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_conns(struct list_head *list, spinlock_t *list_lock)
 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(list_lock);
 208         list_splice(list, &tmp_list);
 209         INIT_LIST_HEAD(list);
 210         spin_unlock_irq(list_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_XLIST_HEAD(&pool->free_list);
 225         INIT_XLIST_HEAD(&pool->drop_list);
 226         INIT_XLIST_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 void refill_local(struct rds_ib_mr_pool *pool, struct xlist_head *xl,
 265                          struct rds_ib_mr **ibmr_ret)
 266 {
 267         struct xlist_head *ibmr_xl;
 268         ibmr_xl = xlist_del_head_fast(xl);
 269         *ibmr_ret = list_entry(ibmr_xl, struct rds_ib_mr, xlist);
 270 }
 271
 272 static inline struct rds_ib_mr *rds_ib_reuse_fmr(struct rds_ib_mr_pool *pool)
 273 {
 274         struct rds_ib_mr *ibmr = NULL;
 275         struct xlist_head *ret;
 276         unsigned long *flag;
 277
 278         preempt_disable();
 279         flag = &__get_cpu_var(clean_list_grace);
 280         set_bit(CLEAN_LIST_BUSY_BIT, flag);
 281         ret = xlist_del_head(&pool->clean_list);
 282         if (ret)
 283                 ibmr = list_entry(ret, struct rds_ib_mr, xlist);
 284
 285         clear_bit(CLEAN_LIST_BUSY_BIT, flag);
 286         preempt_enable();
 287         return ibmr;
 288 }
 289
 290 static inline void wait_clean_list_grace(void)
 291 {
 292         int cpu;
 293         unsigned long *flag;
 294
 295         for_each_online_cpu(cpu) {
 296                 flag = &per_cpu(clean_list_grace, cpu);
 297                 while (test_bit(CLEAN_LIST_BUSY_BIT, flag))
 298                         cpu_relax();
 299         }
 300 }
 301
 302 static struct rds_ib_mr *rds_ib_alloc_fmr(struct rds_ib_device *rds_ibdev)
 303 {
 304         struct rds_ib_mr_pool *pool = rds_ibdev->mr_pool;
 305         struct rds_ib_mr *ibmr = NULL;
 306         int err = 0, iter = 0;
 307
 308         while (1) {
 309                 ibmr = rds_ib_reuse_fmr(pool);
 310                 if (ibmr)
 311                         return ibmr;
 312
 313                 /* No clean MRs - now we have the choice of either
 314                  * allocating a fresh MR up to the limit imposed by the
 315                  * driver, or flush any dirty unused MRs.
 316                  * We try to avoid stalling in the send path if possible,
 317                  * so we allocate as long as we're allowed to.
 318                  *
 319                  * We're fussy with enforcing the FMR limit, though. If the driver
 320                  * tells us we can't use more than N fmrs, we shouldn't start
 321                  * arguing with it */
 322                 if (atomic_inc_return(&pool->item_count) <= pool->max_items)
 323                         break;
 324
 325                 atomic_dec(&pool->item_count);
 326
 327                 if (++iter > 2) {
 328                         rds_ib_stats_inc(s_ib_rdma_mr_pool_depleted);
 329                         return ERR_PTR(-EAGAIN);
 330                 }
 331
 332                 /* We do have some empty MRs. Flush them out. */
 333                 rds_ib_stats_inc(s_ib_rdma_mr_pool_wait);
 334                 rds_ib_flush_mr_pool(pool, 0, &ibmr);
 335                 if (ibmr)
 336                         return ibmr;
 337         }
 338
 339         ibmr = kzalloc_node(sizeof(*ibmr), GFP_KERNEL, rdsibdev_to_node(rds_ibdev));
 340         if (!ibmr) {
 341                 err = -ENOMEM;
 342                 goto out_no_cigar;
 343         }
 344
 345         memset(ibmr, 0, sizeof(*ibmr));
 346
 347         ibmr->fmr = ib_alloc_fmr(rds_ibdev->pd,
 348                         (IB_ACCESS_LOCAL_WRITE |
 349                          IB_ACCESS_REMOTE_READ |
 350                          IB_ACCESS_REMOTE_WRITE|
 351                          IB_ACCESS_REMOTE_ATOMIC),
 352
 353                         &pool->fmr_attr);
 354         if (IS_ERR(ibmr->fmr)) {
 355                 err = PTR_ERR(ibmr->fmr);
 356                 ibmr->fmr = NULL;
 357                 printk(KERN_WARNING "RDS/IB: ib_alloc_fmr failed (err=%d)\n", err);
 358                 goto out_no_cigar;
 359         }
 360
 361         rds_ib_stats_inc(s_ib_rdma_mr_alloc);
 362         return ibmr;
 363
 364 out_no_cigar:
 365         if (ibmr) {
 366                 if (ibmr->fmr)
 367                         ib_dealloc_fmr(ibmr->fmr);
 368                 kfree(ibmr);
 369         }
 370         atomic_dec(&pool->item_count);
 371         return ERR_PTR(err);
 372 }
 373
 374 static int rds_ib_map_fmr(struct rds_ib_device *rds_ibdev, struct rds_ib_mr *ibmr,
 375                struct scatterlist *sg, unsigned int nents)
 376 {
 377         struct ib_device *dev = rds_ibdev->dev;
 378         struct scatterlist *scat = sg;
 379         u64 io_addr = 0;
 380         u64 *dma_pages;
 381         u32 len;
 382         int page_cnt, sg_dma_len;
 383         int i, j;
 384         int ret;
 385
 386         sg_dma_len = ib_dma_map_sg(dev, sg, nents,
 387                                  DMA_BIDIRECTIONAL);
 388         if (unlikely(!sg_dma_len)) {
 389                 printk(KERN_WARNING "RDS/IB: dma_map_sg failed!\n");
 390                 return -EBUSY;
 391         }
 392
 393         len = 0;
 394         page_cnt = 0;
 395
 396         for (i = 0; i < sg_dma_len; ++i) {
 397                 unsigned int dma_len = ib_sg_dma_len(dev, &scat[i]);
 398                 u64 dma_addr = ib_sg_dma_address(dev, &scat[i]);
 399
 400                 if (dma_addr & ~PAGE_MASK) {
 401                         if (i > 0)
 402                                 return -EINVAL;
 403                         else
 404                                 ++page_cnt;
 405                 }
 406                 if ((dma_addr + dma_len) & ~PAGE_MASK) {
 407                         if (i < sg_dma_len - 1)
 408                                 return -EINVAL;
 409                         else
 410                                 ++page_cnt;
 411                 }
 412
 413                 len += dma_len;
 414         }
 415
 416         page_cnt += len >> PAGE_SHIFT;
 417         if (page_cnt > fmr_message_size)
 418                 return -EINVAL;
 419
 420         dma_pages = kmalloc_node(sizeof(u64) * page_cnt, GFP_ATOMIC,
 421                                  rdsibdev_to_node(rds_ibdev));
 422         if (!dma_pages)
 423                 return -ENOMEM;
 424
 425         page_cnt = 0;
 426         for (i = 0; i < sg_dma_len; ++i) {
 427                 unsigned int dma_len = ib_sg_dma_len(dev, &scat[i]);
 428                 u64 dma_addr = ib_sg_dma_address(dev, &scat[i]);
 429
 430                 for (j = 0; j < dma_len; j += PAGE_SIZE)
 431                         dma_pages[page_cnt++] =
 432                                 (dma_addr & PAGE_MASK) + j;
 433         }
 434
 435         ret = ib_map_phys_fmr(ibmr->fmr,
 436                                    dma_pages, page_cnt, io_addr);
 437         if (ret)
 438                 goto out;
 439
 440         /* Success - we successfully remapped the MR, so we can
 441          * safely tear down the old mapping. */
 442         rds_ib_teardown_mr(ibmr);
 443
 444         ibmr->sg = scat;
 445         ibmr->sg_len = nents;
 446         ibmr->sg_dma_len = sg_dma_len;
 447         ibmr->remap_count++;
 448
 449         rds_ib_stats_inc(s_ib_rdma_mr_used);
 450         ret = 0;
 451
 452 out:
 453         kfree(dma_pages);
 454
 455         return ret;
 456 }
 457
 458 void rds_ib_sync_mr(void *trans_private, int direction)
 459 {
 460         struct rds_ib_mr *ibmr = trans_private;
 461         struct rds_ib_device *rds_ibdev = ibmr->device;
 462
 463         switch (direction) {
 464         case DMA_FROM_DEVICE:
 465                 ib_dma_sync_sg_for_cpu(rds_ibdev->dev, ibmr->sg,
 466                         ibmr->sg_dma_len, DMA_BIDIRECTIONAL);
 467                 break;
 468         case DMA_TO_DEVICE:
 469                 ib_dma_sync_sg_for_device(rds_ibdev->dev, ibmr->sg,
 470                         ibmr->sg_dma_len, DMA_BIDIRECTIONAL);
 471                 break;
 472         }
 473 }
 474
 475 static void __rds_ib_teardown_mr(struct rds_ib_mr *ibmr)
 476 {
 477         struct rds_ib_device *rds_ibdev = ibmr->device;
 478
 479         if (ibmr->sg_dma_len) {
 480                 ib_dma_unmap_sg(rds_ibdev->dev,
 481                                 ibmr->sg, ibmr->sg_len,
 482                                 DMA_BIDIRECTIONAL);
 483                 ibmr->sg_dma_len = 0;
 484         }
 485
 486         /* Release the s/g list */
 487         if (ibmr->sg_len) {
 488                 unsigned int i;
 489
 490                 for (i = 0; i < ibmr->sg_len; ++i) {
 491                         struct page *page = sg_page(&ibmr->sg[i]);
 492
 493                         /* FIXME we need a way to tell a r/w MR
 494                          * from a r/o MR */
 495                         BUG_ON(irqs_disabled());
 496                         set_page_dirty(page);
 497                         put_page(page);
 498                 }
 499                 kfree(ibmr->sg);
 500
 501                 ibmr->sg = NULL;
 502                 ibmr->sg_len = 0;
 503         }
 504 }
 505
 506 static void rds_ib_teardown_mr(struct rds_ib_mr *ibmr)
 507 {
 508         unsigned int pinned = ibmr->sg_len;
 509
 510         __rds_ib_teardown_mr(ibmr);
 511         if (pinned) {
 512                 struct rds_ib_device *rds_ibdev = ibmr->device;
 513                 struct rds_ib_mr_pool *pool = rds_ibdev->mr_pool;
 514
 515                 atomic_sub(pinned, &pool->free_pinned);
 516         }
 517 }
 518
 519 static inline unsigned int rds_ib_flush_goal(struct rds_ib_mr_pool *pool, int free_all)
 520 {
 521         unsigned int item_count;
 522
 523         item_count = atomic_read(&pool->item_count);
 524         if (free_all)
 525                 return item_count;
 526
 527         return 0;
 528 }
 529
 530 /*
 531  * given an xlist of mrs, put them all into the list_head for more processing
 532  */
 533 static void xlist_append_to_list(struct xlist_head *xlist, struct list_head *list)
 534 {
 535         struct rds_ib_mr *ibmr;
 536         struct xlist_head splice;
 537         struct xlist_head *cur;
 538         struct xlist_head *next;
 539
 540         splice.next = NULL;
 541         xlist_splice(xlist, &splice);
 542         cur = splice.next;
 543         while (cur) {
 544                 next = cur->next;
 545                 ibmr = list_entry(cur, struct rds_ib_mr, xlist);
 546                 list_add_tail(&ibmr->unmap_list, list);
 547                 cur = next;
 548         }
 549 }
 550
 551 /*
 552  * this takes a list head of mrs and turns it into an xlist of clusters.
 553  * each cluster has an xlist of MR_CLUSTER_SIZE mrs that are ready for
 554  * reuse.
 555  */
 556 static void list_append_to_xlist(struct rds_ib_mr_pool *pool,
 557                                 struct list_head *list, struct xlist_head *xlist,
 558                                 struct xlist_head **tail_ret)
 559 {
 560         struct rds_ib_mr *ibmr;
 561         struct xlist_head *cur_mr = xlist;
 562         struct xlist_head *tail_mr = NULL;
 563
 564         list_for_each_entry(ibmr, list, unmap_list) {
 565                 tail_mr = &ibmr->xlist;
 566                 tail_mr->next = NULL;
 567                 cur_mr->next = tail_mr;
 568                 cur_mr = tail_mr;
 569         }
 570         *tail_ret = tail_mr;
 571 }
 572
 573 /*
 574  * Flush our pool of MRs.
 575  * At a minimum, all currently unused MRs are unmapped.
 576  * If the number of MRs allocated exceeds the limit, we also try
 577  * to free as many MRs as needed to get back to this limit.
 578  */
 579 static int rds_ib_flush_mr_pool(struct rds_ib_mr_pool *pool,
 580                                 int free_all, struct rds_ib_mr **ibmr_ret)
 581 {
 582         struct rds_ib_mr *ibmr, *next;
 583         struct xlist_head clean_xlist;
 584         struct xlist_head *clean_tail;
 585         LIST_HEAD(unmap_list);
 586         LIST_HEAD(fmr_list);
 587         unsigned long unpinned = 0;
 588         unsigned int nfreed = 0, ncleaned = 0, free_goal;
 589         int ret = 0;
 590
 591         rds_ib_stats_inc(s_ib_rdma_mr_pool_flush);
 592
 593         if (ibmr_ret) {
 594                 DEFINE_WAIT(wait);
 595                 while(!mutex_trylock(&pool->flush_lock)) {
 596                         ibmr = rds_ib_reuse_fmr(pool);
 597                         if (ibmr) {
 598                                 *ibmr_ret = ibmr;
 599                                 finish_wait(&pool->flush_wait, &wait);
 600                                 goto out_nolock;
 601                         }
 602
 603                         prepare_to_wait(&pool->flush_wait, &wait,
 604                                         TASK_UNINTERRUPTIBLE);
 605                         if (xlist_empty(&pool->clean_list))
 606                                 schedule();
 607
 608                         ibmr = rds_ib_reuse_fmr(pool);
 609                         if (ibmr) {
 610                                 *ibmr_ret = ibmr;
 611                                 finish_wait(&pool->flush_wait, &wait);
 612                                 goto out_nolock;
 613                         }
 614                 }
 615                 finish_wait(&pool->flush_wait, &wait);
 616         } else
 617                 mutex_lock(&pool->flush_lock);
 618
 619         if (ibmr_ret) {
 620                 ibmr = rds_ib_reuse_fmr(pool);
 621                 if (ibmr) {
 622                         *ibmr_ret = ibmr;
 623                         goto out;
 624                 }
 625         }
 626
 627         /* Get the list of all MRs to be dropped. Ordering matters -
 628          * we want to put drop_list ahead of free_list.
 629          */
 630         xlist_append_to_list(&pool->drop_list, &unmap_list);
 631         xlist_append_to_list(&pool->free_list, &unmap_list);
 632         if (free_all)
 633                 xlist_append_to_list(&pool->clean_list, &unmap_list);
 634
 635         free_goal = rds_ib_flush_goal(pool, free_all);
 636
 637         if (list_empty(&unmap_list))
 638                 goto out;
 639
 640         /* String all ib_mr's onto one list and hand them to ib_unmap_fmr */
 641         list_for_each_entry(ibmr, &unmap_list, unmap_list)
 642                 list_add(&ibmr->fmr->list, &fmr_list);
 643
 644         ret = ib_unmap_fmr(&fmr_list);
 645         if (ret)
 646                 printk(KERN_WARNING "RDS/IB: ib_unmap_fmr failed (err=%d)\n", ret);
 647
 648         /* Now we can destroy the DMA mapping and unpin any pages */
 649         list_for_each_entry_safe(ibmr, next, &unmap_list, unmap_list) {
 650                 unpinned += ibmr->sg_len;
 651                 __rds_ib_teardown_mr(ibmr);
 652                 if (nfreed < free_goal || ibmr->remap_count >= pool->fmr_attr.max_maps) {
 653                         rds_ib_stats_inc(s_ib_rdma_mr_free);
 654                         list_del(&ibmr->unmap_list);
 655                         ib_dealloc_fmr(ibmr->fmr);
 656                         kfree(ibmr);
 657                         nfreed++;
 658                 }
 659                 ncleaned++;
 660         }
 661
 662         if (!list_empty(&unmap_list)) {
 663                 /* we have to make sure that none of the things we're about
 664                  * to put on the clean list would race with other cpus trying
 665                  * to pull items off.  The xlist would explode if we managed to
 666                  * remove something from the clean list and then add it back again
 667                  * while another CPU was spinning on that same item in xlist_del_head.
 668                  *
 669                  * This is pretty unlikely, but just in case  wait for an xlist grace period
 670                  * here before adding anything back into the clean list.
 671                  */
 672                 wait_clean_list_grace();
 673
 674                 list_append_to_xlist(pool, &unmap_list, &clean_xlist, &clean_tail);
 675                 if (ibmr_ret)
 676                         refill_local(pool, &clean_xlist, ibmr_ret);
 677
 678                 /* refill_local may have emptied our list */
 679                 if (!xlist_empty(&clean_xlist))
 680                         xlist_add(clean_xlist.next, clean_tail, &pool->clean_list);
 681
 682         }
 683
 684         atomic_sub(unpinned, &pool->free_pinned);
 685         atomic_sub(ncleaned, &pool->dirty_count);
 686         atomic_sub(nfreed, &pool->item_count);
 687
 688 out:
 689         mutex_unlock(&pool->flush_lock);
 690         if (waitqueue_active(&pool->flush_wait))
 691                 wake_up(&pool->flush_wait);
 692 out_nolock:
 693         return ret;
 694 }
 695
 696 static void rds_ib_mr_pool_flush_worker(struct work_struct *work)
 697 {
 698         struct rds_ib_mr_pool *pool = container_of(work, struct rds_ib_mr_pool, flush_worker.work);
 699
 700         rds_ib_flush_mr_pool(pool, 0, NULL);
 701 }
 702
 703 void rds_ib_free_mr(void *trans_private, int invalidate)
 704 {
 705         struct rds_ib_mr *ibmr = trans_private;
 706         struct rds_ib_device *rds_ibdev = ibmr->device;
 707         struct rds_ib_mr_pool *pool = rds_ibdev->mr_pool;
 708
 709         rdsdebug("RDS/IB: free_mr nents %u\n", ibmr->sg_len);
 710
 711         /* Return it to the pool's free list */
 712         if (ibmr->remap_count >= pool->fmr_attr.max_maps)
 713                 xlist_add(&ibmr->xlist, &ibmr->xlist, &pool->drop_list);
 714         else
 715                 xlist_add(&ibmr->xlist, &ibmr->xlist, &pool->free_list);
 716
 717         atomic_add(ibmr->sg_len, &pool->free_pinned);
 718         atomic_inc(&pool->dirty_count);
 719
 720         /* If we've pinned too many pages, request a flush */
 721         if (atomic_read(&pool->free_pinned) >= pool->max_free_pinned ||
 722             atomic_read(&pool->dirty_count) >= pool->max_items / 10)
 723                 queue_delayed_work(rds_wq, &pool->flush_worker, 10);
 724
 725         if (invalidate) {
 726                 if (likely(!in_interrupt())) {
 727                         rds_ib_flush_mr_pool(pool, 0, NULL);
 728                 } else {
 729                         /* We get here if the user created a MR marked
 730                          * as use_once and invalidate at the same time. */
 731                         queue_delayed_work(rds_wq, &pool->flush_worker, 10);
 732                 }
 733         }
 734
 735         rds_ib_dev_put(rds_ibdev);
 736 }
 737
 738 void rds_ib_flush_mrs(void)
 739 {
 740         struct rds_ib_device *rds_ibdev;
 741
 742         list_for_each_entry(rds_ibdev, &rds_ib_devices, list) {
 743                 struct rds_ib_mr_pool *pool = rds_ibdev->mr_pool;
 744
 745                 if (pool)
 746                         rds_ib_flush_mr_pool(pool, 0, NULL);
 747         }
 748 }
 749
 750 void *rds_ib_get_mr(struct scatterlist *sg, unsigned long nents,
 751                     struct rds_sock *rs, u32 *key_ret)
 752 {
 753         struct rds_ib_device *rds_ibdev;
 754         struct rds_ib_mr *ibmr = NULL;
 755         int ret;
 756
 757         rds_ibdev = rds_ib_get_device(rs->rs_bound_addr);
 758         if (!rds_ibdev) {
 759                 ret = -ENODEV;
 760                 goto out;
 761         }
 762
 763         if (!rds_ibdev->mr_pool) {
 764                 ret = -ENODEV;
 765                 goto out;
 766         }
 767
 768         ibmr = rds_ib_alloc_fmr(rds_ibdev);
 769         if (IS_ERR(ibmr))
 770                 return ibmr;
 771
 772         ret = rds_ib_map_fmr(rds_ibdev, ibmr, sg, nents);
 773         if (ret == 0)
 774                 *key_ret = ibmr->fmr->rkey;
 775         else
 776                 printk(KERN_WARNING "RDS/IB: map_fmr failed (errno=%d)\n", ret);
 777
 778         ibmr->device = rds_ibdev;
 779         rds_ibdev = NULL;
 780
 781  out:
 782         if (ret) {
 783                 if (ibmr)
 784                         rds_ib_free_mr(ibmr, 0);
 785                 ibmr = ERR_PTR(ret);
 786         }
 787         if (rds_ibdev)
 788                 rds_ib_dev_put(rds_ibdev);
 789         return ibmr;
 790 }
 791