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         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(npages_to_free * sizeof(struct page *),
 352                                         GFP_KERNEL);
 353         if (!pages_to_free) {
 354                 pr_debug("Failed to allocate memory for pool free operation\n");
 355                 return 0;
 356         }
 357
 358 restart:
 359         spin_lock_irqsave(&pool->lock, irq_flags);
 360
 361         list_for_each_entry_reverse(p, &pool->list, lru) {
 362                 if (freed_pages >= npages_to_free)
 363                         break;
 364
 365                 pages_to_free[freed_pages++] = p;
 366                 /* We can only remove NUM_PAGES_TO_ALLOC at a time. */
 367                 if (freed_pages >= NUM_PAGES_TO_ALLOC) {
 368                         /* remove range of pages from the pool */
 369                         __list_del(p->lru.prev, &pool->list);
 370
 371                         ttm_pool_update_free_locked(pool, freed_pages);
 372                         /**
 373                          * Because changing page caching is costly
 374                          * we unlock the pool to prevent stalling.
 375                          */
 376                         spin_unlock_irqrestore(&pool->lock, irq_flags);
 377
 378                         ttm_pages_put(pages_to_free, freed_pages, pool->order);
 379                         if (likely(nr_free != FREE_ALL_PAGES))
 380                                 nr_free -= freed_pages;
 381
 382                         if (NUM_PAGES_TO_ALLOC >= nr_free)
 383                                 npages_to_free = nr_free;
 384                         else
 385                                 npages_to_free = NUM_PAGES_TO_ALLOC;
 386
 387                         freed_pages = 0;
 388
 389                         /* free all so restart the processing */
 390                         if (nr_free)
 391                                 goto restart;
 392
 393                         /* Not allowed to fall through or break because
 394                          * following context is inside spinlock while we are
 395                          * outside here.
 396                          */
 397                         goto out;
 398
 399                 }
 400         }
 401
 402         /* remove range of pages from the pool */
 403         if (freed_pages) {
 404                 __list_del(&p->lru, &pool->list);
 405
 406                 ttm_pool_update_free_locked(pool, freed_pages);
 407                 nr_free -= freed_pages;
 408         }
 409
 410         spin_unlock_irqrestore(&pool->lock, irq_flags);
 411
 412         if (freed_pages)
 413                 ttm_pages_put(pages_to_free, freed_pages, pool->order);
 414 out:
 415         if (pages_to_free != static_buf)
 416                 kfree(pages_to_free);
 417         return nr_free;
 418 }
 419
 420 /**
 421  * Callback for mm to request pool to reduce number of page held.
 422  *
 423  * XXX: (dchinner) Deadlock warning!
 424  *
 425  * This code is crying out for a shrinker per pool....
 426  */
 427 static unsigned long
 428 ttm_pool_shrink_scan(struct shrinker *shrink, struct shrink_control *sc)
 429 {
 430         static DEFINE_MUTEX(lock);
 431         static unsigned start_pool;
 432         unsigned i;
 433         unsigned pool_offset;
 434         struct ttm_page_pool *pool;
 435         int shrink_pages = sc->nr_to_scan;
 436         unsigned long freed = 0;
 437         unsigned int nr_free_pool;
 438
 439         if (!mutex_trylock(&lock))
 440                 return SHRINK_STOP;
 441         pool_offset = ++start_pool % NUM_POOLS;
 442         /* select start pool in round robin fashion */
 443         for (i = 0; i < NUM_POOLS; ++i) {
 444                 unsigned nr_free = shrink_pages;
 445                 unsigned page_nr;
 446
 447                 if (shrink_pages == 0)
 448                         break;
 449
 450                 pool = &_manager->pools[(i + pool_offset)%NUM_POOLS];
 451                 page_nr = (1 << pool->order);
 452                 /* OK to use static buffer since global mutex is held. */
 453                 nr_free_pool = roundup(nr_free, page_nr) >> pool->order;
 454                 shrink_pages = ttm_page_pool_free(pool, nr_free_pool, true);
 455                 freed += (nr_free_pool - shrink_pages) << pool->order;
 456                 if (freed >= sc->nr_to_scan)
 457                         break;
 458                 shrink_pages <<= pool->order;
 459         }
 460         mutex_unlock(&lock);
 461         return freed;
 462 }
 463
 464
 465 static unsigned long
 466 ttm_pool_shrink_count(struct shrinker *shrink, struct shrink_control *sc)
 467 {
 468         unsigned i;
 469         unsigned long count = 0;
 470         struct ttm_page_pool *pool;
 471
 472         for (i = 0; i < NUM_POOLS; ++i) {
 473                 pool = &_manager->pools[i];
 474                 count += (pool->npages << pool->order);
 475         }
 476
 477         return count;
 478 }
 479
 480 static int ttm_pool_mm_shrink_init(struct ttm_pool_manager *manager)
 481 {
 482         manager->mm_shrink.count_objects = ttm_pool_shrink_count;
 483         manager->mm_shrink.scan_objects = ttm_pool_shrink_scan;
 484         manager->mm_shrink.seeks = 1;
 485         return register_shrinker(&manager->mm_shrink);
 486 }
 487
 488 static void ttm_pool_mm_shrink_fini(struct ttm_pool_manager *manager)
 489 {
 490         unregister_shrinker(&manager->mm_shrink);
 491 }
 492
 493 static int ttm_set_pages_caching(struct page **pages,
 494                 enum ttm_caching_state cstate, unsigned cpages)
 495 {
 496         int r = 0;
 497         /* Set page caching */
 498         switch (cstate) {
 499         case tt_uncached:
 500                 r = set_pages_array_uc(pages, cpages);
 501                 if (r)
 502                         pr_err("Failed to set %d pages to uc!\n", cpages);
 503                 break;
 504         case tt_wc:
 505                 r = set_pages_array_wc(pages, cpages);
 506                 if (r)
 507                         pr_err("Failed to set %d pages to wc!\n", cpages);
 508                 break;
 509         default:
 510                 break;
 511         }
 512         return r;
 513 }
 514
 515 /**
 516  * Free pages the pages that failed to change the caching state. If there is
 517  * any pages that have changed their caching state already put them to the
 518  * pool.
 519  */
 520 static void ttm_handle_caching_state_failure(struct list_head *pages,
 521                 int ttm_flags, enum ttm_caching_state cstate,
 522                 struct page **failed_pages, unsigned cpages)
 523 {
 524         unsigned i;
 525         /* Failed pages have to be freed */
 526         for (i = 0; i < cpages; ++i) {
 527                 list_del(&failed_pages[i]->lru);
 528                 __free_page(failed_pages[i]);
 529         }
 530 }
 531
 532 /**
 533  * Allocate new pages with correct caching.
 534  *
 535  * This function is reentrant if caller updates count depending on number of
 536  * pages returned in pages array.
 537  */
 538 static int ttm_alloc_new_pages(struct list_head *pages, gfp_t gfp_flags,
 539                                int ttm_flags, enum ttm_caching_state cstate,
 540                                unsigned count, unsigned order)
 541 {
 542         struct page **caching_array;
 543         struct page *p;
 544         int r = 0;
 545         unsigned i, j, cpages;
 546         unsigned npages = 1 << order;
 547         unsigned max_cpages = min(count << order, (unsigned)NUM_PAGES_TO_ALLOC);
 548
 549         /* allocate array for page caching change */
 550         caching_array = kmalloc(max_cpages*sizeof(struct page *), GFP_KERNEL);
 551
 552         if (!caching_array) {
 553                 pr_debug("Unable to allocate table for new pages\n");
 554                 return -ENOMEM;
 555         }
 556
 557         for (i = 0, cpages = 0; i < count; ++i) {
 558                 p = alloc_pages(gfp_flags, order);
 559
 560                 if (!p) {
 561                         pr_debug("Unable to get page %u\n", i);
 562
 563                         /* store already allocated pages in the pool after
 564                          * setting the caching state */
 565                         if (cpages) {
 566                                 r = ttm_set_pages_caching(caching_array,
 567                                                           cstate, cpages);
 568                                 if (r)
 569                                         ttm_handle_caching_state_failure(pages,
 570                                                 ttm_flags, cstate,
 571                                                 caching_array, cpages);
 572                         }
 573                         r = -ENOMEM;
 574                         goto out;
 575                 }
 576
 577                 list_add(&p->lru, pages);
 578
 579 #ifdef CONFIG_HIGHMEM
 580                 /* gfp flags of highmem page should never be dma32 so we
 581                  * we should be fine in such case
 582                  */
 583                 if (PageHighMem(p))
 584                         continue;
 585
 586 #endif
 587                 for (j = 0; j < npages; ++j) {
 588                         caching_array[cpages++] = p++;
 589                         if (cpages == max_cpages) {
 590
 591                                 r = ttm_set_pages_caching(caching_array,
 592                                                 cstate, cpages);
 593                                 if (r) {
 594                                         ttm_handle_caching_state_failure(pages,
 595                                                 ttm_flags, cstate,
 596                                                 caching_array, cpages);
 597                                         goto out;
 598                                 }
 599                                 cpages = 0;
 600                         }
 601                 }
 602         }
 603
 604         if (cpages) {
 605                 r = ttm_set_pages_caching(caching_array, cstate, cpages);
 606                 if (r)
 607                         ttm_handle_caching_state_failure(pages,
 608                                         ttm_flags, cstate,
 609                                         caching_array, cpages);
 610         }
 611 out:
 612         kfree(caching_array);
 613
 614         return r;
 615 }
 616
 617 /**
 618  * Fill the given pool if there aren't enough pages and the requested number of
 619  * pages is small.
 620  */
 621 static void ttm_page_pool_fill_locked(struct ttm_page_pool *pool, int ttm_flags,
 622                                       enum ttm_caching_state cstate,
 623                                       unsigned count, unsigned long *irq_flags)
 624 {
 625         struct page *p;
 626         int r;
 627         unsigned cpages = 0;
 628         /**
 629          * Only allow one pool fill operation at a time.
 630          * If pool doesn't have enough pages for the allocation new pages are
 631          * allocated from outside of pool.
 632          */
 633         if (pool->fill_lock)
 634                 return;
 635
 636         pool->fill_lock = true;
 637
 638         /* If allocation request is small and there are not enough
 639          * pages in a pool we fill the pool up first. */
 640         if (count < _manager->options.small
 641                 && count > pool->npages) {
 642                 struct list_head new_pages;
 643                 unsigned alloc_size = _manager->options.alloc_size;
 644
 645                 /**
 646                  * Can't change page caching if in irqsave context. We have to
 647                  * drop the pool->lock.
 648                  */
 649                 spin_unlock_irqrestore(&pool->lock, *irq_flags);
 650
 651                 INIT_LIST_HEAD(&new_pages);
 652                 r = ttm_alloc_new_pages(&new_pages, pool->gfp_flags, ttm_flags,
 653                                         cstate, alloc_size, 0);
 654                 spin_lock_irqsave(&pool->lock, *irq_flags);
 655
 656                 if (!r) {
 657                         list_splice(&new_pages, &pool->list);
 658                         ++pool->nrefills;
 659                         pool->npages += alloc_size;
 660                 } else {
 661                         pr_debug("Failed to fill pool (%p)\n", pool);
 662                         /* If we have any pages left put them to the pool. */
 663                         list_for_each_entry(p, &new_pages, lru) {
 664                                 ++cpages;
 665                         }
 666                         list_splice(&new_pages, &pool->list);
 667                         pool->npages += cpages;
 668                 }
 669
 670         }
 671         pool->fill_lock = false;
 672 }
 673
 674 /**
 675  * Allocate pages from the pool and put them on the return list.
 676  *
 677  * @return zero for success or negative error code.
 678  */
 679 static int ttm_page_pool_get_pages(struct ttm_page_pool *pool,
 680                                    struct list_head *pages,
 681                                    int ttm_flags,
 682                                    enum ttm_caching_state cstate,
 683                                    unsigned count, unsigned order)
 684 {
 685         unsigned long irq_flags;
 686         struct list_head *p;
 687         unsigned i;
 688         int r = 0;
 689
 690         spin_lock_irqsave(&pool->lock, irq_flags);
 691         if (!order)
 692                 ttm_page_pool_fill_locked(pool, ttm_flags, cstate, count,
 693                                           &irq_flags);
 694
 695         if (count >= pool->npages) {
 696                 /* take all pages from the pool */
 697                 list_splice_init(&pool->list, pages);
 698                 count -= pool->npages;
 699                 pool->npages = 0;
 700                 goto out;
 701         }
 702         /* find the last pages to include for requested number of pages. Split
 703          * pool to begin and halve it to reduce search space. */
 704         if (count <= pool->npages/2) {
 705                 i = 0;
 706                 list_for_each(p, &pool->list) {
 707                         if (++i == count)
 708                                 break;
 709                 }
 710         } else {
 711                 i = pool->npages + 1;
 712                 list_for_each_prev(p, &pool->list) {
 713                         if (--i == count)
 714                                 break;
 715                 }
 716         }
 717         /* Cut 'count' number of pages from the pool */
 718         list_cut_position(pages, &pool->list, p);
 719         pool->npages -= count;
 720         count = 0;
 721 out:
 722         spin_unlock_irqrestore(&pool->lock, irq_flags);
 723
 724         /* clear the pages coming from the pool if requested */
 725         if (ttm_flags & TTM_PAGE_FLAG_ZERO_ALLOC) {
 726                 struct page *page;
 727
 728                 list_for_each_entry(page, pages, lru) {
 729                         if (PageHighMem(page))
 730                                 clear_highpage(page);
 731                         else
 732                                 clear_page(page_address(page));
 733                 }
 734         }
 735
 736         /* If pool didn't have enough pages allocate new one. */
 737         if (count) {
 738                 gfp_t gfp_flags = pool->gfp_flags;
 739
 740                 /* set zero flag for page allocation if required */
 741                 if (ttm_flags & TTM_PAGE_FLAG_ZERO_ALLOC)
 742                         gfp_flags |= __GFP_ZERO;
 743
 744                 if (ttm_flags & TTM_PAGE_FLAG_NO_RETRY)
 745                         gfp_flags |= __GFP_RETRY_MAYFAIL;
 746
 747                 /* ttm_alloc_new_pages doesn't reference pool so we can run
 748                  * multiple requests in parallel.
 749                  **/
 750                 r = ttm_alloc_new_pages(pages, gfp_flags, ttm_flags, cstate,
 751                                         count, order);
 752         }
 753
 754         return r;
 755 }
 756
 757 /* Put all pages in pages list to correct pool to wait for reuse */
 758 static void ttm_put_pages(struct page **pages, unsigned npages, int flags,
 759                           enum ttm_caching_state cstate)
 760 {
 761         struct ttm_page_pool *pool = ttm_get_pool(flags, false, cstate);
 762 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 763         struct ttm_page_pool *huge = ttm_get_pool(flags, true, cstate);
 764 #endif
 765         unsigned long irq_flags;
 766         unsigned i;
 767
 768         if (pool == NULL) {
 769                 /* No pool for this memory type so free the pages */
 770                 i = 0;
 771                 while (i < npages) {
 772 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 773                         struct page *p = pages[i];
 774 #endif
 775                         unsigned order = 0, j;
 776
 777                         if (!pages[i]) {
 778                                 ++i;
 779                                 continue;
 780                         }
 781
 782 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 783                         if (!(flags & TTM_PAGE_FLAG_DMA32)) {
 784                                 for (j = 0; j < HPAGE_PMD_NR; ++j)
 785                                         if (p++ != pages[i + j])
 786                                             break;
 787
 788                                 if (j == HPAGE_PMD_NR)
 789                                         order = HPAGE_PMD_ORDER;
 790                         }
 791 #endif
 792
 793                         if (page_count(pages[i]) != 1)
 794                                 pr_err("Erroneous page count. Leaking pages.\n");
 795                         __free_pages(pages[i], order);
 796
 797                         j = 1 << order;
 798                         while (j) {
 799                                 pages[i++] = NULL;
 800                                 --j;
 801                         }
 802                 }
 803                 return;
 804         }
 805
 806         i = 0;
 807 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 808         if (huge) {
 809                 unsigned max_size, n2free;
 810
 811                 spin_lock_irqsave(&huge->lock, irq_flags);
 812                 while (i < npages) {
 813                         struct page *p = pages[i];
 814                         unsigned j;
 815
 816                         if (!p)
 817                                 break;
 818
 819                         for (j = 0; j < HPAGE_PMD_NR; ++j)
 820                                 if (p++ != pages[i + j])
 821                                     break;
 822
 823                         if (j != HPAGE_PMD_NR)
 824                                 break;
 825
 826                         list_add_tail(&pages[i]->lru, &huge->list);
 827
 828                         for (j = 0; j < HPAGE_PMD_NR; ++j)
 829                                 pages[i++] = NULL;
 830                         huge->npages++;
 831                 }
 832
 833                 /* Check that we don't go over the pool limit */
 834                 max_size = _manager->options.max_size;
 835                 max_size /= HPAGE_PMD_NR;
 836                 if (huge->npages > max_size)
 837                         n2free = huge->npages - max_size;
 838                 else
 839                         n2free = 0;
 840                 spin_unlock_irqrestore(&huge->lock, irq_flags);
 841                 if (n2free)
 842                         ttm_page_pool_free(huge, n2free, false);
 843         }
 844 #endif
 845
 846         spin_lock_irqsave(&pool->lock, irq_flags);
 847         while (i < npages) {
 848                 if (pages[i]) {
 849                         if (page_count(pages[i]) != 1)
 850                                 pr_err("Erroneous page count. Leaking pages.\n");
 851                         list_add_tail(&pages[i]->lru, &pool->list);
 852                         pages[i] = NULL;
 853                         pool->npages++;
 854                 }
 855                 ++i;
 856         }
 857         /* Check that we don't go over the pool limit */
 858         npages = 0;
 859         if (pool->npages > _manager->options.max_size) {
 860                 npages = pool->npages - _manager->options.max_size;
 861                 /* free at least NUM_PAGES_TO_ALLOC number of pages
 862                  * to reduce calls to set_memory_wb */
 863                 if (npages < NUM_PAGES_TO_ALLOC)
 864                         npages = NUM_PAGES_TO_ALLOC;
 865         }
 866         spin_unlock_irqrestore(&pool->lock, irq_flags);
 867         if (npages)
 868                 ttm_page_pool_free(pool, npages, false);
 869 }
 870
 871 /*
 872  * On success pages list will hold count number of correctly
 873  * cached pages.
 874  */
 875 static int ttm_get_pages(struct page **pages, unsigned npages, int flags,
 876                          enum ttm_caching_state cstate)
 877 {
 878         struct ttm_page_pool *pool = ttm_get_pool(flags, false, cstate);
 879 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 880         struct ttm_page_pool *huge = ttm_get_pool(flags, true, cstate);
 881 #endif
 882         struct list_head plist;
 883         struct page *p = NULL;
 884         unsigned count, first;
 885         int r;
 886
 887         /* No pool for cached pages */
 888         if (pool == NULL) {
 889                 gfp_t gfp_flags = GFP_USER;
 890                 unsigned i;
 891 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 892                 unsigned j;
 893 #endif
 894
 895                 /* set zero flag for page allocation if required */
 896                 if (flags & TTM_PAGE_FLAG_ZERO_ALLOC)
 897                         gfp_flags |= __GFP_ZERO;
 898
 899                 if (flags & TTM_PAGE_FLAG_NO_RETRY)
 900                         gfp_flags |= __GFP_RETRY_MAYFAIL;
 901
 902                 if (flags & TTM_PAGE_FLAG_DMA32)
 903                         gfp_flags |= GFP_DMA32;
 904                 else
 905                         gfp_flags |= GFP_HIGHUSER;
 906
 907                 i = 0;
 908 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 909                 if (!(gfp_flags & GFP_DMA32)) {
 910                         while (npages >= HPAGE_PMD_NR) {
 911                                 gfp_t huge_flags = gfp_flags;
 912
 913                                 huge_flags |= GFP_TRANSHUGE_LIGHT | __GFP_NORETRY |
 914                                         __GFP_KSWAPD_RECLAIM;
 915                                 huge_flags &= ~__GFP_MOVABLE;
 916                                 huge_flags &= ~__GFP_COMP;
 917                                 p = alloc_pages(huge_flags, HPAGE_PMD_ORDER);
 918                                 if (!p)
 919                                         break;
 920
 921                                 for (j = 0; j < HPAGE_PMD_NR; ++j)
 922                                         pages[i++] = p++;
 923
 924                                 npages -= HPAGE_PMD_NR;
 925                         }
 926                 }
 927 #endif
 928
 929                 first = i;
 930                 while (npages) {
 931                         p = alloc_page(gfp_flags);
 932                         if (!p) {
 933                                 pr_debug("Unable to allocate page\n");
 934                                 return -ENOMEM;
 935                         }
 936
 937                         /* Swap the pages if we detect consecutive order */
 938                         if (i > first && pages[i - 1] == p - 1)
 939                                 swap(p, pages[i - 1]);
 940
 941                         pages[i++] = p;
 942                         --npages;
 943                 }
 944                 return 0;
 945         }
 946
 947         count = 0;
 948
 949 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 950         if (huge && npages >= HPAGE_PMD_NR) {
 951                 INIT_LIST_HEAD(&plist);
 952                 ttm_page_pool_get_pages(huge, &plist, flags, cstate,
 953                                         npages / HPAGE_PMD_NR,
 954                                         HPAGE_PMD_ORDER);
 955
 956                 list_for_each_entry(p, &plist, lru) {
 957                         unsigned j;
 958
 959                         for (j = 0; j < HPAGE_PMD_NR; ++j)
 960                                 pages[count++] = &p[j];
 961                 }
 962         }
 963 #endif
 964
 965         INIT_LIST_HEAD(&plist);
 966         r = ttm_page_pool_get_pages(pool, &plist, flags, cstate,
 967                                     npages - count, 0);
 968
 969         first = count;
 970         list_for_each_entry(p, &plist, lru) {
 971                 struct page *tmp = p;
 972
 973                 /* Swap the pages if we detect consecutive order */
 974                 if (count > first && pages[count - 1] == tmp - 1)
 975                         swap(tmp, pages[count - 1]);
 976                 pages[count++] = tmp;
 977         }
 978
 979         if (r) {
 980                 /* If there is any pages in the list put them back to
 981                  * the pool.
 982                  */
 983                 pr_debug("Failed to allocate extra pages for large request\n");
 984                 ttm_put_pages(pages, count, flags, cstate);
 985                 return r;
 986         }
 987
 988         return 0;
 989 }
 990
 991 static void ttm_page_pool_init_locked(struct ttm_page_pool *pool, gfp_t flags,
 992                 char *name, unsigned int order)
 993 {
 994         spin_lock_init(&pool->lock);
 995         pool->fill_lock = false;
 996         INIT_LIST_HEAD(&pool->list);
 997         pool->npages = pool->nfrees = 0;
 998         pool->gfp_flags = flags;
 999         pool->name = name;
1000         pool->order = order;
1001 }
1002
1003 int ttm_page_alloc_init(struct ttm_mem_global *glob, unsigned max_pages)
1004 {
1005         int ret;
1006 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1007         unsigned order = HPAGE_PMD_ORDER;
1008 #else
1009         unsigned order = 0;
1010 #endif
1011
1012         WARN_ON(_manager);
1013
1014         pr_info("Initializing pool allocator\n");
1015
1016         _manager = kzalloc(sizeof(*_manager), GFP_KERNEL);
1017         if (!_manager)
1018                 return -ENOMEM;
1019
1020         ttm_page_pool_init_locked(&_manager->wc_pool, GFP_HIGHUSER, "wc", 0);
1021
1022         ttm_page_pool_init_locked(&_manager->uc_pool, GFP_HIGHUSER, "uc", 0);
1023
1024         ttm_page_pool_init_locked(&_manager->wc_pool_dma32,
1025                                   GFP_USER | GFP_DMA32, "wc dma", 0);
1026
1027         ttm_page_pool_init_locked(&_manager->uc_pool_dma32,
1028                                   GFP_USER | GFP_DMA32, "uc dma", 0);
1029
1030         ttm_page_pool_init_locked(&_manager->wc_pool_huge,
1031                                   (GFP_TRANSHUGE_LIGHT | __GFP_NORETRY |
1032                                    __GFP_KSWAPD_RECLAIM) &
1033                                   ~(__GFP_MOVABLE | __GFP_COMP),
1034                                   "wc huge", order);
1035
1036         ttm_page_pool_init_locked(&_manager->uc_pool_huge,
1037                                   (GFP_TRANSHUGE_LIGHT | __GFP_NORETRY |
1038                                    __GFP_KSWAPD_RECLAIM) &
1039                                   ~(__GFP_MOVABLE | __GFP_COMP)
1040                                   , "uc huge", order);
1041
1042         _manager->options.max_size = max_pages;
1043         _manager->options.small = SMALL_ALLOCATION;
1044         _manager->options.alloc_size = NUM_PAGES_TO_ALLOC;
1045
1046         ret = kobject_init_and_add(&_manager->kobj, &ttm_pool_kobj_type,
1047                                    &glob->kobj, "pool");
1048         if (unlikely(ret != 0))
1049                 goto error;
1050
1051         ret = ttm_pool_mm_shrink_init(_manager);
1052         if (unlikely(ret != 0))
1053                 goto error;
1054         return 0;
1055
1056 error:
1057         kobject_put(&_manager->kobj);
1058         _manager = NULL;
1059         return ret;
1060 }
1061
1062 void ttm_page_alloc_fini(void)
1063 {
1064         int i;
1065
1066         pr_info("Finalizing pool allocator\n");
1067         ttm_pool_mm_shrink_fini(_manager);
1068
1069         /* OK to use static buffer since global mutex is no longer used. */
1070         for (i = 0; i < NUM_POOLS; ++i)
1071                 ttm_page_pool_free(&_manager->pools[i], FREE_ALL_PAGES, true);
1072
1073         kobject_put(&_manager->kobj);
1074         _manager = NULL;
1075 }
1076
1077 static void
1078 ttm_pool_unpopulate_helper(struct ttm_tt *ttm, unsigned mem_count_update)
1079 {
1080         struct ttm_mem_global *mem_glob = ttm->bdev->glob->mem_glob;
1081         unsigned i;
1082
1083         if (mem_count_update == 0)
1084                 goto put_pages;
1085
1086         for (i = 0; i < mem_count_update; ++i) {
1087                 if (!ttm->pages[i])
1088                         continue;
1089
1090                 ttm_mem_global_free_page(mem_glob, ttm->pages[i], PAGE_SIZE);
1091         }
1092
1093 put_pages:
1094         ttm_put_pages(ttm->pages, ttm->num_pages, ttm->page_flags,
1095                       ttm->caching_state);
1096         ttm->state = tt_unpopulated;
1097 }
1098
1099 int ttm_pool_populate(struct ttm_tt *ttm, struct ttm_operation_ctx *ctx)
1100 {
1101         struct ttm_mem_global *mem_glob = ttm->bdev->glob->mem_glob;
1102         unsigned i;
1103         int ret;
1104
1105         if (ttm->state != tt_unpopulated)
1106                 return 0;
1107
1108         if (ttm_check_under_lowerlimit(mem_glob, ttm->num_pages, ctx))
1109                 return -ENOMEM;
1110
1111         ret = ttm_get_pages(ttm->pages, ttm->num_pages, ttm->page_flags,
1112                             ttm->caching_state);
1113         if (unlikely(ret != 0)) {
1114                 ttm_pool_unpopulate_helper(ttm, 0);
1115                 return ret;
1116         }
1117
1118         for (i = 0; i < ttm->num_pages; ++i) {
1119                 ret = ttm_mem_global_alloc_page(mem_glob, ttm->pages[i],
1120                                                 PAGE_SIZE, ctx);
1121                 if (unlikely(ret != 0)) {
1122                         ttm_pool_unpopulate_helper(ttm, i);
1123                         return -ENOMEM;
1124                 }
1125         }
1126
1127         if (unlikely(ttm->page_flags & TTM_PAGE_FLAG_SWAPPED)) {
1128                 ret = ttm_tt_swapin(ttm);
1129                 if (unlikely(ret != 0)) {
1130                         ttm_pool_unpopulate(ttm);
1131                         return ret;
1132                 }
1133         }
1134
1135         ttm->state = tt_unbound;
1136         return 0;
1137 }
1138 EXPORT_SYMBOL(ttm_pool_populate);
1139
1140 void ttm_pool_unpopulate(struct ttm_tt *ttm)
1141 {
1142         ttm_pool_unpopulate_helper(ttm, ttm->num_pages);
1143 }
1144 EXPORT_SYMBOL(ttm_pool_unpopulate);
1145
1146 int ttm_populate_and_map_pages(struct device *dev, struct ttm_dma_tt *tt,
1147                                         struct ttm_operation_ctx *ctx)
1148 {
1149         unsigned i, j;
1150         int r;
1151
1152         r = ttm_pool_populate(&tt->ttm, ctx);
1153         if (r)
1154                 return r;
1155
1156         for (i = 0; i < tt->ttm.num_pages; ++i) {
1157                 struct page *p = tt->ttm.pages[i];
1158                 size_t num_pages = 1;
1159
1160                 for (j = i + 1; j < tt->ttm.num_pages; ++j) {
1161                         if (++p != tt->ttm.pages[j])
1162                                 break;
1163
1164                         ++num_pages;
1165                 }
1166
1167                 tt->dma_address[i] = dma_map_page(dev, tt->ttm.pages[i],
1168                                                   0, num_pages * PAGE_SIZE,
1169                                                   DMA_BIDIRECTIONAL);
1170                 if (dma_mapping_error(dev, tt->dma_address[i])) {
1171                         while (i--) {
1172                                 dma_unmap_page(dev, tt->dma_address[i],
1173                                                PAGE_SIZE, DMA_BIDIRECTIONAL);
1174                                 tt->dma_address[i] = 0;
1175                         }
1176                         ttm_pool_unpopulate(&tt->ttm);
1177                         return -EFAULT;
1178                 }
1179
1180                 for (j = 1; j < num_pages; ++j) {
1181                         tt->dma_address[i + 1] = tt->dma_address[i] + PAGE_SIZE;
1182                         ++i;
1183                 }
1184         }
1185         return 0;
1186 }
1187 EXPORT_SYMBOL(ttm_populate_and_map_pages);
1188
1189 void ttm_unmap_and_unpopulate_pages(struct device *dev, struct ttm_dma_tt *tt)
1190 {
1191         unsigned i, j;
1192
1193         for (i = 0; i < tt->ttm.num_pages;) {
1194                 struct page *p = tt->ttm.pages[i];
1195                 size_t num_pages = 1;
1196
1197                 if (!tt->dma_address[i] || !tt->ttm.pages[i]) {
1198                         ++i;
1199                         continue;
1200                 }
1201
1202                 for (j = i + 1; j < tt->ttm.num_pages; ++j) {
1203                         if (++p != tt->ttm.pages[j])
1204                                 break;
1205
1206                         ++num_pages;
1207                 }
1208
1209                 dma_unmap_page(dev, tt->dma_address[i], num_pages * PAGE_SIZE,
1210                                DMA_BIDIRECTIONAL);
1211
1212                 i += num_pages;
1213         }
1214         ttm_pool_unpopulate(&tt->ttm);
1215 }
1216 EXPORT_SYMBOL(ttm_unmap_and_unpopulate_pages);
1217
1218 int ttm_page_alloc_debugfs(struct seq_file *m, void *data)
1219 {
1220         struct ttm_page_pool *p;
1221         unsigned i;
1222         char *h[] = {"pool", "refills", "pages freed", "size"};
1223         if (!_manager) {
1224                 seq_printf(m, "No pool allocator running.\n");
1225                 return 0;
1226         }
1227         seq_printf(m, "%7s %12s %13s %8s\n",
1228                         h[0], h[1], h[2], h[3]);
1229         for (i = 0; i < NUM_POOLS; ++i) {
1230                 p = &_manager->pools[i];
1231
1232                 seq_printf(m, "%7s %12ld %13ld %8d\n",
1233                                 p->name, p->nrefills,
1234                                 p->nfrees, p->npages);
1235         }
1236         return 0;
1237 }
1238 EXPORT_SYMBOL(ttm_page_alloc_debugfs);