drivers/infiniband/core/fmr_pool.c

   1 /*
   2  * Copyright (c) 2004 Topspin Communications.  All rights reserved.
   3  * Copyright (c) 2005 Sun Microsystems, Inc. All rights reserved.
   4  *
   5  * This software is available to you under a choice of one of two
   6  * licenses.  You may choose to be licensed under the terms of the GNU
   7  * General Public License (GPL) Version 2, available from the file
   8  * COPYING in the main directory of this source tree, or the
   9  * OpenIB.org BSD license below:
  10  *
  11  *     Redistribution and use in source and binary forms, with or
  12  *     without modification, are permitted provided that the following
  13  *     conditions are met:
  14  *
  15  *      - Redistributions of source code must retain the above
  16  *        copyright notice, this list of conditions and the following
  17  *        disclaimer.
  18  *
  19  *      - Redistributions in binary form must reproduce the above
  20  *        copyright notice, this list of conditions and the following
  21  *        disclaimer in the documentation and/or other materials
  22  *        provided with the distribution.
  23  *
  24  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  25  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  26  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  27  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  28  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  29  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  30  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  31  * SOFTWARE.
  32  *
  33  * $Id: fmr_pool.c 2730 2005-06-28 16:43:03Z sean.hefty $
  34  */
  35
  36 #include <linux/errno.h>
  37 #include <linux/spinlock.h>
  38 #include <linux/slab.h>
  39 #include <linux/jhash.h>
  40 #include <linux/kthread.h>
  41
  42 #include <rdma/ib_fmr_pool.h>
  43
  44 #include "core_priv.h"
  45
  46 #define PFX "fmr_pool: "
  47
  48 enum {
  49         IB_FMR_MAX_REMAPS = 32,
  50
  51         IB_FMR_HASH_BITS  = 8,
  52         IB_FMR_HASH_SIZE  = 1 << IB_FMR_HASH_BITS,
  53         IB_FMR_HASH_MASK  = IB_FMR_HASH_SIZE - 1
  54 };
  55
  56 /*
  57  * If an FMR is not in use, then the list member will point to either
  58  * its pool's free_list (if the FMR can be mapped again; that is,
  59  * remap_count < pool->max_remaps) or its pool's dirty_list (if the
  60  * FMR needs to be unmapped before being remapped).  In either of
  61  * these cases it is a bug if the ref_count is not 0.  In other words,
  62  * if ref_count is > 0, then the list member must not be linked into
  63  * either free_list or dirty_list.
  64  *
  65  * The cache_node member is used to link the FMR into a cache bucket
  66  * (if caching is enabled).  This is independent of the reference
  67  * count of the FMR.  When a valid FMR is released, its ref_count is
  68  * decremented, and if ref_count reaches 0, the FMR is placed in
  69  * either free_list or dirty_list as appropriate.  However, it is not
  70  * removed from the cache and may be "revived" if a call to
  71  * ib_fmr_register_physical() occurs before the FMR is remapped.  In
  72  * this case we just increment the ref_count and remove the FMR from
  73  * free_list/dirty_list.
  74  *
  75  * Before we remap an FMR from free_list, we remove it from the cache
  76  * (to prevent another user from obtaining a stale FMR).  When an FMR
  77  * is released, we add it to the tail of the free list, so that our
  78  * cache eviction policy is "least recently used."
  79  *
  80  * All manipulation of ref_count, list and cache_node is protected by
  81  * pool_lock to maintain consistency.
  82  */
  83
  84 struct ib_fmr_pool {
  85         spinlock_t                pool_lock;
  86
  87         int                       pool_size;
  88         int                       max_pages;
  89         int                       max_remaps;
  90         int                       dirty_watermark;
  91         int                       dirty_len;
  92         struct list_head          free_list;
  93         struct list_head          dirty_list;
  94         struct hlist_head        *cache_bucket;
  95
  96         void                     (*flush_function)(struct ib_fmr_pool *pool,
  97                                                    void *              arg);
  98         void                     *flush_arg;
  99
 100         struct task_struct       *thread;
 101
 102         atomic_t                  req_ser;
 103         atomic_t                  flush_ser;
 104
 105         wait_queue_head_t         force_wait;
 106 };
 107
 108 static inline u32 ib_fmr_hash(u64 first_page)
 109 {
 110         return jhash_2words((u32) first_page, (u32) (first_page >> 32), 0) &
 111                 (IB_FMR_HASH_SIZE - 1);
 112 }
 113
 114 /* Caller must hold pool_lock */
 115 static inline struct ib_pool_fmr *ib_fmr_cache_lookup(struct ib_fmr_pool *pool,
 116                                                       u64 *page_list,
 117                                                       int  page_list_len,
 118                                                       u64  io_virtual_address)
 119 {
 120         struct hlist_head *bucket;
 121         struct ib_pool_fmr *fmr;
 122         struct hlist_node *pos;
 123
 124         if (!pool->cache_bucket)
 125                 return NULL;
 126
 127         bucket = pool->cache_bucket + ib_fmr_hash(*page_list);
 128
 129         hlist_for_each_entry(fmr, pos, bucket, cache_node)
 130                 if (io_virtual_address == fmr->io_virtual_address &&
 131                     page_list_len      == fmr->page_list_len      &&
 132                     !memcmp(page_list, fmr->page_list,
 133                             page_list_len * sizeof *page_list))
 134                         return fmr;
 135
 136         return NULL;
 137 }
 138
 139 static void ib_fmr_batch_release(struct ib_fmr_pool *pool)
 140 {
 141         int                 ret;
 142         struct ib_pool_fmr *fmr;
 143         LIST_HEAD(unmap_list);
 144         LIST_HEAD(fmr_list);
 145
 146         spin_lock_irq(&pool->pool_lock);
 147
 148         list_for_each_entry(fmr, &pool->dirty_list, list) {
 149                 hlist_del_init(&fmr->cache_node);
 150                 fmr->remap_count = 0;
 151                 list_add_tail(&fmr->fmr->list, &fmr_list);
 152
 153 #ifdef DEBUG
 154                 if (fmr->ref_count !=0) {
 155                         printk(KERN_WARNING PFX "Unmapping FMR 0x%08x with ref count %d\n",
 156                                fmr, fmr->ref_count);
 157                 }
 158 #endif
 159         }
 160
 161         list_splice_init(&pool->dirty_list, &unmap_list);
 162         pool->dirty_len = 0;
 163
 164         spin_unlock_irq(&pool->pool_lock);
 165
 166         if (list_empty(&unmap_list)) {
 167                 return;
 168         }
 169
 170         ret = ib_unmap_fmr(&fmr_list);
 171         if (ret)
 172                 printk(KERN_WARNING PFX "ib_unmap_fmr returned %d\n", ret);
 173
 174         spin_lock_irq(&pool->pool_lock);
 175         list_splice(&unmap_list, &pool->free_list);
 176         spin_unlock_irq(&pool->pool_lock);
 177 }
 178
 179 static int ib_fmr_cleanup_thread(void *pool_ptr)
 180 {
 181         struct ib_fmr_pool *pool = pool_ptr;
 182
 183         do {
 184                 if (atomic_read(&pool->flush_ser) - atomic_read(&pool->req_ser) < 0) {
 185                         ib_fmr_batch_release(pool);
 186
 187                         atomic_inc(&pool->flush_ser);
 188                         wake_up_interruptible(&pool->force_wait);
 189
 190                         if (pool->flush_function)
 191                                 pool->flush_function(pool, pool->flush_arg);
 192                 }
 193
 194                 set_current_state(TASK_INTERRUPTIBLE);
 195                 if (atomic_read(&pool->flush_ser) - atomic_read(&pool->req_ser) >= 0 &&
 196                     !kthread_should_stop())
 197                         schedule();
 198                 __set_current_state(TASK_RUNNING);
 199         } while (!kthread_should_stop());
 200
 201         return 0;
 202 }
 203
 204 /**
 205  * ib_create_fmr_pool - Create an FMR pool
 206  * @pd:Protection domain for FMRs
 207  * @params:FMR pool parameters
 208  *
 209  * Create a pool of FMRs.  Return value is pointer to new pool or
 210  * error code if creation failed.
 211  */
 212 struct ib_fmr_pool *ib_create_fmr_pool(struct ib_pd             *pd,
 213                                        struct ib_fmr_pool_param *params)
 214 {
 215         struct ib_device   *device;
 216         struct ib_fmr_pool *pool;
 217         struct ib_device_attr *attr;
 218         int i;
 219         int ret;
 220         int max_remaps;
 221
 222         if (!params)
 223                 return ERR_PTR(-EINVAL);
 224
 225         device = pd->device;
 226         if (!device->alloc_fmr    || !device->dealloc_fmr  ||
 227             !device->map_phys_fmr || !device->unmap_fmr) {
 228                 printk(KERN_INFO PFX "Device %s does not support FMRs\n",
 229                        device->name);
 230                 return ERR_PTR(-ENOSYS);
 231         }
 232
 233         attr = kmalloc(sizeof *attr, GFP_KERNEL);
 234         if (!attr) {
 235                 printk(KERN_WARNING PFX "couldn't allocate device attr struct\n");
 236                 return ERR_PTR(-ENOMEM);
 237         }
 238
 239         ret = ib_query_device(device, attr);
 240         if (ret) {
 241                 printk(KERN_WARNING PFX "couldn't query device: %d\n", ret);
 242                 kfree(attr);
 243                 return ERR_PTR(ret);
 244         }
 245
 246         if (!attr->max_map_per_fmr)
 247                 max_remaps = IB_FMR_MAX_REMAPS;
 248         else
 249                 max_remaps = attr->max_map_per_fmr;
 250
 251         kfree(attr);
 252
 253         pool = kmalloc(sizeof *pool, GFP_KERNEL);
 254         if (!pool) {
 255                 printk(KERN_WARNING PFX "couldn't allocate pool struct\n");
 256                 return ERR_PTR(-ENOMEM);
 257         }
 258
 259         pool->cache_bucket   = NULL;
 260
 261         pool->flush_function = params->flush_function;
 262         pool->flush_arg      = params->flush_arg;
 263
 264         INIT_LIST_HEAD(&pool->free_list);
 265         INIT_LIST_HEAD(&pool->dirty_list);
 266
 267         if (params->cache) {
 268                 pool->cache_bucket =
 269                         kmalloc(IB_FMR_HASH_SIZE * sizeof *pool->cache_bucket,
 270                                 GFP_KERNEL);
 271                 if (!pool->cache_bucket) {
 272                         printk(KERN_WARNING PFX "Failed to allocate cache in pool\n");
 273                         ret = -ENOMEM;
 274                         goto out_free_pool;
 275                 }
 276
 277                 for (i = 0; i < IB_FMR_HASH_SIZE; ++i)
 278                         INIT_HLIST_HEAD(pool->cache_bucket + i);
 279         }
 280
 281         pool->pool_size       = 0;
 282         pool->max_pages       = params->max_pages_per_fmr;
 283         pool->max_remaps      = max_remaps;
 284         pool->dirty_watermark = params->dirty_watermark;
 285         pool->dirty_len       = 0;
 286         spin_lock_init(&pool->pool_lock);
 287         atomic_set(&pool->req_ser,   0);
 288         atomic_set(&pool->flush_ser, 0);
 289         init_waitqueue_head(&pool->force_wait);
 290
 291         pool->thread = kthread_run(ib_fmr_cleanup_thread,
 292                                    pool,
 293                                    "ib_fmr(%s)",
 294                                    device->name);
 295         if (IS_ERR(pool->thread)) {
 296                 printk(KERN_WARNING PFX "couldn't start cleanup thread\n");
 297                 ret = PTR_ERR(pool->thread);
 298                 goto out_free_pool;
 299         }
 300
 301         {
 302                 struct ib_pool_fmr *fmr;
 303                 struct ib_fmr_attr fmr_attr = {
 304                         .max_pages  = params->max_pages_per_fmr,
 305                         .max_maps   = pool->max_remaps,
 306                         .page_shift = params->page_shift
 307                 };
 308                 int bytes_per_fmr = sizeof *fmr;
 309
 310                 if (pool->cache_bucket)
 311                         bytes_per_fmr += params->max_pages_per_fmr * sizeof (u64);
 312
 313                 for (i = 0; i < params->pool_size; ++i) {
 314                         fmr = kmalloc(bytes_per_fmr, GFP_KERNEL);
 315                         if (!fmr) {
 316                                 printk(KERN_WARNING PFX "failed to allocate fmr "
 317                                        "struct for FMR %d\n", i);
 318                                 goto out_fail;
 319                         }
 320
 321                         fmr->pool             = pool;
 322                         fmr->remap_count      = 0;
 323                         fmr->ref_count        = 0;
 324                         INIT_HLIST_NODE(&fmr->cache_node);
 325
 326                         fmr->fmr = ib_alloc_fmr(pd, params->access, &fmr_attr);
 327                         if (IS_ERR(fmr->fmr)) {
 328                                 printk(KERN_WARNING PFX "fmr_create failed "
 329                                        "for FMR %d\n", i);
 330                                 kfree(fmr);
 331                                 goto out_fail;
 332                         }
 333
 334                         list_add_tail(&fmr->list, &pool->free_list);
 335                         ++pool->pool_size;
 336                 }
 337         }
 338
 339         return pool;
 340
 341  out_free_pool:
 342         kfree(pool->cache_bucket);
 343         kfree(pool);
 344
 345         return ERR_PTR(ret);
 346
 347  out_fail:
 348         ib_destroy_fmr_pool(pool);
 349
 350         return ERR_PTR(-ENOMEM);
 351 }
 352 EXPORT_SYMBOL(ib_create_fmr_pool);
 353
 354 /**
 355  * ib_destroy_fmr_pool - Free FMR pool
 356  * @pool:FMR pool to free
 357  *
 358  * Destroy an FMR pool and free all associated resources.
 359  */
 360 void ib_destroy_fmr_pool(struct ib_fmr_pool *pool)
 361 {
 362         struct ib_pool_fmr *fmr;
 363         struct ib_pool_fmr *tmp;
 364         LIST_HEAD(fmr_list);
 365         int                 i;
 366
 367         kthread_stop(pool->thread);
 368         ib_fmr_batch_release(pool);
 369
 370         i = 0;
 371         list_for_each_entry_safe(fmr, tmp, &pool->free_list, list) {
 372                 if (fmr->remap_count) {
 373                         INIT_LIST_HEAD(&fmr_list);
 374                         list_add_tail(&fmr->fmr->list, &fmr_list);
 375                         ib_unmap_fmr(&fmr_list);
 376                 }
 377                 ib_dealloc_fmr(fmr->fmr);
 378                 list_del(&fmr->list);
 379                 kfree(fmr);
 380                 ++i;
 381         }
 382
 383         if (i < pool->pool_size)
 384                 printk(KERN_WARNING PFX "pool still has %d regions registered\n",
 385                        pool->pool_size - i);
 386
 387         kfree(pool->cache_bucket);
 388         kfree(pool);
 389 }
 390 EXPORT_SYMBOL(ib_destroy_fmr_pool);
 391
 392 /**
 393  * ib_flush_fmr_pool - Invalidate all unmapped FMRs
 394  * @pool:FMR pool to flush
 395  *
 396  * Ensure that all unmapped FMRs are fully invalidated.
 397  */
 398 int ib_flush_fmr_pool(struct ib_fmr_pool *pool)
 399 {
 400         int serial;
 401         struct ib_pool_fmr *fmr, *next;
 402
 403         /*
 404          * The free_list holds FMRs that may have been used
 405          * but have not been remapped enough times to be dirty.
 406          * Put them on the dirty list now so that the cleanup
 407          * thread will reap them too.
 408          */
 409         spin_lock_irq(&pool->pool_lock);
 410         list_for_each_entry_safe(fmr, next, &pool->free_list, list) {
 411                 if (fmr->remap_count > 0)
 412                         list_move(&fmr->list, &pool->dirty_list);
 413         }
 414         spin_unlock_irq(&pool->pool_lock);
 415
 416         serial = atomic_inc_return(&pool->req_ser);
 417         wake_up_process(pool->thread);
 418
 419         if (wait_event_interruptible(pool->force_wait,
 420                                      atomic_read(&pool->flush_ser) - serial >= 0))
 421                 return -EINTR;
 422
 423         return 0;
 424 }
 425 EXPORT_SYMBOL(ib_flush_fmr_pool);
 426
 427 /**
 428  * ib_fmr_pool_map_phys -
 429  * @pool:FMR pool to allocate FMR from
 430  * @page_list:List of pages to map
 431  * @list_len:Number of pages in @page_list
 432  * @io_virtual_address:I/O virtual address for new FMR
 433  *
 434  * Map an FMR from an FMR pool.
 435  */
 436 struct ib_pool_fmr *ib_fmr_pool_map_phys(struct ib_fmr_pool *pool_handle,
 437                                          u64                *page_list,
 438                                          int                 list_len,
 439                                          u64                 io_virtual_address)
 440 {
 441         struct ib_fmr_pool *pool = pool_handle;
 442         struct ib_pool_fmr *fmr;
 443         unsigned long       flags;
 444         int                 result;
 445
 446         if (list_len < 1 || list_len > pool->max_pages)
 447                 return ERR_PTR(-EINVAL);
 448
 449         spin_lock_irqsave(&pool->pool_lock, flags);
 450         fmr = ib_fmr_cache_lookup(pool,
 451                                   page_list,
 452                                   list_len,
 453                                   io_virtual_address);
 454         if (fmr) {
 455                 /* found in cache */
 456                 ++fmr->ref_count;
 457                 if (fmr->ref_count == 1) {
 458                         list_del(&fmr->list);
 459                 }
 460
 461                 spin_unlock_irqrestore(&pool->pool_lock, flags);
 462
 463                 return fmr;
 464         }
 465
 466         if (list_empty(&pool->free_list)) {
 467                 spin_unlock_irqrestore(&pool->pool_lock, flags);
 468                 return ERR_PTR(-EAGAIN);
 469         }
 470
 471         fmr = list_entry(pool->free_list.next, struct ib_pool_fmr, list);
 472         list_del(&fmr->list);
 473         hlist_del_init(&fmr->cache_node);
 474         spin_unlock_irqrestore(&pool->pool_lock, flags);
 475
 476         result = ib_map_phys_fmr(fmr->fmr, page_list, list_len,
 477                                  io_virtual_address);
 478
 479         if (result) {
 480                 spin_lock_irqsave(&pool->pool_lock, flags);
 481                 list_add(&fmr->list, &pool->free_list);
 482                 spin_unlock_irqrestore(&pool->pool_lock, flags);
 483
 484                 printk(KERN_WARNING PFX "fmr_map returns %d\n", result);
 485
 486                 return ERR_PTR(result);
 487         }
 488
 489         ++fmr->remap_count;
 490         fmr->ref_count = 1;
 491
 492         if (pool->cache_bucket) {
 493                 fmr->io_virtual_address = io_virtual_address;
 494                 fmr->page_list_len      = list_len;
 495                 memcpy(fmr->page_list, page_list, list_len * sizeof(*page_list));
 496
 497                 spin_lock_irqsave(&pool->pool_lock, flags);
 498                 hlist_add_head(&fmr->cache_node,
 499                                pool->cache_bucket + ib_fmr_hash(fmr->page_list[0]));
 500                 spin_unlock_irqrestore(&pool->pool_lock, flags);
 501         }
 502
 503         return fmr;
 504 }
 505 EXPORT_SYMBOL(ib_fmr_pool_map_phys);
 506
 507 /**
 508  * ib_fmr_pool_unmap - Unmap FMR
 509  * @fmr:FMR to unmap
 510  *
 511  * Unmap an FMR.  The FMR mapping may remain valid until the FMR is
 512  * reused (or until ib_flush_fmr_pool() is called).
 513  */
 514 int ib_fmr_pool_unmap(struct ib_pool_fmr *fmr)
 515 {
 516         struct ib_fmr_pool *pool;
 517         unsigned long flags;
 518
 519         pool = fmr->pool;
 520
 521         spin_lock_irqsave(&pool->pool_lock, flags);
 522
 523         --fmr->ref_count;
 524         if (!fmr->ref_count) {
 525                 if (fmr->remap_count < pool->max_remaps) {
 526                         list_add_tail(&fmr->list, &pool->free_list);
 527                 } else {
 528                         list_add_tail(&fmr->list, &pool->dirty_list);
 529                         if (++pool->dirty_len >= pool->dirty_watermark) {
 530                                 atomic_inc(&pool->req_ser);
 531                                 wake_up_process(pool->thread);
 532                         }
 533                 }
 534         }
 535
 536 #ifdef DEBUG
 537         if (fmr->ref_count < 0)
 538                 printk(KERN_WARNING PFX "FMR %p has ref count %d < 0\n",
 539                        fmr, fmr->ref_count);
 540 #endif
 541
 542         spin_unlock_irqrestore(&pool->pool_lock, flags);
 543
 544         return 0;
 545 }
 546 EXPORT_SYMBOL(ib_fmr_pool_unmap);