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.
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>
51 #if IS_ENABLED(CONFIG_AGP)
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)
222 #if IS_ENABLED(CONFIG_AGP)
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)
233 #if IS_ENABLED(CONFIG_AGP)
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)
244 #if IS_ENABLED(CONFIG_AGP)
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
300 * @use_static: Safe to use static buffer
302 static int ttm_page_pool_free(struct ttm_page_pool *pool, unsigned nr_free,
305 static struct page *static_buf[NUM_PAGES_TO_ALLOC];
306 unsigned long irq_flags;
308 struct page **pages_to_free;
309 unsigned freed_pages = 0,
310 npages_to_free = nr_free;
312 if (NUM_PAGES_TO_ALLOC < nr_free)
313 npages_to_free = NUM_PAGES_TO_ALLOC;
316 pages_to_free = static_buf;
318 pages_to_free = kmalloc(npages_to_free * sizeof(struct page *),
320 if (!pages_to_free) {
321 pr_err("Failed to allocate memory for pool free operation\n");
326 spin_lock_irqsave(&pool->lock, irq_flags);
328 list_for_each_entry_reverse(p, &pool->list, lru) {
329 if (freed_pages >= npages_to_free)
332 pages_to_free[freed_pages++] = p;
333 /* We can only remove NUM_PAGES_TO_ALLOC at a time. */
334 if (freed_pages >= NUM_PAGES_TO_ALLOC) {
335 /* remove range of pages from the pool */
336 __list_del(p->lru.prev, &pool->list);
338 ttm_pool_update_free_locked(pool, freed_pages);
340 * Because changing page caching is costly
341 * we unlock the pool to prevent stalling.
343 spin_unlock_irqrestore(&pool->lock, irq_flags);
345 ttm_pages_put(pages_to_free, freed_pages);
346 if (likely(nr_free != FREE_ALL_PAGES))
347 nr_free -= freed_pages;
349 if (NUM_PAGES_TO_ALLOC >= nr_free)
350 npages_to_free = nr_free;
352 npages_to_free = NUM_PAGES_TO_ALLOC;
356 /* free all so restart the processing */
360 /* Not allowed to fall through or break because
361 * following context is inside spinlock while we are
369 /* remove range of pages from the pool */
371 __list_del(&p->lru, &pool->list);
373 ttm_pool_update_free_locked(pool, freed_pages);
374 nr_free -= freed_pages;
377 spin_unlock_irqrestore(&pool->lock, irq_flags);
380 ttm_pages_put(pages_to_free, freed_pages);
382 if (pages_to_free != static_buf)
383 kfree(pages_to_free);
388 * Callback for mm to request pool to reduce number of page held.
390 * XXX: (dchinner) Deadlock warning!
392 * This code is crying out for a shrinker per pool....
395 ttm_pool_shrink_scan(struct shrinker *shrink, struct shrink_control *sc)
397 static DEFINE_MUTEX(lock);
398 static unsigned start_pool;
400 unsigned pool_offset;
401 struct ttm_page_pool *pool;
402 int shrink_pages = sc->nr_to_scan;
403 unsigned long freed = 0;
405 if (!mutex_trylock(&lock))
407 pool_offset = ++start_pool % NUM_POOLS;
408 /* select start pool in round robin fashion */
409 for (i = 0; i < NUM_POOLS; ++i) {
410 unsigned nr_free = shrink_pages;
411 if (shrink_pages == 0)
413 pool = &_manager->pools[(i + pool_offset)%NUM_POOLS];
414 /* OK to use static buffer since global mutex is held. */
415 shrink_pages = ttm_page_pool_free(pool, nr_free, true);
416 freed += nr_free - shrink_pages;
424 ttm_pool_shrink_count(struct shrinker *shrink, struct shrink_control *sc)
427 unsigned long count = 0;
429 for (i = 0; i < NUM_POOLS; ++i)
430 count += _manager->pools[i].npages;
435 static void ttm_pool_mm_shrink_init(struct ttm_pool_manager *manager)
437 manager->mm_shrink.count_objects = ttm_pool_shrink_count;
438 manager->mm_shrink.scan_objects = ttm_pool_shrink_scan;
439 manager->mm_shrink.seeks = 1;
440 register_shrinker(&manager->mm_shrink);
443 static void ttm_pool_mm_shrink_fini(struct ttm_pool_manager *manager)
445 unregister_shrinker(&manager->mm_shrink);
448 static int ttm_set_pages_caching(struct page **pages,
449 enum ttm_caching_state cstate, unsigned cpages)
452 /* Set page caching */
455 r = set_pages_array_uc(pages, cpages);
457 pr_err("Failed to set %d pages to uc!\n", cpages);
460 r = set_pages_array_wc(pages, cpages);
462 pr_err("Failed to set %d pages to wc!\n", cpages);
471 * Free pages the pages that failed to change the caching state. If there is
472 * any pages that have changed their caching state already put them to the
475 static void ttm_handle_caching_state_failure(struct list_head *pages,
476 int ttm_flags, enum ttm_caching_state cstate,
477 struct page **failed_pages, unsigned cpages)
480 /* Failed pages have to be freed */
481 for (i = 0; i < cpages; ++i) {
482 list_del(&failed_pages[i]->lru);
483 __free_page(failed_pages[i]);
488 * Allocate new pages with correct caching.
490 * This function is reentrant if caller updates count depending on number of
491 * pages returned in pages array.
493 static int ttm_alloc_new_pages(struct list_head *pages, gfp_t gfp_flags,
494 int ttm_flags, enum ttm_caching_state cstate, unsigned count)
496 struct page **caching_array;
500 unsigned max_cpages = min(count,
501 (unsigned)(PAGE_SIZE/sizeof(struct page *)));
503 /* allocate array for page caching change */
504 caching_array = kmalloc(max_cpages*sizeof(struct page *), GFP_KERNEL);
506 if (!caching_array) {
507 pr_err("Unable to allocate table for new pages\n");
511 for (i = 0, cpages = 0; i < count; ++i) {
512 p = alloc_page(gfp_flags);
515 pr_err("Unable to get page %u\n", i);
517 /* store already allocated pages in the pool after
518 * setting the caching state */
520 r = ttm_set_pages_caching(caching_array,
523 ttm_handle_caching_state_failure(pages,
525 caching_array, cpages);
531 #ifdef CONFIG_HIGHMEM
532 /* gfp flags of highmem page should never be dma32 so we
533 * we should be fine in such case
538 caching_array[cpages++] = p;
539 if (cpages == max_cpages) {
541 r = ttm_set_pages_caching(caching_array,
544 ttm_handle_caching_state_failure(pages,
546 caching_array, cpages);
553 list_add(&p->lru, pages);
557 r = ttm_set_pages_caching(caching_array, cstate, cpages);
559 ttm_handle_caching_state_failure(pages,
561 caching_array, cpages);
564 kfree(caching_array);
570 * Fill the given pool if there aren't enough pages and the requested number of
573 static void ttm_page_pool_fill_locked(struct ttm_page_pool *pool,
574 int ttm_flags, enum ttm_caching_state cstate, unsigned count,
575 unsigned long *irq_flags)
581 * Only allow one pool fill operation at a time.
582 * If pool doesn't have enough pages for the allocation new pages are
583 * allocated from outside of pool.
588 pool->fill_lock = true;
590 /* If allocation request is small and there are not enough
591 * pages in a pool we fill the pool up first. */
592 if (count < _manager->options.small
593 && count > pool->npages) {
594 struct list_head new_pages;
595 unsigned alloc_size = _manager->options.alloc_size;
598 * Can't change page caching if in irqsave context. We have to
599 * drop the pool->lock.
601 spin_unlock_irqrestore(&pool->lock, *irq_flags);
603 INIT_LIST_HEAD(&new_pages);
604 r = ttm_alloc_new_pages(&new_pages, pool->gfp_flags, ttm_flags,
606 spin_lock_irqsave(&pool->lock, *irq_flags);
609 list_splice(&new_pages, &pool->list);
611 pool->npages += alloc_size;
613 pr_err("Failed to fill pool (%p)\n", pool);
614 /* If we have any pages left put them to the pool. */
615 list_for_each_entry(p, &pool->list, lru) {
618 list_splice(&new_pages, &pool->list);
619 pool->npages += cpages;
623 pool->fill_lock = false;
627 * Cut 'count' number of pages from the pool and put them on the return list.
629 * @return count of pages still required to fulfill the request.
631 static unsigned ttm_page_pool_get_pages(struct ttm_page_pool *pool,
632 struct list_head *pages,
634 enum ttm_caching_state cstate,
637 unsigned long irq_flags;
641 spin_lock_irqsave(&pool->lock, irq_flags);
642 ttm_page_pool_fill_locked(pool, ttm_flags, cstate, count, &irq_flags);
644 if (count >= pool->npages) {
645 /* take all pages from the pool */
646 list_splice_init(&pool->list, pages);
647 count -= pool->npages;
651 /* find the last pages to include for requested number of pages. Split
652 * pool to begin and halve it to reduce search space. */
653 if (count <= pool->npages/2) {
655 list_for_each(p, &pool->list) {
660 i = pool->npages + 1;
661 list_for_each_prev(p, &pool->list) {
666 /* Cut 'count' number of pages from the pool */
667 list_cut_position(pages, &pool->list, p);
668 pool->npages -= count;
671 spin_unlock_irqrestore(&pool->lock, irq_flags);
675 /* Put all pages in pages list to correct pool to wait for reuse */
676 static void ttm_put_pages(struct page **pages, unsigned npages, int flags,
677 enum ttm_caching_state cstate)
679 unsigned long irq_flags;
680 struct ttm_page_pool *pool = ttm_get_pool(flags, cstate);
684 /* No pool for this memory type so free the pages */
685 for (i = 0; i < npages; i++) {
687 if (page_count(pages[i]) != 1)
688 pr_err("Erroneous page count. Leaking pages.\n");
689 __free_page(pages[i]);
696 spin_lock_irqsave(&pool->lock, irq_flags);
697 for (i = 0; i < npages; i++) {
699 if (page_count(pages[i]) != 1)
700 pr_err("Erroneous page count. Leaking pages.\n");
701 list_add_tail(&pages[i]->lru, &pool->list);
706 /* Check that we don't go over the pool limit */
708 if (pool->npages > _manager->options.max_size) {
709 npages = pool->npages - _manager->options.max_size;
710 /* free at least NUM_PAGES_TO_ALLOC number of pages
711 * to reduce calls to set_memory_wb */
712 if (npages < NUM_PAGES_TO_ALLOC)
713 npages = NUM_PAGES_TO_ALLOC;
715 spin_unlock_irqrestore(&pool->lock, irq_flags);
717 ttm_page_pool_free(pool, npages, false);
721 * On success pages list will hold count number of correctly
724 static int ttm_get_pages(struct page **pages, unsigned npages, int flags,
725 enum ttm_caching_state cstate)
727 struct ttm_page_pool *pool = ttm_get_pool(flags, cstate);
728 struct list_head plist;
729 struct page *p = NULL;
730 gfp_t gfp_flags = GFP_USER;
734 /* set zero flag for page allocation if required */
735 if (flags & TTM_PAGE_FLAG_ZERO_ALLOC)
736 gfp_flags |= __GFP_ZERO;
738 /* No pool for cached pages */
740 if (flags & TTM_PAGE_FLAG_DMA32)
741 gfp_flags |= GFP_DMA32;
743 gfp_flags |= GFP_HIGHUSER;
745 for (r = 0; r < npages; ++r) {
746 p = alloc_page(gfp_flags);
749 pr_err("Unable to allocate page\n");
758 /* combine zero flag to pool flags */
759 gfp_flags |= pool->gfp_flags;
761 /* First we take pages from the pool */
762 INIT_LIST_HEAD(&plist);
763 npages = ttm_page_pool_get_pages(pool, &plist, flags, cstate, npages);
765 list_for_each_entry(p, &plist, lru) {
769 /* clear the pages coming from the pool if requested */
770 if (flags & TTM_PAGE_FLAG_ZERO_ALLOC) {
771 list_for_each_entry(p, &plist, lru) {
775 clear_page(page_address(p));
779 /* If pool didn't have enough pages allocate new one. */
781 /* ttm_alloc_new_pages doesn't reference pool so we can run
782 * multiple requests in parallel.
784 INIT_LIST_HEAD(&plist);
785 r = ttm_alloc_new_pages(&plist, gfp_flags, flags, cstate, npages);
786 list_for_each_entry(p, &plist, lru) {
790 /* If there is any pages in the list put them back to
792 pr_err("Failed to allocate extra pages for large request\n");
793 ttm_put_pages(pages, count, flags, cstate);
801 static void ttm_page_pool_init_locked(struct ttm_page_pool *pool, gfp_t flags,
804 spin_lock_init(&pool->lock);
805 pool->fill_lock = false;
806 INIT_LIST_HEAD(&pool->list);
807 pool->npages = pool->nfrees = 0;
808 pool->gfp_flags = flags;
812 int ttm_page_alloc_init(struct ttm_mem_global *glob, unsigned max_pages)
818 pr_info("Initializing pool allocator\n");
820 _manager = kzalloc(sizeof(*_manager), GFP_KERNEL);
822 ttm_page_pool_init_locked(&_manager->wc_pool, GFP_HIGHUSER, "wc");
824 ttm_page_pool_init_locked(&_manager->uc_pool, GFP_HIGHUSER, "uc");
826 ttm_page_pool_init_locked(&_manager->wc_pool_dma32,
827 GFP_USER | GFP_DMA32, "wc dma");
829 ttm_page_pool_init_locked(&_manager->uc_pool_dma32,
830 GFP_USER | GFP_DMA32, "uc dma");
832 _manager->options.max_size = max_pages;
833 _manager->options.small = SMALL_ALLOCATION;
834 _manager->options.alloc_size = NUM_PAGES_TO_ALLOC;
836 ret = kobject_init_and_add(&_manager->kobj, &ttm_pool_kobj_type,
837 &glob->kobj, "pool");
838 if (unlikely(ret != 0)) {
839 kobject_put(&_manager->kobj);
844 ttm_pool_mm_shrink_init(_manager);
849 void ttm_page_alloc_fini(void)
853 pr_info("Finalizing pool allocator\n");
854 ttm_pool_mm_shrink_fini(_manager);
856 /* OK to use static buffer since global mutex is no longer used. */
857 for (i = 0; i < NUM_POOLS; ++i)
858 ttm_page_pool_free(&_manager->pools[i], FREE_ALL_PAGES, true);
860 kobject_put(&_manager->kobj);
864 int ttm_pool_populate(struct ttm_tt *ttm)
866 struct ttm_mem_global *mem_glob = ttm->glob->mem_glob;
870 if (ttm->state != tt_unpopulated)
873 for (i = 0; i < ttm->num_pages; ++i) {
874 ret = ttm_get_pages(&ttm->pages[i], 1,
878 ttm_pool_unpopulate(ttm);
882 ret = ttm_mem_global_alloc_page(mem_glob, ttm->pages[i],
884 if (unlikely(ret != 0)) {
885 ttm_pool_unpopulate(ttm);
890 if (unlikely(ttm->page_flags & TTM_PAGE_FLAG_SWAPPED)) {
891 ret = ttm_tt_swapin(ttm);
892 if (unlikely(ret != 0)) {
893 ttm_pool_unpopulate(ttm);
898 ttm->state = tt_unbound;
901 EXPORT_SYMBOL(ttm_pool_populate);
903 void ttm_pool_unpopulate(struct ttm_tt *ttm)
907 for (i = 0; i < ttm->num_pages; ++i) {
909 ttm_mem_global_free_page(ttm->glob->mem_glob,
911 ttm_put_pages(&ttm->pages[i], 1,
916 ttm->state = tt_unpopulated;
918 EXPORT_SYMBOL(ttm_pool_unpopulate);
920 int ttm_page_alloc_debugfs(struct seq_file *m, void *data)
922 struct ttm_page_pool *p;
924 char *h[] = {"pool", "refills", "pages freed", "size"};
926 seq_printf(m, "No pool allocator running.\n");
929 seq_printf(m, "%6s %12s %13s %8s\n",
930 h[0], h[1], h[2], h[3]);
931 for (i = 0; i < NUM_POOLS; ++i) {
932 p = &_manager->pools[i];
934 seq_printf(m, "%6s %12ld %13ld %8d\n",
935 p->name, p->nrefills,
936 p->nfrees, p->npages);
940 EXPORT_SYMBOL(ttm_page_alloc_debugfs);