kernel/net/rds/ib_rdma.c

   1 /*
   2  * Copyright (c) 2006 Oracle.  All rights reserved.
   3  *
   4  * This software is available to you under a choice of one of two
   5  * licenses.  You may choose to be licensed under the terms of the GNU
   6  * General Public License (GPL) Version 2, available from the file
   7  * COPYING in the main directory of this source tree, or the
   8  * OpenIB.org BSD license below:
   9  *
  10  *     Redistribution and use in source and binary forms, with or
  11  *     without modification, are permitted provided that the following
  12  *     conditions are met:
  13  *
  14  *      - Redistributions of source code must retain the above
  15  *        copyright notice, this list of conditions and the following
  16  *        disclaimer.
  17  *
  18  *      - Redistributions in binary form must reproduce the above
  19  *        copyright notice, this list of conditions and the following
  20  *        disclaimer in the documentation and/or other materials
  21  *        provided with the distribution.
  22  *
  23  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  24  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  25  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  26  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  27  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  28  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  29  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  30  * SOFTWARE.
  31  *
  32  */
  33 #include <linux/kernel.h>
  34 #include <linux/slab.h>
  35 #include <linux/rculist.h>
  36 #include <linux/llist.h>
  37 #include <linux/delay.h>
  38
  39 #include "rds.h"
  40 #include "ib.h"
  41
  42 static DEFINE_PER_CPU(unsigned long, clean_list_grace);
  43 #define CLEAN_LIST_BUSY_BIT 0
  44
  45 /*
  46  * This is stored as mr->r_trans_private.
  47  */
  48 struct rds_ib_mr {
  49         struct rds_ib_device    *device;
  50         struct rds_ib_mr_pool   *pool;
  51         struct ib_fmr           *fmr;
  52
  53         struct llist_node       llnode;
  54
  55         /* unmap_list is for freeing */
  56         struct list_head        unmap_list;
  57         unsigned int            remap_count;
  58
  59         struct scatterlist      *sg;
  60         unsigned int            sg_len;
  61         u64                     *dma;
  62         int                     sg_dma_len;
  63 };
  64
  65 /*
  66  * Our own little FMR pool
  67  */
  68 struct rds_ib_mr_pool {
  69         unsigned int            pool_type;
  70         struct mutex            flush_lock;             /* serialize fmr invalidate */
  71         struct delayed_work     flush_worker;           /* flush worker */
  72
  73         atomic_t                item_count;             /* total # of MRs */
  74         atomic_t                dirty_count;            /* # dirty of MRs */
  75
  76         struct llist_head       drop_list;              /* MRs that have reached their max_maps limit */
  77         struct llist_head       free_list;              /* unused MRs */
  78         struct llist_head       clean_list;             /* global unused & unamapped MRs */
  79         wait_queue_head_t       flush_wait;
  80
  81         atomic_t                free_pinned;            /* memory pinned by free MRs */
  82         unsigned long           max_items;
  83         unsigned long           max_items_soft;
  84         unsigned long           max_free_pinned;
  85         struct ib_fmr_attr      fmr_attr;
  86 };
  87
  88 static struct workqueue_struct *rds_ib_fmr_wq;
  89
  90 int rds_ib_fmr_init(void)
  91 {
  92         rds_ib_fmr_wq = create_workqueue("rds_fmr_flushd");
  93         if (!rds_ib_fmr_wq)
  94                 return -ENOMEM;
  95         return 0;
  96 }
  97
  98 /* By the time this is called all the IB devices should have been torn down and
  99  * had their pools freed.  As each pool is freed its work struct is waited on,
 100  * so the pool flushing work queue should be idle by the time we get here.
 101  */
 102 void rds_ib_fmr_exit(void)
 103 {
 104         destroy_workqueue(rds_ib_fmr_wq);
 105 }
 106
 107 static int rds_ib_flush_mr_pool(struct rds_ib_mr_pool *pool, int free_all, struct rds_ib_mr **);
 108 static void rds_ib_teardown_mr(struct rds_ib_mr *ibmr);
 109 static void rds_ib_mr_pool_flush_worker(struct work_struct *work);
 110
 111 static struct rds_ib_device *rds_ib_get_device(__be32 ipaddr)
 112 {
 113         struct rds_ib_device *rds_ibdev;
 114         struct rds_ib_ipaddr *i_ipaddr;
 115
 116         rcu_read_lock();
 117         list_for_each_entry_rcu(rds_ibdev, &rds_ib_devices, list) {
 118                 list_for_each_entry_rcu(i_ipaddr, &rds_ibdev->ipaddr_list, list) {
 119                         if (i_ipaddr->ipaddr == ipaddr) {
 120                                 atomic_inc(&rds_ibdev->refcount);
 121                                 rcu_read_unlock();
 122                                 return rds_ibdev;
 123                         }
 124                 }
 125         }
 126         rcu_read_unlock();
 127
 128         return NULL;
 129 }
 130
 131 static int rds_ib_add_ipaddr(struct rds_ib_device *rds_ibdev, __be32 ipaddr)
 132 {
 133         struct rds_ib_ipaddr *i_ipaddr;
 134
 135         i_ipaddr = kmalloc(sizeof *i_ipaddr, GFP_KERNEL);
 136         if (!i_ipaddr)
 137                 return -ENOMEM;
 138
 139         i_ipaddr->ipaddr = ipaddr;
 140
 141         spin_lock_irq(&rds_ibdev->spinlock);
 142         list_add_tail_rcu(&i_ipaddr->list, &rds_ibdev->ipaddr_list);
 143         spin_unlock_irq(&rds_ibdev->spinlock);
 144
 145         return 0;
 146 }
 147
 148 static void rds_ib_remove_ipaddr(struct rds_ib_device *rds_ibdev, __be32 ipaddr)
 149 {
 150         struct rds_ib_ipaddr *i_ipaddr;
 151         struct rds_ib_ipaddr *to_free = NULL;
 152
 153
 154         spin_lock_irq(&rds_ibdev->spinlock);
 155         list_for_each_entry_rcu(i_ipaddr, &rds_ibdev->ipaddr_list, list) {
 156                 if (i_ipaddr->ipaddr == ipaddr) {
 157                         list_del_rcu(&i_ipaddr->list);
 158                         to_free = i_ipaddr;
 159                         break;
 160                 }
 161         }
 162         spin_unlock_irq(&rds_ibdev->spinlock);
 163
 164         if (to_free)
 165                 kfree_rcu(to_free, rcu);
 166 }
 167
 168 int rds_ib_update_ipaddr(struct rds_ib_device *rds_ibdev, __be32 ipaddr)
 169 {
 170         struct rds_ib_device *rds_ibdev_old;
 171
 172         rds_ibdev_old = rds_ib_get_device(ipaddr);
 173         if (!rds_ibdev_old)
 174                 return rds_ib_add_ipaddr(rds_ibdev, ipaddr);
 175
 176         if (rds_ibdev_old != rds_ibdev) {
 177                 rds_ib_remove_ipaddr(rds_ibdev_old, ipaddr);
 178                 rds_ib_dev_put(rds_ibdev_old);
 179                 return rds_ib_add_ipaddr(rds_ibdev, ipaddr);
 180         }
 181         rds_ib_dev_put(rds_ibdev_old);
 182
 183         return 0;
 184 }
 185
 186 void rds_ib_add_conn(struct rds_ib_device *rds_ibdev, struct rds_connection *conn)
 187 {
 188         struct rds_ib_connection *ic = conn->c_transport_data;
 189
 190         /* conn was previously on the nodev_conns_list */
 191         spin_lock_irq(&ib_nodev_conns_lock);
 192         BUG_ON(list_empty(&ib_nodev_conns));
 193         BUG_ON(list_empty(&ic->ib_node));
 194         list_del(&ic->ib_node);
 195
 196         spin_lock(&rds_ibdev->spinlock);
 197         list_add_tail(&ic->ib_node, &rds_ibdev->conn_list);
 198         spin_unlock(&rds_ibdev->spinlock);
 199         spin_unlock_irq(&ib_nodev_conns_lock);
 200
 201         ic->rds_ibdev = rds_ibdev;
 202         atomic_inc(&rds_ibdev->refcount);
 203 }
 204
 205 void rds_ib_remove_conn(struct rds_ib_device *rds_ibdev, struct rds_connection *conn)
 206 {
 207         struct rds_ib_connection *ic = conn->c_transport_data;
 208
 209         /* place conn on nodev_conns_list */
 210         spin_lock(&ib_nodev_conns_lock);
 211
 212         spin_lock_irq(&rds_ibdev->spinlock);
 213         BUG_ON(list_empty(&ic->ib_node));
 214         list_del(&ic->ib_node);
 215         spin_unlock_irq(&rds_ibdev->spinlock);
 216
 217         list_add_tail(&ic->ib_node, &ib_nodev_conns);
 218
 219         spin_unlock(&ib_nodev_conns_lock);
 220
 221         ic->rds_ibdev = NULL;
 222         rds_ib_dev_put(rds_ibdev);
 223 }
 224
 225 void rds_ib_destroy_nodev_conns(void)
 226 {
 227         struct rds_ib_connection *ic, *_ic;
 228         LIST_HEAD(tmp_list);
 229
 230         /* avoid calling conn_destroy with irqs off */
 231         spin_lock_irq(&ib_nodev_conns_lock);
 232         list_splice(&ib_nodev_conns, &tmp_list);
 233         spin_unlock_irq(&ib_nodev_conns_lock);
 234
 235         list_for_each_entry_safe(ic, _ic, &tmp_list, ib_node)
 236                 rds_conn_destroy(ic->conn);
 237 }
 238
 239 struct rds_ib_mr_pool *rds_ib_create_mr_pool(struct rds_ib_device *rds_ibdev,
 240                                              int pool_type)
 241 {
 242         struct rds_ib_mr_pool *pool;
 243
 244         pool = kzalloc(sizeof(*pool), GFP_KERNEL);
 245         if (!pool)
 246                 return ERR_PTR(-ENOMEM);
 247
 248         pool->pool_type = pool_type;
 249         init_llist_head(&pool->free_list);
 250         init_llist_head(&pool->drop_list);
 251         init_llist_head(&pool->clean_list);
 252         mutex_init(&pool->flush_lock);
 253         init_waitqueue_head(&pool->flush_wait);
 254         INIT_DELAYED_WORK(&pool->flush_worker, rds_ib_mr_pool_flush_worker);
 255
 256         if (pool_type == RDS_IB_MR_1M_POOL) {
 257                 /* +1 allows for unaligned MRs */
 258                 pool->fmr_attr.max_pages = RDS_FMR_1M_MSG_SIZE + 1;
 259                 pool->max_items = RDS_FMR_1M_POOL_SIZE;
 260         } else {
 261                 /* pool_type == RDS_IB_MR_8K_POOL */
 262                 pool->fmr_attr.max_pages = RDS_FMR_8K_MSG_SIZE + 1;
 263                 pool->max_items = RDS_FMR_8K_POOL_SIZE;
 264         }
 265
 266         pool->max_free_pinned = pool->max_items * pool->fmr_attr.max_pages / 4;
 267         pool->fmr_attr.max_maps = rds_ibdev->fmr_max_remaps;
 268         pool->fmr_attr.page_shift = PAGE_SHIFT;
 269         pool->max_items_soft = rds_ibdev->max_fmrs * 3 / 4;
 270
 271         return pool;
 272 }
 273
 274 void rds_ib_get_mr_info(struct rds_ib_device *rds_ibdev, struct rds_info_rdma_connection *iinfo)
 275 {
 276         struct rds_ib_mr_pool *pool_1m = rds_ibdev->mr_1m_pool;
 277
 278         iinfo->rdma_mr_max = pool_1m->max_items;
 279         iinfo->rdma_mr_size = pool_1m->fmr_attr.max_pages;
 280 }
 281
 282 void rds_ib_destroy_mr_pool(struct rds_ib_mr_pool *pool)
 283 {
 284         cancel_delayed_work_sync(&pool->flush_worker);
 285         rds_ib_flush_mr_pool(pool, 1, NULL);
 286         WARN_ON(atomic_read(&pool->item_count));
 287         WARN_ON(atomic_read(&pool->free_pinned));
 288         kfree(pool);
 289 }
 290
 291 static inline struct rds_ib_mr *rds_ib_reuse_fmr(struct rds_ib_mr_pool *pool)
 292 {
 293         struct rds_ib_mr *ibmr = NULL;
 294         struct llist_node *ret;
 295         unsigned long *flag;
 296
 297         preempt_disable();
 298         flag = this_cpu_ptr(&clean_list_grace);
 299         set_bit(CLEAN_LIST_BUSY_BIT, flag);
 300         ret = llist_del_first(&pool->clean_list);
 301         if (ret)
 302                 ibmr = llist_entry(ret, struct rds_ib_mr, llnode);
 303
 304         clear_bit(CLEAN_LIST_BUSY_BIT, flag);
 305         preempt_enable();
 306         return ibmr;
 307 }
 308
 309 static inline void wait_clean_list_grace(void)
 310 {
 311         int cpu;
 312         unsigned long *flag;
 313
 314         for_each_online_cpu(cpu) {
 315                 flag = &per_cpu(clean_list_grace, cpu);
 316                 while (test_bit(CLEAN_LIST_BUSY_BIT, flag))
 317                         cpu_chill();
 318         }
 319 }
 320
 321 static struct rds_ib_mr *rds_ib_alloc_fmr(struct rds_ib_device *rds_ibdev,
 322                                           int npages)
 323 {
 324         struct rds_ib_mr_pool *pool;
 325         struct rds_ib_mr *ibmr = NULL;
 326         int err = 0, iter = 0;
 327
 328         if (npages <= RDS_FMR_8K_MSG_SIZE)
 329                 pool = rds_ibdev->mr_8k_pool;
 330         else
 331                 pool = rds_ibdev->mr_1m_pool;
 332
 333         if (atomic_read(&pool->dirty_count) >= pool->max_items / 10)
 334                 queue_delayed_work(rds_ib_fmr_wq, &pool->flush_worker, 10);
 335
 336         /* Switch pools if one of the pool is reaching upper limit */
 337         if (atomic_read(&pool->dirty_count) >=  pool->max_items * 9 / 10) {
 338                 if (pool->pool_type == RDS_IB_MR_8K_POOL)
 339                         pool = rds_ibdev->mr_1m_pool;
 340                 else
 341                         pool = rds_ibdev->mr_8k_pool;
 342         }
 343
 344         while (1) {
 345                 ibmr = rds_ib_reuse_fmr(pool);
 346                 if (ibmr)
 347                         return ibmr;
 348
 349                 /* No clean MRs - now we have the choice of either
 350                  * allocating a fresh MR up to the limit imposed by the
 351                  * driver, or flush any dirty unused MRs.
 352                  * We try to avoid stalling in the send path if possible,
 353                  * so we allocate as long as we're allowed to.
 354                  *
 355                  * We're fussy with enforcing the FMR limit, though. If the driver
 356                  * tells us we can't use more than N fmrs, we shouldn't start
 357                  * arguing with it */
 358                 if (atomic_inc_return(&pool->item_count) <= pool->max_items)
 359                         break;
 360
 361                 atomic_dec(&pool->item_count);
 362
 363                 if (++iter > 2) {
 364                         if (pool->pool_type == RDS_IB_MR_8K_POOL)
 365                                 rds_ib_stats_inc(s_ib_rdma_mr_8k_pool_depleted);
 366                         else
 367                                 rds_ib_stats_inc(s_ib_rdma_mr_1m_pool_depleted);
 368                         return ERR_PTR(-EAGAIN);
 369                 }
 370
 371                 /* We do have some empty MRs. Flush them out. */
 372                 if (pool->pool_type == RDS_IB_MR_8K_POOL)
 373                         rds_ib_stats_inc(s_ib_rdma_mr_8k_pool_wait);
 374                 else
 375                         rds_ib_stats_inc(s_ib_rdma_mr_1m_pool_wait);
 376                 rds_ib_flush_mr_pool(pool, 0, &ibmr);
 377                 if (ibmr)
 378                         return ibmr;
 379         }
 380
 381         ibmr = kzalloc_node(sizeof(*ibmr), GFP_KERNEL, rdsibdev_to_node(rds_ibdev));
 382         if (!ibmr) {
 383                 err = -ENOMEM;
 384                 goto out_no_cigar;
 385         }
 386
 387         ibmr->fmr = ib_alloc_fmr(rds_ibdev->pd,
 388                         (IB_ACCESS_LOCAL_WRITE |
 389                          IB_ACCESS_REMOTE_READ |
 390                          IB_ACCESS_REMOTE_WRITE|
 391                          IB_ACCESS_REMOTE_ATOMIC),
 392                         &pool->fmr_attr);
 393         if (IS_ERR(ibmr->fmr)) {
 394                 err = PTR_ERR(ibmr->fmr);
 395                 ibmr->fmr = NULL;
 396                 printk(KERN_WARNING "RDS/IB: ib_alloc_fmr failed (err=%d)\n", err);
 397                 goto out_no_cigar;
 398         }
 399
 400         ibmr->pool = pool;
 401         if (pool->pool_type == RDS_IB_MR_8K_POOL)
 402                 rds_ib_stats_inc(s_ib_rdma_mr_8k_alloc);
 403         else
 404                 rds_ib_stats_inc(s_ib_rdma_mr_1m_alloc);
 405
 406         return ibmr;
 407
 408 out_no_cigar:
 409         if (ibmr) {
 410                 if (ibmr->fmr)
 411                         ib_dealloc_fmr(ibmr->fmr);
 412                 kfree(ibmr);
 413         }
 414         atomic_dec(&pool->item_count);
 415         return ERR_PTR(err);
 416 }
 417
 418 static int rds_ib_map_fmr(struct rds_ib_device *rds_ibdev, struct rds_ib_mr *ibmr,
 419                struct scatterlist *sg, unsigned int nents)
 420 {
 421         struct ib_device *dev = rds_ibdev->dev;
 422         struct scatterlist *scat = sg;
 423         u64 io_addr = 0;
 424         u64 *dma_pages;
 425         u32 len;
 426         int page_cnt, sg_dma_len;
 427         int i, j;
 428         int ret;
 429
 430         sg_dma_len = ib_dma_map_sg(dev, sg, nents,
 431                                  DMA_BIDIRECTIONAL);
 432         if (unlikely(!sg_dma_len)) {
 433                 printk(KERN_WARNING "RDS/IB: dma_map_sg failed!\n");
 434                 return -EBUSY;
 435         }
 436
 437         len = 0;
 438         page_cnt = 0;
 439
 440         for (i = 0; i < sg_dma_len; ++i) {
 441                 unsigned int dma_len = ib_sg_dma_len(dev, &scat[i]);
 442                 u64 dma_addr = ib_sg_dma_address(dev, &scat[i]);
 443
 444                 if (dma_addr & ~PAGE_MASK) {
 445                         if (i > 0)
 446                                 return -EINVAL;
 447                         else
 448                                 ++page_cnt;
 449                 }
 450                 if ((dma_addr + dma_len) & ~PAGE_MASK) {
 451                         if (i < sg_dma_len - 1)
 452                                 return -EINVAL;
 453                         else
 454                                 ++page_cnt;
 455                 }
 456
 457                 len += dma_len;
 458         }
 459
 460         page_cnt += len >> PAGE_SHIFT;
 461         if (page_cnt > ibmr->pool->fmr_attr.max_pages)
 462                 return -EINVAL;
 463
 464         dma_pages = kmalloc_node(sizeof(u64) * page_cnt, GFP_ATOMIC,
 465                                  rdsibdev_to_node(rds_ibdev));
 466         if (!dma_pages)
 467                 return -ENOMEM;
 468
 469         page_cnt = 0;
 470         for (i = 0; i < sg_dma_len; ++i) {
 471                 unsigned int dma_len = ib_sg_dma_len(dev, &scat[i]);
 472                 u64 dma_addr = ib_sg_dma_address(dev, &scat[i]);
 473
 474                 for (j = 0; j < dma_len; j += PAGE_SIZE)
 475                         dma_pages[page_cnt++] =
 476                                 (dma_addr & PAGE_MASK) + j;
 477         }
 478
 479         ret = ib_map_phys_fmr(ibmr->fmr,
 480                                    dma_pages, page_cnt, io_addr);
 481         if (ret)
 482                 goto out;
 483
 484         /* Success - we successfully remapped the MR, so we can
 485          * safely tear down the old mapping. */
 486         rds_ib_teardown_mr(ibmr);
 487
 488         ibmr->sg = scat;
 489         ibmr->sg_len = nents;
 490         ibmr->sg_dma_len = sg_dma_len;
 491         ibmr->remap_count++;
 492
 493         if (ibmr->pool->pool_type == RDS_IB_MR_8K_POOL)
 494                 rds_ib_stats_inc(s_ib_rdma_mr_8k_used);
 495         else
 496                 rds_ib_stats_inc(s_ib_rdma_mr_1m_used);
 497         ret = 0;
 498
 499 out:
 500         kfree(dma_pages);
 501
 502         return ret;
 503 }
 504
 505 void rds_ib_sync_mr(void *trans_private, int direction)
 506 {
 507         struct rds_ib_mr *ibmr = trans_private;
 508         struct rds_ib_device *rds_ibdev = ibmr->device;
 509
 510         switch (direction) {
 511         case DMA_FROM_DEVICE:
 512                 ib_dma_sync_sg_for_cpu(rds_ibdev->dev, ibmr->sg,
 513                         ibmr->sg_dma_len, DMA_BIDIRECTIONAL);
 514                 break;
 515         case DMA_TO_DEVICE:
 516                 ib_dma_sync_sg_for_device(rds_ibdev->dev, ibmr->sg,
 517                         ibmr->sg_dma_len, DMA_BIDIRECTIONAL);
 518                 break;
 519         }
 520 }
 521
 522 static void __rds_ib_teardown_mr(struct rds_ib_mr *ibmr)
 523 {
 524         struct rds_ib_device *rds_ibdev = ibmr->device;
 525
 526         if (ibmr->sg_dma_len) {
 527                 ib_dma_unmap_sg(rds_ibdev->dev,
 528                                 ibmr->sg, ibmr->sg_len,
 529                                 DMA_BIDIRECTIONAL);
 530                 ibmr->sg_dma_len = 0;
 531         }
 532
 533         /* Release the s/g list */
 534         if (ibmr->sg_len) {
 535                 unsigned int i;
 536
 537                 for (i = 0; i < ibmr->sg_len; ++i) {
 538                         struct page *page = sg_page(&ibmr->sg[i]);
 539
 540                         /* FIXME we need a way to tell a r/w MR
 541                          * from a r/o MR */
 542                         WARN_ON(!page->mapping && irqs_disabled());
 543                         set_page_dirty(page);
 544                         put_page(page);
 545                 }
 546                 kfree(ibmr->sg);
 547
 548                 ibmr->sg = NULL;
 549                 ibmr->sg_len = 0;
 550         }
 551 }
 552
 553 static void rds_ib_teardown_mr(struct rds_ib_mr *ibmr)
 554 {
 555         unsigned int pinned = ibmr->sg_len;
 556
 557         __rds_ib_teardown_mr(ibmr);
 558         if (pinned) {
 559                 struct rds_ib_mr_pool *pool = ibmr->pool;
 560
 561                 atomic_sub(pinned, &pool->free_pinned);
 562         }
 563 }
 564
 565 static inline unsigned int rds_ib_flush_goal(struct rds_ib_mr_pool *pool, int free_all)
 566 {
 567         unsigned int item_count;
 568
 569         item_count = atomic_read(&pool->item_count);
 570         if (free_all)
 571                 return item_count;
 572
 573         return 0;
 574 }
 575
 576 /*
 577  * given an llist of mrs, put them all into the list_head for more processing
 578  */
 579 static unsigned int llist_append_to_list(struct llist_head *llist,
 580                                          struct list_head *list)
 581 {
 582         struct rds_ib_mr *ibmr;
 583         struct llist_node *node;
 584         struct llist_node *next;
 585         unsigned int count = 0;
 586
 587         node = llist_del_all(llist);
 588         while (node) {
 589                 next = node->next;
 590                 ibmr = llist_entry(node, struct rds_ib_mr, llnode);
 591                 list_add_tail(&ibmr->unmap_list, list);
 592                 node = next;
 593                 count++;
 594         }
 595         return count;
 596 }
 597
 598 /*
 599  * this takes a list head of mrs and turns it into linked llist nodes
 600  * of clusters.  Each cluster has linked llist nodes of
 601  * MR_CLUSTER_SIZE mrs that are ready for reuse.
 602  */
 603 static void list_to_llist_nodes(struct rds_ib_mr_pool *pool,
 604                                 struct list_head *list,
 605                                 struct llist_node **nodes_head,
 606                                 struct llist_node **nodes_tail)
 607 {
 608         struct rds_ib_mr *ibmr;
 609         struct llist_node *cur = NULL;
 610         struct llist_node **next = nodes_head;
 611
 612         list_for_each_entry(ibmr, list, unmap_list) {
 613                 cur = &ibmr->llnode;
 614                 *next = cur;
 615                 next = &cur->next;
 616         }
 617         *next = NULL;
 618         *nodes_tail = cur;
 619 }
 620
 621 /*
 622  * Flush our pool of MRs.
 623  * At a minimum, all currently unused MRs are unmapped.
 624  * If the number of MRs allocated exceeds the limit, we also try
 625  * to free as many MRs as needed to get back to this limit.
 626  */
 627 static int rds_ib_flush_mr_pool(struct rds_ib_mr_pool *pool,
 628                                 int free_all, struct rds_ib_mr **ibmr_ret)
 629 {
 630         struct rds_ib_mr *ibmr, *next;
 631         struct llist_node *clean_nodes;
 632         struct llist_node *clean_tail;
 633         LIST_HEAD(unmap_list);
 634         LIST_HEAD(fmr_list);
 635         unsigned long unpinned = 0;
 636         unsigned int nfreed = 0, dirty_to_clean = 0, free_goal;
 637         int ret = 0;
 638
 639         if (pool->pool_type == RDS_IB_MR_8K_POOL)
 640                 rds_ib_stats_inc(s_ib_rdma_mr_8k_pool_flush);
 641         else
 642                 rds_ib_stats_inc(s_ib_rdma_mr_1m_pool_flush);
 643
 644         if (ibmr_ret) {
 645                 DEFINE_WAIT(wait);
 646                 while (!mutex_trylock(&pool->flush_lock)) {
 647                         ibmr = rds_ib_reuse_fmr(pool);
 648                         if (ibmr) {
 649                                 *ibmr_ret = ibmr;
 650                                 finish_wait(&pool->flush_wait, &wait);
 651                                 goto out_nolock;
 652                         }
 653
 654                         prepare_to_wait(&pool->flush_wait, &wait,
 655                                         TASK_UNINTERRUPTIBLE);
 656                         if (llist_empty(&pool->clean_list))
 657                                 schedule();
 658
 659                         ibmr = rds_ib_reuse_fmr(pool);
 660                         if (ibmr) {
 661                                 *ibmr_ret = ibmr;
 662                                 finish_wait(&pool->flush_wait, &wait);
 663                                 goto out_nolock;
 664                         }
 665                 }
 666                 finish_wait(&pool->flush_wait, &wait);
 667         } else
 668                 mutex_lock(&pool->flush_lock);
 669
 670         if (ibmr_ret) {
 671                 ibmr = rds_ib_reuse_fmr(pool);
 672                 if (ibmr) {
 673                         *ibmr_ret = ibmr;
 674                         goto out;
 675                 }
 676         }
 677
 678         /* Get the list of all MRs to be dropped. Ordering matters -
 679          * we want to put drop_list ahead of free_list.
 680          */
 681         dirty_to_clean = llist_append_to_list(&pool->drop_list, &unmap_list);
 682         dirty_to_clean += llist_append_to_list(&pool->free_list, &unmap_list);
 683         if (free_all)
 684                 llist_append_to_list(&pool->clean_list, &unmap_list);
 685
 686         free_goal = rds_ib_flush_goal(pool, free_all);
 687
 688         if (list_empty(&unmap_list))
 689                 goto out;
 690
 691         /* String all ib_mr's onto one list and hand them to ib_unmap_fmr */
 692         list_for_each_entry(ibmr, &unmap_list, unmap_list)
 693                 list_add(&ibmr->fmr->list, &fmr_list);
 694
 695         ret = ib_unmap_fmr(&fmr_list);
 696         if (ret)
 697                 printk(KERN_WARNING "RDS/IB: ib_unmap_fmr failed (err=%d)\n", ret);
 698
 699         /* Now we can destroy the DMA mapping and unpin any pages */
 700         list_for_each_entry_safe(ibmr, next, &unmap_list, unmap_list) {
 701                 unpinned += ibmr->sg_len;
 702                 __rds_ib_teardown_mr(ibmr);
 703                 if (nfreed < free_goal ||
 704                     ibmr->remap_count >= pool->fmr_attr.max_maps) {
 705                         if (ibmr->pool->pool_type == RDS_IB_MR_8K_POOL)
 706                                 rds_ib_stats_inc(s_ib_rdma_mr_8k_free);
 707                         else
 708                                 rds_ib_stats_inc(s_ib_rdma_mr_1m_free);
 709                         list_del(&ibmr->unmap_list);
 710                         ib_dealloc_fmr(ibmr->fmr);
 711                         kfree(ibmr);
 712                         nfreed++;
 713                 }
 714         }
 715
 716         if (!list_empty(&unmap_list)) {
 717                 /* we have to make sure that none of the things we're about
 718                  * to put on the clean list would race with other cpus trying
 719                  * to pull items off.  The llist would explode if we managed to
 720                  * remove something from the clean list and then add it back again
 721                  * while another CPU was spinning on that same item in llist_del_first.
 722                  *
 723                  * This is pretty unlikely, but just in case  wait for an llist grace period
 724                  * here before adding anything back into the clean list.
 725                  */
 726                 wait_clean_list_grace();
 727
 728                 list_to_llist_nodes(pool, &unmap_list, &clean_nodes, &clean_tail);
 729                 if (ibmr_ret)
 730                         *ibmr_ret = llist_entry(clean_nodes, struct rds_ib_mr, llnode);
 731
 732                 /* more than one entry in llist nodes */
 733                 if (clean_nodes->next)
 734                         llist_add_batch(clean_nodes->next, clean_tail, &pool->clean_list);
 735
 736         }
 737
 738         atomic_sub(unpinned, &pool->free_pinned);
 739         atomic_sub(dirty_to_clean, &pool->dirty_count);
 740         atomic_sub(nfreed, &pool->item_count);
 741
 742 out:
 743         mutex_unlock(&pool->flush_lock);
 744         if (waitqueue_active(&pool->flush_wait))
 745                 wake_up(&pool->flush_wait);
 746 out_nolock:
 747         return ret;
 748 }
 749
 750 static void rds_ib_mr_pool_flush_worker(struct work_struct *work)
 751 {
 752         struct rds_ib_mr_pool *pool = container_of(work, struct rds_ib_mr_pool, flush_worker.work);
 753
 754         rds_ib_flush_mr_pool(pool, 0, NULL);
 755 }
 756
 757 void rds_ib_free_mr(void *trans_private, int invalidate)
 758 {
 759         struct rds_ib_mr *ibmr = trans_private;
 760         struct rds_ib_mr_pool *pool = ibmr->pool;
 761         struct rds_ib_device *rds_ibdev = ibmr->device;
 762
 763         rdsdebug("RDS/IB: free_mr nents %u\n", ibmr->sg_len);
 764
 765         /* Return it to the pool's free list */
 766         if (ibmr->remap_count >= pool->fmr_attr.max_maps)
 767                 llist_add(&ibmr->llnode, &pool->drop_list);
 768         else
 769                 llist_add(&ibmr->llnode, &pool->free_list);
 770
 771         atomic_add(ibmr->sg_len, &pool->free_pinned);
 772         atomic_inc(&pool->dirty_count);
 773
 774         /* If we've pinned too many pages, request a flush */
 775         if (atomic_read(&pool->free_pinned) >= pool->max_free_pinned ||
 776             atomic_read(&pool->dirty_count) >= pool->max_items / 5)
 777                 queue_delayed_work(rds_ib_fmr_wq, &pool->flush_worker, 10);
 778
 779         if (invalidate) {
 780                 if (likely(!in_interrupt())) {
 781                         rds_ib_flush_mr_pool(pool, 0, NULL);
 782                 } else {
 783                         /* We get here if the user created a MR marked
 784                          * as use_once and invalidate at the same time.
 785                          */
 786                         queue_delayed_work(rds_ib_fmr_wq,
 787                                            &pool->flush_worker, 10);
 788                 }
 789         }
 790
 791         rds_ib_dev_put(rds_ibdev);
 792 }
 793
 794 void rds_ib_flush_mrs(void)
 795 {
 796         struct rds_ib_device *rds_ibdev;
 797
 798         down_read(&rds_ib_devices_lock);
 799         list_for_each_entry(rds_ibdev, &rds_ib_devices, list) {
 800                 if (rds_ibdev->mr_8k_pool)
 801                         rds_ib_flush_mr_pool(rds_ibdev->mr_8k_pool, 0, NULL);
 802
 803                 if (rds_ibdev->mr_1m_pool)
 804                         rds_ib_flush_mr_pool(rds_ibdev->mr_1m_pool, 0, NULL);
 805         }
 806         up_read(&rds_ib_devices_lock);
 807 }
 808
 809 void *rds_ib_get_mr(struct scatterlist *sg, unsigned long nents,
 810                     struct rds_sock *rs, u32 *key_ret)
 811 {
 812         struct rds_ib_device *rds_ibdev;
 813         struct rds_ib_mr *ibmr = NULL;
 814         int ret;
 815
 816         rds_ibdev = rds_ib_get_device(rs->rs_bound_addr);
 817         if (!rds_ibdev) {
 818                 ret = -ENODEV;
 819                 goto out;
 820         }
 821
 822         if (!rds_ibdev->mr_8k_pool || !rds_ibdev->mr_1m_pool) {
 823                 ret = -ENODEV;
 824                 goto out;
 825         }
 826
 827         ibmr = rds_ib_alloc_fmr(rds_ibdev, nents);
 828         if (IS_ERR(ibmr)) {
 829                 rds_ib_dev_put(rds_ibdev);
 830                 return ibmr;
 831         }
 832
 833         ret = rds_ib_map_fmr(rds_ibdev, ibmr, sg, nents);
 834         if (ret == 0)
 835                 *key_ret = ibmr->fmr->rkey;
 836         else
 837                 printk(KERN_WARNING "RDS/IB: map_fmr failed (errno=%d)\n", ret);
 838
 839         ibmr->device = rds_ibdev;
 840         rds_ibdev = NULL;
 841
 842  out:
 843         if (ret) {
 844                 if (ibmr)
 845                         rds_ib_free_mr(ibmr, 0);
 846                 ibmr = ERR_PTR(ret);
 847         }
 848         if (rds_ibdev)
 849                 rds_ib_dev_put(rds_ibdev);
 850         return ibmr;
 851 }
 852