net/rds/ib_rdma.c

   1 /*
   2  * Copyright (c) 2006 Oracle.  All rights reserved.
   3  *
   4  * This software is available to you under a choice of one of two
   5  * licenses.  You may choose to be licensed under the terms of the GNU
   6  * General Public License (GPL) Version 2, available from the file
   7  * COPYING in the main directory of this source tree, or the
   8  * OpenIB.org BSD license below:
   9  *
  10  *     Redistribution and use in source and binary forms, with or
  11  *     without modification, are permitted provided that the following
  12  *     conditions are met:
  13  *
  14  *      - Redistributions of source code must retain the above
  15  *        copyright notice, this list of conditions and the following
  16  *        disclaimer.
  17  *
  18  *      - Redistributions in binary form must reproduce the above
  19  *        copyright notice, this list of conditions and the following
  20  *        disclaimer in the documentation and/or other materials
  21  *        provided with the distribution.
  22  *
  23  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  24  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  25  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  26  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  27  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  28  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  29  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  30  * SOFTWARE.
  31  *
  32  */
  33 #include <linux/kernel.h>
  34 #include <linux/slab.h>
  35 #include <linux/rculist.h>
  36
  37 #include "rds.h"
  38 #include "ib.h"
  39 #include "xlist.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 xlist_head       xlist;
  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 xlist_head       drop_list;              /* MRs that have reached their max_maps limit */
  75         struct xlist_head       free_list;              /* unused MRs */
  76         struct xlist_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 static int rds_ib_flush_mr_pool(struct rds_ib_mr_pool *pool, int free_all, struct rds_ib_mr **);
  87 static void rds_ib_teardown_mr(struct rds_ib_mr *ibmr);
  88 static void rds_ib_mr_pool_flush_worker(struct work_struct *work);
  89
  90 static struct rds_ib_device *rds_ib_get_device(__be32 ipaddr)
  91 {
  92         struct rds_ib_device *rds_ibdev;
  93         struct rds_ib_ipaddr *i_ipaddr;
  94
  95         list_for_each_entry(rds_ibdev, &rds_ib_devices, list) {
  96                 rcu_read_lock();
  97                 list_for_each_entry_rcu(i_ipaddr, &rds_ibdev->ipaddr_list, list) {
  98                         if (i_ipaddr->ipaddr == ipaddr) {
  99                                 atomic_inc(&rds_ibdev->refcount);
 100                                 rcu_read_unlock();
 101                                 return rds_ibdev;
 102                         }
 103                 }
 104                 rcu_read_unlock();
 105         }
 106
 107         return NULL;
 108 }
 109
 110 static int rds_ib_add_ipaddr(struct rds_ib_device *rds_ibdev, __be32 ipaddr)
 111 {
 112         struct rds_ib_ipaddr *i_ipaddr;
 113
 114         i_ipaddr = kmalloc(sizeof *i_ipaddr, GFP_KERNEL);
 115         if (!i_ipaddr)
 116                 return -ENOMEM;
 117
 118         i_ipaddr->ipaddr = ipaddr;
 119
 120         spin_lock_irq(&rds_ibdev->spinlock);
 121         list_add_tail_rcu(&i_ipaddr->list, &rds_ibdev->ipaddr_list);
 122         spin_unlock_irq(&rds_ibdev->spinlock);
 123
 124         return 0;
 125 }
 126
 127 static void rds_ib_remove_ipaddr(struct rds_ib_device *rds_ibdev, __be32 ipaddr)
 128 {
 129         struct rds_ib_ipaddr *i_ipaddr;
 130         struct rds_ib_ipaddr *to_free = NULL;
 131
 132
 133         spin_lock_irq(&rds_ibdev->spinlock);
 134         list_for_each_entry_rcu(i_ipaddr, &rds_ibdev->ipaddr_list, list) {
 135                 if (i_ipaddr->ipaddr == ipaddr) {
 136                         list_del_rcu(&i_ipaddr->list);
 137                         to_free = i_ipaddr;
 138                         break;
 139                 }
 140         }
 141         spin_unlock_irq(&rds_ibdev->spinlock);
 142
 143         if (to_free) {
 144                 synchronize_rcu();
 145                 kfree(to_free);
 146         }
 147 }
 148
 149 int rds_ib_update_ipaddr(struct rds_ib_device *rds_ibdev, __be32 ipaddr)
 150 {
 151         struct rds_ib_device *rds_ibdev_old;
 152
 153         rds_ibdev_old = rds_ib_get_device(ipaddr);
 154         if (rds_ibdev_old) {
 155                 rds_ib_remove_ipaddr(rds_ibdev_old, ipaddr);
 156                 rds_ib_dev_put(rds_ibdev_old);
 157         }
 158
 159         return rds_ib_add_ipaddr(rds_ibdev, ipaddr);
 160 }
 161
 162 void rds_ib_add_conn(struct rds_ib_device *rds_ibdev, struct rds_connection *conn)
 163 {
 164         struct rds_ib_connection *ic = conn->c_transport_data;
 165
 166         /* conn was previously on the nodev_conns_list */
 167         spin_lock_irq(&ib_nodev_conns_lock);
 168         BUG_ON(list_empty(&ib_nodev_conns));
 169         BUG_ON(list_empty(&ic->ib_node));
 170         list_del(&ic->ib_node);
 171
 172         spin_lock_irq(&rds_ibdev->spinlock);
 173         list_add_tail(&ic->ib_node, &rds_ibdev->conn_list);
 174         spin_unlock_irq(&rds_ibdev->spinlock);
 175         spin_unlock_irq(&ib_nodev_conns_lock);
 176
 177         ic->rds_ibdev = rds_ibdev;
 178         atomic_inc(&rds_ibdev->refcount);
 179 }
 180
 181 void rds_ib_remove_conn(struct rds_ib_device *rds_ibdev, struct rds_connection *conn)
 182 {
 183         struct rds_ib_connection *ic = conn->c_transport_data;
 184
 185         /* place conn on nodev_conns_list */
 186         spin_lock(&ib_nodev_conns_lock);
 187
 188         spin_lock_irq(&rds_ibdev->spinlock);
 189         BUG_ON(list_empty(&ic->ib_node));
 190         list_del(&ic->ib_node);
 191         spin_unlock_irq(&rds_ibdev->spinlock);
 192
 193         list_add_tail(&ic->ib_node, &ib_nodev_conns);
 194
 195         spin_unlock(&ib_nodev_conns_lock);
 196
 197         ic->rds_ibdev = NULL;
 198         rds_ib_dev_put(rds_ibdev);
 199 }
 200
 201 void rds_ib_destroy_nodev_conns(void)
 202 {
 203         struct rds_ib_connection *ic, *_ic;
 204         LIST_HEAD(tmp_list);
 205
 206         /* avoid calling conn_destroy with irqs off */
 207         spin_lock_irq(&ib_nodev_conns_lock);
 208         list_splice(&ib_nodev_conns, &tmp_list);
 209         spin_unlock_irq(&ib_nodev_conns_lock);
 210
 211         list_for_each_entry_safe(ic, _ic, &tmp_list, ib_node)
 212                 rds_conn_destroy(ic->conn);
 213 }
 214
 215 struct rds_ib_mr_pool *rds_ib_create_mr_pool(struct rds_ib_device *rds_ibdev)
 216 {
 217         struct rds_ib_mr_pool *pool;
 218
 219         pool = kzalloc(sizeof(*pool), GFP_KERNEL);
 220         if (!pool)
 221                 return ERR_PTR(-ENOMEM);
 222
 223         INIT_XLIST_HEAD(&pool->free_list);
 224         INIT_XLIST_HEAD(&pool->drop_list);
 225         INIT_XLIST_HEAD(&pool->clean_list);
 226         mutex_init(&pool->flush_lock);
 227         init_waitqueue_head(&pool->flush_wait);
 228         INIT_DELAYED_WORK(&pool->flush_worker, rds_ib_mr_pool_flush_worker);
 229
 230         pool->fmr_attr.max_pages = fmr_message_size;
 231         pool->fmr_attr.max_maps = rds_ibdev->fmr_max_remaps;
 232         pool->fmr_attr.page_shift = PAGE_SHIFT;
 233         pool->max_free_pinned = rds_ibdev->max_fmrs * fmr_message_size / 4;
 234
 235         /* We never allow more than max_items MRs to be allocated.
 236          * When we exceed more than max_items_soft, we start freeing
 237          * items more aggressively.
 238          * Make sure that max_items > max_items_soft > max_items / 2
 239          */
 240         pool->max_items_soft = rds_ibdev->max_fmrs * 3 / 4;
 241         pool->max_items = rds_ibdev->max_fmrs;
 242
 243         return pool;
 244 }
 245
 246 void rds_ib_get_mr_info(struct rds_ib_device *rds_ibdev, struct rds_info_rdma_connection *iinfo)
 247 {
 248         struct rds_ib_mr_pool *pool = rds_ibdev->mr_pool;
 249
 250         iinfo->rdma_mr_max = pool->max_items;
 251         iinfo->rdma_mr_size = pool->fmr_attr.max_pages;
 252 }
 253
 254 void rds_ib_destroy_mr_pool(struct rds_ib_mr_pool *pool)
 255 {
 256         cancel_delayed_work_sync(&pool->flush_worker);
 257         rds_ib_flush_mr_pool(pool, 1, NULL);
 258         WARN_ON(atomic_read(&pool->item_count));
 259         WARN_ON(atomic_read(&pool->free_pinned));
 260         kfree(pool);
 261 }
 262
 263 static void refill_local(struct rds_ib_mr_pool *pool, struct xlist_head *xl,
 264                          struct rds_ib_mr **ibmr_ret)
 265 {
 266         struct xlist_head *ibmr_xl;
 267         ibmr_xl = xlist_del_head_fast(xl);
 268         *ibmr_ret = list_entry(ibmr_xl, struct rds_ib_mr, xlist);
 269 }
 270
 271 static inline struct rds_ib_mr *rds_ib_reuse_fmr(struct rds_ib_mr_pool *pool)
 272 {
 273         struct rds_ib_mr *ibmr = NULL;
 274         struct xlist_head *ret;
 275         unsigned long *flag;
 276
 277         preempt_disable();
 278         flag = &__get_cpu_var(clean_list_grace);
 279         set_bit(CLEAN_LIST_BUSY_BIT, flag);
 280         ret = xlist_del_head(&pool->clean_list);
 281         if (ret)
 282                 ibmr = list_entry(ret, struct rds_ib_mr, xlist);
 283
 284         clear_bit(CLEAN_LIST_BUSY_BIT, flag);
 285         preempt_enable();
 286         return ibmr;
 287 }
 288
 289 static inline void wait_clean_list_grace(void)
 290 {
 291         int cpu;
 292         unsigned long *flag;
 293
 294         for_each_online_cpu(cpu) {
 295                 flag = &per_cpu(clean_list_grace, cpu);
 296                 while (test_bit(CLEAN_LIST_BUSY_BIT, flag))
 297                         cpu_relax();
 298         }
 299 }
 300
 301 static struct rds_ib_mr *rds_ib_alloc_fmr(struct rds_ib_device *rds_ibdev)
 302 {
 303         struct rds_ib_mr_pool *pool = rds_ibdev->mr_pool;
 304         struct rds_ib_mr *ibmr = NULL;
 305         int err = 0, iter = 0;
 306
 307         while (1) {
 308                 ibmr = rds_ib_reuse_fmr(pool);
 309                 if (ibmr)
 310                         return ibmr;
 311
 312                 /* No clean MRs - now we have the choice of either
 313                  * allocating a fresh MR up to the limit imposed by the
 314                  * driver, or flush any dirty unused MRs.
 315                  * We try to avoid stalling in the send path if possible,
 316                  * so we allocate as long as we're allowed to.
 317                  *
 318                  * We're fussy with enforcing the FMR limit, though. If the driver
 319                  * tells us we can't use more than N fmrs, we shouldn't start
 320                  * arguing with it */
 321                 if (atomic_inc_return(&pool->item_count) <= pool->max_items)
 322                         break;
 323
 324                 atomic_dec(&pool->item_count);
 325
 326                 if (++iter > 2) {
 327                         rds_ib_stats_inc(s_ib_rdma_mr_pool_depleted);
 328                         return ERR_PTR(-EAGAIN);
 329                 }
 330
 331                 /* We do have some empty MRs. Flush them out. */
 332                 rds_ib_stats_inc(s_ib_rdma_mr_pool_wait);
 333                 rds_ib_flush_mr_pool(pool, 0, &ibmr);
 334                 if (ibmr)
 335                         return ibmr;
 336         }
 337
 338         ibmr = kzalloc_node(sizeof(*ibmr), GFP_KERNEL, rdsibdev_to_node(rds_ibdev));
 339         if (!ibmr) {
 340                 err = -ENOMEM;
 341                 goto out_no_cigar;
 342         }
 343
 344         memset(ibmr, 0, sizeof(*ibmr));
 345
 346         ibmr->fmr = ib_alloc_fmr(rds_ibdev->pd,
 347                         (IB_ACCESS_LOCAL_WRITE |
 348                          IB_ACCESS_REMOTE_READ |
 349                          IB_ACCESS_REMOTE_WRITE|
 350                          IB_ACCESS_REMOTE_ATOMIC),
 351                         &pool->fmr_attr);
 352         if (IS_ERR(ibmr->fmr)) {
 353                 err = PTR_ERR(ibmr->fmr);
 354                 ibmr->fmr = NULL;
 355                 printk(KERN_WARNING "RDS/IB: ib_alloc_fmr failed (err=%d)\n", err);
 356                 goto out_no_cigar;
 357         }
 358
 359         rds_ib_stats_inc(s_ib_rdma_mr_alloc);
 360         return ibmr;
 361
 362 out_no_cigar:
 363         if (ibmr) {
 364                 if (ibmr->fmr)
 365                         ib_dealloc_fmr(ibmr->fmr);
 366                 kfree(ibmr);
 367         }
 368         atomic_dec(&pool->item_count);
 369         return ERR_PTR(err);
 370 }
 371
 372 static int rds_ib_map_fmr(struct rds_ib_device *rds_ibdev, struct rds_ib_mr *ibmr,
 373                struct scatterlist *sg, unsigned int nents)
 374 {
 375         struct ib_device *dev = rds_ibdev->dev;
 376         struct scatterlist *scat = sg;
 377         u64 io_addr = 0;
 378         u64 *dma_pages;
 379         u32 len;
 380         int page_cnt, sg_dma_len;
 381         int i, j;
 382         int ret;
 383
 384         sg_dma_len = ib_dma_map_sg(dev, sg, nents,
 385                                  DMA_BIDIRECTIONAL);
 386         if (unlikely(!sg_dma_len)) {
 387                 printk(KERN_WARNING "RDS/IB: dma_map_sg failed!\n");
 388                 return -EBUSY;
 389         }
 390
 391         len = 0;
 392         page_cnt = 0;
 393
 394         for (i = 0; i < sg_dma_len; ++i) {
 395                 unsigned int dma_len = ib_sg_dma_len(dev, &scat[i]);
 396                 u64 dma_addr = ib_sg_dma_address(dev, &scat[i]);
 397
 398                 if (dma_addr & ~PAGE_MASK) {
 399                         if (i > 0)
 400                                 return -EINVAL;
 401                         else
 402                                 ++page_cnt;
 403                 }
 404                 if ((dma_addr + dma_len) & ~PAGE_MASK) {
 405                         if (i < sg_dma_len - 1)
 406                                 return -EINVAL;
 407                         else
 408                                 ++page_cnt;
 409                 }
 410
 411                 len += dma_len;
 412         }
 413
 414         page_cnt += len >> PAGE_SHIFT;
 415         if (page_cnt > fmr_message_size)
 416                 return -EINVAL;
 417
 418         dma_pages = kmalloc_node(sizeof(u64) * page_cnt, GFP_ATOMIC,
 419                                  rdsibdev_to_node(rds_ibdev));
 420         if (!dma_pages)
 421                 return -ENOMEM;
 422
 423         page_cnt = 0;
 424         for (i = 0; i < sg_dma_len; ++i) {
 425                 unsigned int dma_len = ib_sg_dma_len(dev, &scat[i]);
 426                 u64 dma_addr = ib_sg_dma_address(dev, &scat[i]);
 427
 428                 for (j = 0; j < dma_len; j += PAGE_SIZE)
 429                         dma_pages[page_cnt++] =
 430                                 (dma_addr & PAGE_MASK) + j;
 431         }
 432
 433         ret = ib_map_phys_fmr(ibmr->fmr,
 434                                    dma_pages, page_cnt, io_addr);
 435         if (ret)
 436                 goto out;
 437
 438         /* Success - we successfully remapped the MR, so we can
 439          * safely tear down the old mapping. */
 440         rds_ib_teardown_mr(ibmr);
 441
 442         ibmr->sg = scat;
 443         ibmr->sg_len = nents;
 444         ibmr->sg_dma_len = sg_dma_len;
 445         ibmr->remap_count++;
 446
 447         rds_ib_stats_inc(s_ib_rdma_mr_used);
 448         ret = 0;
 449
 450 out:
 451         kfree(dma_pages);
 452
 453         return ret;
 454 }
 455
 456 void rds_ib_sync_mr(void *trans_private, int direction)
 457 {
 458         struct rds_ib_mr *ibmr = trans_private;
 459         struct rds_ib_device *rds_ibdev = ibmr->device;
 460
 461         switch (direction) {
 462         case DMA_FROM_DEVICE:
 463                 ib_dma_sync_sg_for_cpu(rds_ibdev->dev, ibmr->sg,
 464                         ibmr->sg_dma_len, DMA_BIDIRECTIONAL);
 465                 break;
 466         case DMA_TO_DEVICE:
 467                 ib_dma_sync_sg_for_device(rds_ibdev->dev, ibmr->sg,
 468                         ibmr->sg_dma_len, DMA_BIDIRECTIONAL);
 469                 break;
 470         }
 471 }
 472
 473 static void __rds_ib_teardown_mr(struct rds_ib_mr *ibmr)
 474 {
 475         struct rds_ib_device *rds_ibdev = ibmr->device;
 476
 477         if (ibmr->sg_dma_len) {
 478                 ib_dma_unmap_sg(rds_ibdev->dev,
 479                                 ibmr->sg, ibmr->sg_len,
 480                                 DMA_BIDIRECTIONAL);
 481                 ibmr->sg_dma_len = 0;
 482         }
 483
 484         /* Release the s/g list */
 485         if (ibmr->sg_len) {
 486                 unsigned int i;
 487
 488                 for (i = 0; i < ibmr->sg_len; ++i) {
 489                         struct page *page = sg_page(&ibmr->sg[i]);
 490
 491                         /* FIXME we need a way to tell a r/w MR
 492                          * from a r/o MR */
 493                         BUG_ON(irqs_disabled());
 494                         set_page_dirty(page);
 495                         put_page(page);
 496                 }
 497                 kfree(ibmr->sg);
 498
 499                 ibmr->sg = NULL;
 500                 ibmr->sg_len = 0;
 501         }
 502 }
 503
 504 static void rds_ib_teardown_mr(struct rds_ib_mr *ibmr)
 505 {
 506         unsigned int pinned = ibmr->sg_len;
 507
 508         __rds_ib_teardown_mr(ibmr);
 509         if (pinned) {
 510                 struct rds_ib_device *rds_ibdev = ibmr->device;
 511                 struct rds_ib_mr_pool *pool = rds_ibdev->mr_pool;
 512
 513                 atomic_sub(pinned, &pool->free_pinned);
 514         }
 515 }
 516
 517 static inline unsigned int rds_ib_flush_goal(struct rds_ib_mr_pool *pool, int free_all)
 518 {
 519         unsigned int item_count;
 520
 521         item_count = atomic_read(&pool->item_count);
 522         if (free_all)
 523                 return item_count;
 524
 525         return 0;
 526 }
 527
 528 /*
 529  * given an xlist of mrs, put them all into the list_head for more processing
 530  */
 531 static void xlist_append_to_list(struct xlist_head *xlist, struct list_head *list)
 532 {
 533         struct rds_ib_mr *ibmr;
 534         struct xlist_head splice;
 535         struct xlist_head *cur;
 536         struct xlist_head *next;
 537
 538         splice.next = NULL;
 539         xlist_splice(xlist, &splice);
 540         cur = splice.next;
 541         while (cur) {
 542                 next = cur->next;
 543                 ibmr = list_entry(cur, struct rds_ib_mr, xlist);
 544                 list_add_tail(&ibmr->unmap_list, list);
 545                 cur = next;
 546         }
 547 }
 548
 549 /*
 550  * this takes a list head of mrs and turns it into an xlist of clusters.
 551  * each cluster has an xlist of MR_CLUSTER_SIZE mrs that are ready for
 552  * reuse.
 553  */
 554 static void list_append_to_xlist(struct rds_ib_mr_pool *pool,
 555                                 struct list_head *list, struct xlist_head *xlist,
 556                                 struct xlist_head **tail_ret)
 557 {
 558         struct rds_ib_mr *ibmr;
 559         struct xlist_head *cur_mr = xlist;
 560         struct xlist_head *tail_mr = NULL;
 561
 562         list_for_each_entry(ibmr, list, unmap_list) {
 563                 tail_mr = &ibmr->xlist;
 564                 tail_mr->next = NULL;
 565                 cur_mr->next = tail_mr;
 566                 cur_mr = tail_mr;
 567         }
 568         *tail_ret = tail_mr;
 569 }
 570
 571 /*
 572  * Flush our pool of MRs.
 573  * At a minimum, all currently unused MRs are unmapped.
 574  * If the number of MRs allocated exceeds the limit, we also try
 575  * to free as many MRs as needed to get back to this limit.
 576  */
 577 static int rds_ib_flush_mr_pool(struct rds_ib_mr_pool *pool,
 578                                 int free_all, struct rds_ib_mr **ibmr_ret)
 579 {
 580         struct rds_ib_mr *ibmr, *next;
 581         struct xlist_head clean_xlist;
 582         struct xlist_head *clean_tail;
 583         LIST_HEAD(unmap_list);
 584         LIST_HEAD(fmr_list);
 585         unsigned long unpinned = 0;
 586         unsigned int nfreed = 0, ncleaned = 0, free_goal;
 587         int ret = 0;
 588
 589         rds_ib_stats_inc(s_ib_rdma_mr_pool_flush);
 590
 591         if (ibmr_ret) {
 592                 DEFINE_WAIT(wait);
 593                 while(!mutex_trylock(&pool->flush_lock)) {
 594                         ibmr = rds_ib_reuse_fmr(pool);
 595                         if (ibmr) {
 596                                 *ibmr_ret = ibmr;
 597                                 finish_wait(&pool->flush_wait, &wait);
 598                                 goto out_nolock;
 599                         }
 600
 601                         prepare_to_wait(&pool->flush_wait, &wait,
 602                                         TASK_UNINTERRUPTIBLE);
 603                         if (xlist_empty(&pool->clean_list))
 604                                 schedule();
 605
 606                         ibmr = rds_ib_reuse_fmr(pool);
 607                         if (ibmr) {
 608                                 *ibmr_ret = ibmr;
 609                                 finish_wait(&pool->flush_wait, &wait);
 610                                 goto out_nolock;
 611                         }
 612                 }
 613                 finish_wait(&pool->flush_wait, &wait);
 614         } else
 615                 mutex_lock(&pool->flush_lock);
 616
 617         if (ibmr_ret) {
 618                 ibmr = rds_ib_reuse_fmr(pool);
 619                 if (ibmr) {
 620                         *ibmr_ret = ibmr;
 621                         goto out;
 622                 }
 623         }
 624
 625         /* Get the list of all MRs to be dropped. Ordering matters -
 626          * we want to put drop_list ahead of free_list.
 627          */
 628         xlist_append_to_list(&pool->drop_list, &unmap_list);
 629         xlist_append_to_list(&pool->free_list, &unmap_list);
 630         if (free_all)
 631                 xlist_append_to_list(&pool->clean_list, &unmap_list);
 632
 633         free_goal = rds_ib_flush_goal(pool, free_all);
 634
 635         if (list_empty(&unmap_list))
 636                 goto out;
 637
 638         /* String all ib_mr's onto one list and hand them to ib_unmap_fmr */
 639         list_for_each_entry(ibmr, &unmap_list, unmap_list)
 640                 list_add(&ibmr->fmr->list, &fmr_list);
 641
 642         ret = ib_unmap_fmr(&fmr_list);
 643         if (ret)
 644                 printk(KERN_WARNING "RDS/IB: ib_unmap_fmr failed (err=%d)\n", ret);
 645
 646         /* Now we can destroy the DMA mapping and unpin any pages */
 647         list_for_each_entry_safe(ibmr, next, &unmap_list, unmap_list) {
 648                 unpinned += ibmr->sg_len;
 649                 __rds_ib_teardown_mr(ibmr);
 650                 if (nfreed < free_goal || ibmr->remap_count >= pool->fmr_attr.max_maps) {
 651                         rds_ib_stats_inc(s_ib_rdma_mr_free);
 652                         list_del(&ibmr->unmap_list);
 653                         ib_dealloc_fmr(ibmr->fmr);
 654                         kfree(ibmr);
 655                         nfreed++;
 656                 }
 657                 ncleaned++;
 658         }
 659
 660         if (!list_empty(&unmap_list)) {
 661                 /* we have to make sure that none of the things we're about
 662                  * to put on the clean list would race with other cpus trying
 663                  * to pull items off.  The xlist would explode if we managed to
 664                  * remove something from the clean list and then add it back again
 665                  * while another CPU was spinning on that same item in xlist_del_head.
 666                  *
 667                  * This is pretty unlikely, but just in case  wait for an xlist grace period
 668                  * here before adding anything back into the clean list.
 669                  */
 670                 wait_clean_list_grace();
 671
 672                 list_append_to_xlist(pool, &unmap_list, &clean_xlist, &clean_tail);
 673                 if (ibmr_ret)
 674                         refill_local(pool, &clean_xlist, ibmr_ret);
 675
 676                 /* refill_local may have emptied our list */
 677                 if (!xlist_empty(&clean_xlist))
 678                         xlist_add(clean_xlist.next, clean_tail, &pool->clean_list);
 679
 680         }
 681
 682         atomic_sub(unpinned, &pool->free_pinned);
 683         atomic_sub(ncleaned, &pool->dirty_count);
 684         atomic_sub(nfreed, &pool->item_count);
 685
 686 out:
 687         mutex_unlock(&pool->flush_lock);
 688         if (waitqueue_active(&pool->flush_wait))
 689                 wake_up(&pool->flush_wait);
 690 out_nolock:
 691         return ret;
 692 }
 693
 694 struct workqueue_struct *rds_ib_fmr_wq;
 695
 696 int rds_ib_fmr_init(void)
 697 {
 698         rds_ib_fmr_wq = create_workqueue("rds_fmr_flushd");
 699         if (!rds_ib_fmr_wq)
 700                 return -ENOMEM;
 701         return 0;
 702 }
 703
 704 /*
 705  * By the time this is called all the IB devices should have been torn down and
 706  * had their pools freed.  As each pool is freed its work struct is waited on,
 707  * so the pool flushing work queue should be idle by the time we get here.
 708  */
 709 void rds_ib_fmr_exit(void)
 710 {
 711         destroy_workqueue(rds_ib_fmr_wq);
 712 }
 713
 714 static void rds_ib_mr_pool_flush_worker(struct work_struct *work)
 715 {
 716         struct rds_ib_mr_pool *pool = container_of(work, struct rds_ib_mr_pool, flush_worker.work);
 717
 718         rds_ib_flush_mr_pool(pool, 0, NULL);
 719 }
 720
 721 void rds_ib_free_mr(void *trans_private, int invalidate)
 722 {
 723         struct rds_ib_mr *ibmr = trans_private;
 724         struct rds_ib_device *rds_ibdev = ibmr->device;
 725         struct rds_ib_mr_pool *pool = rds_ibdev->mr_pool;
 726
 727         rdsdebug("RDS/IB: free_mr nents %u\n", ibmr->sg_len);
 728
 729         /* Return it to the pool's free list */
 730         if (ibmr->remap_count >= pool->fmr_attr.max_maps)
 731                 xlist_add(&ibmr->xlist, &ibmr->xlist, &pool->drop_list);
 732         else
 733                 xlist_add(&ibmr->xlist, &ibmr->xlist, &pool->free_list);
 734
 735         atomic_add(ibmr->sg_len, &pool->free_pinned);
 736         atomic_inc(&pool->dirty_count);
 737
 738         /* If we've pinned too many pages, request a flush */
 739         if (atomic_read(&pool->free_pinned) >= pool->max_free_pinned ||
 740             atomic_read(&pool->dirty_count) >= pool->max_items / 10)
 741                 queue_delayed_work(rds_ib_fmr_wq, &pool->flush_worker, 10);
 742
 743         if (invalidate) {
 744                 if (likely(!in_interrupt())) {
 745                         rds_ib_flush_mr_pool(pool, 0, NULL);
 746                 } else {
 747                         /* We get here if the user created a MR marked
 748                          * as use_once and invalidate at the same time. */
 749                         queue_delayed_work(rds_ib_fmr_wq,
 750                                            &pool->flush_worker, 10);
 751                 }
 752         }
 753
 754         rds_ib_dev_put(rds_ibdev);
 755 }
 756
 757 void rds_ib_flush_mrs(void)
 758 {
 759         struct rds_ib_device *rds_ibdev;
 760
 761         list_for_each_entry(rds_ibdev, &rds_ib_devices, list) {
 762                 struct rds_ib_mr_pool *pool = rds_ibdev->mr_pool;
 763
 764                 if (pool)
 765                         rds_ib_flush_mr_pool(pool, 0, NULL);
 766         }
 767 }
 768
 769 void *rds_ib_get_mr(struct scatterlist *sg, unsigned long nents,
 770                     struct rds_sock *rs, u32 *key_ret)
 771 {
 772         struct rds_ib_device *rds_ibdev;
 773         struct rds_ib_mr *ibmr = NULL;
 774         int ret;
 775
 776         rds_ibdev = rds_ib_get_device(rs->rs_bound_addr);
 777         if (!rds_ibdev) {
 778                 ret = -ENODEV;
 779                 goto out;
 780         }
 781
 782         if (!rds_ibdev->mr_pool) {
 783                 ret = -ENODEV;
 784                 goto out;
 785         }
 786
 787         ibmr = rds_ib_alloc_fmr(rds_ibdev);
 788         if (IS_ERR(ibmr))
 789                 return ibmr;
 790
 791         ret = rds_ib_map_fmr(rds_ibdev, ibmr, sg, nents);
 792         if (ret == 0)
 793                 *key_ret = ibmr->fmr->rkey;
 794         else
 795                 printk(KERN_WARNING "RDS/IB: map_fmr failed (errno=%d)\n", ret);
 796
 797         ibmr->device = rds_ibdev;
 798         rds_ibdev = NULL;
 799
 800  out:
 801         if (ret) {
 802                 if (ibmr)
 803                         rds_ib_free_mr(ibmr, 0);
 804                 ibmr = ERR_PTR(ret);
 805         }
 806         if (rds_ibdev)
 807                 rds_ib_dev_put(rds_ibdev);
 808         return ibmr;
 809 }
 810