drivers/gpu/drm/ttm/ttm_page_alloc_dma.c

   1 /*
   2  * Copyright 2011 (c) Oracle Corp.
   3
   4  * Permission is hereby granted, free of charge, to any person obtaining a
   5  * copy of this software and associated documentation files (the "Software"),
   6  * to deal in the Software without restriction, including without limitation
   7  * the rights to use, copy, modify, merge, publish, distribute, sub license,
   8  * and/or sell copies of the Software, and to permit persons to whom the
   9  * Software is furnished to do so, subject to the following conditions:
  10  *
  11  * The above copyright notice and this permission notice (including the
  12  * next paragraph) shall be included in all copies or substantial portions
  13  * of the Software.
  14  *
  15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  16  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  17  * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
  18  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  19  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
  20  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
  21  * DEALINGS IN THE SOFTWARE.
  22  *
  23  * Author: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
  24  */
  25
  26 /*
  27  * A simple DMA pool losely based on dmapool.c. It has certain advantages
  28  * over the DMA pools:
  29  * - Pool collects resently freed pages for reuse (and hooks up to
  30  *   the shrinker).
  31  * - Tracks currently in use pages
  32  * - Tracks whether the page is UC, WB or cached (and reverts to WB
  33  *   when freed).
  34  */
  35
  36 #define pr_fmt(fmt) "[TTM] " fmt
  37
  38 #include <linux/dma-mapping.h>
  39 #include <linux/list.h>
  40 #include <linux/seq_file.h> /* for seq_printf */
  41 #include <linux/slab.h>
  42 #include <linux/spinlock.h>
  43 #include <linux/highmem.h>
  44 #include <linux/mm_types.h>
  45 #include <linux/module.h>
  46 #include <linux/mm.h>
  47 #include <linux/atomic.h>
  48 #include <linux/device.h>
  49 #include <linux/kthread.h>
  50 #include <drm/ttm/ttm_bo_driver.h>
  51 #include <drm/ttm/ttm_page_alloc.h>
  52 #include <drm/ttm/ttm_set_memory.h>
  53
  54 #define NUM_PAGES_TO_ALLOC              (PAGE_SIZE/sizeof(struct page *))
  55 #define SMALL_ALLOCATION                4
  56 #define FREE_ALL_PAGES                  (~0U)
  57 #define VADDR_FLAG_HUGE_POOL            1UL
  58 #define VADDR_FLAG_UPDATED_COUNT        2UL
  59
  60 enum pool_type {
  61         IS_UNDEFINED    = 0,
  62         IS_WC           = 1 << 1,
  63         IS_UC           = 1 << 2,
  64         IS_CACHED       = 1 << 3,
  65         IS_DMA32        = 1 << 4,
  66         IS_HUGE         = 1 << 5
  67 };
  68
  69 /*
  70  * The pool structure. There are up to nine pools:
  71  *  - generic (not restricted to DMA32):
  72  *      - write combined, uncached, cached.
  73  *  - dma32 (up to 2^32 - so up 4GB):
  74  *      - write combined, uncached, cached.
  75  *  - huge (not restricted to DMA32):
  76  *      - write combined, uncached, cached.
  77  * for each 'struct device'. The 'cached' is for pages that are actively used.
  78  * The other ones can be shrunk by the shrinker API if neccessary.
  79  * @pools: The 'struct device->dma_pools' link.
  80  * @type: Type of the pool
  81  * @lock: Protects the free_list from concurrnet access. Must be
  82  * used with irqsave/irqrestore variants because pool allocator maybe called
  83  * from delayed work.
  84  * @free_list: Pool of pages that are free to be used. No order requirements.
  85  * @dev: The device that is associated with these pools.
  86  * @size: Size used during DMA allocation.
  87  * @npages_free: Count of available pages for re-use.
  88  * @npages_in_use: Count of pages that are in use.
  89  * @nfrees: Stats when pool is shrinking.
  90  * @nrefills: Stats when the pool is grown.
  91  * @gfp_flags: Flags to pass for alloc_page.
  92  * @name: Name of the pool.
  93  * @dev_name: Name derieved from dev - similar to how dev_info works.
  94  *   Used during shutdown as the dev_info during release is unavailable.
  95  */
  96 struct dma_pool {
  97         struct list_head pools; /* The 'struct device->dma_pools link */
  98         enum pool_type type;
  99         spinlock_t lock;
 100         struct list_head free_list;
 101         struct device *dev;
 102         unsigned size;
 103         unsigned npages_free;
 104         unsigned npages_in_use;
 105         unsigned long nfrees; /* Stats when shrunk. */
 106         unsigned long nrefills; /* Stats when grown. */
 107         gfp_t gfp_flags;
 108         char name[13]; /* "cached dma32" */
 109         char dev_name[64]; /* Constructed from dev */
 110 };
 111
 112 /*
 113  * The accounting page keeping track of the allocated page along with
 114  * the DMA address.
 115  * @page_list: The link to the 'page_list' in 'struct dma_pool'.
 116  * @vaddr: The virtual address of the page and a flag if the page belongs to a
 117  * huge pool
 118  * @dma: The bus address of the page. If the page is not allocated
 119  *   via the DMA API, it will be -1.
 120  */
 121 struct dma_page {
 122         struct list_head page_list;
 123         unsigned long vaddr;
 124         struct page *p;
 125         dma_addr_t dma;
 126 };
 127
 128 /*
 129  * Limits for the pool. They are handled without locks because only place where
 130  * they may change is in sysfs store. They won't have immediate effect anyway
 131  * so forcing serialization to access them is pointless.
 132  */
 133
 134 struct ttm_pool_opts {
 135         unsigned        alloc_size;
 136         unsigned        max_size;
 137         unsigned        small;
 138 };
 139
 140 /*
 141  * Contains the list of all of the 'struct device' and their corresponding
 142  * DMA pools. Guarded by _mutex->lock.
 143  * @pools: The link to 'struct ttm_pool_manager->pools'
 144  * @dev: The 'struct device' associated with the 'pool'
 145  * @pool: The 'struct dma_pool' associated with the 'dev'
 146  */
 147 struct device_pools {
 148         struct list_head pools;
 149         struct device *dev;
 150         struct dma_pool *pool;
 151 };
 152
 153 /*
 154  * struct ttm_pool_manager - Holds memory pools for fast allocation
 155  *
 156  * @lock: Lock used when adding/removing from pools
 157  * @pools: List of 'struct device' and 'struct dma_pool' tuples.
 158  * @options: Limits for the pool.
 159  * @npools: Total amount of pools in existence.
 160  * @shrinker: The structure used by [un|]register_shrinker
 161  */
 162 struct ttm_pool_manager {
 163         struct mutex            lock;
 164         struct list_head        pools;
 165         struct ttm_pool_opts    options;
 166         unsigned                npools;
 167         struct shrinker         mm_shrink;
 168         struct kobject          kobj;
 169 };
 170
 171 static struct ttm_pool_manager *_manager;
 172
 173 static struct attribute ttm_page_pool_max = {
 174         .name = "pool_max_size",
 175         .mode = S_IRUGO | S_IWUSR
 176 };
 177 static struct attribute ttm_page_pool_small = {
 178         .name = "pool_small_allocation",
 179         .mode = S_IRUGO | S_IWUSR
 180 };
 181 static struct attribute ttm_page_pool_alloc_size = {
 182         .name = "pool_allocation_size",
 183         .mode = S_IRUGO | S_IWUSR
 184 };
 185
 186 static struct attribute *ttm_pool_attrs[] = {
 187         &ttm_page_pool_max,
 188         &ttm_page_pool_small,
 189         &ttm_page_pool_alloc_size,
 190         NULL
 191 };
 192
 193 static void ttm_pool_kobj_release(struct kobject *kobj)
 194 {
 195         struct ttm_pool_manager *m =
 196                 container_of(kobj, struct ttm_pool_manager, kobj);
 197         kfree(m);
 198 }
 199
 200 static ssize_t ttm_pool_store(struct kobject *kobj, struct attribute *attr,
 201                               const char *buffer, size_t size)
 202 {
 203         struct ttm_pool_manager *m =
 204                 container_of(kobj, struct ttm_pool_manager, kobj);
 205         int chars;
 206         unsigned val;
 207
 208         chars = sscanf(buffer, "%u", &val);
 209         if (chars == 0)
 210                 return size;
 211
 212         /* Convert kb to number of pages */
 213         val = val / (PAGE_SIZE >> 10);
 214
 215         if (attr == &ttm_page_pool_max) {
 216                 m->options.max_size = val;
 217         } else if (attr == &ttm_page_pool_small) {
 218                 m->options.small = val;
 219         } else if (attr == &ttm_page_pool_alloc_size) {
 220                 if (val > NUM_PAGES_TO_ALLOC*8) {
 221                         pr_err("Setting allocation size to %lu is not allowed. Recommended size is %lu\n",
 222                                NUM_PAGES_TO_ALLOC*(PAGE_SIZE >> 7),
 223                                NUM_PAGES_TO_ALLOC*(PAGE_SIZE >> 10));
 224                         return size;
 225                 } else if (val > NUM_PAGES_TO_ALLOC) {
 226                         pr_warn("Setting allocation size to larger than %lu is not recommended\n",
 227                                 NUM_PAGES_TO_ALLOC*(PAGE_SIZE >> 10));
 228                 }
 229                 m->options.alloc_size = val;
 230         }
 231
 232         return size;
 233 }
 234
 235 static ssize_t ttm_pool_show(struct kobject *kobj, struct attribute *attr,
 236                              char *buffer)
 237 {
 238         struct ttm_pool_manager *m =
 239                 container_of(kobj, struct ttm_pool_manager, kobj);
 240         unsigned val = 0;
 241
 242         if (attr == &ttm_page_pool_max)
 243                 val = m->options.max_size;
 244         else if (attr == &ttm_page_pool_small)
 245                 val = m->options.small;
 246         else if (attr == &ttm_page_pool_alloc_size)
 247                 val = m->options.alloc_size;
 248
 249         val = val * (PAGE_SIZE >> 10);
 250
 251         return snprintf(buffer, PAGE_SIZE, "%u\n", val);
 252 }
 253
 254 static const struct sysfs_ops ttm_pool_sysfs_ops = {
 255         .show = &ttm_pool_show,
 256         .store = &ttm_pool_store,
 257 };
 258
 259 static struct kobj_type ttm_pool_kobj_type = {
 260         .release = &ttm_pool_kobj_release,
 261         .sysfs_ops = &ttm_pool_sysfs_ops,
 262         .default_attrs = ttm_pool_attrs,
 263 };
 264
 265 static int ttm_set_pages_caching(struct dma_pool *pool,
 266                                  struct page **pages, unsigned cpages)
 267 {
 268         int r = 0;
 269         /* Set page caching */
 270         if (pool->type & IS_UC) {
 271                 r = ttm_set_pages_array_uc(pages, cpages);
 272                 if (r)
 273                         pr_err("%s: Failed to set %d pages to uc!\n",
 274                                pool->dev_name, cpages);
 275         }
 276         if (pool->type & IS_WC) {
 277                 r = ttm_set_pages_array_wc(pages, cpages);
 278                 if (r)
 279                         pr_err("%s: Failed to set %d pages to wc!\n",
 280                                pool->dev_name, cpages);
 281         }
 282         return r;
 283 }
 284
 285 static void __ttm_dma_free_page(struct dma_pool *pool, struct dma_page *d_page)
 286 {
 287         unsigned long attrs = 0;
 288         dma_addr_t dma = d_page->dma;
 289         d_page->vaddr &= ~VADDR_FLAG_HUGE_POOL;
 290         if (pool->type & IS_HUGE)
 291                 attrs = DMA_ATTR_NO_WARN;
 292
 293         dma_free_attrs(pool->dev, pool->size, (void *)d_page->vaddr, dma, attrs);
 294
 295         kfree(d_page);
 296         d_page = NULL;
 297 }
 298 static struct dma_page *__ttm_dma_alloc_page(struct dma_pool *pool)
 299 {
 300         struct dma_page *d_page;
 301         unsigned long attrs = 0;
 302         void *vaddr;
 303
 304         d_page = kmalloc(sizeof(struct dma_page), GFP_KERNEL);
 305         if (!d_page)
 306                 return NULL;
 307
 308         if (pool->type & IS_HUGE)
 309                 attrs = DMA_ATTR_NO_WARN;
 310
 311         vaddr = dma_alloc_attrs(pool->dev, pool->size, &d_page->dma,
 312                                 pool->gfp_flags, attrs);
 313         if (vaddr) {
 314                 if (is_vmalloc_addr(vaddr))
 315                         d_page->p = vmalloc_to_page(vaddr);
 316                 else
 317                         d_page->p = virt_to_page(vaddr);
 318                 d_page->vaddr = (unsigned long)vaddr;
 319                 if (pool->type & IS_HUGE)
 320                         d_page->vaddr |= VADDR_FLAG_HUGE_POOL;
 321         } else {
 322                 kfree(d_page);
 323                 d_page = NULL;
 324         }
 325         return d_page;
 326 }
 327 static enum pool_type ttm_to_type(int flags, enum ttm_caching_state cstate)
 328 {
 329         enum pool_type type = IS_UNDEFINED;
 330
 331         if (flags & TTM_PAGE_FLAG_DMA32)
 332                 type |= IS_DMA32;
 333         if (cstate == tt_cached)
 334                 type |= IS_CACHED;
 335         else if (cstate == tt_uncached)
 336                 type |= IS_UC;
 337         else
 338                 type |= IS_WC;
 339
 340         return type;
 341 }
 342
 343 static void ttm_pool_update_free_locked(struct dma_pool *pool,
 344                                         unsigned freed_pages)
 345 {
 346         pool->npages_free -= freed_pages;
 347         pool->nfrees += freed_pages;
 348
 349 }
 350
 351 /* set memory back to wb and free the pages. */
 352 static void ttm_dma_page_put(struct dma_pool *pool, struct dma_page *d_page)
 353 {
 354         struct page *page = d_page->p;
 355         unsigned num_pages;
 356
 357         /* Don't set WB on WB page pool. */
 358         if (!(pool->type & IS_CACHED)) {
 359                 num_pages = pool->size / PAGE_SIZE;
 360                 if (ttm_set_pages_wb(page, num_pages))
 361                         pr_err("%s: Failed to set %d pages to wb!\n",
 362                                pool->dev_name, num_pages);
 363         }
 364
 365         list_del(&d_page->page_list);
 366         __ttm_dma_free_page(pool, d_page);
 367 }
 368
 369 static void ttm_dma_pages_put(struct dma_pool *pool, struct list_head *d_pages,
 370                               struct page *pages[], unsigned npages)
 371 {
 372         struct dma_page *d_page, *tmp;
 373
 374         if (pool->type & IS_HUGE) {
 375                 list_for_each_entry_safe(d_page, tmp, d_pages, page_list)
 376                         ttm_dma_page_put(pool, d_page);
 377
 378                 return;
 379         }
 380
 381         /* Don't set WB on WB page pool. */
 382         if (npages && !(pool->type & IS_CACHED) &&
 383             ttm_set_pages_array_wb(pages, npages))
 384                 pr_err("%s: Failed to set %d pages to wb!\n",
 385                        pool->dev_name, npages);
 386
 387         list_for_each_entry_safe(d_page, tmp, d_pages, page_list) {
 388                 list_del(&d_page->page_list);
 389                 __ttm_dma_free_page(pool, d_page);
 390         }
 391 }
 392
 393 /*
 394  * Free pages from pool.
 395  *
 396  * To prevent hogging the ttm_swap process we only free NUM_PAGES_TO_ALLOC
 397  * number of pages in one go.
 398  *
 399  * @pool: to free the pages from
 400  * @nr_free: If set to true will free all pages in pool
 401  * @use_static: Safe to use static buffer
 402  **/
 403 static unsigned ttm_dma_page_pool_free(struct dma_pool *pool, unsigned nr_free,
 404                                        bool use_static)
 405 {
 406         static struct page *static_buf[NUM_PAGES_TO_ALLOC];
 407         unsigned long irq_flags;
 408         struct dma_page *dma_p, *tmp;
 409         struct page **pages_to_free;
 410         struct list_head d_pages;
 411         unsigned freed_pages = 0,
 412                  npages_to_free = nr_free;
 413
 414         if (NUM_PAGES_TO_ALLOC < nr_free)
 415                 npages_to_free = NUM_PAGES_TO_ALLOC;
 416
 417         if (use_static)
 418                 pages_to_free = static_buf;
 419         else
 420                 pages_to_free = kmalloc_array(npages_to_free,
 421                                               sizeof(struct page *),
 422                                               GFP_KERNEL);
 423
 424         if (!pages_to_free) {
 425                 pr_debug("%s: Failed to allocate memory for pool free operation\n",
 426                        pool->dev_name);
 427                 return 0;
 428         }
 429         INIT_LIST_HEAD(&d_pages);
 430 restart:
 431         spin_lock_irqsave(&pool->lock, irq_flags);
 432
 433         /* We picking the oldest ones off the list */
 434         list_for_each_entry_safe_reverse(dma_p, tmp, &pool->free_list,
 435                                          page_list) {
 436                 if (freed_pages >= npages_to_free)
 437                         break;
 438
 439                 /* Move the dma_page from one list to another. */
 440                 list_move(&dma_p->page_list, &d_pages);
 441
 442                 pages_to_free[freed_pages++] = dma_p->p;
 443                 /* We can only remove NUM_PAGES_TO_ALLOC at a time. */
 444                 if (freed_pages >= NUM_PAGES_TO_ALLOC) {
 445
 446                         ttm_pool_update_free_locked(pool, freed_pages);
 447                         /**
 448                          * Because changing page caching is costly
 449                          * we unlock the pool to prevent stalling.
 450                          */
 451                         spin_unlock_irqrestore(&pool->lock, irq_flags);
 452
 453                         ttm_dma_pages_put(pool, &d_pages, pages_to_free,
 454                                           freed_pages);
 455
 456                         INIT_LIST_HEAD(&d_pages);
 457
 458                         if (likely(nr_free != FREE_ALL_PAGES))
 459                                 nr_free -= freed_pages;
 460
 461                         if (NUM_PAGES_TO_ALLOC >= nr_free)
 462                                 npages_to_free = nr_free;
 463                         else
 464                                 npages_to_free = NUM_PAGES_TO_ALLOC;
 465
 466                         freed_pages = 0;
 467
 468                         /* free all so restart the processing */
 469                         if (nr_free)
 470                                 goto restart;
 471
 472                         /* Not allowed to fall through or break because
 473                          * following context is inside spinlock while we are
 474                          * outside here.
 475                          */
 476                         goto out;
 477
 478                 }
 479         }
 480
 481         /* remove range of pages from the pool */
 482         if (freed_pages) {
 483                 ttm_pool_update_free_locked(pool, freed_pages);
 484                 nr_free -= freed_pages;
 485         }
 486
 487         spin_unlock_irqrestore(&pool->lock, irq_flags);
 488
 489         if (freed_pages)
 490                 ttm_dma_pages_put(pool, &d_pages, pages_to_free, freed_pages);
 491 out:
 492         if (pages_to_free != static_buf)
 493                 kfree(pages_to_free);
 494         return nr_free;
 495 }
 496
 497 static void ttm_dma_free_pool(struct device *dev, enum pool_type type)
 498 {
 499         struct device_pools *p;
 500         struct dma_pool *pool;
 501
 502         if (!dev)
 503                 return;
 504
 505         mutex_lock(&_manager->lock);
 506         list_for_each_entry_reverse(p, &_manager->pools, pools) {
 507                 if (p->dev != dev)
 508                         continue;
 509                 pool = p->pool;
 510                 if (pool->type != type)
 511                         continue;
 512
 513                 list_del(&p->pools);
 514                 kfree(p);
 515                 _manager->npools--;
 516                 break;
 517         }
 518         list_for_each_entry_reverse(pool, &dev->dma_pools, pools) {
 519                 if (pool->type != type)
 520                         continue;
 521                 /* Takes a spinlock.. */
 522                 /* OK to use static buffer since global mutex is held. */
 523                 ttm_dma_page_pool_free(pool, FREE_ALL_PAGES, true);
 524                 WARN_ON(((pool->npages_in_use + pool->npages_free) != 0));
 525                 /* This code path is called after _all_ references to the
 526                  * struct device has been dropped - so nobody should be
 527                  * touching it. In case somebody is trying to _add_ we are
 528                  * guarded by the mutex. */
 529                 list_del(&pool->pools);
 530                 kfree(pool);
 531                 break;
 532         }
 533         mutex_unlock(&_manager->lock);
 534 }
 535
 536 /*
 537  * On free-ing of the 'struct device' this deconstructor is run.
 538  * Albeit the pool might have already been freed earlier.
 539  */
 540 static void ttm_dma_pool_release(struct device *dev, void *res)
 541 {
 542         struct dma_pool *pool = *(struct dma_pool **)res;
 543
 544         if (pool)
 545                 ttm_dma_free_pool(dev, pool->type);
 546 }
 547
 548 static int ttm_dma_pool_match(struct device *dev, void *res, void *match_data)
 549 {
 550         return *(struct dma_pool **)res == match_data;
 551 }
 552
 553 static struct dma_pool *ttm_dma_pool_init(struct device *dev, gfp_t flags,
 554                                           enum pool_type type)
 555 {
 556         const char *n[] = {"wc", "uc", "cached", " dma32", "huge"};
 557         enum pool_type t[] = {IS_WC, IS_UC, IS_CACHED, IS_DMA32, IS_HUGE};
 558         struct device_pools *sec_pool = NULL;
 559         struct dma_pool *pool = NULL, **ptr;
 560         unsigned i;
 561         int ret = -ENODEV;
 562         char *p;
 563
 564         if (!dev)
 565                 return NULL;
 566
 567         ptr = devres_alloc(ttm_dma_pool_release, sizeof(*ptr), GFP_KERNEL);
 568         if (!ptr)
 569                 return NULL;
 570
 571         ret = -ENOMEM;
 572
 573         pool = kmalloc_node(sizeof(struct dma_pool), GFP_KERNEL,
 574                             dev_to_node(dev));
 575         if (!pool)
 576                 goto err_mem;
 577
 578         sec_pool = kmalloc_node(sizeof(struct device_pools), GFP_KERNEL,
 579                                 dev_to_node(dev));
 580         if (!sec_pool)
 581                 goto err_mem;
 582
 583         INIT_LIST_HEAD(&sec_pool->pools);
 584         sec_pool->dev = dev;
 585         sec_pool->pool =  pool;
 586
 587         INIT_LIST_HEAD(&pool->free_list);
 588         INIT_LIST_HEAD(&pool->pools);
 589         spin_lock_init(&pool->lock);
 590         pool->dev = dev;
 591         pool->npages_free = pool->npages_in_use = 0;
 592         pool->nfrees = 0;
 593         pool->gfp_flags = flags;
 594         if (type & IS_HUGE)
 595 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 596                 pool->size = HPAGE_PMD_SIZE;
 597 #else
 598                 BUG();
 599 #endif
 600         else
 601                 pool->size = PAGE_SIZE;
 602         pool->type = type;
 603         pool->nrefills = 0;
 604         p = pool->name;
 605         for (i = 0; i < ARRAY_SIZE(t); i++) {
 606                 if (type & t[i]) {
 607                         p += scnprintf(p, sizeof(pool->name) - (p - pool->name),
 608                                       "%s", n[i]);
 609                 }
 610         }
 611         *p = 0;
 612         /* We copy the name for pr_ calls b/c when dma_pool_destroy is called
 613          * - the kobj->name has already been deallocated.*/
 614         snprintf(pool->dev_name, sizeof(pool->dev_name), "%s %s",
 615                  dev_driver_string(dev), dev_name(dev));
 616         mutex_lock(&_manager->lock);
 617         /* You can get the dma_pool from either the global: */
 618         list_add(&sec_pool->pools, &_manager->pools);
 619         _manager->npools++;
 620         /* or from 'struct device': */
 621         list_add(&pool->pools, &dev->dma_pools);
 622         mutex_unlock(&_manager->lock);
 623
 624         *ptr = pool;
 625         devres_add(dev, ptr);
 626
 627         return pool;
 628 err_mem:
 629         devres_free(ptr);
 630         kfree(sec_pool);
 631         kfree(pool);
 632         return ERR_PTR(ret);
 633 }
 634
 635 static struct dma_pool *ttm_dma_find_pool(struct device *dev,
 636                                           enum pool_type type)
 637 {
 638         struct dma_pool *pool, *tmp;
 639
 640         if (type == IS_UNDEFINED)
 641                 return NULL;
 642
 643         /* NB: We iterate on the 'struct dev' which has no spinlock, but
 644          * it does have a kref which we have taken. The kref is taken during
 645          * graphic driver loading - in the drm_pci_init it calls either
 646          * pci_dev_get or pci_register_driver which both end up taking a kref
 647          * on 'struct device'.
 648          *
 649          * On teardown, the graphic drivers end up quiescing the TTM (put_pages)
 650          * and calls the dev_res deconstructors: ttm_dma_pool_release. The nice
 651          * thing is at that point of time there are no pages associated with the
 652          * driver so this function will not be called.
 653          */
 654         list_for_each_entry_safe(pool, tmp, &dev->dma_pools, pools)
 655                 if (pool->type == type)
 656                         return pool;
 657         return NULL;
 658 }
 659
 660 /*
 661  * Free pages the pages that failed to change the caching state. If there
 662  * are pages that have changed their caching state already put them to the
 663  * pool.
 664  */
 665 static void ttm_dma_handle_caching_state_failure(struct dma_pool *pool,
 666                                                  struct list_head *d_pages,
 667                                                  struct page **failed_pages,
 668                                                  unsigned cpages)
 669 {
 670         struct dma_page *d_page, *tmp;
 671         struct page *p;
 672         unsigned i = 0;
 673
 674         p = failed_pages[0];
 675         if (!p)
 676                 return;
 677         /* Find the failed page. */
 678         list_for_each_entry_safe(d_page, tmp, d_pages, page_list) {
 679                 if (d_page->p != p)
 680                         continue;
 681                 /* .. and then progress over the full list. */
 682                 list_del(&d_page->page_list);
 683                 __ttm_dma_free_page(pool, d_page);
 684                 if (++i < cpages)
 685                         p = failed_pages[i];
 686                 else
 687                         break;
 688         }
 689
 690 }
 691
 692 /*
 693  * Allocate 'count' pages, and put 'need' number of them on the
 694  * 'pages' and as well on the 'dma_address' starting at 'dma_offset' offset.
 695  * The full list of pages should also be on 'd_pages'.
 696  * We return zero for success, and negative numbers as errors.
 697  */
 698 static int ttm_dma_pool_alloc_new_pages(struct dma_pool *pool,
 699                                         struct list_head *d_pages,
 700                                         unsigned count)
 701 {
 702         struct page **caching_array;
 703         struct dma_page *dma_p;
 704         struct page *p;
 705         int r = 0;
 706         unsigned i, j, npages, cpages;
 707         unsigned max_cpages = min(count,
 708                         (unsigned)(PAGE_SIZE/sizeof(struct page *)));
 709
 710         /* allocate array for page caching change */
 711         caching_array = kmalloc_array(max_cpages, sizeof(struct page *),
 712                                       GFP_KERNEL);
 713
 714         if (!caching_array) {
 715                 pr_debug("%s: Unable to allocate table for new pages\n",
 716                        pool->dev_name);
 717                 return -ENOMEM;
 718         }
 719
 720         if (count > 1)
 721                 pr_debug("%s: (%s:%d) Getting %d pages\n",
 722                          pool->dev_name, pool->name, current->pid, count);
 723
 724         for (i = 0, cpages = 0; i < count; ++i) {
 725                 dma_p = __ttm_dma_alloc_page(pool);
 726                 if (!dma_p) {
 727                         pr_debug("%s: Unable to get page %u\n",
 728                                  pool->dev_name, i);
 729
 730                         /* store already allocated pages in the pool after
 731                          * setting the caching state */
 732                         if (cpages) {
 733                                 r = ttm_set_pages_caching(pool, caching_array,
 734                                                           cpages);
 735                                 if (r)
 736                                         ttm_dma_handle_caching_state_failure(
 737                                                 pool, d_pages, caching_array,
 738                                                 cpages);
 739                         }
 740                         r = -ENOMEM;
 741                         goto out;
 742                 }
 743                 p = dma_p->p;
 744                 list_add(&dma_p->page_list, d_pages);
 745
 746 #ifdef CONFIG_HIGHMEM
 747                 /* gfp flags of highmem page should never be dma32 so we
 748                  * we should be fine in such case
 749                  */
 750                 if (PageHighMem(p))
 751                         continue;
 752 #endif
 753
 754                 npages = pool->size / PAGE_SIZE;
 755                 for (j = 0; j < npages; ++j) {
 756                         caching_array[cpages++] = p + j;
 757                         if (cpages == max_cpages) {
 758                                 /* Note: Cannot hold the spinlock */
 759                                 r = ttm_set_pages_caching(pool, caching_array,
 760                                                           cpages);
 761                                 if (r) {
 762                                         ttm_dma_handle_caching_state_failure(
 763                                              pool, d_pages, caching_array,
 764                                              cpages);
 765                                         goto out;
 766                                 }
 767                                 cpages = 0;
 768                         }
 769                 }
 770         }
 771
 772         if (cpages) {
 773                 r = ttm_set_pages_caching(pool, caching_array, cpages);
 774                 if (r)
 775                         ttm_dma_handle_caching_state_failure(pool, d_pages,
 776                                         caching_array, cpages);
 777         }
 778 out:
 779         kfree(caching_array);
 780         return r;
 781 }
 782
 783 /*
 784  * @return count of pages still required to fulfill the request.
 785  */
 786 static int ttm_dma_page_pool_fill_locked(struct dma_pool *pool,
 787                                          unsigned long *irq_flags)
 788 {
 789         unsigned count = _manager->options.small;
 790         int r = pool->npages_free;
 791
 792         if (count > pool->npages_free) {
 793                 struct list_head d_pages;
 794
 795                 INIT_LIST_HEAD(&d_pages);
 796
 797                 spin_unlock_irqrestore(&pool->lock, *irq_flags);
 798
 799                 /* Returns how many more are neccessary to fulfill the
 800                  * request. */
 801                 r = ttm_dma_pool_alloc_new_pages(pool, &d_pages, count);
 802
 803                 spin_lock_irqsave(&pool->lock, *irq_flags);
 804                 if (!r) {
 805                         /* Add the fresh to the end.. */
 806                         list_splice(&d_pages, &pool->free_list);
 807                         ++pool->nrefills;
 808                         pool->npages_free += count;
 809                         r = count;
 810                 } else {
 811                         struct dma_page *d_page;
 812                         unsigned cpages = 0;
 813
 814                         pr_debug("%s: Failed to fill %s pool (r:%d)!\n",
 815                                  pool->dev_name, pool->name, r);
 816
 817                         list_for_each_entry(d_page, &d_pages, page_list) {
 818                                 cpages++;
 819                         }
 820                         list_splice_tail(&d_pages, &pool->free_list);
 821                         pool->npages_free += cpages;
 822                         r = cpages;
 823                 }
 824         }
 825         return r;
 826 }
 827
 828 /*
 829  * The populate list is actually a stack (not that is matters as TTM
 830  * allocates one page at a time.
 831  * return dma_page pointer if success, otherwise NULL.
 832  */
 833 static struct dma_page *ttm_dma_pool_get_pages(struct dma_pool *pool,
 834                                   struct ttm_dma_tt *ttm_dma,
 835                                   unsigned index)
 836 {
 837         struct dma_page *d_page = NULL;
 838         struct ttm_tt *ttm = &ttm_dma->ttm;
 839         unsigned long irq_flags;
 840         int count;
 841
 842         spin_lock_irqsave(&pool->lock, irq_flags);
 843         count = ttm_dma_page_pool_fill_locked(pool, &irq_flags);
 844         if (count) {
 845                 d_page = list_first_entry(&pool->free_list, struct dma_page, page_list);
 846                 ttm->pages[index] = d_page->p;
 847                 ttm_dma->dma_address[index] = d_page->dma;
 848                 list_move_tail(&d_page->page_list, &ttm_dma->pages_list);
 849                 pool->npages_in_use += 1;
 850                 pool->npages_free -= 1;
 851         }
 852         spin_unlock_irqrestore(&pool->lock, irq_flags);
 853         return d_page;
 854 }
 855
 856 static gfp_t ttm_dma_pool_gfp_flags(struct ttm_dma_tt *ttm_dma, bool huge)
 857 {
 858         struct ttm_tt *ttm = &ttm_dma->ttm;
 859         gfp_t gfp_flags;
 860
 861         if (ttm->page_flags & TTM_PAGE_FLAG_DMA32)
 862                 gfp_flags = GFP_USER | GFP_DMA32;
 863         else
 864                 gfp_flags = GFP_HIGHUSER;
 865         if (ttm->page_flags & TTM_PAGE_FLAG_ZERO_ALLOC)
 866                 gfp_flags |= __GFP_ZERO;
 867
 868         if (huge) {
 869                 gfp_flags |= GFP_TRANSHUGE_LIGHT | __GFP_NORETRY |
 870                         __GFP_KSWAPD_RECLAIM;
 871                 gfp_flags &= ~__GFP_MOVABLE;
 872                 gfp_flags &= ~__GFP_COMP;
 873         }
 874
 875         if (ttm->page_flags & TTM_PAGE_FLAG_NO_RETRY)
 876                 gfp_flags |= __GFP_RETRY_MAYFAIL;
 877
 878         return gfp_flags;
 879 }
 880
 881 /*
 882  * On success pages list will hold count number of correctly
 883  * cached pages. On failure will hold the negative return value (-ENOMEM, etc).
 884  */
 885 int ttm_dma_populate(struct ttm_dma_tt *ttm_dma, struct device *dev,
 886                         struct ttm_operation_ctx *ctx)
 887 {
 888         struct ttm_mem_global *mem_glob = &ttm_mem_glob;
 889         struct ttm_tt *ttm = &ttm_dma->ttm;
 890         unsigned long num_pages = ttm->num_pages;
 891         struct dma_pool *pool;
 892         struct dma_page *d_page;
 893         enum pool_type type;
 894         unsigned i;
 895         int ret;
 896
 897         if (ttm->state != tt_unpopulated)
 898                 return 0;
 899
 900         if (ttm_check_under_lowerlimit(mem_glob, num_pages, ctx))
 901                 return -ENOMEM;
 902
 903         INIT_LIST_HEAD(&ttm_dma->pages_list);
 904         i = 0;
 905
 906         type = ttm_to_type(ttm->page_flags, ttm->caching_state);
 907
 908 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 909         if (ttm->page_flags & TTM_PAGE_FLAG_DMA32)
 910                 goto skip_huge;
 911
 912         pool = ttm_dma_find_pool(dev, type | IS_HUGE);
 913         if (!pool) {
 914                 gfp_t gfp_flags = ttm_dma_pool_gfp_flags(ttm_dma, true);
 915
 916                 pool = ttm_dma_pool_init(dev, gfp_flags, type | IS_HUGE);
 917                 if (IS_ERR_OR_NULL(pool))
 918                         goto skip_huge;
 919         }
 920
 921         while (num_pages >= HPAGE_PMD_NR) {
 922                 unsigned j;
 923
 924                 d_page = ttm_dma_pool_get_pages(pool, ttm_dma, i);
 925                 if (!d_page)
 926                         break;
 927
 928                 ret = ttm_mem_global_alloc_page(mem_glob, ttm->pages[i],
 929                                                 pool->size, ctx);
 930                 if (unlikely(ret != 0)) {
 931                         ttm_dma_unpopulate(ttm_dma, dev);
 932                         return -ENOMEM;
 933                 }
 934
 935                 d_page->vaddr |= VADDR_FLAG_UPDATED_COUNT;
 936                 for (j = i + 1; j < (i + HPAGE_PMD_NR); ++j) {
 937                         ttm->pages[j] = ttm->pages[j - 1] + 1;
 938                         ttm_dma->dma_address[j] = ttm_dma->dma_address[j - 1] +
 939                                 PAGE_SIZE;
 940                 }
 941
 942                 i += HPAGE_PMD_NR;
 943                 num_pages -= HPAGE_PMD_NR;
 944         }
 945
 946 skip_huge:
 947 #endif
 948
 949         pool = ttm_dma_find_pool(dev, type);
 950         if (!pool) {
 951                 gfp_t gfp_flags = ttm_dma_pool_gfp_flags(ttm_dma, false);
 952
 953                 pool = ttm_dma_pool_init(dev, gfp_flags, type);
 954                 if (IS_ERR_OR_NULL(pool))
 955                         return -ENOMEM;
 956         }
 957
 958         while (num_pages) {
 959                 d_page = ttm_dma_pool_get_pages(pool, ttm_dma, i);
 960                 if (!d_page) {
 961                         ttm_dma_unpopulate(ttm_dma, dev);
 962                         return -ENOMEM;
 963                 }
 964
 965                 ret = ttm_mem_global_alloc_page(mem_glob, ttm->pages[i],
 966                                                 pool->size, ctx);
 967                 if (unlikely(ret != 0)) {
 968                         ttm_dma_unpopulate(ttm_dma, dev);
 969                         return -ENOMEM;
 970                 }
 971
 972                 d_page->vaddr |= VADDR_FLAG_UPDATED_COUNT;
 973                 ++i;
 974                 --num_pages;
 975         }
 976
 977         if (unlikely(ttm->page_flags & TTM_PAGE_FLAG_SWAPPED)) {
 978                 ret = ttm_tt_swapin(ttm);
 979                 if (unlikely(ret != 0)) {
 980                         ttm_dma_unpopulate(ttm_dma, dev);
 981                         return ret;
 982                 }
 983         }
 984
 985         ttm->state = tt_unbound;
 986         return 0;
 987 }
 988 EXPORT_SYMBOL_GPL(ttm_dma_populate);
 989
 990 /* Put all pages in pages list to correct pool to wait for reuse */
 991 void ttm_dma_unpopulate(struct ttm_dma_tt *ttm_dma, struct device *dev)
 992 {
 993         struct ttm_mem_global *mem_glob = &ttm_mem_glob;
 994         struct ttm_tt *ttm = &ttm_dma->ttm;
 995         struct dma_pool *pool;
 996         struct dma_page *d_page, *next;
 997         enum pool_type type;
 998         bool is_cached = false;
 999         unsigned count, i, npages = 0;
1000         unsigned long irq_flags;
1001
1002         type = ttm_to_type(ttm->page_flags, ttm->caching_state);
1003
1004 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1005         pool = ttm_dma_find_pool(dev, type | IS_HUGE);
1006         if (pool) {
1007                 count = 0;
1008                 list_for_each_entry_safe(d_page, next, &ttm_dma->pages_list,
1009                                          page_list) {
1010                         if (!(d_page->vaddr & VADDR_FLAG_HUGE_POOL))
1011                                 continue;
1012
1013                         count++;
1014                         if (d_page->vaddr & VADDR_FLAG_UPDATED_COUNT) {
1015                                 ttm_mem_global_free_page(mem_glob, d_page->p,
1016                                                          pool->size);
1017                                 d_page->vaddr &= ~VADDR_FLAG_UPDATED_COUNT;
1018                         }
1019                         ttm_dma_page_put(pool, d_page);
1020                 }
1021
1022                 spin_lock_irqsave(&pool->lock, irq_flags);
1023                 pool->npages_in_use -= count;
1024                 pool->nfrees += count;
1025                 spin_unlock_irqrestore(&pool->lock, irq_flags);
1026         }
1027 #endif
1028
1029         pool = ttm_dma_find_pool(dev, type);
1030         if (!pool)
1031                 return;
1032
1033         is_cached = (ttm_dma_find_pool(pool->dev,
1034                      ttm_to_type(ttm->page_flags, tt_cached)) == pool);
1035
1036         /* make sure pages array match list and count number of pages */
1037         count = 0;
1038         list_for_each_entry_safe(d_page, next, &ttm_dma->pages_list,
1039                                  page_list) {
1040                 ttm->pages[count] = d_page->p;
1041                 count++;
1042
1043                 if (d_page->vaddr & VADDR_FLAG_UPDATED_COUNT) {
1044                         ttm_mem_global_free_page(mem_glob, d_page->p,
1045                                                  pool->size);
1046                         d_page->vaddr &= ~VADDR_FLAG_UPDATED_COUNT;
1047                 }
1048
1049                 if (is_cached)
1050                         ttm_dma_page_put(pool, d_page);
1051         }
1052
1053         spin_lock_irqsave(&pool->lock, irq_flags);
1054         pool->npages_in_use -= count;
1055         if (is_cached) {
1056                 pool->nfrees += count;
1057         } else {
1058                 pool->npages_free += count;
1059                 list_splice(&ttm_dma->pages_list, &pool->free_list);
1060                 /*
1061                  * Wait to have at at least NUM_PAGES_TO_ALLOC number of pages
1062                  * to free in order to minimize calls to set_memory_wb().
1063                  */
1064                 if (pool->npages_free >= (_manager->options.max_size +
1065                                           NUM_PAGES_TO_ALLOC))
1066                         npages = pool->npages_free - _manager->options.max_size;
1067         }
1068         spin_unlock_irqrestore(&pool->lock, irq_flags);
1069
1070         INIT_LIST_HEAD(&ttm_dma->pages_list);
1071         for (i = 0; i < ttm->num_pages; i++) {
1072                 ttm->pages[i] = NULL;
1073                 ttm_dma->dma_address[i] = 0;
1074         }
1075
1076         /* shrink pool if necessary (only on !is_cached pools)*/
1077         if (npages)
1078                 ttm_dma_page_pool_free(pool, npages, false);
1079         ttm->state = tt_unpopulated;
1080 }
1081 EXPORT_SYMBOL_GPL(ttm_dma_unpopulate);
1082
1083 /**
1084  * Callback for mm to request pool to reduce number of page held.
1085  *
1086  * XXX: (dchinner) Deadlock warning!
1087  *
1088  * I'm getting sadder as I hear more pathetical whimpers about needing per-pool
1089  * shrinkers
1090  */
1091 static unsigned long
1092 ttm_dma_pool_shrink_scan(struct shrinker *shrink, struct shrink_control *sc)
1093 {
1094         static unsigned start_pool;
1095         unsigned idx = 0;
1096         unsigned pool_offset;
1097         unsigned shrink_pages = sc->nr_to_scan;
1098         struct device_pools *p;
1099         unsigned long freed = 0;
1100
1101         if (list_empty(&_manager->pools))
1102                 return SHRINK_STOP;
1103
1104         if (!mutex_trylock(&_manager->lock))
1105                 return SHRINK_STOP;
1106         if (!_manager->npools)
1107                 goto out;
1108         pool_offset = ++start_pool % _manager->npools;
1109         list_for_each_entry(p, &_manager->pools, pools) {
1110                 unsigned nr_free;
1111
1112                 if (!p->dev)
1113                         continue;
1114                 if (shrink_pages == 0)
1115                         break;
1116                 /* Do it in round-robin fashion. */
1117                 if (++idx < pool_offset)
1118                         continue;
1119                 nr_free = shrink_pages;
1120                 /* OK to use static buffer since global mutex is held. */
1121                 shrink_pages = ttm_dma_page_pool_free(p->pool, nr_free, true);
1122                 freed += nr_free - shrink_pages;
1123
1124                 pr_debug("%s: (%s:%d) Asked to shrink %d, have %d more to go\n",
1125                          p->pool->dev_name, p->pool->name, current->pid,
1126                          nr_free, shrink_pages);
1127         }
1128 out:
1129         mutex_unlock(&_manager->lock);
1130         return freed;
1131 }
1132
1133 static unsigned long
1134 ttm_dma_pool_shrink_count(struct shrinker *shrink, struct shrink_control *sc)
1135 {
1136         struct device_pools *p;
1137         unsigned long count = 0;
1138
1139         if (!mutex_trylock(&_manager->lock))
1140                 return 0;
1141         list_for_each_entry(p, &_manager->pools, pools)
1142                 count += p->pool->npages_free;
1143         mutex_unlock(&_manager->lock);
1144         return count;
1145 }
1146
1147 static int ttm_dma_pool_mm_shrink_init(struct ttm_pool_manager *manager)
1148 {
1149         manager->mm_shrink.count_objects = ttm_dma_pool_shrink_count;
1150         manager->mm_shrink.scan_objects = &ttm_dma_pool_shrink_scan;
1151         manager->mm_shrink.seeks = 1;
1152         return register_shrinker(&manager->mm_shrink);
1153 }
1154
1155 static void ttm_dma_pool_mm_shrink_fini(struct ttm_pool_manager *manager)
1156 {
1157         unregister_shrinker(&manager->mm_shrink);
1158 }
1159
1160 int ttm_dma_page_alloc_init(struct ttm_mem_global *glob, unsigned max_pages)
1161 {
1162         int ret;
1163
1164         WARN_ON(_manager);
1165
1166         pr_info("Initializing DMA pool allocator\n");
1167
1168         _manager = kzalloc(sizeof(*_manager), GFP_KERNEL);
1169         if (!_manager)
1170                 return -ENOMEM;
1171
1172         mutex_init(&_manager->lock);
1173         INIT_LIST_HEAD(&_manager->pools);
1174
1175         _manager->options.max_size = max_pages;
1176         _manager->options.small = SMALL_ALLOCATION;
1177         _manager->options.alloc_size = NUM_PAGES_TO_ALLOC;
1178
1179         /* This takes care of auto-freeing the _manager */
1180         ret = kobject_init_and_add(&_manager->kobj, &ttm_pool_kobj_type,
1181                                    &glob->kobj, "dma_pool");
1182         if (unlikely(ret != 0))
1183                 goto error;
1184
1185         ret = ttm_dma_pool_mm_shrink_init(_manager);
1186         if (unlikely(ret != 0))
1187                 goto error;
1188         return 0;
1189
1190 error:
1191         kobject_put(&_manager->kobj);
1192         _manager = NULL;
1193         return ret;
1194 }
1195
1196 void ttm_dma_page_alloc_fini(void)
1197 {
1198         struct device_pools *p, *t;
1199
1200         pr_info("Finalizing DMA pool allocator\n");
1201         ttm_dma_pool_mm_shrink_fini(_manager);
1202
1203         list_for_each_entry_safe_reverse(p, t, &_manager->pools, pools) {
1204                 dev_dbg(p->dev, "(%s:%d) Freeing.\n", p->pool->name,
1205                         current->pid);
1206                 WARN_ON(devres_destroy(p->dev, ttm_dma_pool_release,
1207                         ttm_dma_pool_match, p->pool));
1208                 ttm_dma_free_pool(p->dev, p->pool->type);
1209         }
1210         kobject_put(&_manager->kobj);
1211         _manager = NULL;
1212 }
1213
1214 int ttm_dma_page_alloc_debugfs(struct seq_file *m, void *data)
1215 {
1216         struct device_pools *p;
1217         struct dma_pool *pool = NULL;
1218
1219         if (!_manager) {
1220                 seq_printf(m, "No pool allocator running.\n");
1221                 return 0;
1222         }
1223         seq_printf(m, "         pool      refills   pages freed    inuse available     name\n");
1224         mutex_lock(&_manager->lock);
1225         list_for_each_entry(p, &_manager->pools, pools) {
1226                 struct device *dev = p->dev;
1227                 if (!dev)
1228                         continue;
1229                 pool = p->pool;
1230                 seq_printf(m, "%13s %12ld %13ld %8d %8d %8s\n",
1231                                 pool->name, pool->nrefills,
1232                                 pool->nfrees, pool->npages_in_use,
1233                                 pool->npages_free,
1234                                 pool->dev_name);
1235         }
1236         mutex_unlock(&_manager->lock);
1237         return 0;
1238 }
1239 EXPORT_SYMBOL_GPL(ttm_dma_page_alloc_debugfs);