net/rds/ib_recv.c

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

DEFINITIONS
