net/rds/ib_recv.c

   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 = ib_sg_dma_address(ic->i_cm_id->device, &recv->r_frag->f_sg);
 350         sge->length = ib_sg_dma_len(ic->i_cm_id->device, &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         struct ib_recv_wr *failed_wr;
 385         unsigned int posted = 0;
 386         int ret = 0;
 387         bool can_wait = !!(gfp & __GFP_DIRECT_RECLAIM);
 388         u32 pos;
 389
 390         /* the goal here is to just make sure that someone, somewhere
 391          * is posting buffers.  If we can't get the refill lock,
 392          * let them do their thing
 393          */
 394         if (!acquire_refill(conn))
 395                 return;
 396
 397         while ((prefill || rds_conn_up(conn)) &&
 398                rds_ib_ring_alloc(&ic->i_recv_ring, 1, &pos)) {
 399                 if (pos >= ic->i_recv_ring.w_nr) {
 400                         printk(KERN_NOTICE "Argh - ring alloc returned pos=%u\n",
 401                                         pos);
 402                         break;
 403                 }
 404
 405                 recv = &ic->i_recvs[pos];
 406                 ret = rds_ib_recv_refill_one(conn, recv, gfp);
 407                 if (ret) {
 408                         break;
 409                 }
 410
 411                 rdsdebug("recv %p ibinc %p page %p addr %lu\n", recv,
 412                          recv->r_ibinc, sg_page(&recv->r_frag->f_sg),
 413                          (long) ib_sg_dma_address(
 414                                 ic->i_cm_id->device,
 415                                 &recv->r_frag->f_sg));
 416
 417                 /* XXX when can this fail? */
 418                 ret = ib_post_recv(ic->i_cm_id->qp, &recv->r_wr, &failed_wr);
 419                 if (ret) {
 420                         rds_ib_conn_error(conn, "recv post on "
 421                                "%pI6c returned %d, disconnecting and "
 422                                "reconnecting\n", &conn->c_faddr,
 423                                ret);
 424                         break;
 425                 }
 426
 427                 posted++;
 428         }
 429
 430         /* We're doing flow control - update the window. */
 431         if (ic->i_flowctl && posted)
 432                 rds_ib_advertise_credits(conn, posted);
 433
 434         if (ret)
 435                 rds_ib_ring_unalloc(&ic->i_recv_ring, 1);
 436
 437         release_refill(conn);
 438
 439         /* if we're called from the softirq handler, we'll be GFP_NOWAIT.
 440          * in this case the ring being low is going to lead to more interrupts
 441          * and we can safely let the softirq code take care of it unless the
 442          * ring is completely empty.
 443          *
 444          * if we're called from krdsd, we'll be GFP_KERNEL.  In this case
 445          * we might have raced with the softirq code while we had the refill
 446          * lock held.  Use rds_ib_ring_low() instead of ring_empty to decide
 447          * if we should requeue.
 448          */
 449         if (rds_conn_up(conn) &&
 450             ((can_wait && rds_ib_ring_low(&ic->i_recv_ring)) ||
 451             rds_ib_ring_empty(&ic->i_recv_ring))) {
 452                 queue_delayed_work(rds_wq, &conn->c_recv_w, 1);
 453         }
 454 }
 455
 456 /*
 457  * We want to recycle several types of recv allocations, like incs and frags.
 458  * To use this, the *_free() function passes in the ptr to a list_head within
 459  * the recyclee, as well as the cache to put it on.
 460  *
 461  * First, we put the memory on a percpu list. When this reaches a certain size,
 462  * We move it to an intermediate non-percpu list in a lockless manner, with some
 463  * xchg/compxchg wizardry.
 464  *
 465  * N.B. Instead of a list_head as the anchor, we use a single pointer, which can
 466  * be NULL and xchg'd. The list is actually empty when the pointer is NULL, and
 467  * list_empty() will return true with one element is actually present.
 468  */
 469 static void rds_ib_recv_cache_put(struct list_head *new_item,
 470                                  struct rds_ib_refill_cache *cache)
 471 {
 472         unsigned long flags;
 473         struct list_head *old, *chpfirst;
 474
 475         local_irq_save(flags);
 476
 477         chpfirst = __this_cpu_read(cache->percpu->first);
 478         if (!chpfirst)
 479                 INIT_LIST_HEAD(new_item);
 480         else /* put on front */
 481                 list_add_tail(new_item, chpfirst);
 482
 483         __this_cpu_write(cache->percpu->first, new_item);
 484         __this_cpu_inc(cache->percpu->count);
 485
 486         if (__this_cpu_read(cache->percpu->count) < RDS_IB_RECYCLE_BATCH_COUNT)
 487                 goto end;
 488
 489         /*
 490          * Return our per-cpu first list to the cache's xfer by atomically
 491          * grabbing the current xfer list, appending it to our per-cpu list,
 492          * and then atomically returning that entire list back to the
 493          * cache's xfer list as long as it's still empty.
 494          */
 495         do {
 496                 old = xchg(&cache->xfer, NULL);
 497                 if (old)
 498                         list_splice_entire_tail(old, chpfirst);
 499                 old = cmpxchg(&cache->xfer, NULL, chpfirst);
 500         } while (old);
 501
 502
 503         __this_cpu_write(cache->percpu->first, NULL);
 504         __this_cpu_write(cache->percpu->count, 0);
 505 end:
 506         local_irq_restore(flags);
 507 }
 508
 509 static struct list_head *rds_ib_recv_cache_get(struct rds_ib_refill_cache *cache)
 510 {
 511         struct list_head *head = cache->ready;
 512
 513         if (head) {
 514                 if (!list_empty(head)) {
 515                         cache->ready = head->next;
 516                         list_del_init(head);
 517                 } else
 518                         cache->ready = NULL;
 519         }
 520
 521         return head;
 522 }
 523
 524 int rds_ib_inc_copy_to_user(struct rds_incoming *inc, struct iov_iter *to)
 525 {
 526         struct rds_ib_incoming *ibinc;
 527         struct rds_page_frag *frag;
 528         unsigned long to_copy;
 529         unsigned long frag_off = 0;
 530         int copied = 0;
 531         int ret;
 532         u32 len;
 533
 534         ibinc = container_of(inc, struct rds_ib_incoming, ii_inc);
 535         frag = list_entry(ibinc->ii_frags.next, struct rds_page_frag, f_item);
 536         len = be32_to_cpu(inc->i_hdr.h_len);
 537
 538         while (iov_iter_count(to) && copied < len) {
 539                 if (frag_off == RDS_FRAG_SIZE) {
 540                         frag = list_entry(frag->f_item.next,
 541                                           struct rds_page_frag, f_item);
 542                         frag_off = 0;
 543                 }
 544                 to_copy = min_t(unsigned long, iov_iter_count(to),
 545                                 RDS_FRAG_SIZE - frag_off);
 546                 to_copy = min_t(unsigned long, to_copy, len - copied);
 547
 548                 /* XXX needs + offset for multiple recvs per page */
 549                 rds_stats_add(s_copy_to_user, to_copy);
 550                 ret = copy_page_to_iter(sg_page(&frag->f_sg),
 551                                         frag->f_sg.offset + frag_off,
 552                                         to_copy,
 553                                         to);
 554                 if (ret != to_copy)
 555                         return -EFAULT;
 556
 557                 frag_off += to_copy;
 558                 copied += to_copy;
 559         }
 560
 561         return copied;
 562 }
 563
 564 /* ic starts out kzalloc()ed */
 565 void rds_ib_recv_init_ack(struct rds_ib_connection *ic)
 566 {
 567         struct ib_send_wr *wr = &ic->i_ack_wr;
 568         struct ib_sge *sge = &ic->i_ack_sge;
 569
 570         sge->addr = ic->i_ack_dma;
 571         sge->length = sizeof(struct rds_header);
 572         sge->lkey = ic->i_pd->local_dma_lkey;
 573
 574         wr->sg_list = sge;
 575         wr->num_sge = 1;
 576         wr->opcode = IB_WR_SEND;
 577         wr->wr_id = RDS_IB_ACK_WR_ID;
 578         wr->send_flags = IB_SEND_SIGNALED | IB_SEND_SOLICITED;
 579 }
 580
 581 /*
 582  * You'd think that with reliable IB connections you wouldn't need to ack
 583  * messages that have been received.  The problem is that IB hardware generates
 584  * an ack message before it has DMAed the message into memory.  This creates a
 585  * potential message loss if the HCA is disabled for any reason between when it
 586  * sends the ack and before the message is DMAed and processed.  This is only a
 587  * potential issue if another HCA is available for fail-over.
 588  *
 589  * When the remote host receives our ack they'll free the sent message from
 590  * their send queue.  To decrease the latency of this we always send an ack
 591  * immediately after we've received messages.
 592  *
 593  * For simplicity, we only have one ack in flight at a time.  This puts
 594  * pressure on senders to have deep enough send queues to absorb the latency of
 595  * a single ack frame being in flight.  This might not be good enough.
 596  *
 597  * This is implemented by have a long-lived send_wr and sge which point to a
 598  * statically allocated ack frame.  This ack wr does not fall under the ring
 599  * accounting that the tx and rx wrs do.  The QP attribute specifically makes
 600  * room for it beyond the ring size.  Send completion notices its special
 601  * wr_id and avoids working with the ring in that case.
 602  */
 603 #ifndef KERNEL_HAS_ATOMIC64
 604 void rds_ib_set_ack(struct rds_ib_connection *ic, u64 seq, int ack_required)
 605 {
 606         unsigned long flags;
 607
 608         spin_lock_irqsave(&ic->i_ack_lock, flags);
 609         ic->i_ack_next = seq;
 610         if (ack_required)
 611                 set_bit(IB_ACK_REQUESTED, &ic->i_ack_flags);
 612         spin_unlock_irqrestore(&ic->i_ack_lock, flags);
 613 }
 614
 615 static u64 rds_ib_get_ack(struct rds_ib_connection *ic)
 616 {
 617         unsigned long flags;
 618         u64 seq;
 619
 620         clear_bit(IB_ACK_REQUESTED, &ic->i_ack_flags);
 621
 622         spin_lock_irqsave(&ic->i_ack_lock, flags);
 623         seq = ic->i_ack_next;
 624         spin_unlock_irqrestore(&ic->i_ack_lock, flags);
 625
 626         return seq;
 627 }
 628 #else
 629 void rds_ib_set_ack(struct rds_ib_connection *ic, u64 seq, int ack_required)
 630 {
 631         atomic64_set(&ic->i_ack_next, seq);
 632         if (ack_required) {
 633                 smp_mb__before_atomic();
 634                 set_bit(IB_ACK_REQUESTED, &ic->i_ack_flags);
 635         }
 636 }
 637
 638 static u64 rds_ib_get_ack(struct rds_ib_connection *ic)
 639 {
 640         clear_bit(IB_ACK_REQUESTED, &ic->i_ack_flags);
 641         smp_mb__after_atomic();
 642
 643         return atomic64_read(&ic->i_ack_next);
 644 }
 645 #endif
 646
 647
 648 static void rds_ib_send_ack(struct rds_ib_connection *ic, unsigned int adv_credits)
 649 {
 650         struct rds_header *hdr = ic->i_ack;
 651         struct ib_send_wr *failed_wr;
 652         u64 seq;
 653         int ret;
 654
 655         seq = rds_ib_get_ack(ic);
 656
 657         rdsdebug("send_ack: ic %p ack %llu\n", ic, (unsigned long long) seq);
 658         rds_message_populate_header(hdr, 0, 0, 0);
 659         hdr->h_ack = cpu_to_be64(seq);
 660         hdr->h_credit = adv_credits;
 661         rds_message_make_checksum(hdr);
 662         ic->i_ack_queued = jiffies;
 663
 664         ret = ib_post_send(ic->i_cm_id->qp, &ic->i_ack_wr, &failed_wr);
 665         if (unlikely(ret)) {
 666                 /* Failed to send. Release the WR, and
 667                  * force another ACK.
 668                  */
 669                 clear_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags);
 670                 set_bit(IB_ACK_REQUESTED, &ic->i_ack_flags);
 671
 672                 rds_ib_stats_inc(s_ib_ack_send_failure);
 673
 674                 rds_ib_conn_error(ic->conn, "sending ack failed\n");
 675         } else
 676                 rds_ib_stats_inc(s_ib_ack_sent);
 677 }
 678
 679 /*
 680  * There are 3 ways of getting acknowledgements to the peer:
 681  *  1.  We call rds_ib_attempt_ack from the recv completion handler
 682  *      to send an ACK-only frame.
 683  *      However, there can be only one such frame in the send queue
 684  *      at any time, so we may have to postpone it.
 685  *  2.  When another (data) packet is transmitted while there's
 686  *      an ACK in the queue, we piggyback the ACK sequence number
 687  *      on the data packet.
 688  *  3.  If the ACK WR is done sending, we get called from the
 689  *      send queue completion handler, and check whether there's
 690  *      another ACK pending (postponed because the WR was on the
 691  *      queue). If so, we transmit it.
 692  *
 693  * We maintain 2 variables:
 694  *  -   i_ack_flags, which keeps track of whether the ACK WR
 695  *      is currently in the send queue or not (IB_ACK_IN_FLIGHT)
 696  *  -   i_ack_next, which is the last sequence number we received
 697  *
 698  * Potentially, send queue and receive queue handlers can run concurrently.
 699  * It would be nice to not have to use a spinlock to synchronize things,
 700  * but the one problem that rules this out is that 64bit updates are
 701  * not atomic on all platforms. Things would be a lot simpler if
 702  * we had atomic64 or maybe cmpxchg64 everywhere.
 703  *
 704  * Reconnecting complicates this picture just slightly. When we
 705  * reconnect, we may be seeing duplicate packets. The peer
 706  * is retransmitting them, because it hasn't seen an ACK for
 707  * them. It is important that we ACK these.
 708  *
 709  * ACK mitigation adds a header flag "ACK_REQUIRED"; any packet with
 710  * this flag set *MUST* be acknowledged immediately.
 711  */
 712
 713 /*
 714  * When we get here, we're called from the recv queue handler.
 715  * Check whether we ought to transmit an ACK.
 716  */
 717 void rds_ib_attempt_ack(struct rds_ib_connection *ic)
 718 {
 719         unsigned int adv_credits;
 720
 721         if (!test_bit(IB_ACK_REQUESTED, &ic->i_ack_flags))
 722                 return;
 723
 724         if (test_and_set_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags)) {
 725                 rds_ib_stats_inc(s_ib_ack_send_delayed);
 726                 return;
 727         }
 728
 729         /* Can we get a send credit? */
 730         if (!rds_ib_send_grab_credits(ic, 1, &adv_credits, 0, RDS_MAX_ADV_CREDIT)) {
 731                 rds_ib_stats_inc(s_ib_tx_throttle);
 732                 clear_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags);
 733                 return;
 734         }
 735
 736         clear_bit(IB_ACK_REQUESTED, &ic->i_ack_flags);
 737         rds_ib_send_ack(ic, adv_credits);
 738 }
 739
 740 /*
 741  * We get here from the send completion handler, when the
 742  * adapter tells us the ACK frame was sent.
 743  */
 744 void rds_ib_ack_send_complete(struct rds_ib_connection *ic)
 745 {
 746         clear_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags);
 747         rds_ib_attempt_ack(ic);
 748 }
 749
 750 /*
 751  * This is called by the regular xmit code when it wants to piggyback
 752  * an ACK on an outgoing frame.
 753  */
 754 u64 rds_ib_piggyb_ack(struct rds_ib_connection *ic)
 755 {
 756         if (test_and_clear_bit(IB_ACK_REQUESTED, &ic->i_ack_flags))
 757                 rds_ib_stats_inc(s_ib_ack_send_piggybacked);
 758         return rds_ib_get_ack(ic);
 759 }
 760
 761 /*
 762  * It's kind of lame that we're copying from the posted receive pages into
 763  * long-lived bitmaps.  We could have posted the bitmaps and rdma written into
 764  * them.  But receiving new congestion bitmaps should be a *rare* event, so
 765  * hopefully we won't need to invest that complexity in making it more
 766  * efficient.  By copying we can share a simpler core with TCP which has to
 767  * copy.
 768  */
 769 static void rds_ib_cong_recv(struct rds_connection *conn,
 770                               struct rds_ib_incoming *ibinc)
 771 {
 772         struct rds_cong_map *map;
 773         unsigned int map_off;
 774         unsigned int map_page;
 775         struct rds_page_frag *frag;
 776         unsigned long frag_off;
 777         unsigned long to_copy;
 778         unsigned long copied;
 779         uint64_t uncongested = 0;
 780         void *addr;
 781
 782         /* catch completely corrupt packets */
 783         if (be32_to_cpu(ibinc->ii_inc.i_hdr.h_len) != RDS_CONG_MAP_BYTES)
 784                 return;
 785
 786         map = conn->c_fcong;
 787         map_page = 0;
 788         map_off = 0;
 789
 790         frag = list_entry(ibinc->ii_frags.next, struct rds_page_frag, f_item);
 791         frag_off = 0;
 792
 793         copied = 0;
 794
 795         while (copied < RDS_CONG_MAP_BYTES) {
 796                 uint64_t *src, *dst;
 797                 unsigned int k;
 798
 799                 to_copy = min(RDS_FRAG_SIZE - frag_off, PAGE_SIZE - map_off);
 800                 BUG_ON(to_copy & 7); /* Must be 64bit aligned. */
 801
 802                 addr = kmap_atomic(sg_page(&frag->f_sg));
 803
 804                 src = addr + frag->f_sg.offset + frag_off;
 805                 dst = (void *)map->m_page_addrs[map_page] + map_off;
 806                 for (k = 0; k < to_copy; k += 8) {
 807                         /* Record ports that became uncongested, ie
 808                          * bits that changed from 0 to 1. */
 809                         uncongested |= ~(*src) & *dst;
 810                         *dst++ = *src++;
 811                 }
 812                 kunmap_atomic(addr);
 813
 814                 copied += to_copy;
 815
 816                 map_off += to_copy;
 817                 if (map_off == PAGE_SIZE) {
 818                         map_off = 0;
 819                         map_page++;
 820                 }
 821
 822                 frag_off += to_copy;
 823                 if (frag_off == RDS_FRAG_SIZE) {
 824                         frag = list_entry(frag->f_item.next,
 825                                           struct rds_page_frag, f_item);
 826                         frag_off = 0;
 827                 }
 828         }
 829
 830         /* the congestion map is in little endian order */
 831         uncongested = le64_to_cpu(uncongested);
 832
 833         rds_cong_map_updated(map, uncongested);
 834 }
 835
 836 static void rds_ib_process_recv(struct rds_connection *conn,
 837                                 struct rds_ib_recv_work *recv, u32 data_len,
 838                                 struct rds_ib_ack_state *state)
 839 {
 840         struct rds_ib_connection *ic = conn->c_transport_data;
 841         struct rds_ib_incoming *ibinc = ic->i_ibinc;
 842         struct rds_header *ihdr, *hdr;
 843
 844         /* XXX shut down the connection if port 0,0 are seen? */
 845
 846         rdsdebug("ic %p ibinc %p recv %p byte len %u\n", ic, ibinc, recv,
 847                  data_len);
 848
 849         if (data_len < sizeof(struct rds_header)) {
 850                 rds_ib_conn_error(conn, "incoming message "
 851                        "from %pI6c didn't include a "
 852                        "header, disconnecting and "
 853                        "reconnecting\n",
 854                        &conn->c_faddr);
 855                 return;
 856         }
 857         data_len -= sizeof(struct rds_header);
 858
 859         ihdr = &ic->i_recv_hdrs[recv - ic->i_recvs];
 860
 861         /* Validate the checksum. */
 862         if (!rds_message_verify_checksum(ihdr)) {
 863                 rds_ib_conn_error(conn, "incoming message "
 864                        "from %pI6c has corrupted header - "
 865                        "forcing a reconnect\n",
 866                        &conn->c_faddr);
 867                 rds_stats_inc(s_recv_drop_bad_checksum);
 868                 return;
 869         }
 870
 871         /* Process the ACK sequence which comes with every packet */
 872         state->ack_recv = be64_to_cpu(ihdr->h_ack);
 873         state->ack_recv_valid = 1;
 874
 875         /* Process the credits update if there was one */
 876         if (ihdr->h_credit)
 877                 rds_ib_send_add_credits(conn, ihdr->h_credit);
 878
 879         if (ihdr->h_sport == 0 && ihdr->h_dport == 0 && data_len == 0) {
 880                 /* This is an ACK-only packet. The fact that it gets
 881                  * special treatment here is that historically, ACKs
 882                  * were rather special beasts.
 883                  */
 884                 rds_ib_stats_inc(s_ib_ack_received);
 885
 886                 /*
 887                  * Usually the frags make their way on to incs and are then freed as
 888                  * the inc is freed.  We don't go that route, so we have to drop the
 889                  * page ref ourselves.  We can't just leave the page on the recv
 890                  * because that confuses the dma mapping of pages and each recv's use
 891                  * of a partial page.
 892                  *
 893                  * FIXME: Fold this into the code path below.
 894                  */
 895                 rds_ib_frag_free(ic, recv->r_frag);
 896                 recv->r_frag = NULL;
 897                 return;
 898         }
 899
 900         /*
 901          * If we don't already have an inc on the connection then this
 902          * fragment has a header and starts a message.. copy its header
 903          * into the inc and save the inc so we can hang upcoming fragments
 904          * off its list.
 905          */
 906         if (!ibinc) {
 907                 ibinc = recv->r_ibinc;
 908                 recv->r_ibinc = NULL;
 909                 ic->i_ibinc = ibinc;
 910
 911                 hdr = &ibinc->ii_inc.i_hdr;
 912                 ibinc->ii_inc.i_rx_lat_trace[RDS_MSG_RX_HDR] =
 913                                 local_clock();
 914                 memcpy(hdr, ihdr, sizeof(*hdr));
 915                 ic->i_recv_data_rem = be32_to_cpu(hdr->h_len);
 916                 ibinc->ii_inc.i_rx_lat_trace[RDS_MSG_RX_START] =
 917                                 local_clock();
 918
 919                 rdsdebug("ic %p ibinc %p rem %u flag 0x%x\n", ic, ibinc,
 920                          ic->i_recv_data_rem, hdr->h_flags);
 921         } else {
 922                 hdr = &ibinc->ii_inc.i_hdr;
 923                 /* We can't just use memcmp here; fragments of a
 924                  * single message may carry different ACKs */
 925                 if (hdr->h_sequence != ihdr->h_sequence ||
 926                     hdr->h_len != ihdr->h_len ||
 927                     hdr->h_sport != ihdr->h_sport ||
 928                     hdr->h_dport != ihdr->h_dport) {
 929                         rds_ib_conn_error(conn,
 930                                 "fragment header mismatch; forcing reconnect\n");
 931                         return;
 932                 }
 933         }
 934
 935         list_add_tail(&recv->r_frag->f_item, &ibinc->ii_frags);
 936         recv->r_frag = NULL;
 937
 938         if (ic->i_recv_data_rem > RDS_FRAG_SIZE)
 939                 ic->i_recv_data_rem -= RDS_FRAG_SIZE;
 940         else {
 941                 ic->i_recv_data_rem = 0;
 942                 ic->i_ibinc = NULL;
 943
 944                 if (ibinc->ii_inc.i_hdr.h_flags == RDS_FLAG_CONG_BITMAP) {
 945                         rds_ib_cong_recv(conn, ibinc);
 946                 } else {
 947                         rds_recv_incoming(conn, &conn->c_faddr, &conn->c_laddr,
 948                                           &ibinc->ii_inc, GFP_ATOMIC);
 949                         state->ack_next = be64_to_cpu(hdr->h_sequence);
 950                         state->ack_next_valid = 1;
 951                 }
 952
 953                 /* Evaluate the ACK_REQUIRED flag *after* we received
 954                  * the complete frame, and after bumping the next_rx
 955                  * sequence. */
 956                 if (hdr->h_flags & RDS_FLAG_ACK_REQUIRED) {
 957                         rds_stats_inc(s_recv_ack_required);
 958                         state->ack_required = 1;
 959                 }
 960
 961                 rds_inc_put(&ibinc->ii_inc);
 962         }
 963 }
 964
 965 void rds_ib_recv_cqe_handler(struct rds_ib_connection *ic,
 966                              struct ib_wc *wc,
 967                              struct rds_ib_ack_state *state)
 968 {
 969         struct rds_connection *conn = ic->conn;
 970         struct rds_ib_recv_work *recv;
 971
 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                  ib_wc_status_msg(wc->status), wc->byte_len,
 975                  be32_to_cpu(wc->ex.imm_data));
 976
 977         rds_ib_stats_inc(s_ib_rx_cq_event);
 978         recv = &ic->i_recvs[rds_ib_ring_oldest(&ic->i_recv_ring)];
 979         ib_dma_unmap_sg(ic->i_cm_id->device, &recv->r_frag->f_sg, 1,
 980                         DMA_FROM_DEVICE);
 981
 982         /* Also process recvs in connecting state because it is possible
 983          * to get a recv completion _before_ the rdmacm ESTABLISHED
 984          * event is processed.
 985          */
 986         if (wc->status == IB_WC_SUCCESS) {
 987                 rds_ib_process_recv(conn, recv, wc->byte_len, state);
 988         } else {
 989                 /* We expect errors as the qp is drained during shutdown */
 990                 if (rds_conn_up(conn) || rds_conn_connecting(conn))
 991                         rds_ib_conn_error(conn, "recv completion on <%pI6c,%pI6c> had status %u (%s), disconnecting and reconnecting\n",
 992                                           &conn->c_laddr, &conn->c_faddr,
 993                                           wc->status,
 994                                           ib_wc_status_msg(wc->status));
 995         }
 996
 997         /* rds_ib_process_recv() doesn't always consume the frag, and
 998          * we might not have called it at all if the wc didn't indicate
 999          * success. We already unmapped the frag's pages, though, and
1000          * the following rds_ib_ring_free() call tells the refill path
1001          * that it will not find an allocated frag here. Make sure we
1002          * keep that promise by freeing a frag that's still on the ring.
1003          */
1004         if (recv->r_frag) {
1005                 rds_ib_frag_free(ic, recv->r_frag);
1006                 recv->r_frag = NULL;
1007         }
1008         rds_ib_ring_free(&ic->i_recv_ring, 1);
1009
1010         /* If we ever end up with a really empty receive ring, we're
1011          * in deep trouble, as the sender will definitely see RNR
1012          * timeouts. */
1013         if (rds_ib_ring_empty(&ic->i_recv_ring))
1014                 rds_ib_stats_inc(s_ib_rx_ring_empty);
1015
1016         if (rds_ib_ring_low(&ic->i_recv_ring)) {
1017                 rds_ib_recv_refill(conn, 0, GFP_NOWAIT);
1018                 rds_ib_stats_inc(s_ib_rx_refill_from_cq);
1019         }
1020 }
1021
1022 int rds_ib_recv_path(struct rds_conn_path *cp)
1023 {
1024         struct rds_connection *conn = cp->cp_conn;
1025         struct rds_ib_connection *ic = conn->c_transport_data;
1026
1027         rdsdebug("conn %p\n", conn);
1028         if (rds_conn_up(conn)) {
1029                 rds_ib_attempt_ack(ic);
1030                 rds_ib_recv_refill(conn, 0, GFP_KERNEL);
1031                 rds_ib_stats_inc(s_ib_rx_refill_from_thread);
1032         }
1033
1034         return 0;
1035 }
1036
1037 int rds_ib_recv_init(void)
1038 {
1039         struct sysinfo si;
1040         int ret = -ENOMEM;
1041
1042         /* Default to 30% of all available RAM for recv memory */
1043         si_meminfo(&si);
1044         rds_ib_sysctl_max_recv_allocation = si.totalram / 3 * PAGE_SIZE / RDS_FRAG_SIZE;
1045
1046         rds_ib_incoming_slab = kmem_cache_create("rds_ib_incoming",
1047                                         sizeof(struct rds_ib_incoming),
1048                                         0, SLAB_HWCACHE_ALIGN, NULL);
1049         if (!rds_ib_incoming_slab)
1050                 goto out;
1051
1052         rds_ib_frag_slab = kmem_cache_create("rds_ib_frag",
1053                                         sizeof(struct rds_page_frag),
1054                                         0, SLAB_HWCACHE_ALIGN, NULL);
1055         if (!rds_ib_frag_slab) {
1056                 kmem_cache_destroy(rds_ib_incoming_slab);
1057                 rds_ib_incoming_slab = NULL;
1058         } else
1059                 ret = 0;
1060 out:
1061         return ret;
1062 }
1063
1064 void rds_ib_recv_exit(void)
1065 {
1066         kmem_cache_destroy(rds_ib_incoming_slab);
1067         kmem_cache_destroy(rds_ib_frag_slab);
1068 }