drivers/gpu/drm/ttm/ttm_page_alloc_dma.c

   1 /*
   2  * Copyright 2011 (c) Oracle Corp.
   3
   4  * Permission is hereby granted, free of charge, to any person obtaining a
   5  * copy of this software and associated documentation files (the "Software"),
   6  * to deal in the Software without restriction, including without limitation
   7  * the rights to use, copy, modify, merge, publish, distribute, sub license,
   8  * and/or sell copies of the Software, and to permit persons to whom the
   9  * Software is furnished to do so, subject to the following conditions:
  10  *
  11  * The above copyright notice and this permission notice (including the
  12  * next paragraph) shall be included in all copies or substantial portions
  13  * of the Software.
  14  *
  15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  16  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  17  * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
  18  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  19  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
  20  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
  21  * DEALINGS IN THE SOFTWARE.
  22  *
  23  * Author: Konrad Rzeszutek Wilk <[email protected]>
  24  */
  25
  26 /*
  27  * A simple DMA pool losely based on dmapool.c. It has certain advantages
  28  * over the DMA pools:
  29  * - Pool collects resently freed pages for reuse (and hooks up to
  30  *   the shrinker).
  31  * - Tracks currently in use pages
  32  * - Tracks whether the page is UC, WB or cached (and reverts to WB
  33  *   when freed).
  34  */
  35
  36 #define pr_fmt(fmt) "[TTM] " fmt
  37
  38 #include <linux/dma-mapping.h>
  39 #include <linux/list.h>
  40 #include <linux/seq_file.h> /* for seq_printf */
  41 #include <linux/slab.h>
  42 #include <linux/spinlock.h>
  43 #include <linux/highmem.h>
  44 #include <linux/mm_types.h>
  45 #include <linux/module.h>
  46 #include <linux/mm.h>
  47 #include <linux/atomic.h>
  48 #include <linux/device.h>
  49 #include <linux/kthread.h>
  50 #include <drm/ttm/ttm_bo_driver.h>
  51 #include <drm/ttm/ttm_page_alloc.h>
  52 #ifdef TTM_HAS_AGP
  53 #include <asm/agp.h>
  54 #endif
  55
  56 #define NUM_PAGES_TO_ALLOC              (PAGE_SIZE/sizeof(struct page *))
  57 #define SMALL_ALLOCATION                4
  58 #define FREE_ALL_PAGES                  (~0U)
  59 /* times are in msecs */
  60 #define IS_UNDEFINED                    (0)
  61 #define IS_WC                           (1<<1)
  62 #define IS_UC                           (1<<2)
  63 #define IS_CACHED                       (1<<3)
  64 #define IS_DMA32                        (1<<4)
  65
  66 enum pool_type {
  67         POOL_IS_UNDEFINED,
  68         POOL_IS_WC = IS_WC,
  69         POOL_IS_UC = IS_UC,
  70         POOL_IS_CACHED = IS_CACHED,
  71         POOL_IS_WC_DMA32 = IS_WC | IS_DMA32,
  72         POOL_IS_UC_DMA32 = IS_UC | IS_DMA32,
  73         POOL_IS_CACHED_DMA32 = IS_CACHED | IS_DMA32,
  74 };
  75 /*
  76  * The pool structure. There are usually six pools:
  77  *  - generic (not restricted to DMA32):
  78  *      - write combined, uncached, cached.
  79  *  - dma32 (up to 2^32 - so up 4GB):
  80  *      - write combined, uncached, cached.
  81  * for each 'struct device'. The 'cached' is for pages that are actively used.
  82  * The other ones can be shrunk by the shrinker API if neccessary.
  83  * @pools: The 'struct device->dma_pools' link.
  84  * @type: Type of the pool
  85  * @lock: Protects the inuse_list and free_list from concurrnet access. Must be
  86  * used with irqsave/irqrestore variants because pool allocator maybe called
  87  * from delayed work.
  88  * @inuse_list: Pool of pages that are in use. The order is very important and
  89  *   it is in the order that the TTM pages that are put back are in.
  90  * @free_list: Pool of pages that are free to be used. No order requirements.
  91  * @dev: The device that is associated with these pools.
  92  * @size: Size used during DMA allocation.
  93  * @npages_free: Count of available pages for re-use.
  94  * @npages_in_use: Count of pages that are in use.
  95  * @nfrees: Stats when pool is shrinking.
  96  * @nrefills: Stats when the pool is grown.
  97  * @gfp_flags: Flags to pass for alloc_page.
  98  * @name: Name of the pool.
  99  * @dev_name: Name derieved from dev - similar to how dev_info works.
 100  *   Used during shutdown as the dev_info during release is unavailable.
 101  */
 102 struct dma_pool {
 103         struct list_head pools; /* The 'struct device->dma_pools link */
 104         enum pool_type type;
 105         spinlock_t lock;
 106         struct list_head inuse_list;
 107         struct list_head free_list;
 108         struct device *dev;
 109         unsigned size;
 110         unsigned npages_free;
 111         unsigned npages_in_use;
 112         unsigned long nfrees; /* Stats when shrunk. */
 113         unsigned long nrefills; /* Stats when grown. */
 114         gfp_t gfp_flags;
 115         char name[13]; /* "cached dma32" */
 116         char dev_name[64]; /* Constructed from dev */
 117 };
 118
 119 /*
 120  * The accounting page keeping track of the allocated page along with
 121  * the DMA address.
 122  * @page_list: The link to the 'page_list' in 'struct dma_pool'.
 123  * @vaddr: The virtual address of the page
 124  * @dma: The bus address of the page. If the page is not allocated
 125  *   via the DMA API, it will be -1.
 126  */
 127 struct dma_page {
 128         struct list_head page_list;
 129         void *vaddr;
 130         struct page *p;
 131         dma_addr_t dma;
 132 };
 133
 134 /*
 135  * Limits for the pool. They are handled without locks because only place where
 136  * they may change is in sysfs store. They won't have immediate effect anyway
 137  * so forcing serialization to access them is pointless.
 138  */
 139
 140 struct ttm_pool_opts {
 141         unsigned        alloc_size;
 142         unsigned        max_size;
 143         unsigned        small;
 144 };
 145
 146 /*
 147  * Contains the list of all of the 'struct device' and their corresponding
 148  * DMA pools. Guarded by _mutex->lock.
 149  * @pools: The link to 'struct ttm_pool_manager->pools'
 150  * @dev: The 'struct device' associated with the 'pool'
 151  * @pool: The 'struct dma_pool' associated with the 'dev'
 152  */
 153 struct device_pools {
 154         struct list_head pools;
 155         struct device *dev;
 156         struct dma_pool *pool;
 157 };
 158
 159 /*
 160  * struct ttm_pool_manager - Holds memory pools for fast allocation
 161  *
 162  * @lock: Lock used when adding/removing from pools
 163  * @pools: List of 'struct device' and 'struct dma_pool' tuples.
 164  * @options: Limits for the pool.
 165  * @npools: Total amount of pools in existence.
 166  * @shrinker: The structure used by [un|]register_shrinker
 167  */
 168 struct ttm_pool_manager {
 169         struct mutex            lock;
 170         struct list_head        pools;
 171         struct ttm_pool_opts    options;
 172         unsigned                npools;
 173         struct shrinker         mm_shrink;
 174         struct kobject          kobj;
 175 };
 176
 177 static struct ttm_pool_manager *_manager;
 178
 179 static struct attribute ttm_page_pool_max = {
 180         .name = "pool_max_size",
 181         .mode = S_IRUGO | S_IWUSR
 182 };
 183 static struct attribute ttm_page_pool_small = {
 184         .name = "pool_small_allocation",
 185         .mode = S_IRUGO | S_IWUSR
 186 };
 187 static struct attribute ttm_page_pool_alloc_size = {
 188         .name = "pool_allocation_size",
 189         .mode = S_IRUGO | S_IWUSR
 190 };
 191
 192 static struct attribute *ttm_pool_attrs[] = {
 193         &ttm_page_pool_max,
 194         &ttm_page_pool_small,
 195         &ttm_page_pool_alloc_size,
 196         NULL
 197 };
 198
 199 static void ttm_pool_kobj_release(struct kobject *kobj)
 200 {
 201         struct ttm_pool_manager *m =
 202                 container_of(kobj, struct ttm_pool_manager, kobj);
 203         kfree(m);
 204 }
 205
 206 static ssize_t ttm_pool_store(struct kobject *kobj, struct attribute *attr,
 207                               const char *buffer, size_t size)
 208 {
 209         struct ttm_pool_manager *m =
 210                 container_of(kobj, struct ttm_pool_manager, kobj);
 211         int chars;
 212         unsigned val;
 213         chars = sscanf(buffer, "%u", &val);
 214         if (chars == 0)
 215                 return size;
 216
 217         /* Convert kb to number of pages */
 218         val = val / (PAGE_SIZE >> 10);
 219
 220         if (attr == &ttm_page_pool_max)
 221                 m->options.max_size = val;
 222         else if (attr == &ttm_page_pool_small)
 223                 m->options.small = val;
 224         else if (attr == &ttm_page_pool_alloc_size) {
 225                 if (val > NUM_PAGES_TO_ALLOC*8) {
 226                         pr_err("Setting allocation size to %lu is not allowed. Recommended size is %lu\n",
 227                                NUM_PAGES_TO_ALLOC*(PAGE_SIZE >> 7),
 228                                NUM_PAGES_TO_ALLOC*(PAGE_SIZE >> 10));
 229                         return size;
 230                 } else if (val > NUM_PAGES_TO_ALLOC) {
 231                         pr_warn("Setting allocation size to larger than %lu is not recommended\n",
 232                                 NUM_PAGES_TO_ALLOC*(PAGE_SIZE >> 10));
 233                 }
 234                 m->options.alloc_size = val;
 235         }
 236
 237         return size;
 238 }
 239
 240 static ssize_t ttm_pool_show(struct kobject *kobj, struct attribute *attr,
 241                              char *buffer)
 242 {
 243         struct ttm_pool_manager *m =
 244                 container_of(kobj, struct ttm_pool_manager, kobj);
 245         unsigned val = 0;
 246
 247         if (attr == &ttm_page_pool_max)
 248                 val = m->options.max_size;
 249         else if (attr == &ttm_page_pool_small)
 250                 val = m->options.small;
 251         else if (attr == &ttm_page_pool_alloc_size)
 252                 val = m->options.alloc_size;
 253
 254         val = val * (PAGE_SIZE >> 10);
 255
 256         return snprintf(buffer, PAGE_SIZE, "%u\n", val);
 257 }
 258
 259 static const struct sysfs_ops ttm_pool_sysfs_ops = {
 260         .show = &ttm_pool_show,
 261         .store = &ttm_pool_store,
 262 };
 263
 264 static struct kobj_type ttm_pool_kobj_type = {
 265         .release = &ttm_pool_kobj_release,
 266         .sysfs_ops = &ttm_pool_sysfs_ops,
 267         .default_attrs = ttm_pool_attrs,
 268 };
 269
 270 #ifndef CONFIG_X86
 271 static int set_pages_array_wb(struct page **pages, int addrinarray)
 272 {
 273 #ifdef TTM_HAS_AGP
 274         int i;
 275
 276         for (i = 0; i < addrinarray; i++)
 277                 unmap_page_from_agp(pages[i]);
 278 #endif
 279         return 0;
 280 }
 281
 282 static int set_pages_array_wc(struct page **pages, int addrinarray)
 283 {
 284 #ifdef TTM_HAS_AGP
 285         int i;
 286
 287         for (i = 0; i < addrinarray; i++)
 288                 map_page_into_agp(pages[i]);
 289 #endif
 290         return 0;
 291 }
 292
 293 static int set_pages_array_uc(struct page **pages, int addrinarray)
 294 {
 295 #ifdef TTM_HAS_AGP
 296         int i;
 297
 298         for (i = 0; i < addrinarray; i++)
 299                 map_page_into_agp(pages[i]);
 300 #endif
 301         return 0;
 302 }
 303 #endif /* for !CONFIG_X86 */
 304
 305 static int ttm_set_pages_caching(struct dma_pool *pool,
 306                                  struct page **pages, unsigned cpages)
 307 {
 308         int r = 0;
 309         /* Set page caching */
 310         if (pool->type & IS_UC) {
 311                 r = set_pages_array_uc(pages, cpages);
 312                 if (r)
 313                         pr_err("%s: Failed to set %d pages to uc!\n",
 314                                pool->dev_name, cpages);
 315         }
 316         if (pool->type & IS_WC) {
 317                 r = set_pages_array_wc(pages, cpages);
 318                 if (r)
 319                         pr_err("%s: Failed to set %d pages to wc!\n",
 320                                pool->dev_name, cpages);
 321         }
 322         return r;
 323 }
 324
 325 static void __ttm_dma_free_page(struct dma_pool *pool, struct dma_page *d_page)
 326 {
 327         dma_addr_t dma = d_page->dma;
 328         dma_free_coherent(pool->dev, pool->size, d_page->vaddr, dma);
 329
 330         kfree(d_page);
 331         d_page = NULL;
 332 }
 333 static struct dma_page *__ttm_dma_alloc_page(struct dma_pool *pool)
 334 {
 335         struct dma_page *d_page;
 336
 337         d_page = kmalloc(sizeof(struct dma_page), GFP_KERNEL);
 338         if (!d_page)
 339                 return NULL;
 340
 341         d_page->vaddr = dma_alloc_coherent(pool->dev, pool->size,
 342                                            &d_page->dma,
 343                                            pool->gfp_flags);
 344         if (d_page->vaddr)
 345                 d_page->p = virt_to_page(d_page->vaddr);
 346         else {
 347                 kfree(d_page);
 348                 d_page = NULL;
 349         }
 350         return d_page;
 351 }
 352 static enum pool_type ttm_to_type(int flags, enum ttm_caching_state cstate)
 353 {
 354         enum pool_type type = IS_UNDEFINED;
 355
 356         if (flags & TTM_PAGE_FLAG_DMA32)
 357                 type |= IS_DMA32;
 358         if (cstate == tt_cached)
 359                 type |= IS_CACHED;
 360         else if (cstate == tt_uncached)
 361                 type |= IS_UC;
 362         else
 363                 type |= IS_WC;
 364
 365         return type;
 366 }
 367
 368 static void ttm_pool_update_free_locked(struct dma_pool *pool,
 369                                         unsigned freed_pages)
 370 {
 371         pool->npages_free -= freed_pages;
 372         pool->nfrees += freed_pages;
 373
 374 }
 375
 376 /* set memory back to wb and free the pages. */
 377 static void ttm_dma_pages_put(struct dma_pool *pool, struct list_head *d_pages,
 378                               struct page *pages[], unsigned npages)
 379 {
 380         struct dma_page *d_page, *tmp;
 381
 382         /* Don't set WB on WB page pool. */
 383         if (npages && !(pool->type & IS_CACHED) &&
 384             set_pages_array_wb(pages, npages))
 385                 pr_err("%s: Failed to set %d pages to wb!\n",
 386                        pool->dev_name, npages);
 387
 388         list_for_each_entry_safe(d_page, tmp, d_pages, page_list) {
 389                 list_del(&d_page->page_list);
 390                 __ttm_dma_free_page(pool, d_page);
 391         }
 392 }
 393
 394 static void ttm_dma_page_put(struct dma_pool *pool, struct dma_page *d_page)
 395 {
 396         /* Don't set WB on WB page pool. */
 397         if (!(pool->type & IS_CACHED) && set_pages_array_wb(&d_page->p, 1))
 398                 pr_err("%s: Failed to set %d pages to wb!\n",
 399                        pool->dev_name, 1);
 400
 401         list_del(&d_page->page_list);
 402         __ttm_dma_free_page(pool, d_page);
 403 }
 404
 405 /*
 406  * Free pages from pool.
 407  *
 408  * To prevent hogging the ttm_swap process we only free NUM_PAGES_TO_ALLOC
 409  * number of pages in one go.
 410  *
 411  * @pool: to free the pages from
 412  * @nr_free: If set to true will free all pages in pool
 413  **/
 414 static unsigned ttm_dma_page_pool_free(struct dma_pool *pool, unsigned nr_free)
 415 {
 416         unsigned long irq_flags;
 417         struct dma_page *dma_p, *tmp;
 418         struct page **pages_to_free;
 419         struct list_head d_pages;
 420         unsigned freed_pages = 0,
 421                  npages_to_free = nr_free;
 422
 423         if (NUM_PAGES_TO_ALLOC < nr_free)
 424                 npages_to_free = NUM_PAGES_TO_ALLOC;
 425 #if 0
 426         if (nr_free > 1) {
 427                 pr_debug("%s: (%s:%d) Attempting to free %d (%d) pages\n",
 428                          pool->dev_name, pool->name, current->pid,
 429                          npages_to_free, nr_free);
 430         }
 431 #endif
 432         pages_to_free = kmalloc(npages_to_free * sizeof(struct page *),
 433                         GFP_KERNEL);
 434
 435         if (!pages_to_free) {
 436                 pr_err("%s: Failed to allocate memory for pool free operation\n",
 437                        pool->dev_name);
 438                 return 0;
 439         }
 440         INIT_LIST_HEAD(&d_pages);
 441 restart:
 442         spin_lock_irqsave(&pool->lock, irq_flags);
 443
 444         /* We picking the oldest ones off the list */
 445         list_for_each_entry_safe_reverse(dma_p, tmp, &pool->free_list,
 446                                          page_list) {
 447                 if (freed_pages >= npages_to_free)
 448                         break;
 449
 450                 /* Move the dma_page from one list to another. */
 451                 list_move(&dma_p->page_list, &d_pages);
 452
 453                 pages_to_free[freed_pages++] = dma_p->p;
 454                 /* We can only remove NUM_PAGES_TO_ALLOC at a time. */
 455                 if (freed_pages >= NUM_PAGES_TO_ALLOC) {
 456
 457                         ttm_pool_update_free_locked(pool, freed_pages);
 458                         /**
 459                          * Because changing page caching is costly
 460                          * we unlock the pool to prevent stalling.
 461                          */
 462                         spin_unlock_irqrestore(&pool->lock, irq_flags);
 463
 464                         ttm_dma_pages_put(pool, &d_pages, pages_to_free,
 465                                           freed_pages);
 466
 467                         INIT_LIST_HEAD(&d_pages);
 468
 469                         if (likely(nr_free != FREE_ALL_PAGES))
 470                                 nr_free -= freed_pages;
 471
 472                         if (NUM_PAGES_TO_ALLOC >= nr_free)
 473                                 npages_to_free = nr_free;
 474                         else
 475                                 npages_to_free = NUM_PAGES_TO_ALLOC;
 476
 477                         freed_pages = 0;
 478
 479                         /* free all so restart the processing */
 480                         if (nr_free)
 481                                 goto restart;
 482
 483                         /* Not allowed to fall through or break because
 484                          * following context is inside spinlock while we are
 485                          * outside here.
 486                          */
 487                         goto out;
 488
 489                 }
 490         }
 491
 492         /* remove range of pages from the pool */
 493         if (freed_pages) {
 494                 ttm_pool_update_free_locked(pool, freed_pages);
 495                 nr_free -= freed_pages;
 496         }
 497
 498         spin_unlock_irqrestore(&pool->lock, irq_flags);
 499
 500         if (freed_pages)
 501                 ttm_dma_pages_put(pool, &d_pages, pages_to_free, freed_pages);
 502 out:
 503         kfree(pages_to_free);
 504         return nr_free;
 505 }
 506
 507 static void ttm_dma_free_pool(struct device *dev, enum pool_type type)
 508 {
 509         struct device_pools *p;
 510         struct dma_pool *pool;
 511
 512         if (!dev)
 513                 return;
 514
 515         mutex_lock(&_manager->lock);
 516         list_for_each_entry_reverse(p, &_manager->pools, pools) {
 517                 if (p->dev != dev)
 518                         continue;
 519                 pool = p->pool;
 520                 if (pool->type != type)
 521                         continue;
 522
 523                 list_del(&p->pools);
 524                 kfree(p);
 525                 _manager->npools--;
 526                 break;
 527         }
 528         list_for_each_entry_reverse(pool, &dev->dma_pools, pools) {
 529                 if (pool->type != type)
 530                         continue;
 531                 /* Takes a spinlock.. */
 532                 ttm_dma_page_pool_free(pool, FREE_ALL_PAGES);
 533                 WARN_ON(((pool->npages_in_use + pool->npages_free) != 0));
 534                 /* This code path is called after _all_ references to the
 535                  * struct device has been dropped - so nobody should be
 536                  * touching it. In case somebody is trying to _add_ we are
 537                  * guarded by the mutex. */
 538                 list_del(&pool->pools);
 539                 kfree(pool);
 540                 break;
 541         }
 542         mutex_unlock(&_manager->lock);
 543 }
 544
 545 /*
 546  * On free-ing of the 'struct device' this deconstructor is run.
 547  * Albeit the pool might have already been freed earlier.
 548  */
 549 static void ttm_dma_pool_release(struct device *dev, void *res)
 550 {
 551         struct dma_pool *pool = *(struct dma_pool **)res;
 552
 553         if (pool)
 554                 ttm_dma_free_pool(dev, pool->type);
 555 }
 556
 557 static int ttm_dma_pool_match(struct device *dev, void *res, void *match_data)
 558 {
 559         return *(struct dma_pool **)res == match_data;
 560 }
 561
 562 static struct dma_pool *ttm_dma_pool_init(struct device *dev, gfp_t flags,
 563                                           enum pool_type type)
 564 {
 565         char *n[] = {"wc", "uc", "cached", " dma32", "unknown",};
 566         enum pool_type t[] = {IS_WC, IS_UC, IS_CACHED, IS_DMA32, IS_UNDEFINED};
 567         struct device_pools *sec_pool = NULL;
 568         struct dma_pool *pool = NULL, **ptr;
 569         unsigned i;
 570         int ret = -ENODEV;
 571         char *p;
 572
 573         if (!dev)
 574                 return NULL;
 575
 576         ptr = devres_alloc(ttm_dma_pool_release, sizeof(*ptr), GFP_KERNEL);
 577         if (!ptr)
 578                 return NULL;
 579
 580         ret = -ENOMEM;
 581
 582         pool = kmalloc_node(sizeof(struct dma_pool), GFP_KERNEL,
 583                             dev_to_node(dev));
 584         if (!pool)
 585                 goto err_mem;
 586
 587         sec_pool = kmalloc_node(sizeof(struct device_pools), GFP_KERNEL,
 588                                 dev_to_node(dev));
 589         if (!sec_pool)
 590                 goto err_mem;
 591
 592         INIT_LIST_HEAD(&sec_pool->pools);
 593         sec_pool->dev = dev;
 594         sec_pool->pool =  pool;
 595
 596         INIT_LIST_HEAD(&pool->free_list);
 597         INIT_LIST_HEAD(&pool->inuse_list);
 598         INIT_LIST_HEAD(&pool->pools);
 599         spin_lock_init(&pool->lock);
 600         pool->dev = dev;
 601         pool->npages_free = pool->npages_in_use = 0;
 602         pool->nfrees = 0;
 603         pool->gfp_flags = flags;
 604         pool->size = PAGE_SIZE;
 605         pool->type = type;
 606         pool->nrefills = 0;
 607         p = pool->name;
 608         for (i = 0; i < 5; 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, *found = NULL;
 642
 643         if (type == IS_UNDEFINED)
 644                 return found;
 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                         continue;
 660                 found = pool;
 661                 break;
 662         }
 663         return found;
 664 }
 665
 666 /*
 667  * Free pages the pages that failed to change the caching state. If there
 668  * are pages that have changed their caching state already put them to the
 669  * pool.
 670  */
 671 static void ttm_dma_handle_caching_state_failure(struct dma_pool *pool,
 672                                                  struct list_head *d_pages,
 673                                                  struct page **failed_pages,
 674                                                  unsigned cpages)
 675 {
 676         struct dma_page *d_page, *tmp;
 677         struct page *p;
 678         unsigned i = 0;
 679
 680         p = failed_pages[0];
 681         if (!p)
 682                 return;
 683         /* Find the failed page. */
 684         list_for_each_entry_safe(d_page, tmp, d_pages, page_list) {
 685                 if (d_page->p != p)
 686                         continue;
 687                 /* .. and then progress over the full list. */
 688                 list_del(&d_page->page_list);
 689                 __ttm_dma_free_page(pool, d_page);
 690                 if (++i < cpages)
 691                         p = failed_pages[i];
 692                 else
 693                         break;
 694         }
 695
 696 }
 697
 698 /*
 699  * Allocate 'count' pages, and put 'need' number of them on the
 700  * 'pages' and as well on the 'dma_address' starting at 'dma_offset' offset.
 701  * The full list of pages should also be on 'd_pages'.
 702  * We return zero for success, and negative numbers as errors.
 703  */
 704 static int ttm_dma_pool_alloc_new_pages(struct dma_pool *pool,
 705                                         struct list_head *d_pages,
 706                                         unsigned count)
 707 {
 708         struct page **caching_array;
 709         struct dma_page *dma_p;
 710         struct page *p;
 711         int r = 0;
 712         unsigned i, cpages;
 713         unsigned max_cpages = min(count,
 714                         (unsigned)(PAGE_SIZE/sizeof(struct page *)));
 715
 716         /* allocate array for page caching change */
 717         caching_array = kmalloc(max_cpages*sizeof(struct page *), GFP_KERNEL);
 718
 719         if (!caching_array) {
 720                 pr_err("%s: Unable to allocate table for new pages\n",
 721                        pool->dev_name);
 722                 return -ENOMEM;
 723         }
 724
 725         if (count > 1) {
 726                 pr_debug("%s: (%s:%d) Getting %d pages\n",
 727                          pool->dev_name, pool->name, current->pid, count);
 728         }
 729
 730         for (i = 0, cpages = 0; i < count; ++i) {
 731                 dma_p = __ttm_dma_alloc_page(pool);
 732                 if (!dma_p) {
 733                         pr_err("%s: Unable to get page %u\n",
 734                                pool->dev_name, i);
 735
 736                         /* store already allocated pages in the pool after
 737                          * setting the caching state */
 738                         if (cpages) {
 739                                 r = ttm_set_pages_caching(pool, caching_array,
 740                                                           cpages);
 741                                 if (r)
 742                                         ttm_dma_handle_caching_state_failure(
 743                                                 pool, d_pages, caching_array,
 744                                                 cpages);
 745                         }
 746                         r = -ENOMEM;
 747                         goto out;
 748                 }
 749                 p = dma_p->p;
 750 #ifdef CONFIG_HIGHMEM
 751                 /* gfp flags of highmem page should never be dma32 so we
 752                  * we should be fine in such case
 753                  */
 754                 if (!PageHighMem(p))
 755 #endif
 756                 {
 757                         caching_array[cpages++] = p;
 758                         if (cpages == max_cpages) {
 759                                 /* Note: Cannot hold the spinlock */
 760                                 r = ttm_set_pages_caching(pool, caching_array,
 761                                                  cpages);
 762                                 if (r) {
 763                                         ttm_dma_handle_caching_state_failure(
 764                                                 pool, d_pages, caching_array,
 765                                                 cpages);
 766                                         goto out;
 767                                 }
 768                                 cpages = 0;
 769                         }
 770                 }
 771                 list_add(&dma_p->page_list, d_pages);
 772         }
 773
 774         if (cpages) {
 775                 r = ttm_set_pages_caching(pool, caching_array, cpages);
 776                 if (r)
 777                         ttm_dma_handle_caching_state_failure(pool, d_pages,
 778                                         caching_array, cpages);
 779         }
 780 out:
 781         kfree(caching_array);
 782         return r;
 783 }
 784
 785 /*
 786  * @return count of pages still required to fulfill the request.
 787  */
 788 static int ttm_dma_page_pool_fill_locked(struct dma_pool *pool,
 789                                          unsigned long *irq_flags)
 790 {
 791         unsigned count = _manager->options.small;
 792         int r = pool->npages_free;
 793
 794         if (count > pool->npages_free) {
 795                 struct list_head d_pages;
 796
 797                 INIT_LIST_HEAD(&d_pages);
 798
 799                 spin_unlock_irqrestore(&pool->lock, *irq_flags);
 800
 801                 /* Returns how many more are neccessary to fulfill the
 802                  * request. */
 803                 r = ttm_dma_pool_alloc_new_pages(pool, &d_pages, count);
 804
 805                 spin_lock_irqsave(&pool->lock, *irq_flags);
 806                 if (!r) {
 807                         /* Add the fresh to the end.. */
 808                         list_splice(&d_pages, &pool->free_list);
 809                         ++pool->nrefills;
 810                         pool->npages_free += count;
 811                         r = count;
 812                 } else {
 813                         struct dma_page *d_page;
 814                         unsigned cpages = 0;
 815
 816                         pr_err("%s: Failed to fill %s pool (r:%d)!\n",
 817                                pool->dev_name, pool->name, r);
 818
 819                         list_for_each_entry(d_page, &d_pages, page_list) {
 820                                 cpages++;
 821                         }
 822                         list_splice_tail(&d_pages, &pool->free_list);
 823                         pool->npages_free += cpages;
 824                         r = cpages;
 825                 }
 826         }
 827         return r;
 828 }
 829
 830 /*
 831  * @return count of pages still required to fulfill the request.
 832  * The populate list is actually a stack (not that is matters as TTM
 833  * allocates one page at a time.
 834  */
 835 static int ttm_dma_pool_get_pages(struct dma_pool *pool,
 836                                   struct ttm_dma_tt *ttm_dma,
 837                                   unsigned index)
 838 {
 839         struct dma_page *d_page;
 840         struct ttm_tt *ttm = &ttm_dma->ttm;
 841         unsigned long irq_flags;
 842         int count, r = -ENOMEM;
 843
 844         spin_lock_irqsave(&pool->lock, irq_flags);
 845         count = ttm_dma_page_pool_fill_locked(pool, &irq_flags);
 846         if (count) {
 847                 d_page = list_first_entry(&pool->free_list, struct dma_page, page_list);
 848                 ttm->pages[index] = d_page->p;
 849                 ttm_dma->dma_address[index] = d_page->dma;
 850                 list_move_tail(&d_page->page_list, &ttm_dma->pages_list);
 851                 r = 0;
 852                 pool->npages_in_use += 1;
 853                 pool->npages_free -= 1;
 854         }
 855         spin_unlock_irqrestore(&pool->lock, irq_flags);
 856         return r;
 857 }
 858
 859 /*
 860  * On success pages list will hold count number of correctly
 861  * cached pages. On failure will hold the negative return value (-ENOMEM, etc).
 862  */
 863 int ttm_dma_populate(struct ttm_dma_tt *ttm_dma, struct device *dev)
 864 {
 865         struct ttm_tt *ttm = &ttm_dma->ttm;
 866         struct ttm_mem_global *mem_glob = ttm->glob->mem_glob;
 867         struct dma_pool *pool;
 868         enum pool_type type;
 869         unsigned i;
 870         gfp_t gfp_flags;
 871         int ret;
 872
 873         if (ttm->state != tt_unpopulated)
 874                 return 0;
 875
 876         type = ttm_to_type(ttm->page_flags, ttm->caching_state);
 877         if (ttm->page_flags & TTM_PAGE_FLAG_DMA32)
 878                 gfp_flags = GFP_USER | GFP_DMA32;
 879         else
 880                 gfp_flags = GFP_HIGHUSER;
 881         if (ttm->page_flags & TTM_PAGE_FLAG_ZERO_ALLOC)
 882                 gfp_flags |= __GFP_ZERO;
 883
 884         pool = ttm_dma_find_pool(dev, type);
 885         if (!pool) {
 886                 pool = ttm_dma_pool_init(dev, gfp_flags, type);
 887                 if (IS_ERR_OR_NULL(pool)) {
 888                         return -ENOMEM;
 889                 }
 890         }
 891
 892         INIT_LIST_HEAD(&ttm_dma->pages_list);
 893         for (i = 0; i < ttm->num_pages; ++i) {
 894                 ret = ttm_dma_pool_get_pages(pool, ttm_dma, i);
 895                 if (ret != 0) {
 896                         ttm_dma_unpopulate(ttm_dma, dev);
 897                         return -ENOMEM;
 898                 }
 899
 900                 ret = ttm_mem_global_alloc_page(mem_glob, ttm->pages[i],
 901                                                 false, false);
 902                 if (unlikely(ret != 0)) {
 903                         ttm_dma_unpopulate(ttm_dma, dev);
 904                         return -ENOMEM;
 905                 }
 906         }
 907
 908         if (unlikely(ttm->page_flags & TTM_PAGE_FLAG_SWAPPED)) {
 909                 ret = ttm_tt_swapin(ttm);
 910                 if (unlikely(ret != 0)) {
 911                         ttm_dma_unpopulate(ttm_dma, dev);
 912                         return ret;
 913                 }
 914         }
 915
 916         ttm->state = tt_unbound;
 917         return 0;
 918 }
 919 EXPORT_SYMBOL_GPL(ttm_dma_populate);
 920
 921 /* Get good estimation how many pages are free in pools */
 922 static int ttm_dma_pool_get_num_unused_pages(void)
 923 {
 924         struct device_pools *p;
 925         unsigned total = 0;
 926
 927         mutex_lock(&_manager->lock);
 928         list_for_each_entry(p, &_manager->pools, pools)
 929                 total += p->pool->npages_free;
 930         mutex_unlock(&_manager->lock);
 931         return total;
 932 }
 933
 934 /* Put all pages in pages list to correct pool to wait for reuse */
 935 void ttm_dma_unpopulate(struct ttm_dma_tt *ttm_dma, struct device *dev)
 936 {
 937         struct ttm_tt *ttm = &ttm_dma->ttm;
 938         struct dma_pool *pool;
 939         struct dma_page *d_page, *next;
 940         enum pool_type type;
 941         bool is_cached = false;
 942         unsigned count = 0, i, npages = 0;
 943         unsigned long irq_flags;
 944
 945         type = ttm_to_type(ttm->page_flags, ttm->caching_state);
 946         pool = ttm_dma_find_pool(dev, type);
 947         if (!pool)
 948                 return;
 949
 950         is_cached = (ttm_dma_find_pool(pool->dev,
 951                      ttm_to_type(ttm->page_flags, tt_cached)) == pool);
 952
 953         /* make sure pages array match list and count number of pages */
 954         list_for_each_entry(d_page, &ttm_dma->pages_list, page_list) {
 955                 ttm->pages[count] = d_page->p;
 956                 count++;
 957         }
 958
 959         spin_lock_irqsave(&pool->lock, irq_flags);
 960         pool->npages_in_use -= count;
 961         if (is_cached) {
 962                 pool->nfrees += count;
 963         } else {
 964                 pool->npages_free += count;
 965                 list_splice(&ttm_dma->pages_list, &pool->free_list);
 966                 npages = count;
 967                 if (pool->npages_free > _manager->options.max_size) {
 968                         npages = pool->npages_free - _manager->options.max_size;
 969                         /* free at least NUM_PAGES_TO_ALLOC number of pages
 970                          * to reduce calls to set_memory_wb */
 971                         if (npages < NUM_PAGES_TO_ALLOC)
 972                                 npages = NUM_PAGES_TO_ALLOC;
 973                 }
 974         }
 975         spin_unlock_irqrestore(&pool->lock, irq_flags);
 976
 977         if (is_cached) {
 978                 list_for_each_entry_safe(d_page, next, &ttm_dma->pages_list, page_list) {
 979                         ttm_mem_global_free_page(ttm->glob->mem_glob,
 980                                                  d_page->p);
 981                         ttm_dma_page_put(pool, d_page);
 982                 }
 983         } else {
 984                 for (i = 0; i < count; i++) {
 985                         ttm_mem_global_free_page(ttm->glob->mem_glob,
 986                                                  ttm->pages[i]);
 987                 }
 988         }
 989
 990         INIT_LIST_HEAD(&ttm_dma->pages_list);
 991         for (i = 0; i < ttm->num_pages; i++) {
 992                 ttm->pages[i] = NULL;
 993                 ttm_dma->dma_address[i] = 0;
 994         }
 995
 996         /* shrink pool if necessary (only on !is_cached pools)*/
 997         if (npages)
 998                 ttm_dma_page_pool_free(pool, npages);
 999         ttm->state = tt_unpopulated;
1000 }
1001 EXPORT_SYMBOL_GPL(ttm_dma_unpopulate);
1002
1003 /**
1004  * Callback for mm to request pool to reduce number of page held.
1005  */
1006 static int ttm_dma_pool_mm_shrink(struct shrinker *shrink,
1007                                   struct shrink_control *sc)
1008 {
1009         static atomic_t start_pool = ATOMIC_INIT(0);
1010         unsigned idx = 0;
1011         unsigned pool_offset = atomic_add_return(1, &start_pool);
1012         unsigned shrink_pages = sc->nr_to_scan;
1013         struct device_pools *p;
1014
1015         if (list_empty(&_manager->pools))
1016                 return 0;
1017
1018         mutex_lock(&_manager->lock);
1019         pool_offset = pool_offset % _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                 pr_debug("%s: (%s:%d) Asked to shrink %d, have %d more to go\n",
1033                          p->pool->dev_name, p->pool->name, current->pid,
1034                          nr_free, shrink_pages);
1035         }
1036         mutex_unlock(&_manager->lock);
1037         /* return estimated number of unused pages in pool */
1038         return ttm_dma_pool_get_num_unused_pages();
1039 }
1040
1041 static void ttm_dma_pool_mm_shrink_init(struct ttm_pool_manager *manager)
1042 {
1043         manager->mm_shrink.shrink = &ttm_dma_pool_mm_shrink;
1044         manager->mm_shrink.seeks = 1;
1045         register_shrinker(&manager->mm_shrink);
1046 }
1047
1048 static void ttm_dma_pool_mm_shrink_fini(struct ttm_pool_manager *manager)
1049 {
1050         unregister_shrinker(&manager->mm_shrink);
1051 }
1052
1053 int ttm_dma_page_alloc_init(struct ttm_mem_global *glob, unsigned max_pages)
1054 {
1055         int ret = -ENOMEM;
1056
1057         WARN_ON(_manager);
1058
1059         pr_info("Initializing DMA pool allocator\n");
1060
1061         _manager = kzalloc(sizeof(*_manager), GFP_KERNEL);
1062         if (!_manager)
1063                 goto err;
1064
1065         mutex_init(&_manager->lock);
1066         INIT_LIST_HEAD(&_manager->pools);
1067
1068         _manager->options.max_size = max_pages;
1069         _manager->options.small = SMALL_ALLOCATION;
1070         _manager->options.alloc_size = NUM_PAGES_TO_ALLOC;
1071
1072         /* This takes care of auto-freeing the _manager */
1073         ret = kobject_init_and_add(&_manager->kobj, &ttm_pool_kobj_type,
1074                                    &glob->kobj, "dma_pool");
1075         if (unlikely(ret != 0)) {
1076                 kobject_put(&_manager->kobj);
1077                 goto err;
1078         }
1079         ttm_dma_pool_mm_shrink_init(_manager);
1080         return 0;
1081 err:
1082         return ret;
1083 }
1084
1085 void ttm_dma_page_alloc_fini(void)
1086 {
1087         struct device_pools *p, *t;
1088
1089         pr_info("Finalizing DMA pool allocator\n");
1090         ttm_dma_pool_mm_shrink_fini(_manager);
1091
1092         list_for_each_entry_safe_reverse(p, t, &_manager->pools, pools) {
1093                 dev_dbg(p->dev, "(%s:%d) Freeing.\n", p->pool->name,
1094                         current->pid);
1095                 WARN_ON(devres_destroy(p->dev, ttm_dma_pool_release,
1096                         ttm_dma_pool_match, p->pool));
1097                 ttm_dma_free_pool(p->dev, p->pool->type);
1098         }
1099         kobject_put(&_manager->kobj);
1100         _manager = NULL;
1101 }
1102
1103 int ttm_dma_page_alloc_debugfs(struct seq_file *m, void *data)
1104 {
1105         struct device_pools *p;
1106         struct dma_pool *pool = NULL;
1107         char *h[] = {"pool", "refills", "pages freed", "inuse", "available",
1108                      "name", "virt", "busaddr"};
1109
1110         if (!_manager) {
1111                 seq_printf(m, "No pool allocator running.\n");
1112                 return 0;
1113         }
1114         seq_printf(m, "%13s %12s %13s %8s %8s %8s\n",
1115                    h[0], h[1], h[2], h[3], h[4], h[5]);
1116         mutex_lock(&_manager->lock);
1117         list_for_each_entry(p, &_manager->pools, pools) {
1118                 struct device *dev = p->dev;
1119                 if (!dev)
1120                         continue;
1121                 pool = p->pool;
1122                 seq_printf(m, "%13s %12ld %13ld %8d %8d %8s\n",
1123                                 pool->name, pool->nrefills,
1124                                 pool->nfrees, pool->npages_in_use,
1125                                 pool->npages_free,
1126                                 pool->dev_name);
1127         }
1128         mutex_unlock(&_manager->lock);
1129         return 0;
1130 }
1131 EXPORT_SYMBOL_GPL(ttm_dma_page_alloc_debugfs);