ttm_page_alloc.c

/*
 * Copyright (c) Red Hat Inc.

 * Permission is hereby granted, free of charge, to any person obtaining a
 * copy of this software and associated documentation files (the "Software"),
 * to deal in the Software without restriction, including without limitation
 * the rights to use, copy, modify, merge, publish, distribute, sub license,
 * and/or sell copies of the Software, and to permit persons to whom the
 * Software is furnished to do so, subject to the following conditions:
 *
 * The above copyright notice and this permission notice (including the
 * next paragraph) shall be included in all copies or substantial portions
 * of the Software.
 *
 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
 * DEALINGS IN THE SOFTWARE.
 *
 * Authors: Dave Airlie <airlied@redhat.com>
 *          Jerome Glisse <jglisse@redhat.com>
 *          Pauli Nieminen <suokkos@gmail.com>
 */

/* simple list based uncached page pool
 * - Pool collects resently freed pages for reuse
 * - Use page->lru to keep a free list
 * - doesn't track currently in use pages
 */
#include <linux/list.h>
#include <linux/spinlock.h>
#include <linux/highmem.h>
#include <linux/mm_types.h>
#include <linux/mm.h>

#include <asm/atomic.h>
#include <asm/agp.h>

#include "ttm/ttm_bo_driver.h"
#include "ttm/ttm_page_alloc.h"


#define NUM_PAGES_TO_ALLOC		(PAGE_SIZE/sizeof(struct page *))
#define SMALL_ALLOCATION		16
#define FREE_ALL_PAGES			(~0U)
/* times are in msecs */
#define PAGE_FREE_INTERVAL		1000

/**
 * struct ttm_page_pool - Pool to reuse recently allocated uc/wc pages.
 *
 * @lock: Protects the shared pool from concurrnet access. Must be used with
 * irqsave/irqrestore variants because pool allocator maybe called from
 * delayed work.
 * @fill_lock: Prevent concurrent calls to fill.
 * @list: Pool of free uc/wc pages for fast reuse.
 * @gfp_flags: Flags to pass for alloc_page.
 * @npages: Number of pages in pool.
 */
struct ttm_page_pool {
	spinlock_t		lock;
	bool			fill_lock;
	struct list_head	list;
	int			gfp_flags;
	unsigned		npages;
};

struct ttm_pool_opts {
	unsigned	alloc_size;
	unsigned	max_size;
	unsigned	small;
};

#define NUM_POOLS 4

/**
 * struct ttm_pool_manager - Holds memory pools for fst allocation
 *
 * Manager is read only object for pool code so it doesn't need locking.
 *
 * @free_interval: minimum number of jiffies between freeing pages from pool.
 * @page_alloc_inited: reference counting for pool allocation.
 * @work: Work that is used to shrink the pool. Work is only run when there is
 * some pages to free.
 * @small_allocation: Limit in number of pages what is small allocation.
 *
 * @pools: All pool objects in use.
 **/
struct ttm_pool_manager {
	struct shrinker		mm_shrink;
	atomic_t		page_alloc_inited;
	struct ttm_pool_opts	options;

	union {
		struct ttm_page_pool	pools[NUM_POOLS];
		struct {
			struct ttm_page_pool	wc_pool;
			struct ttm_page_pool	uc_pool;
			struct ttm_page_pool	wc_pool_dma32;
			struct ttm_page_pool	uc_pool_dma32;
		} ;
	};
};

static struct ttm_pool_manager _manager = {
	.page_alloc_inited	= ATOMIC_INIT(0)
};

#ifdef CONFIG_X86
/* TODO: add this to x86 like _uc, this version here is inefficient */
static int set_pages_array_wc(struct page **pages, int addrinarray)
{
	int i;

	for (i = 0; i < addrinarray; i++)
		set_memory_wc((unsigned long)page_address(pages[i]), 1);
	return 0;
}
#else
static int set_pages_array_wb(struct page **pages, int addrinarray)
{
#ifdef TTM_HAS_AGP
	int i;

	for (i = 0; i < addrinarray; i++)
		unmap_page_from_agp(pages[i]);
#endif
	return 0;
}

static int set_pages_array_wc(struct page **pages, int addrinarray)
{
#ifdef TTM_HAS_AGP
	int i;

	for (i = 0; i < addrinarray; i++)
		map_page_into_agp(pages[i]);
#endif
	return 0;
}

static int set_pages_array_uc(struct page **pages, int addrinarray)
{
#ifdef TTM_HAS_AGP
	int i;

	for (i = 0; i < addrinarray; i++)
		map_page_into_agp(pages[i]);
#endif
	return 0;
}
#endif

/**
 * Select the right pool or requested caching state and ttm flags. */
static struct ttm_page_pool *ttm_get_pool(int flags,
		enum ttm_caching_state cstate)
{
	int pool_index;

	if (cstate == tt_cached)
		return NULL;

	if (cstate == tt_wc)
		pool_index = 0x0;
	else
		pool_index = 0x1;

	if (flags & TTM_PAGE_FLAG_DMA32)
		pool_index |= 0x2;

	return &_manager.pools[pool_index];
}

/* set memory back to wb and free the pages. */
static void ttm_pages_put(struct page *pages[], unsigned npages)
{
	unsigned i;
	if (set_pages_array_wb(pages, npages))
		printk(KERN_ERR "[ttm] Failed to set %d pages to wb!\n",
				npages);
	for (i = 0; i < npages; ++i)
		__free_page(pages[i]);
}

static void ttm_pool_update_free_locked(struct ttm_page_pool *pool,
		unsigned freed_pages)
{
	pool->npages -= freed_pages;
}

/**
 * Free pages from pool.
 *
 * To prevent hogging the ttm_swap process we only free NUM_PAGES_TO_ALLOC
 * number of pages in one go.
 *
 * @pool: to free the pages from
 * @free_all: If set to true will free all pages in pool
 **/
static int ttm_page_pool_free(struct ttm_page_pool *pool, unsigned nr_free)
{
	unsigned long irq_flags;
	struct page *p;
	struct page **pages_to_free;
	unsigned freed_pages = 0,
		 npages_to_free = nr_free;

	if (NUM_PAGES_TO_ALLOC < nr_free)
		npages_to_free = NUM_PAGES_TO_ALLOC;

	pages_to_free = kmalloc(npages_to_free * sizeof(struct page *),
			GFP_KERNEL);
	if (!pages_to_free) {
		printk(KERN_ERR "Failed to allocate memory for pool free operation.\n");
		return 0;
	}

restart:
	spin_lock_irqsave(&pool->lock, irq_flags);

	list_for_each_entry_reverse(p, &pool->list, lru) {
		if (freed_pages >= npages_to_free)
			break;

		pages_to_free[freed_pages++] = p;
		/* We can only remove NUM_PAGES_TO_ALLOC at a time. */
		if (freed_pages >= NUM_PAGES_TO_ALLOC) {
			/* remove range of pages from the pool */
			__list_del(p->lru.prev, &pool->list);

			ttm_pool_update_free_locked(pool, freed_pages);
			/**
			 * Because changing page caching is costly
			 * we unlock the pool to prevent stalling.
			 */
			spin_unlock_irqrestore(&pool->lock, irq_flags);

			ttm_pages_put(pages_to_free, freed_pages);
			if (likely(nr_free != FREE_ALL_PAGES))
				nr_free -= freed_pages;

			if (NUM_PAGES_TO_ALLOC >= nr_free)
				npages_to_free = nr_free;
			else
				npages_to_free = NUM_PAGES_TO_ALLOC;

			freed_pages = 0;

			/* free all so restart the processing */
			if (nr_free)
				goto restart;

			/* Not allowed to fall tough or break because
			 * following context is inside spinlock while we are
			 * outside here.
			 */
			goto out;

		}
	}


	/* remove range of pages from the pool */
	if (freed_pages) {
		__list_del(&p->lru, &pool->list);

		ttm_pool_update_free_locked(pool, freed_pages);
		nr_free -= freed_pages;
	}

	spin_unlock_irqrestore(&pool->lock, irq_flags);

	if (freed_pages)
		ttm_pages_put(pages_to_free, freed_pages);
out:
	kfree(pages_to_free);
	return nr_free;
}

/* Get good estimation how many pages are free in pools */
static int ttm_pool_get_num_unused_pages(void)
{
	unsigned i;
	int total = 0;
	for (i = 0; i < NUM_POOLS; ++i)
		total += _manager.pools[i].npages;

	return total;
}

/**
 * Calback for mm to request pool to reduce number of page held.
 */
static int ttm_pool_mm_shrink(int shrink_pages, gfp_t gfp_mask)
{
	static atomic_t start_pool = ATOMIC_INIT(0);
	unsigned i;
	unsigned pool_offset = atomic_add_return(1, &start_pool);
	struct ttm_page_pool *pool;

	pool_offset = pool_offset % NUM_POOLS;
	/* select start pool in round robin fashion */
	for (i = 0; i < NUM_POOLS; ++i) {
		unsigned nr_free = shrink_pages;
		if (shrink_pages == 0)
			break;
		pool = &_manager.pools[(i + pool_offset)%NUM_POOLS];
		shrink_pages = ttm_page_pool_free(pool, nr_free);
	}
	/* return estimated number of unused pages in pool */
	return ttm_pool_get_num_unused_pages();
}

static void ttm_pool_mm_shrink_init(struct ttm_pool_manager *manager)
{
	manager->mm_shrink.shrink = &ttm_pool_mm_shrink;
	manager->mm_shrink.seeks = 1;
	register_shrinker(&manager->mm_shrink);
}

static void ttm_pool_mm_shrink_fini(struct ttm_pool_manager *manager)
{
	unregister_shrinker(&manager->mm_shrink);
}

static int ttm_set_pages_caching(struct page **pages,
		enum ttm_caching_state cstate, unsigned cpages)
{
	int r = 0;
	/* Set page caching */
	switch (cstate) {
	case tt_uncached:
		r = set_pages_array_uc(pages, cpages);
		if (r)
			printk(KERN_ERR "[ttm] Failed to set %d pages to uc!\n",
					cpages);
		break;
	case tt_wc:
		r = set_pages_array_wc(pages, cpages);
		if (r)
			printk(KERN_ERR "[ttm] Failed to set %d pages to wc!\n",
					cpages);
		break;
	default:
		break;
	}
	return r;
}

/**
 * Free pages the pages that failed to change the caching state. If there is
 * any pages that have changed their caching state already put them to the
 * pool.
 */
static void ttm_handle_caching_state_failure(struct list_head *pages,
		int ttm_flags, enum ttm_caching_state cstate,
		struct page **failed_pages, unsigned cpages)
{
	unsigned i;
	/* Failed pages has to be reed */
	for (i = 0; i < cpages; ++i) {
		list_del(&failed_pages[i]->lru);
		__free_page(failed_pages[i]);
	}
}

/**
 * Allocate new pages with correct caching.
 *
 * This function is reentrant if caller updates count depending on number of
 * pages returned in pages array.
 */
static int ttm_alloc_new_pages(struct list_head *pages, int gfp_flags,
		int ttm_flags, enum ttm_caching_state cstate, unsigned count)
{
	struct page **caching_array;
	struct page *p;
	int r = 0;
	unsigned i, cpages;
	unsigned max_cpages = min(count,
			(unsigned)(PAGE_SIZE/sizeof(struct page *)));

	/* allocate array for page caching change */
	caching_array = kmalloc(max_cpages*sizeof(struct page *), GFP_KERNEL);

	if (!caching_array) {
		printk(KERN_ERR "[ttm] unable to allocate table for new pages.");
		return -ENOMEM;
	}

	for (i = 0, cpages = 0; i < count; ++i) {
		p = alloc_page(gfp_flags);

		if (!p) {
			printk(KERN_ERR "[ttm] unable to get page %u\n", i);

			/* store already allocated pages in the pool after
			 * setting the caching state */
			if (cpages) {
				r = ttm_set_pages_caching(caching_array, cstate, cpages);
				if (r)
					ttm_handle_caching_state_failure(pages,
						ttm_flags, cstate,
						caching_array, cpages);
			}
			r = -ENOMEM;
			goto out;
		}

#ifdef CONFIG_HIGHMEM
		/* gfp flags of highmem page should never be dma32 so we
		 * we should be fine in such case
		 */
		if (!PageHighMem(p))
#endif
		{
			caching_array[cpages++] = p;
			if (cpages == max_cpages) {

				r = ttm_set_pages_caching(caching_array,
						cstate, cpages);
				if (r) {
					ttm_handle_caching_state_failure(pages,
						ttm_flags, cstate,
						caching_array, cpages);
					goto out;
				}
				cpages = 0;
			}
		}

		list_add(&p->lru, pages);
	}

	if (cpages) {
		r = ttm_set_pages_caching(caching_array, cstate, cpages);
		if (r)
			ttm_handle_caching_state_failure(pages,
					ttm_flags, cstate,
					caching_array, cpages);
	}
out:
	kfree(caching_array);

	return r;
}

/**
 * Fill the given pool if there isn't enough pages and requested number of
 * pages is small.
 */
static void ttm_page_pool_fill_locked(struct ttm_page_pool *pool,
		int ttm_flags, enum ttm_caching_state cstate, unsigned count,
		unsigned long *irq_flags)
{
	struct page *p;
	int r;
	unsigned cpages = 0;
	/**
	 * Only allow one pool fill operation at a time.
	 * If pool doesn't have enough pages for the allocation new pages are
	 * allocated from outside of pool.
	 */
	if (pool->fill_lock)
		return;

	pool->fill_lock = true;

	/* If allocation request is small and there is not enough
	 * pages in pool we fill the pool first */
	if (count < _manager.options.small
		&& count > pool->npages) {
		struct list_head new_pages;
		unsigned alloc_size = _manager.options.alloc_size;

		/**
		 * Can't change page caching if in irqsave context. We have to
		 * drop the pool->lock.
		 */
		spin_unlock_irqrestore(&pool->lock, *irq_flags);

		INIT_LIST_HEAD(&new_pages);
		r = ttm_alloc_new_pages(&new_pages, pool->gfp_flags, ttm_flags,
				cstate,	alloc_size);
		spin_lock_irqsave(&pool->lock, *irq_flags);

		if (!r) {
			list_splice(&new_pages, &pool->list);
			pool->npages += alloc_size;
		} else {
			printk(KERN_ERR "[ttm] Failed to fill pool (%p).", pool);
			/* If we have any pages left put them to the pool. */
			list_for_each_entry(p, &pool->list, lru) {
				++cpages;
			}
			list_splice(&new_pages, &pool->list);
			pool->npages += cpages;
		}

	}
	pool->fill_lock = false;
}

/**
 * Cut count nubmer of pages from the pool and put them to return list
 *
 * @return count of pages still to allocate to fill the request.
 */
static unsigned ttm_page_pool_get_pages(struct ttm_page_pool *pool,
		struct list_head *pages, int ttm_flags,
		enum ttm_caching_state cstate, unsigned count)
{
	unsigned long irq_flags;
	struct list_head *p;
	unsigned i;

	spin_lock_irqsave(&pool->lock, irq_flags);
	ttm_page_pool_fill_locked(pool, ttm_flags, cstate, count, &irq_flags);

	if (count >= pool->npages) {
		/* take all pages from the pool */
		list_splice_init(&pool->list, pages);
		count -= pool->npages;
		pool->npages = 0;
		goto out;
	}
	/* find the last pages to include for requested number of pages. Split
	 * pool to begin and halves to reduce search space. */
	if (count <= pool->npages/2) {
		i = 0;
		list_for_each(p, &pool->list) {
			if (++i == count)
				break;
		}
	} else {
		i = pool->npages + 1;
		list_for_each_prev(p, &pool->list) {
			if (--i == count)
				break;
		}
	}
	/* Cut count number of pages from pool */
	list_cut_position(pages, &pool->list, p);
	pool->npages -= count;
	count = 0;
out:
	spin_unlock_irqrestore(&pool->lock, irq_flags);
	return count;
}

/*
 * On success pages list will hold count number of correctly
 * cached pages.
 */
int ttm_get_pages(struct list_head *pages, int flags,
		enum ttm_caching_state cstate, unsigned count)
{
	struct ttm_page_pool *pool = ttm_get_pool(flags, cstate);
	struct page *p = NULL;
	int gfp_flags = 0;
	int r;

	/* set zero flag for page allocation if required */
	if (flags & TTM_PAGE_FLAG_ZERO_ALLOC)
		gfp_flags |= __GFP_ZERO;

	/* No pool for cached pages */
	if (pool == NULL) {
		if (flags & TTM_PAGE_FLAG_DMA32)
			gfp_flags |= GFP_DMA32;
		else
			gfp_flags |= __GFP_HIGHMEM;

		for (r = 0; r < count; ++r) {
			p = alloc_page(gfp_flags);
			if (!p) {

				printk(KERN_ERR "[ttm] unable to allocate page.");
				return -ENOMEM;
			}

			list_add(&p->lru, pages);
		}
		return 0;
	}


	/* combine zero flag to pool flags */
	gfp_flags |= pool->gfp_flags;

	/* First we take pages from the pool */
	count = ttm_page_pool_get_pages(pool, pages, flags, cstate, count);

	/* clear the pages coming from the pool if requested */
	if (flags & TTM_PAGE_FLAG_ZERO_ALLOC) {
		list_for_each_entry(p, pages, lru) {
			clear_page(page_address(p));
		}
	}

	/* If pool didn't have enough pages allocate new one. */
	if (count > 0) {
		/* ttm_alloc_new_pages doesn't reference pool so we can run
		 * multiple requests in parallel.
		 **/
		r = ttm_alloc_new_pages(pages, gfp_flags, flags, cstate, count);
		if (r) {
			/* If there is any pages in the list put them back to
			 * the pool. */
			printk(KERN_ERR "[ttm] Failed to allocate extra pages "
					"for large request.");
			ttm_put_pages(pages, 0, flags, cstate);
			return r;
		}
	}


	return 0;
}

/* Put all pages in pages list to correct pool to wait for reuse */
void ttm_put_pages(struct list_head *pages, unsigned page_count, int flags,
		enum ttm_caching_state cstate)
{
	unsigned long irq_flags;
	struct ttm_page_pool *pool = ttm_get_pool(flags, cstate);
	struct page *p, *tmp;

	if (pool == NULL) {
		/* No pool for this memory type so free the pages */

		list_for_each_entry_safe(p, tmp, pages, lru) {
			__free_page(p);
		}
		/* Make the pages list empty */
		INIT_LIST_HEAD(pages);
		return;
	}
	if (page_count == 0) {
		list_for_each_entry_safe(p, tmp, pages, lru) {
			++page_count;
		}
	}

	spin_lock_irqsave(&pool->lock, irq_flags);
	list_splice_init(pages, &pool->list);
	pool->npages += page_count;
	/* Check that we don't go over the pool limit */
	page_count = 0;
	if (pool->npages > _manager.options.max_size) {
		page_count = pool->npages - _manager.options.max_size;
		/* free at least NUM_PAGES_TO_ALLOC number of pages
		 * to reduce calls to set_memory_wb */
		if (page_count < NUM_PAGES_TO_ALLOC)
			page_count = NUM_PAGES_TO_ALLOC;
	}
	spin_unlock_irqrestore(&pool->lock, irq_flags);
	if (page_count)
		ttm_page_pool_free(pool, page_count);
}

static void ttm_page_pool_init_locked(struct ttm_page_pool *pool, int flags)
{
	spin_lock_init(&pool->lock);
	pool->fill_lock = false;
	INIT_LIST_HEAD(&pool->list);
	pool->npages = 0;
	pool->gfp_flags = flags;
}

int ttm_page_alloc_init(unsigned max_pages)
{
	if (atomic_add_return(1, &_manager.page_alloc_inited) > 1)
		return 0;

	printk(KERN_INFO "[ttm] Initializing pool allocator.\n");

	ttm_page_pool_init_locked(&_manager.wc_pool, GFP_HIGHUSER);

	ttm_page_pool_init_locked(&_manager.uc_pool, GFP_HIGHUSER);

	ttm_page_pool_init_locked(&_manager.wc_pool_dma32, GFP_USER | GFP_DMA32);

	ttm_page_pool_init_locked(&_manager.uc_pool_dma32, GFP_USER | GFP_DMA32);

	_manager.options.max_size = max_pages;
	_manager.options.small = SMALL_ALLOCATION;
	_manager.options.alloc_size = NUM_PAGES_TO_ALLOC;

	ttm_pool_mm_shrink_init(&_manager);

	return 0;
}

void ttm_page_alloc_fini()
{
	int i;

	if (atomic_sub_return(1, &_manager.page_alloc_inited) > 0)
		return;

	printk(KERN_INFO "[ttm] Finilizing pool allocator.\n");
	ttm_pool_mm_shrink_fini(&_manager);

	for (i = 0; i < NUM_POOLS; ++i)
		ttm_page_pool_free(&_manager.pools[i], FREE_ALL_PAGES);
}