rds_ib_connection *ic, 965 struct rds_ib_ack_state *state) 966{ 967 struct rds_connection *conn = ic->conn; 968 struct ib_wc wc; 969 struct rds_ib_recv_work *recv; 970 971 while (ib_poll_cq(ic->i_recv_cq, 1, &wc) > 0) { 972 rdsdebug("wc wr_id 0x%llx status %u (%s) byte_len %u imm_data %u\n", 973 (unsigned long long)wc.wr_id, wc.status, 974 rds_ib_wc_status_str(wc.status), wc.byte_len, 975 be32_to_cpu(wc.ex.imm_data)); 976 rds_ib_stats_inc(s_ib_rx_cq_event); 977 978 recv = &ic->i_recvs[rds_ib_ring_oldest(&ic->i_recv_ring)]; 979 980 ib_dma_unmap_sg(ic->i_cm_id->device, &recv->r_frag->f_sg, 1, DMA_FROM_DEVICE); 981 982 /* 983 * Also process recvs in connecting state because it is possible 984 * to get a recv completion _before_ the rdmacm ESTABLISHED 985 * event is processed. 986 */ 987 if (wc.status == IB_WC_SUCCESS) { 988 rds_ib_process_recv(conn, recv, wc.byte_len, state); 989 } else { 990 /* We expect errors as the qp is drained during shutdown */ 991 if (rds_conn_up(conn) || rds_conn_connecting(conn)) 992 rds_ib_conn_error(conn, "recv completion on %pI4 had " 993 "status %u (%s), disconnecting and " 994 "reconnecting\n", &conn->c_faddr, 995 wc.status, 996 rds_ib_wc_status_str(wc.status)); 997 } 998 999 /* 1000 * It's very important that we only free this ring entry if we've truly 1001 * freed the resources allocated to the entry. The refilling path can 1002 * leak if we don't. 1003 */ 1004 rds_ib_ring_free(&ic->i_recv_ring, 1); 1005 } 1006} 1007 1008void rds_ib_recv_tasklet_fn(unsigned long data) 1009{ 1010 struct rds_ib_connection *ic = (struct rds_ib_connection *) data; 1011 struct rds_connection *conn = ic->conn; 1012 struct rds_ib_ack_state state = { 0, }; 1013 1014 rds_poll_cq(ic, &state); 1015 ib_req_notify_cq(ic->i_recv_cq, IB_CQ_SOLICITED); 1016 rds_poll_cq(ic, &state); 1017 1018 if (state.ack_next_valid) 1019 rds_ib_set_ack(ic, state.ack_next, state.ack_required); 1020 if (state.ack_recv_valid && state.ack_recv > ic->i_ack_recv) { 1021 rds_send_drop_acked(conn, state.ack_recv, NULL); 1022 ic->i_ack_recv = state.ack_recv; 1023 } 1024 if (rds_conn_up(conn)) 1025 rds_ib_attempt_ack(ic); 1026 1027 /* If we ever end up with a really empty receive ring, we're 1028 * in deep trouble, as the sender will definitely see RNR 1029 * timeouts. */ 1030 if (rds_ib_ring_empty(&ic->i_recv_ring)) 1031 rds_ib_stats_inc(s_ib_rx_ring_empty); 1032 1033 if (rds_ib_ring_low(&ic->i_recv_ring)) 1034 rds_ib_recv_refill(conn, 0); 1035} 1036 1037int rds_ib_recv(struct rds_connection *conn) 1038{ 1039 struct rds_ib_connection *ic = conn->c_transport_data; 1040 int ret = 0; 1041 1042 rdsdebug("conn %p\n", conn); 1043 if (rds_conn_up(conn)) 1044 rds_ib_attempt_ack(ic); 1045 1046 return ret; 1047} 1048 1049int rds_ib_recv_init(void) 1050{ 1051 struct sysinfo si; 1052 int ret = -ENOMEM; 1053 1054 /* Default to 30% of all available RAM for recv memory */ 1055 si_meminfo(&si); 1056 rds_ib_sysctl_max_recv_allocation = si.totalram / 3 * PAGE_SIZE / RDS_FRAG_SIZE; 1057 1058 rds_ib_incoming_slab = kmem_cache_create("rds_ib_incoming", 1059 sizeof(struct rds_ib_incoming), 1060 0, SLAB_HWCACHE_ALIGN, NULL); 1061 if (!rds_ib_incoming_slab) 1062 goto out; 1063 1064 rds_ib_frag_slab = kmem_cache_create("rds_ib_frag", 1065 sizeof(struct rds_page_frag), 1066 0, SLAB_HWCACHE_ALIGN, NULL); 1067 if (!rds_ib_frag_slab) 1068 kmem_cache_destroy(rds_ib_incoming_slab); 1069 else 1070 ret = 0; 1071out: 1072 return ret; 1073} 1074 1075void rds_ib_recv_exit(void) 1076{ 1077 kmem_cache_destroy(rds_ib_incoming_slab); 1078 kmem_cache_destroy(rds_ib_frag_slab); 1079} 1080