net/rds/ib_rdma.c

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