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>
44 #include "ttm/ttm_bo_driver.h"
45 #include "ttm/ttm_page_alloc.h"
51 #define NUM_PAGES_TO_ALLOC (PAGE_SIZE/sizeof(struct page *))
52 #define SMALL_ALLOCATION 16
53 #define FREE_ALL_PAGES (~0U)
54 /* times are in msecs */
55 #define PAGE_FREE_INTERVAL 1000
58 * struct ttm_page_pool - Pool to reuse recently allocated uc/wc pages.
60 * @lock: Protects the shared pool from concurrnet access. Must be used with
61 * irqsave/irqrestore variants because pool allocator maybe called from
63 * @fill_lock: Prevent concurrent calls to fill.
64 * @list: Pool of free uc/wc pages for fast reuse.
65 * @gfp_flags: Flags to pass for alloc_page.
66 * @npages: Number of pages in pool.
68 struct ttm_page_pool {
71 struct list_head list;
76 unsigned long nrefills;
80 * Limits for the pool. They are handled without locks because only place where
81 * they may change is in sysfs store. They won't have immediate effect anyway
82 * so forcing serialization to access them is pointless.
85 struct ttm_pool_opts {
94 * struct ttm_pool_manager - Holds memory pools for fst allocation
96 * Manager is read only object for pool code so it doesn't need locking.
98 * @free_interval: minimum number of jiffies between freeing pages from pool.
99 * @page_alloc_inited: reference counting for pool allocation.
100 * @work: Work that is used to shrink the pool. Work is only run when there is
101 * some pages to free.
102 * @small_allocation: Limit in number of pages what is small allocation.
104 * @pools: All pool objects in use.
106 struct ttm_pool_manager {
108 struct shrinker mm_shrink;
109 struct ttm_pool_opts options;
112 struct ttm_page_pool pools[NUM_POOLS];
114 struct ttm_page_pool wc_pool;
115 struct ttm_page_pool uc_pool;
116 struct ttm_page_pool wc_pool_dma32;
117 struct ttm_page_pool uc_pool_dma32;
122 static struct attribute ttm_page_pool_max = {
123 .name = "pool_max_size",
124 .mode = S_IRUGO | S_IWUSR
126 static struct attribute ttm_page_pool_small = {
127 .name = "pool_small_allocation",
128 .mode = S_IRUGO | S_IWUSR
130 static struct attribute ttm_page_pool_alloc_size = {
131 .name = "pool_allocation_size",
132 .mode = S_IRUGO | S_IWUSR
135 static struct attribute *ttm_pool_attrs[] = {
137 &ttm_page_pool_small,
138 &ttm_page_pool_alloc_size,
142 static void ttm_pool_kobj_release(struct kobject *kobj)
144 struct ttm_pool_manager *m =
145 container_of(kobj, struct ttm_pool_manager, kobj);
149 static ssize_t ttm_pool_store(struct kobject *kobj,
150 struct attribute *attr, const char *buffer, size_t size)
152 struct ttm_pool_manager *m =
153 container_of(kobj, struct ttm_pool_manager, kobj);
156 chars = sscanf(buffer, "%u", &val);
160 /* Convert kb to number of pages */
161 val = val / (PAGE_SIZE >> 10);
163 if (attr == &ttm_page_pool_max)
164 m->options.max_size = val;
165 else if (attr == &ttm_page_pool_small)
166 m->options.small = val;
167 else if (attr == &ttm_page_pool_alloc_size) {
168 if (val > NUM_PAGES_TO_ALLOC*8) {
169 printk(KERN_ERR TTM_PFX
170 "Setting allocation size to %lu "
171 "is not allowed. Recommended size is "
173 NUM_PAGES_TO_ALLOC*(PAGE_SIZE >> 7),
174 NUM_PAGES_TO_ALLOC*(PAGE_SIZE >> 10));
176 } else if (val > NUM_PAGES_TO_ALLOC) {
177 printk(KERN_WARNING TTM_PFX
178 "Setting allocation size to "
179 "larger than %lu is not recommended.\n",
180 NUM_PAGES_TO_ALLOC*(PAGE_SIZE >> 10));
182 m->options.alloc_size = val;
188 static ssize_t ttm_pool_show(struct kobject *kobj,
189 struct attribute *attr, char *buffer)
191 struct ttm_pool_manager *m =
192 container_of(kobj, struct ttm_pool_manager, kobj);
195 if (attr == &ttm_page_pool_max)
196 val = m->options.max_size;
197 else if (attr == &ttm_page_pool_small)
198 val = m->options.small;
199 else if (attr == &ttm_page_pool_alloc_size)
200 val = m->options.alloc_size;
202 val = val * (PAGE_SIZE >> 10);
204 return snprintf(buffer, PAGE_SIZE, "%u\n", val);
207 static const struct sysfs_ops ttm_pool_sysfs_ops = {
208 .show = &ttm_pool_show,
209 .store = &ttm_pool_store,
212 static struct kobj_type ttm_pool_kobj_type = {
213 .release = &ttm_pool_kobj_release,
214 .sysfs_ops = &ttm_pool_sysfs_ops,
215 .default_attrs = ttm_pool_attrs,
218 static struct ttm_pool_manager *_manager;
221 static int set_pages_array_wb(struct page **pages, int addrinarray)
226 for (i = 0; i < addrinarray; i++)
227 unmap_page_from_agp(pages[i]);
232 static int set_pages_array_wc(struct page **pages, int addrinarray)
237 for (i = 0; i < addrinarray; i++)
238 map_page_into_agp(pages[i]);
243 static int set_pages_array_uc(struct page **pages, int addrinarray)
248 for (i = 0; i < addrinarray; i++)
249 map_page_into_agp(pages[i]);
256 * Select the right pool or requested caching state and ttm flags. */
257 static struct ttm_page_pool *ttm_get_pool(int flags,
258 enum ttm_caching_state cstate)
262 if (cstate == tt_cached)
270 if (flags & TTM_PAGE_FLAG_DMA32)
273 return &_manager->pools[pool_index];
276 /* set memory back to wb and free the pages. */
277 static void ttm_pages_put(struct page *pages[], unsigned npages)
280 if (set_pages_array_wb(pages, npages))
281 printk(KERN_ERR TTM_PFX "Failed to set %d pages to wb!\n",
283 for (i = 0; i < npages; ++i)
284 __free_page(pages[i]);
287 static void ttm_pool_update_free_locked(struct ttm_page_pool *pool,
288 unsigned freed_pages)
290 pool->npages -= freed_pages;
291 pool->nfrees += freed_pages;
295 * Free pages from pool.
297 * To prevent hogging the ttm_swap process we only free NUM_PAGES_TO_ALLOC
298 * number of pages in one go.
300 * @pool: to free the pages from
301 * @free_all: If set to true will free all pages in pool
303 static int ttm_page_pool_free(struct ttm_page_pool *pool, unsigned nr_free)
305 unsigned long irq_flags;
307 struct page **pages_to_free;
308 unsigned freed_pages = 0,
309 npages_to_free = nr_free;
311 if (NUM_PAGES_TO_ALLOC < nr_free)
312 npages_to_free = NUM_PAGES_TO_ALLOC;
314 pages_to_free = kmalloc(npages_to_free * sizeof(struct page *),
316 if (!pages_to_free) {
317 printk(KERN_ERR TTM_PFX
318 "Failed to allocate memory for pool free operation.\n");
323 spin_lock_irqsave(&pool->lock, irq_flags);
325 list_for_each_entry_reverse(p, &pool->list, lru) {
326 if (freed_pages >= npages_to_free)
329 pages_to_free[freed_pages++] = p;
330 /* We can only remove NUM_PAGES_TO_ALLOC at a time. */
331 if (freed_pages >= NUM_PAGES_TO_ALLOC) {
332 /* remove range of pages from the pool */
333 __list_del(p->lru.prev, &pool->list);
335 ttm_pool_update_free_locked(pool, freed_pages);
337 * Because changing page caching is costly
338 * we unlock the pool to prevent stalling.
340 spin_unlock_irqrestore(&pool->lock, irq_flags);
342 ttm_pages_put(pages_to_free, freed_pages);
343 if (likely(nr_free != FREE_ALL_PAGES))
344 nr_free -= freed_pages;
346 if (NUM_PAGES_TO_ALLOC >= nr_free)
347 npages_to_free = nr_free;
349 npages_to_free = NUM_PAGES_TO_ALLOC;
353 /* free all so restart the processing */
357 /* Not allowed to fall tough or break because
358 * following context is inside spinlock while we are
366 /* remove range of pages from the pool */
368 __list_del(&p->lru, &pool->list);
370 ttm_pool_update_free_locked(pool, freed_pages);
371 nr_free -= freed_pages;
374 spin_unlock_irqrestore(&pool->lock, irq_flags);
377 ttm_pages_put(pages_to_free, freed_pages);
379 kfree(pages_to_free);
383 /* Get good estimation how many pages are free in pools */
384 static int ttm_pool_get_num_unused_pages(void)
388 for (i = 0; i < NUM_POOLS; ++i)
389 total += _manager->pools[i].npages;
395 * Callback for mm to request pool to reduce number of page held.
397 static int ttm_pool_mm_shrink(struct shrinker *shrink, int shrink_pages, gfp_t gfp_mask)
399 static atomic_t start_pool = ATOMIC_INIT(0);
401 unsigned pool_offset = atomic_add_return(1, &start_pool);
402 struct ttm_page_pool *pool;
404 pool_offset = pool_offset % NUM_POOLS;
405 /* select start pool in round robin fashion */
406 for (i = 0; i < NUM_POOLS; ++i) {
407 unsigned nr_free = shrink_pages;
408 if (shrink_pages == 0)
410 pool = &_manager->pools[(i + pool_offset)%NUM_POOLS];
411 shrink_pages = ttm_page_pool_free(pool, nr_free);
413 /* return estimated number of unused pages in pool */
414 return ttm_pool_get_num_unused_pages();
417 static void ttm_pool_mm_shrink_init(struct ttm_pool_manager *manager)
419 manager->mm_shrink.shrink = &ttm_pool_mm_shrink;
420 manager->mm_shrink.seeks = 1;
421 register_shrinker(&manager->mm_shrink);
424 static void ttm_pool_mm_shrink_fini(struct ttm_pool_manager *manager)
426 unregister_shrinker(&manager->mm_shrink);
429 static int ttm_set_pages_caching(struct page **pages,
430 enum ttm_caching_state cstate, unsigned cpages)
433 /* Set page caching */
436 r = set_pages_array_uc(pages, cpages);
438 printk(KERN_ERR TTM_PFX
439 "Failed to set %d pages to uc!\n",
443 r = set_pages_array_wc(pages, cpages);
445 printk(KERN_ERR TTM_PFX
446 "Failed to set %d pages to wc!\n",
456 * Free pages the pages that failed to change the caching state. If there is
457 * any pages that have changed their caching state already put them to the
460 static void ttm_handle_caching_state_failure(struct list_head *pages,
461 int ttm_flags, enum ttm_caching_state cstate,
462 struct page **failed_pages, unsigned cpages)
465 /* Failed pages have to be freed */
466 for (i = 0; i < cpages; ++i) {
467 list_del(&failed_pages[i]->lru);
468 __free_page(failed_pages[i]);
473 * Allocate new pages with correct caching.
475 * This function is reentrant if caller updates count depending on number of
476 * pages returned in pages array.
478 static int ttm_alloc_new_pages(struct list_head *pages, int gfp_flags,
479 int ttm_flags, enum ttm_caching_state cstate, unsigned count)
481 struct page **caching_array;
485 unsigned max_cpages = min(count,
486 (unsigned)(PAGE_SIZE/sizeof(struct page *)));
488 /* allocate array for page caching change */
489 caching_array = kmalloc(max_cpages*sizeof(struct page *), GFP_KERNEL);
491 if (!caching_array) {
492 printk(KERN_ERR TTM_PFX
493 "Unable to allocate table for new pages.");
497 for (i = 0, cpages = 0; i < count; ++i) {
498 p = alloc_page(gfp_flags);
501 printk(KERN_ERR TTM_PFX "Unable to get page %u.\n", i);
503 /* store already allocated pages in the pool after
504 * setting the caching state */
506 r = ttm_set_pages_caching(caching_array,
509 ttm_handle_caching_state_failure(pages,
511 caching_array, cpages);
517 #ifdef CONFIG_HIGHMEM
518 /* gfp flags of highmem page should never be dma32 so we
519 * we should be fine in such case
524 caching_array[cpages++] = p;
525 if (cpages == max_cpages) {
527 r = ttm_set_pages_caching(caching_array,
530 ttm_handle_caching_state_failure(pages,
532 caching_array, cpages);
539 list_add(&p->lru, pages);
543 r = ttm_set_pages_caching(caching_array, cstate, cpages);
545 ttm_handle_caching_state_failure(pages,
547 caching_array, cpages);
550 kfree(caching_array);
556 * Fill the given pool if there isn't enough pages and requested number of
559 static void ttm_page_pool_fill_locked(struct ttm_page_pool *pool,
560 int ttm_flags, enum ttm_caching_state cstate, unsigned count,
561 unsigned long *irq_flags)
567 * Only allow one pool fill operation at a time.
568 * If pool doesn't have enough pages for the allocation new pages are
569 * allocated from outside of pool.
574 pool->fill_lock = true;
576 /* If allocation request is small and there is not enough
577 * pages in pool we fill the pool first */
578 if (count < _manager->options.small
579 && count > pool->npages) {
580 struct list_head new_pages;
581 unsigned alloc_size = _manager->options.alloc_size;
584 * Can't change page caching if in irqsave context. We have to
585 * drop the pool->lock.
587 spin_unlock_irqrestore(&pool->lock, *irq_flags);
589 INIT_LIST_HEAD(&new_pages);
590 r = ttm_alloc_new_pages(&new_pages, pool->gfp_flags, ttm_flags,
592 spin_lock_irqsave(&pool->lock, *irq_flags);
595 list_splice(&new_pages, &pool->list);
597 pool->npages += alloc_size;
599 printk(KERN_ERR TTM_PFX
600 "Failed to fill pool (%p).", pool);
601 /* If we have any pages left put them to the pool. */
602 list_for_each_entry(p, &pool->list, lru) {
605 list_splice(&new_pages, &pool->list);
606 pool->npages += cpages;
610 pool->fill_lock = false;
614 * Cut count nubmer of pages from the pool and put them to return list
616 * @return count of pages still to allocate to fill the request.
618 static unsigned ttm_page_pool_get_pages(struct ttm_page_pool *pool,
619 struct list_head *pages, int ttm_flags,
620 enum ttm_caching_state cstate, unsigned count)
622 unsigned long irq_flags;
626 spin_lock_irqsave(&pool->lock, irq_flags);
627 ttm_page_pool_fill_locked(pool, ttm_flags, cstate, count, &irq_flags);
629 if (count >= pool->npages) {
630 /* take all pages from the pool */
631 list_splice_init(&pool->list, pages);
632 count -= pool->npages;
636 /* find the last pages to include for requested number of pages. Split
637 * pool to begin and halves to reduce search space. */
638 if (count <= pool->npages/2) {
640 list_for_each(p, &pool->list) {
645 i = pool->npages + 1;
646 list_for_each_prev(p, &pool->list) {
651 /* Cut count number of pages from pool */
652 list_cut_position(pages, &pool->list, p);
653 pool->npages -= count;
656 spin_unlock_irqrestore(&pool->lock, irq_flags);
661 * On success pages list will hold count number of correctly
664 int ttm_get_pages(struct list_head *pages, int flags,
665 enum ttm_caching_state cstate, unsigned count)
667 struct ttm_page_pool *pool = ttm_get_pool(flags, cstate);
668 struct page *p = NULL;
669 int gfp_flags = GFP_USER;
672 /* set zero flag for page allocation if required */
673 if (flags & TTM_PAGE_FLAG_ZERO_ALLOC)
674 gfp_flags |= __GFP_ZERO;
676 /* No pool for cached pages */
678 if (flags & TTM_PAGE_FLAG_DMA32)
679 gfp_flags |= GFP_DMA32;
681 gfp_flags |= GFP_HIGHUSER;
683 for (r = 0; r < count; ++r) {
684 p = alloc_page(gfp_flags);
687 printk(KERN_ERR TTM_PFX
688 "Unable to allocate page.");
692 list_add(&p->lru, pages);
698 /* combine zero flag to pool flags */
699 gfp_flags |= pool->gfp_flags;
701 /* First we take pages from the pool */
702 count = ttm_page_pool_get_pages(pool, pages, flags, cstate, count);
704 /* clear the pages coming from the pool if requested */
705 if (flags & TTM_PAGE_FLAG_ZERO_ALLOC) {
706 list_for_each_entry(p, pages, lru) {
707 clear_page(page_address(p));
711 /* If pool didn't have enough pages allocate new one. */
713 /* ttm_alloc_new_pages doesn't reference pool so we can run
714 * multiple requests in parallel.
716 r = ttm_alloc_new_pages(pages, gfp_flags, flags, cstate, count);
718 /* If there is any pages in the list put them back to
720 printk(KERN_ERR TTM_PFX
721 "Failed to allocate extra pages "
722 "for large request.");
723 ttm_put_pages(pages, 0, flags, cstate);
732 /* Put all pages in pages list to correct pool to wait for reuse */
733 void ttm_put_pages(struct list_head *pages, unsigned page_count, int flags,
734 enum ttm_caching_state cstate)
736 unsigned long irq_flags;
737 struct ttm_page_pool *pool = ttm_get_pool(flags, cstate);
738 struct page *p, *tmp;
741 /* No pool for this memory type so free the pages */
743 list_for_each_entry_safe(p, tmp, pages, lru) {
746 /* Make the pages list empty */
747 INIT_LIST_HEAD(pages);
750 if (page_count == 0) {
751 list_for_each_entry_safe(p, tmp, pages, lru) {
756 spin_lock_irqsave(&pool->lock, irq_flags);
757 list_splice_init(pages, &pool->list);
758 pool->npages += page_count;
759 /* Check that we don't go over the pool limit */
761 if (pool->npages > _manager->options.max_size) {
762 page_count = pool->npages - _manager->options.max_size;
763 /* free at least NUM_PAGES_TO_ALLOC number of pages
764 * to reduce calls to set_memory_wb */
765 if (page_count < NUM_PAGES_TO_ALLOC)
766 page_count = NUM_PAGES_TO_ALLOC;
768 spin_unlock_irqrestore(&pool->lock, irq_flags);
770 ttm_page_pool_free(pool, page_count);
773 static void ttm_page_pool_init_locked(struct ttm_page_pool *pool, int flags,
776 spin_lock_init(&pool->lock);
777 pool->fill_lock = false;
778 INIT_LIST_HEAD(&pool->list);
779 pool->npages = pool->nfrees = 0;
780 pool->gfp_flags = flags;
784 int ttm_page_alloc_init(struct ttm_mem_global *glob, unsigned max_pages)
790 printk(KERN_INFO TTM_PFX "Initializing pool allocator.\n");
792 _manager = kzalloc(sizeof(*_manager), GFP_KERNEL);
794 ttm_page_pool_init_locked(&_manager->wc_pool, GFP_HIGHUSER, "wc");
796 ttm_page_pool_init_locked(&_manager->uc_pool, GFP_HIGHUSER, "uc");
798 ttm_page_pool_init_locked(&_manager->wc_pool_dma32,
799 GFP_USER | GFP_DMA32, "wc dma");
801 ttm_page_pool_init_locked(&_manager->uc_pool_dma32,
802 GFP_USER | GFP_DMA32, "uc dma");
804 _manager->options.max_size = max_pages;
805 _manager->options.small = SMALL_ALLOCATION;
806 _manager->options.alloc_size = NUM_PAGES_TO_ALLOC;
808 ret = kobject_init_and_add(&_manager->kobj, &ttm_pool_kobj_type,
809 &glob->kobj, "pool");
810 if (unlikely(ret != 0)) {
811 kobject_put(&_manager->kobj);
816 ttm_pool_mm_shrink_init(_manager);
821 void ttm_page_alloc_fini()
825 printk(KERN_INFO TTM_PFX "Finalizing pool allocator.\n");
826 ttm_pool_mm_shrink_fini(_manager);
828 for (i = 0; i < NUM_POOLS; ++i)
829 ttm_page_pool_free(&_manager->pools[i], FREE_ALL_PAGES);
831 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"};
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);