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