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 #include <linux/llist.h>
  37
  38 #include "rds_single_path.h"
  39 #include "ib_mr.h"
  40
  41 struct workqueue_struct *rds_ib_mr_wq;
  42
  43 static DEFINE_PER_CPU(unsigned long, clean_list_grace);
  44 #define CLEAN_LIST_BUSY_BIT 0
  45
  46 static struct rds_ib_device *rds_ib_get_device(__be32 ipaddr)
  47 {
  48         struct rds_ib_device *rds_ibdev;
  49         struct rds_ib_ipaddr *i_ipaddr;
  50
  51         rcu_read_lock();
  52         list_for_each_entry_rcu(rds_ibdev, &rds_ib_devices, list) {
  53                 list_for_each_entry_rcu(i_ipaddr, &rds_ibdev->ipaddr_list, list) {
  54                         if (i_ipaddr->ipaddr == ipaddr) {
  55                                 atomic_inc(&rds_ibdev->refcount);
  56                                 rcu_read_unlock();
  57                                 return rds_ibdev;
  58                         }
  59                 }
  60         }
  61         rcu_read_unlock();
  62
  63         return NULL;
  64 }
  65
  66 static int rds_ib_add_ipaddr(struct rds_ib_device *rds_ibdev, __be32 ipaddr)
  67 {
  68         struct rds_ib_ipaddr *i_ipaddr;
  69
  70         i_ipaddr = kmalloc(sizeof *i_ipaddr, GFP_KERNEL);
  71         if (!i_ipaddr)
  72                 return -ENOMEM;
  73
  74         i_ipaddr->ipaddr = ipaddr;
  75
  76         spin_lock_irq(&rds_ibdev->spinlock);
  77         list_add_tail_rcu(&i_ipaddr->list, &rds_ibdev->ipaddr_list);
  78         spin_unlock_irq(&rds_ibdev->spinlock);
  79
  80         return 0;
  81 }
  82
  83 static void rds_ib_remove_ipaddr(struct rds_ib_device *rds_ibdev, __be32 ipaddr)
  84 {
  85         struct rds_ib_ipaddr *i_ipaddr;
  86         struct rds_ib_ipaddr *to_free = NULL;
  87
  88
  89         spin_lock_irq(&rds_ibdev->spinlock);
  90         list_for_each_entry_rcu(i_ipaddr, &rds_ibdev->ipaddr_list, list) {
  91                 if (i_ipaddr->ipaddr == ipaddr) {
  92                         list_del_rcu(&i_ipaddr->list);
  93                         to_free = i_ipaddr;
  94                         break;
  95                 }
  96         }
  97         spin_unlock_irq(&rds_ibdev->spinlock);
  98
  99         if (to_free)
 100                 kfree_rcu(to_free, rcu);
 101 }
 102
 103 int rds_ib_update_ipaddr(struct rds_ib_device *rds_ibdev, __be32 ipaddr)
 104 {
 105         struct rds_ib_device *rds_ibdev_old;
 106
 107         rds_ibdev_old = rds_ib_get_device(ipaddr);
 108         if (!rds_ibdev_old)
 109                 return rds_ib_add_ipaddr(rds_ibdev, ipaddr);
 110
 111         if (rds_ibdev_old != rds_ibdev) {
 112                 rds_ib_remove_ipaddr(rds_ibdev_old, ipaddr);
 113                 rds_ib_dev_put(rds_ibdev_old);
 114                 return rds_ib_add_ipaddr(rds_ibdev, ipaddr);
 115         }
 116         rds_ib_dev_put(rds_ibdev_old);
 117
 118         return 0;
 119 }
 120
 121 void rds_ib_add_conn(struct rds_ib_device *rds_ibdev, struct rds_connection *conn)
 122 {
 123         struct rds_ib_connection *ic = conn->c_transport_data;
 124
 125         /* conn was previously on the nodev_conns_list */
 126         spin_lock_irq(&ib_nodev_conns_lock);
 127         BUG_ON(list_empty(&ib_nodev_conns));
 128         BUG_ON(list_empty(&ic->ib_node));
 129         list_del(&ic->ib_node);
 130
 131         spin_lock(&rds_ibdev->spinlock);
 132         list_add_tail(&ic->ib_node, &rds_ibdev->conn_list);
 133         spin_unlock(&rds_ibdev->spinlock);
 134         spin_unlock_irq(&ib_nodev_conns_lock);
 135
 136         ic->rds_ibdev = rds_ibdev;
 137         atomic_inc(&rds_ibdev->refcount);
 138 }
 139
 140 void rds_ib_remove_conn(struct rds_ib_device *rds_ibdev, struct rds_connection *conn)
 141 {
 142         struct rds_ib_connection *ic = conn->c_transport_data;
 143
 144         /* place conn on nodev_conns_list */
 145         spin_lock(&ib_nodev_conns_lock);
 146
 147         spin_lock_irq(&rds_ibdev->spinlock);
 148         BUG_ON(list_empty(&ic->ib_node));
 149         list_del(&ic->ib_node);
 150         spin_unlock_irq(&rds_ibdev->spinlock);
 151
 152         list_add_tail(&ic->ib_node, &ib_nodev_conns);
 153
 154         spin_unlock(&ib_nodev_conns_lock);
 155
 156         ic->rds_ibdev = NULL;
 157         rds_ib_dev_put(rds_ibdev);
 158 }
 159
 160 void rds_ib_destroy_nodev_conns(void)
 161 {
 162         struct rds_ib_connection *ic, *_ic;
 163         LIST_HEAD(tmp_list);
 164
 165         /* avoid calling conn_destroy with irqs off */
 166         spin_lock_irq(&ib_nodev_conns_lock);
 167         list_splice(&ib_nodev_conns, &tmp_list);
 168         spin_unlock_irq(&ib_nodev_conns_lock);
 169
 170         list_for_each_entry_safe(ic, _ic, &tmp_list, ib_node)
 171                 rds_conn_destroy(ic->conn);
 172 }
 173
 174 void rds_ib_get_mr_info(struct rds_ib_device *rds_ibdev, struct rds_info_rdma_connection *iinfo)
 175 {
 176         struct rds_ib_mr_pool *pool_1m = rds_ibdev->mr_1m_pool;
 177
 178         iinfo->rdma_mr_max = pool_1m->max_items;
 179         iinfo->rdma_mr_size = pool_1m->fmr_attr.max_pages;
 180 }
 181
 182 struct rds_ib_mr *rds_ib_reuse_mr(struct rds_ib_mr_pool *pool)
 183 {
 184         struct rds_ib_mr *ibmr = NULL;
 185         struct llist_node *ret;
 186         unsigned long *flag;
 187
 188         preempt_disable();
 189         flag = this_cpu_ptr(&clean_list_grace);
 190         set_bit(CLEAN_LIST_BUSY_BIT, flag);
 191         ret = llist_del_first(&pool->clean_list);
 192         if (ret) {
 193                 ibmr = llist_entry(ret, struct rds_ib_mr, llnode);
 194                 if (pool->pool_type == RDS_IB_MR_8K_POOL)
 195                         rds_ib_stats_inc(s_ib_rdma_mr_8k_reused);
 196                 else
 197                         rds_ib_stats_inc(s_ib_rdma_mr_1m_reused);
 198         }
 199
 200         clear_bit(CLEAN_LIST_BUSY_BIT, flag);
 201         preempt_enable();
 202         return ibmr;
 203 }
 204
 205 static inline void wait_clean_list_grace(void)
 206 {
 207         int cpu;
 208         unsigned long *flag;
 209
 210         for_each_online_cpu(cpu) {
 211                 flag = &per_cpu(clean_list_grace, cpu);
 212                 while (test_bit(CLEAN_LIST_BUSY_BIT, flag))
 213                         cpu_relax();
 214         }
 215 }
 216
 217 void rds_ib_sync_mr(void *trans_private, int direction)
 218 {
 219         struct rds_ib_mr *ibmr = trans_private;
 220         struct rds_ib_device *rds_ibdev = ibmr->device;
 221
 222         switch (direction) {
 223         case DMA_FROM_DEVICE:
 224                 ib_dma_sync_sg_for_cpu(rds_ibdev->dev, ibmr->sg,
 225                         ibmr->sg_dma_len, DMA_BIDIRECTIONAL);
 226                 break;
 227         case DMA_TO_DEVICE:
 228                 ib_dma_sync_sg_for_device(rds_ibdev->dev, ibmr->sg,
 229                         ibmr->sg_dma_len, DMA_BIDIRECTIONAL);
 230                 break;
 231         }
 232 }
 233
 234 void __rds_ib_teardown_mr(struct rds_ib_mr *ibmr)
 235 {
 236         struct rds_ib_device *rds_ibdev = ibmr->device;
 237
 238         if (ibmr->sg_dma_len) {
 239                 ib_dma_unmap_sg(rds_ibdev->dev,
 240                                 ibmr->sg, ibmr->sg_len,
 241                                 DMA_BIDIRECTIONAL);
 242                 ibmr->sg_dma_len = 0;
 243         }
 244
 245         /* Release the s/g list */
 246         if (ibmr->sg_len) {
 247                 unsigned int i;
 248
 249                 for (i = 0; i < ibmr->sg_len; ++i) {
 250                         struct page *page = sg_page(&ibmr->sg[i]);
 251
 252                         /* FIXME we need a way to tell a r/w MR
 253                          * from a r/o MR */
 254                         WARN_ON(!page->mapping && irqs_disabled());
 255                         set_page_dirty(page);
 256                         put_page(page);
 257                 }
 258                 kfree(ibmr->sg);
 259
 260                 ibmr->sg = NULL;
 261                 ibmr->sg_len = 0;
 262         }
 263 }
 264
 265 void rds_ib_teardown_mr(struct rds_ib_mr *ibmr)
 266 {
 267         unsigned int pinned = ibmr->sg_len;
 268
 269         __rds_ib_teardown_mr(ibmr);
 270         if (pinned) {
 271                 struct rds_ib_mr_pool *pool = ibmr->pool;
 272
 273                 atomic_sub(pinned, &pool->free_pinned);
 274         }
 275 }
 276
 277 static inline unsigned int rds_ib_flush_goal(struct rds_ib_mr_pool *pool, int free_all)
 278 {
 279         unsigned int item_count;
 280
 281         item_count = atomic_read(&pool->item_count);
 282         if (free_all)
 283                 return item_count;
 284
 285         return 0;
 286 }
 287
 288 /*
 289  * given an llist of mrs, put them all into the list_head for more processing
 290  */
 291 static unsigned int llist_append_to_list(struct llist_head *llist,
 292                                          struct list_head *list)
 293 {
 294         struct rds_ib_mr *ibmr;
 295         struct llist_node *node;
 296         struct llist_node *next;
 297         unsigned int count = 0;
 298
 299         node = llist_del_all(llist);
 300         while (node) {
 301                 next = node->next;
 302                 ibmr = llist_entry(node, struct rds_ib_mr, llnode);
 303                 list_add_tail(&ibmr->unmap_list, list);
 304                 node = next;
 305                 count++;
 306         }
 307         return count;
 308 }
 309
 310 /*
 311  * this takes a list head of mrs and turns it into linked llist nodes
 312  * of clusters.  Each cluster has linked llist nodes of
 313  * MR_CLUSTER_SIZE mrs that are ready for reuse.
 314  */
 315 static void list_to_llist_nodes(struct rds_ib_mr_pool *pool,
 316                                 struct list_head *list,
 317                                 struct llist_node **nodes_head,
 318                                 struct llist_node **nodes_tail)
 319 {
 320         struct rds_ib_mr *ibmr;
 321         struct llist_node *cur = NULL;
 322         struct llist_node **next = nodes_head;
 323
 324         list_for_each_entry(ibmr, list, unmap_list) {
 325                 cur = &ibmr->llnode;
 326                 *next = cur;
 327                 next = &cur->next;
 328         }
 329         *next = NULL;
 330         *nodes_tail = cur;
 331 }
 332
 333 /*
 334  * Flush our pool of MRs.
 335  * At a minimum, all currently unused MRs are unmapped.
 336  * If the number of MRs allocated exceeds the limit, we also try
 337  * to free as many MRs as needed to get back to this limit.
 338  */
 339 int rds_ib_flush_mr_pool(struct rds_ib_mr_pool *pool,
 340                          int free_all, struct rds_ib_mr **ibmr_ret)
 341 {
 342         struct rds_ib_mr *ibmr;
 343         struct llist_node *clean_nodes;
 344         struct llist_node *clean_tail;
 345         LIST_HEAD(unmap_list);
 346         unsigned long unpinned = 0;
 347         unsigned int nfreed = 0, dirty_to_clean = 0, free_goal;
 348
 349         if (pool->pool_type == RDS_IB_MR_8K_POOL)
 350                 rds_ib_stats_inc(s_ib_rdma_mr_8k_pool_flush);
 351         else
 352                 rds_ib_stats_inc(s_ib_rdma_mr_1m_pool_flush);
 353
 354         if (ibmr_ret) {
 355                 DEFINE_WAIT(wait);
 356                 while (!mutex_trylock(&pool->flush_lock)) {
 357                         ibmr = rds_ib_reuse_mr(pool);
 358                         if (ibmr) {
 359                                 *ibmr_ret = ibmr;
 360                                 finish_wait(&pool->flush_wait, &wait);
 361                                 goto out_nolock;
 362                         }
 363
 364                         prepare_to_wait(&pool->flush_wait, &wait,
 365                                         TASK_UNINTERRUPTIBLE);
 366                         if (llist_empty(&pool->clean_list))
 367                                 schedule();
 368
 369                         ibmr = rds_ib_reuse_mr(pool);
 370                         if (ibmr) {
 371                                 *ibmr_ret = ibmr;
 372                                 finish_wait(&pool->flush_wait, &wait);
 373                                 goto out_nolock;
 374                         }
 375                 }
 376                 finish_wait(&pool->flush_wait, &wait);
 377         } else
 378                 mutex_lock(&pool->flush_lock);
 379
 380         if (ibmr_ret) {
 381                 ibmr = rds_ib_reuse_mr(pool);
 382                 if (ibmr) {
 383                         *ibmr_ret = ibmr;
 384                         goto out;
 385                 }
 386         }
 387
 388         /* Get the list of all MRs to be dropped. Ordering matters -
 389          * we want to put drop_list ahead of free_list.
 390          */
 391         dirty_to_clean = llist_append_to_list(&pool->drop_list, &unmap_list);
 392         dirty_to_clean += llist_append_to_list(&pool->free_list, &unmap_list);
 393         if (free_all)
 394                 llist_append_to_list(&pool->clean_list, &unmap_list);
 395
 396         free_goal = rds_ib_flush_goal(pool, free_all);
 397
 398         if (list_empty(&unmap_list))
 399                 goto out;
 400
 401         if (pool->use_fastreg)
 402                 rds_ib_unreg_frmr(&unmap_list, &nfreed, &unpinned, free_goal);
 403         else
 404                 rds_ib_unreg_fmr(&unmap_list, &nfreed, &unpinned, free_goal);
 405
 406         if (!list_empty(&unmap_list)) {
 407                 /* we have to make sure that none of the things we're about
 408                  * to put on the clean list would race with other cpus trying
 409                  * to pull items off.  The llist would explode if we managed to
 410                  * remove something from the clean list and then add it back again
 411                  * while another CPU was spinning on that same item in llist_del_first.
 412                  *
 413                  * This is pretty unlikely, but just in case  wait for an llist grace period
 414                  * here before adding anything back into the clean list.
 415                  */
 416                 wait_clean_list_grace();
 417
 418                 list_to_llist_nodes(pool, &unmap_list, &clean_nodes, &clean_tail);
 419                 if (ibmr_ret)
 420                         *ibmr_ret = llist_entry(clean_nodes, struct rds_ib_mr, llnode);
 421
 422                 /* more than one entry in llist nodes */
 423                 if (clean_nodes->next)
 424                         llist_add_batch(clean_nodes->next, clean_tail, &pool->clean_list);
 425
 426         }
 427
 428         atomic_sub(unpinned, &pool->free_pinned);
 429         atomic_sub(dirty_to_clean, &pool->dirty_count);
 430         atomic_sub(nfreed, &pool->item_count);
 431
 432 out:
 433         mutex_unlock(&pool->flush_lock);
 434         if (waitqueue_active(&pool->flush_wait))
 435                 wake_up(&pool->flush_wait);
 436 out_nolock:
 437         return 0;
 438 }
 439
 440 struct rds_ib_mr *rds_ib_try_reuse_ibmr(struct rds_ib_mr_pool *pool)
 441 {
 442         struct rds_ib_mr *ibmr = NULL;
 443         int iter = 0;
 444
 445         if (atomic_read(&pool->dirty_count) >= pool->max_items_soft / 10)
 446                 queue_delayed_work(rds_ib_mr_wq, &pool->flush_worker, 10);
 447
 448         while (1) {
 449                 ibmr = rds_ib_reuse_mr(pool);
 450                 if (ibmr)
 451                         return ibmr;
 452
 453                 if (atomic_inc_return(&pool->item_count) <= pool->max_items)
 454                         break;
 455
 456                 atomic_dec(&pool->item_count);
 457
 458                 if (++iter > 2) {
 459                         if (pool->pool_type == RDS_IB_MR_8K_POOL)
 460                                 rds_ib_stats_inc(s_ib_rdma_mr_8k_pool_depleted);
 461                         else
 462                                 rds_ib_stats_inc(s_ib_rdma_mr_1m_pool_depleted);
 463                         return ERR_PTR(-EAGAIN);
 464                 }
 465
 466                 /* We do have some empty MRs. Flush them out. */
 467                 if (pool->pool_type == RDS_IB_MR_8K_POOL)
 468                         rds_ib_stats_inc(s_ib_rdma_mr_8k_pool_wait);
 469                 else
 470                         rds_ib_stats_inc(s_ib_rdma_mr_1m_pool_wait);
 471
 472                 rds_ib_flush_mr_pool(pool, 0, &ibmr);
 473                 if (ibmr)
 474                         return ibmr;
 475         }
 476
 477         return ibmr;
 478 }
 479
 480 static void rds_ib_mr_pool_flush_worker(struct work_struct *work)
 481 {
 482         struct rds_ib_mr_pool *pool = container_of(work, struct rds_ib_mr_pool, flush_worker.work);
 483
 484         rds_ib_flush_mr_pool(pool, 0, NULL);
 485 }
 486
 487 void rds_ib_free_mr(void *trans_private, int invalidate)
 488 {
 489         struct rds_ib_mr *ibmr = trans_private;
 490         struct rds_ib_mr_pool *pool = ibmr->pool;
 491         struct rds_ib_device *rds_ibdev = ibmr->device;
 492
 493         rdsdebug("RDS/IB: free_mr nents %u\n", ibmr->sg_len);
 494
 495         /* Return it to the pool's free list */
 496         if (rds_ibdev->use_fastreg)
 497                 rds_ib_free_frmr_list(ibmr);
 498         else
 499                 rds_ib_free_fmr_list(ibmr);
 500
 501         atomic_add(ibmr->sg_len, &pool->free_pinned);
 502         atomic_inc(&pool->dirty_count);
 503
 504         /* If we've pinned too many pages, request a flush */
 505         if (atomic_read(&pool->free_pinned) >= pool->max_free_pinned ||
 506             atomic_read(&pool->dirty_count) >= pool->max_items / 5)
 507                 queue_delayed_work(rds_ib_mr_wq, &pool->flush_worker, 10);
 508
 509         if (invalidate) {
 510                 if (likely(!in_interrupt())) {
 511                         rds_ib_flush_mr_pool(pool, 0, NULL);
 512                 } else {
 513                         /* We get here if the user created a MR marked
 514                          * as use_once and invalidate at the same time.
 515                          */
 516                         queue_delayed_work(rds_ib_mr_wq,
 517                                            &pool->flush_worker, 10);
 518                 }
 519         }
 520
 521         rds_ib_dev_put(rds_ibdev);
 522 }
 523
 524 void rds_ib_flush_mrs(void)
 525 {
 526         struct rds_ib_device *rds_ibdev;
 527
 528         down_read(&rds_ib_devices_lock);
 529         list_for_each_entry(rds_ibdev, &rds_ib_devices, list) {
 530                 if (rds_ibdev->mr_8k_pool)
 531                         rds_ib_flush_mr_pool(rds_ibdev->mr_8k_pool, 0, NULL);
 532
 533                 if (rds_ibdev->mr_1m_pool)
 534                         rds_ib_flush_mr_pool(rds_ibdev->mr_1m_pool, 0, NULL);
 535         }
 536         up_read(&rds_ib_devices_lock);
 537 }
 538
 539 void *rds_ib_get_mr(struct scatterlist *sg, unsigned long nents,
 540                     struct rds_sock *rs, u32 *key_ret)
 541 {
 542         struct rds_ib_device *rds_ibdev;
 543         struct rds_ib_mr *ibmr = NULL;
 544         struct rds_ib_connection *ic = rs->rs_conn->c_transport_data;
 545         int ret;
 546
 547         rds_ibdev = rds_ib_get_device(rs->rs_bound_addr);
 548         if (!rds_ibdev) {
 549                 ret = -ENODEV;
 550                 goto out;
 551         }
 552
 553         if (!rds_ibdev->mr_8k_pool || !rds_ibdev->mr_1m_pool) {
 554                 ret = -ENODEV;
 555                 goto out;
 556         }
 557
 558         if (rds_ibdev->use_fastreg)
 559                 ibmr = rds_ib_reg_frmr(rds_ibdev, ic, sg, nents, key_ret);
 560         else
 561                 ibmr = rds_ib_reg_fmr(rds_ibdev, sg, nents, key_ret);
 562         if (ibmr)
 563                 rds_ibdev = NULL;
 564
 565  out:
 566         if (!ibmr)
 567                 pr_warn("RDS/IB: rds_ib_get_mr failed (errno=%d)\n", ret);
 568
 569         if (rds_ibdev)
 570                 rds_ib_dev_put(rds_ibdev);
 571
 572         return ibmr;
 573 }
 574
 575 void rds_ib_destroy_mr_pool(struct rds_ib_mr_pool *pool)
 576 {
 577         cancel_delayed_work_sync(&pool->flush_worker);
 578         rds_ib_flush_mr_pool(pool, 1, NULL);
 579         WARN_ON(atomic_read(&pool->item_count));
 580         WARN_ON(atomic_read(&pool->free_pinned));
 581         kfree(pool);
 582 }
 583
 584 struct rds_ib_mr_pool *rds_ib_create_mr_pool(struct rds_ib_device *rds_ibdev,
 585                                              int pool_type)
 586 {
 587         struct rds_ib_mr_pool *pool;
 588
 589         pool = kzalloc(sizeof(*pool), GFP_KERNEL);
 590         if (!pool)
 591                 return ERR_PTR(-ENOMEM);
 592
 593         pool->pool_type = pool_type;
 594         init_llist_head(&pool->free_list);
 595         init_llist_head(&pool->drop_list);
 596         init_llist_head(&pool->clean_list);
 597         mutex_init(&pool->flush_lock);
 598         init_waitqueue_head(&pool->flush_wait);
 599         INIT_DELAYED_WORK(&pool->flush_worker, rds_ib_mr_pool_flush_worker);
 600
 601         if (pool_type == RDS_IB_MR_1M_POOL) {
 602                 /* +1 allows for unaligned MRs */
 603                 pool->fmr_attr.max_pages = RDS_MR_1M_MSG_SIZE + 1;
 604                 pool->max_items = RDS_MR_1M_POOL_SIZE;
 605         } else {
 606                 /* pool_type == RDS_IB_MR_8K_POOL */
 607                 pool->fmr_attr.max_pages = RDS_MR_8K_MSG_SIZE + 1;
 608                 pool->max_items = RDS_MR_8K_POOL_SIZE;
 609         }
 610
 611         pool->max_free_pinned = pool->max_items * pool->fmr_attr.max_pages / 4;
 612         pool->fmr_attr.max_maps = rds_ibdev->fmr_max_remaps;
 613         pool->fmr_attr.page_shift = PAGE_SHIFT;
 614         pool->max_items_soft = rds_ibdev->max_mrs * 3 / 4;
 615         pool->use_fastreg = rds_ibdev->use_fastreg;
 616
 617         return pool;
 618 }
 619
 620 int rds_ib_mr_init(void)
 621 {
 622         rds_ib_mr_wq = alloc_workqueue("rds_mr_flushd", WQ_MEM_RECLAIM, 0);
 623         if (!rds_ib_mr_wq)
 624                 return -ENOMEM;
 625         return 0;
 626 }
 627
 628 /* By the time this is called all the IB devices should have been torn down and
 629  * had their pools freed.  As each pool is freed its work struct is waited on,
 630  * so the pool flushing work queue should be idle by the time we get here.
 631  */
 632 void rds_ib_mr_exit(void)
 633 {
 634         destroy_workqueue(rds_ib_mr_wq);
 635 }