net/rds/ib_rdma.c

   1 /*
   2  * Copyright (c) 2006, 2018 Oracle and/or its affiliates. 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 #include "rds.h"
  41
  42 struct workqueue_struct *rds_ib_mr_wq;
  43 struct rds_ib_dereg_odp_mr {
  44         struct work_struct work;
  45         struct ib_mr *mr;
  46 };
  47
  48 static void rds_ib_odp_mr_worker(struct work_struct *work);
  49
  50 static struct rds_ib_device *rds_ib_get_device(__be32 ipaddr)
  51 {
  52         struct rds_ib_device *rds_ibdev;
  53         struct rds_ib_ipaddr *i_ipaddr;
  54
  55         rcu_read_lock();
  56         list_for_each_entry_rcu(rds_ibdev, &rds_ib_devices, list) {
  57                 list_for_each_entry_rcu(i_ipaddr, &rds_ibdev->ipaddr_list, list) {
  58                         if (i_ipaddr->ipaddr == ipaddr) {
  59                                 refcount_inc(&rds_ibdev->refcount);
  60                                 rcu_read_unlock();
  61                                 return rds_ibdev;
  62                         }
  63                 }
  64         }
  65         rcu_read_unlock();
  66
  67         return NULL;
  68 }
  69
  70 static int rds_ib_add_ipaddr(struct rds_ib_device *rds_ibdev, __be32 ipaddr)
  71 {
  72         struct rds_ib_ipaddr *i_ipaddr;
  73
  74         i_ipaddr = kmalloc(sizeof *i_ipaddr, GFP_KERNEL);
  75         if (!i_ipaddr)
  76                 return -ENOMEM;
  77
  78         i_ipaddr->ipaddr = ipaddr;
  79
  80         spin_lock_irq(&rds_ibdev->spinlock);
  81         list_add_tail_rcu(&i_ipaddr->list, &rds_ibdev->ipaddr_list);
  82         spin_unlock_irq(&rds_ibdev->spinlock);
  83
  84         return 0;
  85 }
  86
  87 static void rds_ib_remove_ipaddr(struct rds_ib_device *rds_ibdev, __be32 ipaddr)
  88 {
  89         struct rds_ib_ipaddr *i_ipaddr;
  90         struct rds_ib_ipaddr *to_free = NULL;
  91
  92
  93         spin_lock_irq(&rds_ibdev->spinlock);
  94         list_for_each_entry_rcu(i_ipaddr, &rds_ibdev->ipaddr_list, list) {
  95                 if (i_ipaddr->ipaddr == ipaddr) {
  96                         list_del_rcu(&i_ipaddr->list);
  97                         to_free = i_ipaddr;
  98                         break;
  99                 }
 100         }
 101         spin_unlock_irq(&rds_ibdev->spinlock);
 102
 103         if (to_free)
 104                 kfree_rcu(to_free, rcu);
 105 }
 106
 107 int rds_ib_update_ipaddr(struct rds_ib_device *rds_ibdev,
 108                          struct in6_addr *ipaddr)
 109 {
 110         struct rds_ib_device *rds_ibdev_old;
 111
 112         rds_ibdev_old = rds_ib_get_device(ipaddr->s6_addr32[3]);
 113         if (!rds_ibdev_old)
 114                 return rds_ib_add_ipaddr(rds_ibdev, ipaddr->s6_addr32[3]);
 115
 116         if (rds_ibdev_old != rds_ibdev) {
 117                 rds_ib_remove_ipaddr(rds_ibdev_old, ipaddr->s6_addr32[3]);
 118                 rds_ib_dev_put(rds_ibdev_old);
 119                 return rds_ib_add_ipaddr(rds_ibdev, ipaddr->s6_addr32[3]);
 120         }
 121         rds_ib_dev_put(rds_ibdev_old);
 122
 123         return 0;
 124 }
 125
 126 void rds_ib_add_conn(struct rds_ib_device *rds_ibdev, struct rds_connection *conn)
 127 {
 128         struct rds_ib_connection *ic = conn->c_transport_data;
 129
 130         /* conn was previously on the nodev_conns_list */
 131         spin_lock_irq(&ib_nodev_conns_lock);
 132         BUG_ON(list_empty(&ib_nodev_conns));
 133         BUG_ON(list_empty(&ic->ib_node));
 134         list_del(&ic->ib_node);
 135
 136         spin_lock(&rds_ibdev->spinlock);
 137         list_add_tail(&ic->ib_node, &rds_ibdev->conn_list);
 138         spin_unlock(&rds_ibdev->spinlock);
 139         spin_unlock_irq(&ib_nodev_conns_lock);
 140
 141         ic->rds_ibdev = rds_ibdev;
 142         refcount_inc(&rds_ibdev->refcount);
 143 }
 144
 145 void rds_ib_remove_conn(struct rds_ib_device *rds_ibdev, struct rds_connection *conn)
 146 {
 147         struct rds_ib_connection *ic = conn->c_transport_data;
 148
 149         /* place conn on nodev_conns_list */
 150         spin_lock(&ib_nodev_conns_lock);
 151
 152         spin_lock_irq(&rds_ibdev->spinlock);
 153         BUG_ON(list_empty(&ic->ib_node));
 154         list_del(&ic->ib_node);
 155         spin_unlock_irq(&rds_ibdev->spinlock);
 156
 157         list_add_tail(&ic->ib_node, &ib_nodev_conns);
 158
 159         spin_unlock(&ib_nodev_conns_lock);
 160
 161         ic->rds_ibdev = NULL;
 162         rds_ib_dev_put(rds_ibdev);
 163 }
 164
 165 void rds_ib_destroy_nodev_conns(void)
 166 {
 167         struct rds_ib_connection *ic, *_ic;
 168         LIST_HEAD(tmp_list);
 169
 170         /* avoid calling conn_destroy with irqs off */
 171         spin_lock_irq(&ib_nodev_conns_lock);
 172         list_splice(&ib_nodev_conns, &tmp_list);
 173         spin_unlock_irq(&ib_nodev_conns_lock);
 174
 175         list_for_each_entry_safe(ic, _ic, &tmp_list, ib_node)
 176                 rds_conn_destroy(ic->conn);
 177 }
 178
 179 void rds_ib_get_mr_info(struct rds_ib_device *rds_ibdev, struct rds_info_rdma_connection *iinfo)
 180 {
 181         struct rds_ib_mr_pool *pool_1m = rds_ibdev->mr_1m_pool;
 182
 183         iinfo->rdma_mr_max = pool_1m->max_items;
 184         iinfo->rdma_mr_size = pool_1m->fmr_attr.max_pages;
 185 }
 186
 187 #if IS_ENABLED(CONFIG_IPV6)
 188 void rds6_ib_get_mr_info(struct rds_ib_device *rds_ibdev,
 189                          struct rds6_info_rdma_connection *iinfo6)
 190 {
 191         struct rds_ib_mr_pool *pool_1m = rds_ibdev->mr_1m_pool;
 192
 193         iinfo6->rdma_mr_max = pool_1m->max_items;
 194         iinfo6->rdma_mr_size = pool_1m->fmr_attr.max_pages;
 195 }
 196 #endif
 197
 198 struct rds_ib_mr *rds_ib_reuse_mr(struct rds_ib_mr_pool *pool)
 199 {
 200         struct rds_ib_mr *ibmr = NULL;
 201         struct llist_node *ret;
 202         unsigned long flags;
 203
 204         spin_lock_irqsave(&pool->clean_lock, flags);
 205         ret = llist_del_first(&pool->clean_list);
 206         spin_unlock_irqrestore(&pool->clean_lock, flags);
 207         if (ret) {
 208                 ibmr = llist_entry(ret, struct rds_ib_mr, llnode);
 209                 if (pool->pool_type == RDS_IB_MR_8K_POOL)
 210                         rds_ib_stats_inc(s_ib_rdma_mr_8k_reused);
 211                 else
 212                         rds_ib_stats_inc(s_ib_rdma_mr_1m_reused);
 213         }
 214
 215         return ibmr;
 216 }
 217
 218 void rds_ib_sync_mr(void *trans_private, int direction)
 219 {
 220         struct rds_ib_mr *ibmr = trans_private;
 221         struct rds_ib_device *rds_ibdev = ibmr->device;
 222
 223         if (ibmr->odp)
 224                 return;
 225
 226         switch (direction) {
 227         case DMA_FROM_DEVICE:
 228                 ib_dma_sync_sg_for_cpu(rds_ibdev->dev, ibmr->sg,
 229                         ibmr->sg_dma_len, DMA_BIDIRECTIONAL);
 230                 break;
 231         case DMA_TO_DEVICE:
 232                 ib_dma_sync_sg_for_device(rds_ibdev->dev, ibmr->sg,
 233                         ibmr->sg_dma_len, DMA_BIDIRECTIONAL);
 234                 break;
 235         }
 236 }
 237
 238 void __rds_ib_teardown_mr(struct rds_ib_mr *ibmr)
 239 {
 240         struct rds_ib_device *rds_ibdev = ibmr->device;
 241
 242         if (ibmr->sg_dma_len) {
 243                 ib_dma_unmap_sg(rds_ibdev->dev,
 244                                 ibmr->sg, ibmr->sg_len,
 245                                 DMA_BIDIRECTIONAL);
 246                 ibmr->sg_dma_len = 0;
 247         }
 248
 249         /* Release the s/g list */
 250         if (ibmr->sg_len) {
 251                 unsigned int i;
 252
 253                 for (i = 0; i < ibmr->sg_len; ++i) {
 254                         struct page *page = sg_page(&ibmr->sg[i]);
 255
 256                         /* FIXME we need a way to tell a r/w MR
 257                          * from a r/o MR */
 258                         WARN_ON(!page->mapping && irqs_disabled());
 259                         set_page_dirty(page);
 260                         put_page(page);
 261                 }
 262                 kfree(ibmr->sg);
 263
 264                 ibmr->sg = NULL;
 265                 ibmr->sg_len = 0;
 266         }
 267 }
 268
 269 void rds_ib_teardown_mr(struct rds_ib_mr *ibmr)
 270 {
 271         unsigned int pinned = ibmr->sg_len;
 272
 273         __rds_ib_teardown_mr(ibmr);
 274         if (pinned) {
 275                 struct rds_ib_mr_pool *pool = ibmr->pool;
 276
 277                 atomic_sub(pinned, &pool->free_pinned);
 278         }
 279 }
 280
 281 static inline unsigned int rds_ib_flush_goal(struct rds_ib_mr_pool *pool, int free_all)
 282 {
 283         unsigned int item_count;
 284
 285         item_count = atomic_read(&pool->item_count);
 286         if (free_all)
 287                 return item_count;
 288
 289         return 0;
 290 }
 291
 292 /*
 293  * given an llist of mrs, put them all into the list_head for more processing
 294  */
 295 static unsigned int llist_append_to_list(struct llist_head *llist,
 296                                          struct list_head *list)
 297 {
 298         struct rds_ib_mr *ibmr;
 299         struct llist_node *node;
 300         struct llist_node *next;
 301         unsigned int count = 0;
 302
 303         node = llist_del_all(llist);
 304         while (node) {
 305                 next = node->next;
 306                 ibmr = llist_entry(node, struct rds_ib_mr, llnode);
 307                 list_add_tail(&ibmr->unmap_list, list);
 308                 node = next;
 309                 count++;
 310         }
 311         return count;
 312 }
 313
 314 /*
 315  * this takes a list head of mrs and turns it into linked llist nodes
 316  * of clusters.  Each cluster has linked llist nodes of
 317  * MR_CLUSTER_SIZE mrs that are ready for reuse.
 318  */
 319 static void list_to_llist_nodes(struct list_head *list,
 320                                 struct llist_node **nodes_head,
 321                                 struct llist_node **nodes_tail)
 322 {
 323         struct rds_ib_mr *ibmr;
 324         struct llist_node *cur = NULL;
 325         struct llist_node **next = nodes_head;
 326
 327         list_for_each_entry(ibmr, list, unmap_list) {
 328                 cur = &ibmr->llnode;
 329                 *next = cur;
 330                 next = &cur->next;
 331         }
 332         *next = NULL;
 333         *nodes_tail = cur;
 334 }
 335
 336 /*
 337  * Flush our pool of MRs.
 338  * At a minimum, all currently unused MRs are unmapped.
 339  * If the number of MRs allocated exceeds the limit, we also try
 340  * to free as many MRs as needed to get back to this limit.
 341  */
 342 int rds_ib_flush_mr_pool(struct rds_ib_mr_pool *pool,
 343                          int free_all, struct rds_ib_mr **ibmr_ret)
 344 {
 345         struct rds_ib_mr *ibmr;
 346         struct llist_node *clean_nodes;
 347         struct llist_node *clean_tail;
 348         LIST_HEAD(unmap_list);
 349         unsigned long unpinned = 0;
 350         unsigned int nfreed = 0, dirty_to_clean = 0, free_goal;
 351
 352         if (pool->pool_type == RDS_IB_MR_8K_POOL)
 353                 rds_ib_stats_inc(s_ib_rdma_mr_8k_pool_flush);
 354         else
 355                 rds_ib_stats_inc(s_ib_rdma_mr_1m_pool_flush);
 356
 357         if (ibmr_ret) {
 358                 DEFINE_WAIT(wait);
 359                 while (!mutex_trylock(&pool->flush_lock)) {
 360                         ibmr = rds_ib_reuse_mr(pool);
 361                         if (ibmr) {
 362                                 *ibmr_ret = ibmr;
 363                                 finish_wait(&pool->flush_wait, &wait);
 364                                 goto out_nolock;
 365                         }
 366
 367                         prepare_to_wait(&pool->flush_wait, &wait,
 368                                         TASK_UNINTERRUPTIBLE);
 369                         if (llist_empty(&pool->clean_list))
 370                                 schedule();
 371
 372                         ibmr = rds_ib_reuse_mr(pool);
 373                         if (ibmr) {
 374                                 *ibmr_ret = ibmr;
 375                                 finish_wait(&pool->flush_wait, &wait);
 376                                 goto out_nolock;
 377                         }
 378                 }
 379                 finish_wait(&pool->flush_wait, &wait);
 380         } else
 381                 mutex_lock(&pool->flush_lock);
 382
 383         if (ibmr_ret) {
 384                 ibmr = rds_ib_reuse_mr(pool);
 385                 if (ibmr) {
 386                         *ibmr_ret = ibmr;
 387                         goto out;
 388                 }
 389         }
 390
 391         /* Get the list of all MRs to be dropped. Ordering matters -
 392          * we want to put drop_list ahead of free_list.
 393          */
 394         dirty_to_clean = llist_append_to_list(&pool->drop_list, &unmap_list);
 395         dirty_to_clean += llist_append_to_list(&pool->free_list, &unmap_list);
 396         if (free_all) {
 397                 unsigned long flags;
 398
 399                 spin_lock_irqsave(&pool->clean_lock, flags);
 400                 llist_append_to_list(&pool->clean_list, &unmap_list);
 401                 spin_unlock_irqrestore(&pool->clean_lock, flags);
 402         }
 403
 404         free_goal = rds_ib_flush_goal(pool, free_all);
 405
 406         if (list_empty(&unmap_list))
 407                 goto out;
 408
 409         if (pool->use_fastreg)
 410                 rds_ib_unreg_frmr(&unmap_list, &nfreed, &unpinned, free_goal);
 411         else
 412                 rds_ib_unreg_fmr(&unmap_list, &nfreed, &unpinned, free_goal);
 413
 414         if (!list_empty(&unmap_list)) {
 415                 unsigned long flags;
 416
 417                 list_to_llist_nodes(&unmap_list, &clean_nodes, &clean_tail);
 418                 if (ibmr_ret) {
 419                         *ibmr_ret = llist_entry(clean_nodes, struct rds_ib_mr, llnode);
 420                         clean_nodes = clean_nodes->next;
 421                 }
 422                 /* more than one entry in llist nodes */
 423                 if (clean_nodes) {
 424                         spin_lock_irqsave(&pool->clean_lock, flags);
 425                         llist_add_batch(clean_nodes, clean_tail,
 426                                         &pool->clean_list);
 427                         spin_unlock_irqrestore(&pool->clean_lock, flags);
 428                 }
 429         }
 430
 431         atomic_sub(unpinned, &pool->free_pinned);
 432         atomic_sub(dirty_to_clean, &pool->dirty_count);
 433         atomic_sub(nfreed, &pool->item_count);
 434
 435 out:
 436         mutex_unlock(&pool->flush_lock);
 437         if (waitqueue_active(&pool->flush_wait))
 438                 wake_up(&pool->flush_wait);
 439 out_nolock:
 440         return 0;
 441 }
 442
 443 struct rds_ib_mr *rds_ib_try_reuse_ibmr(struct rds_ib_mr_pool *pool)
 444 {
 445         struct rds_ib_mr *ibmr = NULL;
 446         int iter = 0;
 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                         break;
 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 NULL;
 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         if (ibmr->odp) {
 496                 /* A MR created and marked as use_once. We use delayed work,
 497                  * because there is a change that we are in interrupt and can't
 498                  * call to ib_dereg_mr() directly.
 499                  */
 500                 INIT_DELAYED_WORK(&ibmr->work, rds_ib_odp_mr_worker);
 501                 queue_delayed_work(rds_ib_mr_wq, &ibmr->work, 0);
 502                 return;
 503         }
 504
 505         /* Return it to the pool's free list */
 506         if (rds_ibdev->use_fastreg)
 507                 rds_ib_free_frmr_list(ibmr);
 508         else
 509                 rds_ib_free_fmr_list(ibmr);
 510
 511         atomic_add(ibmr->sg_len, &pool->free_pinned);
 512         atomic_inc(&pool->dirty_count);
 513
 514         /* If we've pinned too many pages, request a flush */
 515         if (atomic_read(&pool->free_pinned) >= pool->max_free_pinned ||
 516             atomic_read(&pool->dirty_count) >= pool->max_items / 5)
 517                 queue_delayed_work(rds_ib_mr_wq, &pool->flush_worker, 10);
 518
 519         if (invalidate) {
 520                 if (likely(!in_interrupt())) {
 521                         rds_ib_flush_mr_pool(pool, 0, NULL);
 522                 } else {
 523                         /* We get here if the user created a MR marked
 524                          * as use_once and invalidate at the same time.
 525                          */
 526                         queue_delayed_work(rds_ib_mr_wq,
 527                                            &pool->flush_worker, 10);
 528                 }
 529         }
 530
 531         rds_ib_dev_put(rds_ibdev);
 532 }
 533
 534 void rds_ib_flush_mrs(void)
 535 {
 536         struct rds_ib_device *rds_ibdev;
 537
 538         down_read(&rds_ib_devices_lock);
 539         list_for_each_entry(rds_ibdev, &rds_ib_devices, list) {
 540                 if (rds_ibdev->mr_8k_pool)
 541                         rds_ib_flush_mr_pool(rds_ibdev->mr_8k_pool, 0, NULL);
 542
 543                 if (rds_ibdev->mr_1m_pool)
 544                         rds_ib_flush_mr_pool(rds_ibdev->mr_1m_pool, 0, NULL);
 545         }
 546         up_read(&rds_ib_devices_lock);
 547 }
 548
 549 u32 rds_ib_get_lkey(void *trans_private)
 550 {
 551         struct rds_ib_mr *ibmr = trans_private;
 552
 553         return ibmr->u.mr->lkey;
 554 }
 555
 556 void *rds_ib_get_mr(struct scatterlist *sg, unsigned long nents,
 557                     struct rds_sock *rs, u32 *key_ret,
 558                     struct rds_connection *conn,
 559                     u64 start, u64 length, int need_odp)
 560 {
 561         struct rds_ib_device *rds_ibdev;
 562         struct rds_ib_mr *ibmr = NULL;
 563         struct rds_ib_connection *ic = NULL;
 564         int ret;
 565
 566         rds_ibdev = rds_ib_get_device(rs->rs_bound_addr.s6_addr32[3]);
 567         if (!rds_ibdev) {
 568                 ret = -ENODEV;
 569                 goto out;
 570         }
 571
 572         if (need_odp == ODP_ZEROBASED || need_odp == ODP_VIRTUAL) {
 573                 u64 virt_addr = need_odp == ODP_ZEROBASED ? 0 : start;
 574                 int access_flags =
 575                         (IB_ACCESS_LOCAL_WRITE | IB_ACCESS_REMOTE_READ |
 576                          IB_ACCESS_REMOTE_WRITE | IB_ACCESS_REMOTE_ATOMIC |
 577                          IB_ACCESS_ON_DEMAND);
 578                 struct ib_sge sge = {};
 579                 struct ib_mr *ib_mr;
 580
 581                 if (!rds_ibdev->odp_capable) {
 582                         ret = -EOPNOTSUPP;
 583                         goto out;
 584                 }
 585
 586                 ib_mr = ib_reg_user_mr(rds_ibdev->pd, start, length, virt_addr,
 587                                        access_flags);
 588
 589                 if (IS_ERR(ib_mr)) {
 590                         rdsdebug("rds_ib_get_user_mr returned %d\n",
 591                                  IS_ERR(ib_mr));
 592                         ret = PTR_ERR(ib_mr);
 593                         goto out;
 594                 }
 595                 if (key_ret)
 596                         *key_ret = ib_mr->rkey;
 597
 598                 ibmr = kzalloc(sizeof(*ibmr), GFP_KERNEL);
 599                 if (!ibmr) {
 600                         ib_dereg_mr(ib_mr);
 601                         ret = -ENOMEM;
 602                         goto out;
 603                 }
 604                 ibmr->u.mr = ib_mr;
 605                 ibmr->odp = 1;
 606
 607                 sge.addr = virt_addr;
 608                 sge.length = length;
 609                 sge.lkey = ib_mr->lkey;
 610
 611                 ib_advise_mr(rds_ibdev->pd,
 612                              IB_UVERBS_ADVISE_MR_ADVICE_PREFETCH_WRITE,
 613                              IB_UVERBS_ADVISE_MR_FLAG_FLUSH, &sge, 1);
 614                 return ibmr;
 615         }
 616
 617         if (conn)
 618                 ic = conn->c_transport_data;
 619
 620         if (!rds_ibdev->mr_8k_pool || !rds_ibdev->mr_1m_pool) {
 621                 ret = -ENODEV;
 622                 goto out;
 623         }
 624
 625         if (rds_ibdev->use_fastreg)
 626                 ibmr = rds_ib_reg_frmr(rds_ibdev, ic, sg, nents, key_ret);
 627         else
 628                 ibmr = rds_ib_reg_fmr(rds_ibdev, sg, nents, key_ret);
 629         if (IS_ERR(ibmr)) {
 630                 ret = PTR_ERR(ibmr);
 631                 pr_warn("RDS/IB: rds_ib_get_mr failed (errno=%d)\n", ret);
 632         } else {
 633                 return ibmr;
 634         }
 635
 636  out:
 637         if (rds_ibdev)
 638                 rds_ib_dev_put(rds_ibdev);
 639
 640         return ERR_PTR(ret);
 641 }
 642
 643 void rds_ib_destroy_mr_pool(struct rds_ib_mr_pool *pool)
 644 {
 645         cancel_delayed_work_sync(&pool->flush_worker);
 646         rds_ib_flush_mr_pool(pool, 1, NULL);
 647         WARN_ON(atomic_read(&pool->item_count));
 648         WARN_ON(atomic_read(&pool->free_pinned));
 649         kfree(pool);
 650 }
 651
 652 struct rds_ib_mr_pool *rds_ib_create_mr_pool(struct rds_ib_device *rds_ibdev,
 653                                              int pool_type)
 654 {
 655         struct rds_ib_mr_pool *pool;
 656
 657         pool = kzalloc(sizeof(*pool), GFP_KERNEL);
 658         if (!pool)
 659                 return ERR_PTR(-ENOMEM);
 660
 661         pool->pool_type = pool_type;
 662         init_llist_head(&pool->free_list);
 663         init_llist_head(&pool->drop_list);
 664         init_llist_head(&pool->clean_list);
 665         spin_lock_init(&pool->clean_lock);
 666         mutex_init(&pool->flush_lock);
 667         init_waitqueue_head(&pool->flush_wait);
 668         INIT_DELAYED_WORK(&pool->flush_worker, rds_ib_mr_pool_flush_worker);
 669
 670         if (pool_type == RDS_IB_MR_1M_POOL) {
 671                 /* +1 allows for unaligned MRs */
 672                 pool->fmr_attr.max_pages = RDS_MR_1M_MSG_SIZE + 1;
 673                 pool->max_items = rds_ibdev->max_1m_mrs;
 674         } else {
 675                 /* pool_type == RDS_IB_MR_8K_POOL */
 676                 pool->fmr_attr.max_pages = RDS_MR_8K_MSG_SIZE + 1;
 677                 pool->max_items = rds_ibdev->max_8k_mrs;
 678         }
 679
 680         pool->max_free_pinned = pool->max_items * pool->fmr_attr.max_pages / 4;
 681         pool->fmr_attr.max_maps = rds_ibdev->fmr_max_remaps;
 682         pool->fmr_attr.page_shift = PAGE_SHIFT;
 683         pool->max_items_soft = rds_ibdev->max_mrs * 3 / 4;
 684         pool->use_fastreg = rds_ibdev->use_fastreg;
 685
 686         return pool;
 687 }
 688
 689 int rds_ib_mr_init(void)
 690 {
 691         rds_ib_mr_wq = alloc_workqueue("rds_mr_flushd", WQ_MEM_RECLAIM, 0);
 692         if (!rds_ib_mr_wq)
 693                 return -ENOMEM;
 694         return 0;
 695 }
 696
 697 /* By the time this is called all the IB devices should have been torn down and
 698  * had their pools freed.  As each pool is freed its work struct is waited on,
 699  * so the pool flushing work queue should be idle by the time we get here.
 700  */
 701 void rds_ib_mr_exit(void)
 702 {
 703         destroy_workqueue(rds_ib_mr_wq);
 704 }
 705
 706 static void rds_ib_odp_mr_worker(struct work_struct  *work)
 707 {
 708         struct rds_ib_mr *ibmr;
 709
 710         ibmr = container_of(work, struct rds_ib_mr, work.work);
 711         ib_dereg_mr(ibmr->u.mr);
 712         kfree(ibmr);
 713 }