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 #if 0
 414         if (nr_free > 1) {
 415                 pr_debug("%s: (%s:%d) Attempting to free %d (%d) pages\n",
 416                          pool->dev_name, pool->name, current->pid,
 417                          npages_to_free, nr_free);
 418         }
 419 #endif
 420         if (use_static)
 421                 pages_to_free = static_buf;
 422         else
 423                 pages_to_free = kmalloc_array(npages_to_free,
 424                                               sizeof(struct page *),
 425                                               GFP_KERNEL);
 426
 427         if (!pages_to_free) {
 428                 pr_debug("%s: Failed to allocate memory for pool free operation\n",
 429                        pool->dev_name);
 430                 return 0;
 431         }
 432         INIT_LIST_HEAD(&d_pages);
 433 restart:
 434         spin_lock_irqsave(&pool->lock, irq_flags);
 435
 436         /* We picking the oldest ones off the list */
 437         list_for_each_entry_safe_reverse(dma_p, tmp, &pool->free_list,
 438                                          page_list) {
 439                 if (freed_pages >= npages_to_free)
 440                         break;
 441
 442                 /* Move the dma_page from one list to another. */
 443                 list_move(&dma_p->page_list, &d_pages);
 444
 445                 pages_to_free[freed_pages++] = dma_p->p;
 446                 /* We can only remove NUM_PAGES_TO_ALLOC at a time. */
 447                 if (freed_pages >= NUM_PAGES_TO_ALLOC) {
 448
 449                         ttm_pool_update_free_locked(pool, freed_pages);
 450                         /**
 451                          * Because changing page caching is costly
 452                          * we unlock the pool to prevent stalling.
 453                          */
 454                         spin_unlock_irqrestore(&pool->lock, irq_flags);
 455
 456                         ttm_dma_pages_put(pool, &d_pages, pages_to_free,
 457                                           freed_pages);
 458
 459                         INIT_LIST_HEAD(&d_pages);
 460
 461                         if (likely(nr_free != FREE_ALL_PAGES))
 462                                 nr_free -= freed_pages;
 463
 464                         if (NUM_PAGES_TO_ALLOC >= nr_free)
 465                                 npages_to_free = nr_free;
 466                         else
 467                                 npages_to_free = NUM_PAGES_TO_ALLOC;
 468
 469                         freed_pages = 0;
 470
 471                         /* free all so restart the processing */
 472                         if (nr_free)
 473                                 goto restart;
 474
 475                         /* Not allowed to fall through or break because
 476                          * following context is inside spinlock while we are
 477                          * outside here.
 478                          */
 479                         goto out;
 480
 481                 }
 482         }
 483
 484         /* remove range of pages from the pool */
 485         if (freed_pages) {
 486                 ttm_pool_update_free_locked(pool, freed_pages);
 487                 nr_free -= freed_pages;
 488         }
 489
 490         spin_unlock_irqrestore(&pool->lock, irq_flags);
 491
 492         if (freed_pages)
 493                 ttm_dma_pages_put(pool, &d_pages, pages_to_free, freed_pages);
 494 out:
 495         if (pages_to_free != static_buf)
 496                 kfree(pages_to_free);
 497         return nr_free;
 498 }
 499
 500 static void ttm_dma_free_pool(struct device *dev, enum pool_type type)
 501 {
 502         struct device_pools *p;
 503         struct dma_pool *pool;
 504
 505         if (!dev)
 506                 return;
 507
 508         mutex_lock(&_manager->lock);
 509         list_for_each_entry_reverse(p, &_manager->pools, pools) {
 510                 if (p->dev != dev)
 511                         continue;
 512                 pool = p->pool;
 513                 if (pool->type != type)
 514                         continue;
 515
 516                 list_del(&p->pools);
 517                 kfree(p);
 518                 _manager->npools--;
 519                 break;
 520         }
 521         list_for_each_entry_reverse(pool, &dev->dma_pools, pools) {
 522                 if (pool->type != type)
 523                         continue;
 524                 /* Takes a spinlock.. */
 525                 /* OK to use static buffer since global mutex is held. */
 526                 ttm_dma_page_pool_free(pool, FREE_ALL_PAGES, true);
 527                 WARN_ON(((pool->npages_in_use + pool->npages_free) != 0));
 528                 /* This code path is called after _all_ references to the
 529                  * struct device has been dropped - so nobody should be
 530                  * touching it. In case somebody is trying to _add_ we are
 531                  * guarded by the mutex. */
 532                 list_del(&pool->pools);
 533                 kfree(pool);
 534                 break;
 535         }
 536         mutex_unlock(&_manager->lock);
 537 }
 538
 539 /*
 540  * On free-ing of the 'struct device' this deconstructor is run.
 541  * Albeit the pool might have already been freed earlier.
 542  */
 543 static void ttm_dma_pool_release(struct device *dev, void *res)
 544 {
 545         struct dma_pool *pool = *(struct dma_pool **)res;
 546
 547         if (pool)
 548                 ttm_dma_free_pool(dev, pool->type);
 549 }
 550
 551 static int ttm_dma_pool_match(struct device *dev, void *res, void *match_data)
 552 {
 553         return *(struct dma_pool **)res == match_data;
 554 }
 555
 556 static struct dma_pool *ttm_dma_pool_init(struct device *dev, gfp_t flags,
 557                                           enum pool_type type)
 558 {
 559         const char *n[] = {"wc", "uc", "cached", " dma32", "huge"};
 560         enum pool_type t[] = {IS_WC, IS_UC, IS_CACHED, IS_DMA32, IS_HUGE};
 561         struct device_pools *sec_pool = NULL;
 562         struct dma_pool *pool = NULL, **ptr;
 563         unsigned i;
 564         int ret = -ENODEV;
 565         char *p;
 566
 567         if (!dev)
 568                 return NULL;
 569
 570         ptr = devres_alloc(ttm_dma_pool_release, sizeof(*ptr), GFP_KERNEL);
 571         if (!ptr)
 572                 return NULL;
 573
 574         ret = -ENOMEM;
 575
 576         pool = kmalloc_node(sizeof(struct dma_pool), GFP_KERNEL,
 577                             dev_to_node(dev));
 578         if (!pool)
 579                 goto err_mem;
 580
 581         sec_pool = kmalloc_node(sizeof(struct device_pools), GFP_KERNEL,
 582                                 dev_to_node(dev));
 583         if (!sec_pool)
 584                 goto err_mem;
 585
 586         INIT_LIST_HEAD(&sec_pool->pools);
 587         sec_pool->dev = dev;
 588         sec_pool->pool =  pool;
 589
 590         INIT_LIST_HEAD(&pool->free_list);
 591         INIT_LIST_HEAD(&pool->pools);
 592         spin_lock_init(&pool->lock);
 593         pool->dev = dev;
 594         pool->npages_free = pool->npages_in_use = 0;
 595         pool->nfrees = 0;
 596         pool->gfp_flags = flags;
 597         if (type & IS_HUGE)
 598 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 599                 pool->size = HPAGE_PMD_SIZE;
 600 #else
 601                 BUG();
 602 #endif
 603         else
 604                 pool->size = PAGE_SIZE;
 605         pool->type = type;
 606         pool->nrefills = 0;
 607         p = pool->name;
 608         for (i = 0; i < ARRAY_SIZE(t); i++) {
 609                 if (type & t[i]) {
 610                         p += snprintf(p, sizeof(pool->name) - (p - pool->name),
 611                                       "%s", n[i]);
 612                 }
 613         }
 614         *p = 0;
 615         /* We copy the name for pr_ calls b/c when dma_pool_destroy is called
 616          * - the kobj->name has already been deallocated.*/
 617         snprintf(pool->dev_name, sizeof(pool->dev_name), "%s %s",
 618                  dev_driver_string(dev), dev_name(dev));
 619         mutex_lock(&_manager->lock);
 620         /* You can get the dma_pool from either the global: */
 621         list_add(&sec_pool->pools, &_manager->pools);
 622         _manager->npools++;
 623         /* or from 'struct device': */
 624         list_add(&pool->pools, &dev->dma_pools);
 625         mutex_unlock(&_manager->lock);
 626
 627         *ptr = pool;
 628         devres_add(dev, ptr);
 629
 630         return pool;
 631 err_mem:
 632         devres_free(ptr);
 633         kfree(sec_pool);
 634         kfree(pool);
 635         return ERR_PTR(ret);
 636 }
 637
 638 static struct dma_pool *ttm_dma_find_pool(struct device *dev,
 639                                           enum pool_type type)
 640 {
 641         struct dma_pool *pool, *tmp;
 642
 643         if (type == IS_UNDEFINED)
 644                 return NULL;
 645
 646         /* NB: We iterate on the 'struct dev' which has no spinlock, but
 647          * it does have a kref which we have taken. The kref is taken during
 648          * graphic driver loading - in the drm_pci_init it calls either
 649          * pci_dev_get or pci_register_driver which both end up taking a kref
 650          * on 'struct device'.
 651          *
 652          * On teardown, the graphic drivers end up quiescing the TTM (put_pages)
 653          * and calls the dev_res deconstructors: ttm_dma_pool_release. The nice
 654          * thing is at that point of time there are no pages associated with the
 655          * driver so this function will not be called.
 656          */
 657         list_for_each_entry_safe(pool, tmp, &dev->dma_pools, pools)
 658                 if (pool->type == type)
 659                         return pool;
 660         return NULL;
 661 }
 662
 663 /*
 664  * Free pages the pages that failed to change the caching state. If there
 665  * are pages that have changed their caching state already put them to the
 666  * pool.
 667  */
 668 static void ttm_dma_handle_caching_state_failure(struct dma_pool *pool,
 669                                                  struct list_head *d_pages,
 670                                                  struct page **failed_pages,
 671                                                  unsigned cpages)
 672 {
 673         struct dma_page *d_page, *tmp;
 674         struct page *p;
 675         unsigned i = 0;
 676
 677         p = failed_pages[0];
 678         if (!p)
 679                 return;
 680         /* Find the failed page. */
 681         list_for_each_entry_safe(d_page, tmp, d_pages, page_list) {
 682                 if (d_page->p != p)
 683                         continue;
 684                 /* .. and then progress over the full list. */
 685                 list_del(&d_page->page_list);
 686                 __ttm_dma_free_page(pool, d_page);
 687                 if (++i < cpages)
 688                         p = failed_pages[i];
 689                 else
 690                         break;
 691         }
 692
 693 }
 694
 695 /*
 696  * Allocate 'count' pages, and put 'need' number of them on the
 697  * 'pages' and as well on the 'dma_address' starting at 'dma_offset' offset.
 698  * The full list of pages should also be on 'd_pages'.
 699  * We return zero for success, and negative numbers as errors.
 700  */
 701 static int ttm_dma_pool_alloc_new_pages(struct dma_pool *pool,
 702                                         struct list_head *d_pages,
 703                                         unsigned count)
 704 {
 705         struct page **caching_array;
 706         struct dma_page *dma_p;
 707         struct page *p;
 708         int r = 0;
 709         unsigned i, j, npages, cpages;
 710         unsigned max_cpages = min(count,
 711                         (unsigned)(PAGE_SIZE/sizeof(struct page *)));
 712
 713         /* allocate array for page caching change */
 714         caching_array = kmalloc_array(max_cpages, sizeof(struct page *),
 715                                       GFP_KERNEL);
 716
 717         if (!caching_array) {
 718                 pr_debug("%s: Unable to allocate table for new pages\n",
 719                        pool->dev_name);
 720                 return -ENOMEM;
 721         }
 722
 723         if (count > 1)
 724                 pr_debug("%s: (%s:%d) Getting %d pages\n",
 725                          pool->dev_name, pool->name, current->pid, count);
 726
 727         for (i = 0, cpages = 0; i < count; ++i) {
 728                 dma_p = __ttm_dma_alloc_page(pool);
 729                 if (!dma_p) {
 730                         pr_debug("%s: Unable to get page %u\n",
 731                                  pool->dev_name, i);
 732
 733                         /* store already allocated pages in the pool after
 734                          * setting the caching state */
 735                         if (cpages) {
 736                                 r = ttm_set_pages_caching(pool, caching_array,
 737                                                           cpages);
 738                                 if (r)
 739                                         ttm_dma_handle_caching_state_failure(
 740                                                 pool, d_pages, caching_array,
 741                                                 cpages);
 742                         }
 743                         r = -ENOMEM;
 744                         goto out;
 745                 }
 746                 p = dma_p->p;
 747                 list_add(&dma_p->page_list, d_pages);
 748
 749 #ifdef CONFIG_HIGHMEM
 750                 /* gfp flags of highmem page should never be dma32 so we
 751                  * we should be fine in such case
 752                  */
 753                 if (PageHighMem(p))
 754                         continue;
 755 #endif
 756
 757                 npages = pool->size / PAGE_SIZE;
 758                 for (j = 0; j < npages; ++j) {
 759                         caching_array[cpages++] = p + j;
 760                         if (cpages == max_cpages) {
 761                                 /* Note: Cannot hold the spinlock */
 762                                 r = ttm_set_pages_caching(pool, caching_array,
 763                                                           cpages);
 764                                 if (r) {
 765                                         ttm_dma_handle_caching_state_failure(
 766                                              pool, d_pages, caching_array,
 767                                              cpages);
 768                                         goto out;
 769                                 }
 770                                 cpages = 0;
 771                         }
 772                 }
 773         }
 774
 775         if (cpages) {
 776                 r = ttm_set_pages_caching(pool, caching_array, cpages);
 777                 if (r)
 778                         ttm_dma_handle_caching_state_failure(pool, d_pages,
 779                                         caching_array, cpages);
 780         }
 781 out:
 782         kfree(caching_array);
 783         return r;
 784 }
 785
 786 /*
 787  * @return count of pages still required to fulfill the request.
 788  */
 789 static int ttm_dma_page_pool_fill_locked(struct dma_pool *pool,
 790                                          unsigned long *irq_flags)
 791 {
 792         unsigned count = _manager->options.small;
 793         int r = pool->npages_free;
 794
 795         if (count > pool->npages_free) {
 796                 struct list_head d_pages;
 797
 798                 INIT_LIST_HEAD(&d_pages);
 799
 800                 spin_unlock_irqrestore(&pool->lock, *irq_flags);
 801
 802                 /* Returns how many more are neccessary to fulfill the
 803                  * request. */
 804                 r = ttm_dma_pool_alloc_new_pages(pool, &d_pages, count);
 805
 806                 spin_lock_irqsave(&pool->lock, *irq_flags);
 807                 if (!r) {
 808                         /* Add the fresh to the end.. */
 809                         list_splice(&d_pages, &pool->free_list);
 810                         ++pool->nrefills;
 811                         pool->npages_free += count;
 812                         r = count;
 813                 } else {
 814                         struct dma_page *d_page;
 815                         unsigned cpages = 0;
 816
 817                         pr_debug("%s: Failed to fill %s pool (r:%d)!\n",
 818                                  pool->dev_name, pool->name, r);
 819
 820                         list_for_each_entry(d_page, &d_pages, page_list) {
 821                                 cpages++;
 822                         }
 823                         list_splice_tail(&d_pages, &pool->free_list);
 824                         pool->npages_free += cpages;
 825                         r = cpages;
 826                 }
 827         }
 828         return r;
 829 }
 830
 831 /*
 832  * The populate list is actually a stack (not that is matters as TTM
 833  * allocates one page at a time.
 834  * return dma_page pointer if success, otherwise NULL.
 835  */
 836 static struct dma_page *ttm_dma_pool_get_pages(struct dma_pool *pool,
 837                                   struct ttm_dma_tt *ttm_dma,
 838                                   unsigned index)
 839 {
 840         struct dma_page *d_page = NULL;
 841         struct ttm_tt *ttm = &ttm_dma->ttm;
 842         unsigned long irq_flags;
 843         int count;
 844
 845         spin_lock_irqsave(&pool->lock, irq_flags);
 846         count = ttm_dma_page_pool_fill_locked(pool, &irq_flags);
 847         if (count) {
 848                 d_page = list_first_entry(&pool->free_list, struct dma_page, page_list);
 849                 ttm->pages[index] = d_page->p;
 850                 ttm_dma->dma_address[index] = d_page->dma;
 851                 list_move_tail(&d_page->page_list, &ttm_dma->pages_list);
 852                 pool->npages_in_use += 1;
 853                 pool->npages_free -= 1;
 854         }
 855         spin_unlock_irqrestore(&pool->lock, irq_flags);
 856         return d_page;
 857 }
 858
 859 static gfp_t ttm_dma_pool_gfp_flags(struct ttm_dma_tt *ttm_dma, bool huge)
 860 {
 861         struct ttm_tt *ttm = &ttm_dma->ttm;
 862         gfp_t gfp_flags;
 863
 864         if (ttm->page_flags & TTM_PAGE_FLAG_DMA32)
 865                 gfp_flags = GFP_USER | GFP_DMA32;
 866         else
 867                 gfp_flags = GFP_HIGHUSER;
 868         if (ttm->page_flags & TTM_PAGE_FLAG_ZERO_ALLOC)
 869                 gfp_flags |= __GFP_ZERO;
 870
 871         if (huge) {
 872                 gfp_flags |= GFP_TRANSHUGE_LIGHT | __GFP_NORETRY |
 873                         __GFP_KSWAPD_RECLAIM;
 874                 gfp_flags &= ~__GFP_MOVABLE;
 875                 gfp_flags &= ~__GFP_COMP;
 876         }
 877
 878         if (ttm->page_flags & TTM_PAGE_FLAG_NO_RETRY)
 879                 gfp_flags |= __GFP_RETRY_MAYFAIL;
 880
 881         return gfp_flags;
 882 }
 883
 884 /*
 885  * On success pages list will hold count number of correctly
 886  * cached pages. On failure will hold the negative return value (-ENOMEM, etc).
 887  */
 888 int ttm_dma_populate(struct ttm_dma_tt *ttm_dma, struct device *dev,
 889                         struct ttm_operation_ctx *ctx)
 890 {
 891         struct ttm_tt *ttm = &ttm_dma->ttm;
 892         struct ttm_mem_global *mem_glob = ttm->bdev->glob->mem_glob;
 893         unsigned long num_pages = ttm->num_pages;
 894         struct dma_pool *pool;
 895         struct dma_page *d_page;
 896         enum pool_type type;
 897         unsigned i;
 898         int ret;
 899
 900         if (ttm->state != tt_unpopulated)
 901                 return 0;
 902
 903         if (ttm_check_under_lowerlimit(mem_glob, num_pages, ctx))
 904                 return -ENOMEM;
 905
 906         INIT_LIST_HEAD(&ttm_dma->pages_list);
 907         i = 0;
 908
 909         type = ttm_to_type(ttm->page_flags, ttm->caching_state);
 910
 911 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 912         if (ttm->page_flags & TTM_PAGE_FLAG_DMA32)
 913                 goto skip_huge;
 914
 915         pool = ttm_dma_find_pool(dev, type | IS_HUGE);
 916         if (!pool) {
 917                 gfp_t gfp_flags = ttm_dma_pool_gfp_flags(ttm_dma, true);
 918
 919                 pool = ttm_dma_pool_init(dev, gfp_flags, type | IS_HUGE);
 920                 if (IS_ERR_OR_NULL(pool))
 921                         goto skip_huge;
 922         }
 923
 924         while (num_pages >= HPAGE_PMD_NR) {
 925                 unsigned j;
 926
 927                 d_page = ttm_dma_pool_get_pages(pool, ttm_dma, i);
 928                 if (!d_page)
 929                         break;
 930
 931                 ret = ttm_mem_global_alloc_page(mem_glob, ttm->pages[i],
 932                                                 pool->size, ctx);
 933                 if (unlikely(ret != 0)) {
 934                         ttm_dma_unpopulate(ttm_dma, dev);
 935                         return -ENOMEM;
 936                 }
 937
 938                 d_page->vaddr |= VADDR_FLAG_UPDATED_COUNT;
 939                 for (j = i + 1; j < (i + HPAGE_PMD_NR); ++j) {
 940                         ttm->pages[j] = ttm->pages[j - 1] + 1;
 941                         ttm_dma->dma_address[j] = ttm_dma->dma_address[j - 1] +
 942                                 PAGE_SIZE;
 943                 }
 944
 945                 i += HPAGE_PMD_NR;
 946                 num_pages -= HPAGE_PMD_NR;
 947         }
 948
 949 skip_huge:
 950 #endif
 951
 952         pool = ttm_dma_find_pool(dev, type);
 953         if (!pool) {
 954                 gfp_t gfp_flags = ttm_dma_pool_gfp_flags(ttm_dma, false);
 955
 956                 pool = ttm_dma_pool_init(dev, gfp_flags, type);
 957                 if (IS_ERR_OR_NULL(pool))
 958                         return -ENOMEM;
 959         }
 960
 961         while (num_pages) {
 962                 d_page = ttm_dma_pool_get_pages(pool, ttm_dma, i);
 963                 if (!d_page) {
 964                         ttm_dma_unpopulate(ttm_dma, dev);
 965                         return -ENOMEM;
 966                 }
 967
 968                 ret = ttm_mem_global_alloc_page(mem_glob, ttm->pages[i],
 969                                                 pool->size, ctx);
 970                 if (unlikely(ret != 0)) {
 971                         ttm_dma_unpopulate(ttm_dma, dev);
 972                         return -ENOMEM;
 973                 }
 974
 975                 d_page->vaddr |= VADDR_FLAG_UPDATED_COUNT;
 976                 ++i;
 977                 --num_pages;
 978         }
 979
 980         if (unlikely(ttm->page_flags & TTM_PAGE_FLAG_SWAPPED)) {
 981                 ret = ttm_tt_swapin(ttm);
 982                 if (unlikely(ret != 0)) {
 983                         ttm_dma_unpopulate(ttm_dma, dev);
 984                         return ret;
 985                 }
 986         }
 987
 988         ttm->state = tt_unbound;
 989         return 0;
 990 }
 991 EXPORT_SYMBOL_GPL(ttm_dma_populate);
 992
 993 /* Put all pages in pages list to correct pool to wait for reuse */
 994 void ttm_dma_unpopulate(struct ttm_dma_tt *ttm_dma, struct device *dev)
 995 {
 996         struct ttm_tt *ttm = &ttm_dma->ttm;
 997         struct ttm_mem_global *mem_glob = ttm->bdev->glob->mem_glob;
 998         struct dma_pool *pool;
 999         struct dma_page *d_page, *next;
1000         enum pool_type type;
1001         bool is_cached = false;
1002         unsigned count, i, npages = 0;
1003         unsigned long irq_flags;
1004
1005         type = ttm_to_type(ttm->page_flags, ttm->caching_state);
1006
1007 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1008         pool = ttm_dma_find_pool(dev, type | IS_HUGE);
1009         if (pool) {
1010                 count = 0;
1011                 list_for_each_entry_safe(d_page, next, &ttm_dma->pages_list,
1012                                          page_list) {
1013                         if (!(d_page->vaddr & VADDR_FLAG_HUGE_POOL))
1014                                 continue;
1015
1016                         count++;
1017                         if (d_page->vaddr & VADDR_FLAG_UPDATED_COUNT) {
1018                                 ttm_mem_global_free_page(mem_glob, d_page->p,
1019                                                          pool->size);
1020                                 d_page->vaddr &= ~VADDR_FLAG_UPDATED_COUNT;
1021                         }
1022                         ttm_dma_page_put(pool, d_page);
1023                 }
1024
1025                 spin_lock_irqsave(&pool->lock, irq_flags);
1026                 pool->npages_in_use -= count;
1027                 pool->nfrees += count;
1028                 spin_unlock_irqrestore(&pool->lock, irq_flags);
1029         }
1030 #endif
1031
1032         pool = ttm_dma_find_pool(dev, type);
1033         if (!pool)
1034                 return;
1035
1036         is_cached = (ttm_dma_find_pool(pool->dev,
1037                      ttm_to_type(ttm->page_flags, tt_cached)) == pool);
1038
1039         /* make sure pages array match list and count number of pages */
1040         count = 0;
1041         list_for_each_entry_safe(d_page, next, &ttm_dma->pages_list,
1042                                  page_list) {
1043                 ttm->pages[count] = d_page->p;
1044                 count++;
1045
1046                 if (d_page->vaddr & VADDR_FLAG_UPDATED_COUNT) {
1047                         ttm_mem_global_free_page(mem_glob, d_page->p,
1048                                                  pool->size);
1049                         d_page->vaddr &= ~VADDR_FLAG_UPDATED_COUNT;
1050                 }
1051
1052                 if (is_cached)
1053                         ttm_dma_page_put(pool, d_page);
1054         }
1055
1056         spin_lock_irqsave(&pool->lock, irq_flags);
1057         pool->npages_in_use -= count;
1058         if (is_cached) {
1059                 pool->nfrees += count;
1060         } else {
1061                 pool->npages_free += count;
1062                 list_splice(&ttm_dma->pages_list, &pool->free_list);
1063                 /*
1064                  * Wait to have at at least NUM_PAGES_TO_ALLOC number of pages
1065                  * to free in order to minimize calls to set_memory_wb().
1066                  */
1067                 if (pool->npages_free >= (_manager->options.max_size +
1068                                           NUM_PAGES_TO_ALLOC))
1069                         npages = pool->npages_free - _manager->options.max_size;
1070         }
1071         spin_unlock_irqrestore(&pool->lock, irq_flags);
1072
1073         INIT_LIST_HEAD(&ttm_dma->pages_list);
1074         for (i = 0; i < ttm->num_pages; i++) {
1075                 ttm->pages[i] = NULL;
1076                 ttm_dma->dma_address[i] = 0;
1077         }
1078
1079         /* shrink pool if necessary (only on !is_cached pools)*/
1080         if (npages)
1081                 ttm_dma_page_pool_free(pool, npages, false);
1082         ttm->state = tt_unpopulated;
1083 }
1084 EXPORT_SYMBOL_GPL(ttm_dma_unpopulate);
1085
1086 /**
1087  * Callback for mm to request pool to reduce number of page held.
1088  *
1089  * XXX: (dchinner) Deadlock warning!
1090  *
1091  * I'm getting sadder as I hear more pathetical whimpers about needing per-pool
1092  * shrinkers
1093  */
1094 static unsigned long
1095 ttm_dma_pool_shrink_scan(struct shrinker *shrink, struct shrink_control *sc)
1096 {
1097         static unsigned start_pool;
1098         unsigned idx = 0;
1099         unsigned pool_offset;
1100         unsigned shrink_pages = sc->nr_to_scan;
1101         struct device_pools *p;
1102         unsigned long freed = 0;
1103
1104         if (list_empty(&_manager->pools))
1105                 return SHRINK_STOP;
1106
1107         if (!mutex_trylock(&_manager->lock))
1108                 return SHRINK_STOP;
1109         if (!_manager->npools)
1110                 goto out;
1111         pool_offset = ++start_pool % _manager->npools;
1112         list_for_each_entry(p, &_manager->pools, pools) {
1113                 unsigned nr_free;
1114
1115                 if (!p->dev)
1116                         continue;
1117                 if (shrink_pages == 0)
1118                         break;
1119                 /* Do it in round-robin fashion. */
1120                 if (++idx < pool_offset)
1121                         continue;
1122                 nr_free = shrink_pages;
1123                 /* OK to use static buffer since global mutex is held. */
1124                 shrink_pages = ttm_dma_page_pool_free(p->pool, nr_free, true);
1125                 freed += nr_free - shrink_pages;
1126
1127                 pr_debug("%s: (%s:%d) Asked to shrink %d, have %d more to go\n",
1128                          p->pool->dev_name, p->pool->name, current->pid,
1129                          nr_free, shrink_pages);
1130         }
1131 out:
1132         mutex_unlock(&_manager->lock);
1133         return freed;
1134 }
1135
1136 static unsigned long
1137 ttm_dma_pool_shrink_count(struct shrinker *shrink, struct shrink_control *sc)
1138 {
1139         struct device_pools *p;
1140         unsigned long count = 0;
1141
1142         if (!mutex_trylock(&_manager->lock))
1143                 return 0;
1144         list_for_each_entry(p, &_manager->pools, pools)
1145                 count += p->pool->npages_free;
1146         mutex_unlock(&_manager->lock);
1147         return count;
1148 }
1149
1150 static int ttm_dma_pool_mm_shrink_init(struct ttm_pool_manager *manager)
1151 {
1152         manager->mm_shrink.count_objects = ttm_dma_pool_shrink_count;
1153         manager->mm_shrink.scan_objects = &ttm_dma_pool_shrink_scan;
1154         manager->mm_shrink.seeks = 1;
1155         return register_shrinker(&manager->mm_shrink);
1156 }
1157
1158 static void ttm_dma_pool_mm_shrink_fini(struct ttm_pool_manager *manager)
1159 {
1160         unregister_shrinker(&manager->mm_shrink);
1161 }
1162
1163 int ttm_dma_page_alloc_init(struct ttm_mem_global *glob, unsigned max_pages)
1164 {
1165         int ret;
1166
1167         WARN_ON(_manager);
1168
1169         pr_info("Initializing DMA pool allocator\n");
1170
1171         _manager = kzalloc(sizeof(*_manager), GFP_KERNEL);
1172         if (!_manager)
1173                 return -ENOMEM;
1174
1175         mutex_init(&_manager->lock);
1176         INIT_LIST_HEAD(&_manager->pools);
1177
1178         _manager->options.max_size = max_pages;
1179         _manager->options.small = SMALL_ALLOCATION;
1180         _manager->options.alloc_size = NUM_PAGES_TO_ALLOC;
1181
1182         /* This takes care of auto-freeing the _manager */
1183         ret = kobject_init_and_add(&_manager->kobj, &ttm_pool_kobj_type,
1184                                    &glob->kobj, "dma_pool");
1185         if (unlikely(ret != 0))
1186                 goto error;
1187
1188         ret = ttm_dma_pool_mm_shrink_init(_manager);
1189         if (unlikely(ret != 0))
1190                 goto error;
1191         return 0;
1192
1193 error:
1194         kobject_put(&_manager->kobj);
1195         _manager = NULL;
1196         return ret;
1197 }
1198
1199 void ttm_dma_page_alloc_fini(void)
1200 {
1201         struct device_pools *p, *t;
1202
1203         pr_info("Finalizing DMA pool allocator\n");
1204         ttm_dma_pool_mm_shrink_fini(_manager);
1205
1206         list_for_each_entry_safe_reverse(p, t, &_manager->pools, pools) {
1207                 dev_dbg(p->dev, "(%s:%d) Freeing.\n", p->pool->name,
1208                         current->pid);
1209                 WARN_ON(devres_destroy(p->dev, ttm_dma_pool_release,
1210                         ttm_dma_pool_match, p->pool));
1211                 ttm_dma_free_pool(p->dev, p->pool->type);
1212         }
1213         kobject_put(&_manager->kobj);
1214         _manager = NULL;
1215 }
1216
1217 int ttm_dma_page_alloc_debugfs(struct seq_file *m, void *data)
1218 {
1219         struct device_pools *p;
1220         struct dma_pool *pool = NULL;
1221
1222         if (!_manager) {
1223                 seq_printf(m, "No pool allocator running.\n");
1224                 return 0;
1225         }
1226         seq_printf(m, "         pool      refills   pages freed    inuse available     name\n");
1227         mutex_lock(&_manager->lock);
1228         list_for_each_entry(p, &_manager->pools, pools) {
1229                 struct device *dev = p->dev;
1230                 if (!dev)
1231                         continue;
1232                 pool = p->pool;
1233                 seq_printf(m, "%13s %12ld %13ld %8d %8d %8s\n",
1234                                 pool->name, pool->nrefills,
1235                                 pool->nfrees, pool->npages_in_use,
1236                                 pool->npages_free,
1237                                 pool->dev_name);
1238         }
1239         mutex_unlock(&_manager->lock);
1240         return 0;
1241 }
1242 EXPORT_SYMBOL_GPL(ttm_dma_page_alloc_debugfs);
1243
1244 #endif