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 #include <linux/dma-mapping.h>
  37 #include <linux/list.h>
  38 #include <linux/seq_file.h> /* for seq_printf */
  39 #include <linux/slab.h>
  40 #include <linux/spinlock.h>
  41 #include <linux/highmem.h>
  42 #include <linux/mm_types.h>
  43 #include <linux/module.h>
  44 #include <linux/mm.h>
  45 #include <linux/atomic.h>
  46 #include <linux/device.h>
  47 #include <linux/kthread.h>
  48 #include "ttm/ttm_bo_driver.h"
  49 #include "ttm/ttm_page_alloc.h"
  50 #ifdef TTM_HAS_AGP
  51 #include <asm/agp.h>
  52 #endif
  53
  54 #define NUM_PAGES_TO_ALLOC              (PAGE_SIZE/sizeof(struct page *))
  55 #define SMALL_ALLOCATION                4
  56 #define FREE_ALL_PAGES                  (~0U)
  57 /* times are in msecs */
  58 #define IS_UNDEFINED                    (0)
  59 #define IS_WC                           (1<<1)
  60 #define IS_UC                           (1<<2)
  61 #define IS_CACHED                       (1<<3)
  62 #define IS_DMA32                        (1<<4)
  63
  64 enum pool_type {
  65         POOL_IS_UNDEFINED,
  66         POOL_IS_WC = IS_WC,
  67         POOL_IS_UC = IS_UC,
  68         POOL_IS_CACHED = IS_CACHED,
  69         POOL_IS_WC_DMA32 = IS_WC | IS_DMA32,
  70         POOL_IS_UC_DMA32 = IS_UC | IS_DMA32,
  71         POOL_IS_CACHED_DMA32 = IS_CACHED | IS_DMA32,
  72 };
  73 /*
  74  * The pool structure. There are usually six pools:
  75  *  - generic (not restricted to DMA32):
  76  *      - write combined, uncached, cached.
  77  *  - dma32 (up to 2^32 - so up 4GB):
  78  *      - write combined, uncached, cached.
  79  * for each 'struct device'. The 'cached' is for pages that are actively used.
  80  * The other ones can be shrunk by the shrinker API if neccessary.
  81  * @pools: The 'struct device->dma_pools' link.
  82  * @type: Type of the pool
  83  * @lock: Protects the inuse_list and free_list from concurrnet access. Must be
  84  * used with irqsave/irqrestore variants because pool allocator maybe called
  85  * from delayed work.
  86  * @inuse_list: Pool of pages that are in use. The order is very important and
  87  *   it is in the order that the TTM pages that are put back are in.
  88  * @free_list: Pool of pages that are free to be used. No order requirements.
  89  * @dev: The device that is associated with these pools.
  90  * @size: Size used during DMA allocation.
  91  * @npages_free: Count of available pages for re-use.
  92  * @npages_in_use: Count of pages that are in use.
  93  * @nfrees: Stats when pool is shrinking.
  94  * @nrefills: Stats when the pool is grown.
  95  * @gfp_flags: Flags to pass for alloc_page.
  96  * @name: Name of the pool.
  97  * @dev_name: Name derieved from dev - similar to how dev_info works.
  98  *   Used during shutdown as the dev_info during release is unavailable.
  99  */
 100 struct dma_pool {
 101         struct list_head pools; /* The 'struct device->dma_pools link */
 102         enum pool_type type;
 103         spinlock_t lock;
 104         struct list_head inuse_list;
 105         struct list_head free_list;
 106         struct device *dev;
 107         unsigned size;
 108         unsigned npages_free;
 109         unsigned npages_in_use;
 110         unsigned long nfrees; /* Stats when shrunk. */
 111         unsigned long nrefills; /* Stats when grown. */
 112         gfp_t gfp_flags;
 113         char name[13]; /* "cached dma32" */
 114         char dev_name[64]; /* Constructed from dev */
 115 };
 116
 117 /*
 118  * The accounting page keeping track of the allocated page along with
 119  * the DMA address.
 120  * @page_list: The link to the 'page_list' in 'struct dma_pool'.
 121  * @vaddr: The virtual address of the page
 122  * @dma: The bus address of the page. If the page is not allocated
 123  *   via the DMA API, it will be -1.
 124  */
 125 struct dma_page {
 126         struct list_head page_list;
 127         void *vaddr;
 128         struct page *p;
 129         dma_addr_t dma;
 130 };
 131
 132 /*
 133  * Limits for the pool. They are handled without locks because only place where
 134  * they may change is in sysfs store. They won't have immediate effect anyway
 135  * so forcing serialization to access them is pointless.
 136  */
 137
 138 struct ttm_pool_opts {
 139         unsigned        alloc_size;
 140         unsigned        max_size;
 141         unsigned        small;
 142 };
 143
 144 /*
 145  * Contains the list of all of the 'struct device' and their corresponding
 146  * DMA pools. Guarded by _mutex->lock.
 147  * @pools: The link to 'struct ttm_pool_manager->pools'
 148  * @dev: The 'struct device' associated with the 'pool'
 149  * @pool: The 'struct dma_pool' associated with the 'dev'
 150  */
 151 struct device_pools {
 152         struct list_head pools;
 153         struct device *dev;
 154         struct dma_pool *pool;
 155 };
 156
 157 /*
 158  * struct ttm_pool_manager - Holds memory pools for fast allocation
 159  *
 160  * @lock: Lock used when adding/removing from pools
 161  * @pools: List of 'struct device' and 'struct dma_pool' tuples.
 162  * @options: Limits for the pool.
 163  * @npools: Total amount of pools in existence.
 164  * @shrinker: The structure used by [un|]register_shrinker
 165  */
 166 struct ttm_pool_manager {
 167         struct mutex            lock;
 168         struct list_head        pools;
 169         struct ttm_pool_opts    options;
 170         unsigned                npools;
 171         struct shrinker         mm_shrink;
 172         struct kobject          kobj;
 173 };
 174
 175 static struct ttm_pool_manager *_manager;
 176
 177 static struct attribute ttm_page_pool_max = {
 178         .name = "pool_max_size",
 179         .mode = S_IRUGO | S_IWUSR
 180 };
 181 static struct attribute ttm_page_pool_small = {
 182         .name = "pool_small_allocation",
 183         .mode = S_IRUGO | S_IWUSR
 184 };
 185 static struct attribute ttm_page_pool_alloc_size = {
 186         .name = "pool_allocation_size",
 187         .mode = S_IRUGO | S_IWUSR
 188 };
 189
 190 static struct attribute *ttm_pool_attrs[] = {
 191         &ttm_page_pool_max,
 192         &ttm_page_pool_small,
 193         &ttm_page_pool_alloc_size,
 194         NULL
 195 };
 196
 197 static void ttm_pool_kobj_release(struct kobject *kobj)
 198 {
 199         struct ttm_pool_manager *m =
 200                 container_of(kobj, struct ttm_pool_manager, kobj);
 201         kfree(m);
 202 }
 203
 204 static ssize_t ttm_pool_store(struct kobject *kobj, struct attribute *attr,
 205                               const char *buffer, size_t size)
 206 {
 207         struct ttm_pool_manager *m =
 208                 container_of(kobj, struct ttm_pool_manager, kobj);
 209         int chars;
 210         unsigned val;
 211         chars = sscanf(buffer, "%u", &val);
 212         if (chars == 0)
 213                 return size;
 214
 215         /* Convert kb to number of pages */
 216         val = val / (PAGE_SIZE >> 10);
 217
 218         if (attr == &ttm_page_pool_max)
 219                 m->options.max_size = val;
 220         else if (attr == &ttm_page_pool_small)
 221                 m->options.small = val;
 222         else if (attr == &ttm_page_pool_alloc_size) {
 223                 if (val > NUM_PAGES_TO_ALLOC*8) {
 224                         printk(KERN_ERR TTM_PFX
 225                                "Setting allocation size to %lu "
 226                                "is not allowed. Recommended size is "
 227                                "%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                         printk(KERN_WARNING TTM_PFX
 233                                "Setting allocation size to "
 234                                "larger than %lu is not recommended.\n",
 235                                NUM_PAGES_TO_ALLOC*(PAGE_SIZE >> 10));
 236                 }
 237                 m->options.alloc_size = val;
 238         }
 239
 240         return size;
 241 }
 242
 243 static ssize_t ttm_pool_show(struct kobject *kobj, struct attribute *attr,
 244                              char *buffer)
 245 {
 246         struct ttm_pool_manager *m =
 247                 container_of(kobj, struct ttm_pool_manager, kobj);
 248         unsigned val = 0;
 249
 250         if (attr == &ttm_page_pool_max)
 251                 val = m->options.max_size;
 252         else if (attr == &ttm_page_pool_small)
 253                 val = m->options.small;
 254         else if (attr == &ttm_page_pool_alloc_size)
 255                 val = m->options.alloc_size;
 256
 257         val = val * (PAGE_SIZE >> 10);
 258
 259         return snprintf(buffer, PAGE_SIZE, "%u\n", val);
 260 }
 261
 262 static const struct sysfs_ops ttm_pool_sysfs_ops = {
 263         .show = &ttm_pool_show,
 264         .store = &ttm_pool_store,
 265 };
 266
 267 static struct kobj_type ttm_pool_kobj_type = {
 268         .release = &ttm_pool_kobj_release,
 269         .sysfs_ops = &ttm_pool_sysfs_ops,
 270         .default_attrs = ttm_pool_attrs,
 271 };
 272
 273 #ifndef CONFIG_X86
 274 static int set_pages_array_wb(struct page **pages, int addrinarray)
 275 {
 276 #ifdef TTM_HAS_AGP
 277         int i;
 278
 279         for (i = 0; i < addrinarray; i++)
 280                 unmap_page_from_agp(pages[i]);
 281 #endif
 282         return 0;
 283 }
 284
 285 static int set_pages_array_wc(struct page **pages, int addrinarray)
 286 {
 287 #ifdef TTM_HAS_AGP
 288         int i;
 289
 290         for (i = 0; i < addrinarray; i++)
 291                 map_page_into_agp(pages[i]);
 292 #endif
 293         return 0;
 294 }
 295
 296 static int set_pages_array_uc(struct page **pages, int addrinarray)
 297 {
 298 #ifdef TTM_HAS_AGP
 299         int i;
 300
 301         for (i = 0; i < addrinarray; i++)
 302                 map_page_into_agp(pages[i]);
 303 #endif
 304         return 0;
 305 }
 306 #endif /* for !CONFIG_X86 */
 307
 308 static int ttm_set_pages_caching(struct dma_pool *pool,
 309                                  struct page **pages, unsigned cpages)
 310 {
 311         int r = 0;
 312         /* Set page caching */
 313         if (pool->type & IS_UC) {
 314                 r = set_pages_array_uc(pages, cpages);
 315                 if (r)
 316                         pr_err(TTM_PFX
 317                                "%s: Failed to set %d pages to uc!\n",
 318                                pool->dev_name, cpages);
 319         }
 320         if (pool->type & IS_WC) {
 321                 r = set_pages_array_wc(pages, cpages);
 322                 if (r)
 323                         pr_err(TTM_PFX
 324                                "%s: Failed to set %d pages to wc!\n",
 325                                pool->dev_name, cpages);
 326         }
 327         return r;
 328 }
 329
 330 static void __ttm_dma_free_page(struct dma_pool *pool, struct dma_page *d_page)
 331 {
 332         dma_addr_t dma = d_page->dma;
 333         dma_free_coherent(pool->dev, pool->size, d_page->vaddr, dma);
 334
 335         kfree(d_page);
 336         d_page = NULL;
 337 }
 338 static struct dma_page *__ttm_dma_alloc_page(struct dma_pool *pool)
 339 {
 340         struct dma_page *d_page;
 341
 342         d_page = kmalloc(sizeof(struct dma_page), GFP_KERNEL);
 343         if (!d_page)
 344                 return NULL;
 345
 346         d_page->vaddr = dma_alloc_coherent(pool->dev, pool->size,
 347                                            &d_page->dma,
 348                                            pool->gfp_flags);
 349         if (d_page->vaddr)
 350                 d_page->p = virt_to_page(d_page->vaddr);
 351         else {
 352                 kfree(d_page);
 353                 d_page = NULL;
 354         }
 355         return d_page;
 356 }
 357 static enum pool_type ttm_to_type(int flags, enum ttm_caching_state cstate)
 358 {
 359         enum pool_type type = IS_UNDEFINED;
 360
 361         if (flags & TTM_PAGE_FLAG_DMA32)
 362                 type |= IS_DMA32;
 363         if (cstate == tt_cached)
 364                 type |= IS_CACHED;
 365         else if (cstate == tt_uncached)
 366                 type |= IS_UC;
 367         else
 368                 type |= IS_WC;
 369
 370         return type;
 371 }
 372
 373 static void ttm_pool_update_free_locked(struct dma_pool *pool,
 374                                         unsigned freed_pages)
 375 {
 376         pool->npages_free -= freed_pages;
 377         pool->nfrees += freed_pages;
 378
 379 }
 380
 381 /* set memory back to wb and free the pages. */
 382 static void ttm_dma_pages_put(struct dma_pool *pool, struct list_head *d_pages,
 383                               struct page *pages[], unsigned npages)
 384 {
 385         struct dma_page *d_page, *tmp;
 386
 387         /* Don't set WB on WB page pool. */
 388         if (npages && !(pool->type & IS_CACHED) &&
 389             set_pages_array_wb(pages, npages))
 390                 pr_err(TTM_PFX "%s: Failed to set %d pages to wb!\n",
 391                         pool->dev_name, npages);
 392
 393         list_for_each_entry_safe(d_page, tmp, d_pages, page_list) {
 394                 list_del(&d_page->page_list);
 395                 __ttm_dma_free_page(pool, d_page);
 396         }
 397 }
 398
 399 static void ttm_dma_page_put(struct dma_pool *pool, struct dma_page *d_page)
 400 {
 401         /* Don't set WB on WB page pool. */
 402         if (!(pool->type & IS_CACHED) && set_pages_array_wb(&d_page->p, 1))
 403                 pr_err(TTM_PFX "%s: Failed to set %d pages to wb!\n",
 404                         pool->dev_name, 1);
 405
 406         list_del(&d_page->page_list);
 407         __ttm_dma_free_page(pool, d_page);
 408 }
 409
 410 /*
 411  * Free pages from pool.
 412  *
 413  * To prevent hogging the ttm_swap process we only free NUM_PAGES_TO_ALLOC
 414  * number of pages in one go.
 415  *
 416  * @pool: to free the pages from
 417  * @nr_free: If set to true will free all pages in pool
 418  **/
 419 static unsigned ttm_dma_page_pool_free(struct dma_pool *pool, unsigned nr_free)
 420 {
 421         unsigned long irq_flags;
 422         struct dma_page *dma_p, *tmp;
 423         struct page **pages_to_free;
 424         struct list_head d_pages;
 425         unsigned freed_pages = 0,
 426                  npages_to_free = nr_free;
 427
 428         if (NUM_PAGES_TO_ALLOC < nr_free)
 429                 npages_to_free = NUM_PAGES_TO_ALLOC;
 430 #if 0
 431         if (nr_free > 1) {
 432                 pr_debug("%s: (%s:%d) Attempting to free %d (%d) pages\n",
 433                         pool->dev_name, pool->name, current->pid,
 434                         npages_to_free, nr_free);
 435         }
 436 #endif
 437         pages_to_free = kmalloc(npages_to_free * sizeof(struct page *),
 438                         GFP_KERNEL);
 439
 440         if (!pages_to_free) {
 441                 pr_err(TTM_PFX
 442                        "%s: Failed to allocate memory for pool free operation.\n",
 443                         pool->dev_name);
 444                 return 0;
 445         }
 446         INIT_LIST_HEAD(&d_pages);
 447 restart:
 448         spin_lock_irqsave(&pool->lock, irq_flags);
 449
 450         /* We picking the oldest ones off the list */
 451         list_for_each_entry_safe_reverse(dma_p, tmp, &pool->free_list,
 452                                          page_list) {
 453                 if (freed_pages >= npages_to_free)
 454                         break;
 455
 456                 /* Move the dma_page from one list to another. */
 457                 list_move(&dma_p->page_list, &d_pages);
 458
 459                 pages_to_free[freed_pages++] = dma_p->p;
 460                 /* We can only remove NUM_PAGES_TO_ALLOC at a time. */
 461                 if (freed_pages >= NUM_PAGES_TO_ALLOC) {
 462
 463                         ttm_pool_update_free_locked(pool, freed_pages);
 464                         /**
 465                          * Because changing page caching is costly
 466                          * we unlock the pool to prevent stalling.
 467                          */
 468                         spin_unlock_irqrestore(&pool->lock, irq_flags);
 469
 470                         ttm_dma_pages_put(pool, &d_pages, pages_to_free,
 471                                           freed_pages);
 472
 473                         INIT_LIST_HEAD(&d_pages);
 474
 475                         if (likely(nr_free != FREE_ALL_PAGES))
 476                                 nr_free -= freed_pages;
 477
 478                         if (NUM_PAGES_TO_ALLOC >= nr_free)
 479                                 npages_to_free = nr_free;
 480                         else
 481                                 npages_to_free = NUM_PAGES_TO_ALLOC;
 482
 483                         freed_pages = 0;
 484
 485                         /* free all so restart the processing */
 486                         if (nr_free)
 487                                 goto restart;
 488
 489                         /* Not allowed to fall through or break because
 490                          * following context is inside spinlock while we are
 491                          * outside here.
 492                          */
 493                         goto out;
 494
 495                 }
 496         }
 497
 498         /* remove range of pages from the pool */
 499         if (freed_pages) {
 500                 ttm_pool_update_free_locked(pool, freed_pages);
 501                 nr_free -= freed_pages;
 502         }
 503
 504         spin_unlock_irqrestore(&pool->lock, irq_flags);
 505
 506         if (freed_pages)
 507                 ttm_dma_pages_put(pool, &d_pages, pages_to_free, freed_pages);
 508 out:
 509         kfree(pages_to_free);
 510         return nr_free;
 511 }
 512
 513 static void ttm_dma_free_pool(struct device *dev, enum pool_type type)
 514 {
 515         struct device_pools *p;
 516         struct dma_pool *pool;
 517
 518         if (!dev)
 519                 return;
 520
 521         mutex_lock(&_manager->lock);
 522         list_for_each_entry_reverse(p, &_manager->pools, pools) {
 523                 if (p->dev != dev)
 524                         continue;
 525                 pool = p->pool;
 526                 if (pool->type != type)
 527                         continue;
 528
 529                 list_del(&p->pools);
 530                 kfree(p);
 531                 _manager->npools--;
 532                 break;
 533         }
 534         list_for_each_entry_reverse(pool, &dev->dma_pools, pools) {
 535                 if (pool->type != type)
 536                         continue;
 537                 /* Takes a spinlock.. */
 538                 ttm_dma_page_pool_free(pool, FREE_ALL_PAGES);
 539                 WARN_ON(((pool->npages_in_use + pool->npages_free) != 0));
 540                 /* This code path is called after _all_ references to the
 541                  * struct device has been dropped - so nobody should be
 542                  * touching it. In case somebody is trying to _add_ we are
 543                  * guarded by the mutex. */
 544                 list_del(&pool->pools);
 545                 kfree(pool);
 546                 break;
 547         }
 548         mutex_unlock(&_manager->lock);
 549 }
 550
 551 /*
 552  * On free-ing of the 'struct device' this deconstructor is run.
 553  * Albeit the pool might have already been freed earlier.
 554  */
 555 static void ttm_dma_pool_release(struct device *dev, void *res)
 556 {
 557         struct dma_pool *pool = *(struct dma_pool **)res;
 558
 559         if (pool)
 560                 ttm_dma_free_pool(dev, pool->type);
 561 }
 562
 563 static int ttm_dma_pool_match(struct device *dev, void *res, void *match_data)
 564 {
 565         return *(struct dma_pool **)res == match_data;
 566 }
 567
 568 static struct dma_pool *ttm_dma_pool_init(struct device *dev, gfp_t flags,
 569                                           enum pool_type type)
 570 {
 571         char *n[] = {"wc", "uc", "cached", " dma32", "unknown",};
 572         enum pool_type t[] = {IS_WC, IS_UC, IS_CACHED, IS_DMA32, IS_UNDEFINED};
 573         struct device_pools *sec_pool = NULL;
 574         struct dma_pool *pool = NULL, **ptr;
 575         unsigned i;
 576         int ret = -ENODEV;
 577         char *p;
 578
 579         if (!dev)
 580                 return NULL;
 581
 582         ptr = devres_alloc(ttm_dma_pool_release, sizeof(*ptr), GFP_KERNEL);
 583         if (!ptr)
 584                 return NULL;
 585
 586         ret = -ENOMEM;
 587
 588         pool = kmalloc_node(sizeof(struct dma_pool), GFP_KERNEL,
 589                             dev_to_node(dev));
 590         if (!pool)
 591                 goto err_mem;
 592
 593         sec_pool = kmalloc_node(sizeof(struct device_pools), GFP_KERNEL,
 594                                 dev_to_node(dev));
 595         if (!sec_pool)
 596                 goto err_mem;
 597
 598         INIT_LIST_HEAD(&sec_pool->pools);
 599         sec_pool->dev = dev;
 600         sec_pool->pool =  pool;
 601
 602         INIT_LIST_HEAD(&pool->free_list);
 603         INIT_LIST_HEAD(&pool->inuse_list);
 604         INIT_LIST_HEAD(&pool->pools);
 605         spin_lock_init(&pool->lock);
 606         pool->dev = dev;
 607         pool->npages_free = pool->npages_in_use = 0;
 608         pool->nfrees = 0;
 609         pool->gfp_flags = flags;
 610         pool->size = PAGE_SIZE;
 611         pool->type = type;
 612         pool->nrefills = 0;
 613         p = pool->name;
 614         for (i = 0; i < 5; i++) {
 615                 if (type & t[i]) {
 616                         p += snprintf(p, sizeof(pool->name) - (p - pool->name),
 617                                       "%s", n[i]);
 618                 }
 619         }
 620         *p = 0;
 621         /* We copy the name for pr_ calls b/c when dma_pool_destroy is called
 622          * - the kobj->name has already been deallocated.*/
 623         snprintf(pool->dev_name, sizeof(pool->dev_name), "%s %s",
 624                  dev_driver_string(dev), dev_name(dev));
 625         mutex_lock(&_manager->lock);
 626         /* You can get the dma_pool from either the global: */
 627         list_add(&sec_pool->pools, &_manager->pools);
 628         _manager->npools++;
 629         /* or from 'struct device': */
 630         list_add(&pool->pools, &dev->dma_pools);
 631         mutex_unlock(&_manager->lock);
 632
 633         *ptr = pool;
 634         devres_add(dev, ptr);
 635
 636         return pool;
 637 err_mem:
 638         devres_free(ptr);
 639         kfree(sec_pool);
 640         kfree(pool);
 641         return ERR_PTR(ret);
 642 }
 643
 644 static struct dma_pool *ttm_dma_find_pool(struct device *dev,
 645                                           enum pool_type type)
 646 {
 647         struct dma_pool *pool, *tmp, *found = NULL;
 648
 649         if (type == IS_UNDEFINED)
 650                 return found;
 651
 652         /* NB: We iterate on the 'struct dev' which has no spinlock, but
 653          * it does have a kref which we have taken. The kref is taken during
 654          * graphic driver loading - in the drm_pci_init it calls either
 655          * pci_dev_get or pci_register_driver which both end up taking a kref
 656          * on 'struct device'.
 657          *
 658          * On teardown, the graphic drivers end up quiescing the TTM (put_pages)
 659          * and calls the dev_res deconstructors: ttm_dma_pool_release. The nice
 660          * thing is at that point of time there are no pages associated with the
 661          * driver so this function will not be called.
 662          */
 663         list_for_each_entry_safe(pool, tmp, &dev->dma_pools, pools) {
 664                 if (pool->type != type)
 665                         continue;
 666                 found = pool;
 667                 break;
 668         }
 669         return found;
 670 }
 671
 672 /*
 673  * Free pages the pages that failed to change the caching state. If there
 674  * are pages that have changed their caching state already put them to the
 675  * pool.
 676  */
 677 static void ttm_dma_handle_caching_state_failure(struct dma_pool *pool,
 678                                                  struct list_head *d_pages,
 679                                                  struct page **failed_pages,
 680                                                  unsigned cpages)
 681 {
 682         struct dma_page *d_page, *tmp;
 683         struct page *p;
 684         unsigned i = 0;
 685
 686         p = failed_pages[0];
 687         if (!p)
 688                 return;
 689         /* Find the failed page. */
 690         list_for_each_entry_safe(d_page, tmp, d_pages, page_list) {
 691                 if (d_page->p != p)
 692                         continue;
 693                 /* .. and then progress over the full list. */
 694                 list_del(&d_page->page_list);
 695                 __ttm_dma_free_page(pool, d_page);
 696                 if (++i < cpages)
 697                         p = failed_pages[i];
 698                 else
 699                         break;
 700         }
 701
 702 }
 703
 704 /*
 705  * Allocate 'count' pages, and put 'need' number of them on the
 706  * 'pages' and as well on the 'dma_address' starting at 'dma_offset' offset.
 707  * The full list of pages should also be on 'd_pages'.
 708  * We return zero for success, and negative numbers as errors.
 709  */
 710 static int ttm_dma_pool_alloc_new_pages(struct dma_pool *pool,
 711                                         struct list_head *d_pages,
 712                                         unsigned count)
 713 {
 714         struct page **caching_array;
 715         struct dma_page *dma_p;
 716         struct page *p;
 717         int r = 0;
 718         unsigned i, cpages;
 719         unsigned max_cpages = min(count,
 720                         (unsigned)(PAGE_SIZE/sizeof(struct page *)));
 721
 722         /* allocate array for page caching change */
 723         caching_array = kmalloc(max_cpages*sizeof(struct page *), GFP_KERNEL);
 724
 725         if (!caching_array) {
 726                 pr_err(TTM_PFX
 727                        "%s: Unable to allocate table for new pages.",
 728                         pool->dev_name);
 729                 return -ENOMEM;
 730         }
 731
 732         if (count > 1) {
 733                 pr_debug("%s: (%s:%d) Getting %d pages\n",
 734                         pool->dev_name, pool->name, current->pid,
 735                         count);
 736         }
 737
 738         for (i = 0, cpages = 0; i < count; ++i) {
 739                 dma_p = __ttm_dma_alloc_page(pool);
 740                 if (!dma_p) {
 741                         pr_err(TTM_PFX "%s: Unable to get page %u.\n",
 742                                 pool->dev_name, i);
 743
 744                         /* store already allocated pages in the pool after
 745                          * setting the caching state */
 746                         if (cpages) {
 747                                 r = ttm_set_pages_caching(pool, caching_array,
 748                                                           cpages);
 749                                 if (r)
 750                                         ttm_dma_handle_caching_state_failure(
 751                                                 pool, d_pages, caching_array,
 752                                                 cpages);
 753                         }
 754                         r = -ENOMEM;
 755                         goto out;
 756                 }
 757                 p = dma_p->p;
 758 #ifdef CONFIG_HIGHMEM
 759                 /* gfp flags of highmem page should never be dma32 so we
 760                  * we should be fine in such case
 761                  */
 762                 if (!PageHighMem(p))
 763 #endif
 764                 {
 765                         caching_array[cpages++] = p;
 766                         if (cpages == max_cpages) {
 767                                 /* Note: Cannot hold the spinlock */
 768                                 r = ttm_set_pages_caching(pool, caching_array,
 769                                                  cpages);
 770                                 if (r) {
 771                                         ttm_dma_handle_caching_state_failure(
 772                                                 pool, d_pages, caching_array,
 773                                                 cpages);
 774                                         goto out;
 775                                 }
 776                                 cpages = 0;
 777                         }
 778                 }
 779                 list_add(&dma_p->page_list, d_pages);
 780         }
 781
 782         if (cpages) {
 783                 r = ttm_set_pages_caching(pool, caching_array, cpages);
 784                 if (r)
 785                         ttm_dma_handle_caching_state_failure(pool, d_pages,
 786                                         caching_array, cpages);
 787         }
 788 out:
 789         kfree(caching_array);
 790         return r;
 791 }
 792
 793 /*
 794  * @return count of pages still required to fulfill the request.
 795  */
 796 static int ttm_dma_page_pool_fill_locked(struct dma_pool *pool,
 797                                          unsigned long *irq_flags)
 798 {
 799         unsigned count = _manager->options.small;
 800         int r = pool->npages_free;
 801
 802         if (count > pool->npages_free) {
 803                 struct list_head d_pages;
 804
 805                 INIT_LIST_HEAD(&d_pages);
 806
 807                 spin_unlock_irqrestore(&pool->lock, *irq_flags);
 808
 809                 /* Returns how many more are neccessary to fulfill the
 810                  * request. */
 811                 r = ttm_dma_pool_alloc_new_pages(pool, &d_pages, count);
 812
 813                 spin_lock_irqsave(&pool->lock, *irq_flags);
 814                 if (!r) {
 815                         /* Add the fresh to the end.. */
 816                         list_splice(&d_pages, &pool->free_list);
 817                         ++pool->nrefills;
 818                         pool->npages_free += count;
 819                         r = count;
 820                 } else {
 821                         struct dma_page *d_page;
 822                         unsigned cpages = 0;
 823
 824                         pr_err(TTM_PFX "%s: Failed to fill %s pool (r:%d)!\n",
 825                                 pool->dev_name, pool->name, r);
 826
 827                         list_for_each_entry(d_page, &d_pages, page_list) {
 828                                 cpages++;
 829                         }
 830                         list_splice_tail(&d_pages, &pool->free_list);
 831                         pool->npages_free += cpages;
 832                         r = cpages;
 833                 }
 834         }
 835         return r;
 836 }
 837
 838 /*
 839  * @return count of pages still required to fulfill the request.
 840  * The populate list is actually a stack (not that is matters as TTM
 841  * allocates one page at a time.
 842  */
 843 static int ttm_dma_pool_get_pages(struct dma_pool *pool,
 844                                   struct ttm_dma_tt *ttm_dma,
 845                                   unsigned index)
 846 {
 847         struct dma_page *d_page;
 848         struct ttm_tt *ttm = &ttm_dma->ttm;
 849         unsigned long irq_flags;
 850         int count, r = -ENOMEM;
 851
 852         spin_lock_irqsave(&pool->lock, irq_flags);
 853         count = ttm_dma_page_pool_fill_locked(pool, &irq_flags);
 854         if (count) {
 855                 d_page = list_first_entry(&pool->free_list, struct dma_page, page_list);
 856                 ttm->pages[index] = d_page->p;
 857                 ttm_dma->dma_address[index] = d_page->dma;
 858                 list_move_tail(&d_page->page_list, &ttm_dma->pages_list);
 859                 r = 0;
 860                 pool->npages_in_use += 1;
 861                 pool->npages_free -= 1;
 862         }
 863         spin_unlock_irqrestore(&pool->lock, irq_flags);
 864         return r;
 865 }
 866
 867 /*
 868  * On success pages list will hold count number of correctly
 869  * cached pages. On failure will hold the negative return value (-ENOMEM, etc).
 870  */
 871 int ttm_dma_populate(struct ttm_dma_tt *ttm_dma, struct device *dev)
 872 {
 873         struct ttm_tt *ttm = &ttm_dma->ttm;
 874         struct ttm_mem_global *mem_glob = ttm->glob->mem_glob;
 875         struct dma_pool *pool;
 876         enum pool_type type;
 877         unsigned i;
 878         gfp_t gfp_flags;
 879         int ret;
 880
 881         if (ttm->state != tt_unpopulated)
 882                 return 0;
 883
 884         type = ttm_to_type(ttm->page_flags, ttm->caching_state);
 885         if (ttm->page_flags & TTM_PAGE_FLAG_DMA32)
 886                 gfp_flags = GFP_USER | GFP_DMA32;
 887         else
 888                 gfp_flags = GFP_HIGHUSER;
 889         if (ttm->page_flags & TTM_PAGE_FLAG_ZERO_ALLOC)
 890                 gfp_flags |= __GFP_ZERO;
 891
 892         pool = ttm_dma_find_pool(dev, type);
 893         if (!pool) {
 894                 pool = ttm_dma_pool_init(dev, gfp_flags, type);
 895                 if (IS_ERR_OR_NULL(pool)) {
 896                         return -ENOMEM;
 897                 }
 898         }
 899
 900         INIT_LIST_HEAD(&ttm_dma->pages_list);
 901         for (i = 0; i < ttm->num_pages; ++i) {
 902                 ret = ttm_dma_pool_get_pages(pool, ttm_dma, i);
 903                 if (ret != 0) {
 904                         ttm_dma_unpopulate(ttm_dma, dev);
 905                         return -ENOMEM;
 906                 }
 907
 908                 ret = ttm_mem_global_alloc_page(mem_glob, ttm->pages[i],
 909                                                 false, false);
 910                 if (unlikely(ret != 0)) {
 911                         ttm_dma_unpopulate(ttm_dma, dev);
 912                         return -ENOMEM;
 913                 }
 914         }
 915
 916         if (unlikely(ttm->page_flags & TTM_PAGE_FLAG_SWAPPED)) {
 917                 ret = ttm_tt_swapin(ttm);
 918                 if (unlikely(ret != 0)) {
 919                         ttm_dma_unpopulate(ttm_dma, dev);
 920                         return ret;
 921                 }
 922         }
 923
 924         ttm->state = tt_unbound;
 925         return 0;
 926 }
 927 EXPORT_SYMBOL_GPL(ttm_dma_populate);
 928
 929 /* Get good estimation how many pages are free in pools */
 930 static int ttm_dma_pool_get_num_unused_pages(void)
 931 {
 932         struct device_pools *p;
 933         unsigned total = 0;
 934
 935         mutex_lock(&_manager->lock);
 936         list_for_each_entry(p, &_manager->pools, pools)
 937                 total += p->pool->npages_free;
 938         mutex_unlock(&_manager->lock);
 939         return total;
 940 }
 941
 942 /* Put all pages in pages list to correct pool to wait for reuse */
 943 void ttm_dma_unpopulate(struct ttm_dma_tt *ttm_dma, struct device *dev)
 944 {
 945         struct ttm_tt *ttm = &ttm_dma->ttm;
 946         struct dma_pool *pool;
 947         struct dma_page *d_page, *next;
 948         enum pool_type type;
 949         bool is_cached = false;
 950         unsigned count = 0, i, npages = 0;
 951         unsigned long irq_flags;
 952
 953         type = ttm_to_type(ttm->page_flags, ttm->caching_state);
 954         pool = ttm_dma_find_pool(dev, type);
 955         if (!pool)
 956                 return;
 957
 958         is_cached = (ttm_dma_find_pool(pool->dev,
 959                      ttm_to_type(ttm->page_flags, tt_cached)) == pool);
 960
 961         /* make sure pages array match list and count number of pages */
 962         list_for_each_entry(d_page, &ttm_dma->pages_list, page_list) {
 963                 ttm->pages[count] = d_page->p;
 964                 count++;
 965         }
 966
 967         spin_lock_irqsave(&pool->lock, irq_flags);
 968         pool->npages_in_use -= count;
 969         if (is_cached) {
 970                 pool->nfrees += count;
 971         } else {
 972                 pool->npages_free += count;
 973                 list_splice(&ttm_dma->pages_list, &pool->free_list);
 974                 npages = count;
 975                 if (pool->npages_free > _manager->options.max_size) {
 976                         npages = pool->npages_free - _manager->options.max_size;
 977                         /* free at least NUM_PAGES_TO_ALLOC number of pages
 978                          * to reduce calls to set_memory_wb */
 979                         if (npages < NUM_PAGES_TO_ALLOC)
 980                                 npages = NUM_PAGES_TO_ALLOC;
 981                 }
 982         }
 983         spin_unlock_irqrestore(&pool->lock, irq_flags);
 984
 985         if (is_cached) {
 986                 list_for_each_entry_safe(d_page, next, &ttm_dma->pages_list, page_list) {
 987                         ttm_mem_global_free_page(ttm->glob->mem_glob,
 988                                                  d_page->p);
 989                         ttm_dma_page_put(pool, d_page);
 990                 }
 991         } else {
 992                 for (i = 0; i < count; i++) {
 993                         ttm_mem_global_free_page(ttm->glob->mem_glob,
 994                                                  ttm->pages[i]);
 995                 }
 996         }
 997
 998         INIT_LIST_HEAD(&ttm_dma->pages_list);
 999         for (i = 0; i < ttm->num_pages; i++) {
1000                 ttm->pages[i] = NULL;
1001                 ttm_dma->dma_address[i] = 0;
1002         }
1003
1004         /* shrink pool if necessary (only on !is_cached pools)*/
1005         if (npages)
1006                 ttm_dma_page_pool_free(pool, npages);
1007         ttm->state = tt_unpopulated;
1008 }
1009 EXPORT_SYMBOL_GPL(ttm_dma_unpopulate);
1010
1011 /**
1012  * Callback for mm to request pool to reduce number of page held.
1013  */
1014 static int ttm_dma_pool_mm_shrink(struct shrinker *shrink,
1015                                   struct shrink_control *sc)
1016 {
1017         static atomic_t start_pool = ATOMIC_INIT(0);
1018         unsigned idx = 0;
1019         unsigned pool_offset = atomic_add_return(1, &start_pool);
1020         unsigned shrink_pages = sc->nr_to_scan;
1021         struct device_pools *p;
1022
1023         if (list_empty(&_manager->pools))
1024                 return 0;
1025
1026         mutex_lock(&_manager->lock);
1027         pool_offset = pool_offset % _manager->npools;
1028         list_for_each_entry(p, &_manager->pools, pools) {
1029                 unsigned nr_free;
1030
1031                 if (!p->dev)
1032                         continue;
1033                 if (shrink_pages == 0)
1034                         break;
1035                 /* Do it in round-robin fashion. */
1036                 if (++idx < pool_offset)
1037                         continue;
1038                 nr_free = shrink_pages;
1039                 shrink_pages = ttm_dma_page_pool_free(p->pool, nr_free);
1040                 pr_debug("%s: (%s:%d) Asked to shrink %d, have %d more to go\n",
1041                         p->pool->dev_name, p->pool->name, current->pid, nr_free,
1042                         shrink_pages);
1043         }
1044         mutex_unlock(&_manager->lock);
1045         /* return estimated number of unused pages in pool */
1046         return ttm_dma_pool_get_num_unused_pages();
1047 }
1048
1049 static void ttm_dma_pool_mm_shrink_init(struct ttm_pool_manager *manager)
1050 {
1051         manager->mm_shrink.shrink = &ttm_dma_pool_mm_shrink;
1052         manager->mm_shrink.seeks = 1;
1053         register_shrinker(&manager->mm_shrink);
1054 }
1055
1056 static void ttm_dma_pool_mm_shrink_fini(struct ttm_pool_manager *manager)
1057 {
1058         unregister_shrinker(&manager->mm_shrink);
1059 }
1060
1061 int ttm_dma_page_alloc_init(struct ttm_mem_global *glob, unsigned max_pages)
1062 {
1063         int ret = -ENOMEM;
1064
1065         WARN_ON(_manager);
1066
1067         printk(KERN_INFO TTM_PFX "Initializing DMA pool allocator.\n");
1068
1069         _manager = kzalloc(sizeof(*_manager), GFP_KERNEL);
1070         if (!_manager)
1071                 goto err_manager;
1072
1073         mutex_init(&_manager->lock);
1074         INIT_LIST_HEAD(&_manager->pools);
1075
1076         _manager->options.max_size = max_pages;
1077         _manager->options.small = SMALL_ALLOCATION;
1078         _manager->options.alloc_size = NUM_PAGES_TO_ALLOC;
1079
1080         /* This takes care of auto-freeing the _manager */
1081         ret = kobject_init_and_add(&_manager->kobj, &ttm_pool_kobj_type,
1082                                    &glob->kobj, "dma_pool");
1083         if (unlikely(ret != 0)) {
1084                 kobject_put(&_manager->kobj);
1085                 goto err;
1086         }
1087         ttm_dma_pool_mm_shrink_init(_manager);
1088         return 0;
1089 err_manager:
1090         kfree(_manager);
1091         _manager = NULL;
1092 err:
1093         return ret;
1094 }
1095
1096 void ttm_dma_page_alloc_fini(void)
1097 {
1098         struct device_pools *p, *t;
1099
1100         printk(KERN_INFO TTM_PFX "Finalizing DMA pool allocator.\n");
1101         ttm_dma_pool_mm_shrink_fini(_manager);
1102
1103         list_for_each_entry_safe_reverse(p, t, &_manager->pools, pools) {
1104                 dev_dbg(p->dev, "(%s:%d) Freeing.\n", p->pool->name,
1105                         current->pid);
1106                 WARN_ON(devres_destroy(p->dev, ttm_dma_pool_release,
1107                         ttm_dma_pool_match, p->pool));
1108                 ttm_dma_free_pool(p->dev, p->pool->type);
1109         }
1110         kobject_put(&_manager->kobj);
1111         _manager = NULL;
1112 }
1113
1114 int ttm_dma_page_alloc_debugfs(struct seq_file *m, void *data)
1115 {
1116         struct device_pools *p;
1117         struct dma_pool *pool = NULL;
1118         char *h[] = {"pool", "refills", "pages freed", "inuse", "available",
1119                      "name", "virt", "busaddr"};
1120
1121         if (!_manager) {
1122                 seq_printf(m, "No pool allocator running.\n");
1123                 return 0;
1124         }
1125         seq_printf(m, "%13s %12s %13s %8s %8s %8s\n",
1126                    h[0], h[1], h[2], h[3], h[4], h[5]);
1127         mutex_lock(&_manager->lock);
1128         list_for_each_entry(p, &_manager->pools, pools) {
1129                 struct device *dev = p->dev;
1130                 if (!dev)
1131                         continue;
1132                 pool = p->pool;
1133                 seq_printf(m, "%13s %12ld %13ld %8d %8d %8s\n",
1134                                 pool->name, pool->nrefills,
1135                                 pool->nfrees, pool->npages_in_use,
1136                                 pool->npages_free,
1137                                 pool->dev_name);
1138         }
1139         mutex_unlock(&_manager->lock);
1140         return 0;
1141 }
1142 EXPORT_SYMBOL_GPL(ttm_dma_page_alloc_debugfs);