drivers/gpu/drm/ttm/ttm_page_alloc.c

   1 /*
   2  * Copyright (c) Red Hat Inc.
   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  * Authors: Dave Airlie <airlied@redhat.com>
  24  *          Jerome Glisse <jglisse@redhat.com>
  25  *          Pauli Nieminen <suokkos@gmail.com>
  26  */
  27
  28 /* simple list based uncached page pool
  29  * - Pool collects resently freed pages for reuse
  30  * - Use page->lru to keep a free list
  31  * - doesn't track currently in use pages
  32  */
  33
  34 #define pr_fmt(fmt) "[TTM] " fmt
  35
  36 #include <linux/list.h>
  37 #include <linux/spinlock.h>
  38 #include <linux/highmem.h>
  39 #include <linux/mm_types.h>
  40 #include <linux/module.h>
  41 #include <linux/mm.h>
  42 #include <linux/seq_file.h> /* for seq_printf */
  43 #include <linux/slab.h>
  44 #include <linux/dma-mapping.h>
  45
  46 #include <linux/atomic.h>
  47
  48 #include <drm/ttm/ttm_bo_driver.h>
  49 #include <drm/ttm/ttm_page_alloc.h>
  50
  51 #if IS_ENABLED(CONFIG_AGP)
  52 #include <asm/agp.h>
  53 #endif
  54 #ifdef CONFIG_X86
  55 #include <asm/set_memory.h>
  56 #endif
  57
  58 #define NUM_PAGES_TO_ALLOC              (PAGE_SIZE/sizeof(struct page *))
  59 #define SMALL_ALLOCATION                16
  60 #define FREE_ALL_PAGES                  (~0U)
  61 /* times are in msecs */
  62 #define PAGE_FREE_INTERVAL              1000
  63
  64 /**
  65  * struct ttm_page_pool - Pool to reuse recently allocated uc/wc pages.
  66  *
  67  * @lock: Protects the shared pool from concurrnet access. Must be used with
  68  * irqsave/irqrestore variants because pool allocator maybe called from
  69  * delayed work.
  70  * @fill_lock: Prevent concurrent calls to fill.
  71  * @list: Pool of free uc/wc pages for fast reuse.
  72  * @gfp_flags: Flags to pass for alloc_page.
  73  * @npages: Number of pages in pool.
  74  */
  75 struct ttm_page_pool {
  76         spinlock_t              lock;
  77         bool                    fill_lock;
  78         struct list_head        list;
  79         gfp_t                   gfp_flags;
  80         unsigned                npages;
  81         char                    *name;
  82         unsigned long           nfrees;
  83         unsigned long           nrefills;
  84 };
  85
  86 /**
  87  * Limits for the pool. They are handled without locks because only place where
  88  * they may change is in sysfs store. They won't have immediate effect anyway
  89  * so forcing serialization to access them is pointless.
  90  */
  91
  92 struct ttm_pool_opts {
  93         unsigned        alloc_size;
  94         unsigned        max_size;
  95         unsigned        small;
  96 };
  97
  98 #define NUM_POOLS 6
  99
 100 /**
 101  * struct ttm_pool_manager - Holds memory pools for fst allocation
 102  *
 103  * Manager is read only object for pool code so it doesn't need locking.
 104  *
 105  * @free_interval: minimum number of jiffies between freeing pages from pool.
 106  * @page_alloc_inited: reference counting for pool allocation.
 107  * @work: Work that is used to shrink the pool. Work is only run when there is
 108  * some pages to free.
 109  * @small_allocation: Limit in number of pages what is small allocation.
 110  *
 111  * @pools: All pool objects in use.
 112  **/
 113 struct ttm_pool_manager {
 114         struct kobject          kobj;
 115         struct shrinker         mm_shrink;
 116         struct ttm_pool_opts    options;
 117
 118         union {
 119                 struct ttm_page_pool    pools[NUM_POOLS];
 120                 struct {
 121                         struct ttm_page_pool    wc_pool;
 122                         struct ttm_page_pool    uc_pool;
 123                         struct ttm_page_pool    wc_pool_dma32;
 124                         struct ttm_page_pool    uc_pool_dma32;
 125                         struct ttm_page_pool    wc_pool_huge;
 126                         struct ttm_page_pool    uc_pool_huge;
 127                 } ;
 128         };
 129 };
 130
 131 static struct attribute ttm_page_pool_max = {
 132         .name = "pool_max_size",
 133         .mode = S_IRUGO | S_IWUSR
 134 };
 135 static struct attribute ttm_page_pool_small = {
 136         .name = "pool_small_allocation",
 137         .mode = S_IRUGO | S_IWUSR
 138 };
 139 static struct attribute ttm_page_pool_alloc_size = {
 140         .name = "pool_allocation_size",
 141         .mode = S_IRUGO | S_IWUSR
 142 };
 143
 144 static struct attribute *ttm_pool_attrs[] = {
 145         &ttm_page_pool_max,
 146         &ttm_page_pool_small,
 147         &ttm_page_pool_alloc_size,
 148         NULL
 149 };
 150
 151 static void ttm_pool_kobj_release(struct kobject *kobj)
 152 {
 153         struct ttm_pool_manager *m =
 154                 container_of(kobj, struct ttm_pool_manager, kobj);
 155         kfree(m);
 156 }
 157
 158 static ssize_t ttm_pool_store(struct kobject *kobj,
 159                 struct attribute *attr, const char *buffer, size_t size)
 160 {
 161         struct ttm_pool_manager *m =
 162                 container_of(kobj, struct ttm_pool_manager, kobj);
 163         int chars;
 164         unsigned val;
 165         chars = sscanf(buffer, "%u", &val);
 166         if (chars == 0)
 167                 return size;
 168
 169         /* Convert kb to number of pages */
 170         val = val / (PAGE_SIZE >> 10);
 171
 172         if (attr == &ttm_page_pool_max)
 173                 m->options.max_size = val;
 174         else if (attr == &ttm_page_pool_small)
 175                 m->options.small = val;
 176         else if (attr == &ttm_page_pool_alloc_size) {
 177                 if (val > NUM_PAGES_TO_ALLOC*8) {
 178                         pr_err("Setting allocation size to %lu is not allowed. Recommended size is %lu\n",
 179                                NUM_PAGES_TO_ALLOC*(PAGE_SIZE >> 7),
 180                                NUM_PAGES_TO_ALLOC*(PAGE_SIZE >> 10));
 181                         return size;
 182                 } else if (val > NUM_PAGES_TO_ALLOC) {
 183                         pr_warn("Setting allocation size to larger than %lu is not recommended\n",
 184                                 NUM_PAGES_TO_ALLOC*(PAGE_SIZE >> 10));
 185                 }
 186                 m->options.alloc_size = val;
 187         }
 188
 189         return size;
 190 }
 191
 192 static ssize_t ttm_pool_show(struct kobject *kobj,
 193                 struct attribute *attr, char *buffer)
 194 {
 195         struct ttm_pool_manager *m =
 196                 container_of(kobj, struct ttm_pool_manager, kobj);
 197         unsigned val = 0;
 198
 199         if (attr == &ttm_page_pool_max)
 200                 val = m->options.max_size;
 201         else if (attr == &ttm_page_pool_small)
 202                 val = m->options.small;
 203         else if (attr == &ttm_page_pool_alloc_size)
 204                 val = m->options.alloc_size;
 205
 206         val = val * (PAGE_SIZE >> 10);
 207
 208         return snprintf(buffer, PAGE_SIZE, "%u\n", val);
 209 }
 210
 211 static const struct sysfs_ops ttm_pool_sysfs_ops = {
 212         .show = &ttm_pool_show,
 213         .store = &ttm_pool_store,
 214 };
 215
 216 static struct kobj_type ttm_pool_kobj_type = {
 217         .release = &ttm_pool_kobj_release,
 218         .sysfs_ops = &ttm_pool_sysfs_ops,
 219         .default_attrs = ttm_pool_attrs,
 220 };
 221
 222 static struct ttm_pool_manager *_manager;
 223
 224 #ifndef CONFIG_X86
 225 static int set_pages_array_wb(struct page **pages, int addrinarray)
 226 {
 227 #if IS_ENABLED(CONFIG_AGP)
 228         int i;
 229
 230         for (i = 0; i < addrinarray; i++)
 231                 unmap_page_from_agp(pages[i]);
 232 #endif
 233         return 0;
 234 }
 235
 236 static int set_pages_array_wc(struct page **pages, int addrinarray)
 237 {
 238 #if IS_ENABLED(CONFIG_AGP)
 239         int i;
 240
 241         for (i = 0; i < addrinarray; i++)
 242                 map_page_into_agp(pages[i]);
 243 #endif
 244         return 0;
 245 }
 246
 247 static int set_pages_array_uc(struct page **pages, int addrinarray)
 248 {
 249 #if IS_ENABLED(CONFIG_AGP)
 250         int i;
 251
 252         for (i = 0; i < addrinarray; i++)
 253                 map_page_into_agp(pages[i]);
 254 #endif
 255         return 0;
 256 }
 257 #endif
 258
 259 /**
 260  * Select the right pool or requested caching state and ttm flags. */
 261 static struct ttm_page_pool *ttm_get_pool(int flags, bool huge,
 262                                           enum ttm_caching_state cstate)
 263 {
 264         int pool_index;
 265
 266         if (cstate == tt_cached)
 267                 return NULL;
 268
 269         if (cstate == tt_wc)
 270                 pool_index = 0x0;
 271         else
 272                 pool_index = 0x1;
 273
 274         if (flags & TTM_PAGE_FLAG_DMA32) {
 275                 if (huge)
 276                         return NULL;
 277                 pool_index |= 0x2;
 278
 279         } else if (huge) {
 280                 pool_index |= 0x4;
 281         }
 282
 283         return &_manager->pools[pool_index];
 284 }
 285
 286 /* set memory back to wb and free the pages. */
 287 static void ttm_pages_put(struct page *pages[], unsigned npages)
 288 {
 289         unsigned i;
 290         if (set_pages_array_wb(pages, npages))
 291                 pr_err("Failed to set %d pages to wb!\n", npages);
 292         for (i = 0; i < npages; ++i)
 293                 __free_page(pages[i]);
 294 }
 295
 296 static void ttm_pool_update_free_locked(struct ttm_page_pool *pool,
 297                 unsigned freed_pages)
 298 {
 299         pool->npages -= freed_pages;
 300         pool->nfrees += freed_pages;
 301 }
 302
 303 /**
 304  * Free pages from pool.
 305  *
 306  * To prevent hogging the ttm_swap process we only free NUM_PAGES_TO_ALLOC
 307  * number of pages in one go.
 308  *
 309  * @pool: to free the pages from
 310  * @free_all: If set to true will free all pages in pool
 311  * @use_static: Safe to use static buffer
 312  **/
 313 static int ttm_page_pool_free(struct ttm_page_pool *pool, unsigned nr_free,
 314                               bool use_static)
 315 {
 316         static struct page *static_buf[NUM_PAGES_TO_ALLOC];
 317         unsigned long irq_flags;
 318         struct page *p;
 319         struct page **pages_to_free;
 320         unsigned freed_pages = 0,
 321                  npages_to_free = nr_free;
 322
 323         if (NUM_PAGES_TO_ALLOC < nr_free)
 324                 npages_to_free = NUM_PAGES_TO_ALLOC;
 325
 326         if (use_static)
 327                 pages_to_free = static_buf;
 328         else
 329                 pages_to_free = kmalloc(npages_to_free * sizeof(struct page *),
 330                                         GFP_KERNEL);
 331         if (!pages_to_free) {
 332                 pr_debug("Failed to allocate memory for pool free operation\n");
 333                 return 0;
 334         }
 335
 336 restart:
 337         spin_lock_irqsave(&pool->lock, irq_flags);
 338
 339         list_for_each_entry_reverse(p, &pool->list, lru) {
 340                 if (freed_pages >= npages_to_free)
 341                         break;
 342
 343                 pages_to_free[freed_pages++] = p;
 344                 /* We can only remove NUM_PAGES_TO_ALLOC at a time. */
 345                 if (freed_pages >= NUM_PAGES_TO_ALLOC) {
 346                         /* remove range of pages from the pool */
 347                         __list_del(p->lru.prev, &pool->list);
 348
 349                         ttm_pool_update_free_locked(pool, freed_pages);
 350                         /**
 351                          * Because changing page caching is costly
 352                          * we unlock the pool to prevent stalling.
 353                          */
 354                         spin_unlock_irqrestore(&pool->lock, irq_flags);
 355
 356                         ttm_pages_put(pages_to_free, freed_pages);
 357                         if (likely(nr_free != FREE_ALL_PAGES))
 358                                 nr_free -= freed_pages;
 359
 360                         if (NUM_PAGES_TO_ALLOC >= nr_free)
 361                                 npages_to_free = nr_free;
 362                         else
 363                                 npages_to_free = NUM_PAGES_TO_ALLOC;
 364
 365                         freed_pages = 0;
 366
 367                         /* free all so restart the processing */
 368                         if (nr_free)
 369                                 goto restart;
 370
 371                         /* Not allowed to fall through or break because
 372                          * following context is inside spinlock while we are
 373                          * outside here.
 374                          */
 375                         goto out;
 376
 377                 }
 378         }
 379
 380         /* remove range of pages from the pool */
 381         if (freed_pages) {
 382                 __list_del(&p->lru, &pool->list);
 383
 384                 ttm_pool_update_free_locked(pool, freed_pages);
 385                 nr_free -= freed_pages;
 386         }
 387
 388         spin_unlock_irqrestore(&pool->lock, irq_flags);
 389
 390         if (freed_pages)
 391                 ttm_pages_put(pages_to_free, freed_pages);
 392 out:
 393         if (pages_to_free != static_buf)
 394                 kfree(pages_to_free);
 395         return nr_free;
 396 }
 397
 398 /**
 399  * Callback for mm to request pool to reduce number of page held.
 400  *
 401  * XXX: (dchinner) Deadlock warning!
 402  *
 403  * This code is crying out for a shrinker per pool....
 404  */
 405 static unsigned long
 406 ttm_pool_shrink_scan(struct shrinker *shrink, struct shrink_control *sc)
 407 {
 408         static DEFINE_MUTEX(lock);
 409         static unsigned start_pool;
 410         unsigned i;
 411         unsigned pool_offset;
 412         struct ttm_page_pool *pool;
 413         int shrink_pages = sc->nr_to_scan;
 414         unsigned long freed = 0;
 415
 416         if (!mutex_trylock(&lock))
 417                 return SHRINK_STOP;
 418         pool_offset = ++start_pool % NUM_POOLS;
 419         /* select start pool in round robin fashion */
 420         for (i = 0; i < NUM_POOLS; ++i) {
 421                 unsigned nr_free = shrink_pages;
 422                 if (shrink_pages == 0)
 423                         break;
 424                 pool = &_manager->pools[(i + pool_offset)%NUM_POOLS];
 425                 /* OK to use static buffer since global mutex is held. */
 426                 shrink_pages = ttm_page_pool_free(pool, nr_free, true);
 427                 freed += nr_free - shrink_pages;
 428         }
 429         mutex_unlock(&lock);
 430         return freed;
 431 }
 432
 433
 434 static unsigned long
 435 ttm_pool_shrink_count(struct shrinker *shrink, struct shrink_control *sc)
 436 {
 437         unsigned i;
 438         unsigned long count = 0;
 439
 440         for (i = 0; i < NUM_POOLS; ++i)
 441                 count += _manager->pools[i].npages;
 442
 443         return count;
 444 }
 445
 446 static void ttm_pool_mm_shrink_init(struct ttm_pool_manager *manager)
 447 {
 448         manager->mm_shrink.count_objects = ttm_pool_shrink_count;
 449         manager->mm_shrink.scan_objects = ttm_pool_shrink_scan;
 450         manager->mm_shrink.seeks = 1;
 451         register_shrinker(&manager->mm_shrink);
 452 }
 453
 454 static void ttm_pool_mm_shrink_fini(struct ttm_pool_manager *manager)
 455 {
 456         unregister_shrinker(&manager->mm_shrink);
 457 }
 458
 459 static int ttm_set_pages_caching(struct page **pages,
 460                 enum ttm_caching_state cstate, unsigned cpages)
 461 {
 462         int r = 0;
 463         /* Set page caching */
 464         switch (cstate) {
 465         case tt_uncached:
 466                 r = set_pages_array_uc(pages, cpages);
 467                 if (r)
 468                         pr_err("Failed to set %d pages to uc!\n", cpages);
 469                 break;
 470         case tt_wc:
 471                 r = set_pages_array_wc(pages, cpages);
 472                 if (r)
 473                         pr_err("Failed to set %d pages to wc!\n", cpages);
 474                 break;
 475         default:
 476                 break;
 477         }
 478         return r;
 479 }
 480
 481 /**
 482  * Free pages the pages that failed to change the caching state. If there is
 483  * any pages that have changed their caching state already put them to the
 484  * pool.
 485  */
 486 static void ttm_handle_caching_state_failure(struct list_head *pages,
 487                 int ttm_flags, enum ttm_caching_state cstate,
 488                 struct page **failed_pages, unsigned cpages)
 489 {
 490         unsigned i;
 491         /* Failed pages have to be freed */
 492         for (i = 0; i < cpages; ++i) {
 493                 list_del(&failed_pages[i]->lru);
 494                 __free_page(failed_pages[i]);
 495         }
 496 }
 497
 498 /**
 499  * Allocate new pages with correct caching.
 500  *
 501  * This function is reentrant if caller updates count depending on number of
 502  * pages returned in pages array.
 503  */
 504 static int ttm_alloc_new_pages(struct list_head *pages, gfp_t gfp_flags,
 505                                int ttm_flags, enum ttm_caching_state cstate,
 506                                unsigned count, unsigned order)
 507 {
 508         struct page **caching_array;
 509         struct page *p;
 510         int r = 0;
 511         unsigned i, j, cpages;
 512         unsigned npages = 1 << order;
 513         unsigned max_cpages = min(count,
 514                         (unsigned)(PAGE_SIZE/sizeof(struct page *)));
 515
 516         /* allocate array for page caching change */
 517         caching_array = kmalloc(max_cpages*sizeof(struct page *), GFP_KERNEL);
 518
 519         if (!caching_array) {
 520                 pr_debug("Unable to allocate table for new pages\n");
 521                 return -ENOMEM;
 522         }
 523
 524         for (i = 0, cpages = 0; i < count; ++i) {
 525                 p = alloc_pages(gfp_flags, order);
 526
 527                 if (!p) {
 528                         pr_debug("Unable to get page %u\n", i);
 529
 530                         /* store already allocated pages in the pool after
 531                          * setting the caching state */
 532                         if (cpages) {
 533                                 r = ttm_set_pages_caching(caching_array,
 534                                                           cstate, cpages);
 535                                 if (r)
 536                                         ttm_handle_caching_state_failure(pages,
 537                                                 ttm_flags, cstate,
 538                                                 caching_array, cpages);
 539                         }
 540                         r = -ENOMEM;
 541                         goto out;
 542                 }
 543
 544                 list_add(&p->lru, pages);
 545
 546 #ifdef CONFIG_HIGHMEM
 547                 /* gfp flags of highmem page should never be dma32 so we
 548                  * we should be fine in such case
 549                  */
 550                 if (PageHighMem(p))
 551                         continue;
 552
 553 #endif
 554                 for (j = 0; j < npages; ++j) {
 555                         caching_array[cpages++] = p++;
 556                         if (cpages == max_cpages) {
 557
 558                                 r = ttm_set_pages_caching(caching_array,
 559                                                 cstate, cpages);
 560                                 if (r) {
 561                                         ttm_handle_caching_state_failure(pages,
 562                                                 ttm_flags, cstate,
 563                                                 caching_array, cpages);
 564                                         goto out;
 565                                 }
 566                                 cpages = 0;
 567                         }
 568                 }
 569         }
 570
 571         if (cpages) {
 572                 r = ttm_set_pages_caching(caching_array, cstate, cpages);
 573                 if (r)
 574                         ttm_handle_caching_state_failure(pages,
 575                                         ttm_flags, cstate,
 576                                         caching_array, cpages);
 577         }
 578 out:
 579         kfree(caching_array);
 580
 581         return r;
 582 }
 583
 584 /**
 585  * Fill the given pool if there aren't enough pages and the requested number of
 586  * pages is small.
 587  */
 588 static void ttm_page_pool_fill_locked(struct ttm_page_pool *pool, int ttm_flags,
 589                                       enum ttm_caching_state cstate,
 590                                       unsigned count, unsigned long *irq_flags)
 591 {
 592         struct page *p;
 593         int r;
 594         unsigned cpages = 0;
 595         /**
 596          * Only allow one pool fill operation at a time.
 597          * If pool doesn't have enough pages for the allocation new pages are
 598          * allocated from outside of pool.
 599          */
 600         if (pool->fill_lock)
 601                 return;
 602
 603         pool->fill_lock = true;
 604
 605         /* If allocation request is small and there are not enough
 606          * pages in a pool we fill the pool up first. */
 607         if (count < _manager->options.small
 608                 && count > pool->npages) {
 609                 struct list_head new_pages;
 610                 unsigned alloc_size = _manager->options.alloc_size;
 611
 612                 /**
 613                  * Can't change page caching if in irqsave context. We have to
 614                  * drop the pool->lock.
 615                  */
 616                 spin_unlock_irqrestore(&pool->lock, *irq_flags);
 617
 618                 INIT_LIST_HEAD(&new_pages);
 619                 r = ttm_alloc_new_pages(&new_pages, pool->gfp_flags, ttm_flags,
 620                                         cstate, alloc_size, 0);
 621                 spin_lock_irqsave(&pool->lock, *irq_flags);
 622
 623                 if (!r) {
 624                         list_splice(&new_pages, &pool->list);
 625                         ++pool->nrefills;
 626                         pool->npages += alloc_size;
 627                 } else {
 628                         pr_debug("Failed to fill pool (%p)\n", pool);
 629                         /* If we have any pages left put them to the pool. */
 630                         list_for_each_entry(p, &new_pages, lru) {
 631                                 ++cpages;
 632                         }
 633                         list_splice(&new_pages, &pool->list);
 634                         pool->npages += cpages;
 635                 }
 636
 637         }
 638         pool->fill_lock = false;
 639 }
 640
 641 /**
 642  * Allocate pages from the pool and put them on the return list.
 643  *
 644  * @return zero for success or negative error code.
 645  */
 646 static int ttm_page_pool_get_pages(struct ttm_page_pool *pool,
 647                                    struct list_head *pages,
 648                                    int ttm_flags,
 649                                    enum ttm_caching_state cstate,
 650                                    unsigned count, unsigned order)
 651 {
 652         unsigned long irq_flags;
 653         struct list_head *p;
 654         unsigned i;
 655         int r = 0;
 656
 657         spin_lock_irqsave(&pool->lock, irq_flags);
 658         if (!order)
 659                 ttm_page_pool_fill_locked(pool, ttm_flags, cstate, count,
 660                                           &irq_flags);
 661
 662         if (count >= pool->npages) {
 663                 /* take all pages from the pool */
 664                 list_splice_init(&pool->list, pages);
 665                 count -= pool->npages;
 666                 pool->npages = 0;
 667                 goto out;
 668         }
 669         /* find the last pages to include for requested number of pages. Split
 670          * pool to begin and halve it to reduce search space. */
 671         if (count <= pool->npages/2) {
 672                 i = 0;
 673                 list_for_each(p, &pool->list) {
 674                         if (++i == count)
 675                                 break;
 676                 }
 677         } else {
 678                 i = pool->npages + 1;
 679                 list_for_each_prev(p, &pool->list) {
 680                         if (--i == count)
 681                                 break;
 682                 }
 683         }
 684         /* Cut 'count' number of pages from the pool */
 685         list_cut_position(pages, &pool->list, p);
 686         pool->npages -= count;
 687         count = 0;
 688 out:
 689         spin_unlock_irqrestore(&pool->lock, irq_flags);
 690
 691         /* clear the pages coming from the pool if requested */
 692         if (ttm_flags & TTM_PAGE_FLAG_ZERO_ALLOC) {
 693                 struct page *page;
 694
 695                 list_for_each_entry(page, pages, lru) {
 696                         if (PageHighMem(page))
 697                                 clear_highpage(page);
 698                         else
 699                                 clear_page(page_address(page));
 700                 }
 701         }
 702
 703         /* If pool didn't have enough pages allocate new one. */
 704         if (count) {
 705                 gfp_t gfp_flags = pool->gfp_flags;
 706
 707                 /* set zero flag for page allocation if required */
 708                 if (ttm_flags & TTM_PAGE_FLAG_ZERO_ALLOC)
 709                         gfp_flags |= __GFP_ZERO;
 710
 711                 /* ttm_alloc_new_pages doesn't reference pool so we can run
 712                  * multiple requests in parallel.
 713                  **/
 714                 r = ttm_alloc_new_pages(pages, gfp_flags, ttm_flags, cstate,
 715                                         count, order);
 716         }
 717
 718         return r;
 719 }
 720
 721 /* Put all pages in pages list to correct pool to wait for reuse */
 722 static void ttm_put_pages(struct page **pages, unsigned npages, int flags,
 723                           enum ttm_caching_state cstate)
 724 {
 725         struct ttm_page_pool *pool = ttm_get_pool(flags, false, cstate);
 726 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 727         struct ttm_page_pool *huge = ttm_get_pool(flags, true, cstate);
 728 #endif
 729         unsigned long irq_flags;
 730         unsigned i;
 731
 732         if (pool == NULL) {
 733                 /* No pool for this memory type so free the pages */
 734                 i = 0;
 735                 while (i < npages) {
 736 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 737                         struct page *p = pages[i];
 738 #endif
 739                         unsigned order = 0, j;
 740
 741                         if (!pages[i]) {
 742                                 ++i;
 743                                 continue;
 744                         }
 745
 746 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 747                         if (!(flags & TTM_PAGE_FLAG_DMA32)) {
 748                                 for (j = 0; j < HPAGE_PMD_NR; ++j)
 749                                         if (p++ != pages[i + j])
 750                                             break;
 751
 752                                 if (j == HPAGE_PMD_NR)
 753                                         order = HPAGE_PMD_ORDER;
 754                         }
 755 #endif
 756
 757                         if (page_count(pages[i]) != 1)
 758                                 pr_err("Erroneous page count. Leaking pages.\n");
 759                         __free_pages(pages[i], order);
 760
 761                         j = 1 << order;
 762                         while (j) {
 763                                 pages[i++] = NULL;
 764                                 --j;
 765                         }
 766                 }
 767                 return;
 768         }
 769
 770         i = 0;
 771 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 772         if (huge) {
 773                 unsigned max_size, n2free;
 774
 775                 spin_lock_irqsave(&huge->lock, irq_flags);
 776                 while (i < npages) {
 777                         struct page *p = pages[i];
 778                         unsigned j;
 779
 780                         if (!p)
 781                                 break;
 782
 783                         for (j = 0; j < HPAGE_PMD_NR; ++j)
 784                                 if (p++ != pages[i + j])
 785                                     break;
 786
 787                         if (j != HPAGE_PMD_NR)
 788                                 break;
 789
 790                         list_add_tail(&pages[i]->lru, &huge->list);
 791
 792                         for (j = 0; j < HPAGE_PMD_NR; ++j)
 793                                 pages[i++] = NULL;
 794                         huge->npages++;
 795                 }
 796
 797                 /* Check that we don't go over the pool limit */
 798                 max_size = _manager->options.max_size;
 799                 max_size /= HPAGE_PMD_NR;
 800                 if (huge->npages > max_size)
 801                         n2free = huge->npages - max_size;
 802                 else
 803                         n2free = 0;
 804                 spin_unlock_irqrestore(&huge->lock, irq_flags);
 805                 if (n2free)
 806                         ttm_page_pool_free(huge, n2free, false);
 807         }
 808 #endif
 809
 810         spin_lock_irqsave(&pool->lock, irq_flags);
 811         while (i < npages) {
 812                 if (pages[i]) {
 813                         if (page_count(pages[i]) != 1)
 814                                 pr_err("Erroneous page count. Leaking pages.\n");
 815                         list_add_tail(&pages[i]->lru, &pool->list);
 816                         pages[i] = NULL;
 817                         pool->npages++;
 818                 }
 819                 ++i;
 820         }
 821         /* Check that we don't go over the pool limit */
 822         npages = 0;
 823         if (pool->npages > _manager->options.max_size) {
 824                 npages = pool->npages - _manager->options.max_size;
 825                 /* free at least NUM_PAGES_TO_ALLOC number of pages
 826                  * to reduce calls to set_memory_wb */
 827                 if (npages < NUM_PAGES_TO_ALLOC)
 828                         npages = NUM_PAGES_TO_ALLOC;
 829         }
 830         spin_unlock_irqrestore(&pool->lock, irq_flags);
 831         if (npages)
 832                 ttm_page_pool_free(pool, npages, false);
 833 }
 834
 835 /*
 836  * On success pages list will hold count number of correctly
 837  * cached pages.
 838  */
 839 static int ttm_get_pages(struct page **pages, unsigned npages, int flags,
 840                          enum ttm_caching_state cstate)
 841 {
 842         struct ttm_page_pool *pool = ttm_get_pool(flags, false, cstate);
 843 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 844         struct ttm_page_pool *huge = ttm_get_pool(flags, true, cstate);
 845 #endif
 846         struct list_head plist;
 847         struct page *p = NULL;
 848         unsigned count;
 849         int r;
 850
 851         /* No pool for cached pages */
 852         if (pool == NULL) {
 853                 gfp_t gfp_flags = GFP_USER;
 854                 unsigned i;
 855 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 856                 unsigned j;
 857 #endif
 858
 859                 /* set zero flag for page allocation if required */
 860                 if (flags & TTM_PAGE_FLAG_ZERO_ALLOC)
 861                         gfp_flags |= __GFP_ZERO;
 862
 863                 if (flags & TTM_PAGE_FLAG_DMA32)
 864                         gfp_flags |= GFP_DMA32;
 865                 else
 866                         gfp_flags |= GFP_HIGHUSER;
 867
 868                 i = 0;
 869 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 870                 if (!(gfp_flags & GFP_DMA32)) {
 871                         while (npages >= HPAGE_PMD_NR) {
 872                                 gfp_t huge_flags = gfp_flags;
 873
 874                                 huge_flags |= GFP_TRANSHUGE;
 875                                 huge_flags &= ~__GFP_MOVABLE;
 876                                 huge_flags &= ~__GFP_COMP;
 877                                 p = alloc_pages(huge_flags, HPAGE_PMD_ORDER);
 878                                 if (!p)
 879                                         break;
 880
 881                                 for (j = 0; j < HPAGE_PMD_NR; ++j)
 882                                         pages[i++] = p++;
 883
 884                                 npages -= HPAGE_PMD_NR;
 885                         }
 886                 }
 887 #endif
 888
 889                 while (npages) {
 890                         p = alloc_page(gfp_flags);
 891                         if (!p) {
 892                                 pr_debug("Unable to allocate page\n");
 893                                 return -ENOMEM;
 894                         }
 895
 896                         pages[i++] = p;
 897                         --npages;
 898                 }
 899                 return 0;
 900         }
 901
 902         count = 0;
 903
 904 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 905         if (huge && npages >= HPAGE_PMD_NR) {
 906                 INIT_LIST_HEAD(&plist);
 907                 ttm_page_pool_get_pages(huge, &plist, flags, cstate,
 908                                         npages / HPAGE_PMD_NR,
 909                                         HPAGE_PMD_ORDER);
 910
 911                 list_for_each_entry(p, &plist, lru) {
 912                         unsigned j;
 913
 914                         for (j = 0; j < HPAGE_PMD_NR; ++j)
 915                                 pages[count++] = &p[j];
 916                 }
 917         }
 918 #endif
 919
 920         INIT_LIST_HEAD(&plist);
 921         r = ttm_page_pool_get_pages(pool, &plist, flags, cstate,
 922                                     npages - count, 0);
 923
 924         list_for_each_entry(p, &plist, lru)
 925                 pages[count++] = p;
 926
 927         if (r) {
 928                 /* If there is any pages in the list put them back to
 929                  * the pool.
 930                  */
 931                 pr_debug("Failed to allocate extra pages for large request\n");
 932                 ttm_put_pages(pages, count, flags, cstate);
 933                 return r;
 934         }
 935
 936         return 0;
 937 }
 938
 939 static void ttm_page_pool_init_locked(struct ttm_page_pool *pool, gfp_t flags,
 940                 char *name)
 941 {
 942         spin_lock_init(&pool->lock);
 943         pool->fill_lock = false;
 944         INIT_LIST_HEAD(&pool->list);
 945         pool->npages = pool->nfrees = 0;
 946         pool->gfp_flags = flags;
 947         pool->name = name;
 948 }
 949
 950 int ttm_page_alloc_init(struct ttm_mem_global *glob, unsigned max_pages)
 951 {
 952         int ret;
 953
 954         WARN_ON(_manager);
 955
 956         pr_info("Initializing pool allocator\n");
 957
 958         _manager = kzalloc(sizeof(*_manager), GFP_KERNEL);
 959
 960         ttm_page_pool_init_locked(&_manager->wc_pool, GFP_HIGHUSER, "wc");
 961
 962         ttm_page_pool_init_locked(&_manager->uc_pool, GFP_HIGHUSER, "uc");
 963
 964         ttm_page_pool_init_locked(&_manager->wc_pool_dma32,
 965                                   GFP_USER | GFP_DMA32, "wc dma");
 966
 967         ttm_page_pool_init_locked(&_manager->uc_pool_dma32,
 968                                   GFP_USER | GFP_DMA32, "uc dma");
 969
 970         ttm_page_pool_init_locked(&_manager->wc_pool_huge,
 971                                   GFP_TRANSHUGE & ~(__GFP_MOVABLE | __GFP_COMP),
 972                                   "wc huge");
 973
 974         ttm_page_pool_init_locked(&_manager->uc_pool_huge,
 975                                   GFP_TRANSHUGE & ~(__GFP_MOVABLE | __GFP_COMP)
 976                                   , "uc huge");
 977
 978         _manager->options.max_size = max_pages;
 979         _manager->options.small = SMALL_ALLOCATION;
 980         _manager->options.alloc_size = NUM_PAGES_TO_ALLOC;
 981
 982         ret = kobject_init_and_add(&_manager->kobj, &ttm_pool_kobj_type,
 983                                    &glob->kobj, "pool");
 984         if (unlikely(ret != 0)) {
 985                 kobject_put(&_manager->kobj);
 986                 _manager = NULL;
 987                 return ret;
 988         }
 989
 990         ttm_pool_mm_shrink_init(_manager);
 991
 992         return 0;
 993 }
 994
 995 void ttm_page_alloc_fini(void)
 996 {
 997         int i;
 998
 999         pr_info("Finalizing pool allocator\n");
1000         ttm_pool_mm_shrink_fini(_manager);
1001
1002         /* OK to use static buffer since global mutex is no longer used. */
1003         for (i = 0; i < NUM_POOLS; ++i)
1004                 ttm_page_pool_free(&_manager->pools[i], FREE_ALL_PAGES, true);
1005
1006         kobject_put(&_manager->kobj);
1007         _manager = NULL;
1008 }
1009
1010 int ttm_pool_populate(struct ttm_tt *ttm)
1011 {
1012         struct ttm_mem_global *mem_glob = ttm->glob->mem_glob;
1013         unsigned i;
1014         int ret;
1015
1016         if (ttm->state != tt_unpopulated)
1017                 return 0;
1018
1019         ret = ttm_get_pages(ttm->pages, ttm->num_pages, ttm->page_flags,
1020                             ttm->caching_state);
1021         if (unlikely(ret != 0)) {
1022                 ttm_pool_unpopulate(ttm);
1023                 return ret;
1024         }
1025
1026         for (i = 0; i < ttm->num_pages; ++i) {
1027                 ret = ttm_mem_global_alloc_page(mem_glob, ttm->pages[i],
1028                                                 PAGE_SIZE);
1029                 if (unlikely(ret != 0)) {
1030                         ttm_pool_unpopulate(ttm);
1031                         return -ENOMEM;
1032                 }
1033         }
1034
1035         if (unlikely(ttm->page_flags & TTM_PAGE_FLAG_SWAPPED)) {
1036                 ret = ttm_tt_swapin(ttm);
1037                 if (unlikely(ret != 0)) {
1038                         ttm_pool_unpopulate(ttm);
1039                         return ret;
1040                 }
1041         }
1042
1043         ttm->state = tt_unbound;
1044         return 0;
1045 }
1046 EXPORT_SYMBOL(ttm_pool_populate);
1047
1048 void ttm_pool_unpopulate(struct ttm_tt *ttm)
1049 {
1050         unsigned i;
1051
1052         for (i = 0; i < ttm->num_pages; ++i) {
1053                 if (!ttm->pages[i])
1054                         continue;
1055
1056                 ttm_mem_global_free_page(ttm->glob->mem_glob, ttm->pages[i],
1057                                          PAGE_SIZE);
1058         }
1059         ttm_put_pages(ttm->pages, ttm->num_pages, ttm->page_flags,
1060                       ttm->caching_state);
1061         ttm->state = tt_unpopulated;
1062 }
1063 EXPORT_SYMBOL(ttm_pool_unpopulate);
1064
1065 #if defined(CONFIG_SWIOTLB) || defined(CONFIG_INTEL_IOMMU)
1066 int ttm_populate_and_map_pages(struct device *dev, struct ttm_dma_tt *tt)
1067 {
1068         unsigned i, j;
1069         int r;
1070
1071         r = ttm_pool_populate(&tt->ttm);
1072         if (r)
1073                 return r;
1074
1075         for (i = 0; i < tt->ttm.num_pages; ++i) {
1076                 struct page *p = tt->ttm.pages[i];
1077                 size_t num_pages = 1;
1078
1079                 for (j = i + 1; j < tt->ttm.num_pages; ++j) {
1080                         if (++p != tt->ttm.pages[j])
1081                                 break;
1082
1083                         ++num_pages;
1084                 }
1085
1086                 tt->dma_address[i] = dma_map_page(dev, tt->ttm.pages[i],
1087                                                   0, num_pages * PAGE_SIZE,
1088                                                   DMA_BIDIRECTIONAL);
1089                 if (dma_mapping_error(dev, tt->dma_address[i])) {
1090                         while (i--) {
1091                                 dma_unmap_page(dev, tt->dma_address[i],
1092                                                PAGE_SIZE, DMA_BIDIRECTIONAL);
1093                                 tt->dma_address[i] = 0;
1094                         }
1095                         ttm_pool_unpopulate(&tt->ttm);
1096                         return -EFAULT;
1097                 }
1098
1099                 for (j = 1; j < num_pages; ++j) {
1100                         tt->dma_address[i + 1] = tt->dma_address[i] + PAGE_SIZE;
1101                         ++i;
1102                 }
1103         }
1104         return 0;
1105 }
1106 EXPORT_SYMBOL(ttm_populate_and_map_pages);
1107
1108 void ttm_unmap_and_unpopulate_pages(struct device *dev, struct ttm_dma_tt *tt)
1109 {
1110         unsigned i, j;
1111
1112         for (i = 0; i < tt->ttm.num_pages;) {
1113                 struct page *p = tt->ttm.pages[i];
1114                 size_t num_pages = 1;
1115
1116                 if (!tt->dma_address[i] || !tt->ttm.pages[i]) {
1117                         ++i;
1118                         continue;
1119                 }
1120
1121                 for (j = i + 1; j < tt->ttm.num_pages; ++j) {
1122                         if (++p != tt->ttm.pages[j])
1123                                 break;
1124
1125                         ++num_pages;
1126                 }
1127
1128                 dma_unmap_page(dev, tt->dma_address[i], num_pages * PAGE_SIZE,
1129                                DMA_BIDIRECTIONAL);
1130
1131                 i += num_pages;
1132         }
1133         ttm_pool_unpopulate(&tt->ttm);
1134 }
1135 EXPORT_SYMBOL(ttm_unmap_and_unpopulate_pages);
1136 #endif
1137
1138 int ttm_page_alloc_debugfs(struct seq_file *m, void *data)
1139 {
1140         struct ttm_page_pool *p;
1141         unsigned i;
1142         char *h[] = {"pool", "refills", "pages freed", "size"};
1143         if (!_manager) {
1144                 seq_printf(m, "No pool allocator running.\n");
1145                 return 0;
1146         }
1147         seq_printf(m, "%7s %12s %13s %8s\n",
1148                         h[0], h[1], h[2], h[3]);
1149         for (i = 0; i < NUM_POOLS; ++i) {
1150                 p = &_manager->pools[i];
1151
1152                 seq_printf(m, "%7s %12ld %13ld %8d\n",
1153                                 p->name, p->nrefills,
1154                                 p->nfrees, p->npages);
1155         }
1156         return 0;
1157 }
1158 EXPORT_SYMBOL(ttm_page_alloc_debugfs);