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         dma_addr_t dma = d_page->dma;
 289         d_page->vaddr &= ~VADDR_FLAG_HUGE_POOL;
 290         dma_free_coherent(pool->dev, pool->size, (void *)d_page->vaddr, dma);
 291
 292         kfree(d_page);
 293         d_page = NULL;
 294 }
 295 static struct dma_page *__ttm_dma_alloc_page(struct dma_pool *pool)
 296 {
 297         struct dma_page *d_page;
 298         unsigned long attrs = 0;
 299         void *vaddr;
 300
 301         d_page = kmalloc(sizeof(struct dma_page), GFP_KERNEL);
 302         if (!d_page)
 303                 return NULL;
 304
 305         if (pool->type & IS_HUGE)
 306                 attrs = DMA_ATTR_NO_WARN;
 307
 308         vaddr = dma_alloc_attrs(pool->dev, pool->size, &d_page->dma,
 309                                 pool->gfp_flags, attrs);
 310         if (vaddr) {
 311                 if (is_vmalloc_addr(vaddr))
 312                         d_page->p = vmalloc_to_page(vaddr);
 313                 else
 314                         d_page->p = virt_to_page(vaddr);
 315                 d_page->vaddr = (unsigned long)vaddr;
 316                 if (pool->type & IS_HUGE)
 317                         d_page->vaddr |= VADDR_FLAG_HUGE_POOL;
 318         } else {
 319                 kfree(d_page);
 320                 d_page = NULL;
 321         }
 322         return d_page;
 323 }
 324 static enum pool_type ttm_to_type(int flags, enum ttm_caching_state cstate)
 325 {
 326         enum pool_type type = IS_UNDEFINED;
 327
 328         if (flags & TTM_PAGE_FLAG_DMA32)
 329                 type |= IS_DMA32;
 330         if (cstate == tt_cached)
 331                 type |= IS_CACHED;
 332         else if (cstate == tt_uncached)
 333                 type |= IS_UC;
 334         else
 335                 type |= IS_WC;
 336
 337         return type;
 338 }
 339
 340 static void ttm_pool_update_free_locked(struct dma_pool *pool,
 341                                         unsigned freed_pages)
 342 {
 343         pool->npages_free -= freed_pages;
 344         pool->nfrees += freed_pages;
 345
 346 }
 347
 348 /* set memory back to wb and free the pages. */
 349 static void ttm_dma_page_put(struct dma_pool *pool, struct dma_page *d_page)
 350 {
 351         struct page *page = d_page->p;
 352         unsigned num_pages;
 353
 354         /* Don't set WB on WB page pool. */
 355         if (!(pool->type & IS_CACHED)) {
 356                 num_pages = pool->size / PAGE_SIZE;
 357                 if (ttm_set_pages_wb(page, num_pages))
 358                         pr_err("%s: Failed to set %d pages to wb!\n",
 359                                pool->dev_name, num_pages);
 360         }
 361
 362         list_del(&d_page->page_list);
 363         __ttm_dma_free_page(pool, d_page);
 364 }
 365
 366 static void ttm_dma_pages_put(struct dma_pool *pool, struct list_head *d_pages,
 367                               struct page *pages[], unsigned npages)
 368 {
 369         struct dma_page *d_page, *tmp;
 370
 371         if (pool->type & IS_HUGE) {
 372                 list_for_each_entry_safe(d_page, tmp, d_pages, page_list)
 373                         ttm_dma_page_put(pool, d_page);
 374
 375                 return;
 376         }
 377
 378         /* Don't set WB on WB page pool. */
 379         if (npages && !(pool->type & IS_CACHED) &&
 380             ttm_set_pages_array_wb(pages, npages))
 381                 pr_err("%s: Failed to set %d pages to wb!\n",
 382                        pool->dev_name, npages);
 383
 384         list_for_each_entry_safe(d_page, tmp, d_pages, page_list) {
 385                 list_del(&d_page->page_list);
 386                 __ttm_dma_free_page(pool, d_page);
 387         }
 388 }
 389
 390 /*
 391  * Free pages from pool.
 392  *
 393  * To prevent hogging the ttm_swap process we only free NUM_PAGES_TO_ALLOC
 394  * number of pages in one go.
 395  *
 396  * @pool: to free the pages from
 397  * @nr_free: If set to true will free all pages in pool
 398  * @use_static: Safe to use static buffer
 399  **/
 400 static unsigned ttm_dma_page_pool_free(struct dma_pool *pool, unsigned nr_free,
 401                                        bool use_static)
 402 {
 403         static struct page *static_buf[NUM_PAGES_TO_ALLOC];
 404         unsigned long irq_flags;
 405         struct dma_page *dma_p, *tmp;
 406         struct page **pages_to_free;
 407         struct list_head d_pages;
 408         unsigned freed_pages = 0,
 409                  npages_to_free = nr_free;
 410
 411         if (NUM_PAGES_TO_ALLOC < nr_free)
 412                 npages_to_free = NUM_PAGES_TO_ALLOC;
 413
 414         if (use_static)
 415                 pages_to_free = static_buf;
 416         else
 417                 pages_to_free = kmalloc_array(npages_to_free,
 418                                               sizeof(struct page *),
 419                                               GFP_KERNEL);
 420
 421         if (!pages_to_free) {
 422                 pr_debug("%s: Failed to allocate memory for pool free operation\n",
 423                        pool->dev_name);
 424                 return 0;
 425         }
 426         INIT_LIST_HEAD(&d_pages);
 427 restart:
 428         spin_lock_irqsave(&pool->lock, irq_flags);
 429
 430         /* We picking the oldest ones off the list */
 431         list_for_each_entry_safe_reverse(dma_p, tmp, &pool->free_list,
 432                                          page_list) {
 433                 if (freed_pages >= npages_to_free)
 434                         break;
 435
 436                 /* Move the dma_page from one list to another. */
 437                 list_move(&dma_p->page_list, &d_pages);
 438
 439                 pages_to_free[freed_pages++] = dma_p->p;
 440                 /* We can only remove NUM_PAGES_TO_ALLOC at a time. */
 441                 if (freed_pages >= NUM_PAGES_TO_ALLOC) {
 442
 443                         ttm_pool_update_free_locked(pool, freed_pages);
 444                         /**
 445                          * Because changing page caching is costly
 446                          * we unlock the pool to prevent stalling.
 447                          */
 448                         spin_unlock_irqrestore(&pool->lock, irq_flags);
 449
 450                         ttm_dma_pages_put(pool, &d_pages, pages_to_free,
 451                                           freed_pages);
 452
 453                         INIT_LIST_HEAD(&d_pages);
 454
 455                         if (likely(nr_free != FREE_ALL_PAGES))
 456                                 nr_free -= freed_pages;
 457
 458                         if (NUM_PAGES_TO_ALLOC >= nr_free)
 459                                 npages_to_free = nr_free;
 460                         else
 461                                 npages_to_free = NUM_PAGES_TO_ALLOC;
 462
 463                         freed_pages = 0;
 464
 465                         /* free all so restart the processing */
 466                         if (nr_free)
 467                                 goto restart;
 468
 469                         /* Not allowed to fall through or break because
 470                          * following context is inside spinlock while we are
 471                          * outside here.
 472                          */
 473                         goto out;
 474
 475                 }
 476         }
 477
 478         /* remove range of pages from the pool */
 479         if (freed_pages) {
 480                 ttm_pool_update_free_locked(pool, freed_pages);
 481                 nr_free -= freed_pages;
 482         }
 483
 484         spin_unlock_irqrestore(&pool->lock, irq_flags);
 485
 486         if (freed_pages)
 487                 ttm_dma_pages_put(pool, &d_pages, pages_to_free, freed_pages);
 488 out:
 489         if (pages_to_free != static_buf)
 490                 kfree(pages_to_free);
 491         return nr_free;
 492 }
 493
 494 static void ttm_dma_free_pool(struct device *dev, enum pool_type type)
 495 {
 496         struct device_pools *p;
 497         struct dma_pool *pool;
 498
 499         if (!dev)
 500                 return;
 501
 502         mutex_lock(&_manager->lock);
 503         list_for_each_entry_reverse(p, &_manager->pools, pools) {
 504                 if (p->dev != dev)
 505                         continue;
 506                 pool = p->pool;
 507                 if (pool->type != type)
 508                         continue;
 509
 510                 list_del(&p->pools);
 511                 kfree(p);
 512                 _manager->npools--;
 513                 break;
 514         }
 515         list_for_each_entry_reverse(pool, &dev->dma_pools, pools) {
 516                 if (pool->type != type)
 517                         continue;
 518                 /* Takes a spinlock.. */
 519                 /* OK to use static buffer since global mutex is held. */
 520                 ttm_dma_page_pool_free(pool, FREE_ALL_PAGES, true);
 521                 WARN_ON(((pool->npages_in_use + pool->npages_free) != 0));
 522                 /* This code path is called after _all_ references to the
 523                  * struct device has been dropped - so nobody should be
 524                  * touching it. In case somebody is trying to _add_ we are
 525                  * guarded by the mutex. */
 526                 list_del(&pool->pools);
 527                 kfree(pool);
 528                 break;
 529         }
 530         mutex_unlock(&_manager->lock);
 531 }
 532
 533 /*
 534  * On free-ing of the 'struct device' this deconstructor is run.
 535  * Albeit the pool might have already been freed earlier.
 536  */
 537 static void ttm_dma_pool_release(struct device *dev, void *res)
 538 {
 539         struct dma_pool *pool = *(struct dma_pool **)res;
 540
 541         if (pool)
 542                 ttm_dma_free_pool(dev, pool->type);
 543 }
 544
 545 static int ttm_dma_pool_match(struct device *dev, void *res, void *match_data)
 546 {
 547         return *(struct dma_pool **)res == match_data;
 548 }
 549
 550 static struct dma_pool *ttm_dma_pool_init(struct device *dev, gfp_t flags,
 551                                           enum pool_type type)
 552 {
 553         const char *n[] = {"wc", "uc", "cached", " dma32", "huge"};
 554         enum pool_type t[] = {IS_WC, IS_UC, IS_CACHED, IS_DMA32, IS_HUGE};
 555         struct device_pools *sec_pool = NULL;
 556         struct dma_pool *pool = NULL, **ptr;
 557         unsigned i;
 558         int ret = -ENODEV;
 559         char *p;
 560
 561         if (!dev)
 562                 return NULL;
 563
 564         ptr = devres_alloc(ttm_dma_pool_release, sizeof(*ptr), GFP_KERNEL);
 565         if (!ptr)
 566                 return NULL;
 567
 568         ret = -ENOMEM;
 569
 570         pool = kmalloc_node(sizeof(struct dma_pool), GFP_KERNEL,
 571                             dev_to_node(dev));
 572         if (!pool)
 573                 goto err_mem;
 574
 575         sec_pool = kmalloc_node(sizeof(struct device_pools), GFP_KERNEL,
 576                                 dev_to_node(dev));
 577         if (!sec_pool)
 578                 goto err_mem;
 579
 580         INIT_LIST_HEAD(&sec_pool->pools);
 581         sec_pool->dev = dev;
 582         sec_pool->pool =  pool;
 583
 584         INIT_LIST_HEAD(&pool->free_list);
 585         INIT_LIST_HEAD(&pool->pools);
 586         spin_lock_init(&pool->lock);
 587         pool->dev = dev;
 588         pool->npages_free = pool->npages_in_use = 0;
 589         pool->nfrees = 0;
 590         pool->gfp_flags = flags;
 591         if (type & IS_HUGE)
 592 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 593                 pool->size = HPAGE_PMD_SIZE;
 594 #else
 595                 BUG();
 596 #endif
 597         else
 598                 pool->size = PAGE_SIZE;
 599         pool->type = type;
 600         pool->nrefills = 0;
 601         p = pool->name;
 602         for (i = 0; i < ARRAY_SIZE(t); i++) {
 603                 if (type & t[i]) {
 604                         p += snprintf(p, sizeof(pool->name) - (p - pool->name),
 605                                       "%s", n[i]);
 606                 }
 607         }
 608         *p = 0;
 609         /* We copy the name for pr_ calls b/c when dma_pool_destroy is called
 610          * - the kobj->name has already been deallocated.*/
 611         snprintf(pool->dev_name, sizeof(pool->dev_name), "%s %s",
 612                  dev_driver_string(dev), dev_name(dev));
 613         mutex_lock(&_manager->lock);
 614         /* You can get the dma_pool from either the global: */
 615         list_add(&sec_pool->pools, &_manager->pools);
 616         _manager->npools++;
 617         /* or from 'struct device': */
 618         list_add(&pool->pools, &dev->dma_pools);
 619         mutex_unlock(&_manager->lock);
 620
 621         *ptr = pool;
 622         devres_add(dev, ptr);
 623
 624         return pool;
 625 err_mem:
 626         devres_free(ptr);
 627         kfree(sec_pool);
 628         kfree(pool);
 629         return ERR_PTR(ret);
 630 }
 631
 632 static struct dma_pool *ttm_dma_find_pool(struct device *dev,
 633                                           enum pool_type type)
 634 {
 635         struct dma_pool *pool, *tmp;
 636
 637         if (type == IS_UNDEFINED)
 638                 return NULL;
 639
 640         /* NB: We iterate on the 'struct dev' which has no spinlock, but
 641          * it does have a kref which we have taken. The kref is taken during
 642          * graphic driver loading - in the drm_pci_init it calls either
 643          * pci_dev_get or pci_register_driver which both end up taking a kref
 644          * on 'struct device'.
 645          *
 646          * On teardown, the graphic drivers end up quiescing the TTM (put_pages)
 647          * and calls the dev_res deconstructors: ttm_dma_pool_release. The nice
 648          * thing is at that point of time there are no pages associated with the
 649          * driver so this function will not be called.
 650          */
 651         list_for_each_entry_safe(pool, tmp, &dev->dma_pools, pools)
 652                 if (pool->type == type)
 653                         return pool;
 654         return NULL;
 655 }
 656
 657 /*
 658  * Free pages the pages that failed to change the caching state. If there
 659  * are pages that have changed their caching state already put them to the
 660  * pool.
 661  */
 662 static void ttm_dma_handle_caching_state_failure(struct dma_pool *pool,
 663                                                  struct list_head *d_pages,
 664                                                  struct page **failed_pages,
 665                                                  unsigned cpages)
 666 {
 667         struct dma_page *d_page, *tmp;
 668         struct page *p;
 669         unsigned i = 0;
 670
 671         p = failed_pages[0];
 672         if (!p)
 673                 return;
 674         /* Find the failed page. */
 675         list_for_each_entry_safe(d_page, tmp, d_pages, page_list) {
 676                 if (d_page->p != p)
 677                         continue;
 678                 /* .. and then progress over the full list. */
 679                 list_del(&d_page->page_list);
 680                 __ttm_dma_free_page(pool, d_page);
 681                 if (++i < cpages)
 682                         p = failed_pages[i];
 683                 else
 684                         break;
 685         }
 686
 687 }
 688
 689 /*
 690  * Allocate 'count' pages, and put 'need' number of them on the
 691  * 'pages' and as well on the 'dma_address' starting at 'dma_offset' offset.
 692  * The full list of pages should also be on 'd_pages'.
 693  * We return zero for success, and negative numbers as errors.
 694  */
 695 static int ttm_dma_pool_alloc_new_pages(struct dma_pool *pool,
 696                                         struct list_head *d_pages,
 697                                         unsigned count)
 698 {
 699         struct page **caching_array;
 700         struct dma_page *dma_p;
 701         struct page *p;
 702         int r = 0;
 703         unsigned i, j, npages, cpages;
 704         unsigned max_cpages = min(count,
 705                         (unsigned)(PAGE_SIZE/sizeof(struct page *)));
 706
 707         /* allocate array for page caching change */
 708         caching_array = kmalloc_array(max_cpages, sizeof(struct page *),
 709                                       GFP_KERNEL);
 710
 711         if (!caching_array) {
 712                 pr_debug("%s: Unable to allocate table for new pages\n",
 713                        pool->dev_name);
 714                 return -ENOMEM;
 715         }
 716
 717         if (count > 1)
 718                 pr_debug("%s: (%s:%d) Getting %d pages\n",
 719                          pool->dev_name, pool->name, current->pid, count);
 720
 721         for (i = 0, cpages = 0; i < count; ++i) {
 722                 dma_p = __ttm_dma_alloc_page(pool);
 723                 if (!dma_p) {
 724                         pr_debug("%s: Unable to get page %u\n",
 725                                  pool->dev_name, i);
 726
 727                         /* store already allocated pages in the pool after
 728                          * setting the caching state */
 729                         if (cpages) {
 730                                 r = ttm_set_pages_caching(pool, caching_array,
 731                                                           cpages);
 732                                 if (r)
 733                                         ttm_dma_handle_caching_state_failure(
 734                                                 pool, d_pages, caching_array,
 735                                                 cpages);
 736                         }
 737                         r = -ENOMEM;
 738                         goto out;
 739                 }
 740                 p = dma_p->p;
 741                 list_add(&dma_p->page_list, d_pages);
 742
 743 #ifdef CONFIG_HIGHMEM
 744                 /* gfp flags of highmem page should never be dma32 so we
 745                  * we should be fine in such case
 746                  */
 747                 if (PageHighMem(p))
 748                         continue;
 749 #endif
 750
 751                 npages = pool->size / PAGE_SIZE;
 752                 for (j = 0; j < npages; ++j) {
 753                         caching_array[cpages++] = p + j;
 754                         if (cpages == max_cpages) {
 755                                 /* Note: Cannot hold the spinlock */
 756                                 r = ttm_set_pages_caching(pool, caching_array,
 757                                                           cpages);
 758                                 if (r) {
 759                                         ttm_dma_handle_caching_state_failure(
 760                                              pool, d_pages, caching_array,
 761                                              cpages);
 762                                         goto out;
 763                                 }
 764                                 cpages = 0;
 765                         }
 766                 }
 767         }
 768
 769         if (cpages) {
 770                 r = ttm_set_pages_caching(pool, caching_array, cpages);
 771                 if (r)
 772                         ttm_dma_handle_caching_state_failure(pool, d_pages,
 773                                         caching_array, cpages);
 774         }
 775 out:
 776         kfree(caching_array);
 777         return r;
 778 }
 779
 780 /*
 781  * @return count of pages still required to fulfill the request.
 782  */
 783 static int ttm_dma_page_pool_fill_locked(struct dma_pool *pool,
 784                                          unsigned long *irq_flags)
 785 {
 786         unsigned count = _manager->options.small;
 787         int r = pool->npages_free;
 788
 789         if (count > pool->npages_free) {
 790                 struct list_head d_pages;
 791
 792                 INIT_LIST_HEAD(&d_pages);
 793
 794                 spin_unlock_irqrestore(&pool->lock, *irq_flags);
 795
 796                 /* Returns how many more are neccessary to fulfill the
 797                  * request. */
 798                 r = ttm_dma_pool_alloc_new_pages(pool, &d_pages, count);
 799
 800                 spin_lock_irqsave(&pool->lock, *irq_flags);
 801                 if (!r) {
 802                         /* Add the fresh to the end.. */
 803                         list_splice(&d_pages, &pool->free_list);
 804                         ++pool->nrefills;
 805                         pool->npages_free += count;
 806                         r = count;
 807                 } else {
 808                         struct dma_page *d_page;
 809                         unsigned cpages = 0;
 810
 811                         pr_debug("%s: Failed to fill %s pool (r:%d)!\n",
 812                                  pool->dev_name, pool->name, r);
 813
 814                         list_for_each_entry(d_page, &d_pages, page_list) {
 815                                 cpages++;
 816                         }
 817                         list_splice_tail(&d_pages, &pool->free_list);
 818                         pool->npages_free += cpages;
 819                         r = cpages;
 820                 }
 821         }
 822         return r;
 823 }
 824
 825 /*
 826  * The populate list is actually a stack (not that is matters as TTM
 827  * allocates one page at a time.
 828  * return dma_page pointer if success, otherwise NULL.
 829  */
 830 static struct dma_page *ttm_dma_pool_get_pages(struct dma_pool *pool,
 831                                   struct ttm_dma_tt *ttm_dma,
 832                                   unsigned index)
 833 {
 834         struct dma_page *d_page = NULL;
 835         struct ttm_tt *ttm = &ttm_dma->ttm;
 836         unsigned long irq_flags;
 837         int count;
 838
 839         spin_lock_irqsave(&pool->lock, irq_flags);
 840         count = ttm_dma_page_pool_fill_locked(pool, &irq_flags);
 841         if (count) {
 842                 d_page = list_first_entry(&pool->free_list, struct dma_page, page_list);
 843                 ttm->pages[index] = d_page->p;
 844                 ttm_dma->dma_address[index] = d_page->dma;
 845                 list_move_tail(&d_page->page_list, &ttm_dma->pages_list);
 846                 pool->npages_in_use += 1;
 847                 pool->npages_free -= 1;
 848         }
 849         spin_unlock_irqrestore(&pool->lock, irq_flags);
 850         return d_page;
 851 }
 852
 853 static gfp_t ttm_dma_pool_gfp_flags(struct ttm_dma_tt *ttm_dma, bool huge)
 854 {
 855         struct ttm_tt *ttm = &ttm_dma->ttm;
 856         gfp_t gfp_flags;
 857
 858         if (ttm->page_flags & TTM_PAGE_FLAG_DMA32)
 859                 gfp_flags = GFP_USER | GFP_DMA32;
 860         else
 861                 gfp_flags = GFP_HIGHUSER;
 862         if (ttm->page_flags & TTM_PAGE_FLAG_ZERO_ALLOC)
 863                 gfp_flags |= __GFP_ZERO;
 864
 865         if (huge) {
 866                 gfp_flags |= GFP_TRANSHUGE_LIGHT | __GFP_NORETRY |
 867                         __GFP_KSWAPD_RECLAIM;
 868                 gfp_flags &= ~__GFP_MOVABLE;
 869                 gfp_flags &= ~__GFP_COMP;
 870         }
 871
 872         if (ttm->page_flags & TTM_PAGE_FLAG_NO_RETRY)
 873                 gfp_flags |= __GFP_RETRY_MAYFAIL;
 874
 875         return gfp_flags;
 876 }
 877
 878 /*
 879  * On success pages list will hold count number of correctly
 880  * cached pages. On failure will hold the negative return value (-ENOMEM, etc).
 881  */
 882 int ttm_dma_populate(struct ttm_dma_tt *ttm_dma, struct device *dev,
 883                         struct ttm_operation_ctx *ctx)
 884 {
 885         struct ttm_tt *ttm = &ttm_dma->ttm;
 886         struct ttm_mem_global *mem_glob = ttm->bdev->glob->mem_glob;
 887         unsigned long num_pages = ttm->num_pages;
 888         struct dma_pool *pool;
 889         struct dma_page *d_page;
 890         enum pool_type type;
 891         unsigned i;
 892         int ret;
 893
 894         if (ttm->state != tt_unpopulated)
 895                 return 0;
 896
 897         if (ttm_check_under_lowerlimit(mem_glob, num_pages, ctx))
 898                 return -ENOMEM;
 899
 900         INIT_LIST_HEAD(&ttm_dma->pages_list);
 901         i = 0;
 902
 903         type = ttm_to_type(ttm->page_flags, ttm->caching_state);
 904
 905 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 906         if (ttm->page_flags & TTM_PAGE_FLAG_DMA32)
 907                 goto skip_huge;
 908
 909         pool = ttm_dma_find_pool(dev, type | IS_HUGE);
 910         if (!pool) {
 911                 gfp_t gfp_flags = ttm_dma_pool_gfp_flags(ttm_dma, true);
 912
 913                 pool = ttm_dma_pool_init(dev, gfp_flags, type | IS_HUGE);
 914                 if (IS_ERR_OR_NULL(pool))
 915                         goto skip_huge;
 916         }
 917
 918         while (num_pages >= HPAGE_PMD_NR) {
 919                 unsigned j;
 920
 921                 d_page = ttm_dma_pool_get_pages(pool, ttm_dma, i);
 922                 if (!d_page)
 923                         break;
 924
 925                 ret = ttm_mem_global_alloc_page(mem_glob, ttm->pages[i],
 926                                                 pool->size, ctx);
 927                 if (unlikely(ret != 0)) {
 928                         ttm_dma_unpopulate(ttm_dma, dev);
 929                         return -ENOMEM;
 930                 }
 931
 932                 d_page->vaddr |= VADDR_FLAG_UPDATED_COUNT;
 933                 for (j = i + 1; j < (i + HPAGE_PMD_NR); ++j) {
 934                         ttm->pages[j] = ttm->pages[j - 1] + 1;
 935                         ttm_dma->dma_address[j] = ttm_dma->dma_address[j - 1] +
 936                                 PAGE_SIZE;
 937                 }
 938
 939                 i += HPAGE_PMD_NR;
 940                 num_pages -= HPAGE_PMD_NR;
 941         }
 942
 943 skip_huge:
 944 #endif
 945
 946         pool = ttm_dma_find_pool(dev, type);
 947         if (!pool) {
 948                 gfp_t gfp_flags = ttm_dma_pool_gfp_flags(ttm_dma, false);
 949
 950                 pool = ttm_dma_pool_init(dev, gfp_flags, type);
 951                 if (IS_ERR_OR_NULL(pool))
 952                         return -ENOMEM;
 953         }
 954
 955         while (num_pages) {
 956                 d_page = ttm_dma_pool_get_pages(pool, ttm_dma, i);
 957                 if (!d_page) {
 958                         ttm_dma_unpopulate(ttm_dma, dev);
 959                         return -ENOMEM;
 960                 }
 961
 962                 ret = ttm_mem_global_alloc_page(mem_glob, ttm->pages[i],
 963                                                 pool->size, ctx);
 964                 if (unlikely(ret != 0)) {
 965                         ttm_dma_unpopulate(ttm_dma, dev);
 966                         return -ENOMEM;
 967                 }
 968
 969                 d_page->vaddr |= VADDR_FLAG_UPDATED_COUNT;
 970                 ++i;
 971                 --num_pages;
 972         }
 973
 974         if (unlikely(ttm->page_flags & TTM_PAGE_FLAG_SWAPPED)) {
 975                 ret = ttm_tt_swapin(ttm);
 976                 if (unlikely(ret != 0)) {
 977                         ttm_dma_unpopulate(ttm_dma, dev);
 978                         return ret;
 979                 }
 980         }
 981
 982         ttm->state = tt_unbound;
 983         return 0;
 984 }
 985 EXPORT_SYMBOL_GPL(ttm_dma_populate);
 986
 987 /* Put all pages in pages list to correct pool to wait for reuse */
 988 void ttm_dma_unpopulate(struct ttm_dma_tt *ttm_dma, struct device *dev)
 989 {
 990         struct ttm_tt *ttm = &ttm_dma->ttm;
 991         struct ttm_mem_global *mem_glob = ttm->bdev->glob->mem_glob;
 992         struct dma_pool *pool;
 993         struct dma_page *d_page, *next;
 994         enum pool_type type;
 995         bool is_cached = false;
 996         unsigned count, i, npages = 0;
 997         unsigned long irq_flags;
 998
 999         type = ttm_to_type(ttm->page_flags, ttm->caching_state);
1000
1001 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1002         pool = ttm_dma_find_pool(dev, type | IS_HUGE);
1003         if (pool) {
1004                 count = 0;
1005                 list_for_each_entry_safe(d_page, next, &ttm_dma->pages_list,
1006                                          page_list) {
1007                         if (!(d_page->vaddr & VADDR_FLAG_HUGE_POOL))
1008                                 continue;
1009
1010                         count++;
1011                         if (d_page->vaddr & VADDR_FLAG_UPDATED_COUNT) {
1012                                 ttm_mem_global_free_page(mem_glob, d_page->p,
1013                                                          pool->size);
1014                                 d_page->vaddr &= ~VADDR_FLAG_UPDATED_COUNT;
1015                         }
1016                         ttm_dma_page_put(pool, d_page);
1017                 }
1018
1019                 spin_lock_irqsave(&pool->lock, irq_flags);
1020                 pool->npages_in_use -= count;
1021                 pool->nfrees += count;
1022                 spin_unlock_irqrestore(&pool->lock, irq_flags);
1023         }
1024 #endif
1025
1026         pool = ttm_dma_find_pool(dev, type);
1027         if (!pool)
1028                 return;
1029
1030         is_cached = (ttm_dma_find_pool(pool->dev,
1031                      ttm_to_type(ttm->page_flags, tt_cached)) == pool);
1032
1033         /* make sure pages array match list and count number of pages */
1034         count = 0;
1035         list_for_each_entry_safe(d_page, next, &ttm_dma->pages_list,
1036                                  page_list) {
1037                 ttm->pages[count] = d_page->p;
1038                 count++;
1039
1040                 if (d_page->vaddr & VADDR_FLAG_UPDATED_COUNT) {
1041                         ttm_mem_global_free_page(mem_glob, d_page->p,
1042                                                  pool->size);
1043                         d_page->vaddr &= ~VADDR_FLAG_UPDATED_COUNT;
1044                 }
1045
1046                 if (is_cached)
1047                         ttm_dma_page_put(pool, d_page);
1048         }
1049
1050         spin_lock_irqsave(&pool->lock, irq_flags);
1051         pool->npages_in_use -= count;
1052         if (is_cached) {
1053                 pool->nfrees += count;
1054         } else {
1055                 pool->npages_free += count;
1056                 list_splice(&ttm_dma->pages_list, &pool->free_list);
1057                 /*
1058                  * Wait to have at at least NUM_PAGES_TO_ALLOC number of pages
1059                  * to free in order to minimize calls to set_memory_wb().
1060                  */
1061                 if (pool->npages_free >= (_manager->options.max_size +
1062                                           NUM_PAGES_TO_ALLOC))
1063                         npages = pool->npages_free - _manager->options.max_size;
1064         }
1065         spin_unlock_irqrestore(&pool->lock, irq_flags);
1066
1067         INIT_LIST_HEAD(&ttm_dma->pages_list);
1068         for (i = 0; i < ttm->num_pages; i++) {
1069                 ttm->pages[i] = NULL;
1070                 ttm_dma->dma_address[i] = 0;
1071         }
1072
1073         /* shrink pool if necessary (only on !is_cached pools)*/
1074         if (npages)
1075                 ttm_dma_page_pool_free(pool, npages, false);
1076         ttm->state = tt_unpopulated;
1077 }
1078 EXPORT_SYMBOL_GPL(ttm_dma_unpopulate);
1079
1080 /**
1081  * Callback for mm to request pool to reduce number of page held.
1082  *
1083  * XXX: (dchinner) Deadlock warning!
1084  *
1085  * I'm getting sadder as I hear more pathetical whimpers about needing per-pool
1086  * shrinkers
1087  */
1088 static unsigned long
1089 ttm_dma_pool_shrink_scan(struct shrinker *shrink, struct shrink_control *sc)
1090 {
1091         static unsigned start_pool;
1092         unsigned idx = 0;
1093         unsigned pool_offset;
1094         unsigned shrink_pages = sc->nr_to_scan;
1095         struct device_pools *p;
1096         unsigned long freed = 0;
1097
1098         if (list_empty(&_manager->pools))
1099                 return SHRINK_STOP;
1100
1101         if (!mutex_trylock(&_manager->lock))
1102                 return SHRINK_STOP;
1103         if (!_manager->npools)
1104                 goto out;
1105         pool_offset = ++start_pool % _manager->npools;
1106         list_for_each_entry(p, &_manager->pools, pools) {
1107                 unsigned nr_free;
1108
1109                 if (!p->dev)
1110                         continue;
1111                 if (shrink_pages == 0)
1112                         break;
1113                 /* Do it in round-robin fashion. */
1114                 if (++idx < pool_offset)
1115                         continue;
1116                 nr_free = shrink_pages;
1117                 /* OK to use static buffer since global mutex is held. */
1118                 shrink_pages = ttm_dma_page_pool_free(p->pool, nr_free, true);
1119                 freed += nr_free - shrink_pages;
1120
1121                 pr_debug("%s: (%s:%d) Asked to shrink %d, have %d more to go\n",
1122                          p->pool->dev_name, p->pool->name, current->pid,
1123                          nr_free, shrink_pages);
1124         }
1125 out:
1126         mutex_unlock(&_manager->lock);
1127         return freed;
1128 }
1129
1130 static unsigned long
1131 ttm_dma_pool_shrink_count(struct shrinker *shrink, struct shrink_control *sc)
1132 {
1133         struct device_pools *p;
1134         unsigned long count = 0;
1135
1136         if (!mutex_trylock(&_manager->lock))
1137                 return 0;
1138         list_for_each_entry(p, &_manager->pools, pools)
1139                 count += p->pool->npages_free;
1140         mutex_unlock(&_manager->lock);
1141         return count;
1142 }
1143
1144 static int ttm_dma_pool_mm_shrink_init(struct ttm_pool_manager *manager)
1145 {
1146         manager->mm_shrink.count_objects = ttm_dma_pool_shrink_count;
1147         manager->mm_shrink.scan_objects = &ttm_dma_pool_shrink_scan;
1148         manager->mm_shrink.seeks = 1;
1149         return register_shrinker(&manager->mm_shrink);
1150 }
1151
1152 static void ttm_dma_pool_mm_shrink_fini(struct ttm_pool_manager *manager)
1153 {
1154         unregister_shrinker(&manager->mm_shrink);
1155 }
1156
1157 int ttm_dma_page_alloc_init(struct ttm_mem_global *glob, unsigned max_pages)
1158 {
1159         int ret;
1160
1161         WARN_ON(_manager);
1162
1163         pr_info("Initializing DMA pool allocator\n");
1164
1165         _manager = kzalloc(sizeof(*_manager), GFP_KERNEL);
1166         if (!_manager)
1167                 return -ENOMEM;
1168
1169         mutex_init(&_manager->lock);
1170         INIT_LIST_HEAD(&_manager->pools);
1171
1172         _manager->options.max_size = max_pages;
1173         _manager->options.small = SMALL_ALLOCATION;
1174         _manager->options.alloc_size = NUM_PAGES_TO_ALLOC;
1175
1176         /* This takes care of auto-freeing the _manager */
1177         ret = kobject_init_and_add(&_manager->kobj, &ttm_pool_kobj_type,
1178                                    &glob->kobj, "dma_pool");
1179         if (unlikely(ret != 0))
1180                 goto error;
1181
1182         ret = ttm_dma_pool_mm_shrink_init(_manager);
1183         if (unlikely(ret != 0))
1184                 goto error;
1185         return 0;
1186
1187 error:
1188         kobject_put(&_manager->kobj);
1189         _manager = NULL;
1190         return ret;
1191 }
1192
1193 void ttm_dma_page_alloc_fini(void)
1194 {
1195         struct device_pools *p, *t;
1196
1197         pr_info("Finalizing DMA pool allocator\n");
1198         ttm_dma_pool_mm_shrink_fini(_manager);
1199
1200         list_for_each_entry_safe_reverse(p, t, &_manager->pools, pools) {
1201                 dev_dbg(p->dev, "(%s:%d) Freeing.\n", p->pool->name,
1202                         current->pid);
1203                 WARN_ON(devres_destroy(p->dev, ttm_dma_pool_release,
1204                         ttm_dma_pool_match, p->pool));
1205                 ttm_dma_free_pool(p->dev, p->pool->type);
1206         }
1207         kobject_put(&_manager->kobj);
1208         _manager = NULL;
1209 }
1210
1211 int ttm_dma_page_alloc_debugfs(struct seq_file *m, void *data)
1212 {
1213         struct device_pools *p;
1214         struct dma_pool *pool = NULL;
1215
1216         if (!_manager) {
1217                 seq_printf(m, "No pool allocator running.\n");
1218                 return 0;
1219         }
1220         seq_printf(m, "         pool      refills   pages freed    inuse available     name\n");
1221         mutex_lock(&_manager->lock);
1222         list_for_each_entry(p, &_manager->pools, pools) {
1223                 struct device *dev = p->dev;
1224                 if (!dev)
1225                         continue;
1226                 pool = p->pool;
1227                 seq_printf(m, "%13s %12ld %13ld %8d %8d %8s\n",
1228                                 pool->name, pool->nrefills,
1229                                 pool->nfrees, pool->npages_in_use,
1230                                 pool->npages_free,
1231                                 pool->dev_name);
1232         }
1233         mutex_unlock(&_manager->lock);
1234         return 0;
1235 }
1236 EXPORT_SYMBOL_GPL(ttm_dma_page_alloc_debugfs);
1237
1238 #endif