linux/net/rds/ib_recv.c
<<
>>
Prefs
   1/*
   2 * Copyright (c) 2006 Oracle.  All rights reserved.
   3 *
   4 * This software is available to you under a choice of one of two
   5 * licenses.  You may choose to be licensed under the terms of the GNU
   6 * General Public License (GPL) Version 2, available from the file
   7 * COPYING in the main directory of this source tree, or the
   8 * OpenIB.org BSD license below:
   9 *
  10 *     Redistribution and use in source and binary forms, with or
  11 *     without modification, are permitted provided that the following
  12 *     conditions are met:
  13 *
  14 *      - Redistributions of source code must retain the above
  15 *        copyright notice, this list of conditions and the following
  16 *        disclaimer.
  17 *
  18 *      - Redistributions in binary form must reproduce the above
  19 *        copyright notice, this list of conditions and the following
  20 *        disclaimer in the documentation and/or other materials
  21 *        provided with the distribution.
  22 *
  23 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  24 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  25 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  26 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  27 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  28 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  29 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  30 * SOFTWARE.
  31 *
  32 */
  33#include <linux/kernel.h>
  34#include <linux/slab.h>
  35#include <linux/pci.h>
  36#include <linux/dma-mapping.h>
  37#include <rdma/rdma_cm.h>
  38
  39#include "rds_single_path.h"
  40#include "rds.h"
  41#include "ib.h"
  42
  43static struct kmem_cache *rds_ib_incoming_slab;
  44static struct kmem_cache *rds_ib_frag_slab;
  45static atomic_t rds_ib_allocation = ATOMIC_INIT(0);
  46
  47void 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 */
  79static 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
  88static 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
 101static int rds_ib_recv_alloc_cache(struct rds_ib_refill_cache *cache)
 102{
 103        struct rds_ib_cache_head *head;
 104        int cpu;
 105
 106        cache->percpu = alloc_percpu(struct rds_ib_cache_head);
 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
 121int rds_ib_recv_alloc_caches(struct rds_ib_connection *ic)
 122{
 123        int ret;
 124
 125        ret = rds_ib_recv_alloc_cache(&ic->i_cache_incs);
 126        if (!ret) {
 127                ret = rds_ib_recv_alloc_cache(&ic->i_cache_frags);
 128                if (ret)
 129                        free_percpu(ic->i_cache_incs.percpu);
 130        }
 131
 132        return ret;
 133}
 134
 135static 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
 155void 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 */
 185static void rds_ib_recv_cache_put(struct list_head *new_item,
 186                                  struct rds_ib_refill_cache *cache);
 187static 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 */
 191static 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 */
 202void 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
 222static 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
 236void 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
 244static 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
 274static 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
 306static 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;
 353out:
 354        return ret;
 355}
 356
 357static int acquire_refill(struct rds_connection *conn)
 358{
 359        return test_and_set_bit(RDS_RECV_REFILL, &conn->c_flags) == 0;
 360}
 361
 362static 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 * -1 is returned if posting fails due to temporary resource exhaustion.
 381 */
 382void rds_ib_recv_refill(struct rds_connection *conn, int prefill, gfp_t gfp)
 383{
 384        struct rds_ib_connection *ic = conn->c_transport_data;
 385        struct rds_ib_recv_work *recv;
 386        struct ib_recv_wr *failed_wr;
 387        unsigned int posted = 0;
 388        int ret = 0;
 389        bool can_wait = !!(gfp & __GFP_DIRECT_RECLAIM);
 390        u32 pos;
 391
 392        /* the goal here is to just make sure that someone, somewhere
 393         * is posting buffers.  If we can't get the refill lock,
 394         * let them do their thing
 395         */
 396        if (!acquire_refill(conn))
 397                return;
 398
 399        while ((prefill || rds_conn_up(conn)) &&
 400               rds_ib_ring_alloc(&ic->i_recv_ring, 1, &pos)) {
 401                if (pos >= ic->i_recv_ring.w_nr) {
 402                        printk(KERN_NOTICE "Argh - ring alloc returned pos=%u\n",
 403                                        pos);
 404                        break;
 405                }
 406
 407                recv = &ic->i_recvs[pos];
 408                ret = rds_ib_recv_refill_one(conn, recv, gfp);
 409                if (ret) {
 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) ib_sg_dma_address(
 416                                ic->i_cm_id->device,
 417                                &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, &failed_wr);
 421                if (ret) {
 422                        rds_ib_conn_error(conn, "recv post on "
 423                               "%pI4 returned %d, disconnecting and "
 424                               "reconnecting\n", &conn->c_faddr,
 425                               ret);
 426                        break;
 427                }
 428
 429                posted++;
 430        }
 431
 432        /* We're doing flow control - update the window. */
 433        if (ic->i_flowctl && posted)
 434                rds_ib_advertise_credits(conn, posted);
 435
 436        if (ret)
 437                rds_ib_ring_unalloc(&ic->i_recv_ring, 1);
 438
 439        release_refill(conn);
 440
 441        /* if we're called from the softirq handler, we'll be GFP_NOWAIT.
 442         * in this case the ring being low is going to lead to more interrupts
 443         * and we can safely let the softirq code take care of it unless the
 444         * ring is completely empty.
 445         *
 446         * if we're called from krdsd, we'll be GFP_KERNEL.  In this case
 447         * we might have raced with the softirq code while we had the refill
 448         * lock held.  Use rds_ib_ring_low() instead of ring_empty to decide
 449         * if we should requeue.
 450         */
 451        if (rds_conn_up(conn) &&
 452            ((can_wait && rds_ib_ring_low(&ic->i_recv_ring)) ||
 453            rds_ib_ring_empty(&ic->i_recv_ring))) {
 454                queue_delayed_work(rds_wq, &conn->c_recv_w, 1);
 455        }
 456}
 457
 458/*
 459 * We want to recycle several types of recv allocations, like incs and frags.
 460 * To use this, the *_free() function passes in the ptr to a list_head within
 461 * the recyclee, as well as the cache to put it on.
 462 *
 463 * First, we put the memory on a percpu list. When this reaches a certain size,
 464 * We move it to an intermediate non-percpu list in a lockless manner, with some
 465 * xchg/compxchg wizardry.
 466 *
 467 * N.B. Instead of a list_head as the anchor, we use a single pointer, which can
 468 * be NULL and xchg'd. The list is actually empty when the pointer is NULL, and
 469 * list_empty() will return true with one element is actually present.
 470 */
 471static void rds_ib_recv_cache_put(struct list_head *new_item,
 472                                 struct rds_ib_refill_cache *cache)
 473{
 474        unsigned long flags;
 475        struct list_head *old, *chpfirst;
 476
 477        local_irq_save(flags);
 478
 479        chpfirst = __this_cpu_read(cache->percpu->first);
 480        if (!chpfirst)
 481                INIT_LIST_HEAD(new_item);
 482        else /* put on front */
 483                list_add_tail(new_item, chpfirst);
 484
 485        __this_cpu_write(cache->percpu->first, new_item);
 486        __this_cpu_inc(cache->percpu->count);
 487
 488        if (__this_cpu_read(cache->percpu->count) < RDS_IB_RECYCLE_BATCH_COUNT)
 489                goto end;
 490
 491        /*
 492         * Return our per-cpu first list to the cache's xfer by atomically
 493         * grabbing the current xfer list, appending it to our per-cpu list,
 494         * and then atomically returning that entire list back to the
 495         * cache's xfer list as long as it's still empty.
 496         */
 497        do {
 498                old = xchg(&cache->xfer, NULL);
 499                if (old)
 500                        list_splice_entire_tail(old, chpfirst);
 501                old = cmpxchg(&cache->xfer, NULL, chpfirst);
 502        } while (old);
 503
 504
 505        __this_cpu_write(cache->percpu->first, NULL);
 506        __this_cpu_write(cache->percpu->count, 0);
 507end:
 508        local_irq_restore(flags);
 509}
 510
 511static struct list_head *rds_ib_recv_cache_get(struct rds_ib_refill_cache *cache)
 512{
 513        struct list_head *head = cache->ready;
 514
 515        if (head) {
 516                if (!list_empty(head)) {
 517                        cache->ready = head->next;
 518                        list_del_init(head);
 519                } else
 520                        cache->ready = NULL;
 521        }
 522
 523        return head;
 524}
 525
 526int rds_ib_inc_copy_to_user(struct rds_incoming *inc, struct iov_iter *to)
 527{
 528        struct rds_ib_incoming *ibinc;
 529        struct rds_page_frag *frag;
 530        unsigned long to_copy;
 531        unsigned long frag_off = 0;
 532        int copied = 0;
 533        int ret;
 534        u32 len;
 535
 536        ibinc = container_of(inc, struct rds_ib_incoming, ii_inc);
 537        frag = list_entry(ibinc->ii_frags.next, struct rds_page_frag, f_item);
 538        len = be32_to_cpu(inc->i_hdr.h_len);
 539
 540        while (iov_iter_count(to) && copied < len) {
 541                if (frag_off == RDS_FRAG_SIZE) {
 542                        frag = list_entry(frag->f_item.next,
 543                                          struct rds_page_frag, f_item);
 544                        frag_off = 0;
 545                }
 546                to_copy = min_t(unsigned long, iov_iter_count(to),
 547                                RDS_FRAG_SIZE - frag_off);
 548                to_copy = min_t(unsigned long, to_copy, len - copied);
 549
 550                /* XXX needs + offset for multiple recvs per page */
 551                rds_stats_add(s_copy_to_user, to_copy);
 552                ret = copy_page_to_iter(sg_page(&frag->f_sg),
 553                                        frag->f_sg.offset + frag_off,
 554                                        to_copy,
 555                                        to);
 556                if (ret != to_copy)
 557                        return -EFAULT;
 558
 559                frag_off += to_copy;
 560                copied += to_copy;
 561        }
 562
 563        return copied;
 564}
 565
 566/* ic starts out kzalloc()ed */
 567void rds_ib_recv_init_ack(struct rds_ib_connection *ic)
 568{
 569        struct ib_send_wr *wr = &ic->i_ack_wr;
 570        struct ib_sge *sge = &ic->i_ack_sge;
 571
 572        sge->addr = ic->i_ack_dma;
 573        sge->length = sizeof(struct rds_header);
 574        sge->lkey = ic->i_pd->local_dma_lkey;
 575
 576        wr->sg_list = sge;
 577        wr->num_sge = 1;
 578        wr->opcode = IB_WR_SEND;
 579        wr->wr_id = RDS_IB_ACK_WR_ID;
 580        wr->send_flags = IB_SEND_SIGNALED | IB_SEND_SOLICITED;
 581}
 582
 583/*
 584 * You'd think that with reliable IB connections you wouldn't need to ack
 585 * messages that have been received.  The problem is that IB hardware generates
 586 * an ack message before it has DMAed the message into memory.  This creates a
 587 * potential message loss if the HCA is disabled for any reason between when it
 588 * sends the ack and before the message is DMAed and processed.  This is only a
 589 * potential issue if another HCA is available for fail-over.
 590 *
 591 * When the remote host receives our ack they'll free the sent message from
 592 * their send queue.  To decrease the latency of this we always send an ack
 593 * immediately after we've received messages.
 594 *
 595 * For simplicity, we only have one ack in flight at a time.  This puts
 596 * pressure on senders to have deep enough send queues to absorb the latency of
 597 * a single ack frame being in flight.  This might not be good enough.
 598 *
 599 * This is implemented by have a long-lived send_wr and sge which point to a
 600 * statically allocated ack frame.  This ack wr does not fall under the ring
 601 * accounting that the tx and rx wrs do.  The QP attribute specifically makes
 602 * room for it beyond the ring size.  Send completion notices its special
 603 * wr_id and avoids working with the ring in that case.
 604 */
 605#ifndef KERNEL_HAS_ATOMIC64
 606void rds_ib_set_ack(struct rds_ib_connection *ic, u64 seq, int ack_required)
 607{
 608        unsigned long flags;
 609
 610        spin_lock_irqsave(&ic->i_ack_lock, flags);
 611        ic->i_ack_next = seq;
 612        if (ack_required)
 613                set_bit(IB_ACK_REQUESTED, &ic->i_ack_flags);
 614        spin_unlock_irqrestore(&ic->i_ack_lock, flags);
 615}
 616
 617static u64 rds_ib_get_ack(struct rds_ib_connection *ic)
 618{
 619        unsigned long flags;
 620        u64 seq;
 621
 622        clear_bit(IB_ACK_REQUESTED, &ic->i_ack_flags);
 623
 624        spin_lock_irqsave(&ic->i_ack_lock, flags);
 625        seq = ic->i_ack_next;
 626        spin_unlock_irqrestore(&ic->i_ack_lock, flags);
 627
 628        return seq;
 629}
 630#else
 631void rds_ib_set_ack(struct rds_ib_connection *ic, u64 seq, int ack_required)
 632{
 633        atomic64_set(&ic->i_ack_next, seq);
 634        if (ack_required) {
 635                smp_mb__before_atomic();
 636                set_bit(IB_ACK_REQUESTED, &ic->i_ack_flags);
 637        }
 638}
 639
 640static u64 rds_ib_get_ack(struct rds_ib_connection *ic)
 641{
 642        clear_bit(IB_ACK_REQUESTED, &ic->i_ack_flags);
 643        smp_mb__after_atomic();
 644
 645        return atomic64_read(&ic->i_ack_next);
 646}
 647#endif
 648
 649
 650static void rds_ib_send_ack(struct rds_ib_connection *ic, unsigned int adv_credits)
 651{
 652        struct rds_header *hdr = ic->i_ack;
 653        struct ib_send_wr *failed_wr;
 654        u64 seq;
 655        int ret;
 656
 657        seq = rds_ib_get_ack(ic);
 658
 659        rdsdebug("send_ack: ic %p ack %llu\n", ic, (unsigned long long) seq);
 660        rds_message_populate_header(hdr, 0, 0, 0);
 661        hdr->h_ack = cpu_to_be64(seq);
 662        hdr->h_credit = adv_credits;
 663        rds_message_make_checksum(hdr);
 664        ic->i_ack_queued = jiffies;
 665
 666        ret = ib_post_send(ic->i_cm_id->qp, &ic->i_ack_wr, &failed_wr);
 667        if (unlikely(ret)) {
 668                /* Failed to send. Release the WR, and
 669                 * force another ACK.
 670                 */
 671                clear_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags);
 672                set_bit(IB_ACK_REQUESTED, &ic->i_ack_flags);
 673
 674                rds_ib_stats_inc(s_ib_ack_send_failure);
 675
 676                rds_ib_conn_error(ic->conn, "sending ack failed\n");
 677        } else
 678                rds_ib_stats_inc(s_ib_ack_sent);
 679}
 680
 681/*
 682 * There are 3 ways of getting acknowledgements to the peer:
 683 *  1.  We call rds_ib_attempt_ack from the recv completion handler
 684 *      to send an ACK-only frame.
 685 *      However, there can be only one such frame in the send queue
 686 *      at any time, so we may have to postpone it.
 687 *  2.  When another (data) packet is transmitted while there's
 688 *      an ACK in the queue, we piggyback the ACK sequence number
 689 *      on the data packet.
 690 *  3.  If the ACK WR is done sending, we get called from the
 691 *      send queue completion handler, and check whether there's
 692 *      another ACK pending (postponed because the WR was on the
 693 *      queue). If so, we transmit it.
 694 *
 695 * We maintain 2 variables:
 696 *  -   i_ack_flags, which keeps track of whether the ACK WR
 697 *      is currently in the send queue or not (IB_ACK_IN_FLIGHT)
 698 *  -   i_ack_next, which is the last sequence number we received
 699 *
 700 * Potentially, send queue and receive queue handlers can run concurrently.
 701 * It would be nice to not have to use a spinlock to synchronize things,
 702 * but the one problem that rules this out is that 64bit updates are
 703 * not atomic on all platforms. Things would be a lot simpler if
 704 * we had atomic64 or maybe cmpxchg64 everywhere.
 705 *
 706 * Reconnecting complicates this picture just slightly. When we
 707 * reconnect, we may be seeing duplicate packets. The peer
 708 * is retransmitting them, because it hasn't seen an ACK for
 709 * them. It is important that we ACK these.
 710 *
 711 * ACK mitigation adds a header flag "ACK_REQUIRED"; any packet with
 712 * this flag set *MUST* be acknowledged immediately.
 713 */
 714
 715/*
 716 * When we get here, we're called from the recv queue handler.
 717 * Check whether we ought to transmit an ACK.
 718 */
 719void rds_ib_attempt_ack(struct rds_ib_connection *ic)
 720{
 721        unsigned int adv_credits;
 722
 723        if (!test_bit(IB_ACK_REQUESTED, &ic->i_ack_flags))
 724                return;
 725
 726        if (test_and_set_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags)) {
 727                rds_ib_stats_inc(s_ib_ack_send_delayed);
 728                return;
 729        }
 730
 731        /* Can we get a send credit? */
 732        if (!rds_ib_send_grab_credits(ic, 1, &adv_credits, 0, RDS_MAX_ADV_CREDIT)) {
 733                rds_ib_stats_inc(s_ib_tx_throttle);
 734                clear_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags);
 735                return;
 736        }
 737
 738        clear_bit(IB_ACK_REQUESTED, &ic->i_ack_flags);
 739        rds_ib_send_ack(ic, adv_credits);
 740}
 741
 742/*
 743 * We get here from the send completion handler, when the
 744 * adapter tells us the ACK frame was sent.
 745 */
 746void rds_ib_ack_send_complete(struct rds_ib_connection *ic)
 747{
 748        clear_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags);
 749        rds_ib_attempt_ack(ic);
 750}
 751
 752/*
 753 * This is called by the regular xmit code when it wants to piggyback
 754 * an ACK on an outgoing frame.
 755 */
 756u64 rds_ib_piggyb_ack(struct rds_ib_connection *ic)
 757{
 758        if (test_and_clear_bit(IB_ACK_REQUESTED, &ic->i_ack_flags))
 759                rds_ib_stats_inc(s_ib_ack_send_piggybacked);
 760        return rds_ib_get_ack(ic);
 761}
 762
 763/*
 764 * It's kind of lame that we're copying from the posted receive pages into
 765 * long-lived bitmaps.  We could have posted the bitmaps and rdma written into
 766 * them.  But receiving new congestion bitmaps should be a *rare* event, so
 767 * hopefully we won't need to invest that complexity in making it more
 768 * efficient.  By copying we can share a simpler core with TCP which has to
 769 * copy.
 770 */
 771static void rds_ib_cong_recv(struct rds_connection *conn,
 772                              struct rds_ib_incoming *ibinc)
 773{
 774        struct rds_cong_map *map;
 775        unsigned int map_off;
 776        unsigned int map_page;
 777        struct rds_page_frag *frag;
 778        unsigned long frag_off;
 779        unsigned long to_copy;
 780        unsigned long copied;
 781        uint64_t uncongested = 0;
 782        void *addr;
 783
 784        /* catch completely corrupt packets */
 785        if (be32_to_cpu(ibinc->ii_inc.i_hdr.h_len) != RDS_CONG_MAP_BYTES)
 786                return;
 787
 788        map = conn->c_fcong;
 789        map_page = 0;
 790        map_off = 0;
 791
 792        frag = list_entry(ibinc->ii_frags.next, struct rds_page_frag, f_item);
 793        frag_off = 0;
 794
 795        copied = 0;
 796
 797        while (copied < RDS_CONG_MAP_BYTES) {
 798                uint64_t *src, *dst;
 799                unsigned int k;
 800
 801                to_copy = min(RDS_FRAG_SIZE - frag_off, PAGE_SIZE - map_off);
 802                BUG_ON(to_copy & 7); /* Must be 64bit aligned. */
 803
 804                addr = kmap_atomic(sg_page(&frag->f_sg));
 805
 806                src = addr + frag->f_sg.offset + frag_off;
 807                dst = (void *)map->m_page_addrs[map_page] + map_off;
 808                for (k = 0; k < to_copy; k += 8) {
 809                        /* Record ports that became uncongested, ie
 810                         * bits that changed from 0 to 1. */
 811                        uncongested |= ~(*src) & *dst;
 812                        *dst++ = *src++;
 813                }
 814                kunmap_atomic(addr);
 815
 816                copied += to_copy;
 817
 818                map_off += to_copy;
 819                if (map_off == PAGE_SIZE) {
 820                        map_off = 0;
 821                        map_page++;
 822                }
 823
 824                frag_off += to_copy;
 825                if (frag_off == RDS_FRAG_SIZE) {
 826                        frag = list_entry(frag->f_item.next,
 827                                          struct rds_page_frag, f_item);
 828                        frag_off = 0;
 829                }
 830        }
 831
 832        /* the congestion map is in little endian order */
 833        uncongested = le64_to_cpu(uncongested);
 834
 835        rds_cong_map_updated(map, uncongested);
 836}
 837
 838static void rds_ib_process_recv(struct rds_connection *conn,
 839                                struct rds_ib_recv_work *recv, u32 data_len,
 840                                struct rds_ib_ack_state *state)
 841{
 842        struct rds_ib_connection *ic = conn->c_transport_data;
 843        struct rds_ib_incoming *ibinc = ic->i_ibinc;
 844        struct rds_header *ihdr, *hdr;
 845
 846        /* XXX shut down the connection if port 0,0 are seen? */
 847
 848        rdsdebug("ic %p ibinc %p recv %p byte len %u\n", ic, ibinc, recv,
 849                 data_len);
 850
 851        if (data_len < sizeof(struct rds_header)) {
 852                rds_ib_conn_error(conn, "incoming message "
 853                       "from %pI4 didn't include a "
 854                       "header, disconnecting and "
 855                       "reconnecting\n",
 856                       &conn->c_faddr);
 857                return;
 858        }
 859        data_len -= sizeof(struct rds_header);
 860
 861        ihdr = &ic->i_recv_hdrs[recv - ic->i_recvs];
 862
 863        /* Validate the checksum. */
 864        if (!rds_message_verify_checksum(ihdr)) {
 865                rds_ib_conn_error(conn, "incoming message "
 866                       "from %pI4 has corrupted header - "
 867                       "forcing a reconnect\n",
 868                       &conn->c_faddr);
 869                rds_stats_inc(s_recv_drop_bad_checksum);
 870                return;
 871        }
 872
 873        /* Process the ACK sequence which comes with every packet */
 874        state->ack_recv = be64_to_cpu(ihdr->h_ack);
 875        state->ack_recv_valid = 1;
 876
 877        /* Process the credits update if there was one */
 878        if (ihdr->h_credit)
 879                rds_ib_send_add_credits(conn, ihdr->h_credit);
 880
 881        if (ihdr->h_sport == 0 && ihdr->h_dport == 0 && data_len == 0) {
 882                /* This is an ACK-only packet. The fact that it gets
 883                 * special treatment here is that historically, ACKs
 884                 * were rather special beasts.
 885                 */
 886                rds_ib_stats_inc(s_ib_ack_received);
 887
 888                /*
 889                 * Usually the frags make their way on to incs and are then freed as
 890                 * the inc is freed.  We don't go that route, so we have to drop the
 891                 * page ref ourselves.  We can't just leave the page on the recv
 892                 * because that confuses the dma mapping of pages and each recv's use
 893                 * of a partial page.
 894                 *
 895                 * FIXME: Fold this into the code path below.
 896                 */
 897                rds_ib_frag_free(ic, recv->r_frag);
 898                recv->r_frag = NULL;
 899                return;
 900        }
 901
 902        /*
 903         * If we don't already have an inc on the connection then this
 904         * fragment has a header and starts a message.. copy its header
 905         * into the inc and save the inc so we can hang upcoming fragments
 906         * off its list.
 907         */
 908        if (!ibinc) {
 909                ibinc = recv->r_ibinc;
 910                recv->r_ibinc = NULL;
 911                ic->i_ibinc = ibinc;
 912
 913                hdr = &ibinc->ii_inc.i_hdr;
 914                ibinc->ii_inc.i_rx_lat_trace[RDS_MSG_RX_HDR] =
 915                                local_clock();
 916                memcpy(hdr, ihdr, sizeof(*hdr));
 917                ic->i_recv_data_rem = be32_to_cpu(hdr->h_len);
 918                ibinc->ii_inc.i_rx_lat_trace[RDS_MSG_RX_START] =
 919                                local_clock();
 920
 921                rdsdebug("ic %p ibinc %p rem %u flag 0x%x\n", ic, ibinc,
 922                         ic->i_recv_data_rem, hdr->h_flags);
 923        } else {
 924                hdr = &ibinc->ii_inc.i_hdr;
 925                /* We can't just use memcmp here; fragments of a
 926                 * single message may carry different ACKs */
 927                if (hdr->h_sequence != ihdr->h_sequence ||
 928                    hdr->h_len != ihdr->h_len ||
 929                    hdr->h_sport != ihdr->h_sport ||
 930                    hdr->h_dport != ihdr->h_dport) {
 931                        rds_ib_conn_error(conn,
 932                                "fragment header mismatch; forcing reconnect\n");
 933                        return;
 934                }
 935        }
 936
 937        list_add_tail(&recv->r_frag->f_item, &ibinc->ii_frags);
 938        recv->r_frag = NULL;
 939
 940        if (ic->i_recv_data_rem > RDS_FRAG_SIZE)
 941                ic->i_recv_data_rem -= RDS_FRAG_SIZE;
 942        else {
 943                ic->i_recv_data_rem = 0;
 944                ic->i_ibinc = NULL;
 945
 946                if (ibinc->ii_inc.i_hdr.h_flags == RDS_FLAG_CONG_BITMAP)
 947                        rds_ib_cong_recv(conn, ibinc);
 948                else {
 949                        rds_recv_incoming(conn, conn->c_faddr, conn->c_laddr,
 950                                          &ibinc->ii_inc, GFP_ATOMIC);
 951                        state->ack_next = be64_to_cpu(hdr->h_sequence);
 952                        state->ack_next_valid = 1;
 953                }
 954
 955                /* Evaluate the ACK_REQUIRED flag *after* we received
 956                 * the complete frame, and after bumping the next_rx
 957                 * sequence. */
 958                if (hdr->h_flags & RDS_FLAG_ACK_REQUIRED) {
 959                        rds_stats_inc(s_recv_ack_required);
 960                        state->ack_required = 1;
 961                }
 962
 963                rds_inc_put(&ibinc->ii_inc);
 964        }
 965}
 966
 967void rds_ib_recv_cqe_handler(struct rds_ib_connection *ic,
 968                             struct ib_wc *wc,
 969                             struct rds_ib_ack_state *state)
 970{
 971        struct rds_connection *conn = ic->conn;
 972        struct rds_ib_recv_work *recv;
 973
 974        rdsdebug("wc wr_id 0x%llx status %u (%s) byte_len %u imm_data %u\n",
 975                 (unsigned long long)wc->wr_id, wc->status,
 976                 ib_wc_status_msg(wc->status), wc->byte_len,
 977                 be32_to_cpu(wc->ex.imm_data));
 978
 979        rds_ib_stats_inc(s_ib_rx_cq_event);
 980        recv = &ic->i_recvs[rds_ib_ring_oldest(&ic->i_recv_ring)];
 981        ib_dma_unmap_sg(ic->i_cm_id->device, &recv->r_frag->f_sg, 1,
 982                        DMA_FROM_DEVICE);
 983
 984        /* Also process recvs in connecting state because it is possible
 985         * to get a recv completion _before_ the rdmacm ESTABLISHED
 986         * event is processed.
 987         */
 988        if (wc->status == IB_WC_SUCCESS) {
 989                rds_ib_process_recv(conn, recv, wc->byte_len, state);
 990        } else {
 991                /* We expect errors as the qp is drained during shutdown */
 992                if (rds_conn_up(conn) || rds_conn_connecting(conn))
 993                        rds_ib_conn_error(conn, "recv completion on <%pI4,%pI4> had status %u (%s), disconnecting and reconnecting\n",
 994                                          &conn->c_laddr, &conn->c_faddr,
 995                                          wc->status,
 996                                          ib_wc_status_msg(wc->status));
 997        }
 998
 999        /* rds_ib_process_recv() doesn't always consume the frag, and
1000         * we might not have called it at all if the wc didn't indicate
1001         * success. We already unmapped the frag's pages, though, and
1002         * the following rds_ib_ring_free() call tells the refill path
1003         * that it will not find an allocated frag here. Make sure we
1004         * keep that promise by freeing a frag that's still on the ring.
1005         */
1006        if (recv->r_frag) {
1007                rds_ib_frag_free(ic, recv->r_frag);
1008                recv->r_frag = NULL;
1009        }
1010        rds_ib_ring_free(&ic->i_recv_ring, 1);
1011
1012        /* If we ever end up with a really empty receive ring, we're
1013         * in deep trouble, as the sender will definitely see RNR
1014         * timeouts. */
1015        if (rds_ib_ring_empty(&ic->i_recv_ring))
1016                rds_ib_stats_inc(s_ib_rx_ring_empty);
1017
1018        if (rds_ib_ring_low(&ic->i_recv_ring)) {
1019                rds_ib_recv_refill(conn, 0, GFP_NOWAIT);
1020                rds_ib_stats_inc(s_ib_rx_refill_from_cq);
1021        }
1022}
1023
1024int rds_ib_recv_path(struct rds_conn_path *cp)
1025{
1026        struct rds_connection *conn = cp->cp_conn;
1027        struct rds_ib_connection *ic = conn->c_transport_data;
1028        int ret = 0;
1029
1030        rdsdebug("conn %p\n", conn);
1031        if (rds_conn_up(conn)) {
1032                rds_ib_attempt_ack(ic);
1033                rds_ib_recv_refill(conn, 0, GFP_KERNEL);
1034                rds_ib_stats_inc(s_ib_rx_refill_from_thread);
1035        }
1036
1037        return ret;
1038}
1039
1040int rds_ib_recv_init(void)
1041{
1042        struct sysinfo si;
1043        int ret = -ENOMEM;
1044
1045        /* Default to 30% of all available RAM for recv memory */
1046        si_meminfo(&si);
1047        rds_ib_sysctl_max_recv_allocation = si.totalram / 3 * PAGE_SIZE / RDS_FRAG_SIZE;
1048
1049        rds_ib_incoming_slab = kmem_cache_create("rds_ib_incoming",
1050                                        sizeof(struct rds_ib_incoming),
1051                                        0, SLAB_HWCACHE_ALIGN, NULL);
1052        if (!rds_ib_incoming_slab)
1053                goto out;
1054
1055        rds_ib_frag_slab = kmem_cache_create("rds_ib_frag",
1056                                        sizeof(struct rds_page_frag),
1057                                        0, SLAB_HWCACHE_ALIGN, NULL);
1058        if (!rds_ib_frag_slab) {
1059                kmem_cache_destroy(rds_ib_incoming_slab);
1060                rds_ib_incoming_slab = NULL;
1061        } else
1062                ret = 0;
1063out:
1064        return ret;
1065}
1066
1067void rds_ib_recv_exit(void)
1068{
1069        kmem_cache_destroy(rds_ib_incoming_slab);
1070        kmem_cache_destroy(rds_ib_frag_slab);
1071}
1072