drivers/infiniband/core/fmr_pool.c

   1 /*
   2  * Copyright (c) 2004 Topspin Communications.  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  * $Id: fmr_pool.c 1349 2004-12-16 21:09:43Z roland $
  33  */
  34
  35 #include <linux/errno.h>
  36 #include <linux/spinlock.h>
  37 #include <linux/slab.h>
  38 #include <linux/jhash.h>
  39 #include <linux/kthread.h>
  40
  41 #include <ib_fmr_pool.h>
  42
  43 #include "core_priv.h"
  44
  45 enum {
  46         IB_FMR_MAX_REMAPS = 32,
  47
  48         IB_FMR_HASH_BITS  = 8,
  49         IB_FMR_HASH_SIZE  = 1 << IB_FMR_HASH_BITS,
  50         IB_FMR_HASH_MASK  = IB_FMR_HASH_SIZE - 1
  51 };
  52
  53 /*
  54  * If an FMR is not in use, then the list member will point to either
  55  * its pool's free_list (if the FMR can be mapped again; that is,
  56  * remap_count < IB_FMR_MAX_REMAPS) or its pool's dirty_list (if the
  57  * FMR needs to be unmapped before being remapped).  In either of
  58  * these cases it is a bug if the ref_count is not 0.  In other words,
  59  * if ref_count is > 0, then the list member must not be linked into
  60  * either free_list or dirty_list.
  61  *
  62  * The cache_node member is used to link the FMR into a cache bucket
  63  * (if caching is enabled).  This is independent of the reference
  64  * count of the FMR.  When a valid FMR is released, its ref_count is
  65  * decremented, and if ref_count reaches 0, the FMR is placed in
  66  * either free_list or dirty_list as appropriate.  However, it is not
  67  * removed from the cache and may be "revived" if a call to
  68  * ib_fmr_register_physical() occurs before the FMR is remapped.  In
  69  * this case we just increment the ref_count and remove the FMR from
  70  * free_list/dirty_list.
  71  *
  72  * Before we remap an FMR from free_list, we remove it from the cache
  73  * (to prevent another user from obtaining a stale FMR).  When an FMR
  74  * is released, we add it to the tail of the free list, so that our
  75  * cache eviction policy is "least recently used."
  76  *
  77  * All manipulation of ref_count, list and cache_node is protected by
  78  * pool_lock to maintain consistency.
  79  */
  80
  81 struct ib_fmr_pool {
  82         spinlock_t                pool_lock;
  83
  84         int                       pool_size;
  85         int                       max_pages;
  86         int                       dirty_watermark;
  87         int                       dirty_len;
  88         struct list_head          free_list;
  89         struct list_head          dirty_list;
  90         struct hlist_head        *cache_bucket;
  91
  92         void                     (*flush_function)(struct ib_fmr_pool *pool,
  93                                                    void *              arg);
  94         void                     *flush_arg;
  95
  96         struct task_struct       *thread;
  97
  98         atomic_t                  req_ser;
  99         atomic_t                  flush_ser;
 100
 101         wait_queue_head_t         force_wait;
 102 };
 103
 104 static inline u32 ib_fmr_hash(u64 first_page)
 105 {
 106         return jhash_2words((u32) first_page, (u32) (first_page >> 32), 0) &
 107                 (IB_FMR_HASH_SIZE - 1);
 108 }
 109
 110 /* Caller must hold pool_lock */
 111 static inline struct ib_pool_fmr *ib_fmr_cache_lookup(struct ib_fmr_pool *pool,
 112                                                       u64 *page_list,
 113                                                       int  page_list_len,
 114                                                       u64  io_virtual_address)
 115 {
 116         struct hlist_head *bucket;
 117         struct ib_pool_fmr *fmr;
 118         struct hlist_node *pos;
 119
 120         if (!pool->cache_bucket)
 121                 return NULL;
 122
 123         bucket = pool->cache_bucket + ib_fmr_hash(*page_list);
 124
 125         hlist_for_each_entry(fmr, pos, bucket, cache_node)
 126                 if (io_virtual_address == fmr->io_virtual_address &&
 127                     page_list_len      == fmr->page_list_len      &&
 128                     !memcmp(page_list, fmr->page_list,
 129                             page_list_len * sizeof *page_list))
 130                         return fmr;
 131
 132         return NULL;
 133 }
 134
 135 static void ib_fmr_batch_release(struct ib_fmr_pool *pool)
 136 {
 137         int                 ret;
 138         struct ib_pool_fmr *fmr;
 139         LIST_HEAD(unmap_list);
 140         LIST_HEAD(fmr_list);
 141
 142         spin_lock_irq(&pool->pool_lock);
 143
 144         list_for_each_entry(fmr, &pool->dirty_list, list) {
 145                 hlist_del_init(&fmr->cache_node);
 146                 fmr->remap_count = 0;
 147                 list_add_tail(&fmr->fmr->list, &fmr_list);
 148
 149 #ifdef DEBUG
 150                 if (fmr->ref_count !=0) {
 151                         printk(KERN_WARNING "Unmapping FMR 0x%08x with ref count %d",
 152                                fmr, fmr->ref_count);
 153                 }
 154 #endif
 155         }
 156
 157         list_splice(&pool->dirty_list, &unmap_list);
 158         INIT_LIST_HEAD(&pool->dirty_list);
 159         pool->dirty_len = 0;
 160
 161         spin_unlock_irq(&pool->pool_lock);
 162
 163         if (list_empty(&unmap_list)) {
 164                 return;
 165         }
 166
 167         ret = ib_unmap_fmr(&fmr_list);
 168         if (ret)
 169                 printk(KERN_WARNING "ib_unmap_fmr returned %d", ret);
 170
 171         spin_lock_irq(&pool->pool_lock);
 172         list_splice(&unmap_list, &pool->free_list);
 173         spin_unlock_irq(&pool->pool_lock);
 174 }
 175
 176 static int ib_fmr_cleanup_thread(void *pool_ptr)
 177 {
 178         struct ib_fmr_pool *pool = pool_ptr;
 179
 180         do {
 181                 if (pool->dirty_len >= pool->dirty_watermark ||
 182                     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 (pool->dirty_len < pool->dirty_watermark &&
 194                     atomic_read(&pool->flush_ser) - atomic_read(&pool->req_ser) >= 0 &&
 195                     !kthread_should_stop())
 196                         schedule();
 197                 __set_current_state(TASK_RUNNING);
 198         } while (!kthread_should_stop());
 199
 200         return 0;
 201 }
 202
 203 /**
 204  * ib_create_fmr_pool - Create an FMR pool
 205  * @pd:Protection domain for FMRs
 206  * @params:FMR pool parameters
 207  *
 208  * Create a pool of FMRs.  Return value is pointer to new pool or
 209  * error code if creation failed.
 210  */
 211 struct ib_fmr_pool *ib_create_fmr_pool(struct ib_pd             *pd,
 212                                        struct ib_fmr_pool_param *params)
 213 {
 214         struct ib_device   *device;
 215         struct ib_fmr_pool *pool;
 216         int i;
 217         int ret;
 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_WARNING "Device %s does not support fast memory regions",
 226                        device->name);
 227                 return ERR_PTR(-ENOSYS);
 228         }
 229
 230         pool = kmalloc(sizeof *pool, GFP_KERNEL);
 231         if (!pool) {
 232                 printk(KERN_WARNING "couldn't allocate pool struct");
 233                 return ERR_PTR(-ENOMEM);
 234         }
 235
 236         pool->cache_bucket   = NULL;
 237
 238         pool->flush_function = params->flush_function;
 239         pool->flush_arg      = params->flush_arg;
 240
 241         INIT_LIST_HEAD(&pool->free_list);
 242         INIT_LIST_HEAD(&pool->dirty_list);
 243
 244         if (params->cache) {
 245                 pool->cache_bucket =
 246                         kmalloc(IB_FMR_HASH_SIZE * sizeof *pool->cache_bucket,
 247                                 GFP_KERNEL);
 248                 if (!pool->cache_bucket) {
 249                         printk(KERN_WARNING "Failed to allocate cache in pool");
 250                         ret = -ENOMEM;
 251                         goto out_free_pool;
 252                 }
 253
 254                 for (i = 0; i < IB_FMR_HASH_SIZE; ++i)
 255                         INIT_HLIST_HEAD(pool->cache_bucket + i);
 256         }
 257
 258         pool->pool_size       = 0;
 259         pool->max_pages       = params->max_pages_per_fmr;
 260         pool->dirty_watermark = params->dirty_watermark;
 261         pool->dirty_len       = 0;
 262         spin_lock_init(&pool->pool_lock);
 263         atomic_set(&pool->req_ser,   0);
 264         atomic_set(&pool->flush_ser, 0);
 265         init_waitqueue_head(&pool->force_wait);
 266
 267         pool->thread = kthread_create(ib_fmr_cleanup_thread,
 268                                       pool,
 269                                       "ib_fmr(%s)",
 270                                       device->name);
 271         if (IS_ERR(pool->thread)) {
 272                 printk(KERN_WARNING "couldn't start cleanup thread");
 273                 ret = PTR_ERR(pool->thread);
 274                 goto out_free_pool;
 275         }
 276
 277         {
 278                 struct ib_pool_fmr *fmr;
 279                 struct ib_fmr_attr attr = {
 280                         .max_pages = params->max_pages_per_fmr,
 281                         .max_maps  = IB_FMR_MAX_REMAPS,
 282                         .page_size = PAGE_SHIFT
 283                 };
 284
 285                 for (i = 0; i < params->pool_size; ++i) {
 286                         fmr = kmalloc(sizeof *fmr + params->max_pages_per_fmr * sizeof (u64),
 287                                       GFP_KERNEL);
 288                         if (!fmr) {
 289                                 printk(KERN_WARNING "failed to allocate fmr struct "
 290                                        "for FMR %d", i);
 291                                 goto out_fail;
 292                         }
 293
 294                         fmr->pool             = pool;
 295                         fmr->remap_count      = 0;
 296                         fmr->ref_count        = 0;
 297                         INIT_HLIST_NODE(&fmr->cache_node);
 298
 299                         fmr->fmr = ib_alloc_fmr(pd, params->access, &attr);
 300                         if (IS_ERR(fmr->fmr)) {
 301                                 printk(KERN_WARNING "fmr_create failed for FMR %d", i);
 302                                 kfree(fmr);
 303                                 goto out_fail;
 304                         }
 305
 306                         list_add_tail(&fmr->list, &pool->free_list);
 307                         ++pool->pool_size;
 308                 }
 309         }
 310
 311         return pool;
 312
 313  out_free_pool:
 314         kfree(pool->cache_bucket);
 315         kfree(pool);
 316
 317         return ERR_PTR(ret);
 318
 319  out_fail:
 320         ib_destroy_fmr_pool(pool);
 321
 322         return ERR_PTR(-ENOMEM);
 323 }
 324 EXPORT_SYMBOL(ib_create_fmr_pool);
 325
 326 /**
 327  * ib_destroy_fmr_pool - Free FMR pool
 328  * @pool:FMR pool to free
 329  *
 330  * Destroy an FMR pool and free all associated resources.
 331  */
 332 int ib_destroy_fmr_pool(struct ib_fmr_pool *pool)
 333 {
 334         struct ib_pool_fmr *fmr;
 335         struct ib_pool_fmr *tmp;
 336         int                 i;
 337
 338         kthread_stop(pool->thread);
 339         ib_fmr_batch_release(pool);
 340
 341         i = 0;
 342         list_for_each_entry_safe(fmr, tmp, &pool->free_list, list) {
 343                 ib_dealloc_fmr(fmr->fmr);
 344                 list_del(&fmr->list);
 345                 kfree(fmr);
 346                 ++i;
 347         }
 348
 349         if (i < pool->pool_size)
 350                 printk(KERN_WARNING "pool still has %d regions registered",
 351                        pool->pool_size - i);
 352
 353         kfree(pool->cache_bucket);
 354         kfree(pool);
 355
 356         return 0;
 357 }
 358 EXPORT_SYMBOL(ib_destroy_fmr_pool);
 359
 360 /**
 361  * ib_flush_fmr_pool - Invalidate all unmapped FMRs
 362  * @pool:FMR pool to flush
 363  *
 364  * Ensure that all unmapped FMRs are fully invalidated.
 365  */
 366 int ib_flush_fmr_pool(struct ib_fmr_pool *pool)
 367 {
 368         int serial;
 369
 370         atomic_inc(&pool->req_ser);
 371         /*
 372          * It's OK if someone else bumps req_ser again here -- we'll
 373          * just wait a little longer.
 374          */
 375         serial = atomic_read(&pool->req_ser);
 376
 377         wake_up_process(pool->thread);
 378
 379         if (wait_event_interruptible(pool->force_wait,
 380                                      atomic_read(&pool->flush_ser) -
 381                                      atomic_read(&pool->req_ser) >= 0))
 382                 return -EINTR;
 383
 384         return 0;
 385 }
 386 EXPORT_SYMBOL(ib_flush_fmr_pool);
 387
 388 /**
 389  * ib_fmr_pool_map_phys -
 390  * @pool:FMR pool to allocate FMR from
 391  * @page_list:List of pages to map
 392  * @list_len:Number of pages in @page_list
 393  * @io_virtual_address:I/O virtual address for new FMR
 394  *
 395  * Map an FMR from an FMR pool.
 396  */
 397 struct ib_pool_fmr *ib_fmr_pool_map_phys(struct ib_fmr_pool *pool_handle,
 398                                          u64                *page_list,
 399                                          int                 list_len,
 400                                          u64                *io_virtual_address)
 401 {
 402         struct ib_fmr_pool *pool = pool_handle;
 403         struct ib_pool_fmr *fmr;
 404         unsigned long       flags;
 405         int                 result;
 406
 407         if (list_len < 1 || list_len > pool->max_pages)
 408                 return ERR_PTR(-EINVAL);
 409
 410         spin_lock_irqsave(&pool->pool_lock, flags);
 411         fmr = ib_fmr_cache_lookup(pool,
 412                                   page_list,
 413                                   list_len,
 414                                   *io_virtual_address);
 415         if (fmr) {
 416                 /* found in cache */
 417                 ++fmr->ref_count;
 418                 if (fmr->ref_count == 1) {
 419                         list_del(&fmr->list);
 420                 }
 421
 422                 spin_unlock_irqrestore(&pool->pool_lock, flags);
 423
 424                 return fmr;
 425         }
 426
 427         if (list_empty(&pool->free_list)) {
 428                 spin_unlock_irqrestore(&pool->pool_lock, flags);
 429                 return ERR_PTR(-EAGAIN);
 430         }
 431
 432         fmr = list_entry(pool->free_list.next, struct ib_pool_fmr, list);
 433         list_del(&fmr->list);
 434         hlist_del_init(&fmr->cache_node);
 435         spin_unlock_irqrestore(&pool->pool_lock, flags);
 436
 437         result = ib_map_phys_fmr(fmr->fmr, page_list, list_len,
 438                                  *io_virtual_address);
 439
 440         if (result) {
 441                 spin_lock_irqsave(&pool->pool_lock, flags);
 442                 list_add(&fmr->list, &pool->free_list);
 443                 spin_unlock_irqrestore(&pool->pool_lock, flags);
 444
 445                 printk(KERN_WARNING "fmr_map returns %d\n",
 446                        result);
 447
 448                 return ERR_PTR(result);
 449         }
 450
 451         ++fmr->remap_count;
 452         fmr->ref_count = 1;
 453
 454         if (pool->cache_bucket) {
 455                 fmr->io_virtual_address = *io_virtual_address;
 456                 fmr->page_list_len      = list_len;
 457                 memcpy(fmr->page_list, page_list, list_len * sizeof(*page_list));
 458
 459                 spin_lock_irqsave(&pool->pool_lock, flags);
 460                 hlist_add_head(&fmr->cache_node,
 461                                pool->cache_bucket + ib_fmr_hash(fmr->page_list[0]));
 462                 spin_unlock_irqrestore(&pool->pool_lock, flags);
 463         }
 464
 465         return fmr;
 466 }
 467 EXPORT_SYMBOL(ib_fmr_pool_map_phys);
 468
 469 /**
 470  * ib_fmr_pool_unmap - Unmap FMR
 471  * @fmr:FMR to unmap
 472  *
 473  * Unmap an FMR.  The FMR mapping may remain valid until the FMR is
 474  * reused (or until ib_flush_fmr_pool() is called).
 475  */
 476 int ib_fmr_pool_unmap(struct ib_pool_fmr *fmr)
 477 {
 478         struct ib_fmr_pool *pool;
 479         unsigned long flags;
 480
 481         pool = fmr->pool;
 482
 483         spin_lock_irqsave(&pool->pool_lock, flags);
 484
 485         --fmr->ref_count;
 486         if (!fmr->ref_count) {
 487                 if (fmr->remap_count < IB_FMR_MAX_REMAPS) {
 488                         list_add_tail(&fmr->list, &pool->free_list);
 489                 } else {
 490                         list_add_tail(&fmr->list, &pool->dirty_list);
 491                         ++pool->dirty_len;
 492                         wake_up_process(pool->thread);
 493                 }
 494         }
 495
 496 #ifdef DEBUG
 497         if (fmr->ref_count < 0)
 498                 printk(KERN_WARNING "FMR %p has ref count %d < 0",
 499                        fmr, fmr->ref_count);
 500 #endif
 501
 502         spin_unlock_irqrestore(&pool->pool_lock, flags);
 503
 504         return 0;
 505 }
 506 EXPORT_SYMBOL(ib_fmr_pool_unmap);