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>
42 #include <asm/atomic.h>
45 #include "ttm/ttm_bo_driver.h"
46 #include "ttm/ttm_page_alloc.h"
49 #define NUM_PAGES_TO_ALLOC (PAGE_SIZE/sizeof(struct page *))
50 #define SMALL_ALLOCATION 16
51 #define FREE_ALL_PAGES (~0U)
52 /* times are in msecs */
53 #define PAGE_FREE_INTERVAL 1000
56 * struct ttm_page_pool - Pool to reuse recently allocated uc/wc pages.
58 * @lock: Protects the shared pool from concurrnet access. Must be used with
59 * irqsave/irqrestore variants because pool allocator maybe called from
61 * @fill_lock: Prevent concurrent calls to fill.
62 * @list: Pool of free uc/wc pages for fast reuse.
63 * @gfp_flags: Flags to pass for alloc_page.
64 * @npages: Number of pages in pool.
66 struct ttm_page_pool {
69 struct list_head list;
74 unsigned long nrefills;
78 * Limits for the pool. They are handled without locks because only place where
79 * they may change is in sysfs store. They won't have immediate effect anyway
80 * so forcing serialization to access them is pointless.
83 struct ttm_pool_opts {
92 * struct ttm_pool_manager - Holds memory pools for fst allocation
94 * Manager is read only object for pool code so it doesn't need locking.
96 * @free_interval: minimum number of jiffies between freeing pages from pool.
97 * @page_alloc_inited: reference counting for pool allocation.
98 * @work: Work that is used to shrink the pool. Work is only run when there is
100 * @small_allocation: Limit in number of pages what is small allocation.
102 * @pools: All pool objects in use.
104 struct ttm_pool_manager {
106 struct shrinker mm_shrink;
107 atomic_t page_alloc_inited;
108 struct ttm_pool_opts options;
111 struct ttm_page_pool pools[NUM_POOLS];
113 struct ttm_page_pool wc_pool;
114 struct ttm_page_pool uc_pool;
115 struct ttm_page_pool wc_pool_dma32;
116 struct ttm_page_pool uc_pool_dma32;
121 static struct attribute ttm_page_pool_max = {
122 .name = "pool_max_size",
123 .mode = S_IRUGO | S_IWUSR
125 static struct attribute ttm_page_pool_small = {
126 .name = "pool_small_allocation",
127 .mode = S_IRUGO | S_IWUSR
129 static struct attribute ttm_page_pool_alloc_size = {
130 .name = "pool_allocation_size",
131 .mode = S_IRUGO | S_IWUSR
134 static struct attribute *ttm_pool_attrs[] = {
136 &ttm_page_pool_small,
137 &ttm_page_pool_alloc_size,
141 static void ttm_pool_kobj_release(struct kobject *kobj)
143 struct ttm_pool_manager *m =
144 container_of(kobj, struct ttm_pool_manager, kobj);
148 static ssize_t ttm_pool_store(struct kobject *kobj,
149 struct attribute *attr, const char *buffer, size_t size)
151 struct ttm_pool_manager *m =
152 container_of(kobj, struct ttm_pool_manager, kobj);
155 chars = sscanf(buffer, "%u", &val);
159 /* Convert kb to number of pages */
160 val = val / (PAGE_SIZE >> 10);
162 if (attr == &ttm_page_pool_max)
163 m->options.max_size = val;
164 else if (attr == &ttm_page_pool_small)
165 m->options.small = val;
166 else if (attr == &ttm_page_pool_alloc_size) {
167 if (val > NUM_PAGES_TO_ALLOC*8) {
168 printk(KERN_ERR TTM_PFX
169 "Setting allocation size to %lu "
170 "is not allowed. Recommended size is "
172 NUM_PAGES_TO_ALLOC*(PAGE_SIZE >> 7),
173 NUM_PAGES_TO_ALLOC*(PAGE_SIZE >> 10));
175 } else if (val > NUM_PAGES_TO_ALLOC) {
176 printk(KERN_WARNING TTM_PFX
177 "Setting allocation size to "
178 "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 = {
218 .page_alloc_inited = ATOMIC_INIT(0)
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 tough 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(int shrink_pages, gfp_t gfp_mask)
400 static atomic_t start_pool = ATOMIC_INIT(0);
402 unsigned pool_offset = atomic_add_return(1, &start_pool);
403 struct ttm_page_pool *pool;
405 pool_offset = pool_offset % NUM_POOLS;
406 /* select start pool in round robin fashion */
407 for (i = 0; i < NUM_POOLS; ++i) {
408 unsigned nr_free = shrink_pages;
409 if (shrink_pages == 0)
411 pool = &_manager.pools[(i + pool_offset)%NUM_POOLS];
412 shrink_pages = ttm_page_pool_free(pool, nr_free);
414 /* return estimated number of unused pages in pool */
415 return ttm_pool_get_num_unused_pages();
418 static void ttm_pool_mm_shrink_init(struct ttm_pool_manager *manager)
420 manager->mm_shrink.shrink = &ttm_pool_mm_shrink;
421 manager->mm_shrink.seeks = 1;
422 register_shrinker(&manager->mm_shrink);
425 static void ttm_pool_mm_shrink_fini(struct ttm_pool_manager *manager)
427 unregister_shrinker(&manager->mm_shrink);
430 static int ttm_set_pages_caching(struct page **pages,
431 enum ttm_caching_state cstate, unsigned cpages)
434 /* Set page caching */
437 r = set_pages_array_uc(pages, cpages);
439 printk(KERN_ERR TTM_PFX
440 "Failed to set %d pages to uc!\n",
444 r = set_pages_array_wc(pages, cpages);
446 printk(KERN_ERR TTM_PFX
447 "Failed to set %d pages to wc!\n",
457 * Free pages the pages that failed to change the caching state. If there is
458 * any pages that have changed their caching state already put them to the
461 static void ttm_handle_caching_state_failure(struct list_head *pages,
462 int ttm_flags, enum ttm_caching_state cstate,
463 struct page **failed_pages, unsigned cpages)
466 /* Failed pages have to be freed */
467 for (i = 0; i < cpages; ++i) {
468 list_del(&failed_pages[i]->lru);
469 __free_page(failed_pages[i]);
474 * Allocate new pages with correct caching.
476 * This function is reentrant if caller updates count depending on number of
477 * pages returned in pages array.
479 static int ttm_alloc_new_pages(struct list_head *pages, int gfp_flags,
480 int ttm_flags, enum ttm_caching_state cstate, unsigned count)
482 struct page **caching_array;
486 unsigned max_cpages = min(count,
487 (unsigned)(PAGE_SIZE/sizeof(struct page *)));
489 /* allocate array for page caching change */
490 caching_array = kmalloc(max_cpages*sizeof(struct page *), GFP_KERNEL);
492 if (!caching_array) {
493 printk(KERN_ERR TTM_PFX
494 "Unable to allocate table for new pages.");
498 for (i = 0, cpages = 0; i < count; ++i) {
499 p = alloc_page(gfp_flags);
502 printk(KERN_ERR TTM_PFX "Unable to get page %u.\n", i);
504 /* store already allocated pages in the pool after
505 * setting the caching state */
507 r = ttm_set_pages_caching(caching_array,
510 ttm_handle_caching_state_failure(pages,
512 caching_array, cpages);
518 #ifdef CONFIG_HIGHMEM
519 /* gfp flags of highmem page should never be dma32 so we
520 * we should be fine in such case
525 caching_array[cpages++] = p;
526 if (cpages == max_cpages) {
528 r = ttm_set_pages_caching(caching_array,
531 ttm_handle_caching_state_failure(pages,
533 caching_array, cpages);
540 list_add(&p->lru, pages);
544 r = ttm_set_pages_caching(caching_array, cstate, cpages);
546 ttm_handle_caching_state_failure(pages,
548 caching_array, cpages);
551 kfree(caching_array);
557 * Fill the given pool if there isn't enough pages and requested number of
560 static void ttm_page_pool_fill_locked(struct ttm_page_pool *pool,
561 int ttm_flags, enum ttm_caching_state cstate, unsigned count,
562 unsigned long *irq_flags)
568 * Only allow one pool fill operation at a time.
569 * If pool doesn't have enough pages for the allocation new pages are
570 * allocated from outside of pool.
575 pool->fill_lock = true;
577 /* If allocation request is small and there is not enough
578 * pages in pool we fill the pool first */
579 if (count < _manager.options.small
580 && count > pool->npages) {
581 struct list_head new_pages;
582 unsigned alloc_size = _manager.options.alloc_size;
585 * Can't change page caching if in irqsave context. We have to
586 * drop the pool->lock.
588 spin_unlock_irqrestore(&pool->lock, *irq_flags);
590 INIT_LIST_HEAD(&new_pages);
591 r = ttm_alloc_new_pages(&new_pages, pool->gfp_flags, ttm_flags,
593 spin_lock_irqsave(&pool->lock, *irq_flags);
596 list_splice(&new_pages, &pool->list);
598 pool->npages += alloc_size;
600 printk(KERN_ERR TTM_PFX
601 "Failed to fill pool (%p).", pool);
602 /* If we have any pages left put them to the pool. */
603 list_for_each_entry(p, &pool->list, lru) {
606 list_splice(&new_pages, &pool->list);
607 pool->npages += cpages;
611 pool->fill_lock = false;
615 * Cut count nubmer of pages from the pool and put them to return list
617 * @return count of pages still to allocate to fill the request.
619 static unsigned ttm_page_pool_get_pages(struct ttm_page_pool *pool,
620 struct list_head *pages, int ttm_flags,
621 enum ttm_caching_state cstate, unsigned count)
623 unsigned long irq_flags;
627 spin_lock_irqsave(&pool->lock, irq_flags);
628 ttm_page_pool_fill_locked(pool, ttm_flags, cstate, count, &irq_flags);
630 if (count >= pool->npages) {
631 /* take all pages from the pool */
632 list_splice_init(&pool->list, pages);
633 count -= pool->npages;
637 /* find the last pages to include for requested number of pages. Split
638 * pool to begin and halves to reduce search space. */
639 if (count <= pool->npages/2) {
641 list_for_each(p, &pool->list) {
646 i = pool->npages + 1;
647 list_for_each_prev(p, &pool->list) {
652 /* Cut count number of pages from pool */
653 list_cut_position(pages, &pool->list, p);
654 pool->npages -= count;
657 spin_unlock_irqrestore(&pool->lock, irq_flags);
662 * On success pages list will hold count number of correctly
665 int ttm_get_pages(struct list_head *pages, int flags,
666 enum ttm_caching_state cstate, unsigned count)
668 struct ttm_page_pool *pool = ttm_get_pool(flags, cstate);
669 struct page *p = NULL;
673 /* set zero flag for page allocation if required */
674 if (flags & TTM_PAGE_FLAG_ZERO_ALLOC)
675 gfp_flags |= __GFP_ZERO;
677 /* No pool for cached pages */
679 if (flags & TTM_PAGE_FLAG_DMA32)
680 gfp_flags |= GFP_DMA32;
682 gfp_flags |= GFP_HIGHUSER;
684 for (r = 0; r < count; ++r) {
685 p = alloc_page(gfp_flags);
688 printk(KERN_ERR TTM_PFX
689 "Unable to allocate page.");
693 list_add(&p->lru, pages);
699 /* combine zero flag to pool flags */
700 gfp_flags |= pool->gfp_flags;
702 /* First we take pages from the pool */
703 count = ttm_page_pool_get_pages(pool, pages, flags, cstate, count);
705 /* clear the pages coming from the pool if requested */
706 if (flags & TTM_PAGE_FLAG_ZERO_ALLOC) {
707 list_for_each_entry(p, pages, lru) {
708 clear_page(page_address(p));
712 /* If pool didn't have enough pages allocate new one. */
714 /* ttm_alloc_new_pages doesn't reference pool so we can run
715 * multiple requests in parallel.
717 r = ttm_alloc_new_pages(pages, gfp_flags, flags, cstate, count);
719 /* If there is any pages in the list put them back to
721 printk(KERN_ERR TTM_PFX
722 "Failed to allocate extra pages "
723 "for large request.");
724 ttm_put_pages(pages, 0, flags, cstate);
733 /* Put all pages in pages list to correct pool to wait for reuse */
734 void ttm_put_pages(struct list_head *pages, unsigned page_count, int flags,
735 enum ttm_caching_state cstate)
737 unsigned long irq_flags;
738 struct ttm_page_pool *pool = ttm_get_pool(flags, cstate);
739 struct page *p, *tmp;
742 /* No pool for this memory type so free the pages */
744 list_for_each_entry_safe(p, tmp, pages, lru) {
747 /* Make the pages list empty */
748 INIT_LIST_HEAD(pages);
751 if (page_count == 0) {
752 list_for_each_entry_safe(p, tmp, pages, lru) {
757 spin_lock_irqsave(&pool->lock, irq_flags);
758 list_splice_init(pages, &pool->list);
759 pool->npages += page_count;
760 /* Check that we don't go over the pool limit */
762 if (pool->npages > _manager.options.max_size) {
763 page_count = pool->npages - _manager.options.max_size;
764 /* free at least NUM_PAGES_TO_ALLOC number of pages
765 * to reduce calls to set_memory_wb */
766 if (page_count < NUM_PAGES_TO_ALLOC)
767 page_count = NUM_PAGES_TO_ALLOC;
769 spin_unlock_irqrestore(&pool->lock, irq_flags);
771 ttm_page_pool_free(pool, page_count);
774 static void ttm_page_pool_init_locked(struct ttm_page_pool *pool, int flags,
777 spin_lock_init(&pool->lock);
778 pool->fill_lock = false;
779 INIT_LIST_HEAD(&pool->list);
780 pool->npages = pool->nfrees = 0;
781 pool->gfp_flags = flags;
785 int ttm_page_alloc_init(struct ttm_mem_global *glob, unsigned max_pages)
788 if (atomic_add_return(1, &_manager.page_alloc_inited) > 1)
791 printk(KERN_INFO TTM_PFX "Initializing pool allocator.\n");
793 ttm_page_pool_init_locked(&_manager.wc_pool, GFP_HIGHUSER, "wc");
795 ttm_page_pool_init_locked(&_manager.uc_pool, GFP_HIGHUSER, "uc");
797 ttm_page_pool_init_locked(&_manager.wc_pool_dma32, GFP_USER | GFP_DMA32,
800 ttm_page_pool_init_locked(&_manager.uc_pool_dma32, GFP_USER | GFP_DMA32,
803 _manager.options.max_size = max_pages;
804 _manager.options.small = SMALL_ALLOCATION;
805 _manager.options.alloc_size = NUM_PAGES_TO_ALLOC;
807 kobject_init(&_manager.kobj, &ttm_pool_kobj_type);
808 ret = kobject_add(&_manager.kobj, &glob->kobj, "pool");
809 if (unlikely(ret != 0)) {
810 kobject_put(&_manager.kobj);
814 ttm_pool_mm_shrink_init(&_manager);
819 void ttm_page_alloc_fini()
823 if (atomic_sub_return(1, &_manager.page_alloc_inited) > 0)
826 printk(KERN_INFO TTM_PFX "Finalizing pool allocator.\n");
827 ttm_pool_mm_shrink_fini(&_manager);
829 for (i = 0; i < NUM_POOLS; ++i)
830 ttm_page_pool_free(&_manager.pools[i], FREE_ALL_PAGES);
832 kobject_put(&_manager.kobj);
835 int ttm_page_alloc_debugfs(struct seq_file *m, void *data)
837 struct ttm_page_pool *p;
839 char *h[] = {"pool", "refills", "pages freed", "size"};
840 if (atomic_read(&_manager.page_alloc_inited) == 0) {
841 seq_printf(m, "No pool allocator running.\n");
844 seq_printf(m, "%6s %12s %13s %8s\n",
845 h[0], h[1], h[2], h[3]);
846 for (i = 0; i < NUM_POOLS; ++i) {
847 p = &_manager.pools[i];
849 seq_printf(m, "%6s %12ld %13ld %8d\n",
850 p->name, p->nrefills,
851 p->nfrees, p->npages);
855 EXPORT_SYMBOL(ttm_page_alloc_debugfs);