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
  40
  41 /*
  42  * This is stored as mr->r_trans_private.
  43  */
  44 struct rds_ib_mr {
  45         struct rds_ib_device    *device;
  46         struct rds_ib_mr_pool   *pool;
  47         struct ib_fmr           *fmr;
  48         struct list_head        list;
  49         unsigned int            remap_count;
  50
  51         struct scatterlist      *sg;
  52         unsigned int            sg_len;
  53         u64                     *dma;
  54         int                     sg_dma_len;
  55 };
  56
  57 /*
  58  * Our own little FMR pool
  59  */
  60 struct rds_ib_mr_pool {
  61         struct mutex            flush_lock;             /* serialize fmr invalidate */
  62         struct work_struct      flush_worker;           /* flush worker */
  63
  64         spinlock_t              list_lock;              /* protect variables below */
  65         atomic_t                item_count;             /* total # of MRs */
  66         atomic_t                dirty_count;            /* # dirty of MRs */
  67         struct list_head        drop_list;              /* MRs that have reached their max_maps limit */
  68         struct list_head        free_list;              /* unused MRs */
  69         struct list_head        clean_list;             /* unused & unamapped MRs */
  70         atomic_t                free_pinned;            /* memory pinned by free MRs */
  71         unsigned long           max_items;
  72         unsigned long           max_items_soft;
  73         unsigned long           max_free_pinned;
  74         struct ib_fmr_attr      fmr_attr;
  75 };
  76
  77 static int rds_ib_flush_mr_pool(struct rds_ib_mr_pool *pool, int free_all);
  78 static void rds_ib_teardown_mr(struct rds_ib_mr *ibmr);
  79 static void rds_ib_mr_pool_flush_worker(struct work_struct *work);
  80
  81 static struct rds_ib_device *rds_ib_get_device(__be32 ipaddr)
  82 {
  83         struct rds_ib_device *rds_ibdev;
  84         struct rds_ib_ipaddr *i_ipaddr;
  85
  86         list_for_each_entry(rds_ibdev, &rds_ib_devices, list) {
  87                 rcu_read_lock();
  88                 list_for_each_entry_rcu(i_ipaddr, &rds_ibdev->ipaddr_list, list) {
  89                         if (i_ipaddr->ipaddr == ipaddr) {
  90                                 atomic_inc(&rds_ibdev->refcount);
  91                                 rcu_read_unlock();
  92                                 return rds_ibdev;
  93                         }
  94                 }
  95                 rcu_read_unlock();
  96         }
  97
  98         return NULL;
  99 }
 100
 101 static int rds_ib_add_ipaddr(struct rds_ib_device *rds_ibdev, __be32 ipaddr)
 102 {
 103         struct rds_ib_ipaddr *i_ipaddr;
 104
 105         i_ipaddr = kmalloc(sizeof *i_ipaddr, GFP_KERNEL);
 106         if (!i_ipaddr)
 107                 return -ENOMEM;
 108
 109         i_ipaddr->ipaddr = ipaddr;
 110
 111         spin_lock_irq(&rds_ibdev->spinlock);
 112         list_add_tail_rcu(&i_ipaddr->list, &rds_ibdev->ipaddr_list);
 113         spin_unlock_irq(&rds_ibdev->spinlock);
 114
 115         return 0;
 116 }
 117
 118 static void rds_ib_remove_ipaddr(struct rds_ib_device *rds_ibdev, __be32 ipaddr)
 119 {
 120         struct rds_ib_ipaddr *i_ipaddr;
 121         struct rds_ib_ipaddr *to_free = NULL;
 122
 123
 124         spin_lock_irq(&rds_ibdev->spinlock);
 125         list_for_each_entry_rcu(i_ipaddr, &rds_ibdev->ipaddr_list, list) {
 126                 if (i_ipaddr->ipaddr == ipaddr) {
 127                         list_del_rcu(&i_ipaddr->list);
 128                         to_free = i_ipaddr;
 129                         break;
 130                 }
 131         }
 132         spin_unlock_irq(&rds_ibdev->spinlock);
 133
 134         if (to_free) {
 135                 synchronize_rcu();
 136                 kfree(to_free);
 137         }
 138 }
 139
 140 int rds_ib_update_ipaddr(struct rds_ib_device *rds_ibdev, __be32 ipaddr)
 141 {
 142         struct rds_ib_device *rds_ibdev_old;
 143
 144         rds_ibdev_old = rds_ib_get_device(ipaddr);
 145         if (rds_ibdev_old) {
 146                 rds_ib_remove_ipaddr(rds_ibdev_old, ipaddr);
 147                 rds_ib_dev_put(rds_ibdev_old);
 148         }
 149
 150         return rds_ib_add_ipaddr(rds_ibdev, ipaddr);
 151 }
 152
 153 void rds_ib_add_conn(struct rds_ib_device *rds_ibdev, struct rds_connection *conn)
 154 {
 155         struct rds_ib_connection *ic = conn->c_transport_data;
 156
 157         /* conn was previously on the nodev_conns_list */
 158         spin_lock_irq(&ib_nodev_conns_lock);
 159         BUG_ON(list_empty(&ib_nodev_conns));
 160         BUG_ON(list_empty(&ic->ib_node));
 161         list_del(&ic->ib_node);
 162
 163         spin_lock_irq(&rds_ibdev->spinlock);
 164         list_add_tail(&ic->ib_node, &rds_ibdev->conn_list);
 165         spin_unlock_irq(&rds_ibdev->spinlock);
 166         spin_unlock_irq(&ib_nodev_conns_lock);
 167
 168         ic->rds_ibdev = rds_ibdev;
 169         atomic_inc(&rds_ibdev->refcount);
 170 }
 171
 172 void rds_ib_remove_conn(struct rds_ib_device *rds_ibdev, struct rds_connection *conn)
 173 {
 174         struct rds_ib_connection *ic = conn->c_transport_data;
 175
 176         /* place conn on nodev_conns_list */
 177         spin_lock(&ib_nodev_conns_lock);
 178
 179         spin_lock_irq(&rds_ibdev->spinlock);
 180         BUG_ON(list_empty(&ic->ib_node));
 181         list_del(&ic->ib_node);
 182         spin_unlock_irq(&rds_ibdev->spinlock);
 183
 184         list_add_tail(&ic->ib_node, &ib_nodev_conns);
 185
 186         spin_unlock(&ib_nodev_conns_lock);
 187
 188         ic->rds_ibdev = NULL;
 189         rds_ib_dev_put(rds_ibdev);
 190 }
 191
 192 void __rds_ib_destroy_conns(struct list_head *list, spinlock_t *list_lock)
 193 {
 194         struct rds_ib_connection *ic, *_ic;
 195         LIST_HEAD(tmp_list);
 196
 197         /* avoid calling conn_destroy with irqs off */
 198         spin_lock_irq(list_lock);
 199         list_splice(list, &tmp_list);
 200         INIT_LIST_HEAD(list);
 201         spin_unlock_irq(list_lock);
 202
 203         list_for_each_entry_safe(ic, _ic, &tmp_list, ib_node)
 204                 rds_conn_destroy(ic->conn);
 205 }
 206
 207 struct rds_ib_mr_pool *rds_ib_create_mr_pool(struct rds_ib_device *rds_ibdev)
 208 {
 209         struct rds_ib_mr_pool *pool;
 210
 211         pool = kzalloc(sizeof(*pool), GFP_KERNEL);
 212         if (!pool)
 213                 return ERR_PTR(-ENOMEM);
 214
 215         INIT_LIST_HEAD(&pool->free_list);
 216         INIT_LIST_HEAD(&pool->drop_list);
 217         INIT_LIST_HEAD(&pool->clean_list);
 218         mutex_init(&pool->flush_lock);
 219         spin_lock_init(&pool->list_lock);
 220         INIT_WORK(&pool->flush_worker, rds_ib_mr_pool_flush_worker);
 221
 222         pool->fmr_attr.max_pages = fmr_message_size;
 223         pool->fmr_attr.max_maps = rds_ibdev->fmr_max_remaps;
 224         pool->fmr_attr.page_shift = PAGE_SHIFT;
 225         pool->max_free_pinned = rds_ibdev->max_fmrs * fmr_message_size / 4;
 226
 227         /* We never allow more than max_items MRs to be allocated.
 228          * When we exceed more than max_items_soft, we start freeing
 229          * items more aggressively.
 230          * Make sure that max_items > max_items_soft > max_items / 2
 231          */
 232         pool->max_items_soft = rds_ibdev->max_fmrs * 3 / 4;
 233         pool->max_items = rds_ibdev->max_fmrs;
 234
 235         return pool;
 236 }
 237
 238 void rds_ib_get_mr_info(struct rds_ib_device *rds_ibdev, struct rds_info_rdma_connection *iinfo)
 239 {
 240         struct rds_ib_mr_pool *pool = rds_ibdev->mr_pool;
 241
 242         iinfo->rdma_mr_max = pool->max_items;
 243         iinfo->rdma_mr_size = pool->fmr_attr.max_pages;
 244 }
 245
 246 void rds_ib_destroy_mr_pool(struct rds_ib_mr_pool *pool)
 247 {
 248         cancel_work_sync(&pool->flush_worker);
 249         rds_ib_flush_mr_pool(pool, 1);
 250         WARN_ON(atomic_read(&pool->item_count));
 251         WARN_ON(atomic_read(&pool->free_pinned));
 252         kfree(pool);
 253 }
 254
 255 static inline struct rds_ib_mr *rds_ib_reuse_fmr(struct rds_ib_mr_pool *pool)
 256 {
 257         struct rds_ib_mr *ibmr = NULL;
 258         unsigned long flags;
 259
 260         spin_lock_irqsave(&pool->list_lock, flags);
 261         if (!list_empty(&pool->clean_list)) {
 262                 ibmr = list_entry(pool->clean_list.next, struct rds_ib_mr, list);
 263                 list_del_init(&ibmr->list);
 264         }
 265         spin_unlock_irqrestore(&pool->list_lock, flags);
 266
 267         return ibmr;
 268 }
 269
 270 static struct rds_ib_mr *rds_ib_alloc_fmr(struct rds_ib_device *rds_ibdev)
 271 {
 272         struct rds_ib_mr_pool *pool = rds_ibdev->mr_pool;
 273         struct rds_ib_mr *ibmr = NULL;
 274         int err = 0, iter = 0;
 275
 276         while (1) {
 277                 ibmr = rds_ib_reuse_fmr(pool);
 278                 if (ibmr)
 279                         return ibmr;
 280
 281                 /* No clean MRs - now we have the choice of either
 282                  * allocating a fresh MR up to the limit imposed by the
 283                  * driver, or flush any dirty unused MRs.
 284                  * We try to avoid stalling in the send path if possible,
 285                  * so we allocate as long as we're allowed to.
 286                  *
 287                  * We're fussy with enforcing the FMR limit, though. If the driver
 288                  * tells us we can't use more than N fmrs, we shouldn't start
 289                  * arguing with it */
 290                 if (atomic_inc_return(&pool->item_count) <= pool->max_items)
 291                         break;
 292
 293                 atomic_dec(&pool->item_count);
 294
 295                 if (++iter > 2) {
 296                         rds_ib_stats_inc(s_ib_rdma_mr_pool_depleted);
 297                         return ERR_PTR(-EAGAIN);
 298                 }
 299
 300                 /* We do have some empty MRs. Flush them out. */
 301                 rds_ib_stats_inc(s_ib_rdma_mr_pool_wait);
 302                 rds_ib_flush_mr_pool(pool, 0);
 303         }
 304
 305         ibmr = kzalloc_node(sizeof(*ibmr), GFP_KERNEL, rdsibdev_to_node(rds_ibdev));
 306         if (!ibmr) {
 307                 err = -ENOMEM;
 308                 goto out_no_cigar;
 309         }
 310
 311         memset(ibmr, 0, sizeof(*ibmr));
 312
 313         ibmr->fmr = ib_alloc_fmr(rds_ibdev->pd,
 314                         (IB_ACCESS_LOCAL_WRITE |
 315                          IB_ACCESS_REMOTE_READ |
 316                          IB_ACCESS_REMOTE_WRITE|
 317                          IB_ACCESS_REMOTE_ATOMIC),
 318
 319                         &pool->fmr_attr);
 320         if (IS_ERR(ibmr->fmr)) {
 321                 err = PTR_ERR(ibmr->fmr);
 322                 ibmr->fmr = NULL;
 323                 printk(KERN_WARNING "RDS/IB: ib_alloc_fmr failed (err=%d)\n", err);
 324                 goto out_no_cigar;
 325         }
 326
 327         rds_ib_stats_inc(s_ib_rdma_mr_alloc);
 328         return ibmr;
 329
 330 out_no_cigar:
 331         if (ibmr) {
 332                 if (ibmr->fmr)
 333                         ib_dealloc_fmr(ibmr->fmr);
 334                 kfree(ibmr);
 335         }
 336         atomic_dec(&pool->item_count);
 337         return ERR_PTR(err);
 338 }
 339
 340 static int rds_ib_map_fmr(struct rds_ib_device *rds_ibdev, struct rds_ib_mr *ibmr,
 341                struct scatterlist *sg, unsigned int nents)
 342 {
 343         struct ib_device *dev = rds_ibdev->dev;
 344         struct scatterlist *scat = sg;
 345         u64 io_addr = 0;
 346         u64 *dma_pages;
 347         u32 len;
 348         int page_cnt, sg_dma_len;
 349         int i, j;
 350         int ret;
 351
 352         sg_dma_len = ib_dma_map_sg(dev, sg, nents,
 353                                  DMA_BIDIRECTIONAL);
 354         if (unlikely(!sg_dma_len)) {
 355                 printk(KERN_WARNING "RDS/IB: dma_map_sg failed!\n");
 356                 return -EBUSY;
 357         }
 358
 359         len = 0;
 360         page_cnt = 0;
 361
 362         for (i = 0; i < sg_dma_len; ++i) {
 363                 unsigned int dma_len = ib_sg_dma_len(dev, &scat[i]);
 364                 u64 dma_addr = ib_sg_dma_address(dev, &scat[i]);
 365
 366                 if (dma_addr & ~PAGE_MASK) {
 367                         if (i > 0)
 368                                 return -EINVAL;
 369                         else
 370                                 ++page_cnt;
 371                 }
 372                 if ((dma_addr + dma_len) & ~PAGE_MASK) {
 373                         if (i < sg_dma_len - 1)
 374                                 return -EINVAL;
 375                         else
 376                                 ++page_cnt;
 377                 }
 378
 379                 len += dma_len;
 380         }
 381
 382         page_cnt += len >> PAGE_SHIFT;
 383         if (page_cnt > fmr_message_size)
 384                 return -EINVAL;
 385
 386         dma_pages = kmalloc_node(sizeof(u64) * page_cnt, GFP_ATOMIC,
 387                                  rdsibdev_to_node(rds_ibdev));
 388         if (!dma_pages)
 389                 return -ENOMEM;
 390
 391         page_cnt = 0;
 392         for (i = 0; i < sg_dma_len; ++i) {
 393                 unsigned int dma_len = ib_sg_dma_len(dev, &scat[i]);
 394                 u64 dma_addr = ib_sg_dma_address(dev, &scat[i]);
 395
 396                 for (j = 0; j < dma_len; j += PAGE_SIZE)
 397                         dma_pages[page_cnt++] =
 398                                 (dma_addr & PAGE_MASK) + j;
 399         }
 400
 401         ret = ib_map_phys_fmr(ibmr->fmr,
 402                                    dma_pages, page_cnt, io_addr);
 403         if (ret)
 404                 goto out;
 405
 406         /* Success - we successfully remapped the MR, so we can
 407          * safely tear down the old mapping. */
 408         rds_ib_teardown_mr(ibmr);
 409
 410         ibmr->sg = scat;
 411         ibmr->sg_len = nents;
 412         ibmr->sg_dma_len = sg_dma_len;
 413         ibmr->remap_count++;
 414
 415         rds_ib_stats_inc(s_ib_rdma_mr_used);
 416         ret = 0;
 417
 418 out:
 419         kfree(dma_pages);
 420
 421         return ret;
 422 }
 423
 424 void rds_ib_sync_mr(void *trans_private, int direction)
 425 {
 426         struct rds_ib_mr *ibmr = trans_private;
 427         struct rds_ib_device *rds_ibdev = ibmr->device;
 428
 429         switch (direction) {
 430         case DMA_FROM_DEVICE:
 431                 ib_dma_sync_sg_for_cpu(rds_ibdev->dev, ibmr->sg,
 432                         ibmr->sg_dma_len, DMA_BIDIRECTIONAL);
 433                 break;
 434         case DMA_TO_DEVICE:
 435                 ib_dma_sync_sg_for_device(rds_ibdev->dev, ibmr->sg,
 436                         ibmr->sg_dma_len, DMA_BIDIRECTIONAL);
 437                 break;
 438         }
 439 }
 440
 441 static void __rds_ib_teardown_mr(struct rds_ib_mr *ibmr)
 442 {
 443         struct rds_ib_device *rds_ibdev = ibmr->device;
 444
 445         if (ibmr->sg_dma_len) {
 446                 ib_dma_unmap_sg(rds_ibdev->dev,
 447                                 ibmr->sg, ibmr->sg_len,
 448                                 DMA_BIDIRECTIONAL);
 449                 ibmr->sg_dma_len = 0;
 450         }
 451
 452         /* Release the s/g list */
 453         if (ibmr->sg_len) {
 454                 unsigned int i;
 455
 456                 for (i = 0; i < ibmr->sg_len; ++i) {
 457                         struct page *page = sg_page(&ibmr->sg[i]);
 458
 459                         /* FIXME we need a way to tell a r/w MR
 460                          * from a r/o MR */
 461                         BUG_ON(irqs_disabled());
 462                         set_page_dirty(page);
 463                         put_page(page);
 464                 }
 465                 kfree(ibmr->sg);
 466
 467                 ibmr->sg = NULL;
 468                 ibmr->sg_len = 0;
 469         }
 470 }
 471
 472 static void rds_ib_teardown_mr(struct rds_ib_mr *ibmr)
 473 {
 474         unsigned int pinned = ibmr->sg_len;
 475
 476         __rds_ib_teardown_mr(ibmr);
 477         if (pinned) {
 478                 struct rds_ib_device *rds_ibdev = ibmr->device;
 479                 struct rds_ib_mr_pool *pool = rds_ibdev->mr_pool;
 480
 481                 atomic_sub(pinned, &pool->free_pinned);
 482         }
 483 }
 484
 485 static inline unsigned int rds_ib_flush_goal(struct rds_ib_mr_pool *pool, int free_all)
 486 {
 487         unsigned int item_count;
 488
 489         item_count = atomic_read(&pool->item_count);
 490         if (free_all)
 491                 return item_count;
 492
 493         return 0;
 494 }
 495
 496 /*
 497  * Flush our pool of MRs.
 498  * At a minimum, all currently unused MRs are unmapped.
 499  * If the number of MRs allocated exceeds the limit, we also try
 500  * to free as many MRs as needed to get back to this limit.
 501  */
 502 static int rds_ib_flush_mr_pool(struct rds_ib_mr_pool *pool, int free_all)
 503 {
 504         struct rds_ib_mr *ibmr, *next;
 505         LIST_HEAD(unmap_list);
 506         LIST_HEAD(fmr_list);
 507         unsigned long unpinned = 0;
 508         unsigned long flags;
 509         unsigned int nfreed = 0, ncleaned = 0, free_goal;
 510         int ret = 0;
 511
 512         rds_ib_stats_inc(s_ib_rdma_mr_pool_flush);
 513
 514         mutex_lock(&pool->flush_lock);
 515
 516         spin_lock_irqsave(&pool->list_lock, flags);
 517         /* Get the list of all MRs to be dropped. Ordering matters -
 518          * we want to put drop_list ahead of free_list. */
 519         list_splice_init(&pool->free_list, &unmap_list);
 520         list_splice_init(&pool->drop_list, &unmap_list);
 521         if (free_all)
 522                 list_splice_init(&pool->clean_list, &unmap_list);
 523         spin_unlock_irqrestore(&pool->list_lock, flags);
 524
 525         free_goal = rds_ib_flush_goal(pool, free_all);
 526
 527         if (list_empty(&unmap_list))
 528                 goto out;
 529
 530         /* String all ib_mr's onto one list and hand them to ib_unmap_fmr */
 531         list_for_each_entry(ibmr, &unmap_list, list)
 532                 list_add(&ibmr->fmr->list, &fmr_list);
 533         ret = ib_unmap_fmr(&fmr_list);
 534         if (ret)
 535                 printk(KERN_WARNING "RDS/IB: ib_unmap_fmr failed (err=%d)\n", ret);
 536
 537         /* Now we can destroy the DMA mapping and unpin any pages */
 538         list_for_each_entry_safe(ibmr, next, &unmap_list, list) {
 539                 unpinned += ibmr->sg_len;
 540                 __rds_ib_teardown_mr(ibmr);
 541                 if (nfreed < free_goal || ibmr->remap_count >= pool->fmr_attr.max_maps) {
 542                         rds_ib_stats_inc(s_ib_rdma_mr_free);
 543                         list_del(&ibmr->list);
 544                         ib_dealloc_fmr(ibmr->fmr);
 545                         kfree(ibmr);
 546                         nfreed++;
 547                 }
 548                 ncleaned++;
 549         }
 550
 551         spin_lock_irqsave(&pool->list_lock, flags);
 552         list_splice(&unmap_list, &pool->clean_list);
 553         spin_unlock_irqrestore(&pool->list_lock, flags);
 554
 555         atomic_sub(unpinned, &pool->free_pinned);
 556         atomic_sub(ncleaned, &pool->dirty_count);
 557         atomic_sub(nfreed, &pool->item_count);
 558
 559 out:
 560         mutex_unlock(&pool->flush_lock);
 561         return ret;
 562 }
 563
 564 static void rds_ib_mr_pool_flush_worker(struct work_struct *work)
 565 {
 566         struct rds_ib_mr_pool *pool = container_of(work, struct rds_ib_mr_pool, flush_worker);
 567
 568         rds_ib_flush_mr_pool(pool, 0);
 569 }
 570
 571 void rds_ib_free_mr(void *trans_private, int invalidate)
 572 {
 573         struct rds_ib_mr *ibmr = trans_private;
 574         struct rds_ib_device *rds_ibdev = ibmr->device;
 575         struct rds_ib_mr_pool *pool = rds_ibdev->mr_pool;
 576         unsigned long flags;
 577
 578         rdsdebug("RDS/IB: free_mr nents %u\n", ibmr->sg_len);
 579
 580         /* Return it to the pool's free list */
 581         spin_lock_irqsave(&pool->list_lock, flags);
 582         if (ibmr->remap_count >= pool->fmr_attr.max_maps)
 583                 list_add(&ibmr->list, &pool->drop_list);
 584         else
 585                 list_add(&ibmr->list, &pool->free_list);
 586
 587         atomic_add(ibmr->sg_len, &pool->free_pinned);
 588         atomic_inc(&pool->dirty_count);
 589         spin_unlock_irqrestore(&pool->list_lock, flags);
 590
 591         /* If we've pinned too many pages, request a flush */
 592         if (atomic_read(&pool->free_pinned) >= pool->max_free_pinned ||
 593             atomic_read(&pool->dirty_count) >= pool->max_items / 10)
 594                 queue_work(rds_wq, &pool->flush_worker);
 595
 596         if (invalidate) {
 597                 if (likely(!in_interrupt())) {
 598                         rds_ib_flush_mr_pool(pool, 0);
 599                 } else {
 600                         /* We get here if the user created a MR marked
 601                          * as use_once and invalidate at the same time. */
 602                         queue_work(rds_wq, &pool->flush_worker);
 603                 }
 604         }
 605
 606         rds_ib_dev_put(rds_ibdev);
 607 }
 608
 609 void rds_ib_flush_mrs(void)
 610 {
 611         struct rds_ib_device *rds_ibdev;
 612
 613         list_for_each_entry(rds_ibdev, &rds_ib_devices, list) {
 614                 struct rds_ib_mr_pool *pool = rds_ibdev->mr_pool;
 615
 616                 if (pool)
 617                         rds_ib_flush_mr_pool(pool, 0);
 618         }
 619 }
 620
 621 void *rds_ib_get_mr(struct scatterlist *sg, unsigned long nents,
 622                     struct rds_sock *rs, u32 *key_ret)
 623 {
 624         struct rds_ib_device *rds_ibdev;
 625         struct rds_ib_mr *ibmr = NULL;
 626         int ret;
 627
 628         rds_ibdev = rds_ib_get_device(rs->rs_bound_addr);
 629         if (!rds_ibdev) {
 630                 ret = -ENODEV;
 631                 goto out;
 632         }
 633
 634         if (!rds_ibdev->mr_pool) {
 635                 ret = -ENODEV;
 636                 goto out;
 637         }
 638
 639         ibmr = rds_ib_alloc_fmr(rds_ibdev);
 640         if (IS_ERR(ibmr))
 641                 return ibmr;
 642
 643         ret = rds_ib_map_fmr(rds_ibdev, ibmr, sg, nents);
 644         if (ret == 0)
 645                 *key_ret = ibmr->fmr->rkey;
 646         else
 647                 printk(KERN_WARNING "RDS/IB: map_fmr failed (errno=%d)\n", ret);
 648
 649         ibmr->device = rds_ibdev;
 650         rds_ibdev = NULL;
 651
 652  out:
 653         if (ret) {
 654                 if (ibmr)
 655                         rds_ib_free_mr(ibmr, 0);
 656                 ibmr = ERR_PTR(ret);
 657         }
 658         if (rds_ibdev)
 659                 rds_ib_dev_put(rds_ibdev);
 660         return ibmr;
 661 }