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 <[email protected]>
  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 #if IS_ENABLED(CONFIG_AGP)
  54 #include <asm/agp.h>
  55 #endif
  56 #ifdef CONFIG_X86
  57 #include <asm/set_memory.h>
  58 #endif
  59
  60 #define NUM_PAGES_TO_ALLOC              (PAGE_SIZE/sizeof(struct page *))
  61 #define SMALL_ALLOCATION                4
  62 #define FREE_ALL_PAGES                  (~0U)
  63 #define VADDR_FLAG_HUGE_POOL            1UL
  64
  65 enum pool_type {
  66         IS_UNDEFINED    = 0,
  67         IS_WC           = 1 << 1,
  68         IS_UC           = 1 << 2,
  69         IS_CACHED       = 1 << 3,
  70         IS_DMA32        = 1 << 4,
  71         IS_HUGE         = 1 << 5
  72 };
  73
  74 /*
  75  * The pool structure. There are up to nine pools:
  76  *  - generic (not restricted to DMA32):
  77  *      - write combined, uncached, cached.
  78  *  - dma32 (up to 2^32 - so up 4GB):
  79  *      - write combined, uncached, cached.
  80  *  - huge (not restricted to DMA32):
  81  *      - write combined, uncached, cached.
  82  * for each 'struct device'. The 'cached' is for pages that are actively used.
  83  * The other ones can be shrunk by the shrinker API if neccessary.
  84  * @pools: The 'struct device->dma_pools' link.
  85  * @type: Type of the pool
  86  * @lock: Protects the free_list from concurrnet access. Must be
  87  * used with irqsave/irqrestore variants because pool allocator maybe called
  88  * from delayed work.
  89  * @free_list: Pool of pages that are free to be used. No order requirements.
  90  * @dev: The device that is associated with these pools.
  91  * @size: Size used during DMA allocation.
  92  * @npages_free: Count of available pages for re-use.
  93  * @npages_in_use: Count of pages that are in use.
  94  * @nfrees: Stats when pool is shrinking.
  95  * @nrefills: Stats when the pool is grown.
  96  * @gfp_flags: Flags to pass for alloc_page.
  97  * @name: Name of the pool.
  98  * @dev_name: Name derieved from dev - similar to how dev_info works.
  99  *   Used during shutdown as the dev_info during release is unavailable.
 100  */
 101 struct dma_pool {
 102         struct list_head pools; /* The 'struct device->dma_pools link */
 103         enum pool_type type;
 104         spinlock_t lock;
 105         struct list_head free_list;
 106         struct device *dev;
 107         unsigned size;
 108         unsigned npages_free;
 109         unsigned npages_in_use;
 110         unsigned long nfrees; /* Stats when shrunk. */
 111         unsigned long nrefills; /* Stats when grown. */
 112         gfp_t gfp_flags;
 113         char name[13]; /* "cached dma32" */
 114         char dev_name[64]; /* Constructed from dev */
 115 };
 116
 117 /*
 118  * The accounting page keeping track of the allocated page along with
 119  * the DMA address.
 120  * @page_list: The link to the 'page_list' in 'struct dma_pool'.
 121  * @vaddr: The virtual address of the page and a flag if the page belongs to a
 122  * huge pool
 123  * @dma: The bus address of the page. If the page is not allocated
 124  *   via the DMA API, it will be -1.
 125  */
 126 struct dma_page {
 127         struct list_head page_list;
 128         unsigned long vaddr;
 129         struct page *p;
 130         dma_addr_t dma;
 131 };
 132
 133 /*
 134  * Limits for the pool. They are handled without locks because only place where
 135  * they may change is in sysfs store. They won't have immediate effect anyway
 136  * so forcing serialization to access them is pointless.
 137  */
 138
 139 struct ttm_pool_opts {
 140         unsigned        alloc_size;
 141         unsigned        max_size;
 142         unsigned        small;
 143 };
 144
 145 /*
 146  * Contains the list of all of the 'struct device' and their corresponding
 147  * DMA pools. Guarded by _mutex->lock.
 148  * @pools: The link to 'struct ttm_pool_manager->pools'
 149  * @dev: The 'struct device' associated with the 'pool'
 150  * @pool: The 'struct dma_pool' associated with the 'dev'
 151  */
 152 struct device_pools {
 153         struct list_head pools;
 154         struct device *dev;
 155         struct dma_pool *pool;
 156 };
 157
 158 /*
 159  * struct ttm_pool_manager - Holds memory pools for fast allocation
 160  *
 161  * @lock: Lock used when adding/removing from pools
 162  * @pools: List of 'struct device' and 'struct dma_pool' tuples.
 163  * @options: Limits for the pool.
 164  * @npools: Total amount of pools in existence.
 165  * @shrinker: The structure used by [un|]register_shrinker
 166  */
 167 struct ttm_pool_manager {
 168         struct mutex            lock;
 169         struct list_head        pools;
 170         struct ttm_pool_opts    options;
 171         unsigned                npools;
 172         struct shrinker         mm_shrink;
 173         struct kobject          kobj;
 174 };
 175
 176 static struct ttm_pool_manager *_manager;
 177
 178 static struct attribute ttm_page_pool_max = {
 179         .name = "pool_max_size",
 180         .mode = S_IRUGO | S_IWUSR
 181 };
 182 static struct attribute ttm_page_pool_small = {
 183         .name = "pool_small_allocation",
 184         .mode = S_IRUGO | S_IWUSR
 185 };
 186 static struct attribute ttm_page_pool_alloc_size = {
 187         .name = "pool_allocation_size",
 188         .mode = S_IRUGO | S_IWUSR
 189 };
 190
 191 static struct attribute *ttm_pool_attrs[] = {
 192         &ttm_page_pool_max,
 193         &ttm_page_pool_small,
 194         &ttm_page_pool_alloc_size,
 195         NULL
 196 };
 197
 198 static void ttm_pool_kobj_release(struct kobject *kobj)
 199 {
 200         struct ttm_pool_manager *m =
 201                 container_of(kobj, struct ttm_pool_manager, kobj);
 202         kfree(m);
 203 }
 204
 205 static ssize_t ttm_pool_store(struct kobject *kobj, struct attribute *attr,
 206                               const char *buffer, size_t size)
 207 {
 208         struct ttm_pool_manager *m =
 209                 container_of(kobj, struct ttm_pool_manager, kobj);
 210         int chars;
 211         unsigned val;
 212         chars = sscanf(buffer, "%u", &val);
 213         if (chars == 0)
 214                 return size;
 215
 216         /* Convert kb to number of pages */
 217         val = val / (PAGE_SIZE >> 10);
 218
 219         if (attr == &ttm_page_pool_max)
 220                 m->options.max_size = val;
 221         else if (attr == &ttm_page_pool_small)
 222                 m->options.small = val;
 223         else if (attr == &ttm_page_pool_alloc_size) {
 224                 if (val > NUM_PAGES_TO_ALLOC*8) {
 225                         pr_err("Setting allocation size to %lu is not allowed. Recommended size is %lu\n",
 226                                NUM_PAGES_TO_ALLOC*(PAGE_SIZE >> 7),
 227                                NUM_PAGES_TO_ALLOC*(PAGE_SIZE >> 10));
 228                         return size;
 229                 } else if (val > NUM_PAGES_TO_ALLOC) {
 230                         pr_warn("Setting allocation size to larger than %lu is not recommended\n",
 231                                 NUM_PAGES_TO_ALLOC*(PAGE_SIZE >> 10));
 232                 }
 233                 m->options.alloc_size = val;
 234         }
 235
 236         return size;
 237 }
 238
 239 static ssize_t ttm_pool_show(struct kobject *kobj, struct attribute *attr,
 240                              char *buffer)
 241 {
 242         struct ttm_pool_manager *m =
 243                 container_of(kobj, struct ttm_pool_manager, kobj);
 244         unsigned val = 0;
 245
 246         if (attr == &ttm_page_pool_max)
 247                 val = m->options.max_size;
 248         else if (attr == &ttm_page_pool_small)
 249                 val = m->options.small;
 250         else if (attr == &ttm_page_pool_alloc_size)
 251                 val = m->options.alloc_size;
 252
 253         val = val * (PAGE_SIZE >> 10);
 254
 255         return snprintf(buffer, PAGE_SIZE, "%u\n", val);
 256 }
 257
 258 static const struct sysfs_ops ttm_pool_sysfs_ops = {
 259         .show = &ttm_pool_show,
 260         .store = &ttm_pool_store,
 261 };
 262
 263 static struct kobj_type ttm_pool_kobj_type = {
 264         .release = &ttm_pool_kobj_release,
 265         .sysfs_ops = &ttm_pool_sysfs_ops,
 266         .default_attrs = ttm_pool_attrs,
 267 };
 268
 269 #ifndef CONFIG_X86
 270 static int set_pages_array_wb(struct page **pages, int addrinarray)
 271 {
 272 #if IS_ENABLED(CONFIG_AGP)
 273         int i;
 274
 275         for (i = 0; i < addrinarray; i++)
 276                 unmap_page_from_agp(pages[i]);
 277 #endif
 278         return 0;
 279 }
 280
 281 static int set_pages_array_wc(struct page **pages, int addrinarray)
 282 {
 283 #if IS_ENABLED(CONFIG_AGP)
 284         int i;
 285
 286         for (i = 0; i < addrinarray; i++)
 287                 map_page_into_agp(pages[i]);
 288 #endif
 289         return 0;
 290 }
 291
 292 static int set_pages_array_uc(struct page **pages, int addrinarray)
 293 {
 294 #if IS_ENABLED(CONFIG_AGP)
 295         int i;
 296
 297         for (i = 0; i < addrinarray; i++)
 298                 map_page_into_agp(pages[i]);
 299 #endif
 300         return 0;
 301 }
 302 #endif /* for !CONFIG_X86 */
 303
 304 static int ttm_set_pages_caching(struct dma_pool *pool,
 305                                  struct page **pages, unsigned cpages)
 306 {
 307         int r = 0;
 308         /* Set page caching */
 309         if (pool->type & IS_UC) {
 310                 r = set_pages_array_uc(pages, cpages);
 311                 if (r)
 312                         pr_err("%s: Failed to set %d pages to uc!\n",
 313                                pool->dev_name, cpages);
 314         }
 315         if (pool->type & IS_WC) {
 316                 r = set_pages_array_wc(pages, cpages);
 317                 if (r)
 318                         pr_err("%s: Failed to set %d pages to wc!\n",
 319                                pool->dev_name, cpages);
 320         }
 321         return r;
 322 }
 323
 324 static void __ttm_dma_free_page(struct dma_pool *pool, struct dma_page *d_page)
 325 {
 326         dma_addr_t dma = d_page->dma;
 327         d_page->vaddr &= ~VADDR_FLAG_HUGE_POOL;
 328         dma_free_coherent(pool->dev, pool->size, (void *)d_page->vaddr, dma);
 329
 330         kfree(d_page);
 331         d_page = NULL;
 332 }
 333 static struct dma_page *__ttm_dma_alloc_page(struct dma_pool *pool)
 334 {
 335         struct dma_page *d_page;
 336         void *vaddr;
 337
 338         d_page = kmalloc(sizeof(struct dma_page), GFP_KERNEL);
 339         if (!d_page)
 340                 return NULL;
 341
 342         vaddr = dma_alloc_coherent(pool->dev, pool->size, &d_page->dma,
 343                                    pool->gfp_flags);
 344         if (vaddr) {
 345                 if (is_vmalloc_addr(vaddr))
 346                         d_page->p = vmalloc_to_page(vaddr);
 347                 else
 348                         d_page->p = virt_to_page(vaddr);
 349                 d_page->vaddr = (unsigned long)vaddr;
 350                 if (pool->type & IS_HUGE)
 351                         d_page->vaddr |= VADDR_FLAG_HUGE_POOL;
 352         } else {
 353                 kfree(d_page);
 354                 d_page = NULL;
 355         }
 356         return d_page;
 357 }
 358 static enum pool_type ttm_to_type(int flags, enum ttm_caching_state cstate)
 359 {
 360         enum pool_type type = IS_UNDEFINED;
 361
 362         if (flags & TTM_PAGE_FLAG_DMA32)
 363                 type |= IS_DMA32;
 364         if (cstate == tt_cached)
 365                 type |= IS_CACHED;
 366         else if (cstate == tt_uncached)
 367                 type |= IS_UC;
 368         else
 369                 type |= IS_WC;
 370
 371         return type;
 372 }
 373
 374 static void ttm_pool_update_free_locked(struct dma_pool *pool,
 375                                         unsigned freed_pages)
 376 {
 377         pool->npages_free -= freed_pages;
 378         pool->nfrees += freed_pages;
 379
 380 }
 381
 382 /* set memory back to wb and free the pages. */
 383 static void ttm_dma_page_put(struct dma_pool *pool, struct dma_page *d_page)
 384 {
 385         struct page *page = d_page->p;
 386         unsigned i, num_pages;
 387         int ret;
 388
 389         /* Don't set WB on WB page pool. */
 390         if (!(pool->type & IS_CACHED)) {
 391                 num_pages = pool->size / PAGE_SIZE;
 392                 for (i = 0; i < num_pages; ++i, ++page) {
 393                         ret = set_pages_array_wb(&page, 1);
 394                         if (ret) {
 395                                 pr_err("%s: Failed to set %d pages to wb!\n",
 396                                        pool->dev_name, 1);
 397                         }
 398                 }
 399         }
 400
 401         list_del(&d_page->page_list);
 402         __ttm_dma_free_page(pool, d_page);
 403 }
 404
 405 static void ttm_dma_pages_put(struct dma_pool *pool, struct list_head *d_pages,
 406                               struct page *pages[], unsigned npages)
 407 {
 408         struct dma_page *d_page, *tmp;
 409
 410         if (pool->type & IS_HUGE) {
 411                 list_for_each_entry_safe(d_page, tmp, d_pages, page_list)
 412                         ttm_dma_page_put(pool, d_page);
 413
 414                 return;
 415         }
 416
 417         /* Don't set WB on WB page pool. */
 418         if (npages && !(pool->type & IS_CACHED) &&
 419             set_pages_array_wb(pages, npages))
 420                 pr_err("%s: Failed to set %d pages to wb!\n",
 421                        pool->dev_name, npages);
 422
 423         list_for_each_entry_safe(d_page, tmp, d_pages, page_list) {
 424                 list_del(&d_page->page_list);
 425                 __ttm_dma_free_page(pool, d_page);
 426         }
 427 }
 428
 429 /*
 430  * Free pages from pool.
 431  *
 432  * To prevent hogging the ttm_swap process we only free NUM_PAGES_TO_ALLOC
 433  * number of pages in one go.
 434  *
 435  * @pool: to free the pages from
 436  * @nr_free: If set to true will free all pages in pool
 437  * @use_static: Safe to use static buffer
 438  **/
 439 static unsigned ttm_dma_page_pool_free(struct dma_pool *pool, unsigned nr_free,
 440                                        bool use_static)
 441 {
 442         static struct page *static_buf[NUM_PAGES_TO_ALLOC];
 443         unsigned long irq_flags;
 444         struct dma_page *dma_p, *tmp;
 445         struct page **pages_to_free;
 446         struct list_head d_pages;
 447         unsigned freed_pages = 0,
 448                  npages_to_free = nr_free;
 449
 450         if (NUM_PAGES_TO_ALLOC < nr_free)
 451                 npages_to_free = NUM_PAGES_TO_ALLOC;
 452 #if 0
 453         if (nr_free > 1) {
 454                 pr_debug("%s: (%s:%d) Attempting to free %d (%d) pages\n",
 455                          pool->dev_name, pool->name, current->pid,
 456                          npages_to_free, nr_free);
 457         }
 458 #endif
 459         if (use_static)
 460                 pages_to_free = static_buf;
 461         else
 462                 pages_to_free = kmalloc(npages_to_free * sizeof(struct page *),
 463                                         GFP_KERNEL);
 464
 465         if (!pages_to_free) {
 466                 pr_debug("%s: Failed to allocate memory for pool free operation\n",
 467                        pool->dev_name);
 468                 return 0;
 469         }
 470         INIT_LIST_HEAD(&d_pages);
 471 restart:
 472         spin_lock_irqsave(&pool->lock, irq_flags);
 473
 474         /* We picking the oldest ones off the list */
 475         list_for_each_entry_safe_reverse(dma_p, tmp, &pool->free_list,
 476                                          page_list) {
 477                 if (freed_pages >= npages_to_free)
 478                         break;
 479
 480                 /* Move the dma_page from one list to another. */
 481                 list_move(&dma_p->page_list, &d_pages);
 482
 483                 pages_to_free[freed_pages++] = dma_p->p;
 484                 /* We can only remove NUM_PAGES_TO_ALLOC at a time. */
 485                 if (freed_pages >= NUM_PAGES_TO_ALLOC) {
 486
 487                         ttm_pool_update_free_locked(pool, freed_pages);
 488                         /**
 489                          * Because changing page caching is costly
 490                          * we unlock the pool to prevent stalling.
 491                          */
 492                         spin_unlock_irqrestore(&pool->lock, irq_flags);
 493
 494                         ttm_dma_pages_put(pool, &d_pages, pages_to_free,
 495                                           freed_pages);
 496
 497                         INIT_LIST_HEAD(&d_pages);
 498
 499                         if (likely(nr_free != FREE_ALL_PAGES))
 500                                 nr_free -= freed_pages;
 501
 502                         if (NUM_PAGES_TO_ALLOC >= nr_free)
 503                                 npages_to_free = nr_free;
 504                         else
 505                                 npages_to_free = NUM_PAGES_TO_ALLOC;
 506
 507                         freed_pages = 0;
 508
 509                         /* free all so restart the processing */
 510                         if (nr_free)
 511                                 goto restart;
 512
 513                         /* Not allowed to fall through or break because
 514                          * following context is inside spinlock while we are
 515                          * outside here.
 516                          */
 517                         goto out;
 518
 519                 }
 520         }
 521
 522         /* remove range of pages from the pool */
 523         if (freed_pages) {
 524                 ttm_pool_update_free_locked(pool, freed_pages);
 525                 nr_free -= freed_pages;
 526         }
 527
 528         spin_unlock_irqrestore(&pool->lock, irq_flags);
 529
 530         if (freed_pages)
 531                 ttm_dma_pages_put(pool, &d_pages, pages_to_free, freed_pages);
 532 out:
 533         if (pages_to_free != static_buf)
 534                 kfree(pages_to_free);
 535         return nr_free;
 536 }
 537
 538 static void ttm_dma_free_pool(struct device *dev, enum pool_type type)
 539 {
 540         struct device_pools *p;
 541         struct dma_pool *pool;
 542
 543         if (!dev)
 544                 return;
 545
 546         mutex_lock(&_manager->lock);
 547         list_for_each_entry_reverse(p, &_manager->pools, pools) {
 548                 if (p->dev != dev)
 549                         continue;
 550                 pool = p->pool;
 551                 if (pool->type != type)
 552                         continue;
 553
 554                 list_del(&p->pools);
 555                 kfree(p);
 556                 _manager->npools--;
 557                 break;
 558         }
 559         list_for_each_entry_reverse(pool, &dev->dma_pools, pools) {
 560                 if (pool->type != type)
 561                         continue;
 562                 /* Takes a spinlock.. */
 563                 /* OK to use static buffer since global mutex is held. */
 564                 ttm_dma_page_pool_free(pool, FREE_ALL_PAGES, true);
 565                 WARN_ON(((pool->npages_in_use + pool->npages_free) != 0));
 566                 /* This code path is called after _all_ references to the
 567                  * struct device has been dropped - so nobody should be
 568                  * touching it. In case somebody is trying to _add_ we are
 569                  * guarded by the mutex. */
 570                 list_del(&pool->pools);
 571                 kfree(pool);
 572                 break;
 573         }
 574         mutex_unlock(&_manager->lock);
 575 }
 576
 577 /*
 578  * On free-ing of the 'struct device' this deconstructor is run.
 579  * Albeit the pool might have already been freed earlier.
 580  */
 581 static void ttm_dma_pool_release(struct device *dev, void *res)
 582 {
 583         struct dma_pool *pool = *(struct dma_pool **)res;
 584
 585         if (pool)
 586                 ttm_dma_free_pool(dev, pool->type);
 587 }
 588
 589 static int ttm_dma_pool_match(struct device *dev, void *res, void *match_data)
 590 {
 591         return *(struct dma_pool **)res == match_data;
 592 }
 593
 594 static struct dma_pool *ttm_dma_pool_init(struct device *dev, gfp_t flags,
 595                                           enum pool_type type)
 596 {
 597         const char *n[] = {"wc", "uc", "cached", " dma32", "huge"};
 598         enum pool_type t[] = {IS_WC, IS_UC, IS_CACHED, IS_DMA32, IS_HUGE};
 599         struct device_pools *sec_pool = NULL;
 600         struct dma_pool *pool = NULL, **ptr;
 601         unsigned i;
 602         int ret = -ENODEV;
 603         char *p;
 604
 605         if (!dev)
 606                 return NULL;
 607
 608         ptr = devres_alloc(ttm_dma_pool_release, sizeof(*ptr), GFP_KERNEL);
 609         if (!ptr)
 610                 return NULL;
 611
 612         ret = -ENOMEM;
 613
 614         pool = kmalloc_node(sizeof(struct dma_pool), GFP_KERNEL,
 615                             dev_to_node(dev));
 616         if (!pool)
 617                 goto err_mem;
 618
 619         sec_pool = kmalloc_node(sizeof(struct device_pools), GFP_KERNEL,
 620                                 dev_to_node(dev));
 621         if (!sec_pool)
 622                 goto err_mem;
 623
 624         INIT_LIST_HEAD(&sec_pool->pools);
 625         sec_pool->dev = dev;
 626         sec_pool->pool =  pool;
 627
 628         INIT_LIST_HEAD(&pool->free_list);
 629         INIT_LIST_HEAD(&pool->pools);
 630         spin_lock_init(&pool->lock);
 631         pool->dev = dev;
 632         pool->npages_free = pool->npages_in_use = 0;
 633         pool->nfrees = 0;
 634         pool->gfp_flags = flags;
 635         if (type & IS_HUGE)
 636 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 637                 pool->size = HPAGE_PMD_SIZE;
 638 #else
 639                 BUG();
 640 #endif
 641         else
 642                 pool->size = PAGE_SIZE;
 643         pool->type = type;
 644         pool->nrefills = 0;
 645         p = pool->name;
 646         for (i = 0; i < ARRAY_SIZE(t); i++) {
 647                 if (type & t[i]) {
 648                         p += snprintf(p, sizeof(pool->name) - (p - pool->name),
 649                                       "%s", n[i]);
 650                 }
 651         }
 652         *p = 0;
 653         /* We copy the name for pr_ calls b/c when dma_pool_destroy is called
 654          * - the kobj->name has already been deallocated.*/
 655         snprintf(pool->dev_name, sizeof(pool->dev_name), "%s %s",
 656                  dev_driver_string(dev), dev_name(dev));
 657         mutex_lock(&_manager->lock);
 658         /* You can get the dma_pool from either the global: */
 659         list_add(&sec_pool->pools, &_manager->pools);
 660         _manager->npools++;
 661         /* or from 'struct device': */
 662         list_add(&pool->pools, &dev->dma_pools);
 663         mutex_unlock(&_manager->lock);
 664
 665         *ptr = pool;
 666         devres_add(dev, ptr);
 667
 668         return pool;
 669 err_mem:
 670         devres_free(ptr);
 671         kfree(sec_pool);
 672         kfree(pool);
 673         return ERR_PTR(ret);
 674 }
 675
 676 static struct dma_pool *ttm_dma_find_pool(struct device *dev,
 677                                           enum pool_type type)
 678 {
 679         struct dma_pool *pool, *tmp, *found = NULL;
 680
 681         if (type == IS_UNDEFINED)
 682                 return found;
 683
 684         /* NB: We iterate on the 'struct dev' which has no spinlock, but
 685          * it does have a kref which we have taken. The kref is taken during
 686          * graphic driver loading - in the drm_pci_init it calls either
 687          * pci_dev_get or pci_register_driver which both end up taking a kref
 688          * on 'struct device'.
 689          *
 690          * On teardown, the graphic drivers end up quiescing the TTM (put_pages)
 691          * and calls the dev_res deconstructors: ttm_dma_pool_release. The nice
 692          * thing is at that point of time there are no pages associated with the
 693          * driver so this function will not be called.
 694          */
 695         list_for_each_entry_safe(pool, tmp, &dev->dma_pools, pools) {
 696                 if (pool->type != type)
 697                         continue;
 698                 found = pool;
 699                 break;
 700         }
 701         return found;
 702 }
 703
 704 /*
 705  * Free pages the pages that failed to change the caching state. If there
 706  * are pages that have changed their caching state already put them to the
 707  * pool.
 708  */
 709 static void ttm_dma_handle_caching_state_failure(struct dma_pool *pool,
 710                                                  struct list_head *d_pages,
 711                                                  struct page **failed_pages,
 712                                                  unsigned cpages)
 713 {
 714         struct dma_page *d_page, *tmp;
 715         struct page *p;
 716         unsigned i = 0;
 717
 718         p = failed_pages[0];
 719         if (!p)
 720                 return;
 721         /* Find the failed page. */
 722         list_for_each_entry_safe(d_page, tmp, d_pages, page_list) {
 723                 if (d_page->p != p)
 724                         continue;
 725                 /* .. and then progress over the full list. */
 726                 list_del(&d_page->page_list);
 727                 __ttm_dma_free_page(pool, d_page);
 728                 if (++i < cpages)
 729                         p = failed_pages[i];
 730                 else
 731                         break;
 732         }
 733
 734 }
 735
 736 /*
 737  * Allocate 'count' pages, and put 'need' number of them on the
 738  * 'pages' and as well on the 'dma_address' starting at 'dma_offset' offset.
 739  * The full list of pages should also be on 'd_pages'.
 740  * We return zero for success, and negative numbers as errors.
 741  */
 742 static int ttm_dma_pool_alloc_new_pages(struct dma_pool *pool,
 743                                         struct list_head *d_pages,
 744                                         unsigned count)
 745 {
 746         struct page **caching_array;
 747         struct dma_page *dma_p;
 748         struct page *p;
 749         int r = 0;
 750         unsigned i, j, npages, cpages;
 751         unsigned max_cpages = min(count,
 752                         (unsigned)(PAGE_SIZE/sizeof(struct page *)));
 753
 754         /* allocate array for page caching change */
 755         caching_array = kmalloc(max_cpages*sizeof(struct page *), GFP_KERNEL);
 756
 757         if (!caching_array) {
 758                 pr_debug("%s: Unable to allocate table for new pages\n",
 759                        pool->dev_name);
 760                 return -ENOMEM;
 761         }
 762
 763         if (count > 1) {
 764                 pr_debug("%s: (%s:%d) Getting %d pages\n",
 765                          pool->dev_name, pool->name, current->pid, count);
 766         }
 767
 768         for (i = 0, cpages = 0; i < count; ++i) {
 769                 dma_p = __ttm_dma_alloc_page(pool);
 770                 if (!dma_p) {
 771                         pr_debug("%s: Unable to get page %u\n",
 772                                  pool->dev_name, i);
 773
 774                         /* store already allocated pages in the pool after
 775                          * setting the caching state */
 776                         if (cpages) {
 777                                 r = ttm_set_pages_caching(pool, caching_array,
 778                                                           cpages);
 779                                 if (r)
 780                                         ttm_dma_handle_caching_state_failure(
 781                                                 pool, d_pages, caching_array,
 782                                                 cpages);
 783                         }
 784                         r = -ENOMEM;
 785                         goto out;
 786                 }
 787                 p = dma_p->p;
 788                 list_add(&dma_p->page_list, d_pages);
 789
 790 #ifdef CONFIG_HIGHMEM
 791                 /* gfp flags of highmem page should never be dma32 so we
 792                  * we should be fine in such case
 793                  */
 794                 if (PageHighMem(p))
 795                         continue;
 796 #endif
 797
 798                 npages = pool->size / PAGE_SIZE;
 799                 for (j = 0; j < npages; ++j) {
 800                         caching_array[cpages++] = p + j;
 801                         if (cpages == max_cpages) {
 802                                 /* Note: Cannot hold the spinlock */
 803                                 r = ttm_set_pages_caching(pool, caching_array,
 804                                                           cpages);
 805                                 if (r) {
 806                                         ttm_dma_handle_caching_state_failure(
 807                                              pool, d_pages, caching_array,
 808                                              cpages);
 809                                         goto out;
 810                                 }
 811                                 cpages = 0;
 812                         }
 813                 }
 814         }
 815
 816         if (cpages) {
 817                 r = ttm_set_pages_caching(pool, caching_array, cpages);
 818                 if (r)
 819                         ttm_dma_handle_caching_state_failure(pool, d_pages,
 820                                         caching_array, cpages);
 821         }
 822 out:
 823         kfree(caching_array);
 824         return r;
 825 }
 826
 827 /*
 828  * @return count of pages still required to fulfill the request.
 829  */
 830 static int ttm_dma_page_pool_fill_locked(struct dma_pool *pool,
 831                                          unsigned long *irq_flags)
 832 {
 833         unsigned count = _manager->options.small;
 834         int r = pool->npages_free;
 835
 836         if (count > pool->npages_free) {
 837                 struct list_head d_pages;
 838
 839                 INIT_LIST_HEAD(&d_pages);
 840
 841                 spin_unlock_irqrestore(&pool->lock, *irq_flags);
 842
 843                 /* Returns how many more are neccessary to fulfill the
 844                  * request. */
 845                 r = ttm_dma_pool_alloc_new_pages(pool, &d_pages, count);
 846
 847                 spin_lock_irqsave(&pool->lock, *irq_flags);
 848                 if (!r) {
 849                         /* Add the fresh to the end.. */
 850                         list_splice(&d_pages, &pool->free_list);
 851                         ++pool->nrefills;
 852                         pool->npages_free += count;
 853                         r = count;
 854                 } else {
 855                         struct dma_page *d_page;
 856                         unsigned cpages = 0;
 857
 858                         pr_debug("%s: Failed to fill %s pool (r:%d)!\n",
 859                                  pool->dev_name, pool->name, r);
 860
 861                         list_for_each_entry(d_page, &d_pages, page_list) {
 862                                 cpages++;
 863                         }
 864                         list_splice_tail(&d_pages, &pool->free_list);
 865                         pool->npages_free += cpages;
 866                         r = cpages;
 867                 }
 868         }
 869         return r;
 870 }
 871
 872 /*
 873  * @return count of pages still required to fulfill the request.
 874  * The populate list is actually a stack (not that is matters as TTM
 875  * allocates one page at a time.
 876  */
 877 static int ttm_dma_pool_get_pages(struct dma_pool *pool,
 878                                   struct ttm_dma_tt *ttm_dma,
 879                                   unsigned index)
 880 {
 881         struct dma_page *d_page;
 882         struct ttm_tt *ttm = &ttm_dma->ttm;
 883         unsigned long irq_flags;
 884         int count, r = -ENOMEM;
 885
 886         spin_lock_irqsave(&pool->lock, irq_flags);
 887         count = ttm_dma_page_pool_fill_locked(pool, &irq_flags);
 888         if (count) {
 889                 d_page = list_first_entry(&pool->free_list, struct dma_page, page_list);
 890                 ttm->pages[index] = d_page->p;
 891                 ttm_dma->dma_address[index] = d_page->dma;
 892                 list_move_tail(&d_page->page_list, &ttm_dma->pages_list);
 893                 r = 0;
 894                 pool->npages_in_use += 1;
 895                 pool->npages_free -= 1;
 896         }
 897         spin_unlock_irqrestore(&pool->lock, irq_flags);
 898         return r;
 899 }
 900
 901 static gfp_t ttm_dma_pool_gfp_flags(struct ttm_dma_tt *ttm_dma, bool huge)
 902 {
 903         struct ttm_tt *ttm = &ttm_dma->ttm;
 904         gfp_t gfp_flags;
 905
 906         if (ttm->page_flags & TTM_PAGE_FLAG_DMA32)
 907                 gfp_flags = GFP_USER | GFP_DMA32;
 908         else
 909                 gfp_flags = GFP_HIGHUSER;
 910         if (ttm->page_flags & TTM_PAGE_FLAG_ZERO_ALLOC)
 911                 gfp_flags |= __GFP_ZERO;
 912
 913         if (huge) {
 914                 gfp_flags |= GFP_TRANSHUGE;
 915                 gfp_flags &= ~__GFP_MOVABLE;
 916                 gfp_flags &= ~__GFP_COMP;
 917         }
 918
 919         return gfp_flags;
 920 }
 921
 922 /*
 923  * On success pages list will hold count number of correctly
 924  * cached pages. On failure will hold the negative return value (-ENOMEM, etc).
 925  */
 926 int ttm_dma_populate(struct ttm_dma_tt *ttm_dma, struct device *dev)
 927 {
 928         struct ttm_tt *ttm = &ttm_dma->ttm;
 929         struct ttm_mem_global *mem_glob = ttm->glob->mem_glob;
 930         unsigned long num_pages = ttm->num_pages;
 931         struct dma_pool *pool;
 932         enum pool_type type;
 933         unsigned i;
 934         int ret;
 935
 936         if (ttm->state != tt_unpopulated)
 937                 return 0;
 938
 939         INIT_LIST_HEAD(&ttm_dma->pages_list);
 940         i = 0;
 941
 942         type = ttm_to_type(ttm->page_flags, ttm->caching_state);
 943
 944 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 945         if (ttm->page_flags & TTM_PAGE_FLAG_DMA32)
 946                 goto skip_huge;
 947
 948         pool = ttm_dma_find_pool(dev, type | IS_HUGE);
 949         if (!pool) {
 950                 gfp_t gfp_flags = ttm_dma_pool_gfp_flags(ttm_dma, true);
 951
 952                 pool = ttm_dma_pool_init(dev, gfp_flags, type | IS_HUGE);
 953                 if (IS_ERR_OR_NULL(pool))
 954                         goto skip_huge;
 955         }
 956
 957         while (num_pages >= HPAGE_PMD_NR) {
 958                 unsigned j;
 959
 960                 ret = ttm_dma_pool_get_pages(pool, ttm_dma, i);
 961                 if (ret != 0)
 962                         break;
 963
 964                 ret = ttm_mem_global_alloc_page(mem_glob, ttm->pages[i],
 965                                                 pool->size);
 966                 if (unlikely(ret != 0)) {
 967                         ttm_dma_unpopulate(ttm_dma, dev);
 968                         return -ENOMEM;
 969                 }
 970
 971                 for (j = i + 1; j < (i + HPAGE_PMD_NR); ++j) {
 972                         ttm->pages[j] = ttm->pages[j - 1] + 1;
 973                         ttm_dma->dma_address[j] = ttm_dma->dma_address[j - 1] +
 974                                 PAGE_SIZE;
 975                 }
 976
 977                 i += HPAGE_PMD_NR;
 978                 num_pages -= HPAGE_PMD_NR;
 979         }
 980
 981 skip_huge:
 982 #endif
 983
 984         pool = ttm_dma_find_pool(dev, type);
 985         if (!pool) {
 986                 gfp_t gfp_flags = ttm_dma_pool_gfp_flags(ttm_dma, false);
 987
 988                 pool = ttm_dma_pool_init(dev, gfp_flags, type);
 989                 if (IS_ERR_OR_NULL(pool))
 990                         return -ENOMEM;
 991         }
 992
 993         while (num_pages) {
 994                 ret = ttm_dma_pool_get_pages(pool, ttm_dma, i);
 995                 if (ret != 0) {
 996                         ttm_dma_unpopulate(ttm_dma, dev);
 997                         return -ENOMEM;
 998                 }
 999
1000                 ret = ttm_mem_global_alloc_page(mem_glob, ttm->pages[i],
1001                                                 pool->size);
1002                 if (unlikely(ret != 0)) {
1003                         ttm_dma_unpopulate(ttm_dma, dev);
1004                         return -ENOMEM;
1005                 }
1006
1007                 ++i;
1008                 --num_pages;
1009         }
1010
1011         if (unlikely(ttm->page_flags & TTM_PAGE_FLAG_SWAPPED)) {
1012                 ret = ttm_tt_swapin(ttm);
1013                 if (unlikely(ret != 0)) {
1014                         ttm_dma_unpopulate(ttm_dma, dev);
1015                         return ret;
1016                 }
1017         }
1018
1019         ttm->state = tt_unbound;
1020         return 0;
1021 }
1022 EXPORT_SYMBOL_GPL(ttm_dma_populate);
1023
1024 /* Put all pages in pages list to correct pool to wait for reuse */
1025 void ttm_dma_unpopulate(struct ttm_dma_tt *ttm_dma, struct device *dev)
1026 {
1027         struct ttm_tt *ttm = &ttm_dma->ttm;
1028         struct dma_pool *pool;
1029         struct dma_page *d_page, *next;
1030         enum pool_type type;
1031         bool is_cached = false;
1032         unsigned count, i, npages = 0;
1033         unsigned long irq_flags;
1034
1035         type = ttm_to_type(ttm->page_flags, ttm->caching_state);
1036
1037 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1038         pool = ttm_dma_find_pool(dev, type | IS_HUGE);
1039         if (pool) {
1040                 count = 0;
1041                 list_for_each_entry_safe(d_page, next, &ttm_dma->pages_list,
1042                                          page_list) {
1043                         if (!(d_page->vaddr & VADDR_FLAG_HUGE_POOL))
1044                                 continue;
1045
1046                         count++;
1047                         ttm_mem_global_free_page(ttm->glob->mem_glob,
1048                                                  d_page->p, pool->size);
1049                         ttm_dma_page_put(pool, d_page);
1050                 }
1051
1052                 spin_lock_irqsave(&pool->lock, irq_flags);
1053                 pool->npages_in_use -= count;
1054                 pool->nfrees += count;
1055                 spin_unlock_irqrestore(&pool->lock, irq_flags);
1056         }
1057 #endif
1058
1059         pool = ttm_dma_find_pool(dev, type);
1060         if (!pool)
1061                 return;
1062
1063         is_cached = (ttm_dma_find_pool(pool->dev,
1064                      ttm_to_type(ttm->page_flags, tt_cached)) == pool);
1065
1066         /* make sure pages array match list and count number of pages */
1067         count = 0;
1068         list_for_each_entry(d_page, &ttm_dma->pages_list, page_list) {
1069                 ttm->pages[count] = d_page->p;
1070                 count++;
1071         }
1072
1073         spin_lock_irqsave(&pool->lock, irq_flags);
1074         pool->npages_in_use -= count;
1075         if (is_cached) {
1076                 pool->nfrees += count;
1077         } else {
1078                 pool->npages_free += count;
1079                 list_splice(&ttm_dma->pages_list, &pool->free_list);
1080                 /*
1081                  * Wait to have at at least NUM_PAGES_TO_ALLOC number of pages
1082                  * to free in order to minimize calls to set_memory_wb().
1083                  */
1084                 if (pool->npages_free >= (_manager->options.max_size +
1085                                           NUM_PAGES_TO_ALLOC))
1086                         npages = pool->npages_free - _manager->options.max_size;
1087         }
1088         spin_unlock_irqrestore(&pool->lock, irq_flags);
1089
1090         if (is_cached) {
1091                 list_for_each_entry_safe(d_page, next, &ttm_dma->pages_list, page_list) {
1092                         ttm_mem_global_free_page(ttm->glob->mem_glob,
1093                                                  d_page->p, pool->size);
1094                         ttm_dma_page_put(pool, d_page);
1095                 }
1096         } else {
1097                 for (i = 0; i < count; i++) {
1098                         ttm_mem_global_free_page(ttm->glob->mem_glob,
1099                                                  ttm->pages[i], pool->size);
1100                 }
1101         }
1102
1103         INIT_LIST_HEAD(&ttm_dma->pages_list);
1104         for (i = 0; i < ttm->num_pages; i++) {
1105                 ttm->pages[i] = NULL;
1106                 ttm_dma->dma_address[i] = 0;
1107         }
1108
1109         /* shrink pool if necessary (only on !is_cached pools)*/
1110         if (npages)
1111                 ttm_dma_page_pool_free(pool, npages, false);
1112         ttm->state = tt_unpopulated;
1113 }
1114 EXPORT_SYMBOL_GPL(ttm_dma_unpopulate);
1115
1116 /**
1117  * Callback for mm to request pool to reduce number of page held.
1118  *
1119  * XXX: (dchinner) Deadlock warning!
1120  *
1121  * I'm getting sadder as I hear more pathetical whimpers about needing per-pool
1122  * shrinkers
1123  */
1124 static unsigned long
1125 ttm_dma_pool_shrink_scan(struct shrinker *shrink, struct shrink_control *sc)
1126 {
1127         static unsigned start_pool;
1128         unsigned idx = 0;
1129         unsigned pool_offset;
1130         unsigned shrink_pages = sc->nr_to_scan;
1131         struct device_pools *p;
1132         unsigned long freed = 0;
1133
1134         if (list_empty(&_manager->pools))
1135                 return SHRINK_STOP;
1136
1137         if (!mutex_trylock(&_manager->lock))
1138                 return SHRINK_STOP;
1139         if (!_manager->npools)
1140                 goto out;
1141         pool_offset = ++start_pool % _manager->npools;
1142         list_for_each_entry(p, &_manager->pools, pools) {
1143                 unsigned nr_free;
1144
1145                 if (!p->dev)
1146                         continue;
1147                 if (shrink_pages == 0)
1148                         break;
1149                 /* Do it in round-robin fashion. */
1150                 if (++idx < pool_offset)
1151                         continue;
1152                 nr_free = shrink_pages;
1153                 /* OK to use static buffer since global mutex is held. */
1154                 shrink_pages = ttm_dma_page_pool_free(p->pool, nr_free, true);
1155                 freed += nr_free - shrink_pages;
1156
1157                 pr_debug("%s: (%s:%d) Asked to shrink %d, have %d more to go\n",
1158                          p->pool->dev_name, p->pool->name, current->pid,
1159                          nr_free, shrink_pages);
1160         }
1161 out:
1162         mutex_unlock(&_manager->lock);
1163         return freed;
1164 }
1165
1166 static unsigned long
1167 ttm_dma_pool_shrink_count(struct shrinker *shrink, struct shrink_control *sc)
1168 {
1169         struct device_pools *p;
1170         unsigned long count = 0;
1171
1172         if (!mutex_trylock(&_manager->lock))
1173                 return 0;
1174         list_for_each_entry(p, &_manager->pools, pools)
1175                 count += p->pool->npages_free;
1176         mutex_unlock(&_manager->lock);
1177         return count;
1178 }
1179
1180 static void ttm_dma_pool_mm_shrink_init(struct ttm_pool_manager *manager)
1181 {
1182         manager->mm_shrink.count_objects = ttm_dma_pool_shrink_count;
1183         manager->mm_shrink.scan_objects = &ttm_dma_pool_shrink_scan;
1184         manager->mm_shrink.seeks = 1;
1185         register_shrinker(&manager->mm_shrink);
1186 }
1187
1188 static void ttm_dma_pool_mm_shrink_fini(struct ttm_pool_manager *manager)
1189 {
1190         unregister_shrinker(&manager->mm_shrink);
1191 }
1192
1193 int ttm_dma_page_alloc_init(struct ttm_mem_global *glob, unsigned max_pages)
1194 {
1195         int ret = -ENOMEM;
1196
1197         WARN_ON(_manager);
1198
1199         pr_info("Initializing DMA pool allocator\n");
1200
1201         _manager = kzalloc(sizeof(*_manager), GFP_KERNEL);
1202         if (!_manager)
1203                 goto err;
1204
1205         mutex_init(&_manager->lock);
1206         INIT_LIST_HEAD(&_manager->pools);
1207
1208         _manager->options.max_size = max_pages;
1209         _manager->options.small = SMALL_ALLOCATION;
1210         _manager->options.alloc_size = NUM_PAGES_TO_ALLOC;
1211
1212         /* This takes care of auto-freeing the _manager */
1213         ret = kobject_init_and_add(&_manager->kobj, &ttm_pool_kobj_type,
1214                                    &glob->kobj, "dma_pool");
1215         if (unlikely(ret != 0)) {
1216                 kobject_put(&_manager->kobj);
1217                 goto err;
1218         }
1219         ttm_dma_pool_mm_shrink_init(_manager);
1220         return 0;
1221 err:
1222         return ret;
1223 }
1224
1225 void ttm_dma_page_alloc_fini(void)
1226 {
1227         struct device_pools *p, *t;
1228
1229         pr_info("Finalizing DMA pool allocator\n");
1230         ttm_dma_pool_mm_shrink_fini(_manager);
1231
1232         list_for_each_entry_safe_reverse(p, t, &_manager->pools, pools) {
1233                 dev_dbg(p->dev, "(%s:%d) Freeing.\n", p->pool->name,
1234                         current->pid);
1235                 WARN_ON(devres_destroy(p->dev, ttm_dma_pool_release,
1236                         ttm_dma_pool_match, p->pool));
1237                 ttm_dma_free_pool(p->dev, p->pool->type);
1238         }
1239         kobject_put(&_manager->kobj);
1240         _manager = NULL;
1241 }
1242
1243 int ttm_dma_page_alloc_debugfs(struct seq_file *m, void *data)
1244 {
1245         struct device_pools *p;
1246         struct dma_pool *pool = NULL;
1247         char *h[] = {"pool", "refills", "pages freed", "inuse", "available",
1248                      "name", "virt", "busaddr"};
1249
1250         if (!_manager) {
1251                 seq_printf(m, "No pool allocator running.\n");
1252                 return 0;
1253         }
1254         seq_printf(m, "%13s %12s %13s %8s %8s %8s\n",
1255                    h[0], h[1], h[2], h[3], h[4], h[5]);
1256         mutex_lock(&_manager->lock);
1257         list_for_each_entry(p, &_manager->pools, pools) {
1258                 struct device *dev = p->dev;
1259                 if (!dev)
1260                         continue;
1261                 pool = p->pool;
1262                 seq_printf(m, "%13s %12ld %13ld %8d %8d %8s\n",
1263                                 pool->name, pool->nrefills,
1264                                 pool->nfrees, pool->npages_in_use,
1265                                 pool->npages_free,
1266                                 pool->dev_name);
1267         }
1268         mutex_unlock(&_manager->lock);
1269         return 0;
1270 }
1271 EXPORT_SYMBOL_GPL(ttm_dma_page_alloc_debugfs);
1272
1273 #endif