kernel/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 #include <linux/delay.h>
  38
  39 #include "rds.h"
  40 #include "ib.h"
  41
  42 static DEFINE_PER_CPU(unsigned long, clean_list_grace);
  43 #define CLEAN_LIST_BUSY_BIT 0
  44
  45 /*
  46  * This is stored as mr->r_trans_private.
  47  */
  48 struct rds_ib_mr {
  49         struct rds_ib_device    *device;
  50         struct rds_ib_mr_pool   *pool;
  51         struct ib_fmr           *fmr;
  52
  53         struct llist_node       llnode;
  54
  55         /* unmap_list is for freeing */
  56         struct list_head        unmap_list;
  57         unsigned int            remap_count;
  58
  59         struct scatterlist      *sg;
  60         unsigned int            sg_len;
  61         u64                     *dma;
  62         int                     sg_dma_len;
  63 };
  64
  65 /*
  66  * Our own little FMR pool
  67  */
  68 struct rds_ib_mr_pool {
  69         struct mutex            flush_lock;             /* serialize fmr invalidate */
  70         struct delayed_work     flush_worker;           /* flush worker */
  71
  72         atomic_t                item_count;             /* total # of MRs */
  73         atomic_t                dirty_count;            /* # dirty of MRs */
  74
  75         struct llist_head       drop_list;              /* MRs that have reached their max_maps limit */
  76         struct llist_head       free_list;              /* unused MRs */
  77         struct llist_head       clean_list;             /* global unused & unamapped MRs */
  78         wait_queue_head_t       flush_wait;
  79
  80         atomic_t                free_pinned;            /* memory pinned by free MRs */
  81         unsigned long           max_items;
  82         unsigned long           max_items_soft;
  83         unsigned long           max_free_pinned;
  84         struct ib_fmr_attr      fmr_attr;
  85 };
  86
  87 static int rds_ib_flush_mr_pool(struct rds_ib_mr_pool *pool, int free_all, struct rds_ib_mr **);
  88 static void rds_ib_teardown_mr(struct rds_ib_mr *ibmr);
  89 static void rds_ib_mr_pool_flush_worker(struct work_struct *work);
  90
  91 static struct rds_ib_device *rds_ib_get_device(__be32 ipaddr)
  92 {
  93         struct rds_ib_device *rds_ibdev;
  94         struct rds_ib_ipaddr *i_ipaddr;
  95
  96         rcu_read_lock();
  97         list_for_each_entry_rcu(rds_ibdev, &rds_ib_devices, list) {
  98                 list_for_each_entry_rcu(i_ipaddr, &rds_ibdev->ipaddr_list, list) {
  99                         if (i_ipaddr->ipaddr == ipaddr) {
 100                                 atomic_inc(&rds_ibdev->refcount);
 101                                 rcu_read_unlock();
 102                                 return rds_ibdev;
 103                         }
 104                 }
 105         }
 106         rcu_read_unlock();
 107
 108         return NULL;
 109 }
 110
 111 static int rds_ib_add_ipaddr(struct rds_ib_device *rds_ibdev, __be32 ipaddr)
 112 {
 113         struct rds_ib_ipaddr *i_ipaddr;
 114
 115         i_ipaddr = kmalloc(sizeof *i_ipaddr, GFP_KERNEL);
 116         if (!i_ipaddr)
 117                 return -ENOMEM;
 118
 119         i_ipaddr->ipaddr = ipaddr;
 120
 121         spin_lock_irq(&rds_ibdev->spinlock);
 122         list_add_tail_rcu(&i_ipaddr->list, &rds_ibdev->ipaddr_list);
 123         spin_unlock_irq(&rds_ibdev->spinlock);
 124
 125         return 0;
 126 }
 127
 128 static void rds_ib_remove_ipaddr(struct rds_ib_device *rds_ibdev, __be32 ipaddr)
 129 {
 130         struct rds_ib_ipaddr *i_ipaddr;
 131         struct rds_ib_ipaddr *to_free = NULL;
 132
 133
 134         spin_lock_irq(&rds_ibdev->spinlock);
 135         list_for_each_entry_rcu(i_ipaddr, &rds_ibdev->ipaddr_list, list) {
 136                 if (i_ipaddr->ipaddr == ipaddr) {
 137                         list_del_rcu(&i_ipaddr->list);
 138                         to_free = i_ipaddr;
 139                         break;
 140                 }
 141         }
 142         spin_unlock_irq(&rds_ibdev->spinlock);
 143
 144         if (to_free) {
 145                 synchronize_rcu();
 146                 kfree(to_free);
 147         }
 148 }
 149
 150 int rds_ib_update_ipaddr(struct rds_ib_device *rds_ibdev, __be32 ipaddr)
 151 {
 152         struct rds_ib_device *rds_ibdev_old;
 153
 154         rds_ibdev_old = rds_ib_get_device(ipaddr);
 155         if (rds_ibdev_old) {
 156                 rds_ib_remove_ipaddr(rds_ibdev_old, ipaddr);
 157                 rds_ib_dev_put(rds_ibdev_old);
 158         }
 159
 160         return rds_ib_add_ipaddr(rds_ibdev, ipaddr);
 161 }
 162
 163 void rds_ib_add_conn(struct rds_ib_device *rds_ibdev, struct rds_connection *conn)
 164 {
 165         struct rds_ib_connection *ic = conn->c_transport_data;
 166
 167         /* conn was previously on the nodev_conns_list */
 168         spin_lock_irq(&ib_nodev_conns_lock);
 169         BUG_ON(list_empty(&ib_nodev_conns));
 170         BUG_ON(list_empty(&ic->ib_node));
 171         list_del(&ic->ib_node);
 172
 173         spin_lock(&rds_ibdev->spinlock);
 174         list_add_tail(&ic->ib_node, &rds_ibdev->conn_list);
 175         spin_unlock(&rds_ibdev->spinlock);
 176         spin_unlock_irq(&ib_nodev_conns_lock);
 177
 178         ic->rds_ibdev = rds_ibdev;
 179         atomic_inc(&rds_ibdev->refcount);
 180 }
 181
 182 void rds_ib_remove_conn(struct rds_ib_device *rds_ibdev, struct rds_connection *conn)
 183 {
 184         struct rds_ib_connection *ic = conn->c_transport_data;
 185
 186         /* place conn on nodev_conns_list */
 187         spin_lock(&ib_nodev_conns_lock);
 188
 189         spin_lock_irq(&rds_ibdev->spinlock);
 190         BUG_ON(list_empty(&ic->ib_node));
 191         list_del(&ic->ib_node);
 192         spin_unlock_irq(&rds_ibdev->spinlock);
 193
 194         list_add_tail(&ic->ib_node, &ib_nodev_conns);
 195
 196         spin_unlock(&ib_nodev_conns_lock);
 197
 198         ic->rds_ibdev = NULL;
 199         rds_ib_dev_put(rds_ibdev);
 200 }
 201
 202 void rds_ib_destroy_nodev_conns(void)
 203 {
 204         struct rds_ib_connection *ic, *_ic;
 205         LIST_HEAD(tmp_list);
 206
 207         /* avoid calling conn_destroy with irqs off */
 208         spin_lock_irq(&ib_nodev_conns_lock);
 209         list_splice(&ib_nodev_conns, &tmp_list);
 210         spin_unlock_irq(&ib_nodev_conns_lock);
 211
 212         list_for_each_entry_safe(ic, _ic, &tmp_list, ib_node)
 213                 rds_conn_destroy(ic->conn);
 214 }
 215
 216 struct rds_ib_mr_pool *rds_ib_create_mr_pool(struct rds_ib_device *rds_ibdev)
 217 {
 218         struct rds_ib_mr_pool *pool;
 219
 220         pool = kzalloc(sizeof(*pool), GFP_KERNEL);
 221         if (!pool)
 222                 return ERR_PTR(-ENOMEM);
 223
 224         init_llist_head(&pool->free_list);
 225         init_llist_head(&pool->drop_list);
 226         init_llist_head(&pool->clean_list);
 227         mutex_init(&pool->flush_lock);
 228         init_waitqueue_head(&pool->flush_wait);
 229         INIT_DELAYED_WORK(&pool->flush_worker, rds_ib_mr_pool_flush_worker);
 230
 231         pool->fmr_attr.max_pages = fmr_message_size;
 232         pool->fmr_attr.max_maps = rds_ibdev->fmr_max_remaps;
 233         pool->fmr_attr.page_shift = PAGE_SHIFT;
 234         pool->max_free_pinned = rds_ibdev->max_fmrs * fmr_message_size / 4;
 235
 236         /* We never allow more than max_items MRs to be allocated.
 237          * When we exceed more than max_items_soft, we start freeing
 238          * items more aggressively.
 239          * Make sure that max_items > max_items_soft > max_items / 2
 240          */
 241         pool->max_items_soft = rds_ibdev->max_fmrs * 3 / 4;
 242         pool->max_items = rds_ibdev->max_fmrs;
 243
 244         return pool;
 245 }
 246
 247 void rds_ib_get_mr_info(struct rds_ib_device *rds_ibdev, struct rds_info_rdma_connection *iinfo)
 248 {
 249         struct rds_ib_mr_pool *pool = rds_ibdev->mr_pool;
 250
 251         iinfo->rdma_mr_max = pool->max_items;
 252         iinfo->rdma_mr_size = pool->fmr_attr.max_pages;
 253 }
 254
 255 void rds_ib_destroy_mr_pool(struct rds_ib_mr_pool *pool)
 256 {
 257         cancel_delayed_work_sync(&pool->flush_worker);
 258         rds_ib_flush_mr_pool(pool, 1, NULL);
 259         WARN_ON(atomic_read(&pool->item_count));
 260         WARN_ON(atomic_read(&pool->free_pinned));
 261         kfree(pool);
 262 }
 263
 264 static inline struct rds_ib_mr *rds_ib_reuse_fmr(struct rds_ib_mr_pool *pool)
 265 {
 266         struct rds_ib_mr *ibmr = NULL;
 267         struct llist_node *ret;
 268         unsigned long *flag;
 269
 270         preempt_disable();
 271         flag = this_cpu_ptr(&clean_list_grace);
 272         set_bit(CLEAN_LIST_BUSY_BIT, flag);
 273         ret = llist_del_first(&pool->clean_list);
 274         if (ret)
 275                 ibmr = llist_entry(ret, struct rds_ib_mr, llnode);
 276
 277         clear_bit(CLEAN_LIST_BUSY_BIT, flag);
 278         preempt_enable();
 279         return ibmr;
 280 }
 281
 282 static inline void wait_clean_list_grace(void)
 283 {
 284         int cpu;
 285         unsigned long *flag;
 286
 287         for_each_online_cpu(cpu) {
 288                 flag = &per_cpu(clean_list_grace, cpu);
 289                 while (test_bit(CLEAN_LIST_BUSY_BIT, flag))
 290                         cpu_chill();
 291         }
 292 }
 293
 294 static struct rds_ib_mr *rds_ib_alloc_fmr(struct rds_ib_device *rds_ibdev)
 295 {
 296         struct rds_ib_mr_pool *pool = rds_ibdev->mr_pool;
 297         struct rds_ib_mr *ibmr = NULL;
 298         int err = 0, iter = 0;
 299
 300         if (atomic_read(&pool->dirty_count) >= pool->max_items / 10)
 301                 schedule_delayed_work(&pool->flush_worker, 10);
 302
 303         while (1) {
 304                 ibmr = rds_ib_reuse_fmr(pool);
 305                 if (ibmr)
 306                         return ibmr;
 307
 308                 /* No clean MRs - now we have the choice of either
 309                  * allocating a fresh MR up to the limit imposed by the
 310                  * driver, or flush any dirty unused MRs.
 311                  * We try to avoid stalling in the send path if possible,
 312                  * so we allocate as long as we're allowed to.
 313                  *
 314                  * We're fussy with enforcing the FMR limit, though. If the driver
 315                  * tells us we can't use more than N fmrs, we shouldn't start
 316                  * arguing with it */
 317                 if (atomic_inc_return(&pool->item_count) <= pool->max_items)
 318                         break;
 319
 320                 atomic_dec(&pool->item_count);
 321
 322                 if (++iter > 2) {
 323                         rds_ib_stats_inc(s_ib_rdma_mr_pool_depleted);
 324                         return ERR_PTR(-EAGAIN);
 325                 }
 326
 327                 /* We do have some empty MRs. Flush them out. */
 328                 rds_ib_stats_inc(s_ib_rdma_mr_pool_wait);
 329                 rds_ib_flush_mr_pool(pool, 0, &ibmr);
 330                 if (ibmr)
 331                         return ibmr;
 332         }
 333
 334         ibmr = kzalloc_node(sizeof(*ibmr), GFP_KERNEL, rdsibdev_to_node(rds_ibdev));
 335         if (!ibmr) {
 336                 err = -ENOMEM;
 337                 goto out_no_cigar;
 338         }
 339
 340         memset(ibmr, 0, sizeof(*ibmr));
 341
 342         ibmr->fmr = ib_alloc_fmr(rds_ibdev->pd,
 343                         (IB_ACCESS_LOCAL_WRITE |
 344                          IB_ACCESS_REMOTE_READ |
 345                          IB_ACCESS_REMOTE_WRITE|
 346                          IB_ACCESS_REMOTE_ATOMIC),
 347                         &pool->fmr_attr);
 348         if (IS_ERR(ibmr->fmr)) {
 349                 err = PTR_ERR(ibmr->fmr);
 350                 ibmr->fmr = NULL;
 351                 printk(KERN_WARNING "RDS/IB: ib_alloc_fmr failed (err=%d)\n", err);
 352                 goto out_no_cigar;
 353         }
 354
 355         rds_ib_stats_inc(s_ib_rdma_mr_alloc);
 356         return ibmr;
 357
 358 out_no_cigar:
 359         if (ibmr) {
 360                 if (ibmr->fmr)
 361                         ib_dealloc_fmr(ibmr->fmr);
 362                 kfree(ibmr);
 363         }
 364         atomic_dec(&pool->item_count);
 365         return ERR_PTR(err);
 366 }
 367
 368 static int rds_ib_map_fmr(struct rds_ib_device *rds_ibdev, struct rds_ib_mr *ibmr,
 369                struct scatterlist *sg, unsigned int nents)
 370 {
 371         struct ib_device *dev = rds_ibdev->dev;
 372         struct scatterlist *scat = sg;
 373         u64 io_addr = 0;
 374         u64 *dma_pages;
 375         u32 len;
 376         int page_cnt, sg_dma_len;
 377         int i, j;
 378         int ret;
 379
 380         sg_dma_len = ib_dma_map_sg(dev, sg, nents,
 381                                  DMA_BIDIRECTIONAL);
 382         if (unlikely(!sg_dma_len)) {
 383                 printk(KERN_WARNING "RDS/IB: dma_map_sg failed!\n");
 384                 return -EBUSY;
 385         }
 386
 387         len = 0;
 388         page_cnt = 0;
 389
 390         for (i = 0; i < sg_dma_len; ++i) {
 391                 unsigned int dma_len = ib_sg_dma_len(dev, &scat[i]);
 392                 u64 dma_addr = ib_sg_dma_address(dev, &scat[i]);
 393
 394                 if (dma_addr & ~PAGE_MASK) {
 395                         if (i > 0)
 396                                 return -EINVAL;
 397                         else
 398                                 ++page_cnt;
 399                 }
 400                 if ((dma_addr + dma_len) & ~PAGE_MASK) {
 401                         if (i < sg_dma_len - 1)
 402                                 return -EINVAL;
 403                         else
 404                                 ++page_cnt;
 405                 }
 406
 407                 len += dma_len;
 408         }
 409
 410         page_cnt += len >> PAGE_SHIFT;
 411         if (page_cnt > fmr_message_size)
 412                 return -EINVAL;
 413
 414         dma_pages = kmalloc_node(sizeof(u64) * page_cnt, GFP_ATOMIC,
 415                                  rdsibdev_to_node(rds_ibdev));
 416         if (!dma_pages)
 417                 return -ENOMEM;
 418
 419         page_cnt = 0;
 420         for (i = 0; i < sg_dma_len; ++i) {
 421                 unsigned int dma_len = ib_sg_dma_len(dev, &scat[i]);
 422                 u64 dma_addr = ib_sg_dma_address(dev, &scat[i]);
 423
 424                 for (j = 0; j < dma_len; j += PAGE_SIZE)
 425                         dma_pages[page_cnt++] =
 426                                 (dma_addr & PAGE_MASK) + j;
 427         }
 428
 429         ret = ib_map_phys_fmr(ibmr->fmr,
 430                                    dma_pages, page_cnt, io_addr);
 431         if (ret)
 432                 goto out;
 433
 434         /* Success - we successfully remapped the MR, so we can
 435          * safely tear down the old mapping. */
 436         rds_ib_teardown_mr(ibmr);
 437
 438         ibmr->sg = scat;
 439         ibmr->sg_len = nents;
 440         ibmr->sg_dma_len = sg_dma_len;
 441         ibmr->remap_count++;
 442
 443         rds_ib_stats_inc(s_ib_rdma_mr_used);
 444         ret = 0;
 445
 446 out:
 447         kfree(dma_pages);
 448
 449         return ret;
 450 }
 451
 452 void rds_ib_sync_mr(void *trans_private, int direction)
 453 {
 454         struct rds_ib_mr *ibmr = trans_private;
 455         struct rds_ib_device *rds_ibdev = ibmr->device;
 456
 457         switch (direction) {
 458         case DMA_FROM_DEVICE:
 459                 ib_dma_sync_sg_for_cpu(rds_ibdev->dev, ibmr->sg,
 460                         ibmr->sg_dma_len, DMA_BIDIRECTIONAL);
 461                 break;
 462         case DMA_TO_DEVICE:
 463                 ib_dma_sync_sg_for_device(rds_ibdev->dev, ibmr->sg,
 464                         ibmr->sg_dma_len, DMA_BIDIRECTIONAL);
 465                 break;
 466         }
 467 }
 468
 469 static void __rds_ib_teardown_mr(struct rds_ib_mr *ibmr)
 470 {
 471         struct rds_ib_device *rds_ibdev = ibmr->device;
 472
 473         if (ibmr->sg_dma_len) {
 474                 ib_dma_unmap_sg(rds_ibdev->dev,
 475                                 ibmr->sg, ibmr->sg_len,
 476                                 DMA_BIDIRECTIONAL);
 477                 ibmr->sg_dma_len = 0;
 478         }
 479
 480         /* Release the s/g list */
 481         if (ibmr->sg_len) {
 482                 unsigned int i;
 483
 484                 for (i = 0; i < ibmr->sg_len; ++i) {
 485                         struct page *page = sg_page(&ibmr->sg[i]);
 486
 487                         /* FIXME we need a way to tell a r/w MR
 488                          * from a r/o MR */
 489                         BUG_ON(irqs_disabled());
 490                         set_page_dirty(page);
 491                         put_page(page);
 492                 }
 493                 kfree(ibmr->sg);
 494
 495                 ibmr->sg = NULL;
 496                 ibmr->sg_len = 0;
 497         }
 498 }
 499
 500 static void rds_ib_teardown_mr(struct rds_ib_mr *ibmr)
 501 {
 502         unsigned int pinned = ibmr->sg_len;
 503
 504         __rds_ib_teardown_mr(ibmr);
 505         if (pinned) {
 506                 struct rds_ib_device *rds_ibdev = ibmr->device;
 507                 struct rds_ib_mr_pool *pool = rds_ibdev->mr_pool;
 508
 509                 atomic_sub(pinned, &pool->free_pinned);
 510         }
 511 }
 512
 513 static inline unsigned int rds_ib_flush_goal(struct rds_ib_mr_pool *pool, int free_all)
 514 {
 515         unsigned int item_count;
 516
 517         item_count = atomic_read(&pool->item_count);
 518         if (free_all)
 519                 return item_count;
 520
 521         return 0;
 522 }
 523
 524 /*
 525  * given an llist of mrs, put them all into the list_head for more processing
 526  */
 527 static void llist_append_to_list(struct llist_head *llist, struct list_head *list)
 528 {
 529         struct rds_ib_mr *ibmr;
 530         struct llist_node *node;
 531         struct llist_node *next;
 532
 533         node = llist_del_all(llist);
 534         while (node) {
 535                 next = node->next;
 536                 ibmr = llist_entry(node, struct rds_ib_mr, llnode);
 537                 list_add_tail(&ibmr->unmap_list, list);
 538                 node = next;
 539         }
 540 }
 541
 542 /*
 543  * this takes a list head of mrs and turns it into linked llist nodes
 544  * of clusters.  Each cluster has linked llist nodes of
 545  * MR_CLUSTER_SIZE mrs that are ready for reuse.
 546  */
 547 static void list_to_llist_nodes(struct rds_ib_mr_pool *pool,
 548                                 struct list_head *list,
 549                                 struct llist_node **nodes_head,
 550                                 struct llist_node **nodes_tail)
 551 {
 552         struct rds_ib_mr *ibmr;
 553         struct llist_node *cur = NULL;
 554         struct llist_node **next = nodes_head;
 555
 556         list_for_each_entry(ibmr, list, unmap_list) {
 557                 cur = &ibmr->llnode;
 558                 *next = cur;
 559                 next = &cur->next;
 560         }
 561         *next = NULL;
 562         *nodes_tail = cur;
 563 }
 564
 565 /*
 566  * Flush our pool of MRs.
 567  * At a minimum, all currently unused MRs are unmapped.
 568  * If the number of MRs allocated exceeds the limit, we also try
 569  * to free as many MRs as needed to get back to this limit.
 570  */
 571 static int rds_ib_flush_mr_pool(struct rds_ib_mr_pool *pool,
 572                                 int free_all, struct rds_ib_mr **ibmr_ret)
 573 {
 574         struct rds_ib_mr *ibmr, *next;
 575         struct llist_node *clean_nodes;
 576         struct llist_node *clean_tail;
 577         LIST_HEAD(unmap_list);
 578         LIST_HEAD(fmr_list);
 579         unsigned long unpinned = 0;
 580         unsigned int nfreed = 0, ncleaned = 0, free_goal;
 581         int ret = 0;
 582
 583         rds_ib_stats_inc(s_ib_rdma_mr_pool_flush);
 584
 585         if (ibmr_ret) {
 586                 DEFINE_WAIT(wait);
 587                 while(!mutex_trylock(&pool->flush_lock)) {
 588                         ibmr = rds_ib_reuse_fmr(pool);
 589                         if (ibmr) {
 590                                 *ibmr_ret = ibmr;
 591                                 finish_wait(&pool->flush_wait, &wait);
 592                                 goto out_nolock;
 593                         }
 594
 595                         prepare_to_wait(&pool->flush_wait, &wait,
 596                                         TASK_UNINTERRUPTIBLE);
 597                         if (llist_empty(&pool->clean_list))
 598                                 schedule();
 599
 600                         ibmr = rds_ib_reuse_fmr(pool);
 601                         if (ibmr) {
 602                                 *ibmr_ret = ibmr;
 603                                 finish_wait(&pool->flush_wait, &wait);
 604                                 goto out_nolock;
 605                         }
 606                 }
 607                 finish_wait(&pool->flush_wait, &wait);
 608         } else
 609                 mutex_lock(&pool->flush_lock);
 610
 611         if (ibmr_ret) {
 612                 ibmr = rds_ib_reuse_fmr(pool);
 613                 if (ibmr) {
 614                         *ibmr_ret = ibmr;
 615                         goto out;
 616                 }
 617         }
 618
 619         /* Get the list of all MRs to be dropped. Ordering matters -
 620          * we want to put drop_list ahead of free_list.
 621          */
 622         llist_append_to_list(&pool->drop_list, &unmap_list);
 623         llist_append_to_list(&pool->free_list, &unmap_list);
 624         if (free_all)
 625                 llist_append_to_list(&pool->clean_list, &unmap_list);
 626
 627         free_goal = rds_ib_flush_goal(pool, free_all);
 628
 629         if (list_empty(&unmap_list))
 630                 goto out;
 631
 632         /* String all ib_mr's onto one list and hand them to ib_unmap_fmr */
 633         list_for_each_entry(ibmr, &unmap_list, unmap_list)
 634                 list_add(&ibmr->fmr->list, &fmr_list);
 635
 636         ret = ib_unmap_fmr(&fmr_list);
 637         if (ret)
 638                 printk(KERN_WARNING "RDS/IB: ib_unmap_fmr failed (err=%d)\n", ret);
 639
 640         /* Now we can destroy the DMA mapping and unpin any pages */
 641         list_for_each_entry_safe(ibmr, next, &unmap_list, unmap_list) {
 642                 unpinned += ibmr->sg_len;
 643                 __rds_ib_teardown_mr(ibmr);
 644                 if (nfreed < free_goal || ibmr->remap_count >= pool->fmr_attr.max_maps) {
 645                         rds_ib_stats_inc(s_ib_rdma_mr_free);
 646                         list_del(&ibmr->unmap_list);
 647                         ib_dealloc_fmr(ibmr->fmr);
 648                         kfree(ibmr);
 649                         nfreed++;
 650                 }
 651                 ncleaned++;
 652         }
 653
 654         if (!list_empty(&unmap_list)) {
 655                 /* we have to make sure that none of the things we're about
 656                  * to put on the clean list would race with other cpus trying
 657                  * to pull items off.  The llist would explode if we managed to
 658                  * remove something from the clean list and then add it back again
 659                  * while another CPU was spinning on that same item in llist_del_first.
 660                  *
 661                  * This is pretty unlikely, but just in case  wait for an llist grace period
 662                  * here before adding anything back into the clean list.
 663                  */
 664                 wait_clean_list_grace();
 665
 666                 list_to_llist_nodes(pool, &unmap_list, &clean_nodes, &clean_tail);
 667                 if (ibmr_ret)
 668                         *ibmr_ret = llist_entry(clean_nodes, struct rds_ib_mr, llnode);
 669
 670                 /* more than one entry in llist nodes */
 671                 if (clean_nodes->next)
 672                         llist_add_batch(clean_nodes->next, clean_tail, &pool->clean_list);
 673
 674         }
 675
 676         atomic_sub(unpinned, &pool->free_pinned);
 677         atomic_sub(ncleaned, &pool->dirty_count);
 678         atomic_sub(nfreed, &pool->item_count);
 679
 680 out:
 681         mutex_unlock(&pool->flush_lock);
 682         if (waitqueue_active(&pool->flush_wait))
 683                 wake_up(&pool->flush_wait);
 684 out_nolock:
 685         return ret;
 686 }
 687
 688 static void rds_ib_mr_pool_flush_worker(struct work_struct *work)
 689 {
 690         struct rds_ib_mr_pool *pool = container_of(work, struct rds_ib_mr_pool, flush_worker.work);
 691
 692         rds_ib_flush_mr_pool(pool, 0, NULL);
 693 }
 694
 695 void rds_ib_free_mr(void *trans_private, int invalidate)
 696 {
 697         struct rds_ib_mr *ibmr = trans_private;
 698         struct rds_ib_device *rds_ibdev = ibmr->device;
 699         struct rds_ib_mr_pool *pool = rds_ibdev->mr_pool;
 700
 701         rdsdebug("RDS/IB: free_mr nents %u\n", ibmr->sg_len);
 702
 703         /* Return it to the pool's free list */
 704         if (ibmr->remap_count >= pool->fmr_attr.max_maps)
 705                 llist_add(&ibmr->llnode, &pool->drop_list);
 706         else
 707                 llist_add(&ibmr->llnode, &pool->free_list);
 708
 709         atomic_add(ibmr->sg_len, &pool->free_pinned);
 710         atomic_inc(&pool->dirty_count);
 711
 712         /* If we've pinned too many pages, request a flush */
 713         if (atomic_read(&pool->free_pinned) >= pool->max_free_pinned ||
 714             atomic_read(&pool->dirty_count) >= pool->max_items / 10)
 715                 schedule_delayed_work(&pool->flush_worker, 10);
 716
 717         if (invalidate) {
 718                 if (likely(!in_interrupt())) {
 719                         rds_ib_flush_mr_pool(pool, 0, NULL);
 720                 } else {
 721                         /* We get here if the user created a MR marked
 722                          * as use_once and invalidate at the same time. */
 723                         schedule_delayed_work(&pool->flush_worker, 10);
 724                 }
 725         }
 726
 727         rds_ib_dev_put(rds_ibdev);
 728 }
 729
 730 void rds_ib_flush_mrs(void)
 731 {
 732         struct rds_ib_device *rds_ibdev;
 733
 734         down_read(&rds_ib_devices_lock);
 735         list_for_each_entry(rds_ibdev, &rds_ib_devices, list) {
 736                 struct rds_ib_mr_pool *pool = rds_ibdev->mr_pool;
 737
 738                 if (pool)
 739                         rds_ib_flush_mr_pool(pool, 0, NULL);
 740         }
 741         up_read(&rds_ib_devices_lock);
 742 }
 743
 744 void *rds_ib_get_mr(struct scatterlist *sg, unsigned long nents,
 745                     struct rds_sock *rs, u32 *key_ret)
 746 {
 747         struct rds_ib_device *rds_ibdev;
 748         struct rds_ib_mr *ibmr = NULL;
 749         int ret;
 750
 751         rds_ibdev = rds_ib_get_device(rs->rs_bound_addr);
 752         if (!rds_ibdev) {
 753                 ret = -ENODEV;
 754                 goto out;
 755         }
 756
 757         if (!rds_ibdev->mr_pool) {
 758                 ret = -ENODEV;
 759                 goto out;
 760         }
 761
 762         ibmr = rds_ib_alloc_fmr(rds_ibdev);
 763         if (IS_ERR(ibmr))
 764                 return ibmr;
 765
 766         ret = rds_ib_map_fmr(rds_ibdev, ibmr, sg, nents);
 767         if (ret == 0)
 768                 *key_ret = ibmr->fmr->rkey;
 769         else
 770                 printk(KERN_WARNING "RDS/IB: map_fmr failed (errno=%d)\n", ret);
 771
 772         ibmr->device = rds_ibdev;
 773         rds_ibdev = NULL;
 774
 775  out:
 776         if (ret) {
 777                 if (ibmr)
 778                         rds_ib_free_mr(ibmr, 0);
 779                 ibmr = ERR_PTR(ret);
 780         }
 781         if (rds_ibdev)
 782                 rds_ib_dev_put(rds_ibdev);
 783         return ibmr;
 784 }
 785