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 #if defined(CONFIG_SWIOTLB) || defined(CONFIG_INTEL_IOMMU)
  37 #define pr_fmt(fmt) "[TTM] " fmt
  38
  39 #include <linux/dma-mapping.h>
  40 #include <linux/list.h>
  41 #include <linux/seq_file.h> /* for seq_printf */
  42 #include <linux/slab.h>
  43 #include <linux/spinlock.h>
  44 #include <linux/highmem.h>
  45 #include <linux/mm_types.h>
  46 #include <linux/module.h>
  47 #include <linux/mm.h>
  48 #include <linux/atomic.h>
  49 #include <linux/device.h>
  50 #include <linux/kthread.h>
  51 #include <drm/ttm/ttm_bo_driver.h>
  52 #include <drm/ttm/ttm_page_alloc.h>
  53 #include <drm/ttm/ttm_set_memory.h>
  54
  55 #define NUM_PAGES_TO_ALLOC              (PAGE_SIZE/sizeof(struct page *))
  56 #define SMALL_ALLOCATION                4
  57 #define FREE_ALL_PAGES                  (~0U)
  58 #define VADDR_FLAG_HUGE_POOL            1UL
  59 #define VADDR_FLAG_UPDATED_COUNT        2UL
  60
  61 enum pool_type {
  62         IS_UNDEFINED    = 0,
  63         IS_WC           = 1 << 1,
  64         IS_UC           = 1 << 2,
  65         IS_CACHED       = 1 << 3,
  66         IS_DMA32        = 1 << 4,
  67         IS_HUGE         = 1 << 5
  68 };
  69
  70 /*
  71  * The pool structure. There are up to nine pools:
  72  *  - generic (not restricted to DMA32):
  73  *      - write combined, uncached, cached.
  74  *  - dma32 (up to 2^32 - so up 4GB):
  75  *      - write combined, uncached, cached.
  76  *  - huge (not restricted to DMA32):
  77  *      - write combined, uncached, cached.
  78  * for each 'struct device'. The 'cached' is for pages that are actively used.
  79  * The other ones can be shrunk by the shrinker API if neccessary.
  80  * @pools: The 'struct device->dma_pools' link.
  81  * @type: Type of the pool
  82  * @lock: Protects the free_list from concurrnet access. Must be
  83  * used with irqsave/irqrestore variants because pool allocator maybe called
  84  * from delayed work.
  85  * @free_list: Pool of pages that are free to be used. No order requirements.
  86  * @dev: The device that is associated with these pools.
  87  * @size: Size used during DMA allocation.
  88  * @npages_free: Count of available pages for re-use.
  89  * @npages_in_use: Count of pages that are in use.
  90  * @nfrees: Stats when pool is shrinking.
  91  * @nrefills: Stats when the pool is grown.
  92  * @gfp_flags: Flags to pass for alloc_page.
  93  * @name: Name of the pool.
  94  * @dev_name: Name derieved from dev - similar to how dev_info works.
  95  *   Used during shutdown as the dev_info during release is unavailable.
  96  */
  97 struct dma_pool {
  98         struct list_head pools; /* The 'struct device->dma_pools link */
  99         enum pool_type type;
 100         spinlock_t lock;
 101         struct list_head free_list;
 102         struct device *dev;
 103         unsigned size;
 104         unsigned npages_free;
 105         unsigned npages_in_use;
 106         unsigned long nfrees; /* Stats when shrunk. */
 107         unsigned long nrefills; /* Stats when grown. */
 108         gfp_t gfp_flags;
 109         char name[13]; /* "cached dma32" */
 110         char dev_name[64]; /* Constructed from dev */
 111 };
 112
 113 /*
 114  * The accounting page keeping track of the allocated page along with
 115  * the DMA address.
 116  * @page_list: The link to the 'page_list' in 'struct dma_pool'.
 117  * @vaddr: The virtual address of the page and a flag if the page belongs to a
 118  * huge pool
 119  * @dma: The bus address of the page. If the page is not allocated
 120  *   via the DMA API, it will be -1.
 121  */
 122 struct dma_page {
 123         struct list_head page_list;
 124         unsigned long vaddr;
 125         struct page *p;
 126         dma_addr_t dma;
 127 };
 128
 129 /*
 130  * Limits for the pool. They are handled without locks because only place where
 131  * they may change is in sysfs store. They won't have immediate effect anyway
 132  * so forcing serialization to access them is pointless.
 133  */
 134
 135 struct ttm_pool_opts {
 136         unsigned        alloc_size;
 137         unsigned        max_size;
 138         unsigned        small;
 139 };
 140
 141 /*
 142  * Contains the list of all of the 'struct device' and their corresponding
 143  * DMA pools. Guarded by _mutex->lock.
 144  * @pools: The link to 'struct ttm_pool_manager->pools'
 145  * @dev: The 'struct device' associated with the 'pool'
 146  * @pool: The 'struct dma_pool' associated with the 'dev'
 147  */
 148 struct device_pools {
 149         struct list_head pools;
 150         struct device *dev;
 151         struct dma_pool *pool;
 152 };
 153
 154 /*
 155  * struct ttm_pool_manager - Holds memory pools for fast allocation
 156  *
 157  * @lock: Lock used when adding/removing from pools
 158  * @pools: List of 'struct device' and 'struct dma_pool' tuples.
 159  * @options: Limits for the pool.
 160  * @npools: Total amount of pools in existence.
 161  * @shrinker: The structure used by [un|]register_shrinker
 162  */
 163 struct ttm_pool_manager {
 164         struct mutex            lock;
 165         struct list_head        pools;
 166         struct ttm_pool_opts    options;
 167         unsigned                npools;
 168         struct shrinker         mm_shrink;
 169         struct kobject          kobj;
 170 };
 171
 172 static struct ttm_pool_manager *_manager;
 173
 174 static struct attribute ttm_page_pool_max = {
 175         .name = "pool_max_size",
 176         .mode = S_IRUGO | S_IWUSR
 177 };
 178 static struct attribute ttm_page_pool_small = {
 179         .name = "pool_small_allocation",
 180         .mode = S_IRUGO | S_IWUSR
 181 };
 182 static struct attribute ttm_page_pool_alloc_size = {
 183         .name = "pool_allocation_size",
 184         .mode = S_IRUGO | S_IWUSR
 185 };
 186
 187 static struct attribute *ttm_pool_attrs[] = {
 188         &ttm_page_pool_max,
 189         &ttm_page_pool_small,
 190         &ttm_page_pool_alloc_size,
 191         NULL
 192 };
 193
 194 static void ttm_pool_kobj_release(struct kobject *kobj)
 195 {
 196         struct ttm_pool_manager *m =
 197                 container_of(kobj, struct ttm_pool_manager, kobj);
 198         kfree(m);
 199 }
 200
 201 static ssize_t ttm_pool_store(struct kobject *kobj, struct attribute *attr,
 202                               const char *buffer, size_t size)
 203 {
 204         struct ttm_pool_manager *m =
 205                 container_of(kobj, struct ttm_pool_manager, kobj);
 206         int chars;
 207         unsigned val;
 208
 209         chars = sscanf(buffer, "%u", &val);
 210         if (chars == 0)
 211                 return size;
 212
 213         /* Convert kb to number of pages */
 214         val = val / (PAGE_SIZE >> 10);
 215
 216         if (attr == &ttm_page_pool_max) {
 217                 m->options.max_size = val;
 218         } else if (attr == &ttm_page_pool_small) {
 219                 m->options.small = val;
 220         } else if (attr == &ttm_page_pool_alloc_size) {
 221                 if (val > NUM_PAGES_TO_ALLOC*8) {
 222                         pr_err("Setting allocation size to %lu is not allowed. Recommended size is %lu\n",
 223                                NUM_PAGES_TO_ALLOC*(PAGE_SIZE >> 7),
 224                                NUM_PAGES_TO_ALLOC*(PAGE_SIZE >> 10));
 225                         return size;
 226                 } else if (val > NUM_PAGES_TO_ALLOC) {
 227                         pr_warn("Setting allocation size to larger than %lu is not recommended\n",
 228                                 NUM_PAGES_TO_ALLOC*(PAGE_SIZE >> 10));
 229                 }
 230                 m->options.alloc_size = val;
 231         }
 232
 233         return size;
 234 }
 235
 236 static ssize_t ttm_pool_show(struct kobject *kobj, struct attribute *attr,
 237                              char *buffer)
 238 {
 239         struct ttm_pool_manager *m =
 240                 container_of(kobj, struct ttm_pool_manager, kobj);
 241         unsigned val = 0;
 242
 243         if (attr == &ttm_page_pool_max)
 244                 val = m->options.max_size;
 245         else if (attr == &ttm_page_pool_small)
 246                 val = m->options.small;
 247         else if (attr == &ttm_page_pool_alloc_size)
 248                 val = m->options.alloc_size;
 249
 250         val = val * (PAGE_SIZE >> 10);
 251
 252         return snprintf(buffer, PAGE_SIZE, "%u\n", val);
 253 }
 254
 255 static const struct sysfs_ops ttm_pool_sysfs_ops = {
 256         .show = &ttm_pool_show,
 257         .store = &ttm_pool_store,
 258 };
 259
 260 static struct kobj_type ttm_pool_kobj_type = {
 261         .release = &ttm_pool_kobj_release,
 262         .sysfs_ops = &ttm_pool_sysfs_ops,
 263         .default_attrs = ttm_pool_attrs,
 264 };
 265
 266 static int ttm_set_pages_caching(struct dma_pool *pool,
 267                                  struct page **pages, unsigned cpages)
 268 {
 269         int r = 0;
 270         /* Set page caching */
 271         if (pool->type & IS_UC) {
 272                 r = ttm_set_pages_array_uc(pages, cpages);
 273                 if (r)
 274                         pr_err("%s: Failed to set %d pages to uc!\n",
 275                                pool->dev_name, cpages);
 276         }
 277         if (pool->type & IS_WC) {
 278                 r = ttm_set_pages_array_wc(pages, cpages);
 279                 if (r)
 280                         pr_err("%s: Failed to set %d pages to wc!\n",
 281                                pool->dev_name, cpages);
 282         }
 283         return r;
 284 }
 285
 286 static void __ttm_dma_free_page(struct dma_pool *pool, struct dma_page *d_page)
 287 {
 288         unsigned long attrs = 0;
 289         dma_addr_t dma = d_page->dma;
 290         d_page->vaddr &= ~VADDR_FLAG_HUGE_POOL;
 291         if (pool->type & IS_HUGE)
 292                 attrs = DMA_ATTR_NO_WARN;
 293
 294         dma_free_attrs(pool->dev, pool->size, (void *)d_page->vaddr, dma, attrs);
 295
 296         kfree(d_page);
 297         d_page = NULL;
 298 }
 299 static struct dma_page *__ttm_dma_alloc_page(struct dma_pool *pool)
 300 {
 301         struct dma_page *d_page;
 302         unsigned long attrs = 0;
 303         void *vaddr;
 304
 305         d_page = kmalloc(sizeof(struct dma_page), GFP_KERNEL);
 306         if (!d_page)
 307                 return NULL;
 308
 309         if (pool->type & IS_HUGE)
 310                 attrs = DMA_ATTR_NO_WARN;
 311
 312         vaddr = dma_alloc_attrs(pool->dev, pool->size, &d_page->dma,
 313                                 pool->gfp_flags, attrs);
 314         if (vaddr) {
 315                 if (is_vmalloc_addr(vaddr))
 316                         d_page->p = vmalloc_to_page(vaddr);
 317                 else
 318                         d_page->p = virt_to_page(vaddr);
 319                 d_page->vaddr = (unsigned long)vaddr;
 320                 if (pool->type & IS_HUGE)
 321                         d_page->vaddr |= VADDR_FLAG_HUGE_POOL;
 322         } else {
 323                 kfree(d_page);
 324                 d_page = NULL;
 325         }
 326         return d_page;
 327 }
 328 static enum pool_type ttm_to_type(int flags, enum ttm_caching_state cstate)
 329 {
 330         enum pool_type type = IS_UNDEFINED;
 331
 332         if (flags & TTM_PAGE_FLAG_DMA32)
 333                 type |= IS_DMA32;
 334         if (cstate == tt_cached)
 335                 type |= IS_CACHED;
 336         else if (cstate == tt_uncached)
 337                 type |= IS_UC;
 338         else
 339                 type |= IS_WC;
 340
 341         return type;
 342 }
 343
 344 static void ttm_pool_update_free_locked(struct dma_pool *pool,
 345                                         unsigned freed_pages)
 346 {
 347         pool->npages_free -= freed_pages;
 348         pool->nfrees += freed_pages;
 349
 350 }
 351
 352 /* set memory back to wb and free the pages. */
 353 static void ttm_dma_page_put(struct dma_pool *pool, struct dma_page *d_page)
 354 {
 355         struct page *page = d_page->p;
 356         unsigned num_pages;
 357
 358         /* Don't set WB on WB page pool. */
 359         if (!(pool->type & IS_CACHED)) {
 360                 num_pages = pool->size / PAGE_SIZE;
 361                 if (ttm_set_pages_wb(page, num_pages))
 362                         pr_err("%s: Failed to set %d pages to wb!\n",
 363                                pool->dev_name, num_pages);
 364         }
 365
 366         list_del(&d_page->page_list);
 367         __ttm_dma_free_page(pool, d_page);
 368 }
 369
 370 static void ttm_dma_pages_put(struct dma_pool *pool, struct list_head *d_pages,
 371                               struct page *pages[], unsigned npages)
 372 {
 373         struct dma_page *d_page, *tmp;
 374
 375         if (pool->type & IS_HUGE) {
 376                 list_for_each_entry_safe(d_page, tmp, d_pages, page_list)
 377                         ttm_dma_page_put(pool, d_page);
 378
 379                 return;
 380         }
 381
 382         /* Don't set WB on WB page pool. */
 383         if (npages && !(pool->type & IS_CACHED) &&
 384             ttm_set_pages_array_wb(pages, npages))
 385                 pr_err("%s: Failed to set %d pages to wb!\n",
 386                        pool->dev_name, npages);
 387
 388         list_for_each_entry_safe(d_page, tmp, d_pages, page_list) {
 389                 list_del(&d_page->page_list);
 390                 __ttm_dma_free_page(pool, d_page);
 391         }
 392 }
 393
 394 /*
 395  * Free pages from pool.
 396  *
 397  * To prevent hogging the ttm_swap process we only free NUM_PAGES_TO_ALLOC
 398  * number of pages in one go.
 399  *
 400  * @pool: to free the pages from
 401  * @nr_free: If set to true will free all pages in pool
 402  * @use_static: Safe to use static buffer
 403  **/
 404 static unsigned ttm_dma_page_pool_free(struct dma_pool *pool, unsigned nr_free,
 405                                        bool use_static)
 406 {
 407         static struct page *static_buf[NUM_PAGES_TO_ALLOC];
 408         unsigned long irq_flags;
 409         struct dma_page *dma_p, *tmp;
 410         struct page **pages_to_free;
 411         struct list_head d_pages;
 412         unsigned freed_pages = 0,
 413                  npages_to_free = nr_free;
 414
 415         if (NUM_PAGES_TO_ALLOC < nr_free)
 416                 npages_to_free = NUM_PAGES_TO_ALLOC;
 417
 418         if (use_static)
 419                 pages_to_free = static_buf;
 420         else
 421                 pages_to_free = kmalloc_array(npages_to_free,
 422                                               sizeof(struct page *),
 423                                               GFP_KERNEL);
 424
 425         if (!pages_to_free) {
 426                 pr_debug("%s: Failed to allocate memory for pool free operation\n",
 427                        pool->dev_name);
 428                 return 0;
 429         }
 430         INIT_LIST_HEAD(&d_pages);
 431 restart:
 432         spin_lock_irqsave(&pool->lock, irq_flags);
 433
 434         /* We picking the oldest ones off the list */
 435         list_for_each_entry_safe_reverse(dma_p, tmp, &pool->free_list,
 436                                          page_list) {
 437                 if (freed_pages >= npages_to_free)
 438                         break;
 439
 440                 /* Move the dma_page from one list to another. */
 441                 list_move(&dma_p->page_list, &d_pages);
 442
 443                 pages_to_free[freed_pages++] = dma_p->p;
 444                 /* We can only remove NUM_PAGES_TO_ALLOC at a time. */
 445                 if (freed_pages >= NUM_PAGES_TO_ALLOC) {
 446
 447                         ttm_pool_update_free_locked(pool, freed_pages);
 448                         /**
 449                          * Because changing page caching is costly
 450                          * we unlock the pool to prevent stalling.
 451                          */
 452                         spin_unlock_irqrestore(&pool->lock, irq_flags);
 453
 454                         ttm_dma_pages_put(pool, &d_pages, pages_to_free,
 455                                           freed_pages);
 456
 457                         INIT_LIST_HEAD(&d_pages);
 458
 459                         if (likely(nr_free != FREE_ALL_PAGES))
 460                                 nr_free -= freed_pages;
 461
 462                         if (NUM_PAGES_TO_ALLOC >= nr_free)
 463                                 npages_to_free = nr_free;
 464                         else
 465                                 npages_to_free = NUM_PAGES_TO_ALLOC;
 466
 467                         freed_pages = 0;
 468
 469                         /* free all so restart the processing */
 470                         if (nr_free)
 471                                 goto restart;
 472
 473                         /* Not allowed to fall through or break because
 474                          * following context is inside spinlock while we are
 475                          * outside here.
 476                          */
 477                         goto out;
 478
 479                 }
 480         }
 481
 482         /* remove range of pages from the pool */
 483         if (freed_pages) {
 484                 ttm_pool_update_free_locked(pool, freed_pages);
 485                 nr_free -= freed_pages;
 486         }
 487
 488         spin_unlock_irqrestore(&pool->lock, irq_flags);
 489
 490         if (freed_pages)
 491                 ttm_dma_pages_put(pool, &d_pages, pages_to_free, freed_pages);
 492 out:
 493         if (pages_to_free != static_buf)
 494                 kfree(pages_to_free);
 495         return nr_free;
 496 }
 497
 498 static void ttm_dma_free_pool(struct device *dev, enum pool_type type)
 499 {
 500         struct device_pools *p;
 501         struct dma_pool *pool;
 502
 503         if (!dev)
 504                 return;
 505
 506         mutex_lock(&_manager->lock);
 507         list_for_each_entry_reverse(p, &_manager->pools, pools) {
 508                 if (p->dev != dev)
 509                         continue;
 510                 pool = p->pool;
 511                 if (pool->type != type)
 512                         continue;
 513
 514                 list_del(&p->pools);
 515                 kfree(p);
 516                 _manager->npools--;
 517                 break;
 518         }
 519         list_for_each_entry_reverse(pool, &dev->dma_pools, pools) {
 520                 if (pool->type != type)
 521                         continue;
 522                 /* Takes a spinlock.. */
 523                 /* OK to use static buffer since global mutex is held. */
 524                 ttm_dma_page_pool_free(pool, FREE_ALL_PAGES, true);
 525                 WARN_ON(((pool->npages_in_use + pool->npages_free) != 0));
 526                 /* This code path is called after _all_ references to the
 527                  * struct device has been dropped - so nobody should be
 528                  * touching it. In case somebody is trying to _add_ we are
 529                  * guarded by the mutex. */
 530                 list_del(&pool->pools);
 531                 kfree(pool);
 532                 break;
 533         }
 534         mutex_unlock(&_manager->lock);
 535 }
 536
 537 /*
 538  * On free-ing of the 'struct device' this deconstructor is run.
 539  * Albeit the pool might have already been freed earlier.
 540  */
 541 static void ttm_dma_pool_release(struct device *dev, void *res)
 542 {
 543         struct dma_pool *pool = *(struct dma_pool **)res;
 544
 545         if (pool)
 546                 ttm_dma_free_pool(dev, pool->type);
 547 }
 548
 549 static int ttm_dma_pool_match(struct device *dev, void *res, void *match_data)
 550 {
 551         return *(struct dma_pool **)res == match_data;
 552 }
 553
 554 static struct dma_pool *ttm_dma_pool_init(struct device *dev, gfp_t flags,
 555                                           enum pool_type type)
 556 {
 557         const char *n[] = {"wc", "uc", "cached", " dma32", "huge"};
 558         enum pool_type t[] = {IS_WC, IS_UC, IS_CACHED, IS_DMA32, IS_HUGE};
 559         struct device_pools *sec_pool = NULL;
 560         struct dma_pool *pool = NULL, **ptr;
 561         unsigned i;
 562         int ret = -ENODEV;
 563         char *p;
 564
 565         if (!dev)
 566                 return NULL;
 567
 568         ptr = devres_alloc(ttm_dma_pool_release, sizeof(*ptr), GFP_KERNEL);
 569         if (!ptr)
 570                 return NULL;
 571
 572         ret = -ENOMEM;
 573
 574         pool = kmalloc_node(sizeof(struct dma_pool), GFP_KERNEL,
 575                             dev_to_node(dev));
 576         if (!pool)
 577                 goto err_mem;
 578
 579         sec_pool = kmalloc_node(sizeof(struct device_pools), GFP_KERNEL,
 580                                 dev_to_node(dev));
 581         if (!sec_pool)
 582                 goto err_mem;
 583
 584         INIT_LIST_HEAD(&sec_pool->pools);
 585         sec_pool->dev = dev;
 586         sec_pool->pool =  pool;
 587
 588         INIT_LIST_HEAD(&pool->free_list);
 589         INIT_LIST_HEAD(&pool->pools);
 590         spin_lock_init(&pool->lock);
 591         pool->dev = dev;
 592         pool->npages_free = pool->npages_in_use = 0;
 593         pool->nfrees = 0;
 594         pool->gfp_flags = flags;
 595         if (type & IS_HUGE)
 596 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 597                 pool->size = HPAGE_PMD_SIZE;
 598 #else
 599                 BUG();
 600 #endif
 601         else
 602                 pool->size = PAGE_SIZE;
 603         pool->type = type;
 604         pool->nrefills = 0;
 605         p = pool->name;
 606         for (i = 0; i < ARRAY_SIZE(t); i++) {
 607                 if (type & t[i]) {
 608                         p += snprintf(p, sizeof(pool->name) - (p - pool->name),
 609                                       "%s", n[i]);
 610                 }
 611         }
 612         *p = 0;
 613         /* We copy the name for pr_ calls b/c when dma_pool_destroy is called
 614          * - the kobj->name has already been deallocated.*/
 615         snprintf(pool->dev_name, sizeof(pool->dev_name), "%s %s",
 616                  dev_driver_string(dev), dev_name(dev));
 617         mutex_lock(&_manager->lock);
 618         /* You can get the dma_pool from either the global: */
 619         list_add(&sec_pool->pools, &_manager->pools);
 620         _manager->npools++;
 621         /* or from 'struct device': */
 622         list_add(&pool->pools, &dev->dma_pools);
 623         mutex_unlock(&_manager->lock);
 624
 625         *ptr = pool;
 626         devres_add(dev, ptr);
 627
 628         return pool;
 629 err_mem:
 630         devres_free(ptr);
 631         kfree(sec_pool);
 632         kfree(pool);
 633         return ERR_PTR(ret);
 634 }
 635
 636 static struct dma_pool *ttm_dma_find_pool(struct device *dev,
 637                                           enum pool_type type)
 638 {
 639         struct dma_pool *pool, *tmp;
 640
 641         if (type == IS_UNDEFINED)
 642                 return NULL;
 643
 644         /* NB: We iterate on the 'struct dev' which has no spinlock, but
 645          * it does have a kref which we have taken. The kref is taken during
 646          * graphic driver loading - in the drm_pci_init it calls either
 647          * pci_dev_get or pci_register_driver which both end up taking a kref
 648          * on 'struct device'.
 649          *
 650          * On teardown, the graphic drivers end up quiescing the TTM (put_pages)
 651          * and calls the dev_res deconstructors: ttm_dma_pool_release. The nice
 652          * thing is at that point of time there are no pages associated with the
 653          * driver so this function will not be called.
 654          */
 655         list_for_each_entry_safe(pool, tmp, &dev->dma_pools, pools)
 656                 if (pool->type == type)
 657                         return pool;
 658         return NULL;
 659 }
 660
 661 /*
 662  * Free pages the pages that failed to change the caching state. If there
 663  * are pages that have changed their caching state already put them to the
 664  * pool.
 665  */
 666 static void ttm_dma_handle_caching_state_failure(struct dma_pool *pool,
 667                                                  struct list_head *d_pages,
 668                                                  struct page **failed_pages,
 669                                                  unsigned cpages)
 670 {
 671         struct dma_page *d_page, *tmp;
 672         struct page *p;
 673         unsigned i = 0;
 674
 675         p = failed_pages[0];
 676         if (!p)
 677                 return;
 678         /* Find the failed page. */
 679         list_for_each_entry_safe(d_page, tmp, d_pages, page_list) {
 680                 if (d_page->p != p)
 681                         continue;
 682                 /* .. and then progress over the full list. */
 683                 list_del(&d_page->page_list);
 684                 __ttm_dma_free_page(pool, d_page);
 685                 if (++i < cpages)
 686                         p = failed_pages[i];
 687                 else
 688                         break;
 689         }
 690
 691 }
 692
 693 /*
 694  * Allocate 'count' pages, and put 'need' number of them on the
 695  * 'pages' and as well on the 'dma_address' starting at 'dma_offset' offset.
 696  * The full list of pages should also be on 'd_pages'.
 697  * We return zero for success, and negative numbers as errors.
 698  */
 699 static int ttm_dma_pool_alloc_new_pages(struct dma_pool *pool,
 700                                         struct list_head *d_pages,
 701                                         unsigned count)
 702 {
 703         struct page **caching_array;
 704         struct dma_page *dma_p;
 705         struct page *p;
 706         int r = 0;
 707         unsigned i, j, npages, cpages;
 708         unsigned max_cpages = min(count,
 709                         (unsigned)(PAGE_SIZE/sizeof(struct page *)));
 710
 711         /* allocate array for page caching change */
 712         caching_array = kmalloc_array(max_cpages, sizeof(struct page *),
 713                                       GFP_KERNEL);
 714
 715         if (!caching_array) {
 716                 pr_debug("%s: Unable to allocate table for new pages\n",
 717                        pool->dev_name);
 718                 return -ENOMEM;
 719         }
 720
 721         if (count > 1)
 722                 pr_debug("%s: (%s:%d) Getting %d pages\n",
 723                          pool->dev_name, pool->name, current->pid, count);
 724
 725         for (i = 0, cpages = 0; i < count; ++i) {
 726                 dma_p = __ttm_dma_alloc_page(pool);
 727                 if (!dma_p) {
 728                         pr_debug("%s: Unable to get page %u\n",
 729                                  pool->dev_name, i);
 730
 731                         /* store already allocated pages in the pool after
 732                          * setting the caching state */
 733                         if (cpages) {
 734                                 r = ttm_set_pages_caching(pool, caching_array,
 735                                                           cpages);
 736                                 if (r)
 737                                         ttm_dma_handle_caching_state_failure(
 738                                                 pool, d_pages, caching_array,
 739                                                 cpages);
 740                         }
 741                         r = -ENOMEM;
 742                         goto out;
 743                 }
 744                 p = dma_p->p;
 745                 list_add(&dma_p->page_list, d_pages);
 746
 747 #ifdef CONFIG_HIGHMEM
 748                 /* gfp flags of highmem page should never be dma32 so we
 749                  * we should be fine in such case
 750                  */
 751                 if (PageHighMem(p))
 752                         continue;
 753 #endif
 754
 755                 npages = pool->size / PAGE_SIZE;
 756                 for (j = 0; j < npages; ++j) {
 757                         caching_array[cpages++] = p + j;
 758                         if (cpages == max_cpages) {
 759                                 /* Note: Cannot hold the spinlock */
 760                                 r = ttm_set_pages_caching(pool, caching_array,
 761                                                           cpages);
 762                                 if (r) {
 763                                         ttm_dma_handle_caching_state_failure(
 764                                              pool, d_pages, caching_array,
 765                                              cpages);
 766                                         goto out;
 767                                 }
 768                                 cpages = 0;
 769                         }
 770                 }
 771         }
 772
 773         if (cpages) {
 774                 r = ttm_set_pages_caching(pool, caching_array, cpages);
 775                 if (r)
 776                         ttm_dma_handle_caching_state_failure(pool, d_pages,
 777                                         caching_array, cpages);
 778         }
 779 out:
 780         kfree(caching_array);
 781         return r;
 782 }
 783
 784 /*
 785  * @return count of pages still required to fulfill the request.
 786  */
 787 static int ttm_dma_page_pool_fill_locked(struct dma_pool *pool,
 788                                          unsigned long *irq_flags)
 789 {
 790         unsigned count = _manager->options.small;
 791         int r = pool->npages_free;
 792
 793         if (count > pool->npages_free) {
 794                 struct list_head d_pages;
 795
 796                 INIT_LIST_HEAD(&d_pages);
 797
 798                 spin_unlock_irqrestore(&pool->lock, *irq_flags);
 799
 800                 /* Returns how many more are neccessary to fulfill the
 801                  * request. */
 802                 r = ttm_dma_pool_alloc_new_pages(pool, &d_pages, count);
 803
 804                 spin_lock_irqsave(&pool->lock, *irq_flags);
 805                 if (!r) {
 806                         /* Add the fresh to the end.. */
 807                         list_splice(&d_pages, &pool->free_list);
 808                         ++pool->nrefills;
 809                         pool->npages_free += count;
 810                         r = count;
 811                 } else {
 812                         struct dma_page *d_page;
 813                         unsigned cpages = 0;
 814
 815                         pr_debug("%s: Failed to fill %s pool (r:%d)!\n",
 816                                  pool->dev_name, pool->name, r);
 817
 818                         list_for_each_entry(d_page, &d_pages, page_list) {
 819                                 cpages++;
 820                         }
 821                         list_splice_tail(&d_pages, &pool->free_list);
 822                         pool->npages_free += cpages;
 823                         r = cpages;
 824                 }
 825         }
 826         return r;
 827 }
 828
 829 /*
 830  * The populate list is actually a stack (not that is matters as TTM
 831  * allocates one page at a time.
 832  * return dma_page pointer if success, otherwise NULL.
 833  */
 834 static struct dma_page *ttm_dma_pool_get_pages(struct dma_pool *pool,
 835                                   struct ttm_dma_tt *ttm_dma,
 836                                   unsigned index)
 837 {
 838         struct dma_page *d_page = NULL;
 839         struct ttm_tt *ttm = &ttm_dma->ttm;
 840         unsigned long irq_flags;
 841         int count;
 842
 843         spin_lock_irqsave(&pool->lock, irq_flags);
 844         count = ttm_dma_page_pool_fill_locked(pool, &irq_flags);
 845         if (count) {
 846                 d_page = list_first_entry(&pool->free_list, struct dma_page, page_list);
 847                 ttm->pages[index] = d_page->p;
 848                 ttm_dma->dma_address[index] = d_page->dma;
 849                 list_move_tail(&d_page->page_list, &ttm_dma->pages_list);
 850                 pool->npages_in_use += 1;
 851                 pool->npages_free -= 1;
 852         }
 853         spin_unlock_irqrestore(&pool->lock, irq_flags);
 854         return d_page;
 855 }
 856
 857 static gfp_t ttm_dma_pool_gfp_flags(struct ttm_dma_tt *ttm_dma, bool huge)
 858 {
 859         struct ttm_tt *ttm = &ttm_dma->ttm;
 860         gfp_t gfp_flags;
 861
 862         if (ttm->page_flags & TTM_PAGE_FLAG_DMA32)
 863                 gfp_flags = GFP_USER | GFP_DMA32;
 864         else
 865                 gfp_flags = GFP_HIGHUSER;
 866         if (ttm->page_flags & TTM_PAGE_FLAG_ZERO_ALLOC)
 867                 gfp_flags |= __GFP_ZERO;
 868
 869         if (huge) {
 870                 gfp_flags |= GFP_TRANSHUGE_LIGHT | __GFP_NORETRY |
 871                         __GFP_KSWAPD_RECLAIM;
 872                 gfp_flags &= ~__GFP_MOVABLE;
 873                 gfp_flags &= ~__GFP_COMP;
 874         }
 875
 876         if (ttm->page_flags & TTM_PAGE_FLAG_NO_RETRY)
 877                 gfp_flags |= __GFP_RETRY_MAYFAIL;
 878
 879         return gfp_flags;
 880 }
 881
 882 /*
 883  * On success pages list will hold count number of correctly
 884  * cached pages. On failure will hold the negative return value (-ENOMEM, etc).
 885  */
 886 int ttm_dma_populate(struct ttm_dma_tt *ttm_dma, struct device *dev,
 887                         struct ttm_operation_ctx *ctx)
 888 {
 889         struct ttm_tt *ttm = &ttm_dma->ttm;
 890         struct ttm_mem_global *mem_glob = ttm->bdev->glob->mem_glob;
 891         unsigned long num_pages = ttm->num_pages;
 892         struct dma_pool *pool;
 893         struct dma_page *d_page;
 894         enum pool_type type;
 895         unsigned i;
 896         int ret;
 897
 898         if (ttm->state != tt_unpopulated)
 899                 return 0;
 900
 901         if (ttm_check_under_lowerlimit(mem_glob, num_pages, ctx))
 902                 return -ENOMEM;
 903
 904         INIT_LIST_HEAD(&ttm_dma->pages_list);
 905         i = 0;
 906
 907         type = ttm_to_type(ttm->page_flags, ttm->caching_state);
 908
 909 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 910         if (ttm->page_flags & TTM_PAGE_FLAG_DMA32)
 911                 goto skip_huge;
 912
 913         pool = ttm_dma_find_pool(dev, type | IS_HUGE);
 914         if (!pool) {
 915                 gfp_t gfp_flags = ttm_dma_pool_gfp_flags(ttm_dma, true);
 916
 917                 pool = ttm_dma_pool_init(dev, gfp_flags, type | IS_HUGE);
 918                 if (IS_ERR_OR_NULL(pool))
 919                         goto skip_huge;
 920         }
 921
 922         while (num_pages >= HPAGE_PMD_NR) {
 923                 unsigned j;
 924
 925                 d_page = ttm_dma_pool_get_pages(pool, ttm_dma, i);
 926                 if (!d_page)
 927                         break;
 928
 929                 ret = ttm_mem_global_alloc_page(mem_glob, ttm->pages[i],
 930                                                 pool->size, ctx);
 931                 if (unlikely(ret != 0)) {
 932                         ttm_dma_unpopulate(ttm_dma, dev);
 933                         return -ENOMEM;
 934                 }
 935
 936                 d_page->vaddr |= VADDR_FLAG_UPDATED_COUNT;
 937                 for (j = i + 1; j < (i + HPAGE_PMD_NR); ++j) {
 938                         ttm->pages[j] = ttm->pages[j - 1] + 1;
 939                         ttm_dma->dma_address[j] = ttm_dma->dma_address[j - 1] +
 940                                 PAGE_SIZE;
 941                 }
 942
 943                 i += HPAGE_PMD_NR;
 944                 num_pages -= HPAGE_PMD_NR;
 945         }
 946
 947 skip_huge:
 948 #endif
 949
 950         pool = ttm_dma_find_pool(dev, type);
 951         if (!pool) {
 952                 gfp_t gfp_flags = ttm_dma_pool_gfp_flags(ttm_dma, false);
 953
 954                 pool = ttm_dma_pool_init(dev, gfp_flags, type);
 955                 if (IS_ERR_OR_NULL(pool))
 956                         return -ENOMEM;
 957         }
 958
 959         while (num_pages) {
 960                 d_page = ttm_dma_pool_get_pages(pool, ttm_dma, i);
 961                 if (!d_page) {
 962                         ttm_dma_unpopulate(ttm_dma, dev);
 963                         return -ENOMEM;
 964                 }
 965
 966                 ret = ttm_mem_global_alloc_page(mem_glob, ttm->pages[i],
 967                                                 pool->size, ctx);
 968                 if (unlikely(ret != 0)) {
 969                         ttm_dma_unpopulate(ttm_dma, dev);
 970                         return -ENOMEM;
 971                 }
 972
 973                 d_page->vaddr |= VADDR_FLAG_UPDATED_COUNT;
 974                 ++i;
 975                 --num_pages;
 976         }
 977
 978         if (unlikely(ttm->page_flags & TTM_PAGE_FLAG_SWAPPED)) {
 979                 ret = ttm_tt_swapin(ttm);
 980                 if (unlikely(ret != 0)) {
 981                         ttm_dma_unpopulate(ttm_dma, dev);
 982                         return ret;
 983                 }
 984         }
 985
 986         ttm->state = tt_unbound;
 987         return 0;
 988 }
 989 EXPORT_SYMBOL_GPL(ttm_dma_populate);
 990
 991 /* Put all pages in pages list to correct pool to wait for reuse */
 992 void ttm_dma_unpopulate(struct ttm_dma_tt *ttm_dma, struct device *dev)
 993 {
 994         struct ttm_tt *ttm = &ttm_dma->ttm;
 995         struct ttm_mem_global *mem_glob = ttm->bdev->glob->mem_glob;
 996         struct dma_pool *pool;
 997         struct dma_page *d_page, *next;
 998         enum pool_type type;
 999         bool is_cached = false;
1000         unsigned count, i, npages = 0;
1001         unsigned long irq_flags;
1002
1003         type = ttm_to_type(ttm->page_flags, ttm->caching_state);
1004
1005 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1006         pool = ttm_dma_find_pool(dev, type | IS_HUGE);
1007         if (pool) {
1008                 count = 0;
1009                 list_for_each_entry_safe(d_page, next, &ttm_dma->pages_list,
1010                                          page_list) {
1011                         if (!(d_page->vaddr & VADDR_FLAG_HUGE_POOL))
1012                                 continue;
1013
1014                         count++;
1015                         if (d_page->vaddr & VADDR_FLAG_UPDATED_COUNT) {
1016                                 ttm_mem_global_free_page(mem_glob, d_page->p,
1017                                                          pool->size);
1018                                 d_page->vaddr &= ~VADDR_FLAG_UPDATED_COUNT;
1019                         }
1020                         ttm_dma_page_put(pool, d_page);
1021                 }
1022
1023                 spin_lock_irqsave(&pool->lock, irq_flags);
1024                 pool->npages_in_use -= count;
1025                 pool->nfrees += count;
1026                 spin_unlock_irqrestore(&pool->lock, irq_flags);
1027         }
1028 #endif
1029
1030         pool = ttm_dma_find_pool(dev, type);
1031         if (!pool)
1032                 return;
1033
1034         is_cached = (ttm_dma_find_pool(pool->dev,
1035                      ttm_to_type(ttm->page_flags, tt_cached)) == pool);
1036
1037         /* make sure pages array match list and count number of pages */
1038         count = 0;
1039         list_for_each_entry_safe(d_page, next, &ttm_dma->pages_list,
1040                                  page_list) {
1041                 ttm->pages[count] = d_page->p;
1042                 count++;
1043
1044                 if (d_page->vaddr & VADDR_FLAG_UPDATED_COUNT) {
1045                         ttm_mem_global_free_page(mem_glob, d_page->p,
1046                                                  pool->size);
1047                         d_page->vaddr &= ~VADDR_FLAG_UPDATED_COUNT;
1048                 }
1049
1050                 if (is_cached)
1051                         ttm_dma_page_put(pool, d_page);
1052         }
1053
1054         spin_lock_irqsave(&pool->lock, irq_flags);
1055         pool->npages_in_use -= count;
1056         if (is_cached) {
1057                 pool->nfrees += count;
1058         } else {
1059                 pool->npages_free += count;
1060                 list_splice(&ttm_dma->pages_list, &pool->free_list);
1061                 /*
1062                  * Wait to have at at least NUM_PAGES_TO_ALLOC number of pages
1063                  * to free in order to minimize calls to set_memory_wb().
1064                  */
1065                 if (pool->npages_free >= (_manager->options.max_size +
1066                                           NUM_PAGES_TO_ALLOC))
1067                         npages = pool->npages_free - _manager->options.max_size;
1068         }
1069         spin_unlock_irqrestore(&pool->lock, irq_flags);
1070
1071         INIT_LIST_HEAD(&ttm_dma->pages_list);
1072         for (i = 0; i < ttm->num_pages; i++) {
1073                 ttm->pages[i] = NULL;
1074                 ttm_dma->dma_address[i] = 0;
1075         }
1076
1077         /* shrink pool if necessary (only on !is_cached pools)*/
1078         if (npages)
1079                 ttm_dma_page_pool_free(pool, npages, false);
1080         ttm->state = tt_unpopulated;
1081 }
1082 EXPORT_SYMBOL_GPL(ttm_dma_unpopulate);
1083
1084 /**
1085  * Callback for mm to request pool to reduce number of page held.
1086  *
1087  * XXX: (dchinner) Deadlock warning!
1088  *
1089  * I'm getting sadder as I hear more pathetical whimpers about needing per-pool
1090  * shrinkers
1091  */
1092 static unsigned long
1093 ttm_dma_pool_shrink_scan(struct shrinker *shrink, struct shrink_control *sc)
1094 {
1095         static unsigned start_pool;
1096         unsigned idx = 0;
1097         unsigned pool_offset;
1098         unsigned shrink_pages = sc->nr_to_scan;
1099         struct device_pools *p;
1100         unsigned long freed = 0;
1101
1102         if (list_empty(&_manager->pools))
1103                 return SHRINK_STOP;
1104
1105         if (!mutex_trylock(&_manager->lock))
1106                 return SHRINK_STOP;
1107         if (!_manager->npools)
1108                 goto out;
1109         pool_offset = ++start_pool % _manager->npools;
1110         list_for_each_entry(p, &_manager->pools, pools) {
1111                 unsigned nr_free;
1112
1113                 if (!p->dev)
1114                         continue;
1115                 if (shrink_pages == 0)
1116                         break;
1117                 /* Do it in round-robin fashion. */
1118                 if (++idx < pool_offset)
1119                         continue;
1120                 nr_free = shrink_pages;
1121                 /* OK to use static buffer since global mutex is held. */
1122                 shrink_pages = ttm_dma_page_pool_free(p->pool, nr_free, true);
1123                 freed += nr_free - shrink_pages;
1124
1125                 pr_debug("%s: (%s:%d) Asked to shrink %d, have %d more to go\n",
1126                          p->pool->dev_name, p->pool->name, current->pid,
1127                          nr_free, shrink_pages);
1128         }
1129 out:
1130         mutex_unlock(&_manager->lock);
1131         return freed;
1132 }
1133
1134 static unsigned long
1135 ttm_dma_pool_shrink_count(struct shrinker *shrink, struct shrink_control *sc)
1136 {
1137         struct device_pools *p;
1138         unsigned long count = 0;
1139
1140         if (!mutex_trylock(&_manager->lock))
1141                 return 0;
1142         list_for_each_entry(p, &_manager->pools, pools)
1143                 count += p->pool->npages_free;
1144         mutex_unlock(&_manager->lock);
1145         return count;
1146 }
1147
1148 static int ttm_dma_pool_mm_shrink_init(struct ttm_pool_manager *manager)
1149 {
1150         manager->mm_shrink.count_objects = ttm_dma_pool_shrink_count;
1151         manager->mm_shrink.scan_objects = &ttm_dma_pool_shrink_scan;
1152         manager->mm_shrink.seeks = 1;
1153         return register_shrinker(&manager->mm_shrink);
1154 }
1155
1156 static void ttm_dma_pool_mm_shrink_fini(struct ttm_pool_manager *manager)
1157 {
1158         unregister_shrinker(&manager->mm_shrink);
1159 }
1160
1161 int ttm_dma_page_alloc_init(struct ttm_mem_global *glob, unsigned max_pages)
1162 {
1163         int ret;
1164
1165         WARN_ON(_manager);
1166
1167         pr_info("Initializing DMA pool allocator\n");
1168
1169         _manager = kzalloc(sizeof(*_manager), GFP_KERNEL);
1170         if (!_manager)
1171                 return -ENOMEM;
1172
1173         mutex_init(&_manager->lock);
1174         INIT_LIST_HEAD(&_manager->pools);
1175
1176         _manager->options.max_size = max_pages;
1177         _manager->options.small = SMALL_ALLOCATION;
1178         _manager->options.alloc_size = NUM_PAGES_TO_ALLOC;
1179
1180         /* This takes care of auto-freeing the _manager */
1181         ret = kobject_init_and_add(&_manager->kobj, &ttm_pool_kobj_type,
1182                                    &glob->kobj, "dma_pool");
1183         if (unlikely(ret != 0))
1184                 goto error;
1185
1186         ret = ttm_dma_pool_mm_shrink_init(_manager);
1187         if (unlikely(ret != 0))
1188                 goto error;
1189         return 0;
1190
1191 error:
1192         kobject_put(&_manager->kobj);
1193         _manager = NULL;
1194         return ret;
1195 }
1196
1197 void ttm_dma_page_alloc_fini(void)
1198 {
1199         struct device_pools *p, *t;
1200
1201         pr_info("Finalizing DMA pool allocator\n");
1202         ttm_dma_pool_mm_shrink_fini(_manager);
1203
1204         list_for_each_entry_safe_reverse(p, t, &_manager->pools, pools) {
1205                 dev_dbg(p->dev, "(%s:%d) Freeing.\n", p->pool->name,
1206                         current->pid);
1207                 WARN_ON(devres_destroy(p->dev, ttm_dma_pool_release,
1208                         ttm_dma_pool_match, p->pool));
1209                 ttm_dma_free_pool(p->dev, p->pool->type);
1210         }
1211         kobject_put(&_manager->kobj);
1212         _manager = NULL;
1213 }
1214
1215 int ttm_dma_page_alloc_debugfs(struct seq_file *m, void *data)
1216 {
1217         struct device_pools *p;
1218         struct dma_pool *pool = NULL;
1219
1220         if (!_manager) {
1221                 seq_printf(m, "No pool allocator running.\n");
1222                 return 0;
1223         }
1224         seq_printf(m, "         pool      refills   pages freed    inuse available     name\n");
1225         mutex_lock(&_manager->lock);
1226         list_for_each_entry(p, &_manager->pools, pools) {
1227                 struct device *dev = p->dev;
1228                 if (!dev)
1229                         continue;
1230                 pool = p->pool;
1231                 seq_printf(m, "%13s %12ld %13ld %8d %8d %8s\n",
1232                                 pool->name, pool->nrefills,
1233                                 pool->nfrees, pool->npages_in_use,
1234                                 pool->npages_free,
1235                                 pool->dev_name);
1236         }
1237         mutex_unlock(&_manager->lock);
1238         return 0;
1239 }
1240 EXPORT_SYMBOL_GPL(ttm_dma_page_alloc_debugfs);
1241
1242 #endif