net/rds/ib_recv.c

   1 /*
   2  * Copyright (c) 2006, 2019 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/sched/clock.h>
  35 #include <linux/slab.h>
  36 #include <linux/pci.h>
  37 #include <linux/dma-mapping.h>
  38 #include <rdma/rdma_cm.h>
  39
  40 #include "rds_single_path.h"
  41 #include "rds.h"
  42 #include "ib.h"
  43
  44 static struct kmem_cache *rds_ib_incoming_slab;
  45 static struct kmem_cache *rds_ib_frag_slab;
  46 static atomic_t rds_ib_allocation = ATOMIC_INIT(0);
  47
  48 void rds_ib_recv_init_ring(struct rds_ib_connection *ic)
  49 {
  50         struct rds_ib_recv_work *recv;
  51         u32 i;
  52
  53         for (i = 0, recv = ic->i_recvs; i < ic->i_recv_ring.w_nr; i++, recv++) {
  54                 struct ib_sge *sge;
  55
  56                 recv->r_ibinc = NULL;
  57                 recv->r_frag = NULL;
  58
  59                 recv->r_wr.next = NULL;
  60                 recv->r_wr.wr_id = i;
  61                 recv->r_wr.sg_list = recv->r_sge;
  62                 recv->r_wr.num_sge = RDS_IB_RECV_SGE;
  63
  64                 sge = &recv->r_sge[0];
  65                 sge->addr = ic->i_recv_hdrs_dma[i];
  66                 sge->length = sizeof(struct rds_header);
  67                 sge->lkey = ic->i_pd->local_dma_lkey;
  68
  69                 sge = &recv->r_sge[1];
  70                 sge->addr = 0;
  71                 sge->length = RDS_FRAG_SIZE;
  72                 sge->lkey = ic->i_pd->local_dma_lkey;
  73         }
  74 }
  75
  76 /*
  77  * The entire 'from' list, including the from element itself, is put on
  78  * to the tail of the 'to' list.
  79  */
  80 static void list_splice_entire_tail(struct list_head *from,
  81                                     struct list_head *to)
  82 {
  83         struct list_head *from_last = from->prev;
  84
  85         list_splice_tail(from_last, to);
  86         list_add_tail(from_last, to);
  87 }
  88
  89 static void rds_ib_cache_xfer_to_ready(struct rds_ib_refill_cache *cache)
  90 {
  91         struct list_head *tmp;
  92
  93         tmp = xchg(&cache->xfer, NULL);
  94         if (tmp) {
  95                 if (cache->ready)
  96                         list_splice_entire_tail(tmp, cache->ready);
  97                 else
  98                         cache->ready = tmp;
  99         }
 100 }
 101
 102 static int rds_ib_recv_alloc_cache(struct rds_ib_refill_cache *cache, gfp_t gfp)
 103 {
 104         struct rds_ib_cache_head *head;
 105         int cpu;
 106
 107         cache->percpu = alloc_percpu_gfp(struct rds_ib_cache_head, gfp);
 108         if (!cache->percpu)
 109                return -ENOMEM;
 110
 111         for_each_possible_cpu(cpu) {
 112                 head = per_cpu_ptr(cache->percpu, cpu);
 113                 head->first = NULL;
 114                 head->count = 0;
 115         }
 116         cache->xfer = NULL;
 117         cache->ready = NULL;
 118
 119         return 0;
 120 }
 121
 122 int rds_ib_recv_alloc_caches(struct rds_ib_connection *ic, gfp_t gfp)
 123 {
 124         int ret;
 125
 126         ret = rds_ib_recv_alloc_cache(&ic->i_cache_incs, gfp);
 127         if (!ret) {
 128                 ret = rds_ib_recv_alloc_cache(&ic->i_cache_frags, gfp);
 129                 if (ret)
 130                         free_percpu(ic->i_cache_incs.percpu);
 131         }
 132
 133         return ret;
 134 }
 135
 136 static void rds_ib_cache_splice_all_lists(struct rds_ib_refill_cache *cache,
 137                                           struct list_head *caller_list)
 138 {
 139         struct rds_ib_cache_head *head;
 140         int cpu;
 141
 142         for_each_possible_cpu(cpu) {
 143                 head = per_cpu_ptr(cache->percpu, cpu);
 144                 if (head->first) {
 145                         list_splice_entire_tail(head->first, caller_list);
 146                         head->first = NULL;
 147                 }
 148         }
 149
 150         if (cache->ready) {
 151                 list_splice_entire_tail(cache->ready, caller_list);
 152                 cache->ready = NULL;
 153         }
 154 }
 155
 156 void rds_ib_recv_free_caches(struct rds_ib_connection *ic)
 157 {
 158         struct rds_ib_incoming *inc;
 159         struct rds_ib_incoming *inc_tmp;
 160         struct rds_page_frag *frag;
 161         struct rds_page_frag *frag_tmp;
 162         LIST_HEAD(list);
 163
 164         rds_ib_cache_xfer_to_ready(&ic->i_cache_incs);
 165         rds_ib_cache_splice_all_lists(&ic->i_cache_incs, &list);
 166         free_percpu(ic->i_cache_incs.percpu);
 167
 168         list_for_each_entry_safe(inc, inc_tmp, &list, ii_cache_entry) {
 169                 list_del(&inc->ii_cache_entry);
 170                 WARN_ON(!list_empty(&inc->ii_frags));
 171                 kmem_cache_free(rds_ib_incoming_slab, inc);
 172                 atomic_dec(&rds_ib_allocation);
 173         }
 174
 175         rds_ib_cache_xfer_to_ready(&ic->i_cache_frags);
 176         rds_ib_cache_splice_all_lists(&ic->i_cache_frags, &list);
 177         free_percpu(ic->i_cache_frags.percpu);
 178
 179         list_for_each_entry_safe(frag, frag_tmp, &list, f_cache_entry) {
 180                 list_del(&frag->f_cache_entry);
 181                 WARN_ON(!list_empty(&frag->f_item));
 182                 kmem_cache_free(rds_ib_frag_slab, frag);
 183         }
 184 }
 185
 186 /* fwd decl */
 187 static void rds_ib_recv_cache_put(struct list_head *new_item,
 188                                   struct rds_ib_refill_cache *cache);
 189 static struct list_head *rds_ib_recv_cache_get(struct rds_ib_refill_cache *cache);
 190
 191
 192 /* Recycle frag and attached recv buffer f_sg */
 193 static void rds_ib_frag_free(struct rds_ib_connection *ic,
 194                              struct rds_page_frag *frag)
 195 {
 196         rdsdebug("frag %p page %p\n", frag, sg_page(&frag->f_sg));
 197
 198         rds_ib_recv_cache_put(&frag->f_cache_entry, &ic->i_cache_frags);
 199         atomic_add(RDS_FRAG_SIZE / SZ_1K, &ic->i_cache_allocs);
 200         rds_ib_stats_add(s_ib_recv_added_to_cache, RDS_FRAG_SIZE);
 201 }
 202
 203 /* Recycle inc after freeing attached frags */
 204 void rds_ib_inc_free(struct rds_incoming *inc)
 205 {
 206         struct rds_ib_incoming *ibinc;
 207         struct rds_page_frag *frag;
 208         struct rds_page_frag *pos;
 209         struct rds_ib_connection *ic = inc->i_conn->c_transport_data;
 210
 211         ibinc = container_of(inc, struct rds_ib_incoming, ii_inc);
 212
 213         /* Free attached frags */
 214         list_for_each_entry_safe(frag, pos, &ibinc->ii_frags, f_item) {
 215                 list_del_init(&frag->f_item);
 216                 rds_ib_frag_free(ic, frag);
 217         }
 218         BUG_ON(!list_empty(&ibinc->ii_frags));
 219
 220         rdsdebug("freeing ibinc %p inc %p\n", ibinc, inc);
 221         rds_ib_recv_cache_put(&ibinc->ii_cache_entry, &ic->i_cache_incs);
 222 }
 223
 224 static void rds_ib_recv_clear_one(struct rds_ib_connection *ic,
 225                                   struct rds_ib_recv_work *recv)
 226 {
 227         if (recv->r_ibinc) {
 228                 rds_inc_put(&recv->r_ibinc->ii_inc);
 229                 recv->r_ibinc = NULL;
 230         }
 231         if (recv->r_frag) {
 232                 ib_dma_unmap_sg(ic->i_cm_id->device, &recv->r_frag->f_sg, 1, DMA_FROM_DEVICE);
 233                 rds_ib_frag_free(ic, recv->r_frag);
 234                 recv->r_frag = NULL;
 235         }
 236 }
 237
 238 void rds_ib_recv_clear_ring(struct rds_ib_connection *ic)
 239 {
 240         u32 i;
 241
 242         for (i = 0; i < ic->i_recv_ring.w_nr; i++)
 243                 rds_ib_recv_clear_one(ic, &ic->i_recvs[i]);
 244 }
 245
 246 static struct rds_ib_incoming *rds_ib_refill_one_inc(struct rds_ib_connection *ic,
 247                                                      gfp_t slab_mask)
 248 {
 249         struct rds_ib_incoming *ibinc;
 250         struct list_head *cache_item;
 251         int avail_allocs;
 252
 253         cache_item = rds_ib_recv_cache_get(&ic->i_cache_incs);
 254         if (cache_item) {
 255                 ibinc = container_of(cache_item, struct rds_ib_incoming, ii_cache_entry);
 256         } else {
 257                 avail_allocs = atomic_add_unless(&rds_ib_allocation,
 258                                                  1, rds_ib_sysctl_max_recv_allocation);
 259                 if (!avail_allocs) {
 260                         rds_ib_stats_inc(s_ib_rx_alloc_limit);
 261                         return NULL;
 262                 }
 263                 ibinc = kmem_cache_alloc(rds_ib_incoming_slab, slab_mask);
 264                 if (!ibinc) {
 265                         atomic_dec(&rds_ib_allocation);
 266                         return NULL;
 267                 }
 268                 rds_ib_stats_inc(s_ib_rx_total_incs);
 269         }
 270         INIT_LIST_HEAD(&ibinc->ii_frags);
 271         rds_inc_init(&ibinc->ii_inc, ic->conn, &ic->conn->c_faddr);
 272
 273         return ibinc;
 274 }
 275
 276 static struct rds_page_frag *rds_ib_refill_one_frag(struct rds_ib_connection *ic,
 277                                                     gfp_t slab_mask, gfp_t page_mask)
 278 {
 279         struct rds_page_frag *frag;
 280         struct list_head *cache_item;
 281         int ret;
 282
 283         cache_item = rds_ib_recv_cache_get(&ic->i_cache_frags);
 284         if (cache_item) {
 285                 frag = container_of(cache_item, struct rds_page_frag, f_cache_entry);
 286                 atomic_sub(RDS_FRAG_SIZE / SZ_1K, &ic->i_cache_allocs);
 287                 rds_ib_stats_add(s_ib_recv_added_to_cache, RDS_FRAG_SIZE);
 288         } else {
 289                 frag = kmem_cache_alloc(rds_ib_frag_slab, slab_mask);
 290                 if (!frag)
 291                         return NULL;
 292
 293                 sg_init_table(&frag->f_sg, 1);
 294                 ret = rds_page_remainder_alloc(&frag->f_sg,
 295                                                RDS_FRAG_SIZE, page_mask);
 296                 if (ret) {
 297                         kmem_cache_free(rds_ib_frag_slab, frag);
 298                         return NULL;
 299                 }
 300                 rds_ib_stats_inc(s_ib_rx_total_frags);
 301         }
 302
 303         INIT_LIST_HEAD(&frag->f_item);
 304
 305         return frag;
 306 }
 307
 308 static int rds_ib_recv_refill_one(struct rds_connection *conn,
 309                                   struct rds_ib_recv_work *recv, gfp_t gfp)
 310 {
 311         struct rds_ib_connection *ic = conn->c_transport_data;
 312         struct ib_sge *sge;
 313         int ret = -ENOMEM;
 314         gfp_t slab_mask = gfp;
 315         gfp_t page_mask = gfp;
 316
 317         if (gfp & __GFP_DIRECT_RECLAIM) {
 318                 slab_mask = GFP_KERNEL;
 319                 page_mask = GFP_HIGHUSER;
 320         }
 321
 322         if (!ic->i_cache_incs.ready)
 323                 rds_ib_cache_xfer_to_ready(&ic->i_cache_incs);
 324         if (!ic->i_cache_frags.ready)
 325                 rds_ib_cache_xfer_to_ready(&ic->i_cache_frags);
 326
 327         /*
 328          * ibinc was taken from recv if recv contained the start of a message.
 329          * recvs that were continuations will still have this allocated.
 330          */
 331         if (!recv->r_ibinc) {
 332                 recv->r_ibinc = rds_ib_refill_one_inc(ic, slab_mask);
 333                 if (!recv->r_ibinc)
 334                         goto out;
 335         }
 336
 337         WARN_ON(recv->r_frag); /* leak! */
 338         recv->r_frag = rds_ib_refill_one_frag(ic, slab_mask, page_mask);
 339         if (!recv->r_frag)
 340                 goto out;
 341
 342         ret = ib_dma_map_sg(ic->i_cm_id->device, &recv->r_frag->f_sg,
 343                             1, DMA_FROM_DEVICE);
 344         WARN_ON(ret != 1);
 345
 346         sge = &recv->r_sge[0];
 347         sge->addr = ic->i_recv_hdrs_dma[recv - ic->i_recvs];
 348         sge->length = sizeof(struct rds_header);
 349
 350         sge = &recv->r_sge[1];
 351         sge->addr = sg_dma_address(&recv->r_frag->f_sg);
 352         sge->length = sg_dma_len(&recv->r_frag->f_sg);
 353
 354         ret = 0;
 355 out:
 356         return ret;
 357 }
 358
 359 static int acquire_refill(struct rds_connection *conn)
 360 {
 361         return test_and_set_bit(RDS_RECV_REFILL, &conn->c_flags) == 0;
 362 }
 363
 364 static void release_refill(struct rds_connection *conn)
 365 {
 366         clear_bit(RDS_RECV_REFILL, &conn->c_flags);
 367         smp_mb__after_atomic();
 368
 369         /* We don't use wait_on_bit()/wake_up_bit() because our waking is in a
 370          * hot path and finding waiters is very rare.  We don't want to walk
 371          * the system-wide hashed waitqueue buckets in the fast path only to
 372          * almost never find waiters.
 373          */
 374         if (waitqueue_active(&conn->c_waitq))
 375                 wake_up_all(&conn->c_waitq);
 376 }
 377
 378 /*
 379  * This tries to allocate and post unused work requests after making sure that
 380  * they have all the allocations they need to queue received fragments into
 381  * sockets.
 382  */
 383 void rds_ib_recv_refill(struct rds_connection *conn, int prefill, gfp_t gfp)
 384 {
 385         struct rds_ib_connection *ic = conn->c_transport_data;
 386         struct rds_ib_recv_work *recv;
 387         unsigned int posted = 0;
 388         int ret = 0;
 389         bool can_wait = !!(gfp & __GFP_DIRECT_RECLAIM);
 390         bool must_wake = false;
 391         u32 pos;
 392
 393         /* the goal here is to just make sure that someone, somewhere
 394          * is posting buffers.  If we can't get the refill lock,
 395          * let them do their thing
 396          */
 397         if (!acquire_refill(conn))
 398                 return;
 399
 400         while ((prefill || rds_conn_up(conn)) &&
 401                rds_ib_ring_alloc(&ic->i_recv_ring, 1, &pos)) {
 402                 if (pos >= ic->i_recv_ring.w_nr) {
 403                         printk(KERN_NOTICE "Argh - ring alloc returned pos=%u\n",
 404                                         pos);
 405                         break;
 406                 }
 407
 408                 recv = &ic->i_recvs[pos];
 409                 ret = rds_ib_recv_refill_one(conn, recv, gfp);
 410                 if (ret) {
 411                         must_wake = true;
 412                         break;
 413                 }
 414
 415                 rdsdebug("recv %p ibinc %p page %p addr %lu\n", recv,
 416                          recv->r_ibinc, sg_page(&recv->r_frag->f_sg),
 417                          (long)sg_dma_address(&recv->r_frag->f_sg));
 418
 419                 /* XXX when can this fail? */
 420                 ret = ib_post_recv(ic->i_cm_id->qp, &recv->r_wr, NULL);
 421                 if (ret) {
 422                         rds_ib_conn_error(conn, "recv post on "
 423                                "%pI6c returned %d, disconnecting and "
 424                                "reconnecting\n", &conn->c_faddr,
 425                                ret);
 426                         break;
 427                 }
 428
 429                 posted++;
 430
 431                 if ((posted > 128 && need_resched()) || posted > 8192) {
 432                         must_wake = true;
 433                         break;
 434                 }
 435         }
 436
 437         /* We're doing flow control - update the window. */
 438         if (ic->i_flowctl && posted)
 439                 rds_ib_advertise_credits(conn, posted);
 440
 441         if (ret)
 442                 rds_ib_ring_unalloc(&ic->i_recv_ring, 1);
 443
 444         release_refill(conn);
 445
 446         /* if we're called from the softirq handler, we'll be GFP_NOWAIT.
 447          * in this case the ring being low is going to lead to more interrupts
 448          * and we can safely let the softirq code take care of it unless the
 449          * ring is completely empty.
 450          *
 451          * if we're called from krdsd, we'll be GFP_KERNEL.  In this case
 452          * we might have raced with the softirq code while we had the refill
 453          * lock held.  Use rds_ib_ring_low() instead of ring_empty to decide
 454          * if we should requeue.
 455          */
 456         if (rds_conn_up(conn) &&
 457             (must_wake ||
 458             (can_wait && rds_ib_ring_low(&ic->i_recv_ring)) ||
 459             rds_ib_ring_empty(&ic->i_recv_ring))) {
 460                 queue_delayed_work(rds_wq, &conn->c_recv_w, 1);
 461         }
 462         if (can_wait)
 463                 cond_resched();
 464 }
 465
 466 /*
 467  * We want to recycle several types of recv allocations, like incs and frags.
 468  * To use this, the *_free() function passes in the ptr to a list_head within
 469  * the recyclee, as well as the cache to put it on.
 470  *
 471  * First, we put the memory on a percpu list. When this reaches a certain size,
 472  * We move it to an intermediate non-percpu list in a lockless manner, with some
 473  * xchg/compxchg wizardry.
 474  *
 475  * N.B. Instead of a list_head as the anchor, we use a single pointer, which can
 476  * be NULL and xchg'd. The list is actually empty when the pointer is NULL, and
 477  * list_empty() will return true with one element is actually present.
 478  */
 479 static void rds_ib_recv_cache_put(struct list_head *new_item,
 480                                  struct rds_ib_refill_cache *cache)
 481 {
 482         unsigned long flags;
 483         struct list_head *old, *chpfirst;
 484
 485         local_irq_save(flags);
 486
 487         chpfirst = __this_cpu_read(cache->percpu->first);
 488         if (!chpfirst)
 489                 INIT_LIST_HEAD(new_item);
 490         else /* put on front */
 491                 list_add_tail(new_item, chpfirst);
 492
 493         __this_cpu_write(cache->percpu->first, new_item);
 494         __this_cpu_inc(cache->percpu->count);
 495
 496         if (__this_cpu_read(cache->percpu->count) < RDS_IB_RECYCLE_BATCH_COUNT)
 497                 goto end;
 498
 499         /*
 500          * Return our per-cpu first list to the cache's xfer by atomically
 501          * grabbing the current xfer list, appending it to our per-cpu list,
 502          * and then atomically returning that entire list back to the
 503          * cache's xfer list as long as it's still empty.
 504          */
 505         do {
 506                 old = xchg(&cache->xfer, NULL);
 507                 if (old)
 508                         list_splice_entire_tail(old, chpfirst);
 509                 old = cmpxchg(&cache->xfer, NULL, chpfirst);
 510         } while (old);
 511
 512
 513         __this_cpu_write(cache->percpu->first, NULL);
 514         __this_cpu_write(cache->percpu->count, 0);
 515 end:
 516         local_irq_restore(flags);
 517 }
 518
 519 static struct list_head *rds_ib_recv_cache_get(struct rds_ib_refill_cache *cache)
 520 {
 521         struct list_head *head = cache->ready;
 522
 523         if (head) {
 524                 if (!list_empty(head)) {
 525                         cache->ready = head->next;
 526                         list_del_init(head);
 527                 } else
 528                         cache->ready = NULL;
 529         }
 530
 531         return head;
 532 }
 533
 534 int rds_ib_inc_copy_to_user(struct rds_incoming *inc, struct iov_iter *to)
 535 {
 536         struct rds_ib_incoming *ibinc;
 537         struct rds_page_frag *frag;
 538         unsigned long to_copy;
 539         unsigned long frag_off = 0;
 540         int copied = 0;
 541         int ret;
 542         u32 len;
 543
 544         ibinc = container_of(inc, struct rds_ib_incoming, ii_inc);
 545         frag = list_entry(ibinc->ii_frags.next, struct rds_page_frag, f_item);
 546         len = be32_to_cpu(inc->i_hdr.h_len);
 547
 548         while (iov_iter_count(to) && copied < len) {
 549                 if (frag_off == RDS_FRAG_SIZE) {
 550                         frag = list_entry(frag->f_item.next,
 551                                           struct rds_page_frag, f_item);
 552                         frag_off = 0;
 553                 }
 554                 to_copy = min_t(unsigned long, iov_iter_count(to),
 555                                 RDS_FRAG_SIZE - frag_off);
 556                 to_copy = min_t(unsigned long, to_copy, len - copied);
 557
 558                 /* XXX needs + offset for multiple recvs per page */
 559                 rds_stats_add(s_copy_to_user, to_copy);
 560                 ret = copy_page_to_iter(sg_page(&frag->f_sg),
 561                                         frag->f_sg.offset + frag_off,
 562                                         to_copy,
 563                                         to);
 564                 if (ret != to_copy)
 565                         return -EFAULT;
 566
 567                 frag_off += to_copy;
 568                 copied += to_copy;
 569         }
 570
 571         return copied;
 572 }
 573
 574 /* ic starts out kzalloc()ed */
 575 void rds_ib_recv_init_ack(struct rds_ib_connection *ic)
 576 {
 577         struct ib_send_wr *wr = &ic->i_ack_wr;
 578         struct ib_sge *sge = &ic->i_ack_sge;
 579
 580         sge->addr = ic->i_ack_dma;
 581         sge->length = sizeof(struct rds_header);
 582         sge->lkey = ic->i_pd->local_dma_lkey;
 583
 584         wr->sg_list = sge;
 585         wr->num_sge = 1;
 586         wr->opcode = IB_WR_SEND;
 587         wr->wr_id = RDS_IB_ACK_WR_ID;
 588         wr->send_flags = IB_SEND_SIGNALED | IB_SEND_SOLICITED;
 589 }
 590
 591 /*
 592  * You'd think that with reliable IB connections you wouldn't need to ack
 593  * messages that have been received.  The problem is that IB hardware generates
 594  * an ack message before it has DMAed the message into memory.  This creates a
 595  * potential message loss if the HCA is disabled for any reason between when it
 596  * sends the ack and before the message is DMAed and processed.  This is only a
 597  * potential issue if another HCA is available for fail-over.
 598  *
 599  * When the remote host receives our ack they'll free the sent message from
 600  * their send queue.  To decrease the latency of this we always send an ack
 601  * immediately after we've received messages.
 602  *
 603  * For simplicity, we only have one ack in flight at a time.  This puts
 604  * pressure on senders to have deep enough send queues to absorb the latency of
 605  * a single ack frame being in flight.  This might not be good enough.
 606  *
 607  * This is implemented by have a long-lived send_wr and sge which point to a
 608  * statically allocated ack frame.  This ack wr does not fall under the ring
 609  * accounting that the tx and rx wrs do.  The QP attribute specifically makes
 610  * room for it beyond the ring size.  Send completion notices its special
 611  * wr_id and avoids working with the ring in that case.
 612  */
 613 #ifndef KERNEL_HAS_ATOMIC64
 614 void rds_ib_set_ack(struct rds_ib_connection *ic, u64 seq, int ack_required)
 615 {
 616         unsigned long flags;
 617
 618         spin_lock_irqsave(&ic->i_ack_lock, flags);
 619         ic->i_ack_next = seq;
 620         if (ack_required)
 621                 set_bit(IB_ACK_REQUESTED, &ic->i_ack_flags);
 622         spin_unlock_irqrestore(&ic->i_ack_lock, flags);
 623 }
 624
 625 static u64 rds_ib_get_ack(struct rds_ib_connection *ic)
 626 {
 627         unsigned long flags;
 628         u64 seq;
 629
 630         clear_bit(IB_ACK_REQUESTED, &ic->i_ack_flags);
 631
 632         spin_lock_irqsave(&ic->i_ack_lock, flags);
 633         seq = ic->i_ack_next;
 634         spin_unlock_irqrestore(&ic->i_ack_lock, flags);
 635
 636         return seq;
 637 }
 638 #else
 639 void rds_ib_set_ack(struct rds_ib_connection *ic, u64 seq, int ack_required)
 640 {
 641         atomic64_set(&ic->i_ack_next, seq);
 642         if (ack_required) {
 643                 smp_mb__before_atomic();
 644                 set_bit(IB_ACK_REQUESTED, &ic->i_ack_flags);
 645         }
 646 }
 647
 648 static u64 rds_ib_get_ack(struct rds_ib_connection *ic)
 649 {
 650         clear_bit(IB_ACK_REQUESTED, &ic->i_ack_flags);
 651         smp_mb__after_atomic();
 652
 653         return atomic64_read(&ic->i_ack_next);
 654 }
 655 #endif
 656
 657
 658 static void rds_ib_send_ack(struct rds_ib_connection *ic, unsigned int adv_credits)
 659 {
 660         struct rds_header *hdr = ic->i_ack;
 661         u64 seq;
 662         int ret;
 663
 664         seq = rds_ib_get_ack(ic);
 665
 666         rdsdebug("send_ack: ic %p ack %llu\n", ic, (unsigned long long) seq);
 667
 668         ib_dma_sync_single_for_cpu(ic->rds_ibdev->dev, ic->i_ack_dma,
 669                                    sizeof(*hdr), DMA_TO_DEVICE);
 670         rds_message_populate_header(hdr, 0, 0, 0);
 671         hdr->h_ack = cpu_to_be64(seq);
 672         hdr->h_credit = adv_credits;
 673         rds_message_make_checksum(hdr);
 674         ib_dma_sync_single_for_device(ic->rds_ibdev->dev, ic->i_ack_dma,
 675                                       sizeof(*hdr), DMA_TO_DEVICE);
 676
 677         ic->i_ack_queued = jiffies;
 678
 679         ret = ib_post_send(ic->i_cm_id->qp, &ic->i_ack_wr, NULL);
 680         if (unlikely(ret)) {
 681                 /* Failed to send. Release the WR, and
 682                  * force another ACK.
 683                  */
 684                 clear_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags);
 685                 set_bit(IB_ACK_REQUESTED, &ic->i_ack_flags);
 686
 687                 rds_ib_stats_inc(s_ib_ack_send_failure);
 688
 689                 rds_ib_conn_error(ic->conn, "sending ack failed\n");
 690         } else
 691                 rds_ib_stats_inc(s_ib_ack_sent);
 692 }
 693
 694 /*
 695  * There are 3 ways of getting acknowledgements to the peer:
 696  *  1.  We call rds_ib_attempt_ack from the recv completion handler
 697  *      to send an ACK-only frame.
 698  *      However, there can be only one such frame in the send queue
 699  *      at any time, so we may have to postpone it.
 700  *  2.  When another (data) packet is transmitted while there's
 701  *      an ACK in the queue, we piggyback the ACK sequence number
 702  *      on the data packet.
 703  *  3.  If the ACK WR is done sending, we get called from the
 704  *      send queue completion handler, and check whether there's
 705  *      another ACK pending (postponed because the WR was on the
 706  *      queue). If so, we transmit it.
 707  *
 708  * We maintain 2 variables:
 709  *  -   i_ack_flags, which keeps track of whether the ACK WR
 710  *      is currently in the send queue or not (IB_ACK_IN_FLIGHT)
 711  *  -   i_ack_next, which is the last sequence number we received
 712  *
 713  * Potentially, send queue and receive queue handlers can run concurrently.
 714  * It would be nice to not have to use a spinlock to synchronize things,
 715  * but the one problem that rules this out is that 64bit updates are
 716  * not atomic on all platforms. Things would be a lot simpler if
 717  * we had atomic64 or maybe cmpxchg64 everywhere.
 718  *
 719  * Reconnecting complicates this picture just slightly. When we
 720  * reconnect, we may be seeing duplicate packets. The peer
 721  * is retransmitting them, because it hasn't seen an ACK for
 722  * them. It is important that we ACK these.
 723  *
 724  * ACK mitigation adds a header flag "ACK_REQUIRED"; any packet with
 725  * this flag set *MUST* be acknowledged immediately.
 726  */
 727
 728 /*
 729  * When we get here, we're called from the recv queue handler.
 730  * Check whether we ought to transmit an ACK.
 731  */
 732 void rds_ib_attempt_ack(struct rds_ib_connection *ic)
 733 {
 734         unsigned int adv_credits;
 735
 736         if (!test_bit(IB_ACK_REQUESTED, &ic->i_ack_flags))
 737                 return;
 738
 739         if (test_and_set_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags)) {
 740                 rds_ib_stats_inc(s_ib_ack_send_delayed);
 741                 return;
 742         }
 743
 744         /* Can we get a send credit? */
 745         if (!rds_ib_send_grab_credits(ic, 1, &adv_credits, 0, RDS_MAX_ADV_CREDIT)) {
 746                 rds_ib_stats_inc(s_ib_tx_throttle);
 747                 clear_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags);
 748                 return;
 749         }
 750
 751         clear_bit(IB_ACK_REQUESTED, &ic->i_ack_flags);
 752         rds_ib_send_ack(ic, adv_credits);
 753 }
 754
 755 /*
 756  * We get here from the send completion handler, when the
 757  * adapter tells us the ACK frame was sent.
 758  */
 759 void rds_ib_ack_send_complete(struct rds_ib_connection *ic)
 760 {
 761         clear_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags);
 762         rds_ib_attempt_ack(ic);
 763 }
 764
 765 /*
 766  * This is called by the regular xmit code when it wants to piggyback
 767  * an ACK on an outgoing frame.
 768  */
 769 u64 rds_ib_piggyb_ack(struct rds_ib_connection *ic)
 770 {
 771         if (test_and_clear_bit(IB_ACK_REQUESTED, &ic->i_ack_flags))
 772                 rds_ib_stats_inc(s_ib_ack_send_piggybacked);
 773         return rds_ib_get_ack(ic);
 774 }
 775
 776 /*
 777  * It's kind of lame that we're copying from the posted receive pages into
 778  * long-lived bitmaps.  We could have posted the bitmaps and rdma written into
 779  * them.  But receiving new congestion bitmaps should be a *rare* event, so
 780  * hopefully we won't need to invest that complexity in making it more
 781  * efficient.  By copying we can share a simpler core with TCP which has to
 782  * copy.
 783  */
 784 static void rds_ib_cong_recv(struct rds_connection *conn,
 785                               struct rds_ib_incoming *ibinc)
 786 {
 787         struct rds_cong_map *map;
 788         unsigned int map_off;
 789         unsigned int map_page;
 790         struct rds_page_frag *frag;
 791         unsigned long frag_off;
 792         unsigned long to_copy;
 793         unsigned long copied;
 794         __le64 uncongested = 0;
 795         void *addr;
 796
 797         /* catch completely corrupt packets */
 798         if (be32_to_cpu(ibinc->ii_inc.i_hdr.h_len) != RDS_CONG_MAP_BYTES)
 799                 return;
 800
 801         map = conn->c_fcong;
 802         map_page = 0;
 803         map_off = 0;
 804
 805         frag = list_entry(ibinc->ii_frags.next, struct rds_page_frag, f_item);
 806         frag_off = 0;
 807
 808         copied = 0;
 809
 810         while (copied < RDS_CONG_MAP_BYTES) {
 811                 __le64 *src, *dst;
 812                 unsigned int k;
 813
 814                 to_copy = min(RDS_FRAG_SIZE - frag_off, PAGE_SIZE - map_off);
 815                 BUG_ON(to_copy & 7); /* Must be 64bit aligned. */
 816
 817                 addr = kmap_atomic(sg_page(&frag->f_sg));
 818
 819                 src = addr + frag->f_sg.offset + frag_off;
 820                 dst = (void *)map->m_page_addrs[map_page] + map_off;
 821                 for (k = 0; k < to_copy; k += 8) {
 822                         /* Record ports that became uncongested, ie
 823                          * bits that changed from 0 to 1. */
 824                         uncongested |= ~(*src) & *dst;
 825                         *dst++ = *src++;
 826                 }
 827                 kunmap_atomic(addr);
 828
 829                 copied += to_copy;
 830
 831                 map_off += to_copy;
 832                 if (map_off == PAGE_SIZE) {
 833                         map_off = 0;
 834                         map_page++;
 835                 }
 836
 837                 frag_off += to_copy;
 838                 if (frag_off == RDS_FRAG_SIZE) {
 839                         frag = list_entry(frag->f_item.next,
 840                                           struct rds_page_frag, f_item);
 841                         frag_off = 0;
 842                 }
 843         }
 844
 845         /* the congestion map is in little endian order */
 846         rds_cong_map_updated(map, le64_to_cpu(uncongested));
 847 }
 848
 849 static void rds_ib_process_recv(struct rds_connection *conn,
 850                                 struct rds_ib_recv_work *recv, u32 data_len,
 851                                 struct rds_ib_ack_state *state)
 852 {
 853         struct rds_ib_connection *ic = conn->c_transport_data;
 854         struct rds_ib_incoming *ibinc = ic->i_ibinc;
 855         struct rds_header *ihdr, *hdr;
 856         dma_addr_t dma_addr = ic->i_recv_hdrs_dma[recv - ic->i_recvs];
 857
 858         /* XXX shut down the connection if port 0,0 are seen? */
 859
 860         rdsdebug("ic %p ibinc %p recv %p byte len %u\n", ic, ibinc, recv,
 861                  data_len);
 862
 863         if (data_len < sizeof(struct rds_header)) {
 864                 rds_ib_conn_error(conn, "incoming message "
 865                        "from %pI6c didn't include a "
 866                        "header, disconnecting and "
 867                        "reconnecting\n",
 868                        &conn->c_faddr);
 869                 return;
 870         }
 871         data_len -= sizeof(struct rds_header);
 872
 873         ihdr = ic->i_recv_hdrs[recv - ic->i_recvs];
 874
 875         ib_dma_sync_single_for_cpu(ic->rds_ibdev->dev, dma_addr,
 876                                    sizeof(*ihdr), DMA_FROM_DEVICE);
 877         /* Validate the checksum. */
 878         if (!rds_message_verify_checksum(ihdr)) {
 879                 rds_ib_conn_error(conn, "incoming message "
 880                        "from %pI6c has corrupted header - "
 881                        "forcing a reconnect\n",
 882                        &conn->c_faddr);
 883                 rds_stats_inc(s_recv_drop_bad_checksum);
 884                 goto done;
 885         }
 886
 887         /* Process the ACK sequence which comes with every packet */
 888         state->ack_recv = be64_to_cpu(ihdr->h_ack);
 889         state->ack_recv_valid = 1;
 890
 891         /* Process the credits update if there was one */
 892         if (ihdr->h_credit)
 893                 rds_ib_send_add_credits(conn, ihdr->h_credit);
 894
 895         if (ihdr->h_sport == 0 && ihdr->h_dport == 0 && data_len == 0) {
 896                 /* This is an ACK-only packet. The fact that it gets
 897                  * special treatment here is that historically, ACKs
 898                  * were rather special beasts.
 899                  */
 900                 rds_ib_stats_inc(s_ib_ack_received);
 901
 902                 /*
 903                  * Usually the frags make their way on to incs and are then freed as
 904                  * the inc is freed.  We don't go that route, so we have to drop the
 905                  * page ref ourselves.  We can't just leave the page on the recv
 906                  * because that confuses the dma mapping of pages and each recv's use
 907                  * of a partial page.
 908                  *
 909                  * FIXME: Fold this into the code path below.
 910                  */
 911                 rds_ib_frag_free(ic, recv->r_frag);
 912                 recv->r_frag = NULL;
 913                 goto done;
 914         }
 915
 916         /*
 917          * If we don't already have an inc on the connection then this
 918          * fragment has a header and starts a message.. copy its header
 919          * into the inc and save the inc so we can hang upcoming fragments
 920          * off its list.
 921          */
 922         if (!ibinc) {
 923                 ibinc = recv->r_ibinc;
 924                 recv->r_ibinc = NULL;
 925                 ic->i_ibinc = ibinc;
 926
 927                 hdr = &ibinc->ii_inc.i_hdr;
 928                 ibinc->ii_inc.i_rx_lat_trace[RDS_MSG_RX_HDR] =
 929                                 local_clock();
 930                 memcpy(hdr, ihdr, sizeof(*hdr));
 931                 ic->i_recv_data_rem = be32_to_cpu(hdr->h_len);
 932                 ibinc->ii_inc.i_rx_lat_trace[RDS_MSG_RX_START] =
 933                                 local_clock();
 934
 935                 rdsdebug("ic %p ibinc %p rem %u flag 0x%x\n", ic, ibinc,
 936                          ic->i_recv_data_rem, hdr->h_flags);
 937         } else {
 938                 hdr = &ibinc->ii_inc.i_hdr;
 939                 /* We can't just use memcmp here; fragments of a
 940                  * single message may carry different ACKs */
 941                 if (hdr->h_sequence != ihdr->h_sequence ||
 942                     hdr->h_len != ihdr->h_len ||
 943                     hdr->h_sport != ihdr->h_sport ||
 944                     hdr->h_dport != ihdr->h_dport) {
 945                         rds_ib_conn_error(conn,
 946                                 "fragment header mismatch; forcing reconnect\n");
 947                         goto done;
 948                 }
 949         }
 950
 951         list_add_tail(&recv->r_frag->f_item, &ibinc->ii_frags);
 952         recv->r_frag = NULL;
 953
 954         if (ic->i_recv_data_rem > RDS_FRAG_SIZE)
 955                 ic->i_recv_data_rem -= RDS_FRAG_SIZE;
 956         else {
 957                 ic->i_recv_data_rem = 0;
 958                 ic->i_ibinc = NULL;
 959
 960                 if (ibinc->ii_inc.i_hdr.h_flags == RDS_FLAG_CONG_BITMAP) {
 961                         rds_ib_cong_recv(conn, ibinc);
 962                 } else {
 963                         rds_recv_incoming(conn, &conn->c_faddr, &conn->c_laddr,
 964                                           &ibinc->ii_inc, GFP_ATOMIC);
 965                         state->ack_next = be64_to_cpu(hdr->h_sequence);
 966                         state->ack_next_valid = 1;
 967                 }
 968
 969                 /* Evaluate the ACK_REQUIRED flag *after* we received
 970                  * the complete frame, and after bumping the next_rx
 971                  * sequence. */
 972                 if (hdr->h_flags & RDS_FLAG_ACK_REQUIRED) {
 973                         rds_stats_inc(s_recv_ack_required);
 974                         state->ack_required = 1;
 975                 }
 976
 977                 rds_inc_put(&ibinc->ii_inc);
 978         }
 979 done:
 980         ib_dma_sync_single_for_device(ic->rds_ibdev->dev, dma_addr,
 981                                       sizeof(*ihdr), DMA_FROM_DEVICE);
 982 }
 983
 984 void rds_ib_recv_cqe_handler(struct rds_ib_connection *ic,
 985                              struct ib_wc *wc,
 986                              struct rds_ib_ack_state *state)
 987 {
 988         struct rds_connection *conn = ic->conn;
 989         struct rds_ib_recv_work *recv;
 990
 991         rdsdebug("wc wr_id 0x%llx status %u (%s) byte_len %u imm_data %u\n",
 992                  (unsigned long long)wc->wr_id, wc->status,
 993                  ib_wc_status_msg(wc->status), wc->byte_len,
 994                  be32_to_cpu(wc->ex.imm_data));
 995
 996         rds_ib_stats_inc(s_ib_rx_cq_event);
 997         recv = &ic->i_recvs[rds_ib_ring_oldest(&ic->i_recv_ring)];
 998         ib_dma_unmap_sg(ic->i_cm_id->device, &recv->r_frag->f_sg, 1,
 999                         DMA_FROM_DEVICE);
1000
1001         /* Also process recvs in connecting state because it is possible
1002          * to get a recv completion _before_ the rdmacm ESTABLISHED
1003          * event is processed.
1004          */
1005         if (wc->status == IB_WC_SUCCESS) {
1006                 rds_ib_process_recv(conn, recv, wc->byte_len, state);
1007         } else {
1008                 /* We expect errors as the qp is drained during shutdown */
1009                 if (rds_conn_up(conn) || rds_conn_connecting(conn))
1010                         rds_ib_conn_error(conn, "recv completion on <%pI6c,%pI6c, %d> had status %u (%s), vendor err 0x%x, disconnecting and reconnecting\n",
1011                                           &conn->c_laddr, &conn->c_faddr,
1012                                           conn->c_tos, wc->status,
1013                                           ib_wc_status_msg(wc->status),
1014                                           wc->vendor_err);
1015         }
1016
1017         /* rds_ib_process_recv() doesn't always consume the frag, and
1018          * we might not have called it at all if the wc didn't indicate
1019          * success. We already unmapped the frag's pages, though, and
1020          * the following rds_ib_ring_free() call tells the refill path
1021          * that it will not find an allocated frag here. Make sure we
1022          * keep that promise by freeing a frag that's still on the ring.
1023          */
1024         if (recv->r_frag) {
1025                 rds_ib_frag_free(ic, recv->r_frag);
1026                 recv->r_frag = NULL;
1027         }
1028         rds_ib_ring_free(&ic->i_recv_ring, 1);
1029
1030         /* If we ever end up with a really empty receive ring, we're
1031          * in deep trouble, as the sender will definitely see RNR
1032          * timeouts. */
1033         if (rds_ib_ring_empty(&ic->i_recv_ring))
1034                 rds_ib_stats_inc(s_ib_rx_ring_empty);
1035
1036         if (rds_ib_ring_low(&ic->i_recv_ring)) {
1037                 rds_ib_recv_refill(conn, 0, GFP_NOWAIT | __GFP_NOWARN);
1038                 rds_ib_stats_inc(s_ib_rx_refill_from_cq);
1039         }
1040 }
1041
1042 int rds_ib_recv_path(struct rds_conn_path *cp)
1043 {
1044         struct rds_connection *conn = cp->cp_conn;
1045         struct rds_ib_connection *ic = conn->c_transport_data;
1046
1047         rdsdebug("conn %p\n", conn);
1048         if (rds_conn_up(conn)) {
1049                 rds_ib_attempt_ack(ic);
1050                 rds_ib_recv_refill(conn, 0, GFP_KERNEL);
1051                 rds_ib_stats_inc(s_ib_rx_refill_from_thread);
1052         }
1053
1054         return 0;
1055 }
1056
1057 int rds_ib_recv_init(void)
1058 {
1059         struct sysinfo si;
1060         int ret = -ENOMEM;
1061
1062         /* Default to 30% of all available RAM for recv memory */
1063         si_meminfo(&si);
1064         rds_ib_sysctl_max_recv_allocation = si.totalram / 3 * PAGE_SIZE / RDS_FRAG_SIZE;
1065
1066         rds_ib_incoming_slab =
1067                 kmem_cache_create_usercopy("rds_ib_incoming",
1068                                            sizeof(struct rds_ib_incoming),
1069                                            0, SLAB_HWCACHE_ALIGN,
1070                                            offsetof(struct rds_ib_incoming,
1071                                                     ii_inc.i_usercopy),
1072                                            sizeof(struct rds_inc_usercopy),
1073                                            NULL);
1074         if (!rds_ib_incoming_slab)
1075                 goto out;
1076
1077         rds_ib_frag_slab = kmem_cache_create("rds_ib_frag",
1078                                         sizeof(struct rds_page_frag),
1079                                         0, SLAB_HWCACHE_ALIGN, NULL);
1080         if (!rds_ib_frag_slab) {
1081                 kmem_cache_destroy(rds_ib_incoming_slab);
1082                 rds_ib_incoming_slab = NULL;
1083         } else
1084                 ret = 0;
1085 out:
1086         return ret;
1087 }
1088
1089 void rds_ib_recv_exit(void)
1090 {
1091         WARN_ON(atomic_read(&rds_ib_allocation));
1092
1093         kmem_cache_destroy(rds_ib_incoming_slab);
1094         kmem_cache_destroy(rds_ib_frag_slab);
1095 }