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.h"
  39 #include "ib.h"
  40
  41 static DEFINE_PER_CPU(unsigned long, clean_list_grace);
  42 #define CLEAN_LIST_BUSY_BIT 0
  43
  44 /*
  45  * This is stored as mr->r_trans_private.
  46  */
  47 struct rds_ib_mr {
  48         struct rds_ib_device    *device;
  49         struct rds_ib_mr_pool   *pool;
  50         struct ib_fmr           *fmr;
  51
  52         struct llist_node       llnode;
  53
  54         /* unmap_list is for freeing */
  55         struct list_head        unmap_list;
  56         unsigned int            remap_count;
  57
  58         struct scatterlist      *sg;
  59         unsigned int            sg_len;
  60         u64                     *dma;
  61         int                     sg_dma_len;
  62 };
  63
  64 /*
  65  * Our own little FMR pool
  66  */
  67 struct rds_ib_mr_pool {
  68         struct mutex            flush_lock;             /* serialize fmr invalidate */
  69         struct delayed_work     flush_worker;           /* flush worker */
  70
  71         atomic_t                item_count;             /* total # of MRs */
  72         atomic_t                dirty_count;            /* # dirty of MRs */
  73
  74         struct llist_head       drop_list;              /* MRs that have reached their max_maps limit */
  75         struct llist_head       free_list;              /* unused MRs */
  76         struct llist_head       clean_list;             /* global unused & unamapped MRs */
  77         wait_queue_head_t       flush_wait;
  78
  79         atomic_t                free_pinned;            /* memory pinned by free MRs */
  80         unsigned long           max_items;
  81         unsigned long           max_items_soft;
  82         unsigned long           max_free_pinned;
  83         struct ib_fmr_attr      fmr_attr;
  84 };
  85
  86 struct workqueue_struct *rds_ib_fmr_wq;
  87
  88 int rds_ib_fmr_init(void)
  89 {
  90         rds_ib_fmr_wq = create_workqueue("rds_fmr_flushd");
  91         if (!rds_ib_fmr_wq)
  92                 return -ENOMEM;
  93         return 0;
  94 }
  95
  96 /* By the time this is called all the IB devices should have been torn down and
  97  * had their pools freed.  As each pool is freed its work struct is waited on,
  98  * so the pool flushing work queue should be idle by the time we get here.
  99  */
 100 void rds_ib_fmr_exit(void)
 101 {
 102         destroy_workqueue(rds_ib_fmr_wq);
 103 }
 104
 105 static int rds_ib_flush_mr_pool(struct rds_ib_mr_pool *pool, int free_all, struct rds_ib_mr **);
 106 static void rds_ib_teardown_mr(struct rds_ib_mr *ibmr);
 107 static void rds_ib_mr_pool_flush_worker(struct work_struct *work);
 108
 109 static struct rds_ib_device *rds_ib_get_device(__be32 ipaddr)
 110 {
 111         struct rds_ib_device *rds_ibdev;
 112         struct rds_ib_ipaddr *i_ipaddr;
 113
 114         rcu_read_lock();
 115         list_for_each_entry_rcu(rds_ibdev, &rds_ib_devices, list) {
 116                 list_for_each_entry_rcu(i_ipaddr, &rds_ibdev->ipaddr_list, list) {
 117                         if (i_ipaddr->ipaddr == ipaddr) {
 118                                 atomic_inc(&rds_ibdev->refcount);
 119                                 rcu_read_unlock();
 120                                 return rds_ibdev;
 121                         }
 122                 }
 123         }
 124         rcu_read_unlock();
 125
 126         return NULL;
 127 }
 128
 129 static int rds_ib_add_ipaddr(struct rds_ib_device *rds_ibdev, __be32 ipaddr)
 130 {
 131         struct rds_ib_ipaddr *i_ipaddr;
 132
 133         i_ipaddr = kmalloc(sizeof *i_ipaddr, GFP_KERNEL);
 134         if (!i_ipaddr)
 135                 return -ENOMEM;
 136
 137         i_ipaddr->ipaddr = ipaddr;
 138
 139         spin_lock_irq(&rds_ibdev->spinlock);
 140         list_add_tail_rcu(&i_ipaddr->list, &rds_ibdev->ipaddr_list);
 141         spin_unlock_irq(&rds_ibdev->spinlock);
 142
 143         return 0;
 144 }
 145
 146 static void rds_ib_remove_ipaddr(struct rds_ib_device *rds_ibdev, __be32 ipaddr)
 147 {
 148         struct rds_ib_ipaddr *i_ipaddr;
 149         struct rds_ib_ipaddr *to_free = NULL;
 150
 151
 152         spin_lock_irq(&rds_ibdev->spinlock);
 153         list_for_each_entry_rcu(i_ipaddr, &rds_ibdev->ipaddr_list, list) {
 154                 if (i_ipaddr->ipaddr == ipaddr) {
 155                         list_del_rcu(&i_ipaddr->list);
 156                         to_free = i_ipaddr;
 157                         break;
 158                 }
 159         }
 160         spin_unlock_irq(&rds_ibdev->spinlock);
 161
 162         if (to_free)
 163                 kfree_rcu(to_free, rcu);
 164 }
 165
 166 int rds_ib_update_ipaddr(struct rds_ib_device *rds_ibdev, __be32 ipaddr)
 167 {
 168         struct rds_ib_device *rds_ibdev_old;
 169
 170         rds_ibdev_old = rds_ib_get_device(ipaddr);
 171         if (!rds_ibdev_old)
 172                 return rds_ib_add_ipaddr(rds_ibdev, ipaddr);
 173
 174         if (rds_ibdev_old != rds_ibdev) {
 175                 rds_ib_remove_ipaddr(rds_ibdev_old, ipaddr);
 176                 rds_ib_dev_put(rds_ibdev_old);
 177                 return rds_ib_add_ipaddr(rds_ibdev, ipaddr);
 178         }
 179         rds_ib_dev_put(rds_ibdev_old);
 180
 181         return 0;
 182 }
 183
 184 void rds_ib_add_conn(struct rds_ib_device *rds_ibdev, struct rds_connection *conn)
 185 {
 186         struct rds_ib_connection *ic = conn->c_transport_data;
 187
 188         /* conn was previously on the nodev_conns_list */
 189         spin_lock_irq(&ib_nodev_conns_lock);
 190         BUG_ON(list_empty(&ib_nodev_conns));
 191         BUG_ON(list_empty(&ic->ib_node));
 192         list_del(&ic->ib_node);
 193
 194         spin_lock(&rds_ibdev->spinlock);
 195         list_add_tail(&ic->ib_node, &rds_ibdev->conn_list);
 196         spin_unlock(&rds_ibdev->spinlock);
 197         spin_unlock_irq(&ib_nodev_conns_lock);
 198
 199         ic->rds_ibdev = rds_ibdev;
 200         atomic_inc(&rds_ibdev->refcount);
 201 }
 202
 203 void rds_ib_remove_conn(struct rds_ib_device *rds_ibdev, struct rds_connection *conn)
 204 {
 205         struct rds_ib_connection *ic = conn->c_transport_data;
 206
 207         /* place conn on nodev_conns_list */
 208         spin_lock(&ib_nodev_conns_lock);
 209
 210         spin_lock_irq(&rds_ibdev->spinlock);
 211         BUG_ON(list_empty(&ic->ib_node));
 212         list_del(&ic->ib_node);
 213         spin_unlock_irq(&rds_ibdev->spinlock);
 214
 215         list_add_tail(&ic->ib_node, &ib_nodev_conns);
 216
 217         spin_unlock(&ib_nodev_conns_lock);
 218
 219         ic->rds_ibdev = NULL;
 220         rds_ib_dev_put(rds_ibdev);
 221 }
 222
 223 void rds_ib_destroy_nodev_conns(void)
 224 {
 225         struct rds_ib_connection *ic, *_ic;
 226         LIST_HEAD(tmp_list);
 227
 228         /* avoid calling conn_destroy with irqs off */
 229         spin_lock_irq(&ib_nodev_conns_lock);
 230         list_splice(&ib_nodev_conns, &tmp_list);
 231         spin_unlock_irq(&ib_nodev_conns_lock);
 232
 233         list_for_each_entry_safe(ic, _ic, &tmp_list, ib_node)
 234                 rds_conn_destroy(ic->conn);
 235 }
 236
 237 struct rds_ib_mr_pool *rds_ib_create_mr_pool(struct rds_ib_device *rds_ibdev)
 238 {
 239         struct rds_ib_mr_pool *pool;
 240
 241         pool = kzalloc(sizeof(*pool), GFP_KERNEL);
 242         if (!pool)
 243                 return ERR_PTR(-ENOMEM);
 244
 245         init_llist_head(&pool->free_list);
 246         init_llist_head(&pool->drop_list);
 247         init_llist_head(&pool->clean_list);
 248         mutex_init(&pool->flush_lock);
 249         init_waitqueue_head(&pool->flush_wait);
 250         INIT_DELAYED_WORK(&pool->flush_worker, rds_ib_mr_pool_flush_worker);
 251
 252         pool->fmr_attr.max_pages = fmr_message_size;
 253         pool->fmr_attr.max_maps = rds_ibdev->fmr_max_remaps;
 254         pool->fmr_attr.page_shift = PAGE_SHIFT;
 255         pool->max_free_pinned = rds_ibdev->max_fmrs * fmr_message_size / 4;
 256
 257         /* We never allow more than max_items MRs to be allocated.
 258          * When we exceed more than max_items_soft, we start freeing
 259          * items more aggressively.
 260          * Make sure that max_items > max_items_soft > max_items / 2
 261          */
 262         pool->max_items_soft = rds_ibdev->max_fmrs * 3 / 4;
 263         pool->max_items = rds_ibdev->max_fmrs;
 264
 265         return pool;
 266 }
 267
 268 void rds_ib_get_mr_info(struct rds_ib_device *rds_ibdev, struct rds_info_rdma_connection *iinfo)
 269 {
 270         struct rds_ib_mr_pool *pool = rds_ibdev->mr_pool;
 271
 272         iinfo->rdma_mr_max = pool->max_items;
 273         iinfo->rdma_mr_size = pool->fmr_attr.max_pages;
 274 }
 275
 276 void rds_ib_destroy_mr_pool(struct rds_ib_mr_pool *pool)
 277 {
 278         cancel_delayed_work_sync(&pool->flush_worker);
 279         rds_ib_flush_mr_pool(pool, 1, NULL);
 280         WARN_ON(atomic_read(&pool->item_count));
 281         WARN_ON(atomic_read(&pool->free_pinned));
 282         kfree(pool);
 283 }
 284
 285 static inline struct rds_ib_mr *rds_ib_reuse_fmr(struct rds_ib_mr_pool *pool)
 286 {
 287         struct rds_ib_mr *ibmr = NULL;
 288         struct llist_node *ret;
 289         unsigned long *flag;
 290
 291         preempt_disable();
 292         flag = this_cpu_ptr(&clean_list_grace);
 293         set_bit(CLEAN_LIST_BUSY_BIT, flag);
 294         ret = llist_del_first(&pool->clean_list);
 295         if (ret)
 296                 ibmr = llist_entry(ret, struct rds_ib_mr, llnode);
 297
 298         clear_bit(CLEAN_LIST_BUSY_BIT, flag);
 299         preempt_enable();
 300         return ibmr;
 301 }
 302
 303 static inline void wait_clean_list_grace(void)
 304 {
 305         int cpu;
 306         unsigned long *flag;
 307
 308         for_each_online_cpu(cpu) {
 309                 flag = &per_cpu(clean_list_grace, cpu);
 310                 while (test_bit(CLEAN_LIST_BUSY_BIT, flag))
 311                         cpu_relax();
 312         }
 313 }
 314
 315 static struct rds_ib_mr *rds_ib_alloc_fmr(struct rds_ib_device *rds_ibdev)
 316 {
 317         struct rds_ib_mr_pool *pool = rds_ibdev->mr_pool;
 318         struct rds_ib_mr *ibmr = NULL;
 319         int err = 0, iter = 0;
 320
 321         if (atomic_read(&pool->dirty_count) >= pool->max_items / 10)
 322                 queue_delayed_work(rds_ib_fmr_wq, &pool->flush_worker, 10);
 323
 324         while (1) {
 325                 ibmr = rds_ib_reuse_fmr(pool);
 326                 if (ibmr)
 327                         return ibmr;
 328
 329                 /* No clean MRs - now we have the choice of either
 330                  * allocating a fresh MR up to the limit imposed by the
 331                  * driver, or flush any dirty unused MRs.
 332                  * We try to avoid stalling in the send path if possible,
 333                  * so we allocate as long as we're allowed to.
 334                  *
 335                  * We're fussy with enforcing the FMR limit, though. If the driver
 336                  * tells us we can't use more than N fmrs, we shouldn't start
 337                  * arguing with it */
 338                 if (atomic_inc_return(&pool->item_count) <= pool->max_items)
 339                         break;
 340
 341                 atomic_dec(&pool->item_count);
 342
 343                 if (++iter > 2) {
 344                         rds_ib_stats_inc(s_ib_rdma_mr_pool_depleted);
 345                         return ERR_PTR(-EAGAIN);
 346                 }
 347
 348                 /* We do have some empty MRs. Flush them out. */
 349                 rds_ib_stats_inc(s_ib_rdma_mr_pool_wait);
 350                 rds_ib_flush_mr_pool(pool, 0, &ibmr);
 351                 if (ibmr)
 352                         return ibmr;
 353         }
 354
 355         ibmr = kzalloc_node(sizeof(*ibmr), GFP_KERNEL, rdsibdev_to_node(rds_ibdev));
 356         if (!ibmr) {
 357                 err = -ENOMEM;
 358                 goto out_no_cigar;
 359         }
 360
 361         ibmr->fmr = ib_alloc_fmr(rds_ibdev->pd,
 362                         (IB_ACCESS_LOCAL_WRITE |
 363                          IB_ACCESS_REMOTE_READ |
 364                          IB_ACCESS_REMOTE_WRITE|
 365                          IB_ACCESS_REMOTE_ATOMIC),
 366                         &pool->fmr_attr);
 367         if (IS_ERR(ibmr->fmr)) {
 368                 err = PTR_ERR(ibmr->fmr);
 369                 ibmr->fmr = NULL;
 370                 printk(KERN_WARNING "RDS/IB: ib_alloc_fmr failed (err=%d)\n", err);
 371                 goto out_no_cigar;
 372         }
 373
 374         rds_ib_stats_inc(s_ib_rdma_mr_alloc);
 375         return ibmr;
 376
 377 out_no_cigar:
 378         if (ibmr) {
 379                 if (ibmr->fmr)
 380                         ib_dealloc_fmr(ibmr->fmr);
 381                 kfree(ibmr);
 382         }
 383         atomic_dec(&pool->item_count);
 384         return ERR_PTR(err);
 385 }
 386
 387 static int rds_ib_map_fmr(struct rds_ib_device *rds_ibdev, struct rds_ib_mr *ibmr,
 388                struct scatterlist *sg, unsigned int nents)
 389 {
 390         struct ib_device *dev = rds_ibdev->dev;
 391         struct scatterlist *scat = sg;
 392         u64 io_addr = 0;
 393         u64 *dma_pages;
 394         u32 len;
 395         int page_cnt, sg_dma_len;
 396         int i, j;
 397         int ret;
 398
 399         sg_dma_len = ib_dma_map_sg(dev, sg, nents,
 400                                  DMA_BIDIRECTIONAL);
 401         if (unlikely(!sg_dma_len)) {
 402                 printk(KERN_WARNING "RDS/IB: dma_map_sg failed!\n");
 403                 return -EBUSY;
 404         }
 405
 406         len = 0;
 407         page_cnt = 0;
 408
 409         for (i = 0; i < sg_dma_len; ++i) {
 410                 unsigned int dma_len = ib_sg_dma_len(dev, &scat[i]);
 411                 u64 dma_addr = ib_sg_dma_address(dev, &scat[i]);
 412
 413                 if (dma_addr & ~PAGE_MASK) {
 414                         if (i > 0)
 415                                 return -EINVAL;
 416                         else
 417                                 ++page_cnt;
 418                 }
 419                 if ((dma_addr + dma_len) & ~PAGE_MASK) {
 420                         if (i < sg_dma_len - 1)
 421                                 return -EINVAL;
 422                         else
 423                                 ++page_cnt;
 424                 }
 425
 426                 len += dma_len;
 427         }
 428
 429         page_cnt += len >> PAGE_SHIFT;
 430         if (page_cnt > fmr_message_size)
 431                 return -EINVAL;
 432
 433         dma_pages = kmalloc_node(sizeof(u64) * page_cnt, GFP_ATOMIC,
 434                                  rdsibdev_to_node(rds_ibdev));
 435         if (!dma_pages)
 436                 return -ENOMEM;
 437
 438         page_cnt = 0;
 439         for (i = 0; i < sg_dma_len; ++i) {
 440                 unsigned int dma_len = ib_sg_dma_len(dev, &scat[i]);
 441                 u64 dma_addr = ib_sg_dma_address(dev, &scat[i]);
 442
 443                 for (j = 0; j < dma_len; j += PAGE_SIZE)
 444                         dma_pages[page_cnt++] =
 445                                 (dma_addr & PAGE_MASK) + j;
 446         }
 447
 448         ret = ib_map_phys_fmr(ibmr->fmr,
 449                                    dma_pages, page_cnt, io_addr);
 450         if (ret)
 451                 goto out;
 452
 453         /* Success - we successfully remapped the MR, so we can
 454          * safely tear down the old mapping. */
 455         rds_ib_teardown_mr(ibmr);
 456
 457         ibmr->sg = scat;
 458         ibmr->sg_len = nents;
 459         ibmr->sg_dma_len = sg_dma_len;
 460         ibmr->remap_count++;
 461
 462         rds_ib_stats_inc(s_ib_rdma_mr_used);
 463         ret = 0;
 464
 465 out:
 466         kfree(dma_pages);
 467
 468         return ret;
 469 }
 470
 471 void rds_ib_sync_mr(void *trans_private, int direction)
 472 {
 473         struct rds_ib_mr *ibmr = trans_private;
 474         struct rds_ib_device *rds_ibdev = ibmr->device;
 475
 476         switch (direction) {
 477         case DMA_FROM_DEVICE:
 478                 ib_dma_sync_sg_for_cpu(rds_ibdev->dev, ibmr->sg,
 479                         ibmr->sg_dma_len, DMA_BIDIRECTIONAL);
 480                 break;
 481         case DMA_TO_DEVICE:
 482                 ib_dma_sync_sg_for_device(rds_ibdev->dev, ibmr->sg,
 483                         ibmr->sg_dma_len, DMA_BIDIRECTIONAL);
 484                 break;
 485         }
 486 }
 487
 488 static void __rds_ib_teardown_mr(struct rds_ib_mr *ibmr)
 489 {
 490         struct rds_ib_device *rds_ibdev = ibmr->device;
 491
 492         if (ibmr->sg_dma_len) {
 493                 ib_dma_unmap_sg(rds_ibdev->dev,
 494                                 ibmr->sg, ibmr->sg_len,
 495                                 DMA_BIDIRECTIONAL);
 496                 ibmr->sg_dma_len = 0;
 497         }
 498
 499         /* Release the s/g list */
 500         if (ibmr->sg_len) {
 501                 unsigned int i;
 502
 503                 for (i = 0; i < ibmr->sg_len; ++i) {
 504                         struct page *page = sg_page(&ibmr->sg[i]);
 505
 506                         /* FIXME we need a way to tell a r/w MR
 507                          * from a r/o MR */
 508                         WARN_ON(!page->mapping && irqs_disabled());
 509                         set_page_dirty(page);
 510                         put_page(page);
 511                 }
 512                 kfree(ibmr->sg);
 513
 514                 ibmr->sg = NULL;
 515                 ibmr->sg_len = 0;
 516         }
 517 }
 518
 519 static void rds_ib_teardown_mr(struct rds_ib_mr *ibmr)
 520 {
 521         unsigned int pinned = ibmr->sg_len;
 522
 523         __rds_ib_teardown_mr(ibmr);
 524         if (pinned) {
 525                 struct rds_ib_device *rds_ibdev = ibmr->device;
 526                 struct rds_ib_mr_pool *pool = rds_ibdev->mr_pool;
 527
 528                 atomic_sub(pinned, &pool->free_pinned);
 529         }
 530 }
 531
 532 static inline unsigned int rds_ib_flush_goal(struct rds_ib_mr_pool *pool, int free_all)
 533 {
 534         unsigned int item_count;
 535
 536         item_count = atomic_read(&pool->item_count);
 537         if (free_all)
 538                 return item_count;
 539
 540         return 0;
 541 }
 542
 543 /*
 544  * given an llist of mrs, put them all into the list_head for more processing
 545  */
 546 static unsigned int llist_append_to_list(struct llist_head *llist,
 547                                          struct list_head *list)
 548 {
 549         struct rds_ib_mr *ibmr;
 550         struct llist_node *node;
 551         struct llist_node *next;
 552         unsigned int count = 0;
 553
 554         node = llist_del_all(llist);
 555         while (node) {
 556                 next = node->next;
 557                 ibmr = llist_entry(node, struct rds_ib_mr, llnode);
 558                 list_add_tail(&ibmr->unmap_list, list);
 559                 node = next;
 560                 count++;
 561         }
 562         return count;
 563 }
 564
 565 /*
 566  * this takes a list head of mrs and turns it into linked llist nodes
 567  * of clusters.  Each cluster has linked llist nodes of
 568  * MR_CLUSTER_SIZE mrs that are ready for reuse.
 569  */
 570 static void list_to_llist_nodes(struct rds_ib_mr_pool *pool,
 571                                 struct list_head *list,
 572                                 struct llist_node **nodes_head,
 573                                 struct llist_node **nodes_tail)
 574 {
 575         struct rds_ib_mr *ibmr;
 576         struct llist_node *cur = NULL;
 577         struct llist_node **next = nodes_head;
 578
 579         list_for_each_entry(ibmr, list, unmap_list) {
 580                 cur = &ibmr->llnode;
 581                 *next = cur;
 582                 next = &cur->next;
 583         }
 584         *next = NULL;
 585         *nodes_tail = cur;
 586 }
 587
 588 /*
 589  * Flush our pool of MRs.
 590  * At a minimum, all currently unused MRs are unmapped.
 591  * If the number of MRs allocated exceeds the limit, we also try
 592  * to free as many MRs as needed to get back to this limit.
 593  */
 594 static int rds_ib_flush_mr_pool(struct rds_ib_mr_pool *pool,
 595                                 int free_all, struct rds_ib_mr **ibmr_ret)
 596 {
 597         struct rds_ib_mr *ibmr, *next;
 598         struct llist_node *clean_nodes;
 599         struct llist_node *clean_tail;
 600         LIST_HEAD(unmap_list);
 601         LIST_HEAD(fmr_list);
 602         unsigned long unpinned = 0;
 603         unsigned int nfreed = 0, dirty_to_clean = 0, free_goal;
 604         int ret = 0;
 605
 606         rds_ib_stats_inc(s_ib_rdma_mr_pool_flush);
 607
 608         if (ibmr_ret) {
 609                 DEFINE_WAIT(wait);
 610                 while(!mutex_trylock(&pool->flush_lock)) {
 611                         ibmr = rds_ib_reuse_fmr(pool);
 612                         if (ibmr) {
 613                                 *ibmr_ret = ibmr;
 614                                 finish_wait(&pool->flush_wait, &wait);
 615                                 goto out_nolock;
 616                         }
 617
 618                         prepare_to_wait(&pool->flush_wait, &wait,
 619                                         TASK_UNINTERRUPTIBLE);
 620                         if (llist_empty(&pool->clean_list))
 621                                 schedule();
 622
 623                         ibmr = rds_ib_reuse_fmr(pool);
 624                         if (ibmr) {
 625                                 *ibmr_ret = ibmr;
 626                                 finish_wait(&pool->flush_wait, &wait);
 627                                 goto out_nolock;
 628                         }
 629                 }
 630                 finish_wait(&pool->flush_wait, &wait);
 631         } else
 632                 mutex_lock(&pool->flush_lock);
 633
 634         if (ibmr_ret) {
 635                 ibmr = rds_ib_reuse_fmr(pool);
 636                 if (ibmr) {
 637                         *ibmr_ret = ibmr;
 638                         goto out;
 639                 }
 640         }
 641
 642         /* Get the list of all MRs to be dropped. Ordering matters -
 643          * we want to put drop_list ahead of free_list.
 644          */
 645         dirty_to_clean = llist_append_to_list(&pool->drop_list, &unmap_list);
 646         dirty_to_clean += llist_append_to_list(&pool->free_list, &unmap_list);
 647         if (free_all)
 648                 llist_append_to_list(&pool->clean_list, &unmap_list);
 649
 650         free_goal = rds_ib_flush_goal(pool, free_all);
 651
 652         if (list_empty(&unmap_list))
 653                 goto out;
 654
 655         /* String all ib_mr's onto one list and hand them to ib_unmap_fmr */
 656         list_for_each_entry(ibmr, &unmap_list, unmap_list)
 657                 list_add(&ibmr->fmr->list, &fmr_list);
 658
 659         ret = ib_unmap_fmr(&fmr_list);
 660         if (ret)
 661                 printk(KERN_WARNING "RDS/IB: ib_unmap_fmr failed (err=%d)\n", ret);
 662
 663         /* Now we can destroy the DMA mapping and unpin any pages */
 664         list_for_each_entry_safe(ibmr, next, &unmap_list, unmap_list) {
 665                 unpinned += ibmr->sg_len;
 666                 __rds_ib_teardown_mr(ibmr);
 667                 if (nfreed < free_goal || ibmr->remap_count >= pool->fmr_attr.max_maps) {
 668                         rds_ib_stats_inc(s_ib_rdma_mr_free);
 669                         list_del(&ibmr->unmap_list);
 670                         ib_dealloc_fmr(ibmr->fmr);
 671                         kfree(ibmr);
 672                         nfreed++;
 673                 }
 674         }
 675
 676         if (!list_empty(&unmap_list)) {
 677                 /* we have to make sure that none of the things we're about
 678                  * to put on the clean list would race with other cpus trying
 679                  * to pull items off.  The llist would explode if we managed to
 680                  * remove something from the clean list and then add it back again
 681                  * while another CPU was spinning on that same item in llist_del_first.
 682                  *
 683                  * This is pretty unlikely, but just in case  wait for an llist grace period
 684                  * here before adding anything back into the clean list.
 685                  */
 686                 wait_clean_list_grace();
 687
 688                 list_to_llist_nodes(pool, &unmap_list, &clean_nodes, &clean_tail);
 689                 if (ibmr_ret)
 690                         *ibmr_ret = llist_entry(clean_nodes, struct rds_ib_mr, llnode);
 691
 692                 /* more than one entry in llist nodes */
 693                 if (clean_nodes->next)
 694                         llist_add_batch(clean_nodes->next, clean_tail, &pool->clean_list);
 695
 696         }
 697
 698         atomic_sub(unpinned, &pool->free_pinned);
 699         atomic_sub(dirty_to_clean, &pool->dirty_count);
 700         atomic_sub(nfreed, &pool->item_count);
 701
 702 out:
 703         mutex_unlock(&pool->flush_lock);
 704         if (waitqueue_active(&pool->flush_wait))
 705                 wake_up(&pool->flush_wait);
 706 out_nolock:
 707         return ret;
 708 }
 709
 710 static void rds_ib_mr_pool_flush_worker(struct work_struct *work)
 711 {
 712         struct rds_ib_mr_pool *pool = container_of(work, struct rds_ib_mr_pool, flush_worker.work);
 713
 714         rds_ib_flush_mr_pool(pool, 0, NULL);
 715 }
 716
 717 void rds_ib_free_mr(void *trans_private, int invalidate)
 718 {
 719         struct rds_ib_mr *ibmr = trans_private;
 720         struct rds_ib_device *rds_ibdev = ibmr->device;
 721         struct rds_ib_mr_pool *pool = rds_ibdev->mr_pool;
 722
 723         rdsdebug("RDS/IB: free_mr nents %u\n", ibmr->sg_len);
 724
 725         /* Return it to the pool's free list */
 726         if (ibmr->remap_count >= pool->fmr_attr.max_maps)
 727                 llist_add(&ibmr->llnode, &pool->drop_list);
 728         else
 729                 llist_add(&ibmr->llnode, &pool->free_list);
 730
 731         atomic_add(ibmr->sg_len, &pool->free_pinned);
 732         atomic_inc(&pool->dirty_count);
 733
 734         /* If we've pinned too many pages, request a flush */
 735         if (atomic_read(&pool->free_pinned) >= pool->max_free_pinned ||
 736             atomic_read(&pool->dirty_count) >= pool->max_items / 5)
 737                 queue_delayed_work(rds_ib_fmr_wq, &pool->flush_worker, 10);
 738
 739         if (invalidate) {
 740                 if (likely(!in_interrupt())) {
 741                         rds_ib_flush_mr_pool(pool, 0, NULL);
 742                 } else {
 743                         /* We get here if the user created a MR marked
 744                          * as use_once and invalidate at the same time.
 745                          */
 746                         queue_delayed_work(rds_ib_fmr_wq,
 747                                            &pool->flush_worker, 10);
 748                 }
 749         }
 750
 751         rds_ib_dev_put(rds_ibdev);
 752 }
 753
 754 void rds_ib_flush_mrs(void)
 755 {
 756         struct rds_ib_device *rds_ibdev;
 757
 758         down_read(&rds_ib_devices_lock);
 759         list_for_each_entry(rds_ibdev, &rds_ib_devices, list) {
 760                 struct rds_ib_mr_pool *pool = rds_ibdev->mr_pool;
 761
 762                 if (pool)
 763                         rds_ib_flush_mr_pool(pool, 0, NULL);
 764         }
 765         up_read(&rds_ib_devices_lock);
 766 }
 767
 768 void *rds_ib_get_mr(struct scatterlist *sg, unsigned long nents,
 769                     struct rds_sock *rs, u32 *key_ret)
 770 {
 771         struct rds_ib_device *rds_ibdev;
 772         struct rds_ib_mr *ibmr = NULL;
 773         int ret;
 774
 775         rds_ibdev = rds_ib_get_device(rs->rs_bound_addr);
 776         if (!rds_ibdev) {
 777                 ret = -ENODEV;
 778                 goto out;
 779         }
 780
 781         if (!rds_ibdev->mr_pool) {
 782                 ret = -ENODEV;
 783                 goto out;
 784         }
 785
 786         ibmr = rds_ib_alloc_fmr(rds_ibdev);
 787         if (IS_ERR(ibmr)) {
 788                 rds_ib_dev_put(rds_ibdev);
 789                 return ibmr;
 790         }
 791
 792         ret = rds_ib_map_fmr(rds_ibdev, ibmr, sg, nents);
 793         if (ret == 0)
 794                 *key_ret = ibmr->fmr->rkey;
 795         else
 796                 printk(KERN_WARNING "RDS/IB: map_fmr failed (errno=%d)\n", ret);
 797
 798         ibmr->device = rds_ibdev;
 799         rds_ibdev = NULL;
 800
 801  out:
 802         if (ret) {
 803                 if (ibmr)
 804                         rds_ib_free_mr(ibmr, 0);
 805                 ibmr = ERR_PTR(ret);
 806         }
 807         if (rds_ibdev)
 808                 rds_ib_dev_put(rds_ibdev);
 809         return ibmr;
 810 }
 811