1 /* 2 * Copyright (c) 2006, 2017 Oracle and/or its affiliates. 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_single_path.h" 40 #include "rds.h" 41 #include "ib.h" 42 43 static struct kmem_cache *rds_ib_incoming_slab; 44 static struct kmem_cache *rds_ib_frag_slab; 45 static atomic_t rds_ib_allocation = ATOMIC_INIT(0); 46 47 void rds_ib_recv_init_ring(struct rds_ib_connection *ic) 48 { 49 struct rds_ib_recv_work *recv; 50 u32 i; 51 52 for (i = 0, recv = ic->i_recvs; i < ic->i_recv_ring.w_nr; i++, recv++) { 53 struct ib_sge *sge; 54 55 recv->r_ibinc = NULL; 56 recv->r_frag = NULL; 57 58 recv->r_wr.next = NULL; 59 recv->r_wr.wr_id = i; 60 recv->r_wr.sg_list = recv->r_sge; 61 recv->r_wr.num_sge = RDS_IB_RECV_SGE; 62 63 sge = &recv->r_sge[0]; 64 sge->addr = ic->i_recv_hdrs_dma + (i * sizeof(struct rds_header)); 65 sge->length = sizeof(struct rds_header); 66 sge->lkey = ic->i_pd->local_dma_lkey; 67 68 sge = &recv->r_sge[1]; 69 sge->addr = 0; 70 sge->length = RDS_FRAG_SIZE; 71 sge->lkey = ic->i_pd->local_dma_lkey; 72 } 73 } 74 75 /* 76 * The entire 'from' list, including the from element itself, is put on 77 * to the tail of the 'to' list. 78 */ 79 static void list_splice_entire_tail(struct list_head *from, 80 struct list_head *to) 81 { 82 struct list_head *from_last = from->prev; 83 84 list_splice_tail(from_last, to); 85 list_add_tail(from_last, to); 86 } 87 88 static void rds_ib_cache_xfer_to_ready(struct rds_ib_refill_cache *cache) 89 { 90 struct list_head *tmp; 91 92 tmp = xchg(&cache->xfer, NULL); 93 if (tmp) { 94 if (cache->ready) 95 list_splice_entire_tail(tmp, cache->ready); 96 else 97 cache->ready = tmp; 98 } 99 } 100 101 static int rds_ib_recv_alloc_cache(struct rds_ib_refill_cache *cache, gfp_t gfp) 102 { 103 struct rds_ib_cache_head *head; 104 int cpu; 105 106 cache->percpu = alloc_percpu_gfp(struct rds_ib_cache_head, gfp); 107 if (!cache->percpu) 108 return -ENOMEM; 109 110 for_each_possible_cpu(cpu) { 111 head = per_cpu_ptr(cache->percpu, cpu); 112 head->first = NULL; 113 head->count = 0; 114 } 115 cache->xfer = NULL; 116 cache->ready = NULL; 117 118 return 0; 119 } 120 121 int rds_ib_recv_alloc_caches(struct rds_ib_connection *ic, gfp_t gfp) 122 { 123 int ret; 124 125 ret = rds_ib_recv_alloc_cache(&ic->i_cache_incs, gfp); 126 if (!ret) { 127 ret = rds_ib_recv_alloc_cache(&ic->i_cache_frags, gfp); 128 if (ret) 129 free_percpu(ic->i_cache_incs.percpu); 130 } 131 132 return ret; 133 } 134 135 static void rds_ib_cache_splice_all_lists(struct rds_ib_refill_cache *cache, 136 struct list_head *caller_list) 137 { 138 struct rds_ib_cache_head *head; 139 int cpu; 140 141 for_each_possible_cpu(cpu) { 142 head = per_cpu_ptr(cache->percpu, cpu); 143 if (head->first) { 144 list_splice_entire_tail(head->first, caller_list); 145 head->first = NULL; 146 } 147 } 148 149 if (cache->ready) { 150 list_splice_entire_tail(cache->ready, caller_list); 151 cache->ready = NULL; 152 } 153 } 154 155 void rds_ib_recv_free_caches(struct rds_ib_connection *ic) 156 { 157 struct rds_ib_incoming *inc; 158 struct rds_ib_incoming *inc_tmp; 159 struct rds_page_frag *frag; 160 struct rds_page_frag *frag_tmp; 161 LIST_HEAD(list); 162 163 rds_ib_cache_xfer_to_ready(&ic->i_cache_incs); 164 rds_ib_cache_splice_all_lists(&ic->i_cache_incs, &list); 165 free_percpu(ic->i_cache_incs.percpu); 166 167 list_for_each_entry_safe(inc, inc_tmp, &list, ii_cache_entry) { 168 list_del(&inc->ii_cache_entry); 169 WARN_ON(!list_empty(&inc->ii_frags)); 170 kmem_cache_free(rds_ib_incoming_slab, inc); 171 } 172 173 rds_ib_cache_xfer_to_ready(&ic->i_cache_frags); 174 rds_ib_cache_splice_all_lists(&ic->i_cache_frags, &list); 175 free_percpu(ic->i_cache_frags.percpu); 176 177 list_for_each_entry_safe(frag, frag_tmp, &list, f_cache_entry) { 178 list_del(&frag->f_cache_entry); 179 WARN_ON(!list_empty(&frag->f_item)); 180 kmem_cache_free(rds_ib_frag_slab, frag); 181 } 182 } 183 184 /* fwd decl */ 185 static void rds_ib_recv_cache_put(struct list_head *new_item, 186 struct rds_ib_refill_cache *cache); 187 static struct list_head *rds_ib_recv_cache_get(struct rds_ib_refill_cache *cache); 188 189 190 /* Recycle frag and attached recv buffer f_sg */ 191 static void rds_ib_frag_free(struct rds_ib_connection *ic, 192 struct rds_page_frag *frag) 193 { 194 rdsdebug("frag %p page %p\n", frag, sg_page(&frag->f_sg)); 195 196 rds_ib_recv_cache_put(&frag->f_cache_entry, &ic->i_cache_frags); 197 atomic_add(RDS_FRAG_SIZE / SZ_1K, &ic->i_cache_allocs); 198 rds_ib_stats_add(s_ib_recv_added_to_cache, RDS_FRAG_SIZE); 199 } 200 201 /* Recycle inc after freeing attached frags */ 202 void rds_ib_inc_free(struct rds_incoming *inc) 203 { 204 struct rds_ib_incoming *ibinc; 205 struct rds_page_frag *frag; 206 struct rds_page_frag *pos; 207 struct rds_ib_connection *ic = inc->i_conn->c_transport_data; 208 209 ibinc = container_of(inc, struct rds_ib_incoming, ii_inc); 210 211 /* Free attached frags */ 212 list_for_each_entry_safe(frag, pos, &ibinc->ii_frags, f_item) { 213 list_del_init(&frag->f_item); 214 rds_ib_frag_free(ic, frag); 215 } 216 BUG_ON(!list_empty(&ibinc->ii_frags)); 217 218 rdsdebug("freeing ibinc %p inc %p\n", ibinc, inc); 219 rds_ib_recv_cache_put(&ibinc->ii_cache_entry, &ic->i_cache_incs); 220 } 221 222 static void rds_ib_recv_clear_one(struct rds_ib_connection *ic, 223 struct rds_ib_recv_work *recv) 224 { 225 if (recv->r_ibinc) { 226 rds_inc_put(&recv->r_ibinc->ii_inc); 227 recv->r_ibinc = NULL; 228 } 229 if (recv->r_frag) { 230 ib_dma_unmap_sg(ic->i_cm_id->device, &recv->r_frag->f_sg, 1, DMA_FROM_DEVICE); 231 rds_ib_frag_free(ic, recv->r_frag); 232 recv->r_frag = NULL; 233 } 234 } 235 236 void rds_ib_recv_clear_ring(struct rds_ib_connection *ic) 237 { 238 u32 i; 239 240 for (i = 0; i < ic->i_recv_ring.w_nr; i++) 241 rds_ib_recv_clear_one(ic, &ic->i_recvs[i]); 242 } 243 244 static struct rds_ib_incoming *rds_ib_refill_one_inc(struct rds_ib_connection *ic, 245 gfp_t slab_mask) 246 { 247 struct rds_ib_incoming *ibinc; 248 struct list_head *cache_item; 249 int avail_allocs; 250 251 cache_item = rds_ib_recv_cache_get(&ic->i_cache_incs); 252 if (cache_item) { 253 ibinc = container_of(cache_item, struct rds_ib_incoming, ii_cache_entry); 254 } else { 255 avail_allocs = atomic_add_unless(&rds_ib_allocation, 256 1, rds_ib_sysctl_max_recv_allocation); 257 if (!avail_allocs) { 258 rds_ib_stats_inc(s_ib_rx_alloc_limit); 259 return NULL; 260 } 261 ibinc = kmem_cache_alloc(rds_ib_incoming_slab, slab_mask); 262 if (!ibinc) { 263 atomic_dec(&rds_ib_allocation); 264 return NULL; 265 } 266 rds_ib_stats_inc(s_ib_rx_total_incs); 267 } 268 INIT_LIST_HEAD(&ibinc->ii_frags); 269 rds_inc_init(&ibinc->ii_inc, ic->conn, &ic->conn->c_faddr); 270 271 return ibinc; 272 } 273 274 static struct rds_page_frag *rds_ib_refill_one_frag(struct rds_ib_connection *ic, 275 gfp_t slab_mask, gfp_t page_mask) 276 { 277 struct rds_page_frag *frag; 278 struct list_head *cache_item; 279 int ret; 280 281 cache_item = rds_ib_recv_cache_get(&ic->i_cache_frags); 282 if (cache_item) { 283 frag = container_of(cache_item, struct rds_page_frag, f_cache_entry); 284 atomic_sub(RDS_FRAG_SIZE / SZ_1K, &ic->i_cache_allocs); 285 rds_ib_stats_add(s_ib_recv_added_to_cache, RDS_FRAG_SIZE); 286 } else { 287 frag = kmem_cache_alloc(rds_ib_frag_slab, slab_mask); 288 if (!frag) 289 return NULL; 290 291 sg_init_table(&frag->f_sg, 1); 292 ret = rds_page_remainder_alloc(&frag->f_sg, 293 RDS_FRAG_SIZE, page_mask); 294 if (ret) { 295 kmem_cache_free(rds_ib_frag_slab, frag); 296 return NULL; 297 } 298 rds_ib_stats_inc(s_ib_rx_total_frags); 299 } 300 301 INIT_LIST_HEAD(&frag->f_item); 302 303 return frag; 304 } 305 306 static int rds_ib_recv_refill_one(struct rds_connection *conn, 307 struct rds_ib_recv_work *recv, gfp_t gfp) 308 { 309 struct rds_ib_connection *ic = conn->c_transport_data; 310 struct ib_sge *sge; 311 int ret = -ENOMEM; 312 gfp_t slab_mask = GFP_NOWAIT; 313 gfp_t page_mask = GFP_NOWAIT; 314 315 if (gfp & __GFP_DIRECT_RECLAIM) { 316 slab_mask = GFP_KERNEL; 317 page_mask = GFP_HIGHUSER; 318 } 319 320 if (!ic->i_cache_incs.ready) 321 rds_ib_cache_xfer_to_ready(&ic->i_cache_incs); 322 if (!ic->i_cache_frags.ready) 323 rds_ib_cache_xfer_to_ready(&ic->i_cache_frags); 324 325 /* 326 * ibinc was taken from recv if recv contained the start of a message. 327 * recvs that were continuations will still have this allocated. 328 */ 329 if (!recv->r_ibinc) { 330 recv->r_ibinc = rds_ib_refill_one_inc(ic, slab_mask); 331 if (!recv->r_ibinc) 332 goto out; 333 } 334 335 WARN_ON(recv->r_frag); /* leak! */ 336 recv->r_frag = rds_ib_refill_one_frag(ic, slab_mask, page_mask); 337 if (!recv->r_frag) 338 goto out; 339 340 ret = ib_dma_map_sg(ic->i_cm_id->device, &recv->r_frag->f_sg, 341 1, DMA_FROM_DEVICE); 342 WARN_ON(ret != 1); 343 344 sge = &recv->r_sge[0]; 345 sge->addr = ic->i_recv_hdrs_dma + (recv - ic->i_recvs) * sizeof(struct rds_header); 346 sge->length = sizeof(struct rds_header); 347 348 sge = &recv->r_sge[1]; 349 sge->addr = sg_dma_address(&recv->r_frag->f_sg); 350 sge->length = sg_dma_len(&recv->r_frag->f_sg); 351 352 ret = 0; 353 out: 354 return ret; 355 } 356 357 static int acquire_refill(struct rds_connection *conn) 358 { 359 return test_and_set_bit(RDS_RECV_REFILL, &conn->c_flags) == 0; 360 } 361 362 static void release_refill(struct rds_connection *conn) 363 { 364 clear_bit(RDS_RECV_REFILL, &conn->c_flags); 365 366 /* We don't use wait_on_bit()/wake_up_bit() because our waking is in a 367 * hot path and finding waiters is very rare. We don't want to walk 368 * the system-wide hashed waitqueue buckets in the fast path only to 369 * almost never find waiters. 370 */ 371 if (waitqueue_active(&conn->c_waitq)) 372 wake_up_all(&conn->c_waitq); 373 } 374 375 /* 376 * This tries to allocate and post unused work requests after making sure that 377 * they have all the allocations they need to queue received fragments into 378 * sockets. 379 */ 380 void rds_ib_recv_refill(struct rds_connection *conn, int prefill, gfp_t gfp) 381 { 382 struct rds_ib_connection *ic = conn->c_transport_data; 383 struct rds_ib_recv_work *recv; 384 unsigned int posted = 0; 385 int ret = 0; 386 bool can_wait = !!(gfp & __GFP_DIRECT_RECLAIM); 387 u32 pos; 388 389 /* the goal here is to just make sure that someone, somewhere 390 * is posting buffers. If we can't get the refill lock, 391 * let them do their thing 392 */ 393 if (!acquire_refill(conn)) 394 return; 395 396 while ((prefill || rds_conn_up(conn)) && 397 rds_ib_ring_alloc(&ic->i_recv_ring, 1, &pos)) { 398 if (pos >= ic->i_recv_ring.w_nr) { 399 printk(KERN_NOTICE "Argh - ring alloc returned pos=%u\n", 400 pos); 401 break; 402 } 403 404 recv = &ic->i_recvs[pos]; 405 ret = rds_ib_recv_refill_one(conn, recv, gfp); 406 if (ret) { 407 break; 408 } 409 410 rdsdebug("recv %p ibinc %p page %p addr %lu\n", recv, 411 recv->r_ibinc, sg_page(&recv->r_frag->f_sg), 412 (long)sg_dma_address(&recv->r_frag->f_sg)); 413 414 /* XXX when can this fail? */ 415 ret = ib_post_recv(ic->i_cm_id->qp, &recv->r_wr, NULL); 416 if (ret) { 417 rds_ib_conn_error(conn, "recv post on " 418 "%pI6c returned %d, disconnecting and " 419 "reconnecting\n", &conn->c_faddr, 420 ret); 421 break; 422 } 423 424 posted++; 425 } 426 427 /* We're doing flow control - update the window. */ 428 if (ic->i_flowctl && posted) 429 rds_ib_advertise_credits(conn, posted); 430 431 if (ret) 432 rds_ib_ring_unalloc(&ic->i_recv_ring, 1); 433 434 release_refill(conn); 435 436 /* if we're called from the softirq handler, we'll be GFP_NOWAIT. 437 * in this case the ring being low is going to lead to more interrupts 438 * and we can safely let the softirq code take care of it unless the 439 * ring is completely empty. 440 * 441 * if we're called from krdsd, we'll be GFP_KERNEL. In this case 442 * we might have raced with the softirq code while we had the refill 443 * lock held. Use rds_ib_ring_low() instead of ring_empty to decide 444 * if we should requeue. 445 */ 446 if (rds_conn_up(conn) && 447 ((can_wait && rds_ib_ring_low(&ic->i_recv_ring)) || 448 rds_ib_ring_empty(&ic->i_recv_ring))) { 449 queue_delayed_work(rds_wq, &conn->c_recv_w, 1); 450 } 451 } 452 453 /* 454 * We want to recycle several types of recv allocations, like incs and frags. 455 * To use this, the *_free() function passes in the ptr to a list_head within 456 * the recyclee, as well as the cache to put it on. 457 * 458 * First, we put the memory on a percpu list. When this reaches a certain size, 459 * We move it to an intermediate non-percpu list in a lockless manner, with some 460 * xchg/compxchg wizardry. 461 * 462 * N.B. Instead of a list_head as the anchor, we use a single pointer, which can 463 * be NULL and xchg'd. The list is actually empty when the pointer is NULL, and 464 * list_empty() will return true with one element is actually present. 465 */ 466 static void rds_ib_recv_cache_put(struct list_head *new_item, 467 struct rds_ib_refill_cache *cache) 468 { 469 unsigned long flags; 470 struct list_head *old, *chpfirst; 471 472 local_irq_save(flags); 473 474 chpfirst = __this_cpu_read(cache->percpu->first); 475 if (!chpfirst) 476 INIT_LIST_HEAD(new_item); 477 else /* put on front */ 478 list_add_tail(new_item, chpfirst); 479 480 __this_cpu_write(cache->percpu->first, new_item); 481 __this_cpu_inc(cache->percpu->count); 482 483 if (__this_cpu_read(cache->percpu->count) < RDS_IB_RECYCLE_BATCH_COUNT) 484 goto end; 485 486 /* 487 * Return our per-cpu first list to the cache's xfer by atomically 488 * grabbing the current xfer list, appending it to our per-cpu list, 489 * and then atomically returning that entire list back to the 490 * cache's xfer list as long as it's still empty. 491 */ 492 do { 493 old = xchg(&cache->xfer, NULL); 494 if (old) 495 list_splice_entire_tail(old, chpfirst); 496 old = cmpxchg(&cache->xfer, NULL, chpfirst); 497 } while (old); 498 499 500 __this_cpu_write(cache->percpu->first, NULL); 501 __this_cpu_write(cache->percpu->count, 0); 502 end: 503 local_irq_restore(flags); 504 } 505 506 static struct list_head *rds_ib_recv_cache_get(struct rds_ib_refill_cache *cache) 507 { 508 struct list_head *head = cache->ready; 509 510 if (head) { 511 if (!list_empty(head)) { 512 cache->ready = head->next; 513 list_del_init(head); 514 } else 515 cache->ready = NULL; 516 } 517 518 return head; 519 } 520 521 int rds_ib_inc_copy_to_user(struct rds_incoming *inc, struct iov_iter *to) 522 { 523 struct rds_ib_incoming *ibinc; 524 struct rds_page_frag *frag; 525 unsigned long to_copy; 526 unsigned long frag_off = 0; 527 int copied = 0; 528 int ret; 529 u32 len; 530 531 ibinc = container_of(inc, struct rds_ib_incoming, ii_inc); 532 frag = list_entry(ibinc->ii_frags.next, struct rds_page_frag, f_item); 533 len = be32_to_cpu(inc->i_hdr.h_len); 534 535 while (iov_iter_count(to) && copied < len) { 536 if (frag_off == RDS_FRAG_SIZE) { 537 frag = list_entry(frag->f_item.next, 538 struct rds_page_frag, f_item); 539 frag_off = 0; 540 } 541 to_copy = min_t(unsigned long, iov_iter_count(to), 542 RDS_FRAG_SIZE - frag_off); 543 to_copy = min_t(unsigned long, to_copy, len - copied); 544 545 /* XXX needs + offset for multiple recvs per page */ 546 rds_stats_add(s_copy_to_user, to_copy); 547 ret = copy_page_to_iter(sg_page(&frag->f_sg), 548 frag->f_sg.offset + frag_off, 549 to_copy, 550 to); 551 if (ret != to_copy) 552 return -EFAULT; 553 554 frag_off += to_copy; 555 copied += to_copy; 556 } 557 558 return copied; 559 } 560 561 /* ic starts out kzalloc()ed */ 562 void rds_ib_recv_init_ack(struct rds_ib_connection *ic) 563 { 564 struct ib_send_wr *wr = &ic->i_ack_wr; 565 struct ib_sge *sge = &ic->i_ack_sge; 566 567 sge->addr = ic->i_ack_dma; 568 sge->length = sizeof(struct rds_header); 569 sge->lkey = ic->i_pd->local_dma_lkey; 570 571 wr->sg_list = sge; 572 wr->num_sge = 1; 573 wr->opcode = IB_WR_SEND; 574 wr->wr_id = RDS_IB_ACK_WR_ID; 575 wr->send_flags = IB_SEND_SIGNALED | IB_SEND_SOLICITED; 576 } 577 578 /* 579 * You'd think that with reliable IB connections you wouldn't need to ack 580 * messages that have been received. The problem is that IB hardware generates 581 * an ack message before it has DMAed the message into memory. This creates a 582 * potential message loss if the HCA is disabled for any reason between when it 583 * sends the ack and before the message is DMAed and processed. This is only a 584 * potential issue if another HCA is available for fail-over. 585 * 586 * When the remote host receives our ack they'll free the sent message from 587 * their send queue. To decrease the latency of this we always send an ack 588 * immediately after we've received messages. 589 * 590 * For simplicity, we only have one ack in flight at a time. This puts 591 * pressure on senders to have deep enough send queues to absorb the latency of 592 * a single ack frame being in flight. This might not be good enough. 593 * 594 * This is implemented by have a long-lived send_wr and sge which point to a 595 * statically allocated ack frame. This ack wr does not fall under the ring 596 * accounting that the tx and rx wrs do. The QP attribute specifically makes 597 * room for it beyond the ring size. Send completion notices its special 598 * wr_id and avoids working with the ring in that case. 599 */ 600 #ifndef KERNEL_HAS_ATOMIC64 601 void rds_ib_set_ack(struct rds_ib_connection *ic, u64 seq, int ack_required) 602 { 603 unsigned long flags; 604 605 spin_lock_irqsave(&ic->i_ack_lock, flags); 606 ic->i_ack_next = seq; 607 if (ack_required) 608 set_bit(IB_ACK_REQUESTED, &ic->i_ack_flags); 609 spin_unlock_irqrestore(&ic->i_ack_lock, flags); 610 } 611 612 static u64 rds_ib_get_ack(struct rds_ib_connection *ic) 613 { 614 unsigned long flags; 615 u64 seq; 616 617 clear_bit(IB_ACK_REQUESTED, &ic->i_ack_flags); 618 619 spin_lock_irqsave(&ic->i_ack_lock, flags); 620 seq = ic->i_ack_next; 621 spin_unlock_irqrestore(&ic->i_ack_lock, flags); 622 623 return seq; 624 } 625 #else 626 void rds_ib_set_ack(struct rds_ib_connection *ic, u64 seq, int ack_required) 627 { 628 atomic64_set(&ic->i_ack_next, seq); 629 if (ack_required) { 630 smp_mb__before_atomic(); 631 set_bit(IB_ACK_REQUESTED, &ic->i_ack_flags); 632 } 633 } 634 635 static u64 rds_ib_get_ack(struct rds_ib_connection *ic) 636 { 637 clear_bit(IB_ACK_REQUESTED, &ic->i_ack_flags); 638 smp_mb__after_atomic(); 639 640 return atomic64_read(&ic->i_ack_next); 641 } 642 #endif 643 644 645 static void rds_ib_send_ack(struct rds_ib_connection *ic, unsigned int adv_credits) 646 { 647 struct rds_header *hdr = ic->i_ack; 648 u64 seq; 649 int ret; 650 651 seq = rds_ib_get_ack(ic); 652 653 rdsdebug("send_ack: ic %p ack %llu\n", ic, (unsigned long long) seq); 654 rds_message_populate_header(hdr, 0, 0, 0); 655 hdr->h_ack = cpu_to_be64(seq); 656 hdr->h_credit = adv_credits; 657 rds_message_make_checksum(hdr); 658 ic->i_ack_queued = jiffies; 659 660 ret = ib_post_send(ic->i_cm_id->qp, &ic->i_ack_wr, NULL); 661 if (unlikely(ret)) { 662 /* Failed to send. Release the WR, and 663 * force another ACK. 664 */ 665 clear_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags); 666 set_bit(IB_ACK_REQUESTED, &ic->i_ack_flags); 667 668 rds_ib_stats_inc(s_ib_ack_send_failure); 669 670 rds_ib_conn_error(ic->conn, "sending ack failed\n"); 671 } else 672 rds_ib_stats_inc(s_ib_ack_sent); 673 } 674 675 /* 676 * There are 3 ways of getting acknowledgements to the peer: 677 * 1. We call rds_ib_attempt_ack from the recv completion handler 678 * to send an ACK-only frame. 679 * However, there can be only one such frame in the send queue 680 * at any time, so we may have to postpone it. 681 * 2. When another (data) packet is transmitted while there's 682 * an ACK in the queue, we piggyback the ACK sequence number 683 * on the data packet. 684 * 3. If the ACK WR is done sending, we get called from the 685 * send queue completion handler, and check whether there's 686 * another ACK pending (postponed because the WR was on the 687 * queue). If so, we transmit it. 688 * 689 * We maintain 2 variables: 690 * - i_ack_flags, which keeps track of whether the ACK WR 691 * is currently in the send queue or not (IB_ACK_IN_FLIGHT) 692 * - i_ack_next, which is the last sequence number we received 693 * 694 * Potentially, send queue and receive queue handlers can run concurrently. 695 * It would be nice to not have to use a spinlock to synchronize things, 696 * but the one problem that rules this out is that 64bit updates are 697 * not atomic on all platforms. Things would be a lot simpler if 698 * we had atomic64 or maybe cmpxchg64 everywhere. 699 * 700 * Reconnecting complicates this picture just slightly. When we 701 * reconnect, we may be seeing duplicate packets. The peer 702 * is retransmitting them, because it hasn't seen an ACK for 703 * them. It is important that we ACK these. 704 * 705 * ACK mitigation adds a header flag "ACK_REQUIRED"; any packet with 706 * this flag set *MUST* be acknowledged immediately. 707 */ 708 709 /* 710 * When we get here, we're called from the recv queue handler. 711 * Check whether we ought to transmit an ACK. 712 */ 713 void rds_ib_attempt_ack(struct rds_ib_connection *ic) 714 { 715 unsigned int adv_credits; 716 717 if (!test_bit(IB_ACK_REQUESTED, &ic->i_ack_flags)) 718 return; 719 720 if (test_and_set_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags)) { 721 rds_ib_stats_inc(s_ib_ack_send_delayed); 722 return; 723 } 724 725 /* Can we get a send credit? */ 726 if (!rds_ib_send_grab_credits(ic, 1, &adv_credits, 0, RDS_MAX_ADV_CREDIT)) { 727 rds_ib_stats_inc(s_ib_tx_throttle); 728 clear_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags); 729 return; 730 } 731 732 clear_bit(IB_ACK_REQUESTED, &ic->i_ack_flags); 733 rds_ib_send_ack(ic, adv_credits); 734 } 735 736 /* 737 * We get here from the send completion handler, when the 738 * adapter tells us the ACK frame was sent. 739 */ 740 void rds_ib_ack_send_complete(struct rds_ib_connection *ic) 741 { 742 clear_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags); 743 rds_ib_attempt_ack(ic); 744 } 745 746 /* 747 * This is called by the regular xmit code when it wants to piggyback 748 * an ACK on an outgoing frame. 749 */ 750 u64 rds_ib_piggyb_ack(struct rds_ib_connection *ic) 751 { 752 if (test_and_clear_bit(IB_ACK_REQUESTED, &ic->i_ack_flags)) 753 rds_ib_stats_inc(s_ib_ack_send_piggybacked); 754 return rds_ib_get_ack(ic); 755 } 756 757 /* 758 * It's kind of lame that we're copying from the posted receive pages into 759 * long-lived bitmaps. We could have posted the bitmaps and rdma written into 760 * them. But receiving new congestion bitmaps should be a *rare* event, so 761 * hopefully we won't need to invest that complexity in making it more 762 * efficient. By copying we can share a simpler core with TCP which has to 763 * copy. 764 */ 765 static void rds_ib_cong_recv(struct rds_connection *conn, 766 struct rds_ib_incoming *ibinc) 767 { 768 struct rds_cong_map *map; 769 unsigned int map_off; 770 unsigned int map_page; 771 struct rds_page_frag *frag; 772 unsigned long frag_off; 773 unsigned long to_copy; 774 unsigned long copied; 775 __le64 uncongested = 0; 776 void *addr; 777 778 /* catch completely corrupt packets */ 779 if (be32_to_cpu(ibinc->ii_inc.i_hdr.h_len) != RDS_CONG_MAP_BYTES) 780 return; 781 782 map = conn->c_fcong; 783 map_page = 0; 784 map_off = 0; 785 786 frag = list_entry(ibinc->ii_frags.next, struct rds_page_frag, f_item); 787 frag_off = 0; 788 789 copied = 0; 790 791 while (copied < RDS_CONG_MAP_BYTES) { 792 __le64 *src, *dst; 793 unsigned int k; 794 795 to_copy = min(RDS_FRAG_SIZE - frag_off, PAGE_SIZE - map_off); 796 BUG_ON(to_copy & 7); /* Must be 64bit aligned. */ 797 798 addr = kmap_atomic(sg_page(&frag->f_sg)); 799 800 src = addr + frag->f_sg.offset + frag_off; 801 dst = (void *)map->m_page_addrs[map_page] + map_off; 802 for (k = 0; k < to_copy; k += 8) { 803 /* Record ports that became uncongested, ie 804 * bits that changed from 0 to 1. */ 805 uncongested |= ~(*src) & *dst; 806 *dst++ = *src++; 807 } 808 kunmap_atomic(addr); 809 810 copied += to_copy; 811 812 map_off += to_copy; 813 if (map_off == PAGE_SIZE) { 814 map_off = 0; 815 map_page++; 816 } 817 818 frag_off += to_copy; 819 if (frag_off == RDS_FRAG_SIZE) { 820 frag = list_entry(frag->f_item.next, 821 struct rds_page_frag, f_item); 822 frag_off = 0; 823 } 824 } 825 826 /* the congestion map is in little endian order */ 827 rds_cong_map_updated(map, le64_to_cpu(uncongested)); 828 } 829 830 static void rds_ib_process_recv(struct rds_connection *conn, 831 struct rds_ib_recv_work *recv, u32 data_len, 832 struct rds_ib_ack_state *state) 833 { 834 struct rds_ib_connection *ic = conn->c_transport_data; 835 struct rds_ib_incoming *ibinc = ic->i_ibinc; 836 struct rds_header *ihdr, *hdr; 837 838 /* XXX shut down the connection if port 0,0 are seen? */ 839 840 rdsdebug("ic %p ibinc %p recv %p byte len %u\n", ic, ibinc, recv, 841 data_len); 842 843 if (data_len < sizeof(struct rds_header)) { 844 rds_ib_conn_error(conn, "incoming message " 845 "from %pI6c didn't include a " 846 "header, disconnecting and " 847 "reconnecting\n", 848 &conn->c_faddr); 849 return; 850 } 851 data_len -= sizeof(struct rds_header); 852 853 ihdr = &ic->i_recv_hdrs[recv - ic->i_recvs]; 854 855 /* Validate the checksum. */ 856 if (!rds_message_verify_checksum(ihdr)) { 857 rds_ib_conn_error(conn, "incoming message " 858 "from %pI6c has corrupted header - " 859 "forcing a reconnect\n", 860 &conn->c_faddr); 861 rds_stats_inc(s_recv_drop_bad_checksum); 862 return; 863 } 864 865 /* Process the ACK sequence which comes with every packet */ 866 state->ack_recv = be64_to_cpu(ihdr->h_ack); 867 state->ack_recv_valid = 1; 868 869 /* Process the credits update if there was one */ 870 if (ihdr->h_credit) 871 rds_ib_send_add_credits(conn, ihdr->h_credit); 872 873 if (ihdr->h_sport == 0 && ihdr->h_dport == 0 && data_len == 0) { 874 /* This is an ACK-only packet. The fact that it gets 875 * special treatment here is that historically, ACKs 876 * were rather special beasts. 877 */ 878 rds_ib_stats_inc(s_ib_ack_received); 879 880 /* 881 * Usually the frags make their way on to incs and are then freed as 882 * the inc is freed. We don't go that route, so we have to drop the 883 * page ref ourselves. We can't just leave the page on the recv 884 * because that confuses the dma mapping of pages and each recv's use 885 * of a partial page. 886 * 887 * FIXME: Fold this into the code path below. 888 */ 889 rds_ib_frag_free(ic, recv->r_frag); 890 recv->r_frag = NULL; 891 return; 892 } 893 894 /* 895 * If we don't already have an inc on the connection then this 896 * fragment has a header and starts a message.. copy its header 897 * into the inc and save the inc so we can hang upcoming fragments 898 * off its list. 899 */ 900 if (!ibinc) { 901 ibinc = recv->r_ibinc; 902 recv->r_ibinc = NULL; 903 ic->i_ibinc = ibinc; 904 905 hdr = &ibinc->ii_inc.i_hdr; 906 ibinc->ii_inc.i_rx_lat_trace[RDS_MSG_RX_HDR] = 907 local_clock(); 908 memcpy(hdr, ihdr, sizeof(*hdr)); 909 ic->i_recv_data_rem = be32_to_cpu(hdr->h_len); 910 ibinc->ii_inc.i_rx_lat_trace[RDS_MSG_RX_START] = 911 local_clock(); 912 913 rdsdebug("ic %p ibinc %p rem %u flag 0x%x\n", ic, ibinc, 914 ic->i_recv_data_rem, hdr->h_flags); 915 } else { 916 hdr = &ibinc->ii_inc.i_hdr; 917 /* We can't just use memcmp here; fragments of a 918 * single message may carry different ACKs */ 919 if (hdr->h_sequence != ihdr->h_sequence || 920 hdr->h_len != ihdr->h_len || 921 hdr->h_sport != ihdr->h_sport || 922 hdr->h_dport != ihdr->h_dport) { 923 rds_ib_conn_error(conn, 924 "fragment header mismatch; forcing reconnect\n"); 925 return; 926 } 927 } 928 929 list_add_tail(&recv->r_frag->f_item, &ibinc->ii_frags); 930 recv->r_frag = NULL; 931 932 if (ic->i_recv_data_rem > RDS_FRAG_SIZE) 933 ic->i_recv_data_rem -= RDS_FRAG_SIZE; 934 else { 935 ic->i_recv_data_rem = 0; 936 ic->i_ibinc = NULL; 937 938 if (ibinc->ii_inc.i_hdr.h_flags == RDS_FLAG_CONG_BITMAP) { 939 rds_ib_cong_recv(conn, ibinc); 940 } else { 941 rds_recv_incoming(conn, &conn->c_faddr, &conn->c_laddr, 942 &ibinc->ii_inc, GFP_ATOMIC); 943 state->ack_next = be64_to_cpu(hdr->h_sequence); 944 state->ack_next_valid = 1; 945 } 946 947 /* Evaluate the ACK_REQUIRED flag *after* we received 948 * the complete frame, and after bumping the next_rx 949 * sequence. */ 950 if (hdr->h_flags & RDS_FLAG_ACK_REQUIRED) { 951 rds_stats_inc(s_recv_ack_required); 952 state->ack_required = 1; 953 } 954 955 rds_inc_put(&ibinc->ii_inc); 956 } 957 } 958 959 void rds_ib_recv_cqe_handler(struct rds_ib_connection *ic, 960 struct ib_wc *wc, 961 struct rds_ib_ack_state *state) 962 { 963 struct rds_connection *conn = ic->conn; 964 struct rds_ib_recv_work *recv; 965 966 rdsdebug("wc wr_id 0x%llx status %u (%s) byte_len %u imm_data %u\n", 967 (unsigned long long)wc->wr_id, wc->status, 968 ib_wc_status_msg(wc->status), wc->byte_len, 969 be32_to_cpu(wc->ex.imm_data)); 970 971 rds_ib_stats_inc(s_ib_rx_cq_event); 972 recv = &ic->i_recvs[rds_ib_ring_oldest(&ic->i_recv_ring)]; 973 ib_dma_unmap_sg(ic->i_cm_id->device, &recv->r_frag->f_sg, 1, 974 DMA_FROM_DEVICE); 975 976 /* Also process recvs in connecting state because it is possible 977 * to get a recv completion _before_ the rdmacm ESTABLISHED 978 * event is processed. 979 */ 980 if (wc->status == IB_WC_SUCCESS) { 981 rds_ib_process_recv(conn, recv, wc->byte_len, state); 982 } else { 983 /* We expect errors as the qp is drained during shutdown */ 984 if (rds_conn_up(conn) || rds_conn_connecting(conn)) 985 rds_ib_conn_error(conn, "recv completion on <%pI6c,%pI6c, %d> had status %u (%s), disconnecting and reconnecting\n", 986 &conn->c_laddr, &conn->c_faddr, 987 conn->c_tos, wc->status, 988 ib_wc_status_msg(wc->status)); 989 } 990 991 /* rds_ib_process_recv() doesn't always consume the frag, and 992 * we might not have called it at all if the wc didn't indicate 993 * success. We already unmapped the frag's pages, though, and 994 * the following rds_ib_ring_free() call tells the refill path 995 * that it will not find an allocated frag here. Make sure we 996 * keep that promise by freeing a frag that's still on the ring. 997 */ 998 if (recv->r_frag) { 999 rds_ib_frag_free(ic, recv->r_frag); 1000 recv->r_frag = NULL; 1001 } 1002 rds_ib_ring_free(&ic->i_recv_ring, 1); 1003 1004 /* If we ever end up with a really empty receive ring, we're 1005 * in deep trouble, as the sender will definitely see RNR 1006 * timeouts. */ 1007 if (rds_ib_ring_empty(&ic->i_recv_ring)) 1008 rds_ib_stats_inc(s_ib_rx_ring_empty); 1009 1010 if (rds_ib_ring_low(&ic->i_recv_ring)) { 1011 rds_ib_recv_refill(conn, 0, GFP_NOWAIT); 1012 rds_ib_stats_inc(s_ib_rx_refill_from_cq); 1013 } 1014 } 1015 1016 int rds_ib_recv_path(struct rds_conn_path *cp) 1017 { 1018 struct rds_connection *conn = cp->cp_conn; 1019 struct rds_ib_connection *ic = conn->c_transport_data; 1020 1021 rdsdebug("conn %p\n", conn); 1022 if (rds_conn_up(conn)) { 1023 rds_ib_attempt_ack(ic); 1024 rds_ib_recv_refill(conn, 0, GFP_KERNEL); 1025 rds_ib_stats_inc(s_ib_rx_refill_from_thread); 1026 } 1027 1028 return 0; 1029 } 1030 1031 int rds_ib_recv_init(void) 1032 { 1033 struct sysinfo si; 1034 int ret = -ENOMEM; 1035 1036 /* Default to 30% of all available RAM for recv memory */ 1037 si_meminfo(&si); 1038 rds_ib_sysctl_max_recv_allocation = si.totalram / 3 * PAGE_SIZE / RDS_FRAG_SIZE; 1039 1040 rds_ib_incoming_slab = kmem_cache_create("rds_ib_incoming", 1041 sizeof(struct rds_ib_incoming), 1042 0, SLAB_HWCACHE_ALIGN, NULL); 1043 if (!rds_ib_incoming_slab) 1044 goto out; 1045 1046 rds_ib_frag_slab = kmem_cache_create("rds_ib_frag", 1047 sizeof(struct rds_page_frag), 1048 0, SLAB_HWCACHE_ALIGN, NULL); 1049 if (!rds_ib_frag_slab) { 1050 kmem_cache_destroy(rds_ib_incoming_slab); 1051 rds_ib_incoming_slab = NULL; 1052 } else 1053 ret = 0; 1054 out: 1055 return ret; 1056 } 1057 1058 void rds_ib_recv_exit(void) 1059 { 1060 kmem_cache_destroy(rds_ib_incoming_slab); 1061 kmem_cache_destroy(rds_ib_frag_slab); 1062 } 1063