2 * Copyright (c) Red Hat Inc.
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:
11 * The above copyright notice and this permission notice (including the
12 * next paragraph) shall be included in all copies or substantial portions
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.
23 * Authors: Dave Airlie <airlied@redhat.com>
24 * Jerome Glisse <jglisse@redhat.com>
25 * Pauli Nieminen <suokkos@gmail.com>
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
34 #define pr_fmt(fmt) "[TTM] " fmt
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>
42 #include <linux/seq_file.h> /* for seq_printf */
43 #include <linux/slab.h>
44 #include <linux/dma-mapping.h>
46 #include <linux/atomic.h>
48 #include <drm/ttm/ttm_bo_driver.h>
49 #include <drm/ttm/ttm_page_alloc.h>
55 #define NUM_PAGES_TO_ALLOC (PAGE_SIZE/sizeof(struct page *))
56 #define SMALL_ALLOCATION 16
57 #define FREE_ALL_PAGES (~0U)
58 /* times are in msecs */
59 #define PAGE_FREE_INTERVAL 1000
62 * struct ttm_page_pool - Pool to reuse recently allocated uc/wc pages.
64 * @lock: Protects the shared pool from concurrnet access. Must be used with
65 * irqsave/irqrestore variants because pool allocator maybe called from
67 * @fill_lock: Prevent concurrent calls to fill.
68 * @list: Pool of free uc/wc pages for fast reuse.
69 * @gfp_flags: Flags to pass for alloc_page.
70 * @npages: Number of pages in pool.
72 struct ttm_page_pool {
75 struct list_head list;
80 unsigned long nrefills;
84 * Limits for the pool. They are handled without locks because only place where
85 * they may change is in sysfs store. They won't have immediate effect anyway
86 * so forcing serialization to access them is pointless.
89 struct ttm_pool_opts {
98 * struct ttm_pool_manager - Holds memory pools for fst allocation
100 * Manager is read only object for pool code so it doesn't need locking.
102 * @free_interval: minimum number of jiffies between freeing pages from pool.
103 * @page_alloc_inited: reference counting for pool allocation.
104 * @work: Work that is used to shrink the pool. Work is only run when there is
105 * some pages to free.
106 * @small_allocation: Limit in number of pages what is small allocation.
108 * @pools: All pool objects in use.
110 struct ttm_pool_manager {
112 struct shrinker mm_shrink;
113 struct ttm_pool_opts options;
116 struct ttm_page_pool pools[NUM_POOLS];
118 struct ttm_page_pool wc_pool;
119 struct ttm_page_pool uc_pool;
120 struct ttm_page_pool wc_pool_dma32;
121 struct ttm_page_pool uc_pool_dma32;
126 static struct attribute ttm_page_pool_max = {
127 .name = "pool_max_size",
128 .mode = S_IRUGO | S_IWUSR
130 static struct attribute ttm_page_pool_small = {
131 .name = "pool_small_allocation",
132 .mode = S_IRUGO | S_IWUSR
134 static struct attribute ttm_page_pool_alloc_size = {
135 .name = "pool_allocation_size",
136 .mode = S_IRUGO | S_IWUSR
139 static struct attribute *ttm_pool_attrs[] = {
141 &ttm_page_pool_small,
142 &ttm_page_pool_alloc_size,
146 static void ttm_pool_kobj_release(struct kobject *kobj)
148 struct ttm_pool_manager *m =
149 container_of(kobj, struct ttm_pool_manager, kobj);
153 static ssize_t ttm_pool_store(struct kobject *kobj,
154 struct attribute *attr, const char *buffer, size_t size)
156 struct ttm_pool_manager *m =
157 container_of(kobj, struct ttm_pool_manager, kobj);
160 chars = sscanf(buffer, "%u", &val);
164 /* Convert kb to number of pages */
165 val = val / (PAGE_SIZE >> 10);
167 if (attr == &ttm_page_pool_max)
168 m->options.max_size = val;
169 else if (attr == &ttm_page_pool_small)
170 m->options.small = val;
171 else if (attr == &ttm_page_pool_alloc_size) {
172 if (val > NUM_PAGES_TO_ALLOC*8) {
173 pr_err("Setting allocation size to %lu is not allowed. Recommended size is %lu\n",
174 NUM_PAGES_TO_ALLOC*(PAGE_SIZE >> 7),
175 NUM_PAGES_TO_ALLOC*(PAGE_SIZE >> 10));
177 } else if (val > NUM_PAGES_TO_ALLOC) {
178 pr_warn("Setting allocation size to larger than %lu is not recommended\n",
179 NUM_PAGES_TO_ALLOC*(PAGE_SIZE >> 10));
181 m->options.alloc_size = val;
187 static ssize_t ttm_pool_show(struct kobject *kobj,
188 struct attribute *attr, char *buffer)
190 struct ttm_pool_manager *m =
191 container_of(kobj, struct ttm_pool_manager, kobj);
194 if (attr == &ttm_page_pool_max)
195 val = m->options.max_size;
196 else if (attr == &ttm_page_pool_small)
197 val = m->options.small;
198 else if (attr == &ttm_page_pool_alloc_size)
199 val = m->options.alloc_size;
201 val = val * (PAGE_SIZE >> 10);
203 return snprintf(buffer, PAGE_SIZE, "%u\n", val);
206 static const struct sysfs_ops ttm_pool_sysfs_ops = {
207 .show = &ttm_pool_show,
208 .store = &ttm_pool_store,
211 static struct kobj_type ttm_pool_kobj_type = {
212 .release = &ttm_pool_kobj_release,
213 .sysfs_ops = &ttm_pool_sysfs_ops,
214 .default_attrs = ttm_pool_attrs,
217 static struct ttm_pool_manager *_manager;
220 static int set_pages_array_wb(struct page **pages, int addrinarray)
225 for (i = 0; i < addrinarray; i++)
226 unmap_page_from_agp(pages[i]);
231 static int set_pages_array_wc(struct page **pages, int addrinarray)
236 for (i = 0; i < addrinarray; i++)
237 map_page_into_agp(pages[i]);
242 static int set_pages_array_uc(struct page **pages, int addrinarray)
247 for (i = 0; i < addrinarray; i++)
248 map_page_into_agp(pages[i]);
255 * Select the right pool or requested caching state and ttm flags. */
256 static struct ttm_page_pool *ttm_get_pool(int flags,
257 enum ttm_caching_state cstate)
261 if (cstate == tt_cached)
269 if (flags & TTM_PAGE_FLAG_DMA32)
272 return &_manager->pools[pool_index];
275 /* set memory back to wb and free the pages. */
276 static void ttm_pages_put(struct page *pages[], unsigned npages)
279 if (set_pages_array_wb(pages, npages))
280 pr_err("Failed to set %d pages to wb!\n", npages);
281 for (i = 0; i < npages; ++i)
282 __free_page(pages[i]);
285 static void ttm_pool_update_free_locked(struct ttm_page_pool *pool,
286 unsigned freed_pages)
288 pool->npages -= freed_pages;
289 pool->nfrees += freed_pages;
293 * Free pages from pool.
295 * To prevent hogging the ttm_swap process we only free NUM_PAGES_TO_ALLOC
296 * number of pages in one go.
298 * @pool: to free the pages from
299 * @free_all: If set to true will free all pages in pool
301 static int ttm_page_pool_free(struct ttm_page_pool *pool, unsigned nr_free)
303 unsigned long irq_flags;
305 struct page **pages_to_free;
306 unsigned freed_pages = 0,
307 npages_to_free = nr_free;
309 if (NUM_PAGES_TO_ALLOC < nr_free)
310 npages_to_free = NUM_PAGES_TO_ALLOC;
312 pages_to_free = kmalloc(npages_to_free * sizeof(struct page *),
314 if (!pages_to_free) {
315 pr_err("Failed to allocate memory for pool free operation\n");
320 spin_lock_irqsave(&pool->lock, irq_flags);
322 list_for_each_entry_reverse(p, &pool->list, lru) {
323 if (freed_pages >= npages_to_free)
326 pages_to_free[freed_pages++] = p;
327 /* We can only remove NUM_PAGES_TO_ALLOC at a time. */
328 if (freed_pages >= NUM_PAGES_TO_ALLOC) {
329 /* remove range of pages from the pool */
330 __list_del(p->lru.prev, &pool->list);
332 ttm_pool_update_free_locked(pool, freed_pages);
334 * Because changing page caching is costly
335 * we unlock the pool to prevent stalling.
337 spin_unlock_irqrestore(&pool->lock, irq_flags);
339 ttm_pages_put(pages_to_free, freed_pages);
340 if (likely(nr_free != FREE_ALL_PAGES))
341 nr_free -= freed_pages;
343 if (NUM_PAGES_TO_ALLOC >= nr_free)
344 npages_to_free = nr_free;
346 npages_to_free = NUM_PAGES_TO_ALLOC;
350 /* free all so restart the processing */
354 /* Not allowed to fall through or break because
355 * following context is inside spinlock while we are
363 /* remove range of pages from the pool */
365 __list_del(&p->lru, &pool->list);
367 ttm_pool_update_free_locked(pool, freed_pages);
368 nr_free -= freed_pages;
371 spin_unlock_irqrestore(&pool->lock, irq_flags);
374 ttm_pages_put(pages_to_free, freed_pages);
376 kfree(pages_to_free);
381 * Callback for mm to request pool to reduce number of page held.
383 * XXX: (dchinner) Deadlock warning!
385 * ttm_page_pool_free() does memory allocation using GFP_KERNEL. that means
386 * this can deadlock when called a sc->gfp_mask that is not equal to
389 * This code is crying out for a shrinker per pool....
392 ttm_pool_shrink_scan(struct shrinker *shrink, struct shrink_control *sc)
394 static atomic_t start_pool = ATOMIC_INIT(0);
396 unsigned pool_offset = atomic_add_return(1, &start_pool);
397 struct ttm_page_pool *pool;
398 int shrink_pages = sc->nr_to_scan;
399 unsigned long freed = 0;
401 pool_offset = pool_offset % NUM_POOLS;
402 /* select start pool in round robin fashion */
403 for (i = 0; i < NUM_POOLS; ++i) {
404 unsigned nr_free = shrink_pages;
405 if (shrink_pages == 0)
407 pool = &_manager->pools[(i + pool_offset)%NUM_POOLS];
408 shrink_pages = ttm_page_pool_free(pool, nr_free);
409 freed += nr_free - shrink_pages;
416 ttm_pool_shrink_count(struct shrinker *shrink, struct shrink_control *sc)
419 unsigned long count = 0;
421 for (i = 0; i < NUM_POOLS; ++i)
422 count += _manager->pools[i].npages;
427 static void ttm_pool_mm_shrink_init(struct ttm_pool_manager *manager)
429 manager->mm_shrink.count_objects = ttm_pool_shrink_count;
430 manager->mm_shrink.scan_objects = ttm_pool_shrink_scan;
431 manager->mm_shrink.seeks = 1;
432 register_shrinker(&manager->mm_shrink);
435 static void ttm_pool_mm_shrink_fini(struct ttm_pool_manager *manager)
437 unregister_shrinker(&manager->mm_shrink);
440 static int ttm_set_pages_caching(struct page **pages,
441 enum ttm_caching_state cstate, unsigned cpages)
444 /* Set page caching */
447 r = set_pages_array_uc(pages, cpages);
449 pr_err("Failed to set %d pages to uc!\n", cpages);
452 r = set_pages_array_wc(pages, cpages);
454 pr_err("Failed to set %d pages to wc!\n", cpages);
463 * Free pages the pages that failed to change the caching state. If there is
464 * any pages that have changed their caching state already put them to the
467 static void ttm_handle_caching_state_failure(struct list_head *pages,
468 int ttm_flags, enum ttm_caching_state cstate,
469 struct page **failed_pages, unsigned cpages)
472 /* Failed pages have to be freed */
473 for (i = 0; i < cpages; ++i) {
474 list_del(&failed_pages[i]->lru);
475 __free_page(failed_pages[i]);
480 * Allocate new pages with correct caching.
482 * This function is reentrant if caller updates count depending on number of
483 * pages returned in pages array.
485 static int ttm_alloc_new_pages(struct list_head *pages, gfp_t gfp_flags,
486 int ttm_flags, enum ttm_caching_state cstate, unsigned count)
488 struct page **caching_array;
492 unsigned max_cpages = min(count,
493 (unsigned)(PAGE_SIZE/sizeof(struct page *)));
495 /* allocate array for page caching change */
496 caching_array = kmalloc(max_cpages*sizeof(struct page *), GFP_KERNEL);
498 if (!caching_array) {
499 pr_err("Unable to allocate table for new pages\n");
503 for (i = 0, cpages = 0; i < count; ++i) {
504 p = alloc_page(gfp_flags);
507 pr_err("Unable to get page %u\n", i);
509 /* store already allocated pages in the pool after
510 * setting the caching state */
512 r = ttm_set_pages_caching(caching_array,
515 ttm_handle_caching_state_failure(pages,
517 caching_array, cpages);
523 #ifdef CONFIG_HIGHMEM
524 /* gfp flags of highmem page should never be dma32 so we
525 * we should be fine in such case
530 caching_array[cpages++] = p;
531 if (cpages == max_cpages) {
533 r = ttm_set_pages_caching(caching_array,
536 ttm_handle_caching_state_failure(pages,
538 caching_array, cpages);
545 list_add(&p->lru, pages);
549 r = ttm_set_pages_caching(caching_array, cstate, cpages);
551 ttm_handle_caching_state_failure(pages,
553 caching_array, cpages);
556 kfree(caching_array);
562 * Fill the given pool if there aren't enough pages and the requested number of
565 static void ttm_page_pool_fill_locked(struct ttm_page_pool *pool,
566 int ttm_flags, enum ttm_caching_state cstate, unsigned count,
567 unsigned long *irq_flags)
573 * Only allow one pool fill operation at a time.
574 * If pool doesn't have enough pages for the allocation new pages are
575 * allocated from outside of pool.
580 pool->fill_lock = true;
582 /* If allocation request is small and there are not enough
583 * pages in a pool we fill the pool up first. */
584 if (count < _manager->options.small
585 && count > pool->npages) {
586 struct list_head new_pages;
587 unsigned alloc_size = _manager->options.alloc_size;
590 * Can't change page caching if in irqsave context. We have to
591 * drop the pool->lock.
593 spin_unlock_irqrestore(&pool->lock, *irq_flags);
595 INIT_LIST_HEAD(&new_pages);
596 r = ttm_alloc_new_pages(&new_pages, pool->gfp_flags, ttm_flags,
598 spin_lock_irqsave(&pool->lock, *irq_flags);
601 list_splice(&new_pages, &pool->list);
603 pool->npages += alloc_size;
605 pr_err("Failed to fill pool (%p)\n", pool);
606 /* If we have any pages left put them to the pool. */
607 list_for_each_entry(p, &pool->list, lru) {
610 list_splice(&new_pages, &pool->list);
611 pool->npages += cpages;
615 pool->fill_lock = false;
619 * Cut 'count' number of pages from the pool and put them on the return list.
621 * @return count of pages still required to fulfill the request.
623 static unsigned ttm_page_pool_get_pages(struct ttm_page_pool *pool,
624 struct list_head *pages,
626 enum ttm_caching_state cstate,
629 unsigned long irq_flags;
633 spin_lock_irqsave(&pool->lock, irq_flags);
634 ttm_page_pool_fill_locked(pool, ttm_flags, cstate, count, &irq_flags);
636 if (count >= pool->npages) {
637 /* take all pages from the pool */
638 list_splice_init(&pool->list, pages);
639 count -= pool->npages;
643 /* find the last pages to include for requested number of pages. Split
644 * pool to begin and halve it to reduce search space. */
645 if (count <= pool->npages/2) {
647 list_for_each(p, &pool->list) {
652 i = pool->npages + 1;
653 list_for_each_prev(p, &pool->list) {
658 /* Cut 'count' number of pages from the pool */
659 list_cut_position(pages, &pool->list, p);
660 pool->npages -= count;
663 spin_unlock_irqrestore(&pool->lock, irq_flags);
667 /* Put all pages in pages list to correct pool to wait for reuse */
668 static void ttm_put_pages(struct page **pages, unsigned npages, int flags,
669 enum ttm_caching_state cstate)
671 unsigned long irq_flags;
672 struct ttm_page_pool *pool = ttm_get_pool(flags, cstate);
676 /* No pool for this memory type so free the pages */
677 for (i = 0; i < npages; i++) {
679 if (page_count(pages[i]) != 1)
680 pr_err("Erroneous page count. Leaking pages.\n");
681 __free_page(pages[i]);
688 spin_lock_irqsave(&pool->lock, irq_flags);
689 for (i = 0; i < npages; i++) {
691 if (page_count(pages[i]) != 1)
692 pr_err("Erroneous page count. Leaking pages.\n");
693 list_add_tail(&pages[i]->lru, &pool->list);
698 /* Check that we don't go over the pool limit */
700 if (pool->npages > _manager->options.max_size) {
701 npages = pool->npages - _manager->options.max_size;
702 /* free at least NUM_PAGES_TO_ALLOC number of pages
703 * to reduce calls to set_memory_wb */
704 if (npages < NUM_PAGES_TO_ALLOC)
705 npages = NUM_PAGES_TO_ALLOC;
707 spin_unlock_irqrestore(&pool->lock, irq_flags);
709 ttm_page_pool_free(pool, npages);
713 * On success pages list will hold count number of correctly
716 static int ttm_get_pages(struct page **pages, unsigned npages, int flags,
717 enum ttm_caching_state cstate)
719 struct ttm_page_pool *pool = ttm_get_pool(flags, cstate);
720 struct list_head plist;
721 struct page *p = NULL;
722 gfp_t gfp_flags = GFP_USER;
726 /* set zero flag for page allocation if required */
727 if (flags & TTM_PAGE_FLAG_ZERO_ALLOC)
728 gfp_flags |= __GFP_ZERO;
730 /* No pool for cached pages */
732 if (flags & TTM_PAGE_FLAG_DMA32)
733 gfp_flags |= GFP_DMA32;
735 gfp_flags |= GFP_HIGHUSER;
737 for (r = 0; r < npages; ++r) {
738 p = alloc_page(gfp_flags);
741 pr_err("Unable to allocate page\n");
750 /* combine zero flag to pool flags */
751 gfp_flags |= pool->gfp_flags;
753 /* First we take pages from the pool */
754 INIT_LIST_HEAD(&plist);
755 npages = ttm_page_pool_get_pages(pool, &plist, flags, cstate, npages);
757 list_for_each_entry(p, &plist, lru) {
761 /* clear the pages coming from the pool if requested */
762 if (flags & TTM_PAGE_FLAG_ZERO_ALLOC) {
763 list_for_each_entry(p, &plist, lru) {
767 clear_page(page_address(p));
771 /* If pool didn't have enough pages allocate new one. */
773 /* ttm_alloc_new_pages doesn't reference pool so we can run
774 * multiple requests in parallel.
776 INIT_LIST_HEAD(&plist);
777 r = ttm_alloc_new_pages(&plist, gfp_flags, flags, cstate, npages);
778 list_for_each_entry(p, &plist, lru) {
782 /* If there is any pages in the list put them back to
784 pr_err("Failed to allocate extra pages for large request\n");
785 ttm_put_pages(pages, count, flags, cstate);
793 static void ttm_page_pool_init_locked(struct ttm_page_pool *pool, int flags,
796 spin_lock_init(&pool->lock);
797 pool->fill_lock = false;
798 INIT_LIST_HEAD(&pool->list);
799 pool->npages = pool->nfrees = 0;
800 pool->gfp_flags = flags;
804 int ttm_page_alloc_init(struct ttm_mem_global *glob, unsigned max_pages)
810 pr_info("Initializing pool allocator\n");
812 _manager = kzalloc(sizeof(*_manager), GFP_KERNEL);
814 ttm_page_pool_init_locked(&_manager->wc_pool, GFP_HIGHUSER, "wc");
816 ttm_page_pool_init_locked(&_manager->uc_pool, GFP_HIGHUSER, "uc");
818 ttm_page_pool_init_locked(&_manager->wc_pool_dma32,
819 GFP_USER | GFP_DMA32, "wc dma");
821 ttm_page_pool_init_locked(&_manager->uc_pool_dma32,
822 GFP_USER | GFP_DMA32, "uc dma");
824 _manager->options.max_size = max_pages;
825 _manager->options.small = SMALL_ALLOCATION;
826 _manager->options.alloc_size = NUM_PAGES_TO_ALLOC;
828 ret = kobject_init_and_add(&_manager->kobj, &ttm_pool_kobj_type,
829 &glob->kobj, "pool");
830 if (unlikely(ret != 0)) {
831 kobject_put(&_manager->kobj);
836 ttm_pool_mm_shrink_init(_manager);
841 void ttm_page_alloc_fini(void)
845 pr_info("Finalizing pool allocator\n");
846 ttm_pool_mm_shrink_fini(_manager);
848 for (i = 0; i < NUM_POOLS; ++i)
849 ttm_page_pool_free(&_manager->pools[i], FREE_ALL_PAGES);
851 kobject_put(&_manager->kobj);
855 int ttm_pool_populate(struct ttm_tt *ttm)
857 struct ttm_mem_global *mem_glob = ttm->glob->mem_glob;
861 if (ttm->state != tt_unpopulated)
864 for (i = 0; i < ttm->num_pages; ++i) {
865 ret = ttm_get_pages(&ttm->pages[i], 1,
869 ttm_pool_unpopulate(ttm);
873 ret = ttm_mem_global_alloc_page(mem_glob, ttm->pages[i],
875 if (unlikely(ret != 0)) {
876 ttm_pool_unpopulate(ttm);
881 if (unlikely(ttm->page_flags & TTM_PAGE_FLAG_SWAPPED)) {
882 ret = ttm_tt_swapin(ttm);
883 if (unlikely(ret != 0)) {
884 ttm_pool_unpopulate(ttm);
889 ttm->state = tt_unbound;
892 EXPORT_SYMBOL(ttm_pool_populate);
894 void ttm_pool_unpopulate(struct ttm_tt *ttm)
898 for (i = 0; i < ttm->num_pages; ++i) {
900 ttm_mem_global_free_page(ttm->glob->mem_glob,
902 ttm_put_pages(&ttm->pages[i], 1,
907 ttm->state = tt_unpopulated;
909 EXPORT_SYMBOL(ttm_pool_unpopulate);
911 int ttm_page_alloc_debugfs(struct seq_file *m, void *data)
913 struct ttm_page_pool *p;
915 char *h[] = {"pool", "refills", "pages freed", "size"};
917 seq_printf(m, "No pool allocator running.\n");
920 seq_printf(m, "%6s %12s %13s %8s\n",
921 h[0], h[1], h[2], h[3]);
922 for (i = 0; i < NUM_POOLS; ++i) {
923 p = &_manager->pools[i];
925 seq_printf(m, "%6s %12ld %13ld %8d\n",
926 p->name, p->nrefills,
927 p->nfrees, p->npages);
931 EXPORT_SYMBOL(ttm_page_alloc_debugfs);