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
33 #include <linux/list.h>
34 #include <linux/spinlock.h>
35 #include <linux/highmem.h>
36 #include <linux/mm_types.h>
37 #include <linux/module.h>
39 #include <linux/seq_file.h> /* for seq_printf */
40 #include <linux/slab.h>
41 #include <linux/dma-mapping.h>
43 #include <linux/atomic.h>
45 #include "ttm/ttm_bo_driver.h"
46 #include "ttm/ttm_page_alloc.h"
52 #define NUM_PAGES_TO_ALLOC (PAGE_SIZE/sizeof(struct page *))
53 #define SMALL_ALLOCATION 16
54 #define FREE_ALL_PAGES (~0U)
55 /* times are in msecs */
56 #define PAGE_FREE_INTERVAL 1000
59 * struct ttm_page_pool - Pool to reuse recently allocated uc/wc pages.
61 * @lock: Protects the shared pool from concurrnet access. Must be used with
62 * irqsave/irqrestore variants because pool allocator maybe called from
64 * @fill_lock: Prevent concurrent calls to fill.
65 * @list: Pool of free uc/wc pages for fast reuse.
66 * @gfp_flags: Flags to pass for alloc_page.
67 * @npages: Number of pages in pool.
69 struct ttm_page_pool {
72 struct list_head list;
77 unsigned long nrefills;
81 * Limits for the pool. They are handled without locks because only place where
82 * they may change is in sysfs store. They won't have immediate effect anyway
83 * so forcing serialization to access them is pointless.
86 struct ttm_pool_opts {
95 * struct ttm_pool_manager - Holds memory pools for fst allocation
97 * Manager is read only object for pool code so it doesn't need locking.
99 * @free_interval: minimum number of jiffies between freeing pages from pool.
100 * @page_alloc_inited: reference counting for pool allocation.
101 * @work: Work that is used to shrink the pool. Work is only run when there is
102 * some pages to free.
103 * @small_allocation: Limit in number of pages what is small allocation.
105 * @pools: All pool objects in use.
107 struct ttm_pool_manager {
109 struct shrinker mm_shrink;
110 struct ttm_pool_opts options;
113 struct ttm_page_pool pools[NUM_POOLS];
115 struct ttm_page_pool wc_pool;
116 struct ttm_page_pool uc_pool;
117 struct ttm_page_pool wc_pool_dma32;
118 struct ttm_page_pool uc_pool_dma32;
123 static struct attribute ttm_page_pool_max = {
124 .name = "pool_max_size",
125 .mode = S_IRUGO | S_IWUSR
127 static struct attribute ttm_page_pool_small = {
128 .name = "pool_small_allocation",
129 .mode = S_IRUGO | S_IWUSR
131 static struct attribute ttm_page_pool_alloc_size = {
132 .name = "pool_allocation_size",
133 .mode = S_IRUGO | S_IWUSR
136 static struct attribute *ttm_pool_attrs[] = {
138 &ttm_page_pool_small,
139 &ttm_page_pool_alloc_size,
143 static void ttm_pool_kobj_release(struct kobject *kobj)
145 struct ttm_pool_manager *m =
146 container_of(kobj, struct ttm_pool_manager, kobj);
150 static ssize_t ttm_pool_store(struct kobject *kobj,
151 struct attribute *attr, const char *buffer, size_t size)
153 struct ttm_pool_manager *m =
154 container_of(kobj, struct ttm_pool_manager, kobj);
157 chars = sscanf(buffer, "%u", &val);
161 /* Convert kb to number of pages */
162 val = val / (PAGE_SIZE >> 10);
164 if (attr == &ttm_page_pool_max)
165 m->options.max_size = val;
166 else if (attr == &ttm_page_pool_small)
167 m->options.small = val;
168 else if (attr == &ttm_page_pool_alloc_size) {
169 if (val > NUM_PAGES_TO_ALLOC*8) {
170 printk(KERN_ERR TTM_PFX
171 "Setting allocation size to %lu "
172 "is not allowed. Recommended size is "
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 printk(KERN_WARNING TTM_PFX
179 "Setting allocation size to "
180 "larger than %lu is not recommended.\n",
181 NUM_PAGES_TO_ALLOC*(PAGE_SIZE >> 10));
183 m->options.alloc_size = val;
189 static ssize_t ttm_pool_show(struct kobject *kobj,
190 struct attribute *attr, char *buffer)
192 struct ttm_pool_manager *m =
193 container_of(kobj, struct ttm_pool_manager, kobj);
196 if (attr == &ttm_page_pool_max)
197 val = m->options.max_size;
198 else if (attr == &ttm_page_pool_small)
199 val = m->options.small;
200 else if (attr == &ttm_page_pool_alloc_size)
201 val = m->options.alloc_size;
203 val = val * (PAGE_SIZE >> 10);
205 return snprintf(buffer, PAGE_SIZE, "%u\n", val);
208 static const struct sysfs_ops ttm_pool_sysfs_ops = {
209 .show = &ttm_pool_show,
210 .store = &ttm_pool_store,
213 static struct kobj_type ttm_pool_kobj_type = {
214 .release = &ttm_pool_kobj_release,
215 .sysfs_ops = &ttm_pool_sysfs_ops,
216 .default_attrs = ttm_pool_attrs,
219 static struct ttm_pool_manager *_manager;
222 static int set_pages_array_wb(struct page **pages, int addrinarray)
227 for (i = 0; i < addrinarray; i++)
228 unmap_page_from_agp(pages[i]);
233 static int set_pages_array_wc(struct page **pages, int addrinarray)
238 for (i = 0; i < addrinarray; i++)
239 map_page_into_agp(pages[i]);
244 static int set_pages_array_uc(struct page **pages, int addrinarray)
249 for (i = 0; i < addrinarray; i++)
250 map_page_into_agp(pages[i]);
257 * Select the right pool or requested caching state and ttm flags. */
258 static struct ttm_page_pool *ttm_get_pool(int flags,
259 enum ttm_caching_state cstate)
263 if (cstate == tt_cached)
271 if (flags & TTM_PAGE_FLAG_DMA32)
274 return &_manager->pools[pool_index];
277 /* set memory back to wb and free the pages. */
278 static void ttm_pages_put(struct page *pages[], unsigned npages)
281 if (set_pages_array_wb(pages, npages))
282 printk(KERN_ERR TTM_PFX "Failed to set %d pages to wb!\n",
284 for (i = 0; i < npages; ++i)
285 __free_page(pages[i]);
288 static void ttm_pool_update_free_locked(struct ttm_page_pool *pool,
289 unsigned freed_pages)
291 pool->npages -= freed_pages;
292 pool->nfrees += freed_pages;
296 * Free pages from pool.
298 * To prevent hogging the ttm_swap process we only free NUM_PAGES_TO_ALLOC
299 * number of pages in one go.
301 * @pool: to free the pages from
302 * @free_all: If set to true will free all pages in pool
304 static int ttm_page_pool_free(struct ttm_page_pool *pool, unsigned nr_free)
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;
315 pages_to_free = kmalloc(npages_to_free * sizeof(struct page *),
317 if (!pages_to_free) {
318 printk(KERN_ERR TTM_PFX
319 "Failed to allocate memory for pool free operation.\n");
324 spin_lock_irqsave(&pool->lock, irq_flags);
326 list_for_each_entry_reverse(p, &pool->list, lru) {
327 if (freed_pages >= npages_to_free)
330 pages_to_free[freed_pages++] = p;
331 /* We can only remove NUM_PAGES_TO_ALLOC at a time. */
332 if (freed_pages >= NUM_PAGES_TO_ALLOC) {
333 /* remove range of pages from the pool */
334 __list_del(p->lru.prev, &pool->list);
336 ttm_pool_update_free_locked(pool, freed_pages);
338 * Because changing page caching is costly
339 * we unlock the pool to prevent stalling.
341 spin_unlock_irqrestore(&pool->lock, irq_flags);
343 ttm_pages_put(pages_to_free, freed_pages);
344 if (likely(nr_free != FREE_ALL_PAGES))
345 nr_free -= freed_pages;
347 if (NUM_PAGES_TO_ALLOC >= nr_free)
348 npages_to_free = nr_free;
350 npages_to_free = NUM_PAGES_TO_ALLOC;
354 /* free all so restart the processing */
358 /* Not allowed to fall through or break because
359 * following context is inside spinlock while we are
367 /* remove range of pages from the pool */
369 __list_del(&p->lru, &pool->list);
371 ttm_pool_update_free_locked(pool, freed_pages);
372 nr_free -= freed_pages;
375 spin_unlock_irqrestore(&pool->lock, irq_flags);
378 ttm_pages_put(pages_to_free, freed_pages);
380 kfree(pages_to_free);
384 /* Get good estimation how many pages are free in pools */
385 static int ttm_pool_get_num_unused_pages(void)
389 for (i = 0; i < NUM_POOLS; ++i)
390 total += _manager->pools[i].npages;
396 * Callback for mm to request pool to reduce number of page held.
398 static int ttm_pool_mm_shrink(struct shrinker *shrink,
399 struct shrink_control *sc)
401 static DEFINE_MUTEX(lock);
402 static unsigned start_pool;
404 unsigned pool_offset;
405 struct ttm_page_pool *pool;
406 int shrink_pages = sc->nr_to_scan;
408 if (shrink_pages == 0)
410 if (!mutex_trylock(&lock))
412 pool_offset = ++start_pool % NUM_POOLS;
413 /* select start pool in round robin fashion */
414 for (i = 0; i < NUM_POOLS; ++i) {
415 unsigned nr_free = shrink_pages;
416 if (shrink_pages == 0)
418 pool = &_manager->pools[(i + pool_offset)%NUM_POOLS];
419 shrink_pages = ttm_page_pool_free(pool, nr_free);
423 /* return estimated number of unused pages in pool */
424 return ttm_pool_get_num_unused_pages();
427 static void ttm_pool_mm_shrink_init(struct ttm_pool_manager *manager)
429 manager->mm_shrink.shrink = &ttm_pool_mm_shrink;
430 manager->mm_shrink.seeks = 1;
431 register_shrinker(&manager->mm_shrink);
434 static void ttm_pool_mm_shrink_fini(struct ttm_pool_manager *manager)
436 unregister_shrinker(&manager->mm_shrink);
439 static int ttm_set_pages_caching(struct page **pages,
440 enum ttm_caching_state cstate, unsigned cpages)
443 /* Set page caching */
446 r = set_pages_array_uc(pages, cpages);
448 printk(KERN_ERR TTM_PFX
449 "Failed to set %d pages to uc!\n",
453 r = set_pages_array_wc(pages, cpages);
455 printk(KERN_ERR TTM_PFX
456 "Failed to set %d pages to wc!\n",
466 * Free pages the pages that failed to change the caching state. If there is
467 * any pages that have changed their caching state already put them to the
470 static void ttm_handle_caching_state_failure(struct list_head *pages,
471 int ttm_flags, enum ttm_caching_state cstate,
472 struct page **failed_pages, unsigned cpages)
475 /* Failed pages have to be freed */
476 for (i = 0; i < cpages; ++i) {
477 list_del(&failed_pages[i]->lru);
478 __free_page(failed_pages[i]);
483 * Allocate new pages with correct caching.
485 * This function is reentrant if caller updates count depending on number of
486 * pages returned in pages array.
488 static int ttm_alloc_new_pages(struct list_head *pages, gfp_t gfp_flags,
489 int ttm_flags, enum ttm_caching_state cstate, unsigned count)
491 struct page **caching_array;
495 unsigned max_cpages = min(count,
496 (unsigned)(PAGE_SIZE/sizeof(struct page *)));
498 /* allocate array for page caching change */
499 caching_array = kmalloc(max_cpages*sizeof(struct page *), GFP_KERNEL);
501 if (!caching_array) {
502 printk(KERN_ERR TTM_PFX
503 "Unable to allocate table for new pages.");
507 for (i = 0, cpages = 0; i < count; ++i) {
508 p = alloc_page(gfp_flags);
511 printk(KERN_ERR TTM_PFX "Unable to get page %u.\n", i);
513 /* store already allocated pages in the pool after
514 * setting the caching state */
516 r = ttm_set_pages_caching(caching_array,
519 ttm_handle_caching_state_failure(pages,
521 caching_array, cpages);
527 #ifdef CONFIG_HIGHMEM
528 /* gfp flags of highmem page should never be dma32 so we
529 * we should be fine in such case
534 caching_array[cpages++] = p;
535 if (cpages == max_cpages) {
537 r = ttm_set_pages_caching(caching_array,
540 ttm_handle_caching_state_failure(pages,
542 caching_array, cpages);
549 list_add(&p->lru, pages);
553 r = ttm_set_pages_caching(caching_array, cstate, cpages);
555 ttm_handle_caching_state_failure(pages,
557 caching_array, cpages);
560 kfree(caching_array);
566 * Fill the given pool if there aren't enough pages and the requested number of
569 static void ttm_page_pool_fill_locked(struct ttm_page_pool *pool,
570 int ttm_flags, enum ttm_caching_state cstate, unsigned count,
571 unsigned long *irq_flags)
577 * Only allow one pool fill operation at a time.
578 * If pool doesn't have enough pages for the allocation new pages are
579 * allocated from outside of pool.
584 pool->fill_lock = true;
586 /* If allocation request is small and there are not enough
587 * pages in a pool we fill the pool up first. */
588 if (count < _manager->options.small
589 && count > pool->npages) {
590 struct list_head new_pages;
591 unsigned alloc_size = _manager->options.alloc_size;
594 * Can't change page caching if in irqsave context. We have to
595 * drop the pool->lock.
597 spin_unlock_irqrestore(&pool->lock, *irq_flags);
599 INIT_LIST_HEAD(&new_pages);
600 r = ttm_alloc_new_pages(&new_pages, pool->gfp_flags, ttm_flags,
602 spin_lock_irqsave(&pool->lock, *irq_flags);
605 list_splice(&new_pages, &pool->list);
607 pool->npages += alloc_size;
609 printk(KERN_ERR TTM_PFX
610 "Failed to fill pool (%p).", pool);
611 /* If we have any pages left put them to the pool. */
612 list_for_each_entry(p, &new_pages, lru) {
615 list_splice(&new_pages, &pool->list);
616 pool->npages += cpages;
620 pool->fill_lock = false;
624 * Cut 'count' number of pages from the pool and put them on the return list.
626 * @return count of pages still required to fulfill the request.
628 static unsigned ttm_page_pool_get_pages(struct ttm_page_pool *pool,
629 struct list_head *pages, int ttm_flags,
630 enum ttm_caching_state cstate, unsigned count)
632 unsigned long irq_flags;
636 spin_lock_irqsave(&pool->lock, irq_flags);
637 ttm_page_pool_fill_locked(pool, ttm_flags, cstate, count, &irq_flags);
639 if (count >= pool->npages) {
640 /* take all pages from the pool */
641 list_splice_init(&pool->list, pages);
642 count -= pool->npages;
646 /* find the last pages to include for requested number of pages. Split
647 * pool to begin and halve it to reduce search space. */
648 if (count <= pool->npages/2) {
650 list_for_each(p, &pool->list) {
655 i = pool->npages + 1;
656 list_for_each_prev(p, &pool->list) {
661 /* Cut 'count' number of pages from the pool */
662 list_cut_position(pages, &pool->list, p);
663 pool->npages -= count;
666 spin_unlock_irqrestore(&pool->lock, irq_flags);
671 * On success pages list will hold count number of correctly
674 int ttm_get_pages(struct list_head *pages, int flags,
675 enum ttm_caching_state cstate, unsigned count,
676 dma_addr_t *dma_address)
678 struct ttm_page_pool *pool = ttm_get_pool(flags, cstate);
679 struct page *p = NULL;
680 gfp_t gfp_flags = GFP_USER;
683 /* set zero flag for page allocation if required */
684 if (flags & TTM_PAGE_FLAG_ZERO_ALLOC)
685 gfp_flags |= __GFP_ZERO;
687 /* No pool for cached pages */
689 if (flags & TTM_PAGE_FLAG_DMA32)
690 gfp_flags |= GFP_DMA32;
692 gfp_flags |= GFP_HIGHUSER;
694 for (r = 0; r < count; ++r) {
695 p = alloc_page(gfp_flags);
698 printk(KERN_ERR TTM_PFX
699 "Unable to allocate page.");
703 list_add(&p->lru, pages);
709 /* combine zero flag to pool flags */
710 gfp_flags |= pool->gfp_flags;
712 /* First we take pages from the pool */
713 count = ttm_page_pool_get_pages(pool, pages, flags, cstate, count);
715 /* clear the pages coming from the pool if requested */
716 if (flags & TTM_PAGE_FLAG_ZERO_ALLOC) {
717 list_for_each_entry(p, pages, lru) {
721 clear_page(page_address(p));
725 /* If pool didn't have enough pages allocate new one. */
727 /* ttm_alloc_new_pages doesn't reference pool so we can run
728 * multiple requests in parallel.
730 r = ttm_alloc_new_pages(pages, gfp_flags, flags, cstate, count);
732 /* If there is any pages in the list put them back to
734 printk(KERN_ERR TTM_PFX
735 "Failed to allocate extra pages "
736 "for large request.");
737 ttm_put_pages(pages, 0, flags, cstate, NULL);
746 /* Put all pages in pages list to correct pool to wait for reuse */
747 void ttm_put_pages(struct list_head *pages, unsigned page_count, int flags,
748 enum ttm_caching_state cstate, dma_addr_t *dma_address)
750 unsigned long irq_flags;
751 struct ttm_page_pool *pool = ttm_get_pool(flags, cstate);
752 struct page *p, *tmp;
755 /* No pool for this memory type so free the pages */
757 list_for_each_entry_safe(p, tmp, pages, lru) {
760 /* Make the pages list empty */
761 INIT_LIST_HEAD(pages);
764 if (page_count == 0) {
765 list_for_each_entry_safe(p, tmp, pages, lru) {
770 spin_lock_irqsave(&pool->lock, irq_flags);
771 list_splice_init(pages, &pool->list);
772 pool->npages += page_count;
773 /* Check that we don't go over the pool limit */
775 if (pool->npages > _manager->options.max_size) {
776 page_count = pool->npages - _manager->options.max_size;
777 /* free at least NUM_PAGES_TO_ALLOC number of pages
778 * to reduce calls to set_memory_wb */
779 if (page_count < NUM_PAGES_TO_ALLOC)
780 page_count = NUM_PAGES_TO_ALLOC;
782 spin_unlock_irqrestore(&pool->lock, irq_flags);
784 ttm_page_pool_free(pool, page_count);
787 static void ttm_page_pool_init_locked(struct ttm_page_pool *pool, int flags,
790 spin_lock_init(&pool->lock);
791 pool->fill_lock = false;
792 INIT_LIST_HEAD(&pool->list);
793 pool->npages = pool->nfrees = 0;
794 pool->gfp_flags = flags;
798 int ttm_page_alloc_init(struct ttm_mem_global *glob, unsigned max_pages)
804 printk(KERN_INFO TTM_PFX "Initializing pool allocator.\n");
806 _manager = kzalloc(sizeof(*_manager), GFP_KERNEL);
808 ttm_page_pool_init_locked(&_manager->wc_pool, GFP_HIGHUSER, "wc");
810 ttm_page_pool_init_locked(&_manager->uc_pool, GFP_HIGHUSER, "uc");
812 ttm_page_pool_init_locked(&_manager->wc_pool_dma32,
813 GFP_USER | GFP_DMA32, "wc dma");
815 ttm_page_pool_init_locked(&_manager->uc_pool_dma32,
816 GFP_USER | GFP_DMA32, "uc dma");
818 _manager->options.max_size = max_pages;
819 _manager->options.small = SMALL_ALLOCATION;
820 _manager->options.alloc_size = NUM_PAGES_TO_ALLOC;
822 ret = kobject_init_and_add(&_manager->kobj, &ttm_pool_kobj_type,
823 &glob->kobj, "pool");
824 if (unlikely(ret != 0)) {
825 kobject_put(&_manager->kobj);
830 ttm_pool_mm_shrink_init(_manager);
835 void ttm_page_alloc_fini(void)
839 printk(KERN_INFO TTM_PFX "Finalizing pool allocator.\n");
840 ttm_pool_mm_shrink_fini(_manager);
842 for (i = 0; i < NUM_POOLS; ++i)
843 ttm_page_pool_free(&_manager->pools[i], FREE_ALL_PAGES);
845 kobject_put(&_manager->kobj);
849 int ttm_page_alloc_debugfs(struct seq_file *m, void *data)
851 struct ttm_page_pool *p;
853 char *h[] = {"pool", "refills", "pages freed", "size"};
855 seq_printf(m, "No pool allocator running.\n");
858 seq_printf(m, "%6s %12s %13s %8s\n",
859 h[0], h[1], h[2], h[3]);
860 for (i = 0; i < NUM_POOLS; ++i) {
861 p = &_manager->pools[i];
863 seq_printf(m, "%6s %12ld %13ld %8d\n",
864 p->name, p->nrefills,
865 p->nfrees, p->npages);
869 EXPORT_SYMBOL(ttm_page_alloc_debugfs);