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