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 <[email protected]>
  24  *          Jerome Glisse <[email protected]>
  25  *          Pauli Nieminen <[email protected]>
  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         unsigned int            order;
  85 };
  86
  87 /**
  88  * Limits for the pool. They are handled without locks because only place where
  89  * they may change is in sysfs store. They won't have immediate effect anyway
  90  * so forcing serialization to access them is pointless.
  91  */
  92
  93 struct ttm_pool_opts {
  94         unsigned        alloc_size;
  95         unsigned        max_size;
  96         unsigned        small;
  97 };
  98
  99 #define NUM_POOLS 6
 100
 101 /**
 102  * struct ttm_pool_manager - Holds memory pools for fst allocation
 103  *
 104  * Manager is read only object for pool code so it doesn't need locking.
 105  *
 106  * @free_interval: minimum number of jiffies between freeing pages from pool.
 107  * @page_alloc_inited: reference counting for pool allocation.
 108  * @work: Work that is used to shrink the pool. Work is only run when there is
 109  * some pages to free.
 110  * @small_allocation: Limit in number of pages what is small allocation.
 111  *
 112  * @pools: All pool objects in use.
 113  **/
 114 struct ttm_pool_manager {
 115         struct kobject          kobj;
 116         struct shrinker         mm_shrink;
 117         struct ttm_pool_opts    options;
 118
 119         union {
 120                 struct ttm_page_pool    pools[NUM_POOLS];
 121                 struct {
 122                         struct ttm_page_pool    wc_pool;
 123                         struct ttm_page_pool    uc_pool;
 124                         struct ttm_page_pool    wc_pool_dma32;
 125                         struct ttm_page_pool    uc_pool_dma32;
 126                         struct ttm_page_pool    wc_pool_huge;
 127                         struct ttm_page_pool    uc_pool_huge;
 128                 } ;
 129         };
 130 };
 131
 132 static struct attribute ttm_page_pool_max = {
 133         .name = "pool_max_size",
 134         .mode = S_IRUGO | S_IWUSR
 135 };
 136 static struct attribute ttm_page_pool_small = {
 137         .name = "pool_small_allocation",
 138         .mode = S_IRUGO | S_IWUSR
 139 };
 140 static struct attribute ttm_page_pool_alloc_size = {
 141         .name = "pool_allocation_size",
 142         .mode = S_IRUGO | S_IWUSR
 143 };
 144
 145 static struct attribute *ttm_pool_attrs[] = {
 146         &ttm_page_pool_max,
 147         &ttm_page_pool_small,
 148         &ttm_page_pool_alloc_size,
 149         NULL
 150 };
 151
 152 static void ttm_pool_kobj_release(struct kobject *kobj)
 153 {
 154         struct ttm_pool_manager *m =
 155                 container_of(kobj, struct ttm_pool_manager, kobj);
 156         kfree(m);
 157 }
 158
 159 static ssize_t ttm_pool_store(struct kobject *kobj,
 160                 struct attribute *attr, const char *buffer, size_t size)
 161 {
 162         struct ttm_pool_manager *m =
 163                 container_of(kobj, struct ttm_pool_manager, kobj);
 164         int chars;
 165         unsigned val;
 166         chars = sscanf(buffer, "%u", &val);
 167         if (chars == 0)
 168                 return size;
 169
 170         /* Convert kb to number of pages */
 171         val = val / (PAGE_SIZE >> 10);
 172
 173         if (attr == &ttm_page_pool_max)
 174                 m->options.max_size = val;
 175         else if (attr == &ttm_page_pool_small)
 176                 m->options.small = val;
 177         else if (attr == &ttm_page_pool_alloc_size) {
 178                 if (val > NUM_PAGES_TO_ALLOC*8) {
 179                         pr_err("Setting allocation size to %lu is not allowed. Recommended size is %lu\n",
 180                                NUM_PAGES_TO_ALLOC*(PAGE_SIZE >> 7),
 181                                NUM_PAGES_TO_ALLOC*(PAGE_SIZE >> 10));
 182                         return size;
 183                 } else if (val > NUM_PAGES_TO_ALLOC) {
 184                         pr_warn("Setting allocation size to larger than %lu is not recommended\n",
 185                                 NUM_PAGES_TO_ALLOC*(PAGE_SIZE >> 10));
 186                 }
 187                 m->options.alloc_size = val;
 188         }
 189
 190         return size;
 191 }
 192
 193 static ssize_t ttm_pool_show(struct kobject *kobj,
 194                 struct attribute *attr, char *buffer)
 195 {
 196         struct ttm_pool_manager *m =
 197                 container_of(kobj, struct ttm_pool_manager, kobj);
 198         unsigned val = 0;
 199
 200         if (attr == &ttm_page_pool_max)
 201                 val = m->options.max_size;
 202         else if (attr == &ttm_page_pool_small)
 203                 val = m->options.small;
 204         else if (attr == &ttm_page_pool_alloc_size)
 205                 val = m->options.alloc_size;
 206
 207         val = val * (PAGE_SIZE >> 10);
 208
 209         return snprintf(buffer, PAGE_SIZE, "%u\n", val);
 210 }
 211
 212 static const struct sysfs_ops ttm_pool_sysfs_ops = {
 213         .show = &ttm_pool_show,
 214         .store = &ttm_pool_store,
 215 };
 216
 217 static struct kobj_type ttm_pool_kobj_type = {
 218         .release = &ttm_pool_kobj_release,
 219         .sysfs_ops = &ttm_pool_sysfs_ops,
 220         .default_attrs = ttm_pool_attrs,
 221 };
 222
 223 static struct ttm_pool_manager *_manager;
 224
 225 #ifndef CONFIG_X86
 226 static int set_pages_wb(struct page *page, int numpages)
 227 {
 228 #if IS_ENABLED(CONFIG_AGP)
 229         int i;
 230
 231         for (i = 0; i < numpages; i++)
 232                 unmap_page_from_agp(page++);
 233 #endif
 234         return 0;
 235 }
 236
 237 static int set_pages_array_wb(struct page **pages, int addrinarray)
 238 {
 239 #if IS_ENABLED(CONFIG_AGP)
 240         int i;
 241
 242         for (i = 0; i < addrinarray; i++)
 243                 unmap_page_from_agp(pages[i]);
 244 #endif
 245         return 0;
 246 }
 247
 248 static int set_pages_array_wc(struct page **pages, int addrinarray)
 249 {
 250 #if IS_ENABLED(CONFIG_AGP)
 251         int i;
 252
 253         for (i = 0; i < addrinarray; i++)
 254                 map_page_into_agp(pages[i]);
 255 #endif
 256         return 0;
 257 }
 258
 259 static int set_pages_array_uc(struct page **pages, int addrinarray)
 260 {
 261 #if IS_ENABLED(CONFIG_AGP)
 262         int i;
 263
 264         for (i = 0; i < addrinarray; i++)
 265                 map_page_into_agp(pages[i]);
 266 #endif
 267         return 0;
 268 }
 269 #endif
 270
 271 /**
 272  * Select the right pool or requested caching state and ttm flags. */
 273 static struct ttm_page_pool *ttm_get_pool(int flags, bool huge,
 274                                           enum ttm_caching_state cstate)
 275 {
 276         int pool_index;
 277
 278         if (cstate == tt_cached)
 279                 return NULL;
 280
 281         if (cstate == tt_wc)
 282                 pool_index = 0x0;
 283         else
 284                 pool_index = 0x1;
 285
 286         if (flags & TTM_PAGE_FLAG_DMA32) {
 287                 if (huge)
 288                         return NULL;
 289                 pool_index |= 0x2;
 290
 291         } else if (huge) {
 292                 pool_index |= 0x4;
 293         }
 294
 295         return &_manager->pools[pool_index];
 296 }
 297
 298 /* set memory back to wb and free the pages. */
 299 static void ttm_pages_put(struct page *pages[], unsigned npages,
 300                 unsigned int order)
 301 {
 302         unsigned int i, pages_nr = (1 << order);
 303
 304         if (order == 0) {
 305                 if (set_pages_array_wb(pages, npages))
 306                         pr_err("Failed to set %d pages to wb!\n", npages);
 307         }
 308
 309         for (i = 0; i < npages; ++i) {
 310                 if (order > 0) {
 311                         if (set_pages_wb(pages[i], pages_nr))
 312                                 pr_err("Failed to set %d pages to wb!\n", pages_nr);
 313                 }
 314                 __free_pages(pages[i], order);
 315         }
 316 }
 317
 318 static void ttm_pool_update_free_locked(struct ttm_page_pool *pool,
 319                 unsigned freed_pages)
 320 {
 321         pool->npages -= freed_pages;
 322         pool->nfrees += freed_pages;
 323 }
 324
 325 /**
 326  * Free pages from pool.
 327  *
 328  * To prevent hogging the ttm_swap process we only free NUM_PAGES_TO_ALLOC
 329  * number of pages in one go.
 330  *
 331  * @pool: to free the pages from
 332  * @free_all: If set to true will free all pages in pool
 333  * @use_static: Safe to use static buffer
 334  **/
 335 static int ttm_page_pool_free(struct ttm_page_pool *pool, unsigned nr_free,
 336                               bool use_static)
 337 {
 338         static struct page *static_buf[NUM_PAGES_TO_ALLOC];
 339         unsigned long irq_flags;
 340         struct page *p;
 341         struct page **pages_to_free;
 342         unsigned freed_pages = 0,
 343                  npages_to_free = nr_free;
 344
 345         if (NUM_PAGES_TO_ALLOC < nr_free)
 346                 npages_to_free = NUM_PAGES_TO_ALLOC;
 347
 348         if (use_static)
 349                 pages_to_free = static_buf;
 350         else
 351                 pages_to_free = kmalloc_array(npages_to_free,
 352                                               sizeof(struct page *),
 353                                               GFP_KERNEL);
 354         if (!pages_to_free) {
 355                 pr_debug("Failed to allocate memory for pool free operation\n");
 356                 return 0;
 357         }
 358
 359 restart:
 360         spin_lock_irqsave(&pool->lock, irq_flags);
 361
 362         list_for_each_entry_reverse(p, &pool->list, lru) {
 363                 if (freed_pages >= npages_to_free)
 364                         break;
 365
 366                 pages_to_free[freed_pages++] = p;
 367                 /* We can only remove NUM_PAGES_TO_ALLOC at a time. */
 368                 if (freed_pages >= NUM_PAGES_TO_ALLOC) {
 369                         /* remove range of pages from the pool */
 370                         __list_del(p->lru.prev, &pool->list);
 371
 372                         ttm_pool_update_free_locked(pool, freed_pages);
 373                         /**
 374                          * Because changing page caching is costly
 375                          * we unlock the pool to prevent stalling.
 376                          */
 377                         spin_unlock_irqrestore(&pool->lock, irq_flags);
 378
 379                         ttm_pages_put(pages_to_free, freed_pages, pool->order);
 380                         if (likely(nr_free != FREE_ALL_PAGES))
 381                                 nr_free -= freed_pages;
 382
 383                         if (NUM_PAGES_TO_ALLOC >= nr_free)
 384                                 npages_to_free = nr_free;
 385                         else
 386                                 npages_to_free = NUM_PAGES_TO_ALLOC;
 387
 388                         freed_pages = 0;
 389
 390                         /* free all so restart the processing */
 391                         if (nr_free)
 392                                 goto restart;
 393
 394                         /* Not allowed to fall through or break because
 395                          * following context is inside spinlock while we are
 396                          * outside here.
 397                          */
 398                         goto out;
 399
 400                 }
 401         }
 402
 403         /* remove range of pages from the pool */
 404         if (freed_pages) {
 405                 __list_del(&p->lru, &pool->list);
 406
 407                 ttm_pool_update_free_locked(pool, freed_pages);
 408                 nr_free -= freed_pages;
 409         }
 410
 411         spin_unlock_irqrestore(&pool->lock, irq_flags);
 412
 413         if (freed_pages)
 414                 ttm_pages_put(pages_to_free, freed_pages, pool->order);
 415 out:
 416         if (pages_to_free != static_buf)
 417                 kfree(pages_to_free);
 418         return nr_free;
 419 }
 420
 421 /**
 422  * Callback for mm to request pool to reduce number of page held.
 423  *
 424  * XXX: (dchinner) Deadlock warning!
 425  *
 426  * This code is crying out for a shrinker per pool....
 427  */
 428 static unsigned long
 429 ttm_pool_shrink_scan(struct shrinker *shrink, struct shrink_control *sc)
 430 {
 431         static DEFINE_MUTEX(lock);
 432         static unsigned start_pool;
 433         unsigned i;
 434         unsigned pool_offset;
 435         struct ttm_page_pool *pool;
 436         int shrink_pages = sc->nr_to_scan;
 437         unsigned long freed = 0;
 438         unsigned int nr_free_pool;
 439
 440         if (!mutex_trylock(&lock))
 441                 return SHRINK_STOP;
 442         pool_offset = ++start_pool % NUM_POOLS;
 443         /* select start pool in round robin fashion */
 444         for (i = 0; i < NUM_POOLS; ++i) {
 445                 unsigned nr_free = shrink_pages;
 446                 unsigned page_nr;
 447
 448                 if (shrink_pages == 0)
 449                         break;
 450
 451                 pool = &_manager->pools[(i + pool_offset)%NUM_POOLS];
 452                 page_nr = (1 << pool->order);
 453                 /* OK to use static buffer since global mutex is held. */
 454                 nr_free_pool = roundup(nr_free, page_nr) >> pool->order;
 455                 shrink_pages = ttm_page_pool_free(pool, nr_free_pool, true);
 456                 freed += (nr_free_pool - shrink_pages) << pool->order;
 457                 if (freed >= sc->nr_to_scan)
 458                         break;
 459                 shrink_pages <<= pool->order;
 460         }
 461         mutex_unlock(&lock);
 462         return freed;
 463 }
 464
 465
 466 static unsigned long
 467 ttm_pool_shrink_count(struct shrinker *shrink, struct shrink_control *sc)
 468 {
 469         unsigned i;
 470         unsigned long count = 0;
 471         struct ttm_page_pool *pool;
 472
 473         for (i = 0; i < NUM_POOLS; ++i) {
 474                 pool = &_manager->pools[i];
 475                 count += (pool->npages << pool->order);
 476         }
 477
 478         return count;
 479 }
 480
 481 static int ttm_pool_mm_shrink_init(struct ttm_pool_manager *manager)
 482 {
 483         manager->mm_shrink.count_objects = ttm_pool_shrink_count;
 484         manager->mm_shrink.scan_objects = ttm_pool_shrink_scan;
 485         manager->mm_shrink.seeks = 1;
 486         return register_shrinker(&manager->mm_shrink);
 487 }
 488
 489 static void ttm_pool_mm_shrink_fini(struct ttm_pool_manager *manager)
 490 {
 491         unregister_shrinker(&manager->mm_shrink);
 492 }
 493
 494 static int ttm_set_pages_caching(struct page **pages,
 495                 enum ttm_caching_state cstate, unsigned cpages)
 496 {
 497         int r = 0;
 498         /* Set page caching */
 499         switch (cstate) {
 500         case tt_uncached:
 501                 r = set_pages_array_uc(pages, cpages);
 502                 if (r)
 503                         pr_err("Failed to set %d pages to uc!\n", cpages);
 504                 break;
 505         case tt_wc:
 506                 r = set_pages_array_wc(pages, cpages);
 507                 if (r)
 508                         pr_err("Failed to set %d pages to wc!\n", cpages);
 509                 break;
 510         default:
 511                 break;
 512         }
 513         return r;
 514 }
 515
 516 /**
 517  * Free pages the pages that failed to change the caching state. If there is
 518  * any pages that have changed their caching state already put them to the
 519  * pool.
 520  */
 521 static void ttm_handle_caching_state_failure(struct list_head *pages,
 522                 int ttm_flags, enum ttm_caching_state cstate,
 523                 struct page **failed_pages, unsigned cpages)
 524 {
 525         unsigned i;
 526         /* Failed pages have to be freed */
 527         for (i = 0; i < cpages; ++i) {
 528                 list_del(&failed_pages[i]->lru);
 529                 __free_page(failed_pages[i]);
 530         }
 531 }
 532
 533 /**
 534  * Allocate new pages with correct caching.
 535  *
 536  * This function is reentrant if caller updates count depending on number of
 537  * pages returned in pages array.
 538  */
 539 static int ttm_alloc_new_pages(struct list_head *pages, gfp_t gfp_flags,
 540                                int ttm_flags, enum ttm_caching_state cstate,
 541                                unsigned count, unsigned order)
 542 {
 543         struct page **caching_array;
 544         struct page *p;
 545         int r = 0;
 546         unsigned i, j, cpages;
 547         unsigned npages = 1 << order;
 548         unsigned max_cpages = min(count << order, (unsigned)NUM_PAGES_TO_ALLOC);
 549
 550         /* allocate array for page caching change */
 551         caching_array = kmalloc_array(max_cpages, sizeof(struct page *),
 552                                       GFP_KERNEL);
 553
 554         if (!caching_array) {
 555                 pr_debug("Unable to allocate table for new pages\n");
 556                 return -ENOMEM;
 557         }
 558
 559         for (i = 0, cpages = 0; i < count; ++i) {
 560                 p = alloc_pages(gfp_flags, order);
 561
 562                 if (!p) {
 563                         pr_debug("Unable to get page %u\n", i);
 564
 565                         /* store already allocated pages in the pool after
 566                          * setting the caching state */
 567                         if (cpages) {
 568                                 r = ttm_set_pages_caching(caching_array,
 569                                                           cstate, cpages);
 570                                 if (r)
 571                                         ttm_handle_caching_state_failure(pages,
 572                                                 ttm_flags, cstate,
 573                                                 caching_array, cpages);
 574                         }
 575                         r = -ENOMEM;
 576                         goto out;
 577                 }
 578
 579                 list_add(&p->lru, pages);
 580
 581 #ifdef CONFIG_HIGHMEM
 582                 /* gfp flags of highmem page should never be dma32 so we
 583                  * we should be fine in such case
 584                  */
 585                 if (PageHighMem(p))
 586                         continue;
 587
 588 #endif
 589                 for (j = 0; j < npages; ++j) {
 590                         caching_array[cpages++] = p++;
 591                         if (cpages == max_cpages) {
 592
 593                                 r = ttm_set_pages_caching(caching_array,
 594                                                 cstate, cpages);
 595                                 if (r) {
 596                                         ttm_handle_caching_state_failure(pages,
 597                                                 ttm_flags, cstate,
 598                                                 caching_array, cpages);
 599                                         goto out;
 600                                 }
 601                                 cpages = 0;
 602                         }
 603                 }
 604         }
 605
 606         if (cpages) {
 607                 r = ttm_set_pages_caching(caching_array, cstate, cpages);
 608                 if (r)
 609                         ttm_handle_caching_state_failure(pages,
 610                                         ttm_flags, cstate,
 611                                         caching_array, cpages);
 612         }
 613 out:
 614         kfree(caching_array);
 615
 616         return r;
 617 }
 618
 619 /**
 620  * Fill the given pool if there aren't enough pages and the requested number of
 621  * pages is small.
 622  */
 623 static void ttm_page_pool_fill_locked(struct ttm_page_pool *pool, int ttm_flags,
 624                                       enum ttm_caching_state cstate,
 625                                       unsigned count, unsigned long *irq_flags)
 626 {
 627         struct page *p;
 628         int r;
 629         unsigned cpages = 0;
 630         /**
 631          * Only allow one pool fill operation at a time.
 632          * If pool doesn't have enough pages for the allocation new pages are
 633          * allocated from outside of pool.
 634          */
 635         if (pool->fill_lock)
 636                 return;
 637
 638         pool->fill_lock = true;
 639
 640         /* If allocation request is small and there are not enough
 641          * pages in a pool we fill the pool up first. */
 642         if (count < _manager->options.small
 643                 && count > pool->npages) {
 644                 struct list_head new_pages;
 645                 unsigned alloc_size = _manager->options.alloc_size;
 646
 647                 /**
 648                  * Can't change page caching if in irqsave context. We have to
 649                  * drop the pool->lock.
 650                  */
 651                 spin_unlock_irqrestore(&pool->lock, *irq_flags);
 652
 653                 INIT_LIST_HEAD(&new_pages);
 654                 r = ttm_alloc_new_pages(&new_pages, pool->gfp_flags, ttm_flags,
 655                                         cstate, alloc_size, 0);
 656                 spin_lock_irqsave(&pool->lock, *irq_flags);
 657
 658                 if (!r) {
 659                         list_splice(&new_pages, &pool->list);
 660                         ++pool->nrefills;
 661                         pool->npages += alloc_size;
 662                 } else {
 663                         pr_debug("Failed to fill pool (%p)\n", pool);
 664                         /* If we have any pages left put them to the pool. */
 665                         list_for_each_entry(p, &new_pages, lru) {
 666                                 ++cpages;
 667                         }
 668                         list_splice(&new_pages, &pool->list);
 669                         pool->npages += cpages;
 670                 }
 671
 672         }
 673         pool->fill_lock = false;
 674 }
 675
 676 /**
 677  * Allocate pages from the pool and put them on the return list.
 678  *
 679  * @return zero for success or negative error code.
 680  */
 681 static int ttm_page_pool_get_pages(struct ttm_page_pool *pool,
 682                                    struct list_head *pages,
 683                                    int ttm_flags,
 684                                    enum ttm_caching_state cstate,
 685                                    unsigned count, unsigned order)
 686 {
 687         unsigned long irq_flags;
 688         struct list_head *p;
 689         unsigned i;
 690         int r = 0;
 691
 692         spin_lock_irqsave(&pool->lock, irq_flags);
 693         if (!order)
 694                 ttm_page_pool_fill_locked(pool, ttm_flags, cstate, count,
 695                                           &irq_flags);
 696
 697         if (count >= pool->npages) {
 698                 /* take all pages from the pool */
 699                 list_splice_init(&pool->list, pages);
 700                 count -= pool->npages;
 701                 pool->npages = 0;
 702                 goto out;
 703         }
 704         /* find the last pages to include for requested number of pages. Split
 705          * pool to begin and halve it to reduce search space. */
 706         if (count <= pool->npages/2) {
 707                 i = 0;
 708                 list_for_each(p, &pool->list) {
 709                         if (++i == count)
 710                                 break;
 711                 }
 712         } else {
 713                 i = pool->npages + 1;
 714                 list_for_each_prev(p, &pool->list) {
 715                         if (--i == count)
 716                                 break;
 717                 }
 718         }
 719         /* Cut 'count' number of pages from the pool */
 720         list_cut_position(pages, &pool->list, p);
 721         pool->npages -= count;
 722         count = 0;
 723 out:
 724         spin_unlock_irqrestore(&pool->lock, irq_flags);
 725
 726         /* clear the pages coming from the pool if requested */
 727         if (ttm_flags & TTM_PAGE_FLAG_ZERO_ALLOC) {
 728                 struct page *page;
 729
 730                 list_for_each_entry(page, pages, lru) {
 731                         if (PageHighMem(page))
 732                                 clear_highpage(page);
 733                         else
 734                                 clear_page(page_address(page));
 735                 }
 736         }
 737
 738         /* If pool didn't have enough pages allocate new one. */
 739         if (count) {
 740                 gfp_t gfp_flags = pool->gfp_flags;
 741
 742                 /* set zero flag for page allocation if required */
 743                 if (ttm_flags & TTM_PAGE_FLAG_ZERO_ALLOC)
 744                         gfp_flags |= __GFP_ZERO;
 745
 746                 if (ttm_flags & TTM_PAGE_FLAG_NO_RETRY)
 747                         gfp_flags |= __GFP_RETRY_MAYFAIL;
 748
 749                 /* ttm_alloc_new_pages doesn't reference pool so we can run
 750                  * multiple requests in parallel.
 751                  **/
 752                 r = ttm_alloc_new_pages(pages, gfp_flags, ttm_flags, cstate,
 753                                         count, order);
 754         }
 755
 756         return r;
 757 }
 758
 759 /* Put all pages in pages list to correct pool to wait for reuse */
 760 static void ttm_put_pages(struct page **pages, unsigned npages, int flags,
 761                           enum ttm_caching_state cstate)
 762 {
 763         struct ttm_page_pool *pool = ttm_get_pool(flags, false, cstate);
 764 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 765         struct ttm_page_pool *huge = ttm_get_pool(flags, true, cstate);
 766 #endif
 767         unsigned long irq_flags;
 768         unsigned i;
 769
 770         if (pool == NULL) {
 771                 /* No pool for this memory type so free the pages */
 772                 i = 0;
 773                 while (i < npages) {
 774 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 775                         struct page *p = pages[i];
 776 #endif
 777                         unsigned order = 0, j;
 778
 779                         if (!pages[i]) {
 780                                 ++i;
 781                                 continue;
 782                         }
 783
 784 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 785                         if (!(flags & TTM_PAGE_FLAG_DMA32)) {
 786                                 for (j = 0; j < HPAGE_PMD_NR; ++j)
 787                                         if (p++ != pages[i + j])
 788                                             break;
 789
 790                                 if (j == HPAGE_PMD_NR)
 791                                         order = HPAGE_PMD_ORDER;
 792                         }
 793 #endif
 794
 795                         if (page_count(pages[i]) != 1)
 796                                 pr_err("Erroneous page count. Leaking pages.\n");
 797                         __free_pages(pages[i], order);
 798
 799                         j = 1 << order;
 800                         while (j) {
 801                                 pages[i++] = NULL;
 802                                 --j;
 803                         }
 804                 }
 805                 return;
 806         }
 807
 808         i = 0;
 809 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 810         if (huge) {
 811                 unsigned max_size, n2free;
 812
 813                 spin_lock_irqsave(&huge->lock, irq_flags);
 814                 while (i < npages) {
 815                         struct page *p = pages[i];
 816                         unsigned j;
 817
 818                         if (!p)
 819                                 break;
 820
 821                         for (j = 0; j < HPAGE_PMD_NR; ++j)
 822                                 if (p++ != pages[i + j])
 823                                     break;
 824
 825                         if (j != HPAGE_PMD_NR)
 826                                 break;
 827
 828                         list_add_tail(&pages[i]->lru, &huge->list);
 829
 830                         for (j = 0; j < HPAGE_PMD_NR; ++j)
 831                                 pages[i++] = NULL;
 832                         huge->npages++;
 833                 }
 834
 835                 /* Check that we don't go over the pool limit */
 836                 max_size = _manager->options.max_size;
 837                 max_size /= HPAGE_PMD_NR;
 838                 if (huge->npages > max_size)
 839                         n2free = huge->npages - max_size;
 840                 else
 841                         n2free = 0;
 842                 spin_unlock_irqrestore(&huge->lock, irq_flags);
 843                 if (n2free)
 844                         ttm_page_pool_free(huge, n2free, false);
 845         }
 846 #endif
 847
 848         spin_lock_irqsave(&pool->lock, irq_flags);
 849         while (i < npages) {
 850                 if (pages[i]) {
 851                         if (page_count(pages[i]) != 1)
 852                                 pr_err("Erroneous page count. Leaking pages.\n");
 853                         list_add_tail(&pages[i]->lru, &pool->list);
 854                         pages[i] = NULL;
 855                         pool->npages++;
 856                 }
 857                 ++i;
 858         }
 859         /* Check that we don't go over the pool limit */
 860         npages = 0;
 861         if (pool->npages > _manager->options.max_size) {
 862                 npages = pool->npages - _manager->options.max_size;
 863                 /* free at least NUM_PAGES_TO_ALLOC number of pages
 864                  * to reduce calls to set_memory_wb */
 865                 if (npages < NUM_PAGES_TO_ALLOC)
 866                         npages = NUM_PAGES_TO_ALLOC;
 867         }
 868         spin_unlock_irqrestore(&pool->lock, irq_flags);
 869         if (npages)
 870                 ttm_page_pool_free(pool, npages, false);
 871 }
 872
 873 /*
 874  * On success pages list will hold count number of correctly
 875  * cached pages.
 876  */
 877 static int ttm_get_pages(struct page **pages, unsigned npages, int flags,
 878                          enum ttm_caching_state cstate)
 879 {
 880         struct ttm_page_pool *pool = ttm_get_pool(flags, false, cstate);
 881 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 882         struct ttm_page_pool *huge = ttm_get_pool(flags, true, cstate);
 883 #endif
 884         struct list_head plist;
 885         struct page *p = NULL;
 886         unsigned count, first;
 887         int r;
 888
 889         /* No pool for cached pages */
 890         if (pool == NULL) {
 891                 gfp_t gfp_flags = GFP_USER;
 892                 unsigned i;
 893 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 894                 unsigned j;
 895 #endif
 896
 897                 /* set zero flag for page allocation if required */
 898                 if (flags & TTM_PAGE_FLAG_ZERO_ALLOC)
 899                         gfp_flags |= __GFP_ZERO;
 900
 901                 if (flags & TTM_PAGE_FLAG_NO_RETRY)
 902                         gfp_flags |= __GFP_RETRY_MAYFAIL;
 903
 904                 if (flags & TTM_PAGE_FLAG_DMA32)
 905                         gfp_flags |= GFP_DMA32;
 906                 else
 907                         gfp_flags |= GFP_HIGHUSER;
 908
 909                 i = 0;
 910 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 911                 if (!(gfp_flags & GFP_DMA32)) {
 912                         while (npages >= HPAGE_PMD_NR) {
 913                                 gfp_t huge_flags = gfp_flags;
 914
 915                                 huge_flags |= GFP_TRANSHUGE_LIGHT | __GFP_NORETRY |
 916                                         __GFP_KSWAPD_RECLAIM;
 917                                 huge_flags &= ~__GFP_MOVABLE;
 918                                 huge_flags &= ~__GFP_COMP;
 919                                 p = alloc_pages(huge_flags, HPAGE_PMD_ORDER);
 920                                 if (!p)
 921                                         break;
 922
 923                                 for (j = 0; j < HPAGE_PMD_NR; ++j)
 924                                         pages[i++] = p++;
 925
 926                                 npages -= HPAGE_PMD_NR;
 927                         }
 928                 }
 929 #endif
 930
 931                 first = i;
 932                 while (npages) {
 933                         p = alloc_page(gfp_flags);
 934                         if (!p) {
 935                                 pr_debug("Unable to allocate page\n");
 936                                 return -ENOMEM;
 937                         }
 938
 939                         /* Swap the pages if we detect consecutive order */
 940                         if (i > first && pages[i - 1] == p - 1)
 941                                 swap(p, pages[i - 1]);
 942
 943                         pages[i++] = p;
 944                         --npages;
 945                 }
 946                 return 0;
 947         }
 948
 949         count = 0;
 950
 951 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 952         if (huge && npages >= HPAGE_PMD_NR) {
 953                 INIT_LIST_HEAD(&plist);
 954                 ttm_page_pool_get_pages(huge, &plist, flags, cstate,
 955                                         npages / HPAGE_PMD_NR,
 956                                         HPAGE_PMD_ORDER);
 957
 958                 list_for_each_entry(p, &plist, lru) {
 959                         unsigned j;
 960
 961                         for (j = 0; j < HPAGE_PMD_NR; ++j)
 962                                 pages[count++] = &p[j];
 963                 }
 964         }
 965 #endif
 966
 967         INIT_LIST_HEAD(&plist);
 968         r = ttm_page_pool_get_pages(pool, &plist, flags, cstate,
 969                                     npages - count, 0);
 970
 971         first = count;
 972         list_for_each_entry(p, &plist, lru) {
 973                 struct page *tmp = p;
 974
 975                 /* Swap the pages if we detect consecutive order */
 976                 if (count > first && pages[count - 1] == tmp - 1)
 977                         swap(tmp, pages[count - 1]);
 978                 pages[count++] = tmp;
 979         }
 980
 981         if (r) {
 982                 /* If there is any pages in the list put them back to
 983                  * the pool.
 984                  */
 985                 pr_debug("Failed to allocate extra pages for large request\n");
 986                 ttm_put_pages(pages, count, flags, cstate);
 987                 return r;
 988         }
 989
 990         return 0;
 991 }
 992
 993 static void ttm_page_pool_init_locked(struct ttm_page_pool *pool, gfp_t flags,
 994                 char *name, unsigned int order)
 995 {
 996         spin_lock_init(&pool->lock);
 997         pool->fill_lock = false;
 998         INIT_LIST_HEAD(&pool->list);
 999         pool->npages = pool->nfrees = 0;
1000         pool->gfp_flags = flags;
1001         pool->name = name;
1002         pool->order = order;
1003 }
1004
1005 int ttm_page_alloc_init(struct ttm_mem_global *glob, unsigned max_pages)
1006 {
1007         int ret;
1008 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1009         unsigned order = HPAGE_PMD_ORDER;
1010 #else
1011         unsigned order = 0;
1012 #endif
1013
1014         WARN_ON(_manager);
1015
1016         pr_info("Initializing pool allocator\n");
1017
1018         _manager = kzalloc(sizeof(*_manager), GFP_KERNEL);
1019         if (!_manager)
1020                 return -ENOMEM;
1021
1022         ttm_page_pool_init_locked(&_manager->wc_pool, GFP_HIGHUSER, "wc", 0);
1023
1024         ttm_page_pool_init_locked(&_manager->uc_pool, GFP_HIGHUSER, "uc", 0);
1025
1026         ttm_page_pool_init_locked(&_manager->wc_pool_dma32,
1027                                   GFP_USER | GFP_DMA32, "wc dma", 0);
1028
1029         ttm_page_pool_init_locked(&_manager->uc_pool_dma32,
1030                                   GFP_USER | GFP_DMA32, "uc dma", 0);
1031
1032         ttm_page_pool_init_locked(&_manager->wc_pool_huge,
1033                                   (GFP_TRANSHUGE_LIGHT | __GFP_NORETRY |
1034                                    __GFP_KSWAPD_RECLAIM) &
1035                                   ~(__GFP_MOVABLE | __GFP_COMP),
1036                                   "wc huge", order);
1037
1038         ttm_page_pool_init_locked(&_manager->uc_pool_huge,
1039                                   (GFP_TRANSHUGE_LIGHT | __GFP_NORETRY |
1040                                    __GFP_KSWAPD_RECLAIM) &
1041                                   ~(__GFP_MOVABLE | __GFP_COMP)
1042                                   , "uc huge", order);
1043
1044         _manager->options.max_size = max_pages;
1045         _manager->options.small = SMALL_ALLOCATION;
1046         _manager->options.alloc_size = NUM_PAGES_TO_ALLOC;
1047
1048         ret = kobject_init_and_add(&_manager->kobj, &ttm_pool_kobj_type,
1049                                    &glob->kobj, "pool");
1050         if (unlikely(ret != 0))
1051                 goto error;
1052
1053         ret = ttm_pool_mm_shrink_init(_manager);
1054         if (unlikely(ret != 0))
1055                 goto error;
1056         return 0;
1057
1058 error:
1059         kobject_put(&_manager->kobj);
1060         _manager = NULL;
1061         return ret;
1062 }
1063
1064 void ttm_page_alloc_fini(void)
1065 {
1066         int i;
1067
1068         pr_info("Finalizing pool allocator\n");
1069         ttm_pool_mm_shrink_fini(_manager);
1070
1071         /* OK to use static buffer since global mutex is no longer used. */
1072         for (i = 0; i < NUM_POOLS; ++i)
1073                 ttm_page_pool_free(&_manager->pools[i], FREE_ALL_PAGES, true);
1074
1075         kobject_put(&_manager->kobj);
1076         _manager = NULL;
1077 }
1078
1079 static void
1080 ttm_pool_unpopulate_helper(struct ttm_tt *ttm, unsigned mem_count_update)
1081 {
1082         struct ttm_mem_global *mem_glob = ttm->bdev->glob->mem_glob;
1083         unsigned i;
1084
1085         if (mem_count_update == 0)
1086                 goto put_pages;
1087
1088         for (i = 0; i < mem_count_update; ++i) {
1089                 if (!ttm->pages[i])
1090                         continue;
1091
1092                 ttm_mem_global_free_page(mem_glob, ttm->pages[i], PAGE_SIZE);
1093         }
1094
1095 put_pages:
1096         ttm_put_pages(ttm->pages, ttm->num_pages, ttm->page_flags,
1097                       ttm->caching_state);
1098         ttm->state = tt_unpopulated;
1099 }
1100
1101 int ttm_pool_populate(struct ttm_tt *ttm, struct ttm_operation_ctx *ctx)
1102 {
1103         struct ttm_mem_global *mem_glob = ttm->bdev->glob->mem_glob;
1104         unsigned i;
1105         int ret;
1106
1107         if (ttm->state != tt_unpopulated)
1108                 return 0;
1109
1110         if (ttm_check_under_lowerlimit(mem_glob, ttm->num_pages, ctx))
1111                 return -ENOMEM;
1112
1113         ret = ttm_get_pages(ttm->pages, ttm->num_pages, ttm->page_flags,
1114                             ttm->caching_state);
1115         if (unlikely(ret != 0)) {
1116                 ttm_pool_unpopulate_helper(ttm, 0);
1117                 return ret;
1118         }
1119
1120         for (i = 0; i < ttm->num_pages; ++i) {
1121                 ret = ttm_mem_global_alloc_page(mem_glob, ttm->pages[i],
1122                                                 PAGE_SIZE, ctx);
1123                 if (unlikely(ret != 0)) {
1124                         ttm_pool_unpopulate_helper(ttm, i);
1125                         return -ENOMEM;
1126                 }
1127         }
1128
1129         if (unlikely(ttm->page_flags & TTM_PAGE_FLAG_SWAPPED)) {
1130                 ret = ttm_tt_swapin(ttm);
1131                 if (unlikely(ret != 0)) {
1132                         ttm_pool_unpopulate(ttm);
1133                         return ret;
1134                 }
1135         }
1136
1137         ttm->state = tt_unbound;
1138         return 0;
1139 }
1140 EXPORT_SYMBOL(ttm_pool_populate);
1141
1142 void ttm_pool_unpopulate(struct ttm_tt *ttm)
1143 {
1144         ttm_pool_unpopulate_helper(ttm, ttm->num_pages);
1145 }
1146 EXPORT_SYMBOL(ttm_pool_unpopulate);
1147
1148 int ttm_populate_and_map_pages(struct device *dev, struct ttm_dma_tt *tt,
1149                                         struct ttm_operation_ctx *ctx)
1150 {
1151         unsigned i, j;
1152         int r;
1153
1154         r = ttm_pool_populate(&tt->ttm, ctx);
1155         if (r)
1156                 return r;
1157
1158         for (i = 0; i < tt->ttm.num_pages; ++i) {
1159                 struct page *p = tt->ttm.pages[i];
1160                 size_t num_pages = 1;
1161
1162                 for (j = i + 1; j < tt->ttm.num_pages; ++j) {
1163                         if (++p != tt->ttm.pages[j])
1164                                 break;
1165
1166                         ++num_pages;
1167                 }
1168
1169                 tt->dma_address[i] = dma_map_page(dev, tt->ttm.pages[i],
1170                                                   0, num_pages * PAGE_SIZE,
1171                                                   DMA_BIDIRECTIONAL);
1172                 if (dma_mapping_error(dev, tt->dma_address[i])) {
1173                         while (i--) {
1174                                 dma_unmap_page(dev, tt->dma_address[i],
1175                                                PAGE_SIZE, DMA_BIDIRECTIONAL);
1176                                 tt->dma_address[i] = 0;
1177                         }
1178                         ttm_pool_unpopulate(&tt->ttm);
1179                         return -EFAULT;
1180                 }
1181
1182                 for (j = 1; j < num_pages; ++j) {
1183                         tt->dma_address[i + 1] = tt->dma_address[i] + PAGE_SIZE;
1184                         ++i;
1185                 }
1186         }
1187         return 0;
1188 }
1189 EXPORT_SYMBOL(ttm_populate_and_map_pages);
1190
1191 void ttm_unmap_and_unpopulate_pages(struct device *dev, struct ttm_dma_tt *tt)
1192 {
1193         unsigned i, j;
1194
1195         for (i = 0; i < tt->ttm.num_pages;) {
1196                 struct page *p = tt->ttm.pages[i];
1197                 size_t num_pages = 1;
1198
1199                 if (!tt->dma_address[i] || !tt->ttm.pages[i]) {
1200                         ++i;
1201                         continue;
1202                 }
1203
1204                 for (j = i + 1; j < tt->ttm.num_pages; ++j) {
1205                         if (++p != tt->ttm.pages[j])
1206                                 break;
1207
1208                         ++num_pages;
1209                 }
1210
1211                 dma_unmap_page(dev, tt->dma_address[i], num_pages * PAGE_SIZE,
1212                                DMA_BIDIRECTIONAL);
1213
1214                 i += num_pages;
1215         }
1216         ttm_pool_unpopulate(&tt->ttm);
1217 }
1218 EXPORT_SYMBOL(ttm_unmap_and_unpopulate_pages);
1219
1220 int ttm_page_alloc_debugfs(struct seq_file *m, void *data)
1221 {
1222         struct ttm_page_pool *p;
1223         unsigned i;
1224         char *h[] = {"pool", "refills", "pages freed", "size"};
1225         if (!_manager) {
1226                 seq_printf(m, "No pool allocator running.\n");
1227                 return 0;
1228         }
1229         seq_printf(m, "%7s %12s %13s %8s\n",
1230                         h[0], h[1], h[2], h[3]);
1231         for (i = 0; i < NUM_POOLS; ++i) {
1232                 p = &_manager->pools[i];
1233
1234                 seq_printf(m, "%7s %12ld %13ld %8d\n",
1235                                 p->name, p->nrefills,
1236                                 p->nfrees, p->npages);
1237         }
1238         return 0;
1239 }
1240 EXPORT_SYMBOL(ttm_page_alloc_debugfs);