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
  35 #include "rds.h"
  36 #include "rdma.h"
  37 #include "ib.h"
  38
  39
  40 /*
  41  * This is stored as mr->r_trans_private.
  42  */
  43 struct rds_ib_mr {
  44         struct rds_ib_device    *device;
  45         struct rds_ib_mr_pool   *pool;
  46         struct ib_fmr           *fmr;
  47         struct list_head        list;
  48         unsigned int            remap_count;
  49
  50         struct scatterlist      *sg;
  51         unsigned int            sg_len;
  52         u64                     *dma;
  53         int                     sg_dma_len;
  54 };
  55
  56 /*
  57  * Our own little FMR pool
  58  */
  59 struct rds_ib_mr_pool {
  60         struct mutex            flush_lock;             /* serialize fmr invalidate */
  61         struct work_struct      flush_worker;           /* flush worker */
  62
  63         spinlock_t              list_lock;              /* protect variables below */
  64         atomic_t                item_count;             /* total # of MRs */
  65         atomic_t                dirty_count;            /* # dirty of MRs */
  66         struct list_head        drop_list;              /* MRs that have reached their max_maps limit */
  67         struct list_head        free_list;              /* unused MRs */
  68         struct list_head        clean_list;             /* unused & unamapped MRs */
  69         atomic_t                free_pinned;            /* memory pinned by free MRs */
  70         unsigned long           max_items;
  71         unsigned long           max_items_soft;
  72         unsigned long           max_free_pinned;
  73         struct ib_fmr_attr      fmr_attr;
  74 };
  75
  76 static int rds_ib_flush_mr_pool(struct rds_ib_mr_pool *pool, int free_all);
  77 static void rds_ib_teardown_mr(struct rds_ib_mr *ibmr);
  78 static void rds_ib_mr_pool_flush_worker(struct work_struct *work);
  79
  80 static struct rds_ib_device *rds_ib_get_device(__be32 ipaddr)
  81 {
  82         struct rds_ib_device *rds_ibdev;
  83         struct rds_ib_ipaddr *i_ipaddr;
  84
  85         list_for_each_entry(rds_ibdev, &rds_ib_devices, list) {
  86                 spin_lock_irq(&rds_ibdev->spinlock);
  87                 list_for_each_entry(i_ipaddr, &rds_ibdev->ipaddr_list, list) {
  88                         if (i_ipaddr->ipaddr == ipaddr) {
  89                                 spin_unlock_irq(&rds_ibdev->spinlock);
  90                                 return rds_ibdev;
  91                         }
  92                 }
  93                 spin_unlock_irq(&rds_ibdev->spinlock);
  94         }
  95
  96         return NULL;
  97 }
  98
  99 static int rds_ib_add_ipaddr(struct rds_ib_device *rds_ibdev, __be32 ipaddr)
 100 {
 101         struct rds_ib_ipaddr *i_ipaddr;
 102
 103         i_ipaddr = kmalloc(sizeof *i_ipaddr, GFP_KERNEL);
 104         if (!i_ipaddr)
 105                 return -ENOMEM;
 106
 107         i_ipaddr->ipaddr = ipaddr;
 108
 109         spin_lock_irq(&rds_ibdev->spinlock);
 110         list_add_tail(&i_ipaddr->list, &rds_ibdev->ipaddr_list);
 111         spin_unlock_irq(&rds_ibdev->spinlock);
 112
 113         return 0;
 114 }
 115
 116 static void rds_ib_remove_ipaddr(struct rds_ib_device *rds_ibdev, __be32 ipaddr)
 117 {
 118         struct rds_ib_ipaddr *i_ipaddr, *next;
 119
 120         spin_lock_irq(&rds_ibdev->spinlock);
 121         list_for_each_entry_safe(i_ipaddr, next, &rds_ibdev->ipaddr_list, list) {
 122                 if (i_ipaddr->ipaddr == ipaddr) {
 123                         list_del(&i_ipaddr->list);
 124                         kfree(i_ipaddr);
 125                         break;
 126                 }
 127         }
 128         spin_unlock_irq(&rds_ibdev->spinlock);
 129 }
 130
 131 int rds_ib_update_ipaddr(struct rds_ib_device *rds_ibdev, __be32 ipaddr)
 132 {
 133         struct rds_ib_device *rds_ibdev_old;
 134
 135         rds_ibdev_old = rds_ib_get_device(ipaddr);
 136         if (rds_ibdev_old)
 137                 rds_ib_remove_ipaddr(rds_ibdev_old, ipaddr);
 138
 139         return rds_ib_add_ipaddr(rds_ibdev, ipaddr);
 140 }
 141
 142 int rds_ib_add_conn(struct rds_ib_device *rds_ibdev, struct rds_connection *conn)
 143 {
 144         struct rds_ib_connection *ic = conn->c_transport_data;
 145
 146         /* conn was previously on the nodev_conns_list */
 147         spin_lock_irq(&ib_nodev_conns_lock);
 148         BUG_ON(list_empty(&ib_nodev_conns));
 149         BUG_ON(list_empty(&ic->ib_node));
 150         list_del(&ic->ib_node);
 151         spin_unlock_irq(&ib_nodev_conns_lock);
 152
 153         spin_lock_irq(&rds_ibdev->spinlock);
 154         list_add_tail(&ic->ib_node, &rds_ibdev->conn_list);
 155         spin_unlock_irq(&rds_ibdev->spinlock);
 156
 157         ic->rds_ibdev = rds_ibdev;
 158
 159         return 0;
 160 }
 161
 162 void rds_ib_remove_nodev_conns(void)
 163 {
 164         struct rds_ib_connection *ic, *_ic;
 165         LIST_HEAD(tmp_list);
 166
 167         /* avoid calling conn_destroy with irqs off */
 168         spin_lock_irq(&ib_nodev_conns_lock);
 169         list_splice(&ib_nodev_conns, &tmp_list);
 170         INIT_LIST_HEAD(&ib_nodev_conns);
 171         spin_unlock_irq(&ib_nodev_conns_lock);
 172
 173         list_for_each_entry_safe(ic, _ic, &tmp_list, ib_node) {
 174                 if (ic->conn->c_passive)
 175                         rds_conn_destroy(ic->conn->c_passive);
 176                 rds_conn_destroy(ic->conn);
 177         }
 178 }
 179
 180 void rds_ib_remove_conns(struct rds_ib_device *rds_ibdev)
 181 {
 182         struct rds_ib_connection *ic, *_ic;
 183         LIST_HEAD(tmp_list);
 184
 185         /* avoid calling conn_destroy with irqs off */
 186         spin_lock_irq(&rds_ibdev->spinlock);
 187         list_splice(&rds_ibdev->conn_list, &tmp_list);
 188         INIT_LIST_HEAD(&rds_ibdev->conn_list);
 189         spin_unlock_irq(&rds_ibdev->spinlock);
 190
 191         list_for_each_entry_safe(ic, _ic, &tmp_list, ib_node) {
 192                 if (ic->conn->c_passive)
 193                         rds_conn_destroy(ic->conn->c_passive);
 194                 rds_conn_destroy(ic->conn);
 195         }
 196 }
 197
 198 struct rds_ib_mr_pool *rds_ib_create_mr_pool(struct rds_ib_device *rds_ibdev)
 199 {
 200         struct rds_ib_mr_pool *pool;
 201
 202         pool = kzalloc(sizeof(*pool), GFP_KERNEL);
 203         if (!pool)
 204                 return ERR_PTR(-ENOMEM);
 205
 206         INIT_LIST_HEAD(&pool->free_list);
 207         INIT_LIST_HEAD(&pool->drop_list);
 208         INIT_LIST_HEAD(&pool->clean_list);
 209         mutex_init(&pool->flush_lock);
 210         spin_lock_init(&pool->list_lock);
 211         INIT_WORK(&pool->flush_worker, rds_ib_mr_pool_flush_worker);
 212
 213         pool->fmr_attr.max_pages = fmr_message_size;
 214         pool->fmr_attr.max_maps = rds_ibdev->fmr_max_remaps;
 215         pool->fmr_attr.page_shift = rds_ibdev->fmr_page_shift;
 216         pool->max_free_pinned = rds_ibdev->max_fmrs * fmr_message_size / 4;
 217
 218         /* We never allow more than max_items MRs to be allocated.
 219          * When we exceed more than max_items_soft, we start freeing
 220          * items more aggressively.
 221          * Make sure that max_items > max_items_soft > max_items / 2
 222          */
 223         pool->max_items_soft = rds_ibdev->max_fmrs * 3 / 4;
 224         pool->max_items = rds_ibdev->max_fmrs;
 225
 226         return pool;
 227 }
 228
 229 void rds_ib_get_mr_info(struct rds_ib_device *rds_ibdev, struct rds_info_rdma_connection *iinfo)
 230 {
 231         struct rds_ib_mr_pool *pool = rds_ibdev->mr_pool;
 232
 233         iinfo->rdma_mr_max = pool->max_items;
 234         iinfo->rdma_mr_size = pool->fmr_attr.max_pages;
 235 }
 236
 237 void rds_ib_destroy_mr_pool(struct rds_ib_mr_pool *pool)
 238 {
 239         flush_workqueue(rds_wq);
 240         rds_ib_flush_mr_pool(pool, 1);
 241         BUG_ON(atomic_read(&pool->item_count));
 242         BUG_ON(atomic_read(&pool->free_pinned));
 243         kfree(pool);
 244 }
 245
 246 static inline struct rds_ib_mr *rds_ib_reuse_fmr(struct rds_ib_mr_pool *pool)
 247 {
 248         struct rds_ib_mr *ibmr = NULL;
 249         unsigned long flags;
 250
 251         spin_lock_irqsave(&pool->list_lock, flags);
 252         if (!list_empty(&pool->clean_list)) {
 253                 ibmr = list_entry(pool->clean_list.next, struct rds_ib_mr, list);
 254                 list_del_init(&ibmr->list);
 255         }
 256         spin_unlock_irqrestore(&pool->list_lock, flags);
 257
 258         return ibmr;
 259 }
 260
 261 static struct rds_ib_mr *rds_ib_alloc_fmr(struct rds_ib_device *rds_ibdev)
 262 {
 263         struct rds_ib_mr_pool *pool = rds_ibdev->mr_pool;
 264         struct rds_ib_mr *ibmr = NULL;
 265         int err = 0, iter = 0;
 266
 267         while (1) {
 268                 ibmr = rds_ib_reuse_fmr(pool);
 269                 if (ibmr)
 270                         return ibmr;
 271
 272                 /* No clean MRs - now we have the choice of either
 273                  * allocating a fresh MR up to the limit imposed by the
 274                  * driver, or flush any dirty unused MRs.
 275                  * We try to avoid stalling in the send path if possible,
 276                  * so we allocate as long as we're allowed to.
 277                  *
 278                  * We're fussy with enforcing the FMR limit, though. If the driver
 279                  * tells us we can't use more than N fmrs, we shouldn't start
 280                  * arguing with it */
 281                 if (atomic_inc_return(&pool->item_count) <= pool->max_items)
 282                         break;
 283
 284                 atomic_dec(&pool->item_count);
 285
 286                 if (++iter > 2) {
 287                         rds_ib_stats_inc(s_ib_rdma_mr_pool_depleted);
 288                         return ERR_PTR(-EAGAIN);
 289                 }
 290
 291                 /* We do have some empty MRs. Flush them out. */
 292                 rds_ib_stats_inc(s_ib_rdma_mr_pool_wait);
 293                 rds_ib_flush_mr_pool(pool, 0);
 294         }
 295
 296         ibmr = kzalloc(sizeof(*ibmr), GFP_KERNEL);
 297         if (!ibmr) {
 298                 err = -ENOMEM;
 299                 goto out_no_cigar;
 300         }
 301
 302         ibmr->fmr = ib_alloc_fmr(rds_ibdev->pd,
 303                         (IB_ACCESS_LOCAL_WRITE |
 304                          IB_ACCESS_REMOTE_READ |
 305                          IB_ACCESS_REMOTE_WRITE),
 306                         &pool->fmr_attr);
 307         if (IS_ERR(ibmr->fmr)) {
 308                 err = PTR_ERR(ibmr->fmr);
 309                 ibmr->fmr = NULL;
 310                 printk(KERN_WARNING "RDS/IB: ib_alloc_fmr failed (err=%d)\n", err);
 311                 goto out_no_cigar;
 312         }
 313
 314         rds_ib_stats_inc(s_ib_rdma_mr_alloc);
 315         return ibmr;
 316
 317 out_no_cigar:
 318         if (ibmr) {
 319                 if (ibmr->fmr)
 320                         ib_dealloc_fmr(ibmr->fmr);
 321                 kfree(ibmr);
 322         }
 323         atomic_dec(&pool->item_count);
 324         return ERR_PTR(err);
 325 }
 326
 327 static int rds_ib_map_fmr(struct rds_ib_device *rds_ibdev, struct rds_ib_mr *ibmr,
 328                struct scatterlist *sg, unsigned int nents)
 329 {
 330         struct ib_device *dev = rds_ibdev->dev;
 331         struct scatterlist *scat = sg;
 332         u64 io_addr = 0;
 333         u64 *dma_pages;
 334         u32 len;
 335         int page_cnt, sg_dma_len;
 336         int i, j;
 337         int ret;
 338
 339         sg_dma_len = ib_dma_map_sg(dev, sg, nents,
 340                                  DMA_BIDIRECTIONAL);
 341         if (unlikely(!sg_dma_len)) {
 342                 printk(KERN_WARNING "RDS/IB: dma_map_sg failed!\n");
 343                 return -EBUSY;
 344         }
 345
 346         len = 0;
 347         page_cnt = 0;
 348
 349         for (i = 0; i < sg_dma_len; ++i) {
 350                 unsigned int dma_len = ib_sg_dma_len(dev, &scat[i]);
 351                 u64 dma_addr = ib_sg_dma_address(dev, &scat[i]);
 352
 353                 if (dma_addr & ~rds_ibdev->fmr_page_mask) {
 354                         if (i > 0)
 355                                 return -EINVAL;
 356                         else
 357                                 ++page_cnt;
 358                 }
 359                 if ((dma_addr + dma_len) & ~rds_ibdev->fmr_page_mask) {
 360                         if (i < sg_dma_len - 1)
 361                                 return -EINVAL;
 362                         else
 363                                 ++page_cnt;
 364                 }
 365
 366                 len += dma_len;
 367         }
 368
 369         page_cnt += len >> rds_ibdev->fmr_page_shift;
 370         if (page_cnt > fmr_message_size)
 371                 return -EINVAL;
 372
 373         dma_pages = kmalloc(sizeof(u64) * page_cnt, GFP_ATOMIC);
 374         if (!dma_pages)
 375                 return -ENOMEM;
 376
 377         page_cnt = 0;
 378         for (i = 0; i < sg_dma_len; ++i) {
 379                 unsigned int dma_len = ib_sg_dma_len(dev, &scat[i]);
 380                 u64 dma_addr = ib_sg_dma_address(dev, &scat[i]);
 381
 382                 for (j = 0; j < dma_len; j += rds_ibdev->fmr_page_size)
 383                         dma_pages[page_cnt++] =
 384                                 (dma_addr & rds_ibdev->fmr_page_mask) + j;
 385         }
 386
 387         ret = ib_map_phys_fmr(ibmr->fmr,
 388                                    dma_pages, page_cnt, io_addr);
 389         if (ret)
 390                 goto out;
 391
 392         /* Success - we successfully remapped the MR, so we can
 393          * safely tear down the old mapping. */
 394         rds_ib_teardown_mr(ibmr);
 395
 396         ibmr->sg = scat;
 397         ibmr->sg_len = nents;
 398         ibmr->sg_dma_len = sg_dma_len;
 399         ibmr->remap_count++;
 400
 401         rds_ib_stats_inc(s_ib_rdma_mr_used);
 402         ret = 0;
 403
 404 out:
 405         kfree(dma_pages);
 406
 407         return ret;
 408 }
 409
 410 void rds_ib_sync_mr(void *trans_private, int direction)
 411 {
 412         struct rds_ib_mr *ibmr = trans_private;
 413         struct rds_ib_device *rds_ibdev = ibmr->device;
 414
 415         switch (direction) {
 416         case DMA_FROM_DEVICE:
 417                 ib_dma_sync_sg_for_cpu(rds_ibdev->dev, ibmr->sg,
 418                         ibmr->sg_dma_len, DMA_BIDIRECTIONAL);
 419                 break;
 420         case DMA_TO_DEVICE:
 421                 ib_dma_sync_sg_for_device(rds_ibdev->dev, ibmr->sg,
 422                         ibmr->sg_dma_len, DMA_BIDIRECTIONAL);
 423                 break;
 424         }
 425 }
 426
 427 static void __rds_ib_teardown_mr(struct rds_ib_mr *ibmr)
 428 {
 429         struct rds_ib_device *rds_ibdev = ibmr->device;
 430
 431         if (ibmr->sg_dma_len) {
 432                 ib_dma_unmap_sg(rds_ibdev->dev,
 433                                 ibmr->sg, ibmr->sg_len,
 434                                 DMA_BIDIRECTIONAL);
 435                 ibmr->sg_dma_len = 0;
 436         }
 437
 438         /* Release the s/g list */
 439         if (ibmr->sg_len) {
 440                 unsigned int i;
 441
 442                 for (i = 0; i < ibmr->sg_len; ++i) {
 443                         struct page *page = sg_page(&ibmr->sg[i]);
 444
 445                         /* FIXME we need a way to tell a r/w MR
 446                          * from a r/o MR */
 447                         set_page_dirty(page);
 448                         put_page(page);
 449                 }
 450                 kfree(ibmr->sg);
 451
 452                 ibmr->sg = NULL;
 453                 ibmr->sg_len = 0;
 454         }
 455 }
 456
 457 static void rds_ib_teardown_mr(struct rds_ib_mr *ibmr)
 458 {
 459         unsigned int pinned = ibmr->sg_len;
 460
 461         __rds_ib_teardown_mr(ibmr);
 462         if (pinned) {
 463                 struct rds_ib_device *rds_ibdev = ibmr->device;
 464                 struct rds_ib_mr_pool *pool = rds_ibdev->mr_pool;
 465
 466                 atomic_sub(pinned, &pool->free_pinned);
 467         }
 468 }
 469
 470 static inline unsigned int rds_ib_flush_goal(struct rds_ib_mr_pool *pool, int free_all)
 471 {
 472         unsigned int item_count;
 473
 474         item_count = atomic_read(&pool->item_count);
 475         if (free_all)
 476                 return item_count;
 477
 478         return 0;
 479 }
 480
 481 /*
 482  * Flush our pool of MRs.
 483  * At a minimum, all currently unused MRs are unmapped.
 484  * If the number of MRs allocated exceeds the limit, we also try
 485  * to free as many MRs as needed to get back to this limit.
 486  */
 487 static int rds_ib_flush_mr_pool(struct rds_ib_mr_pool *pool, int free_all)
 488 {
 489         struct rds_ib_mr *ibmr, *next;
 490         LIST_HEAD(unmap_list);
 491         LIST_HEAD(fmr_list);
 492         unsigned long unpinned = 0;
 493         unsigned long flags;
 494         unsigned int nfreed = 0, ncleaned = 0, free_goal;
 495         int ret = 0;
 496
 497         rds_ib_stats_inc(s_ib_rdma_mr_pool_flush);
 498
 499         mutex_lock(&pool->flush_lock);
 500
 501         spin_lock_irqsave(&pool->list_lock, flags);
 502         /* Get the list of all MRs to be dropped. Ordering matters -
 503          * we want to put drop_list ahead of free_list. */
 504         list_splice_init(&pool->free_list, &unmap_list);
 505         list_splice_init(&pool->drop_list, &unmap_list);
 506         if (free_all)
 507                 list_splice_init(&pool->clean_list, &unmap_list);
 508         spin_unlock_irqrestore(&pool->list_lock, flags);
 509
 510         free_goal = rds_ib_flush_goal(pool, free_all);
 511
 512         if (list_empty(&unmap_list))
 513                 goto out;
 514
 515         /* String all ib_mr's onto one list and hand them to ib_unmap_fmr */
 516         list_for_each_entry(ibmr, &unmap_list, list)
 517                 list_add(&ibmr->fmr->list, &fmr_list);
 518         ret = ib_unmap_fmr(&fmr_list);
 519         if (ret)
 520                 printk(KERN_WARNING "RDS/IB: ib_unmap_fmr failed (err=%d)\n", ret);
 521
 522         /* Now we can destroy the DMA mapping and unpin any pages */
 523         list_for_each_entry_safe(ibmr, next, &unmap_list, list) {
 524                 unpinned += ibmr->sg_len;
 525                 __rds_ib_teardown_mr(ibmr);
 526                 if (nfreed < free_goal || ibmr->remap_count >= pool->fmr_attr.max_maps) {
 527                         rds_ib_stats_inc(s_ib_rdma_mr_free);
 528                         list_del(&ibmr->list);
 529                         ib_dealloc_fmr(ibmr->fmr);
 530                         kfree(ibmr);
 531                         nfreed++;
 532                 }
 533                 ncleaned++;
 534         }
 535
 536         spin_lock_irqsave(&pool->list_lock, flags);
 537         list_splice(&unmap_list, &pool->clean_list);
 538         spin_unlock_irqrestore(&pool->list_lock, flags);
 539
 540         atomic_sub(unpinned, &pool->free_pinned);
 541         atomic_sub(ncleaned, &pool->dirty_count);
 542         atomic_sub(nfreed, &pool->item_count);
 543
 544 out:
 545         mutex_unlock(&pool->flush_lock);
 546         return ret;
 547 }
 548
 549 static void rds_ib_mr_pool_flush_worker(struct work_struct *work)
 550 {
 551         struct rds_ib_mr_pool *pool = container_of(work, struct rds_ib_mr_pool, flush_worker);
 552
 553         rds_ib_flush_mr_pool(pool, 0);
 554 }
 555
 556 void rds_ib_free_mr(void *trans_private, int invalidate)
 557 {
 558         struct rds_ib_mr *ibmr = trans_private;
 559         struct rds_ib_device *rds_ibdev = ibmr->device;
 560         struct rds_ib_mr_pool *pool = rds_ibdev->mr_pool;
 561         unsigned long flags;
 562
 563         rdsdebug("RDS/IB: free_mr nents %u\n", ibmr->sg_len);
 564
 565         /* Return it to the pool's free list */
 566         spin_lock_irqsave(&pool->list_lock, flags);
 567         if (ibmr->remap_count >= pool->fmr_attr.max_maps)
 568                 list_add(&ibmr->list, &pool->drop_list);
 569         else
 570                 list_add(&ibmr->list, &pool->free_list);
 571
 572         atomic_add(ibmr->sg_len, &pool->free_pinned);
 573         atomic_inc(&pool->dirty_count);
 574         spin_unlock_irqrestore(&pool->list_lock, flags);
 575
 576         /* If we've pinned too many pages, request a flush */
 577         if (atomic_read(&pool->free_pinned) >= pool->max_free_pinned
 578          || atomic_read(&pool->dirty_count) >= pool->max_items / 10)
 579                 queue_work(rds_wq, &pool->flush_worker);
 580
 581         if (invalidate) {
 582                 if (likely(!in_interrupt())) {
 583                         rds_ib_flush_mr_pool(pool, 0);
 584                 } else {
 585                         /* We get here if the user created a MR marked
 586                          * as use_once and invalidate at the same time. */
 587                         queue_work(rds_wq, &pool->flush_worker);
 588                 }
 589         }
 590 }
 591
 592 void rds_ib_flush_mrs(void)
 593 {
 594         struct rds_ib_device *rds_ibdev;
 595
 596         list_for_each_entry(rds_ibdev, &rds_ib_devices, list) {
 597                 struct rds_ib_mr_pool *pool = rds_ibdev->mr_pool;
 598
 599                 if (pool)
 600                         rds_ib_flush_mr_pool(pool, 0);
 601         }
 602 }
 603
 604 void *rds_ib_get_mr(struct scatterlist *sg, unsigned long nents,
 605                     struct rds_sock *rs, u32 *key_ret)
 606 {
 607         struct rds_ib_device *rds_ibdev;
 608         struct rds_ib_mr *ibmr = NULL;
 609         int ret;
 610
 611         rds_ibdev = rds_ib_get_device(rs->rs_bound_addr);
 612         if (!rds_ibdev) {
 613                 ret = -ENODEV;
 614                 goto out;
 615         }
 616
 617         if (!rds_ibdev->mr_pool) {
 618                 ret = -ENODEV;
 619                 goto out;
 620         }
 621
 622         ibmr = rds_ib_alloc_fmr(rds_ibdev);
 623         if (IS_ERR(ibmr))
 624                 return ibmr;
 625
 626         ret = rds_ib_map_fmr(rds_ibdev, ibmr, sg, nents);
 627         if (ret == 0)
 628                 *key_ret = ibmr->fmr->rkey;
 629         else
 630                 printk(KERN_WARNING "RDS/IB: map_fmr failed (errno=%d)\n", ret);
 631
 632         ibmr->device = rds_ibdev;
 633
 634  out:
 635         if (ret) {
 636                 if (ibmr)
 637                         rds_ib_free_mr(ibmr, 0);
 638                 ibmr = ERR_PTR(ret);
 639         }
 640         return ibmr;
 641 }