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         rcu_read_lock();
  96         list_for_each_entry_rcu(rds_ibdev, &rds_ib_devices, list) {
  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         }
 105         rcu_read_unlock();
 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(&rds_ibdev->spinlock);
 173         list_add_tail(&ic->ib_node, &rds_ibdev->conn_list);
 174         spin_unlock(&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         if (atomic_read(&pool->dirty_count) >= pool->max_items / 10)
 308                 schedule_delayed_work(&pool->flush_worker, 10);
 309
 310         while (1) {
 311                 ibmr = rds_ib_reuse_fmr(pool);
 312                 if (ibmr)
 313                         return ibmr;
 314
 315                 /* No clean MRs - now we have the choice of either
 316                  * allocating a fresh MR up to the limit imposed by the
 317                  * driver, or flush any dirty unused MRs.
 318                  * We try to avoid stalling in the send path if possible,
 319                  * so we allocate as long as we're allowed to.
 320                  *
 321                  * We're fussy with enforcing the FMR limit, though. If the driver
 322                  * tells us we can't use more than N fmrs, we shouldn't start
 323                  * arguing with it */
 324                 if (atomic_inc_return(&pool->item_count) <= pool->max_items)
 325                         break;
 326
 327                 atomic_dec(&pool->item_count);
 328
 329                 if (++iter > 2) {
 330                         rds_ib_stats_inc(s_ib_rdma_mr_pool_depleted);
 331                         return ERR_PTR(-EAGAIN);
 332                 }
 333
 334                 /* We do have some empty MRs. Flush them out. */
 335                 rds_ib_stats_inc(s_ib_rdma_mr_pool_wait);
 336                 rds_ib_flush_mr_pool(pool, 0, &ibmr);
 337                 if (ibmr)
 338                         return ibmr;
 339         }
 340
 341         ibmr = kzalloc_node(sizeof(*ibmr), GFP_KERNEL, rdsibdev_to_node(rds_ibdev));
 342         if (!ibmr) {
 343                 err = -ENOMEM;
 344                 goto out_no_cigar;
 345         }
 346
 347         memset(ibmr, 0, sizeof(*ibmr));
 348
 349         ibmr->fmr = ib_alloc_fmr(rds_ibdev->pd,
 350                         (IB_ACCESS_LOCAL_WRITE |
 351                          IB_ACCESS_REMOTE_READ |
 352                          IB_ACCESS_REMOTE_WRITE|
 353                          IB_ACCESS_REMOTE_ATOMIC),
 354                         &pool->fmr_attr);
 355         if (IS_ERR(ibmr->fmr)) {
 356                 err = PTR_ERR(ibmr->fmr);
 357                 ibmr->fmr = NULL;
 358                 printk(KERN_WARNING "RDS/IB: ib_alloc_fmr failed (err=%d)\n", err);
 359                 goto out_no_cigar;
 360         }
 361
 362         rds_ib_stats_inc(s_ib_rdma_mr_alloc);
 363         return ibmr;
 364
 365 out_no_cigar:
 366         if (ibmr) {
 367                 if (ibmr->fmr)
 368                         ib_dealloc_fmr(ibmr->fmr);
 369                 kfree(ibmr);
 370         }
 371         atomic_dec(&pool->item_count);
 372         return ERR_PTR(err);
 373 }
 374
 375 static int rds_ib_map_fmr(struct rds_ib_device *rds_ibdev, struct rds_ib_mr *ibmr,
 376                struct scatterlist *sg, unsigned int nents)
 377 {
 378         struct ib_device *dev = rds_ibdev->dev;
 379         struct scatterlist *scat = sg;
 380         u64 io_addr = 0;
 381         u64 *dma_pages;
 382         u32 len;
 383         int page_cnt, sg_dma_len;
 384         int i, j;
 385         int ret;
 386
 387         sg_dma_len = ib_dma_map_sg(dev, sg, nents,
 388                                  DMA_BIDIRECTIONAL);
 389         if (unlikely(!sg_dma_len)) {
 390                 printk(KERN_WARNING "RDS/IB: dma_map_sg failed!\n");
 391                 return -EBUSY;
 392         }
 393
 394         len = 0;
 395         page_cnt = 0;
 396
 397         for (i = 0; i < sg_dma_len; ++i) {
 398                 unsigned int dma_len = ib_sg_dma_len(dev, &scat[i]);
 399                 u64 dma_addr = ib_sg_dma_address(dev, &scat[i]);
 400
 401                 if (dma_addr & ~PAGE_MASK) {
 402                         if (i > 0)
 403                                 return -EINVAL;
 404                         else
 405                                 ++page_cnt;
 406                 }
 407                 if ((dma_addr + dma_len) & ~PAGE_MASK) {
 408                         if (i < sg_dma_len - 1)
 409                                 return -EINVAL;
 410                         else
 411                                 ++page_cnt;
 412                 }
 413
 414                 len += dma_len;
 415         }
 416
 417         page_cnt += len >> PAGE_SHIFT;
 418         if (page_cnt > fmr_message_size)
 419                 return -EINVAL;
 420
 421         dma_pages = kmalloc_node(sizeof(u64) * page_cnt, GFP_ATOMIC,
 422                                  rdsibdev_to_node(rds_ibdev));
 423         if (!dma_pages)
 424                 return -ENOMEM;
 425
 426         page_cnt = 0;
 427         for (i = 0; i < sg_dma_len; ++i) {
 428                 unsigned int dma_len = ib_sg_dma_len(dev, &scat[i]);
 429                 u64 dma_addr = ib_sg_dma_address(dev, &scat[i]);
 430
 431                 for (j = 0; j < dma_len; j += PAGE_SIZE)
 432                         dma_pages[page_cnt++] =
 433                                 (dma_addr & PAGE_MASK) + j;
 434         }
 435
 436         ret = ib_map_phys_fmr(ibmr->fmr,
 437                                    dma_pages, page_cnt, io_addr);
 438         if (ret)
 439                 goto out;
 440
 441         /* Success - we successfully remapped the MR, so we can
 442          * safely tear down the old mapping. */
 443         rds_ib_teardown_mr(ibmr);
 444
 445         ibmr->sg = scat;
 446         ibmr->sg_len = nents;
 447         ibmr->sg_dma_len = sg_dma_len;
 448         ibmr->remap_count++;
 449
 450         rds_ib_stats_inc(s_ib_rdma_mr_used);
 451         ret = 0;
 452
 453 out:
 454         kfree(dma_pages);
 455
 456         return ret;
 457 }
 458
 459 void rds_ib_sync_mr(void *trans_private, int direction)
 460 {
 461         struct rds_ib_mr *ibmr = trans_private;
 462         struct rds_ib_device *rds_ibdev = ibmr->device;
 463
 464         switch (direction) {
 465         case DMA_FROM_DEVICE:
 466                 ib_dma_sync_sg_for_cpu(rds_ibdev->dev, ibmr->sg,
 467                         ibmr->sg_dma_len, DMA_BIDIRECTIONAL);
 468                 break;
 469         case DMA_TO_DEVICE:
 470                 ib_dma_sync_sg_for_device(rds_ibdev->dev, ibmr->sg,
 471                         ibmr->sg_dma_len, DMA_BIDIRECTIONAL);
 472                 break;
 473         }
 474 }
 475
 476 static void __rds_ib_teardown_mr(struct rds_ib_mr *ibmr)
 477 {
 478         struct rds_ib_device *rds_ibdev = ibmr->device;
 479
 480         if (ibmr->sg_dma_len) {
 481                 ib_dma_unmap_sg(rds_ibdev->dev,
 482                                 ibmr->sg, ibmr->sg_len,
 483                                 DMA_BIDIRECTIONAL);
 484                 ibmr->sg_dma_len = 0;
 485         }
 486
 487         /* Release the s/g list */
 488         if (ibmr->sg_len) {
 489                 unsigned int i;
 490
 491                 for (i = 0; i < ibmr->sg_len; ++i) {
 492                         struct page *page = sg_page(&ibmr->sg[i]);
 493
 494                         /* FIXME we need a way to tell a r/w MR
 495                          * from a r/o MR */
 496                         BUG_ON(irqs_disabled());
 497                         set_page_dirty(page);
 498                         put_page(page);
 499                 }
 500                 kfree(ibmr->sg);
 501
 502                 ibmr->sg = NULL;
 503                 ibmr->sg_len = 0;
 504         }
 505 }
 506
 507 static void rds_ib_teardown_mr(struct rds_ib_mr *ibmr)
 508 {
 509         unsigned int pinned = ibmr->sg_len;
 510
 511         __rds_ib_teardown_mr(ibmr);
 512         if (pinned) {
 513                 struct rds_ib_device *rds_ibdev = ibmr->device;
 514                 struct rds_ib_mr_pool *pool = rds_ibdev->mr_pool;
 515
 516                 atomic_sub(pinned, &pool->free_pinned);
 517         }
 518 }
 519
 520 static inline unsigned int rds_ib_flush_goal(struct rds_ib_mr_pool *pool, int free_all)
 521 {
 522         unsigned int item_count;
 523
 524         item_count = atomic_read(&pool->item_count);
 525         if (free_all)
 526                 return item_count;
 527
 528         return 0;
 529 }
 530
 531 /*
 532  * given an xlist of mrs, put them all into the list_head for more processing
 533  */
 534 static void xlist_append_to_list(struct xlist_head *xlist, struct list_head *list)
 535 {
 536         struct rds_ib_mr *ibmr;
 537         struct xlist_head splice;
 538         struct xlist_head *cur;
 539         struct xlist_head *next;
 540
 541         splice.next = NULL;
 542         xlist_splice(xlist, &splice);
 543         cur = splice.next;
 544         while (cur) {
 545                 next = cur->next;
 546                 ibmr = list_entry(cur, struct rds_ib_mr, xlist);
 547                 list_add_tail(&ibmr->unmap_list, list);
 548                 cur = next;
 549         }
 550 }
 551
 552 /*
 553  * this takes a list head of mrs and turns it into an xlist of clusters.
 554  * each cluster has an xlist of MR_CLUSTER_SIZE mrs that are ready for
 555  * reuse.
 556  */
 557 static void list_append_to_xlist(struct rds_ib_mr_pool *pool,
 558                                 struct list_head *list, struct xlist_head *xlist,
 559                                 struct xlist_head **tail_ret)
 560 {
 561         struct rds_ib_mr *ibmr;
 562         struct xlist_head *cur_mr = xlist;
 563         struct xlist_head *tail_mr = NULL;
 564
 565         list_for_each_entry(ibmr, list, unmap_list) {
 566                 tail_mr = &ibmr->xlist;
 567                 tail_mr->next = NULL;
 568                 cur_mr->next = tail_mr;
 569                 cur_mr = tail_mr;
 570         }
 571         *tail_ret = tail_mr;
 572 }
 573
 574 /*
 575  * Flush our pool of MRs.
 576  * At a minimum, all currently unused MRs are unmapped.
 577  * If the number of MRs allocated exceeds the limit, we also try
 578  * to free as many MRs as needed to get back to this limit.
 579  */
 580 static int rds_ib_flush_mr_pool(struct rds_ib_mr_pool *pool,
 581                                 int free_all, struct rds_ib_mr **ibmr_ret)
 582 {
 583         struct rds_ib_mr *ibmr, *next;
 584         struct xlist_head clean_xlist;
 585         struct xlist_head *clean_tail;
 586         LIST_HEAD(unmap_list);
 587         LIST_HEAD(fmr_list);
 588         unsigned long unpinned = 0;
 589         unsigned int nfreed = 0, ncleaned = 0, free_goal;
 590         int ret = 0;
 591
 592         rds_ib_stats_inc(s_ib_rdma_mr_pool_flush);
 593
 594         if (ibmr_ret) {
 595                 DEFINE_WAIT(wait);
 596                 while(!mutex_trylock(&pool->flush_lock)) {
 597                         ibmr = rds_ib_reuse_fmr(pool);
 598                         if (ibmr) {
 599                                 *ibmr_ret = ibmr;
 600                                 finish_wait(&pool->flush_wait, &wait);
 601                                 goto out_nolock;
 602                         }
 603
 604                         prepare_to_wait(&pool->flush_wait, &wait,
 605                                         TASK_UNINTERRUPTIBLE);
 606                         if (xlist_empty(&pool->clean_list))
 607                                 schedule();
 608
 609                         ibmr = rds_ib_reuse_fmr(pool);
 610                         if (ibmr) {
 611                                 *ibmr_ret = ibmr;
 612                                 finish_wait(&pool->flush_wait, &wait);
 613                                 goto out_nolock;
 614                         }
 615                 }
 616                 finish_wait(&pool->flush_wait, &wait);
 617         } else
 618                 mutex_lock(&pool->flush_lock);
 619
 620         if (ibmr_ret) {
 621                 ibmr = rds_ib_reuse_fmr(pool);
 622                 if (ibmr) {
 623                         *ibmr_ret = ibmr;
 624                         goto out;
 625                 }
 626         }
 627
 628         /* Get the list of all MRs to be dropped. Ordering matters -
 629          * we want to put drop_list ahead of free_list.
 630          */
 631         xlist_append_to_list(&pool->drop_list, &unmap_list);
 632         xlist_append_to_list(&pool->free_list, &unmap_list);
 633         if (free_all)
 634                 xlist_append_to_list(&pool->clean_list, &unmap_list);
 635
 636         free_goal = rds_ib_flush_goal(pool, free_all);
 637
 638         if (list_empty(&unmap_list))
 639                 goto out;
 640
 641         /* String all ib_mr's onto one list and hand them to ib_unmap_fmr */
 642         list_for_each_entry(ibmr, &unmap_list, unmap_list)
 643                 list_add(&ibmr->fmr->list, &fmr_list);
 644
 645         ret = ib_unmap_fmr(&fmr_list);
 646         if (ret)
 647                 printk(KERN_WARNING "RDS/IB: ib_unmap_fmr failed (err=%d)\n", ret);
 648
 649         /* Now we can destroy the DMA mapping and unpin any pages */
 650         list_for_each_entry_safe(ibmr, next, &unmap_list, unmap_list) {
 651                 unpinned += ibmr->sg_len;
 652                 __rds_ib_teardown_mr(ibmr);
 653                 if (nfreed < free_goal || ibmr->remap_count >= pool->fmr_attr.max_maps) {
 654                         rds_ib_stats_inc(s_ib_rdma_mr_free);
 655                         list_del(&ibmr->unmap_list);
 656                         ib_dealloc_fmr(ibmr->fmr);
 657                         kfree(ibmr);
 658                         nfreed++;
 659                 }
 660                 ncleaned++;
 661         }
 662
 663         if (!list_empty(&unmap_list)) {
 664                 /* we have to make sure that none of the things we're about
 665                  * to put on the clean list would race with other cpus trying
 666                  * to pull items off.  The xlist would explode if we managed to
 667                  * remove something from the clean list and then add it back again
 668                  * while another CPU was spinning on that same item in xlist_del_head.
 669                  *
 670                  * This is pretty unlikely, but just in case  wait for an xlist grace period
 671                  * here before adding anything back into the clean list.
 672                  */
 673                 wait_clean_list_grace();
 674
 675                 list_append_to_xlist(pool, &unmap_list, &clean_xlist, &clean_tail);
 676                 if (ibmr_ret)
 677                         refill_local(pool, &clean_xlist, ibmr_ret);
 678
 679                 /* refill_local may have emptied our list */
 680                 if (!xlist_empty(&clean_xlist))
 681                         xlist_add(clean_xlist.next, clean_tail, &pool->clean_list);
 682
 683         }
 684
 685         atomic_sub(unpinned, &pool->free_pinned);
 686         atomic_sub(ncleaned, &pool->dirty_count);
 687         atomic_sub(nfreed, &pool->item_count);
 688
 689 out:
 690         mutex_unlock(&pool->flush_lock);
 691         if (waitqueue_active(&pool->flush_wait))
 692                 wake_up(&pool->flush_wait);
 693 out_nolock:
 694         return ret;
 695 }
 696
 697 static void rds_ib_mr_pool_flush_worker(struct work_struct *work)
 698 {
 699         struct rds_ib_mr_pool *pool = container_of(work, struct rds_ib_mr_pool, flush_worker.work);
 700
 701         rds_ib_flush_mr_pool(pool, 0, NULL);
 702 }
 703
 704 void rds_ib_free_mr(void *trans_private, int invalidate)
 705 {
 706         struct rds_ib_mr *ibmr = trans_private;
 707         struct rds_ib_device *rds_ibdev = ibmr->device;
 708         struct rds_ib_mr_pool *pool = rds_ibdev->mr_pool;
 709
 710         rdsdebug("RDS/IB: free_mr nents %u\n", ibmr->sg_len);
 711
 712         /* Return it to the pool's free list */
 713         if (ibmr->remap_count >= pool->fmr_attr.max_maps)
 714                 xlist_add(&ibmr->xlist, &ibmr->xlist, &pool->drop_list);
 715         else
 716                 xlist_add(&ibmr->xlist, &ibmr->xlist, &pool->free_list);
 717
 718         atomic_add(ibmr->sg_len, &pool->free_pinned);
 719         atomic_inc(&pool->dirty_count);
 720
 721         /* If we've pinned too many pages, request a flush */
 722         if (atomic_read(&pool->free_pinned) >= pool->max_free_pinned ||
 723             atomic_read(&pool->dirty_count) >= pool->max_items / 10)
 724                 schedule_delayed_work(&pool->flush_worker, 10);
 725
 726         if (invalidate) {
 727                 if (likely(!in_interrupt())) {
 728                         rds_ib_flush_mr_pool(pool, 0, NULL);
 729                 } else {
 730                         /* We get here if the user created a MR marked
 731                          * as use_once and invalidate at the same time. */
 732                         schedule_delayed_work(&pool->flush_worker, 10);
 733                 }
 734         }
 735
 736         rds_ib_dev_put(rds_ibdev);
 737 }
 738
 739 void rds_ib_flush_mrs(void)
 740 {
 741         struct rds_ib_device *rds_ibdev;
 742
 743         down_read(&rds_ib_devices_lock);
 744         list_for_each_entry(rds_ibdev, &rds_ib_devices, list) {
 745                 struct rds_ib_mr_pool *pool = rds_ibdev->mr_pool;
 746
 747                 if (pool)
 748                         rds_ib_flush_mr_pool(pool, 0, NULL);
 749         }
 750         up_read(&rds_ib_devices_lock);
 751 }
 752
 753 void *rds_ib_get_mr(struct scatterlist *sg, unsigned long nents,
 754                     struct rds_sock *rs, u32 *key_ret)
 755 {
 756         struct rds_ib_device *rds_ibdev;
 757         struct rds_ib_mr *ibmr = NULL;
 758         int ret;
 759
 760         rds_ibdev = rds_ib_get_device(rs->rs_bound_addr);
 761         if (!rds_ibdev) {
 762                 ret = -ENODEV;
 763                 goto out;
 764         }
 765
 766         if (!rds_ibdev->mr_pool) {
 767                 ret = -ENODEV;
 768                 goto out;
 769         }
 770
 771         ibmr = rds_ib_alloc_fmr(rds_ibdev);
 772         if (IS_ERR(ibmr))
 773                 return ibmr;
 774
 775         ret = rds_ib_map_fmr(rds_ibdev, ibmr, sg, nents);
 776         if (ret == 0)
 777                 *key_ret = ibmr->fmr->rkey;
 778         else
 779                 printk(KERN_WARNING "RDS/IB: map_fmr failed (errno=%d)\n", ret);
 780
 781         ibmr->device = rds_ibdev;
 782         rds_ibdev = NULL;
 783
 784  out:
 785         if (ret) {
 786                 if (ibmr)
 787                         rds_ib_free_mr(ibmr, 0);
 788                 ibmr = ERR_PTR(ret);
 789         }
 790         if (rds_ibdev)
 791                 rds_ib_dev_put(rds_ibdev);
 792         return ibmr;
 793 }
 794