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 kthread_worker     *worker;
 100         struct kthread_work       work;
 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
 123         if (!pool->cache_bucket)
 124                 return NULL;
 125
 126         bucket = pool->cache_bucket + ib_fmr_hash(*page_list);
 127
 128         hlist_for_each_entry(fmr, bucket, cache_node)
 129                 if (io_virtual_address == fmr->io_virtual_address &&
 130                     page_list_len      == fmr->page_list_len      &&
 131                     !memcmp(page_list, fmr->page_list,
 132                             page_list_len * sizeof *page_list))
 133                         return fmr;
 134
 135         return NULL;
 136 }
 137
 138 static void ib_fmr_batch_release(struct ib_fmr_pool *pool)
 139 {
 140         int                 ret;
 141         struct ib_pool_fmr *fmr;
 142         LIST_HEAD(unmap_list);
 143         LIST_HEAD(fmr_list);
 144
 145         spin_lock_irq(&pool->pool_lock);
 146
 147         list_for_each_entry(fmr, &pool->dirty_list, list) {
 148                 hlist_del_init(&fmr->cache_node);
 149                 fmr->remap_count = 0;
 150                 list_add_tail(&fmr->fmr->list, &fmr_list);
 151         }
 152
 153         list_splice_init(&pool->dirty_list, &unmap_list);
 154         pool->dirty_len = 0;
 155
 156         spin_unlock_irq(&pool->pool_lock);
 157
 158         if (list_empty(&unmap_list)) {
 159                 return;
 160         }
 161
 162         ret = ib_unmap_fmr(&fmr_list);
 163         if (ret)
 164                 pr_warn(PFX "ib_unmap_fmr returned %d\n", ret);
 165
 166         spin_lock_irq(&pool->pool_lock);
 167         list_splice(&unmap_list, &pool->free_list);
 168         spin_unlock_irq(&pool->pool_lock);
 169 }
 170
 171 static void ib_fmr_cleanup_func(struct kthread_work *work)
 172 {
 173         struct ib_fmr_pool *pool = container_of(work, struct ib_fmr_pool, work);
 174
 175         ib_fmr_batch_release(pool);
 176         atomic_inc(&pool->flush_ser);
 177         wake_up_interruptible(&pool->force_wait);
 178
 179         if (pool->flush_function)
 180                 pool->flush_function(pool, pool->flush_arg);
 181
 182         if (atomic_read(&pool->flush_ser) - atomic_read(&pool->req_ser) < 0)
 183                 kthread_queue_work(pool->worker, &pool->work);
 184 }
 185
 186 /**
 187  * ib_create_fmr_pool - Create an FMR pool
 188  * @pd:Protection domain for FMRs
 189  * @params:FMR pool parameters
 190  *
 191  * Create a pool of FMRs.  Return value is pointer to new pool or
 192  * error code if creation failed.
 193  */
 194 struct ib_fmr_pool *ib_create_fmr_pool(struct ib_pd             *pd,
 195                                        struct ib_fmr_pool_param *params)
 196 {
 197         struct ib_device   *device;
 198         struct ib_fmr_pool *pool;
 199         int i;
 200         int ret;
 201         int max_remaps;
 202
 203         if (!params)
 204                 return ERR_PTR(-EINVAL);
 205
 206         device = pd->device;
 207         if (!device->ops.alloc_fmr    || !device->ops.dealloc_fmr  ||
 208             !device->ops.map_phys_fmr || !device->ops.unmap_fmr) {
 209                 dev_info(&device->dev, "Device does not support FMRs\n");
 210                 return ERR_PTR(-ENOSYS);
 211         }
 212
 213         if (!device->attrs.max_map_per_fmr)
 214                 max_remaps = IB_FMR_MAX_REMAPS;
 215         else
 216                 max_remaps = device->attrs.max_map_per_fmr;
 217
 218         pool = kmalloc(sizeof *pool, GFP_KERNEL);
 219         if (!pool)
 220                 return ERR_PTR(-ENOMEM);
 221
 222         pool->cache_bucket   = NULL;
 223         pool->flush_function = params->flush_function;
 224         pool->flush_arg      = params->flush_arg;
 225
 226         INIT_LIST_HEAD(&pool->free_list);
 227         INIT_LIST_HEAD(&pool->dirty_list);
 228
 229         if (params->cache) {
 230                 pool->cache_bucket =
 231                         kmalloc_array(IB_FMR_HASH_SIZE,
 232                                       sizeof(*pool->cache_bucket),
 233                                       GFP_KERNEL);
 234                 if (!pool->cache_bucket) {
 235                         ret = -ENOMEM;
 236                         goto out_free_pool;
 237                 }
 238
 239                 for (i = 0; i < IB_FMR_HASH_SIZE; ++i)
 240                         INIT_HLIST_HEAD(pool->cache_bucket + i);
 241         }
 242
 243         pool->pool_size       = 0;
 244         pool->max_pages       = params->max_pages_per_fmr;
 245         pool->max_remaps      = max_remaps;
 246         pool->dirty_watermark = params->dirty_watermark;
 247         pool->dirty_len       = 0;
 248         spin_lock_init(&pool->pool_lock);
 249         atomic_set(&pool->req_ser,   0);
 250         atomic_set(&pool->flush_ser, 0);
 251         init_waitqueue_head(&pool->force_wait);
 252
 253         pool->worker =
 254                 kthread_create_worker(0, "ib_fmr(%s)", dev_name(&device->dev));
 255         if (IS_ERR(pool->worker)) {
 256                 pr_warn(PFX "couldn't start cleanup kthread worker\n");
 257                 ret = PTR_ERR(pool->worker);
 258                 goto out_free_pool;
 259         }
 260         kthread_init_work(&pool->work, ib_fmr_cleanup_func);
 261
 262         {
 263                 struct ib_pool_fmr *fmr;
 264                 struct ib_fmr_attr fmr_attr = {
 265                         .max_pages  = params->max_pages_per_fmr,
 266                         .max_maps   = pool->max_remaps,
 267                         .page_shift = params->page_shift
 268                 };
 269                 int bytes_per_fmr = sizeof *fmr;
 270
 271                 if (pool->cache_bucket)
 272                         bytes_per_fmr += params->max_pages_per_fmr * sizeof (u64);
 273
 274                 for (i = 0; i < params->pool_size; ++i) {
 275                         fmr = kmalloc(bytes_per_fmr, GFP_KERNEL);
 276                         if (!fmr)
 277                                 goto out_fail;
 278
 279                         fmr->pool             = pool;
 280                         fmr->remap_count      = 0;
 281                         fmr->ref_count        = 0;
 282                         INIT_HLIST_NODE(&fmr->cache_node);
 283
 284                         fmr->fmr = ib_alloc_fmr(pd, params->access, &fmr_attr);
 285                         if (IS_ERR(fmr->fmr)) {
 286                                 pr_warn(PFX "fmr_create failed for FMR %d\n",
 287                                         i);
 288                                 kfree(fmr);
 289                                 goto out_fail;
 290                         }
 291
 292                         list_add_tail(&fmr->list, &pool->free_list);
 293                         ++pool->pool_size;
 294                 }
 295         }
 296
 297         return pool;
 298
 299  out_free_pool:
 300         kfree(pool->cache_bucket);
 301         kfree(pool);
 302
 303         return ERR_PTR(ret);
 304
 305  out_fail:
 306         ib_destroy_fmr_pool(pool);
 307
 308         return ERR_PTR(-ENOMEM);
 309 }
 310 EXPORT_SYMBOL(ib_create_fmr_pool);
 311
 312 /**
 313  * ib_destroy_fmr_pool - Free FMR pool
 314  * @pool:FMR pool to free
 315  *
 316  * Destroy an FMR pool and free all associated resources.
 317  */
 318 void ib_destroy_fmr_pool(struct ib_fmr_pool *pool)
 319 {
 320         struct ib_pool_fmr *fmr;
 321         struct ib_pool_fmr *tmp;
 322         LIST_HEAD(fmr_list);
 323         int                 i;
 324
 325         kthread_destroy_worker(pool->worker);
 326         ib_fmr_batch_release(pool);
 327
 328         i = 0;
 329         list_for_each_entry_safe(fmr, tmp, &pool->free_list, list) {
 330                 if (fmr->remap_count) {
 331                         INIT_LIST_HEAD(&fmr_list);
 332                         list_add_tail(&fmr->fmr->list, &fmr_list);
 333                         ib_unmap_fmr(&fmr_list);
 334                 }
 335                 ib_dealloc_fmr(fmr->fmr);
 336                 list_del(&fmr->list);
 337                 kfree(fmr);
 338                 ++i;
 339         }
 340
 341         if (i < pool->pool_size)
 342                 pr_warn(PFX "pool still has %d regions registered\n",
 343                         pool->pool_size - i);
 344
 345         kfree(pool->cache_bucket);
 346         kfree(pool);
 347 }
 348 EXPORT_SYMBOL(ib_destroy_fmr_pool);
 349
 350 /**
 351  * ib_flush_fmr_pool - Invalidate all unmapped FMRs
 352  * @pool:FMR pool to flush
 353  *
 354  * Ensure that all unmapped FMRs are fully invalidated.
 355  */
 356 int ib_flush_fmr_pool(struct ib_fmr_pool *pool)
 357 {
 358         int serial;
 359         struct ib_pool_fmr *fmr, *next;
 360
 361         /*
 362          * The free_list holds FMRs that may have been used
 363          * but have not been remapped enough times to be dirty.
 364          * Put them on the dirty list now so that the cleanup
 365          * thread will reap them too.
 366          */
 367         spin_lock_irq(&pool->pool_lock);
 368         list_for_each_entry_safe(fmr, next, &pool->free_list, list) {
 369                 if (fmr->remap_count > 0)
 370                         list_move(&fmr->list, &pool->dirty_list);
 371         }
 372         spin_unlock_irq(&pool->pool_lock);
 373
 374         serial = atomic_inc_return(&pool->req_ser);
 375         kthread_queue_work(pool->worker, &pool->work);
 376
 377         if (wait_event_interruptible(pool->force_wait,
 378                                      atomic_read(&pool->flush_ser) - serial >= 0))
 379                 return -EINTR;
 380
 381         return 0;
 382 }
 383 EXPORT_SYMBOL(ib_flush_fmr_pool);
 384
 385 /**
 386  * ib_fmr_pool_map_phys - Map an FMR from an FMR pool.
 387  * @pool_handle: FMR pool to allocate FMR from
 388  * @page_list: List of pages to map
 389  * @list_len: Number of pages in @page_list
 390  * @io_virtual_address: I/O virtual address for new FMR
 391  */
 392 struct ib_pool_fmr *ib_fmr_pool_map_phys(struct ib_fmr_pool *pool_handle,
 393                                          u64                *page_list,
 394                                          int                 list_len,
 395                                          u64                 io_virtual_address)
 396 {
 397         struct ib_fmr_pool *pool = pool_handle;
 398         struct ib_pool_fmr *fmr;
 399         unsigned long       flags;
 400         int                 result;
 401
 402         if (list_len < 1 || list_len > pool->max_pages)
 403                 return ERR_PTR(-EINVAL);
 404
 405         spin_lock_irqsave(&pool->pool_lock, flags);
 406         fmr = ib_fmr_cache_lookup(pool,
 407                                   page_list,
 408                                   list_len,
 409                                   io_virtual_address);
 410         if (fmr) {
 411                 /* found in cache */
 412                 ++fmr->ref_count;
 413                 if (fmr->ref_count == 1) {
 414                         list_del(&fmr->list);
 415                 }
 416
 417                 spin_unlock_irqrestore(&pool->pool_lock, flags);
 418
 419                 return fmr;
 420         }
 421
 422         if (list_empty(&pool->free_list)) {
 423                 spin_unlock_irqrestore(&pool->pool_lock, flags);
 424                 return ERR_PTR(-EAGAIN);
 425         }
 426
 427         fmr = list_entry(pool->free_list.next, struct ib_pool_fmr, list);
 428         list_del(&fmr->list);
 429         hlist_del_init(&fmr->cache_node);
 430         spin_unlock_irqrestore(&pool->pool_lock, flags);
 431
 432         result = ib_map_phys_fmr(fmr->fmr, page_list, list_len,
 433                                  io_virtual_address);
 434
 435         if (result) {
 436                 spin_lock_irqsave(&pool->pool_lock, flags);
 437                 list_add(&fmr->list, &pool->free_list);
 438                 spin_unlock_irqrestore(&pool->pool_lock, flags);
 439
 440                 pr_warn(PFX "fmr_map returns %d\n", result);
 441
 442                 return ERR_PTR(result);
 443         }
 444
 445         ++fmr->remap_count;
 446         fmr->ref_count = 1;
 447
 448         if (pool->cache_bucket) {
 449                 fmr->io_virtual_address = io_virtual_address;
 450                 fmr->page_list_len      = list_len;
 451                 memcpy(fmr->page_list, page_list, list_len * sizeof(*page_list));
 452
 453                 spin_lock_irqsave(&pool->pool_lock, flags);
 454                 hlist_add_head(&fmr->cache_node,
 455                                pool->cache_bucket + ib_fmr_hash(fmr->page_list[0]));
 456                 spin_unlock_irqrestore(&pool->pool_lock, flags);
 457         }
 458
 459         return fmr;
 460 }
 461 EXPORT_SYMBOL(ib_fmr_pool_map_phys);
 462
 463 /**
 464  * ib_fmr_pool_unmap - Unmap FMR
 465  * @fmr:FMR to unmap
 466  *
 467  * Unmap an FMR.  The FMR mapping may remain valid until the FMR is
 468  * reused (or until ib_flush_fmr_pool() is called).
 469  */
 470 void ib_fmr_pool_unmap(struct ib_pool_fmr *fmr)
 471 {
 472         struct ib_fmr_pool *pool;
 473         unsigned long flags;
 474
 475         pool = fmr->pool;
 476
 477         spin_lock_irqsave(&pool->pool_lock, flags);
 478
 479         --fmr->ref_count;
 480         if (!fmr->ref_count) {
 481                 if (fmr->remap_count < pool->max_remaps) {
 482                         list_add_tail(&fmr->list, &pool->free_list);
 483                 } else {
 484                         list_add_tail(&fmr->list, &pool->dirty_list);
 485                         if (++pool->dirty_len >= pool->dirty_watermark) {
 486                                 atomic_inc(&pool->req_ser);
 487                                 kthread_queue_work(pool->worker, &pool->work);
 488                         }
 489                 }
 490         }
 491
 492         spin_unlock_irqrestore(&pool->pool_lock, flags);
 493 }
 494 EXPORT_SYMBOL(ib_fmr_pool_unmap);