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 serialiazation 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] Setting allocation size to %lu "
169 "is not allowed. Recomended size is "
171 NUM_PAGES_TO_ALLOC
*(PAGE_SIZE
>> 7),
172 NUM_PAGES_TO_ALLOC
*(PAGE_SIZE
>> 10));
174 } else if (val
> NUM_PAGES_TO_ALLOC
) {
175 printk(KERN_WARNING
"[ttm] Setting allocation size to "
176 "larger than %lu is not recomended.\n",
177 NUM_PAGES_TO_ALLOC
*(PAGE_SIZE
>> 10));
179 m
->options
.alloc_size
= val
;
185 static ssize_t
ttm_pool_show(struct kobject
*kobj
,
186 struct attribute
*attr
, char *buffer
)
188 struct ttm_pool_manager
*m
=
189 container_of(kobj
, struct ttm_pool_manager
, kobj
);
192 if (attr
== &ttm_page_pool_max
)
193 val
= m
->options
.max_size
;
194 else if (attr
== &ttm_page_pool_small
)
195 val
= m
->options
.small
;
196 else if (attr
== &ttm_page_pool_alloc_size
)
197 val
= m
->options
.alloc_size
;
199 val
= val
* (PAGE_SIZE
>> 10);
201 return snprintf(buffer
, PAGE_SIZE
, "%u\n", val
);
204 static const struct sysfs_ops ttm_pool_sysfs_ops
= {
205 .show
= &ttm_pool_show
,
206 .store
= &ttm_pool_store
,
209 static struct kobj_type ttm_pool_kobj_type
= {
210 .release
= &ttm_pool_kobj_release
,
211 .sysfs_ops
= &ttm_pool_sysfs_ops
,
212 .default_attrs
= ttm_pool_attrs
,
215 static struct ttm_pool_manager _manager
= {
216 .page_alloc_inited
= ATOMIC_INIT(0)
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 printk(KERN_ERR
"[ttm] Failed to set %d pages to wb!\n",
282 for (i
= 0; i
< npages
; ++i
)
283 __free_page(pages
[i
]);
286 static void ttm_pool_update_free_locked(struct ttm_page_pool
*pool
,
287 unsigned freed_pages
)
289 pool
->npages
-= freed_pages
;
290 pool
->nfrees
+= freed_pages
;
294 * Free pages from pool.
296 * To prevent hogging the ttm_swap process we only free NUM_PAGES_TO_ALLOC
297 * number of pages in one go.
299 * @pool: to free the pages from
300 * @free_all: If set to true will free all pages in pool
302 static int ttm_page_pool_free(struct ttm_page_pool
*pool
, unsigned nr_free
)
304 unsigned long irq_flags
;
306 struct page
**pages_to_free
;
307 unsigned freed_pages
= 0,
308 npages_to_free
= nr_free
;
310 if (NUM_PAGES_TO_ALLOC
< nr_free
)
311 npages_to_free
= NUM_PAGES_TO_ALLOC
;
313 pages_to_free
= kmalloc(npages_to_free
* sizeof(struct page
*),
315 if (!pages_to_free
) {
316 printk(KERN_ERR
"Failed to allocate memory for pool free operation.\n");
321 spin_lock_irqsave(&pool
->lock
, irq_flags
);
323 list_for_each_entry_reverse(p
, &pool
->list
, lru
) {
324 if (freed_pages
>= npages_to_free
)
327 pages_to_free
[freed_pages
++] = p
;
328 /* We can only remove NUM_PAGES_TO_ALLOC at a time. */
329 if (freed_pages
>= NUM_PAGES_TO_ALLOC
) {
330 /* remove range of pages from the pool */
331 __list_del(p
->lru
.prev
, &pool
->list
);
333 ttm_pool_update_free_locked(pool
, freed_pages
);
335 * Because changing page caching is costly
336 * we unlock the pool to prevent stalling.
338 spin_unlock_irqrestore(&pool
->lock
, irq_flags
);
340 ttm_pages_put(pages_to_free
, freed_pages
);
341 if (likely(nr_free
!= FREE_ALL_PAGES
))
342 nr_free
-= freed_pages
;
344 if (NUM_PAGES_TO_ALLOC
>= nr_free
)
345 npages_to_free
= nr_free
;
347 npages_to_free
= NUM_PAGES_TO_ALLOC
;
351 /* free all so restart the processing */
355 /* Not allowed to fall tough or break because
356 * following context is inside spinlock while we are
364 /* remove range of pages from the pool */
366 __list_del(&p
->lru
, &pool
->list
);
368 ttm_pool_update_free_locked(pool
, freed_pages
);
369 nr_free
-= freed_pages
;
372 spin_unlock_irqrestore(&pool
->lock
, irq_flags
);
375 ttm_pages_put(pages_to_free
, freed_pages
);
377 kfree(pages_to_free
);
381 /* Get good estimation how many pages are free in pools */
382 static int ttm_pool_get_num_unused_pages(void)
386 for (i
= 0; i
< NUM_POOLS
; ++i
)
387 total
+= _manager
.pools
[i
].npages
;
393 * Calback for mm to request pool to reduce number of page held.
395 static int ttm_pool_mm_shrink(int shrink_pages
, gfp_t gfp_mask
)
397 static atomic_t start_pool
= ATOMIC_INIT(0);
399 unsigned pool_offset
= atomic_add_return(1, &start_pool
);
400 struct ttm_page_pool
*pool
;
402 pool_offset
= pool_offset
% NUM_POOLS
;
403 /* select start pool in round robin fashion */
404 for (i
= 0; i
< NUM_POOLS
; ++i
) {
405 unsigned nr_free
= shrink_pages
;
406 if (shrink_pages
== 0)
408 pool
= &_manager
.pools
[(i
+ pool_offset
)%NUM_POOLS
];
409 shrink_pages
= ttm_page_pool_free(pool
, nr_free
);
411 /* return estimated number of unused pages in pool */
412 return ttm_pool_get_num_unused_pages();
415 static void ttm_pool_mm_shrink_init(struct ttm_pool_manager
*manager
)
417 manager
->mm_shrink
.shrink
= &ttm_pool_mm_shrink
;
418 manager
->mm_shrink
.seeks
= 1;
419 register_shrinker(&manager
->mm_shrink
);
422 static void ttm_pool_mm_shrink_fini(struct ttm_pool_manager
*manager
)
424 unregister_shrinker(&manager
->mm_shrink
);
427 static int ttm_set_pages_caching(struct page
**pages
,
428 enum ttm_caching_state cstate
, unsigned cpages
)
431 /* Set page caching */
434 r
= set_pages_array_uc(pages
, cpages
);
436 printk(KERN_ERR
"[ttm] Failed to set %d pages to uc!\n",
440 r
= set_pages_array_wc(pages
, cpages
);
442 printk(KERN_ERR
"[ttm] Failed to set %d pages to wc!\n",
452 * Free pages the pages that failed to change the caching state. If there is
453 * any pages that have changed their caching state already put them to the
456 static void ttm_handle_caching_state_failure(struct list_head
*pages
,
457 int ttm_flags
, enum ttm_caching_state cstate
,
458 struct page
**failed_pages
, unsigned cpages
)
461 /* Failed pages has to be reed */
462 for (i
= 0; i
< cpages
; ++i
) {
463 list_del(&failed_pages
[i
]->lru
);
464 __free_page(failed_pages
[i
]);
469 * Allocate new pages with correct caching.
471 * This function is reentrant if caller updates count depending on number of
472 * pages returned in pages array.
474 static int ttm_alloc_new_pages(struct list_head
*pages
, int gfp_flags
,
475 int ttm_flags
, enum ttm_caching_state cstate
, unsigned count
)
477 struct page
**caching_array
;
481 unsigned max_cpages
= min(count
,
482 (unsigned)(PAGE_SIZE
/sizeof(struct page
*)));
484 /* allocate array for page caching change */
485 caching_array
= kmalloc(max_cpages
*sizeof(struct page
*), GFP_KERNEL
);
487 if (!caching_array
) {
488 printk(KERN_ERR
"[ttm] unable to allocate table for new pages.");
492 for (i
= 0, cpages
= 0; i
< count
; ++i
) {
493 p
= alloc_page(gfp_flags
);
496 printk(KERN_ERR
"[ttm] unable to get page %u\n", i
);
498 /* store already allocated pages in the pool after
499 * setting the caching state */
501 r
= ttm_set_pages_caching(caching_array
, cstate
, cpages
);
503 ttm_handle_caching_state_failure(pages
,
505 caching_array
, cpages
);
511 #ifdef CONFIG_HIGHMEM
512 /* gfp flags of highmem page should never be dma32 so we
513 * we should be fine in such case
518 caching_array
[cpages
++] = p
;
519 if (cpages
== max_cpages
) {
521 r
= ttm_set_pages_caching(caching_array
,
524 ttm_handle_caching_state_failure(pages
,
526 caching_array
, cpages
);
533 list_add(&p
->lru
, pages
);
537 r
= ttm_set_pages_caching(caching_array
, cstate
, cpages
);
539 ttm_handle_caching_state_failure(pages
,
541 caching_array
, cpages
);
544 kfree(caching_array
);
550 * Fill the given pool if there isn't enough pages and requested number of
553 static void ttm_page_pool_fill_locked(struct ttm_page_pool
*pool
,
554 int ttm_flags
, enum ttm_caching_state cstate
, unsigned count
,
555 unsigned long *irq_flags
)
561 * Only allow one pool fill operation at a time.
562 * If pool doesn't have enough pages for the allocation new pages are
563 * allocated from outside of pool.
568 pool
->fill_lock
= true;
570 /* If allocation request is small and there is not enough
571 * pages in pool we fill the pool first */
572 if (count
< _manager
.options
.small
573 && count
> pool
->npages
) {
574 struct list_head new_pages
;
575 unsigned alloc_size
= _manager
.options
.alloc_size
;
578 * Can't change page caching if in irqsave context. We have to
579 * drop the pool->lock.
581 spin_unlock_irqrestore(&pool
->lock
, *irq_flags
);
583 INIT_LIST_HEAD(&new_pages
);
584 r
= ttm_alloc_new_pages(&new_pages
, pool
->gfp_flags
, ttm_flags
,
586 spin_lock_irqsave(&pool
->lock
, *irq_flags
);
589 list_splice(&new_pages
, &pool
->list
);
591 pool
->npages
+= alloc_size
;
593 printk(KERN_ERR
"[ttm] Failed to fill pool (%p).", pool
);
594 /* If we have any pages left put them to the pool. */
595 list_for_each_entry(p
, &pool
->list
, lru
) {
598 list_splice(&new_pages
, &pool
->list
);
599 pool
->npages
+= cpages
;
603 pool
->fill_lock
= false;
607 * Cut count nubmer of pages from the pool and put them to return list
609 * @return count of pages still to allocate to fill the request.
611 static unsigned ttm_page_pool_get_pages(struct ttm_page_pool
*pool
,
612 struct list_head
*pages
, int ttm_flags
,
613 enum ttm_caching_state cstate
, unsigned count
)
615 unsigned long irq_flags
;
619 spin_lock_irqsave(&pool
->lock
, irq_flags
);
620 ttm_page_pool_fill_locked(pool
, ttm_flags
, cstate
, count
, &irq_flags
);
622 if (count
>= pool
->npages
) {
623 /* take all pages from the pool */
624 list_splice_init(&pool
->list
, pages
);
625 count
-= pool
->npages
;
629 /* find the last pages to include for requested number of pages. Split
630 * pool to begin and halves to reduce search space. */
631 if (count
<= pool
->npages
/2) {
633 list_for_each(p
, &pool
->list
) {
638 i
= pool
->npages
+ 1;
639 list_for_each_prev(p
, &pool
->list
) {
644 /* Cut count number of pages from pool */
645 list_cut_position(pages
, &pool
->list
, p
);
646 pool
->npages
-= count
;
649 spin_unlock_irqrestore(&pool
->lock
, irq_flags
);
654 * On success pages list will hold count number of correctly
657 int ttm_get_pages(struct list_head
*pages
, int flags
,
658 enum ttm_caching_state cstate
, unsigned count
)
660 struct ttm_page_pool
*pool
= ttm_get_pool(flags
, cstate
);
661 struct page
*p
= NULL
;
665 /* set zero flag for page allocation if required */
666 if (flags
& TTM_PAGE_FLAG_ZERO_ALLOC
)
667 gfp_flags
|= __GFP_ZERO
;
669 /* No pool for cached pages */
671 if (flags
& TTM_PAGE_FLAG_DMA32
)
672 gfp_flags
|= GFP_DMA32
;
674 gfp_flags
|= __GFP_HIGHMEM
;
676 for (r
= 0; r
< count
; ++r
) {
677 p
= alloc_page(gfp_flags
);
680 printk(KERN_ERR
"[ttm] unable to allocate page.");
684 list_add(&p
->lru
, pages
);
690 /* combine zero flag to pool flags */
691 gfp_flags
|= pool
->gfp_flags
;
693 /* First we take pages from the pool */
694 count
= ttm_page_pool_get_pages(pool
, pages
, flags
, cstate
, count
);
696 /* clear the pages coming from the pool if requested */
697 if (flags
& TTM_PAGE_FLAG_ZERO_ALLOC
) {
698 list_for_each_entry(p
, pages
, lru
) {
699 clear_page(page_address(p
));
703 /* If pool didn't have enough pages allocate new one. */
705 /* ttm_alloc_new_pages doesn't reference pool so we can run
706 * multiple requests in parallel.
708 r
= ttm_alloc_new_pages(pages
, gfp_flags
, flags
, cstate
, count
);
710 /* If there is any pages in the list put them back to
712 printk(KERN_ERR
"[ttm] Failed to allocate extra pages "
713 "for large request.");
714 ttm_put_pages(pages
, 0, flags
, cstate
);
723 /* Put all pages in pages list to correct pool to wait for reuse */
724 void ttm_put_pages(struct list_head
*pages
, unsigned page_count
, int flags
,
725 enum ttm_caching_state cstate
)
727 unsigned long irq_flags
;
728 struct ttm_page_pool
*pool
= ttm_get_pool(flags
, cstate
);
729 struct page
*p
, *tmp
;
732 /* No pool for this memory type so free the pages */
734 list_for_each_entry_safe(p
, tmp
, pages
, lru
) {
737 /* Make the pages list empty */
738 INIT_LIST_HEAD(pages
);
741 if (page_count
== 0) {
742 list_for_each_entry_safe(p
, tmp
, pages
, lru
) {
747 spin_lock_irqsave(&pool
->lock
, irq_flags
);
748 list_splice_init(pages
, &pool
->list
);
749 pool
->npages
+= page_count
;
750 /* Check that we don't go over the pool limit */
752 if (pool
->npages
> _manager
.options
.max_size
) {
753 page_count
= pool
->npages
- _manager
.options
.max_size
;
754 /* free at least NUM_PAGES_TO_ALLOC number of pages
755 * to reduce calls to set_memory_wb */
756 if (page_count
< NUM_PAGES_TO_ALLOC
)
757 page_count
= NUM_PAGES_TO_ALLOC
;
759 spin_unlock_irqrestore(&pool
->lock
, irq_flags
);
761 ttm_page_pool_free(pool
, page_count
);
764 static void ttm_page_pool_init_locked(struct ttm_page_pool
*pool
, int flags
,
767 spin_lock_init(&pool
->lock
);
768 pool
->fill_lock
= false;
769 INIT_LIST_HEAD(&pool
->list
);
770 pool
->npages
= pool
->nfrees
= 0;
771 pool
->gfp_flags
= flags
;
775 int ttm_page_alloc_init(struct ttm_mem_global
*glob
, unsigned max_pages
)
778 if (atomic_add_return(1, &_manager
.page_alloc_inited
) > 1)
781 printk(KERN_INFO
"[ttm] Initializing pool allocator.\n");
783 ttm_page_pool_init_locked(&_manager
.wc_pool
, GFP_HIGHUSER
, "wc");
785 ttm_page_pool_init_locked(&_manager
.uc_pool
, GFP_HIGHUSER
, "uc");
787 ttm_page_pool_init_locked(&_manager
.wc_pool_dma32
, GFP_USER
| GFP_DMA32
,
790 ttm_page_pool_init_locked(&_manager
.uc_pool_dma32
, GFP_USER
| GFP_DMA32
,
793 _manager
.options
.max_size
= max_pages
;
794 _manager
.options
.small
= SMALL_ALLOCATION
;
795 _manager
.options
.alloc_size
= NUM_PAGES_TO_ALLOC
;
797 kobject_init(&_manager
.kobj
, &ttm_pool_kobj_type
);
798 ret
= kobject_add(&_manager
.kobj
, &glob
->kobj
, "pool");
799 if (unlikely(ret
!= 0)) {
800 kobject_put(&_manager
.kobj
);
804 ttm_pool_mm_shrink_init(&_manager
);
809 void ttm_page_alloc_fini()
813 if (atomic_sub_return(1, &_manager
.page_alloc_inited
) > 0)
816 printk(KERN_INFO
"[ttm] Finilizing pool allocator.\n");
817 ttm_pool_mm_shrink_fini(&_manager
);
819 for (i
= 0; i
< NUM_POOLS
; ++i
)
820 ttm_page_pool_free(&_manager
.pools
[i
], FREE_ALL_PAGES
);
822 kobject_put(&_manager
.kobj
);
825 int ttm_page_alloc_debugfs(struct seq_file
*m
, void *data
)
827 struct ttm_page_pool
*p
;
829 char *h
[] = {"pool", "refills", "pages freed", "size"};
830 if (atomic_read(&_manager
.page_alloc_inited
) == 0) {
831 seq_printf(m
, "No pool allocator running.\n");
834 seq_printf(m
, "%6s %12s %13s %8s\n",
835 h
[0], h
[1], h
[2], h
[3]);
836 for (i
= 0; i
< NUM_POOLS
; ++i
) {
837 p
= &_manager
.pools
[i
];
839 seq_printf(m
, "%6s %12ld %13ld %8d\n",
840 p
->name
, p
->nrefills
,
841 p
->nfrees
, p
->npages
);
845 EXPORT_SYMBOL(ttm_page_alloc_debugfs
);