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 #ifdef TTM_HAS_AGP
  54 #include <asm/agp.h>
  55 #endif
  56
  57 #define NUM_PAGES_TO_ALLOC              (PAGE_SIZE/sizeof(struct page *))
  58 #define SMALL_ALLOCATION                4
  59 #define FREE_ALL_PAGES                  (~0U)
  60 /* times are in msecs */
  61 #define IS_UNDEFINED                    (0)
  62 #define IS_WC                           (1<<1)
  63 #define IS_UC                           (1<<2)
  64 #define IS_CACHED                       (1<<3)
  65 #define IS_DMA32                        (1<<4)
  66
  67 enum pool_type {
  68         POOL_IS_UNDEFINED,
  69         POOL_IS_WC = IS_WC,
  70         POOL_IS_UC = IS_UC,
  71         POOL_IS_CACHED = IS_CACHED,
  72         POOL_IS_WC_DMA32 = IS_WC | IS_DMA32,
  73         POOL_IS_UC_DMA32 = IS_UC | IS_DMA32,
  74         POOL_IS_CACHED_DMA32 = IS_CACHED | IS_DMA32,
  75 };
  76 /*
  77  * The pool structure. There are usually six pools:
  78  *  - generic (not restricted to DMA32):
  79  *      - write combined, uncached, cached.
  80  *  - dma32 (up to 2^32 - so up 4GB):
  81  *      - write combined, uncached, cached.
  82  * for each 'struct device'. The 'cached' is for pages that are actively used.
  83  * The other ones can be shrunk by the shrinker API if neccessary.
  84  * @pools: The 'struct device->dma_pools' link.
  85  * @type: Type of the pool
  86  * @lock: Protects the inuse_list and free_list from concurrnet access. Must be
  87  * used with irqsave/irqrestore variants because pool allocator maybe called
  88  * from delayed work.
  89  * @inuse_list: Pool of pages that are in use. The order is very important and
  90  *   it is in the order that the TTM pages that are put back are in.
  91  * @free_list: Pool of pages that are free to be used. No order requirements.
  92  * @dev: The device that is associated with these pools.
  93  * @size: Size used during DMA allocation.
  94  * @npages_free: Count of available pages for re-use.
  95  * @npages_in_use: Count of pages that are in use.
  96  * @nfrees: Stats when pool is shrinking.
  97  * @nrefills: Stats when the pool is grown.
  98  * @gfp_flags: Flags to pass for alloc_page.
  99  * @name: Name of the pool.
 100  * @dev_name: Name derieved from dev - similar to how dev_info works.
 101  *   Used during shutdown as the dev_info during release is unavailable.
 102  */
 103 struct dma_pool {
 104         struct list_head pools; /* The 'struct device->dma_pools link */
 105         enum pool_type type;
 106         spinlock_t lock;
 107         struct list_head inuse_list;
 108         struct list_head free_list;
 109         struct device *dev;
 110         unsigned size;
 111         unsigned npages_free;
 112         unsigned npages_in_use;
 113         unsigned long nfrees; /* Stats when shrunk. */
 114         unsigned long nrefills; /* Stats when grown. */
 115         gfp_t gfp_flags;
 116         char name[13]; /* "cached dma32" */
 117         char dev_name[64]; /* Constructed from dev */
 118 };
 119
 120 /*
 121  * The accounting page keeping track of the allocated page along with
 122  * the DMA address.
 123  * @page_list: The link to the 'page_list' in 'struct dma_pool'.
 124  * @vaddr: The virtual address of the page
 125  * @dma: The bus address of the page. If the page is not allocated
 126  *   via the DMA API, it will be -1.
 127  */
 128 struct dma_page {
 129         struct list_head page_list;
 130         void *vaddr;
 131         struct page *p;
 132         dma_addr_t dma;
 133 };
 134
 135 /*
 136  * Limits for the pool. They are handled without locks because only place where
 137  * they may change is in sysfs store. They won't have immediate effect anyway
 138  * so forcing serialization to access them is pointless.
 139  */
 140
 141 struct ttm_pool_opts {
 142         unsigned        alloc_size;
 143         unsigned        max_size;
 144         unsigned        small;
 145 };
 146
 147 /*
 148  * Contains the list of all of the 'struct device' and their corresponding
 149  * DMA pools. Guarded by _mutex->lock.
 150  * @pools: The link to 'struct ttm_pool_manager->pools'
 151  * @dev: The 'struct device' associated with the 'pool'
 152  * @pool: The 'struct dma_pool' associated with the 'dev'
 153  */
 154 struct device_pools {
 155         struct list_head pools;
 156         struct device *dev;
 157         struct dma_pool *pool;
 158 };
 159
 160 /*
 161  * struct ttm_pool_manager - Holds memory pools for fast allocation
 162  *
 163  * @lock: Lock used when adding/removing from pools
 164  * @pools: List of 'struct device' and 'struct dma_pool' tuples.
 165  * @options: Limits for the pool.
 166  * @npools: Total amount of pools in existence.
 167  * @shrinker: The structure used by [un|]register_shrinker
 168  */
 169 struct ttm_pool_manager {
 170         struct mutex            lock;
 171         struct list_head        pools;
 172         struct ttm_pool_opts    options;
 173         unsigned                npools;
 174         struct shrinker         mm_shrink;
 175         struct kobject          kobj;
 176 };
 177
 178 static struct ttm_pool_manager *_manager;
 179
 180 static struct attribute ttm_page_pool_max = {
 181         .name = "pool_max_size",
 182         .mode = S_IRUGO | S_IWUSR
 183 };
 184 static struct attribute ttm_page_pool_small = {
 185         .name = "pool_small_allocation",
 186         .mode = S_IRUGO | S_IWUSR
 187 };
 188 static struct attribute ttm_page_pool_alloc_size = {
 189         .name = "pool_allocation_size",
 190         .mode = S_IRUGO | S_IWUSR
 191 };
 192
 193 static struct attribute *ttm_pool_attrs[] = {
 194         &ttm_page_pool_max,
 195         &ttm_page_pool_small,
 196         &ttm_page_pool_alloc_size,
 197         NULL
 198 };
 199
 200 static void ttm_pool_kobj_release(struct kobject *kobj)
 201 {
 202         struct ttm_pool_manager *m =
 203                 container_of(kobj, struct ttm_pool_manager, kobj);
 204         kfree(m);
 205 }
 206
 207 static ssize_t ttm_pool_store(struct kobject *kobj, struct attribute *attr,
 208                               const char *buffer, size_t size)
 209 {
 210         struct ttm_pool_manager *m =
 211                 container_of(kobj, struct ttm_pool_manager, kobj);
 212         int chars;
 213         unsigned val;
 214         chars = sscanf(buffer, "%u", &val);
 215         if (chars == 0)
 216                 return size;
 217
 218         /* Convert kb to number of pages */
 219         val = val / (PAGE_SIZE >> 10);
 220
 221         if (attr == &ttm_page_pool_max)
 222                 m->options.max_size = val;
 223         else if (attr == &ttm_page_pool_small)
 224                 m->options.small = val;
 225         else if (attr == &ttm_page_pool_alloc_size) {
 226                 if (val > NUM_PAGES_TO_ALLOC*8) {
 227                         pr_err("Setting allocation size to %lu is not allowed. Recommended size is %lu\n",
 228                                NUM_PAGES_TO_ALLOC*(PAGE_SIZE >> 7),
 229                                NUM_PAGES_TO_ALLOC*(PAGE_SIZE >> 10));
 230                         return size;
 231                 } else if (val > NUM_PAGES_TO_ALLOC) {
 232                         pr_warn("Setting allocation size to larger than %lu is not recommended\n",
 233                                 NUM_PAGES_TO_ALLOC*(PAGE_SIZE >> 10));
 234                 }
 235                 m->options.alloc_size = val;
 236         }
 237
 238         return size;
 239 }
 240
 241 static ssize_t ttm_pool_show(struct kobject *kobj, struct attribute *attr,
 242                              char *buffer)
 243 {
 244         struct ttm_pool_manager *m =
 245                 container_of(kobj, struct ttm_pool_manager, kobj);
 246         unsigned val = 0;
 247
 248         if (attr == &ttm_page_pool_max)
 249                 val = m->options.max_size;
 250         else if (attr == &ttm_page_pool_small)
 251                 val = m->options.small;
 252         else if (attr == &ttm_page_pool_alloc_size)
 253                 val = m->options.alloc_size;
 254
 255         val = val * (PAGE_SIZE >> 10);
 256
 257         return snprintf(buffer, PAGE_SIZE, "%u\n", val);
 258 }
 259
 260 static const struct sysfs_ops ttm_pool_sysfs_ops = {
 261         .show = &ttm_pool_show,
 262         .store = &ttm_pool_store,
 263 };
 264
 265 static struct kobj_type ttm_pool_kobj_type = {
 266         .release = &ttm_pool_kobj_release,
 267         .sysfs_ops = &ttm_pool_sysfs_ops,
 268         .default_attrs = ttm_pool_attrs,
 269 };
 270
 271 #ifndef CONFIG_X86
 272 static int set_pages_array_wb(struct page **pages, int addrinarray)
 273 {
 274 #ifdef TTM_HAS_AGP
 275         int i;
 276
 277         for (i = 0; i < addrinarray; i++)
 278                 unmap_page_from_agp(pages[i]);
 279 #endif
 280         return 0;
 281 }
 282
 283 static int set_pages_array_wc(struct page **pages, int addrinarray)
 284 {
 285 #ifdef TTM_HAS_AGP
 286         int i;
 287
 288         for (i = 0; i < addrinarray; i++)
 289                 map_page_into_agp(pages[i]);
 290 #endif
 291         return 0;
 292 }
 293
 294 static int set_pages_array_uc(struct page **pages, int addrinarray)
 295 {
 296 #ifdef TTM_HAS_AGP
 297         int i;
 298
 299         for (i = 0; i < addrinarray; i++)
 300                 map_page_into_agp(pages[i]);
 301 #endif
 302         return 0;
 303 }
 304 #endif /* for !CONFIG_X86 */
 305
 306 static int ttm_set_pages_caching(struct dma_pool *pool,
 307                                  struct page **pages, unsigned cpages)
 308 {
 309         int r = 0;
 310         /* Set page caching */
 311         if (pool->type & IS_UC) {
 312                 r = set_pages_array_uc(pages, cpages);
 313                 if (r)
 314                         pr_err("%s: Failed to set %d pages to uc!\n",
 315                                pool->dev_name, cpages);
 316         }
 317         if (pool->type & IS_WC) {
 318                 r = set_pages_array_wc(pages, cpages);
 319                 if (r)
 320                         pr_err("%s: Failed to set %d pages to wc!\n",
 321                                pool->dev_name, cpages);
 322         }
 323         return r;
 324 }
 325
 326 static void __ttm_dma_free_page(struct dma_pool *pool, struct dma_page *d_page)
 327 {
 328         dma_addr_t dma = d_page->dma;
 329         dma_free_coherent(pool->dev, pool->size, d_page->vaddr, dma);
 330
 331         kfree(d_page);
 332         d_page = NULL;
 333 }
 334 static struct dma_page *__ttm_dma_alloc_page(struct dma_pool *pool)
 335 {
 336         struct dma_page *d_page;
 337
 338         d_page = kmalloc(sizeof(struct dma_page), GFP_KERNEL);
 339         if (!d_page)
 340                 return NULL;
 341
 342         d_page->vaddr = dma_alloc_coherent(pool->dev, pool->size,
 343                                            &d_page->dma,
 344                                            pool->gfp_flags);
 345         if (d_page->vaddr)
 346                 d_page->p = virt_to_page(d_page->vaddr);
 347         else {
 348                 kfree(d_page);
 349                 d_page = NULL;
 350         }
 351         return d_page;
 352 }
 353 static enum pool_type ttm_to_type(int flags, enum ttm_caching_state cstate)
 354 {
 355         enum pool_type type = IS_UNDEFINED;
 356
 357         if (flags & TTM_PAGE_FLAG_DMA32)
 358                 type |= IS_DMA32;
 359         if (cstate == tt_cached)
 360                 type |= IS_CACHED;
 361         else if (cstate == tt_uncached)
 362                 type |= IS_UC;
 363         else
 364                 type |= IS_WC;
 365
 366         return type;
 367 }
 368
 369 static void ttm_pool_update_free_locked(struct dma_pool *pool,
 370                                         unsigned freed_pages)
 371 {
 372         pool->npages_free -= freed_pages;
 373         pool->nfrees += freed_pages;
 374
 375 }
 376
 377 /* set memory back to wb and free the pages. */
 378 static void ttm_dma_pages_put(struct dma_pool *pool, struct list_head *d_pages,
 379                               struct page *pages[], unsigned npages)
 380 {
 381         struct dma_page *d_page, *tmp;
 382
 383         /* Don't set WB on WB page pool. */
 384         if (npages && !(pool->type & IS_CACHED) &&
 385             set_pages_array_wb(pages, npages))
 386                 pr_err("%s: Failed to set %d pages to wb!\n",
 387                        pool->dev_name, npages);
 388
 389         list_for_each_entry_safe(d_page, tmp, d_pages, page_list) {
 390                 list_del(&d_page->page_list);
 391                 __ttm_dma_free_page(pool, d_page);
 392         }
 393 }
 394
 395 static void ttm_dma_page_put(struct dma_pool *pool, struct dma_page *d_page)
 396 {
 397         /* Don't set WB on WB page pool. */
 398         if (!(pool->type & IS_CACHED) && set_pages_array_wb(&d_page->p, 1))
 399                 pr_err("%s: Failed to set %d pages to wb!\n",
 400                        pool->dev_name, 1);
 401
 402         list_del(&d_page->page_list);
 403         __ttm_dma_free_page(pool, d_page);
 404 }
 405
 406 /*
 407  * Free pages from pool.
 408  *
 409  * To prevent hogging the ttm_swap process we only free NUM_PAGES_TO_ALLOC
 410  * number of pages in one go.
 411  *
 412  * @pool: to free the pages from
 413  * @nr_free: If set to true will free all pages in pool
 414  * @gfp: GFP flags.
 415  **/
 416 static unsigned ttm_dma_page_pool_free(struct dma_pool *pool, unsigned nr_free,
 417                                        gfp_t gfp)
 418 {
 419         unsigned long irq_flags;
 420         struct dma_page *dma_p, *tmp;
 421         struct page **pages_to_free;
 422         struct list_head d_pages;
 423         unsigned freed_pages = 0,
 424                  npages_to_free = nr_free;
 425
 426         if (NUM_PAGES_TO_ALLOC < nr_free)
 427                 npages_to_free = NUM_PAGES_TO_ALLOC;
 428 #if 0
 429         if (nr_free > 1) {
 430                 pr_debug("%s: (%s:%d) Attempting to free %d (%d) pages\n",
 431                          pool->dev_name, pool->name, current->pid,
 432                          npages_to_free, nr_free);
 433         }
 434 #endif
 435         pages_to_free = kmalloc(npages_to_free * sizeof(struct page *), gfp);
 436
 437         if (!pages_to_free) {
 438                 pr_err("%s: Failed to allocate memory for pool free operation\n",
 439                        pool->dev_name);
 440                 return 0;
 441         }
 442         INIT_LIST_HEAD(&d_pages);
 443 restart:
 444         spin_lock_irqsave(&pool->lock, irq_flags);
 445
 446         /* We picking the oldest ones off the list */
 447         list_for_each_entry_safe_reverse(dma_p, tmp, &pool->free_list,
 448                                          page_list) {
 449                 if (freed_pages >= npages_to_free)
 450                         break;
 451
 452                 /* Move the dma_page from one list to another. */
 453                 list_move(&dma_p->page_list, &d_pages);
 454
 455                 pages_to_free[freed_pages++] = dma_p->p;
 456                 /* We can only remove NUM_PAGES_TO_ALLOC at a time. */
 457                 if (freed_pages >= NUM_PAGES_TO_ALLOC) {
 458
 459                         ttm_pool_update_free_locked(pool, freed_pages);
 460                         /**
 461                          * Because changing page caching is costly
 462                          * we unlock the pool to prevent stalling.
 463                          */
 464                         spin_unlock_irqrestore(&pool->lock, irq_flags);
 465
 466                         ttm_dma_pages_put(pool, &d_pages, pages_to_free,
 467                                           freed_pages);
 468
 469                         INIT_LIST_HEAD(&d_pages);
 470
 471                         if (likely(nr_free != FREE_ALL_PAGES))
 472                                 nr_free -= freed_pages;
 473
 474                         if (NUM_PAGES_TO_ALLOC >= nr_free)
 475                                 npages_to_free = nr_free;
 476                         else
 477                                 npages_to_free = NUM_PAGES_TO_ALLOC;
 478
 479                         freed_pages = 0;
 480
 481                         /* free all so restart the processing */
 482                         if (nr_free)
 483                                 goto restart;
 484
 485                         /* Not allowed to fall through or break because
 486                          * following context is inside spinlock while we are
 487                          * outside here.
 488                          */
 489                         goto out;
 490
 491                 }
 492         }
 493
 494         /* remove range of pages from the pool */
 495         if (freed_pages) {
 496                 ttm_pool_update_free_locked(pool, freed_pages);
 497                 nr_free -= freed_pages;
 498         }
 499
 500         spin_unlock_irqrestore(&pool->lock, irq_flags);
 501
 502         if (freed_pages)
 503                 ttm_dma_pages_put(pool, &d_pages, pages_to_free, freed_pages);
 504 out:
 505         kfree(pages_to_free);
 506         return nr_free;
 507 }
 508
 509 static void ttm_dma_free_pool(struct device *dev, enum pool_type type)
 510 {
 511         struct device_pools *p;
 512         struct dma_pool *pool;
 513
 514         if (!dev)
 515                 return;
 516
 517         mutex_lock(&_manager->lock);
 518         list_for_each_entry_reverse(p, &_manager->pools, pools) {
 519                 if (p->dev != dev)
 520                         continue;
 521                 pool = p->pool;
 522                 if (pool->type != type)
 523                         continue;
 524
 525                 list_del(&p->pools);
 526                 kfree(p);
 527                 _manager->npools--;
 528                 break;
 529         }
 530         list_for_each_entry_reverse(pool, &dev->dma_pools, pools) {
 531                 if (pool->type != type)
 532                         continue;
 533                 /* Takes a spinlock.. */
 534                 ttm_dma_page_pool_free(pool, FREE_ALL_PAGES, GFP_KERNEL);
 535                 WARN_ON(((pool->npages_in_use + pool->npages_free) != 0));
 536                 /* This code path is called after _all_ references to the
 537                  * struct device has been dropped - so nobody should be
 538                  * touching it. In case somebody is trying to _add_ we are
 539                  * guarded by the mutex. */
 540                 list_del(&pool->pools);
 541                 kfree(pool);
 542                 break;
 543         }
 544         mutex_unlock(&_manager->lock);
 545 }
 546
 547 /*
 548  * On free-ing of the 'struct device' this deconstructor is run.
 549  * Albeit the pool might have already been freed earlier.
 550  */
 551 static void ttm_dma_pool_release(struct device *dev, void *res)
 552 {
 553         struct dma_pool *pool = *(struct dma_pool **)res;
 554
 555         if (pool)
 556                 ttm_dma_free_pool(dev, pool->type);
 557 }
 558
 559 static int ttm_dma_pool_match(struct device *dev, void *res, void *match_data)
 560 {
 561         return *(struct dma_pool **)res == match_data;
 562 }
 563
 564 static struct dma_pool *ttm_dma_pool_init(struct device *dev, gfp_t flags,
 565                                           enum pool_type type)
 566 {
 567         char *n[] = {"wc", "uc", "cached", " dma32", "unknown",};
 568         enum pool_type t[] = {IS_WC, IS_UC, IS_CACHED, IS_DMA32, IS_UNDEFINED};
 569         struct device_pools *sec_pool = NULL;
 570         struct dma_pool *pool = NULL, **ptr;
 571         unsigned i;
 572         int ret = -ENODEV;
 573         char *p;
 574
 575         if (!dev)
 576                 return NULL;
 577
 578         ptr = devres_alloc(ttm_dma_pool_release, sizeof(*ptr), GFP_KERNEL);
 579         if (!ptr)
 580                 return NULL;
 581
 582         ret = -ENOMEM;
 583
 584         pool = kmalloc_node(sizeof(struct dma_pool), GFP_KERNEL,
 585                             dev_to_node(dev));
 586         if (!pool)
 587                 goto err_mem;
 588
 589         sec_pool = kmalloc_node(sizeof(struct device_pools), GFP_KERNEL,
 590                                 dev_to_node(dev));
 591         if (!sec_pool)
 592                 goto err_mem;
 593
 594         INIT_LIST_HEAD(&sec_pool->pools);
 595         sec_pool->dev = dev;
 596         sec_pool->pool =  pool;
 597
 598         INIT_LIST_HEAD(&pool->free_list);
 599         INIT_LIST_HEAD(&pool->inuse_list);
 600         INIT_LIST_HEAD(&pool->pools);
 601         spin_lock_init(&pool->lock);
 602         pool->dev = dev;
 603         pool->npages_free = pool->npages_in_use = 0;
 604         pool->nfrees = 0;
 605         pool->gfp_flags = flags;
 606         pool->size = PAGE_SIZE;
 607         pool->type = type;
 608         pool->nrefills = 0;
 609         p = pool->name;
 610         for (i = 0; i < 5; i++) {
 611                 if (type & t[i]) {
 612                         p += snprintf(p, sizeof(pool->name) - (p - pool->name),
 613                                       "%s", n[i]);
 614                 }
 615         }
 616         *p = 0;
 617         /* We copy the name for pr_ calls b/c when dma_pool_destroy is called
 618          * - the kobj->name has already been deallocated.*/
 619         snprintf(pool->dev_name, sizeof(pool->dev_name), "%s %s",
 620                  dev_driver_string(dev), dev_name(dev));
 621         mutex_lock(&_manager->lock);
 622         /* You can get the dma_pool from either the global: */
 623         list_add(&sec_pool->pools, &_manager->pools);
 624         _manager->npools++;
 625         /* or from 'struct device': */
 626         list_add(&pool->pools, &dev->dma_pools);
 627         mutex_unlock(&_manager->lock);
 628
 629         *ptr = pool;
 630         devres_add(dev, ptr);
 631
 632         return pool;
 633 err_mem:
 634         devres_free(ptr);
 635         kfree(sec_pool);
 636         kfree(pool);
 637         return ERR_PTR(ret);
 638 }
 639
 640 static struct dma_pool *ttm_dma_find_pool(struct device *dev,
 641                                           enum pool_type type)
 642 {
 643         struct dma_pool *pool, *tmp, *found = NULL;
 644
 645         if (type == IS_UNDEFINED)
 646                 return found;
 647
 648         /* NB: We iterate on the 'struct dev' which has no spinlock, but
 649          * it does have a kref which we have taken. The kref is taken during
 650          * graphic driver loading - in the drm_pci_init it calls either
 651          * pci_dev_get or pci_register_driver which both end up taking a kref
 652          * on 'struct device'.
 653          *
 654          * On teardown, the graphic drivers end up quiescing the TTM (put_pages)
 655          * and calls the dev_res deconstructors: ttm_dma_pool_release. The nice
 656          * thing is at that point of time there are no pages associated with the
 657          * driver so this function will not be called.
 658          */
 659         list_for_each_entry_safe(pool, tmp, &dev->dma_pools, pools) {
 660                 if (pool->type != type)
 661                         continue;
 662                 found = pool;
 663                 break;
 664         }
 665         return found;
 666 }
 667
 668 /*
 669  * Free pages the pages that failed to change the caching state. If there
 670  * are pages that have changed their caching state already put them to the
 671  * pool.
 672  */
 673 static void ttm_dma_handle_caching_state_failure(struct dma_pool *pool,
 674                                                  struct list_head *d_pages,
 675                                                  struct page **failed_pages,
 676                                                  unsigned cpages)
 677 {
 678         struct dma_page *d_page, *tmp;
 679         struct page *p;
 680         unsigned i = 0;
 681
 682         p = failed_pages[0];
 683         if (!p)
 684                 return;
 685         /* Find the failed page. */
 686         list_for_each_entry_safe(d_page, tmp, d_pages, page_list) {
 687                 if (d_page->p != p)
 688                         continue;
 689                 /* .. and then progress over the full list. */
 690                 list_del(&d_page->page_list);
 691                 __ttm_dma_free_page(pool, d_page);
 692                 if (++i < cpages)
 693                         p = failed_pages[i];
 694                 else
 695                         break;
 696         }
 697
 698 }
 699
 700 /*
 701  * Allocate 'count' pages, and put 'need' number of them on the
 702  * 'pages' and as well on the 'dma_address' starting at 'dma_offset' offset.
 703  * The full list of pages should also be on 'd_pages'.
 704  * We return zero for success, and negative numbers as errors.
 705  */
 706 static int ttm_dma_pool_alloc_new_pages(struct dma_pool *pool,
 707                                         struct list_head *d_pages,
 708                                         unsigned count)
 709 {
 710         struct page **caching_array;
 711         struct dma_page *dma_p;
 712         struct page *p;
 713         int r = 0;
 714         unsigned i, cpages;
 715         unsigned max_cpages = min(count,
 716                         (unsigned)(PAGE_SIZE/sizeof(struct page *)));
 717
 718         /* allocate array for page caching change */
 719         caching_array = kmalloc(max_cpages*sizeof(struct page *), GFP_KERNEL);
 720
 721         if (!caching_array) {
 722                 pr_err("%s: Unable to allocate table for new pages\n",
 723                        pool->dev_name);
 724                 return -ENOMEM;
 725         }
 726
 727         if (count > 1) {
 728                 pr_debug("%s: (%s:%d) Getting %d pages\n",
 729                          pool->dev_name, pool->name, current->pid, count);
 730         }
 731
 732         for (i = 0, cpages = 0; i < count; ++i) {
 733                 dma_p = __ttm_dma_alloc_page(pool);
 734                 if (!dma_p) {
 735                         pr_err("%s: Unable to get page %u\n",
 736                                pool->dev_name, i);
 737
 738                         /* store already allocated pages in the pool after
 739                          * setting the caching state */
 740                         if (cpages) {
 741                                 r = ttm_set_pages_caching(pool, caching_array,
 742                                                           cpages);
 743                                 if (r)
 744                                         ttm_dma_handle_caching_state_failure(
 745                                                 pool, d_pages, caching_array,
 746                                                 cpages);
 747                         }
 748                         r = -ENOMEM;
 749                         goto out;
 750                 }
 751                 p = dma_p->p;
 752 #ifdef CONFIG_HIGHMEM
 753                 /* gfp flags of highmem page should never be dma32 so we
 754                  * we should be fine in such case
 755                  */
 756                 if (!PageHighMem(p))
 757 #endif
 758                 {
 759                         caching_array[cpages++] = p;
 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                 list_add(&dma_p->page_list, d_pages);
 774         }
 775
 776         if (cpages) {
 777                 r = ttm_set_pages_caching(pool, caching_array, cpages);
 778                 if (r)
 779                         ttm_dma_handle_caching_state_failure(pool, d_pages,
 780                                         caching_array, cpages);
 781         }
 782 out:
 783         kfree(caching_array);
 784         return r;
 785 }
 786
 787 /*
 788  * @return count of pages still required to fulfill the request.
 789  */
 790 static int ttm_dma_page_pool_fill_locked(struct dma_pool *pool,
 791                                          unsigned long *irq_flags)
 792 {
 793         unsigned count = _manager->options.small;
 794         int r = pool->npages_free;
 795
 796         if (count > pool->npages_free) {
 797                 struct list_head d_pages;
 798
 799                 INIT_LIST_HEAD(&d_pages);
 800
 801                 spin_unlock_irqrestore(&pool->lock, *irq_flags);
 802
 803                 /* Returns how many more are neccessary to fulfill the
 804                  * request. */
 805                 r = ttm_dma_pool_alloc_new_pages(pool, &d_pages, count);
 806
 807                 spin_lock_irqsave(&pool->lock, *irq_flags);
 808                 if (!r) {
 809                         /* Add the fresh to the end.. */
 810                         list_splice(&d_pages, &pool->free_list);
 811                         ++pool->nrefills;
 812                         pool->npages_free += count;
 813                         r = count;
 814                 } else {
 815                         struct dma_page *d_page;
 816                         unsigned cpages = 0;
 817
 818                         pr_err("%s: Failed to fill %s pool (r:%d)!\n",
 819                                pool->dev_name, pool->name, r);
 820
 821                         list_for_each_entry(d_page, &d_pages, page_list) {
 822                                 cpages++;
 823                         }
 824                         list_splice_tail(&d_pages, &pool->free_list);
 825                         pool->npages_free += cpages;
 826                         r = cpages;
 827                 }
 828         }
 829         return r;
 830 }
 831
 832 /*
 833  * @return count of pages still required to fulfill the request.
 834  * The populate list is actually a stack (not that is matters as TTM
 835  * allocates one page at a time.
 836  */
 837 static int ttm_dma_pool_get_pages(struct dma_pool *pool,
 838                                   struct ttm_dma_tt *ttm_dma,
 839                                   unsigned index)
 840 {
 841         struct dma_page *d_page;
 842         struct ttm_tt *ttm = &ttm_dma->ttm;
 843         unsigned long irq_flags;
 844         int count, r = -ENOMEM;
 845
 846         spin_lock_irqsave(&pool->lock, irq_flags);
 847         count = ttm_dma_page_pool_fill_locked(pool, &irq_flags);
 848         if (count) {
 849                 d_page = list_first_entry(&pool->free_list, struct dma_page, page_list);
 850                 ttm->pages[index] = d_page->p;
 851                 ttm_dma->dma_address[index] = d_page->dma;
 852                 list_move_tail(&d_page->page_list, &ttm_dma->pages_list);
 853                 r = 0;
 854                 pool->npages_in_use += 1;
 855                 pool->npages_free -= 1;
 856         }
 857         spin_unlock_irqrestore(&pool->lock, irq_flags);
 858         return r;
 859 }
 860
 861 /*
 862  * On success pages list will hold count number of correctly
 863  * cached pages. On failure will hold the negative return value (-ENOMEM, etc).
 864  */
 865 int ttm_dma_populate(struct ttm_dma_tt *ttm_dma, struct device *dev)
 866 {
 867         struct ttm_tt *ttm = &ttm_dma->ttm;
 868         struct ttm_mem_global *mem_glob = ttm->glob->mem_glob;
 869         struct dma_pool *pool;
 870         enum pool_type type;
 871         unsigned i;
 872         gfp_t gfp_flags;
 873         int ret;
 874
 875         if (ttm->state != tt_unpopulated)
 876                 return 0;
 877
 878         type = ttm_to_type(ttm->page_flags, ttm->caching_state);
 879         if (ttm->page_flags & TTM_PAGE_FLAG_DMA32)
 880                 gfp_flags = GFP_USER | GFP_DMA32;
 881         else
 882                 gfp_flags = GFP_HIGHUSER;
 883         if (ttm->page_flags & TTM_PAGE_FLAG_ZERO_ALLOC)
 884                 gfp_flags |= __GFP_ZERO;
 885
 886         pool = ttm_dma_find_pool(dev, type);
 887         if (!pool) {
 888                 pool = ttm_dma_pool_init(dev, gfp_flags, type);
 889                 if (IS_ERR_OR_NULL(pool)) {
 890                         return -ENOMEM;
 891                 }
 892         }
 893
 894         INIT_LIST_HEAD(&ttm_dma->pages_list);
 895         for (i = 0; i < ttm->num_pages; ++i) {
 896                 ret = ttm_dma_pool_get_pages(pool, ttm_dma, i);
 897                 if (ret != 0) {
 898                         ttm_dma_unpopulate(ttm_dma, dev);
 899                         return -ENOMEM;
 900                 }
 901
 902                 ret = ttm_mem_global_alloc_page(mem_glob, ttm->pages[i],
 903                                                 false, false);
 904                 if (unlikely(ret != 0)) {
 905                         ttm_dma_unpopulate(ttm_dma, dev);
 906                         return -ENOMEM;
 907                 }
 908         }
 909
 910         if (unlikely(ttm->page_flags & TTM_PAGE_FLAG_SWAPPED)) {
 911                 ret = ttm_tt_swapin(ttm);
 912                 if (unlikely(ret != 0)) {
 913                         ttm_dma_unpopulate(ttm_dma, dev);
 914                         return ret;
 915                 }
 916         }
 917
 918         ttm->state = tt_unbound;
 919         return 0;
 920 }
 921 EXPORT_SYMBOL_GPL(ttm_dma_populate);
 922
 923 /* Put all pages in pages list to correct pool to wait for reuse */
 924 void ttm_dma_unpopulate(struct ttm_dma_tt *ttm_dma, struct device *dev)
 925 {
 926         struct ttm_tt *ttm = &ttm_dma->ttm;
 927         struct dma_pool *pool;
 928         struct dma_page *d_page, *next;
 929         enum pool_type type;
 930         bool is_cached = false;
 931         unsigned count = 0, i, npages = 0;
 932         unsigned long irq_flags;
 933
 934         type = ttm_to_type(ttm->page_flags, ttm->caching_state);
 935         pool = ttm_dma_find_pool(dev, type);
 936         if (!pool)
 937                 return;
 938
 939         is_cached = (ttm_dma_find_pool(pool->dev,
 940                      ttm_to_type(ttm->page_flags, tt_cached)) == pool);
 941
 942         /* make sure pages array match list and count number of pages */
 943         list_for_each_entry(d_page, &ttm_dma->pages_list, page_list) {
 944                 ttm->pages[count] = d_page->p;
 945                 count++;
 946         }
 947
 948         spin_lock_irqsave(&pool->lock, irq_flags);
 949         pool->npages_in_use -= count;
 950         if (is_cached) {
 951                 pool->nfrees += count;
 952         } else {
 953                 pool->npages_free += count;
 954                 list_splice(&ttm_dma->pages_list, &pool->free_list);
 955                 npages = count;
 956                 if (pool->npages_free > _manager->options.max_size) {
 957                         npages = pool->npages_free - _manager->options.max_size;
 958                         /* free at least NUM_PAGES_TO_ALLOC number of pages
 959                          * to reduce calls to set_memory_wb */
 960                         if (npages < NUM_PAGES_TO_ALLOC)
 961                                 npages = NUM_PAGES_TO_ALLOC;
 962                 }
 963         }
 964         spin_unlock_irqrestore(&pool->lock, irq_flags);
 965
 966         if (is_cached) {
 967                 list_for_each_entry_safe(d_page, next, &ttm_dma->pages_list, page_list) {
 968                         ttm_mem_global_free_page(ttm->glob->mem_glob,
 969                                                  d_page->p);
 970                         ttm_dma_page_put(pool, d_page);
 971                 }
 972         } else {
 973                 for (i = 0; i < count; i++) {
 974                         ttm_mem_global_free_page(ttm->glob->mem_glob,
 975                                                  ttm->pages[i]);
 976                 }
 977         }
 978
 979         INIT_LIST_HEAD(&ttm_dma->pages_list);
 980         for (i = 0; i < ttm->num_pages; i++) {
 981                 ttm->pages[i] = NULL;
 982                 ttm_dma->dma_address[i] = 0;
 983         }
 984
 985         /* shrink pool if necessary (only on !is_cached pools)*/
 986         if (npages)
 987                 ttm_dma_page_pool_free(pool, npages, GFP_KERNEL);
 988         ttm->state = tt_unpopulated;
 989 }
 990 EXPORT_SYMBOL_GPL(ttm_dma_unpopulate);
 991
 992 /**
 993  * Callback for mm to request pool to reduce number of page held.
 994  *
 995  * XXX: (dchinner) Deadlock warning!
 996  *
 997  * We need to pass sc->gfp_mask to ttm_dma_page_pool_free().
 998  *
 999  * I'm getting sadder as I hear more pathetical whimpers about needing per-pool
1000  * shrinkers
1001  */
1002 static unsigned long
1003 ttm_dma_pool_shrink_scan(struct shrinker *shrink, struct shrink_control *sc)
1004 {
1005         static unsigned start_pool;
1006         unsigned idx = 0;
1007         unsigned pool_offset;
1008         unsigned shrink_pages = sc->nr_to_scan;
1009         struct device_pools *p;
1010         unsigned long freed = 0;
1011
1012         if (list_empty(&_manager->pools))
1013                 return SHRINK_STOP;
1014
1015         if (!mutex_trylock(&_manager->lock))
1016                 return SHRINK_STOP;
1017         if (!_manager->npools)
1018                 goto out;
1019         pool_offset = ++start_pool % _manager->npools;
1020         list_for_each_entry(p, &_manager->pools, pools) {
1021                 unsigned nr_free;
1022
1023                 if (!p->dev)
1024                         continue;
1025                 if (shrink_pages == 0)
1026                         break;
1027                 /* Do it in round-robin fashion. */
1028                 if (++idx < pool_offset)
1029                         continue;
1030                 nr_free = shrink_pages;
1031                 shrink_pages = ttm_dma_page_pool_free(p->pool, nr_free,
1032                                                       sc->gfp_mask);
1033                 freed += nr_free - shrink_pages;
1034
1035                 pr_debug("%s: (%s:%d) Asked to shrink %d, have %d more to go\n",
1036                          p->pool->dev_name, p->pool->name, current->pid,
1037                          nr_free, shrink_pages);
1038         }
1039 out:
1040         mutex_unlock(&_manager->lock);
1041         return freed;
1042 }
1043
1044 static unsigned long
1045 ttm_dma_pool_shrink_count(struct shrinker *shrink, struct shrink_control *sc)
1046 {
1047         struct device_pools *p;
1048         unsigned long count = 0;
1049
1050         if (!mutex_trylock(&_manager->lock))
1051                 return 0;
1052         list_for_each_entry(p, &_manager->pools, pools)
1053                 count += p->pool->npages_free;
1054         mutex_unlock(&_manager->lock);
1055         return count;
1056 }
1057
1058 static void ttm_dma_pool_mm_shrink_init(struct ttm_pool_manager *manager)
1059 {
1060         manager->mm_shrink.count_objects = ttm_dma_pool_shrink_count;
1061         manager->mm_shrink.scan_objects = &ttm_dma_pool_shrink_scan;
1062         manager->mm_shrink.seeks = 1;
1063         register_shrinker(&manager->mm_shrink);
1064 }
1065
1066 static void ttm_dma_pool_mm_shrink_fini(struct ttm_pool_manager *manager)
1067 {
1068         unregister_shrinker(&manager->mm_shrink);
1069 }
1070
1071 int ttm_dma_page_alloc_init(struct ttm_mem_global *glob, unsigned max_pages)
1072 {
1073         int ret = -ENOMEM;
1074
1075         WARN_ON(_manager);
1076
1077         pr_info("Initializing DMA pool allocator\n");
1078
1079         _manager = kzalloc(sizeof(*_manager), GFP_KERNEL);
1080         if (!_manager)
1081                 goto err;
1082
1083         mutex_init(&_manager->lock);
1084         INIT_LIST_HEAD(&_manager->pools);
1085
1086         _manager->options.max_size = max_pages;
1087         _manager->options.small = SMALL_ALLOCATION;
1088         _manager->options.alloc_size = NUM_PAGES_TO_ALLOC;
1089
1090         /* This takes care of auto-freeing the _manager */
1091         ret = kobject_init_and_add(&_manager->kobj, &ttm_pool_kobj_type,
1092                                    &glob->kobj, "dma_pool");
1093         if (unlikely(ret != 0)) {
1094                 kobject_put(&_manager->kobj);
1095                 goto err;
1096         }
1097         ttm_dma_pool_mm_shrink_init(_manager);
1098         return 0;
1099 err:
1100         return ret;
1101 }
1102
1103 void ttm_dma_page_alloc_fini(void)
1104 {
1105         struct device_pools *p, *t;
1106
1107         pr_info("Finalizing DMA pool allocator\n");
1108         ttm_dma_pool_mm_shrink_fini(_manager);
1109
1110         list_for_each_entry_safe_reverse(p, t, &_manager->pools, pools) {
1111                 dev_dbg(p->dev, "(%s:%d) Freeing.\n", p->pool->name,
1112                         current->pid);
1113                 WARN_ON(devres_destroy(p->dev, ttm_dma_pool_release,
1114                         ttm_dma_pool_match, p->pool));
1115                 ttm_dma_free_pool(p->dev, p->pool->type);
1116         }
1117         kobject_put(&_manager->kobj);
1118         _manager = NULL;
1119 }
1120
1121 int ttm_dma_page_alloc_debugfs(struct seq_file *m, void *data)
1122 {
1123         struct device_pools *p;
1124         struct dma_pool *pool = NULL;
1125         char *h[] = {"pool", "refills", "pages freed", "inuse", "available",
1126                      "name", "virt", "busaddr"};
1127
1128         if (!_manager) {
1129                 seq_printf(m, "No pool allocator running.\n");
1130                 return 0;
1131         }
1132         seq_printf(m, "%13s %12s %13s %8s %8s %8s\n",
1133                    h[0], h[1], h[2], h[3], h[4], h[5]);
1134         mutex_lock(&_manager->lock);
1135         list_for_each_entry(p, &_manager->pools, pools) {
1136                 struct device *dev = p->dev;
1137                 if (!dev)
1138                         continue;
1139                 pool = p->pool;
1140                 seq_printf(m, "%13s %12ld %13ld %8d %8d %8s\n",
1141                                 pool->name, pool->nrefills,
1142                                 pool->nfrees, pool->npages_in_use,
1143                                 pool->npages_free,
1144                                 pool->dev_name);
1145         }
1146         mutex_unlock(&_manager->lock);
1147         return 0;
1148 }
1149 EXPORT_SYMBOL_GPL(ttm_dma_page_alloc_debugfs);
1150
1151 #endif