net/rds/ib_rdma.c

   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/rculist.h>
  36
  37 #include "rds.h"
  38 #include "ib.h"
  39 #include "xlist.h"
  40
  41 static struct workqueue_struct *rds_ib_fmr_wq;
  42
  43 static DEFINE_PER_CPU(unsigned long, clean_list_grace);
  44 #define CLEAN_LIST_BUSY_BIT 0
  45
  46 /*
  47  * This is stored as mr->r_trans_private.
  48  */
  49 struct rds_ib_mr {
  50         struct rds_ib_device    *device;
  51         struct rds_ib_mr_pool   *pool;
  52         struct ib_fmr           *fmr;
  53
  54         struct xlist_head       xlist;
  55
  56         /* unmap_list is for freeing */
  57         struct list_head        unmap_list;
  58         unsigned int            remap_count;
  59
  60         struct scatterlist      *sg;
  61         unsigned int            sg_len;
  62         u64                     *dma;
  63         int                     sg_dma_len;
  64 };
  65
  66 /*
  67  * Our own little FMR pool
  68  */
  69 struct rds_ib_mr_pool {
  70         struct mutex            flush_lock;             /* serialize fmr invalidate */
  71         struct delayed_work     flush_worker;           /* flush worker */
  72
  73         atomic_t                item_count;             /* total # of MRs */
  74         atomic_t                dirty_count;            /* # dirty of MRs */
  75
  76         struct xlist_head       drop_list;              /* MRs that have reached their max_maps limit */
  77         struct xlist_head       free_list;              /* unused MRs */
  78         struct xlist_head       clean_list;             /* global unused & unamapped MRs */
  79         wait_queue_head_t       flush_wait;
  80
  81         atomic_t                free_pinned;            /* memory pinned by free MRs */
  82         unsigned long           max_items;
  83         unsigned long           max_items_soft;
  84         unsigned long           max_free_pinned;
  85         struct ib_fmr_attr      fmr_attr;
  86 };
  87
  88 static int rds_ib_flush_mr_pool(struct rds_ib_mr_pool *pool, int free_all, struct rds_ib_mr **);
  89 static void rds_ib_teardown_mr(struct rds_ib_mr *ibmr);
  90 static void rds_ib_mr_pool_flush_worker(struct work_struct *work);
  91
  92 static struct rds_ib_device *rds_ib_get_device(__be32 ipaddr)
  93 {
  94         struct rds_ib_device *rds_ibdev;
  95         struct rds_ib_ipaddr *i_ipaddr;
  96
  97         rcu_read_lock();
  98         list_for_each_entry_rcu(rds_ibdev, &rds_ib_devices, list) {
  99                 list_for_each_entry_rcu(i_ipaddr, &rds_ibdev->ipaddr_list, list) {
 100                         if (i_ipaddr->ipaddr == ipaddr) {
 101                                 atomic_inc(&rds_ibdev->refcount);
 102                                 rcu_read_unlock();
 103                                 return rds_ibdev;
 104                         }
 105                 }
 106         }
 107         rcu_read_unlock();
 108
 109         return NULL;
 110 }
 111
 112 static int rds_ib_add_ipaddr(struct rds_ib_device *rds_ibdev, __be32 ipaddr)
 113 {
 114         struct rds_ib_ipaddr *i_ipaddr;
 115
 116         i_ipaddr = kmalloc(sizeof *i_ipaddr, GFP_KERNEL);
 117         if (!i_ipaddr)
 118                 return -ENOMEM;
 119
 120         i_ipaddr->ipaddr = ipaddr;
 121
 122         spin_lock_irq(&rds_ibdev->spinlock);
 123         list_add_tail_rcu(&i_ipaddr->list, &rds_ibdev->ipaddr_list);
 124         spin_unlock_irq(&rds_ibdev->spinlock);
 125
 126         return 0;
 127 }
 128
 129 static void rds_ib_remove_ipaddr(struct rds_ib_device *rds_ibdev, __be32 ipaddr)
 130 {
 131         struct rds_ib_ipaddr *i_ipaddr;
 132         struct rds_ib_ipaddr *to_free = NULL;
 133
 134
 135         spin_lock_irq(&rds_ibdev->spinlock);
 136         list_for_each_entry_rcu(i_ipaddr, &rds_ibdev->ipaddr_list, list) {
 137                 if (i_ipaddr->ipaddr == ipaddr) {
 138                         list_del_rcu(&i_ipaddr->list);
 139                         to_free = i_ipaddr;
 140                         break;
 141                 }
 142         }
 143         spin_unlock_irq(&rds_ibdev->spinlock);
 144
 145         if (to_free) {
 146                 synchronize_rcu();
 147                 kfree(to_free);
 148         }
 149 }
 150
 151 int rds_ib_update_ipaddr(struct rds_ib_device *rds_ibdev, __be32 ipaddr)
 152 {
 153         struct rds_ib_device *rds_ibdev_old;
 154
 155         rds_ibdev_old = rds_ib_get_device(ipaddr);
 156         if (rds_ibdev_old) {
 157                 rds_ib_remove_ipaddr(rds_ibdev_old, ipaddr);
 158                 rds_ib_dev_put(rds_ibdev_old);
 159         }
 160
 161         return rds_ib_add_ipaddr(rds_ibdev, ipaddr);
 162 }
 163
 164 void rds_ib_add_conn(struct rds_ib_device *rds_ibdev, struct rds_connection *conn)
 165 {
 166         struct rds_ib_connection *ic = conn->c_transport_data;
 167
 168         /* conn was previously on the nodev_conns_list */
 169         spin_lock_irq(&ib_nodev_conns_lock);
 170         BUG_ON(list_empty(&ib_nodev_conns));
 171         BUG_ON(list_empty(&ic->ib_node));
 172         list_del(&ic->ib_node);
 173
 174         spin_lock(&rds_ibdev->spinlock);
 175         list_add_tail(&ic->ib_node, &rds_ibdev->conn_list);
 176         spin_unlock(&rds_ibdev->spinlock);
 177         spin_unlock_irq(&ib_nodev_conns_lock);
 178
 179         ic->rds_ibdev = rds_ibdev;
 180         atomic_inc(&rds_ibdev->refcount);
 181 }
 182
 183 void rds_ib_remove_conn(struct rds_ib_device *rds_ibdev, struct rds_connection *conn)
 184 {
 185         struct rds_ib_connection *ic = conn->c_transport_data;
 186
 187         /* place conn on nodev_conns_list */
 188         spin_lock(&ib_nodev_conns_lock);
 189
 190         spin_lock_irq(&rds_ibdev->spinlock);
 191         BUG_ON(list_empty(&ic->ib_node));
 192         list_del(&ic->ib_node);
 193         spin_unlock_irq(&rds_ibdev->spinlock);
 194
 195         list_add_tail(&ic->ib_node, &ib_nodev_conns);
 196
 197         spin_unlock(&ib_nodev_conns_lock);
 198
 199         ic->rds_ibdev = NULL;
 200         rds_ib_dev_put(rds_ibdev);
 201 }
 202
 203 void rds_ib_destroy_nodev_conns(void)
 204 {
 205         struct rds_ib_connection *ic, *_ic;
 206         LIST_HEAD(tmp_list);
 207
 208         /* avoid calling conn_destroy with irqs off */
 209         spin_lock_irq(&ib_nodev_conns_lock);
 210         list_splice(&ib_nodev_conns, &tmp_list);
 211         spin_unlock_irq(&ib_nodev_conns_lock);
 212
 213         list_for_each_entry_safe(ic, _ic, &tmp_list, ib_node)
 214                 rds_conn_destroy(ic->conn);
 215 }
 216
 217 struct rds_ib_mr_pool *rds_ib_create_mr_pool(struct rds_ib_device *rds_ibdev)
 218 {
 219         struct rds_ib_mr_pool *pool;
 220
 221         pool = kzalloc(sizeof(*pool), GFP_KERNEL);
 222         if (!pool)
 223                 return ERR_PTR(-ENOMEM);
 224
 225         INIT_XLIST_HEAD(&pool->free_list);
 226         INIT_XLIST_HEAD(&pool->drop_list);
 227         INIT_XLIST_HEAD(&pool->clean_list);
 228         mutex_init(&pool->flush_lock);
 229         init_waitqueue_head(&pool->flush_wait);
 230         INIT_DELAYED_WORK(&pool->flush_worker, rds_ib_mr_pool_flush_worker);
 231
 232         pool->fmr_attr.max_pages = fmr_message_size;
 233         pool->fmr_attr.max_maps = rds_ibdev->fmr_max_remaps;
 234         pool->fmr_attr.page_shift = PAGE_SHIFT;
 235         pool->max_free_pinned = rds_ibdev->max_fmrs * fmr_message_size / 4;
 236
 237         /* We never allow more than max_items MRs to be allocated.
 238          * When we exceed more than max_items_soft, we start freeing
 239          * items more aggressively.
 240          * Make sure that max_items > max_items_soft > max_items / 2
 241          */
 242         pool->max_items_soft = rds_ibdev->max_fmrs * 3 / 4;
 243         pool->max_items = rds_ibdev->max_fmrs;
 244
 245         return pool;
 246 }
 247
 248 void rds_ib_get_mr_info(struct rds_ib_device *rds_ibdev, struct rds_info_rdma_connection *iinfo)
 249 {
 250         struct rds_ib_mr_pool *pool = rds_ibdev->mr_pool;
 251
 252         iinfo->rdma_mr_max = pool->max_items;
 253         iinfo->rdma_mr_size = pool->fmr_attr.max_pages;
 254 }
 255
 256 void rds_ib_destroy_mr_pool(struct rds_ib_mr_pool *pool)
 257 {
 258         cancel_delayed_work_sync(&pool->flush_worker);
 259         rds_ib_flush_mr_pool(pool, 1, NULL);
 260         WARN_ON(atomic_read(&pool->item_count));
 261         WARN_ON(atomic_read(&pool->free_pinned));
 262         kfree(pool);
 263 }
 264
 265 static void refill_local(struct rds_ib_mr_pool *pool, struct xlist_head *xl,
 266                          struct rds_ib_mr **ibmr_ret)
 267 {
 268         struct xlist_head *ibmr_xl;
 269         ibmr_xl = xlist_del_head_fast(xl);
 270         *ibmr_ret = list_entry(ibmr_xl, struct rds_ib_mr, xlist);
 271 }
 272
 273 static inline struct rds_ib_mr *rds_ib_reuse_fmr(struct rds_ib_mr_pool *pool)
 274 {
 275         struct rds_ib_mr *ibmr = NULL;
 276         struct xlist_head *ret;
 277         unsigned long *flag;
 278
 279         preempt_disable();
 280         flag = &__get_cpu_var(clean_list_grace);
 281         set_bit(CLEAN_LIST_BUSY_BIT, flag);
 282         ret = xlist_del_head(&pool->clean_list);
 283         if (ret)
 284                 ibmr = list_entry(ret, struct rds_ib_mr, xlist);
 285
 286         clear_bit(CLEAN_LIST_BUSY_BIT, flag);
 287         preempt_enable();
 288         return ibmr;
 289 }
 290
 291 static inline void wait_clean_list_grace(void)
 292 {
 293         int cpu;
 294         unsigned long *flag;
 295
 296         for_each_online_cpu(cpu) {
 297                 flag = &per_cpu(clean_list_grace, cpu);
 298                 while (test_bit(CLEAN_LIST_BUSY_BIT, flag))
 299                         cpu_relax();
 300         }
 301 }
 302
 303 static struct rds_ib_mr *rds_ib_alloc_fmr(struct rds_ib_device *rds_ibdev)
 304 {
 305         struct rds_ib_mr_pool *pool = rds_ibdev->mr_pool;
 306         struct rds_ib_mr *ibmr = NULL;
 307         int err = 0, iter = 0;
 308
 309         if (atomic_read(&pool->dirty_count) >= pool->max_items / 10)
 310                 queue_delayed_work(rds_ib_fmr_wq, &pool->flush_worker, 10);
 311
 312         while (1) {
 313                 ibmr = rds_ib_reuse_fmr(pool);
 314                 if (ibmr)
 315                         return ibmr;
 316
 317                 /* No clean MRs - now we have the choice of either
 318                  * allocating a fresh MR up to the limit imposed by the
 319                  * driver, or flush any dirty unused MRs.
 320                  * We try to avoid stalling in the send path if possible,
 321                  * so we allocate as long as we're allowed to.
 322                  *
 323                  * We're fussy with enforcing the FMR limit, though. If the driver
 324                  * tells us we can't use more than N fmrs, we shouldn't start
 325                  * arguing with it */
 326                 if (atomic_inc_return(&pool->item_count) <= pool->max_items)
 327                         break;
 328
 329                 atomic_dec(&pool->item_count);
 330
 331                 if (++iter > 2) {
 332                         rds_ib_stats_inc(s_ib_rdma_mr_pool_depleted);
 333                         return ERR_PTR(-EAGAIN);
 334                 }
 335
 336                 /* We do have some empty MRs. Flush them out. */
 337                 rds_ib_stats_inc(s_ib_rdma_mr_pool_wait);
 338                 rds_ib_flush_mr_pool(pool, 0, &ibmr);
 339                 if (ibmr)
 340                         return ibmr;
 341         }
 342
 343         ibmr = kzalloc_node(sizeof(*ibmr), GFP_KERNEL, rdsibdev_to_node(rds_ibdev));
 344         if (!ibmr) {
 345                 err = -ENOMEM;
 346                 goto out_no_cigar;
 347         }
 348
 349         memset(ibmr, 0, sizeof(*ibmr));
 350
 351         ibmr->fmr = ib_alloc_fmr(rds_ibdev->pd,
 352                         (IB_ACCESS_LOCAL_WRITE |
 353                          IB_ACCESS_REMOTE_READ |
 354                          IB_ACCESS_REMOTE_WRITE|
 355                          IB_ACCESS_REMOTE_ATOMIC),
 356                         &pool->fmr_attr);
 357         if (IS_ERR(ibmr->fmr)) {
 358                 err = PTR_ERR(ibmr->fmr);
 359                 ibmr->fmr = NULL;
 360                 printk(KERN_WARNING "RDS/IB: ib_alloc_fmr failed (err=%d)\n", err);
 361                 goto out_no_cigar;
 362         }
 363
 364         rds_ib_stats_inc(s_ib_rdma_mr_alloc);
 365         return ibmr;
 366
 367 out_no_cigar:
 368         if (ibmr) {
 369                 if (ibmr->fmr)
 370                         ib_dealloc_fmr(ibmr->fmr);
 371                 kfree(ibmr);
 372         }
 373         atomic_dec(&pool->item_count);
 374         return ERR_PTR(err);
 375 }
 376
 377 static int rds_ib_map_fmr(struct rds_ib_device *rds_ibdev, struct rds_ib_mr *ibmr,
 378                struct scatterlist *sg, unsigned int nents)
 379 {
 380         struct ib_device *dev = rds_ibdev->dev;
 381         struct scatterlist *scat = sg;
 382         u64 io_addr = 0;
 383         u64 *dma_pages;
 384         u32 len;
 385         int page_cnt, sg_dma_len;
 386         int i, j;
 387         int ret;
 388
 389         sg_dma_len = ib_dma_map_sg(dev, sg, nents,
 390                                  DMA_BIDIRECTIONAL);
 391         if (unlikely(!sg_dma_len)) {
 392                 printk(KERN_WARNING "RDS/IB: dma_map_sg failed!\n");
 393                 return -EBUSY;
 394         }
 395
 396         len = 0;
 397         page_cnt = 0;
 398
 399         for (i = 0; i < sg_dma_len; ++i) {
 400                 unsigned int dma_len = ib_sg_dma_len(dev, &scat[i]);
 401                 u64 dma_addr = ib_sg_dma_address(dev, &scat[i]);
 402
 403                 if (dma_addr & ~PAGE_MASK) {
 404                         if (i > 0)
 405                                 return -EINVAL;
 406                         else
 407                                 ++page_cnt;
 408                 }
 409                 if ((dma_addr + dma_len) & ~PAGE_MASK) {
 410                         if (i < sg_dma_len - 1)
 411                                 return -EINVAL;
 412                         else
 413                                 ++page_cnt;
 414                 }
 415
 416                 len += dma_len;
 417         }
 418
 419         page_cnt += len >> PAGE_SHIFT;
 420         if (page_cnt > fmr_message_size)
 421                 return -EINVAL;
 422
 423         dma_pages = kmalloc_node(sizeof(u64) * page_cnt, GFP_ATOMIC,
 424                                  rdsibdev_to_node(rds_ibdev));
 425         if (!dma_pages)
 426                 return -ENOMEM;
 427
 428         page_cnt = 0;
 429         for (i = 0; i < sg_dma_len; ++i) {
 430                 unsigned int dma_len = ib_sg_dma_len(dev, &scat[i]);
 431                 u64 dma_addr = ib_sg_dma_address(dev, &scat[i]);
 432
 433                 for (j = 0; j < dma_len; j += PAGE_SIZE)
 434                         dma_pages[page_cnt++] =
 435                                 (dma_addr & PAGE_MASK) + j;
 436         }
 437
 438         ret = ib_map_phys_fmr(ibmr->fmr,
 439                                    dma_pages, page_cnt, io_addr);
 440         if (ret)
 441                 goto out;
 442
 443         /* Success - we successfully remapped the MR, so we can
 444          * safely tear down the old mapping. */
 445         rds_ib_teardown_mr(ibmr);
 446
 447         ibmr->sg = scat;
 448         ibmr->sg_len = nents;
 449         ibmr->sg_dma_len = sg_dma_len;
 450         ibmr->remap_count++;
 451
 452         rds_ib_stats_inc(s_ib_rdma_mr_used);
 453         ret = 0;
 454
 455 out:
 456         kfree(dma_pages);
 457
 458         return ret;
 459 }
 460
 461 void rds_ib_sync_mr(void *trans_private, int direction)
 462 {
 463         struct rds_ib_mr *ibmr = trans_private;
 464         struct rds_ib_device *rds_ibdev = ibmr->device;
 465
 466         switch (direction) {
 467         case DMA_FROM_DEVICE:
 468                 ib_dma_sync_sg_for_cpu(rds_ibdev->dev, ibmr->sg,
 469                         ibmr->sg_dma_len, DMA_BIDIRECTIONAL);
 470                 break;
 471         case DMA_TO_DEVICE:
 472                 ib_dma_sync_sg_for_device(rds_ibdev->dev, ibmr->sg,
 473                         ibmr->sg_dma_len, DMA_BIDIRECTIONAL);
 474                 break;
 475         }
 476 }
 477
 478 static void __rds_ib_teardown_mr(struct rds_ib_mr *ibmr)
 479 {
 480         struct rds_ib_device *rds_ibdev = ibmr->device;
 481
 482         if (ibmr->sg_dma_len) {
 483                 ib_dma_unmap_sg(rds_ibdev->dev,
 484                                 ibmr->sg, ibmr->sg_len,
 485                                 DMA_BIDIRECTIONAL);
 486                 ibmr->sg_dma_len = 0;
 487         }
 488
 489         /* Release the s/g list */
 490         if (ibmr->sg_len) {
 491                 unsigned int i;
 492
 493                 for (i = 0; i < ibmr->sg_len; ++i) {
 494                         struct page *page = sg_page(&ibmr->sg[i]);
 495
 496                         /* FIXME we need a way to tell a r/w MR
 497                          * from a r/o MR */
 498                         BUG_ON(irqs_disabled());
 499                         set_page_dirty(page);
 500                         put_page(page);
 501                 }
 502                 kfree(ibmr->sg);
 503
 504                 ibmr->sg = NULL;
 505                 ibmr->sg_len = 0;
 506         }
 507 }
 508
 509 static void rds_ib_teardown_mr(struct rds_ib_mr *ibmr)
 510 {
 511         unsigned int pinned = ibmr->sg_len;
 512
 513         __rds_ib_teardown_mr(ibmr);
 514         if (pinned) {
 515                 struct rds_ib_device *rds_ibdev = ibmr->device;
 516                 struct rds_ib_mr_pool *pool = rds_ibdev->mr_pool;
 517
 518                 atomic_sub(pinned, &pool->free_pinned);
 519         }
 520 }
 521
 522 static inline unsigned int rds_ib_flush_goal(struct rds_ib_mr_pool *pool, int free_all)
 523 {
 524         unsigned int item_count;
 525
 526         item_count = atomic_read(&pool->item_count);
 527         if (free_all)
 528                 return item_count;
 529
 530         return 0;
 531 }
 532
 533 /*
 534  * given an xlist of mrs, put them all into the list_head for more processing
 535  */
 536 static void xlist_append_to_list(struct xlist_head *xlist, struct list_head *list)
 537 {
 538         struct rds_ib_mr *ibmr;
 539         struct xlist_head splice;
 540         struct xlist_head *cur;
 541         struct xlist_head *next;
 542
 543         splice.next = NULL;
 544         xlist_splice(xlist, &splice);
 545         cur = splice.next;
 546         while (cur) {
 547                 next = cur->next;
 548                 ibmr = list_entry(cur, struct rds_ib_mr, xlist);
 549                 list_add_tail(&ibmr->unmap_list, list);
 550                 cur = next;
 551         }
 552 }
 553
 554 /*
 555  * this takes a list head of mrs and turns it into an xlist of clusters.
 556  * each cluster has an xlist of MR_CLUSTER_SIZE mrs that are ready for
 557  * reuse.
 558  */
 559 static void list_append_to_xlist(struct rds_ib_mr_pool *pool,
 560                                 struct list_head *list, struct xlist_head *xlist,
 561                                 struct xlist_head **tail_ret)
 562 {
 563         struct rds_ib_mr *ibmr;
 564         struct xlist_head *cur_mr = xlist;
 565         struct xlist_head *tail_mr = NULL;
 566
 567         list_for_each_entry(ibmr, list, unmap_list) {
 568                 tail_mr = &ibmr->xlist;
 569                 tail_mr->next = NULL;
 570                 cur_mr->next = tail_mr;
 571                 cur_mr = tail_mr;
 572         }
 573         *tail_ret = tail_mr;
 574 }
 575
 576 /*
 577  * Flush our pool of MRs.
 578  * At a minimum, all currently unused MRs are unmapped.
 579  * If the number of MRs allocated exceeds the limit, we also try
 580  * to free as many MRs as needed to get back to this limit.
 581  */
 582 static int rds_ib_flush_mr_pool(struct rds_ib_mr_pool *pool,
 583                                 int free_all, struct rds_ib_mr **ibmr_ret)
 584 {
 585         struct rds_ib_mr *ibmr, *next;
 586         struct xlist_head clean_xlist;
 587         struct xlist_head *clean_tail;
 588         LIST_HEAD(unmap_list);
 589         LIST_HEAD(fmr_list);
 590         unsigned long unpinned = 0;
 591         unsigned int nfreed = 0, ncleaned = 0, free_goal;
 592         int ret = 0;
 593
 594         rds_ib_stats_inc(s_ib_rdma_mr_pool_flush);
 595
 596         if (ibmr_ret) {
 597                 DEFINE_WAIT(wait);
 598                 while(!mutex_trylock(&pool->flush_lock)) {
 599                         ibmr = rds_ib_reuse_fmr(pool);
 600                         if (ibmr) {
 601                                 *ibmr_ret = ibmr;
 602                                 finish_wait(&pool->flush_wait, &wait);
 603                                 goto out_nolock;
 604                         }
 605
 606                         prepare_to_wait(&pool->flush_wait, &wait,
 607                                         TASK_UNINTERRUPTIBLE);
 608                         if (xlist_empty(&pool->clean_list))
 609                                 schedule();
 610
 611                         ibmr = rds_ib_reuse_fmr(pool);
 612                         if (ibmr) {
 613                                 *ibmr_ret = ibmr;
 614                                 finish_wait(&pool->flush_wait, &wait);
 615                                 goto out_nolock;
 616                         }
 617                 }
 618                 finish_wait(&pool->flush_wait, &wait);
 619         } else
 620                 mutex_lock(&pool->flush_lock);
 621
 622         if (ibmr_ret) {
 623                 ibmr = rds_ib_reuse_fmr(pool);
 624                 if (ibmr) {
 625                         *ibmr_ret = ibmr;
 626                         goto out;
 627                 }
 628         }
 629
 630         /* Get the list of all MRs to be dropped. Ordering matters -
 631          * we want to put drop_list ahead of free_list.
 632          */
 633         xlist_append_to_list(&pool->drop_list, &unmap_list);
 634         xlist_append_to_list(&pool->free_list, &unmap_list);
 635         if (free_all)
 636                 xlist_append_to_list(&pool->clean_list, &unmap_list);
 637
 638         free_goal = rds_ib_flush_goal(pool, free_all);
 639
 640         if (list_empty(&unmap_list))
 641                 goto out;
 642
 643         /* String all ib_mr's onto one list and hand them to ib_unmap_fmr */
 644         list_for_each_entry(ibmr, &unmap_list, unmap_list)
 645                 list_add(&ibmr->fmr->list, &fmr_list);
 646
 647         ret = ib_unmap_fmr(&fmr_list);
 648         if (ret)
 649                 printk(KERN_WARNING "RDS/IB: ib_unmap_fmr failed (err=%d)\n", ret);
 650
 651         /* Now we can destroy the DMA mapping and unpin any pages */
 652         list_for_each_entry_safe(ibmr, next, &unmap_list, unmap_list) {
 653                 unpinned += ibmr->sg_len;
 654                 __rds_ib_teardown_mr(ibmr);
 655                 if (nfreed < free_goal || ibmr->remap_count >= pool->fmr_attr.max_maps) {
 656                         rds_ib_stats_inc(s_ib_rdma_mr_free);
 657                         list_del(&ibmr->unmap_list);
 658                         ib_dealloc_fmr(ibmr->fmr);
 659                         kfree(ibmr);
 660                         nfreed++;
 661                 }
 662                 ncleaned++;
 663         }
 664
 665         if (!list_empty(&unmap_list)) {
 666                 /* we have to make sure that none of the things we're about
 667                  * to put on the clean list would race with other cpus trying
 668                  * to pull items off.  The xlist would explode if we managed to
 669                  * remove something from the clean list and then add it back again
 670                  * while another CPU was spinning on that same item in xlist_del_head.
 671                  *
 672                  * This is pretty unlikely, but just in case  wait for an xlist grace period
 673                  * here before adding anything back into the clean list.
 674                  */
 675                 wait_clean_list_grace();
 676
 677                 list_append_to_xlist(pool, &unmap_list, &clean_xlist, &clean_tail);
 678                 if (ibmr_ret)
 679                         refill_local(pool, &clean_xlist, ibmr_ret);
 680
 681                 /* refill_local may have emptied our list */
 682                 if (!xlist_empty(&clean_xlist))
 683                         xlist_add(clean_xlist.next, clean_tail, &pool->clean_list);
 684
 685         }
 686
 687         atomic_sub(unpinned, &pool->free_pinned);
 688         atomic_sub(ncleaned, &pool->dirty_count);
 689         atomic_sub(nfreed, &pool->item_count);
 690
 691 out:
 692         mutex_unlock(&pool->flush_lock);
 693         if (waitqueue_active(&pool->flush_wait))
 694                 wake_up(&pool->flush_wait);
 695 out_nolock:
 696         return ret;
 697 }
 698
 699 int rds_ib_fmr_init(void)
 700 {
 701         rds_ib_fmr_wq = create_workqueue("rds_fmr_flushd");
 702         if (!rds_ib_fmr_wq)
 703                 return -ENOMEM;
 704         return 0;
 705 }
 706
 707 /*
 708  * By the time this is called all the IB devices should have been torn down and
 709  * had their pools freed.  As each pool is freed its work struct is waited on,
 710  * so the pool flushing work queue should be idle by the time we get here.
 711  */
 712 void rds_ib_fmr_exit(void)
 713 {
 714         destroy_workqueue(rds_ib_fmr_wq);
 715 }
 716
 717 static void rds_ib_mr_pool_flush_worker(struct work_struct *work)
 718 {
 719         struct rds_ib_mr_pool *pool = container_of(work, struct rds_ib_mr_pool, flush_worker.work);
 720
 721         rds_ib_flush_mr_pool(pool, 0, NULL);
 722 }
 723
 724 void rds_ib_free_mr(void *trans_private, int invalidate)
 725 {
 726         struct rds_ib_mr *ibmr = trans_private;
 727         struct rds_ib_device *rds_ibdev = ibmr->device;
 728         struct rds_ib_mr_pool *pool = rds_ibdev->mr_pool;
 729
 730         rdsdebug("RDS/IB: free_mr nents %u\n", ibmr->sg_len);
 731
 732         /* Return it to the pool's free list */
 733         if (ibmr->remap_count >= pool->fmr_attr.max_maps)
 734                 xlist_add(&ibmr->xlist, &ibmr->xlist, &pool->drop_list);
 735         else
 736                 xlist_add(&ibmr->xlist, &ibmr->xlist, &pool->free_list);
 737
 738         atomic_add(ibmr->sg_len, &pool->free_pinned);
 739         atomic_inc(&pool->dirty_count);
 740
 741         /* If we've pinned too many pages, request a flush */
 742         if (atomic_read(&pool->free_pinned) >= pool->max_free_pinned ||
 743             atomic_read(&pool->dirty_count) >= pool->max_items / 10)
 744                 queue_delayed_work(rds_ib_fmr_wq, &pool->flush_worker, 10);
 745
 746         if (invalidate) {
 747                 if (likely(!in_interrupt())) {
 748                         rds_ib_flush_mr_pool(pool, 0, NULL);
 749                 } else {
 750                         /* We get here if the user created a MR marked
 751                          * as use_once and invalidate at the same time. */
 752                         queue_delayed_work(rds_ib_fmr_wq,
 753                                            &pool->flush_worker, 10);
 754                 }
 755         }
 756
 757         rds_ib_dev_put(rds_ibdev);
 758 }
 759
 760 void rds_ib_flush_mrs(void)
 761 {
 762         struct rds_ib_device *rds_ibdev;
 763
 764         down_read(&rds_ib_devices_lock);
 765         list_for_each_entry(rds_ibdev, &rds_ib_devices, list) {
 766                 struct rds_ib_mr_pool *pool = rds_ibdev->mr_pool;
 767
 768                 if (pool)
 769                         rds_ib_flush_mr_pool(pool, 0, NULL);
 770         }
 771         up_read(&rds_ib_devices_lock);
 772 }
 773
 774 void *rds_ib_get_mr(struct scatterlist *sg, unsigned long nents,
 775                     struct rds_sock *rs, u32 *key_ret)
 776 {
 777         struct rds_ib_device *rds_ibdev;
 778         struct rds_ib_mr *ibmr = NULL;
 779         int ret;
 780
 781         rds_ibdev = rds_ib_get_device(rs->rs_bound_addr);
 782         if (!rds_ibdev) {
 783                 ret = -ENODEV;
 784                 goto out;
 785         }
 786
 787         if (!rds_ibdev->mr_pool) {
 788                 ret = -ENODEV;
 789                 goto out;
 790         }
 791
 792         ibmr = rds_ib_alloc_fmr(rds_ibdev);
 793         if (IS_ERR(ibmr))
 794                 return ibmr;
 795
 796         ret = rds_ib_map_fmr(rds_ibdev, ibmr, sg, nents);
 797         if (ret == 0)
 798                 *key_ret = ibmr->fmr->rkey;
 799         else
 800                 printk(KERN_WARNING "RDS/IB: map_fmr failed (errno=%d)\n", ret);
 801
 802         ibmr->device = rds_ibdev;
 803         rds_ibdev = NULL;
 804
 805  out:
 806         if (ret) {
 807                 if (ibmr)
 808                         rds_ib_free_mr(ibmr, 0);
 809                 ibmr = ERR_PTR(ret);
 810         }
 811         if (rds_ibdev)
 812                 rds_ib_dev_put(rds_ibdev);
 813         return ibmr;
 814 }
 815