nvmap: page pools: replace background allocator with background zeroer

The background allocator rapidly becomes useless once the system
has filled memory with cached pages.  It refuses to allocate when
free memory < 128MB (which it always is, the kernel is aiming to
keep very little free), and freed pages are not returned to the
page pool when zero_memory=1.

Remove the background allocator completely, and instead return
freed memory to the page pool in a separate list to be zeroed
in the background.  This results in a self-balancing pool of
memory available to graphics, and reduces presure on the kernel's
page allocator.  If the pool grows too big it will get reduced
by the shrinker.  If it gets too small, the next allocation will
fall back to the page allocator, and then later return those pages
to the pool.

Before this change incremental page pool hit rate reported by
/d/nvmap/pagepool/page_pool_hits vs. /d/nvmap/pagepool/page_pool_misses
goes to 0% after boot.  After this change it is near 100% for small
app launches and 75% for larger app launches.

Change-Id: I4bc914498d7d0369eef9e621bda110d9b8be90b2
Signed-off-by: Colin Cross <ccross@android.com>
Signed-off-by: Krishna Reddy <vdumpa@nvidia.com>
Reviewed-on: http://git-master/r/664674
GVS: Gerrit_Virtual_Submit
Reviewed-on: http://git-master/r/736428
Reviewed-by: Alex Waterman <alexw@nvidia.com>
Tested-by: Alex Waterman <alexw@nvidia.com>

3 files changed, 79 insertions, 155 deletions

diff --git a/drivers/video/tegra/nvmap/nvmap_handle.c b/drivers/video/tegra/nvmap/nvmap_handle.c
index 59bb6db2b835..813496596a14 100644
--- a/drivers/video/tegra/nvmap/nvmap_handle.c
+++ b/drivers/video/tegra/nvmap/nvmap_handle.c
@@ -122,8 +122,7 @@ void _nvmap_handle_free(struct nvmap_handle *h)
 	for (i = 0; i < nr_page; i++)
 		h->pgalloc.pages[i] = nvmap_to_page(h->pgalloc.pages[i]);
 
-	if (!zero_memory)
-		page_index = nvmap_page_pool_fill_lots(&nvmap_dev->pool,
+	page_index = nvmap_page_pool_fill_lots(&nvmap_dev->pool,
 				h->pgalloc.pages, nr_page);
 
 	for (i = page_index; i < nr_page; i++)
diff --git a/drivers/video/tegra/nvmap/nvmap_pp.c b/drivers/video/tegra/nvmap/nvmap_pp.c
index c00d9f20edd4..599b76a74b6d 100644
--- a/drivers/video/tegra/nvmap/nvmap_pp.c
+++ b/drivers/video/tegra/nvmap/nvmap_pp.c
@@ -34,8 +34,7 @@
 #include "nvmap_priv.h"
 
 #define NVMAP_TEST_PAGE_POOL_SHRINKER     1
-#define PENDING_PAGES_SIZE                128
-#define MIN_AVAILABLE_MB                  128
+#define PENDING_PAGES_SIZE                32
 
 static bool enable_pp = 1;
 static int pool_size;
@@ -43,7 +42,6 @@ static int pool_size;
 static struct task_struct *background_allocator;
 static DECLARE_WAIT_QUEUE_HEAD(nvmap_bg_wait);
 static struct page *pending_pages[PENDING_PAGES_SIZE];
-static atomic_t bg_pages_to_fill;
 
 #ifdef CONFIG_NVMAP_PAGE_POOL_DEBUG
 static inline void __pp_dbg_var_add(u64 *dbg_var, u32 nr)
@@ -62,6 +60,21 @@ static inline void __pp_dbg_var_add(u64 *dbg_var, u32 nr)
 static int __nvmap_page_pool_fill_lots_locked(struct nvmap_page_pool *pool,
 				       struct page **pages, u32 nr);
 
+static inline struct page *get_zero_list_page(struct nvmap_page_pool *pool)
+{
+	struct page *page;
+
+	if (list_empty(&pool->zero_list))
+		return NULL;
+
+	page = list_first_entry(&pool->zero_list, struct page, lru);
+	list_del(&page->lru);
+
+	pool->to_zero--;
+
+	return page;
+}
+
 static inline struct page *get_page_list_page(struct nvmap_page_pool *pool)
 {
 	struct page *page;
@@ -82,81 +95,50 @@ static inline bool nvmap_bg_should_run(struct nvmap_page_pool *pool)
 	bool ret;
 
 	mutex_lock(&pool->lock);
-	ret = (pool->to_zero > 0 || atomic_read(&bg_pages_to_fill));
+	ret = (pool->to_zero > 0);
 	mutex_unlock(&pool->lock);
 
 	return ret;
 }
 
-/*
- * Allocate n pages one by one. Not the most efficient allocation scheme ever;
- * however, it will make it easier later on to handle single or small number of
- * page allocations from the page pool being individually freed.
- */
-static int __nvmap_pp_alloc_n_pages(struct page **pages, int n, gfp_t flags)
+static int nvmap_pp_zero_pages(struct page **pages, int nr)
 {
 	int i;
 
-	for (i = 0; i < n; i++) {
-		pages[i] = alloc_page(flags);
-		if (!pages[i])
-			goto no_mem;
-	}
+	for (i = 0; i < nr; i++)
+		clear_highpage(pages[i]);
 
 	return 0;
-
-no_mem:
-	for (i -= 1; i >= 0; i--)
-		__free_page(pages[i]);
-	return -ENOMEM;
 }
 
-/*
- * Actually do the fill. This requires a few steps:
- *
- *  1. Allocate a bunch of pages.
- *
- *  2. Fill the page pool with the allocated pages. We don't want to hold the
- *     PP lock for too long so this is the only time we hold the PP lock.
- *
- *  3. Rinse and repeat until we have allocated all the pages we think we need
- *     or the page pool is full. Since we are not holding the lock for the
- *     entire fill it is possible that other pages were filled into the pool.
- *
- *  4. Free any left over pages if the pool is filled before we finish.
- */
-static void nvmap_pp_do_background_fill(struct nvmap_page_pool *pool)
+static void nvmap_pp_do_background_zero_pages(struct nvmap_page_pool *pool)
 {
-	int err;
-	u32 pages = 0, nr, i;
-	gfp_t gfp = GFP_NVMAP | __GFP_NOMEMALLOC |
-		    __GFP_NORETRY | __GFP_NO_KSWAPD;
-
-	pages = (u32)atomic_xchg(&bg_pages_to_fill, pages);
-
-	if (!pages || !enable_pp)
-		return;
+	int i;
+	struct page *page;
+	int ret;
 
-	/* If this param is set, force zero page allocation. */
-	if (zero_memory)
-		gfp |= __GFP_ZERO;
+	mutex_lock(&pool->lock);
+	for (i = 0; i < PENDING_PAGES_SIZE; i++) {
+		page = get_zero_list_page(pool);
+		if (page == NULL)
+			break;
+		pending_pages[i] = page;
+	}
+	mutex_unlock(&pool->lock);
 
-	do {
-		nr = min_t(u32, PENDING_PAGES_SIZE, pages);
-		err = __nvmap_pp_alloc_n_pages(pending_pages, nr, gfp);
-		if (err) {
-			pr_info("Failed to alloc %u pages for PP!\n", pages);
-			return;
-		}
+	ret = nvmap_pp_zero_pages(pending_pages, i);
+	if (ret < 0) {
+		ret = 0;
+		goto out;
+	}
 
-		mutex_lock(&pool->lock);
-		i = __nvmap_page_pool_fill_lots_locked(pool, pending_pages, nr);
-		mutex_unlock(&pool->lock);
-		pages -= nr;
-	} while (pages && i == nr);
+	mutex_lock(&pool->lock);
+	ret = __nvmap_page_pool_fill_lots_locked(pool, pending_pages, i);
+	mutex_unlock(&pool->lock);
 
-	for (; i < nr; i++)
-		__free_page(pending_pages[i]);
+out:
+	for (; ret < i; ret++)
+		__free_page(pending_pages[ret]);
 }
 
 /*
@@ -167,18 +149,20 @@ static void nvmap_pp_do_background_fill(struct nvmap_page_pool *pool)
  * not directly seen by userspace. Of course if the page pools are empty user
  * space will suffer.
  */
-static int nvmap_background_zero_allocator(void *arg)
+static int nvmap_background_zero_thread(void *arg)
 {
 	struct nvmap_page_pool *pool = &nvmap_dev->pool;
 	struct sched_param param = { .sched_priority = 0 };
 
-	pr_info("PP alloc thread starting.\n");
+	pr_info("PP zeroing thread starting.\n");
 
 	set_freezable();
 	sched_setscheduler(current, SCHED_IDLE, &param);
 
 	while (!kthread_should_stop()) {
-		nvmap_pp_do_background_fill(pool);
+		while (nvmap_bg_should_run(pool)) {
+			nvmap_pp_do_background_zero_pages(pool);
+		}
 
 		wait_event_freezable(nvmap_bg_wait,
 				nvmap_bg_should_run(pool) ||
@@ -189,44 +173,6 @@ static int nvmap_background_zero_allocator(void *arg)
 }
 
 /*
- * Call this if the background allocator should possibly wake up. This function
- * will check to make sure its actually a good idea for that to happen before
- * waking the allocator up.
- */
-static inline void nvmap_pp_wake_up_allocator(void)
-{
-	struct nvmap_page_pool *pool = &nvmap_dev->pool;
-	struct sysinfo info;
-	int free_pages, tmp;
-
-	if (!enable_pp)
-		return;
-
-	/* Hueristic: if we don't need to prefill explicitly zero'ed memory then
-	 * lots of memory can be placed back in the pools by possible frees.
-	 * Therefor don't fill the pool unless we really need to as we may get
-	 * more memory without needing to alloc pages.
-	 */
-	if (!zero_memory && pool->count > NVMAP_PP_ZERO_MEM_FILL_MIN)
-		return;
-
-	if (pool->max - pool->count < NVMAP_PP_DEF_FILL_THRESH)
-		return;
-
-	si_meminfo(&info);
-	free_pages = (int)info.freeram;
-
-	tmp = free_pages - (MIN_AVAILABLE_MB << (20 - PAGE_SHIFT));
-	if (tmp <= 0)
-		return;
-
-	/* Let the background thread know how much memory to fill. */
-	atomic_set(&bg_pages_to_fill,
-		   min(tmp, (int)(pool->max - pool->count)));
-	wake_up_interruptible(&nvmap_bg_wait);
-}
-
-/*
  * This removes a page from the page pool. If ignore_disable is set, then
  * the enable_pp flag is ignored.
  */
@@ -295,7 +241,6 @@ static int __nvmap_page_pool_alloc_lots_locked(struct nvmap_page_pool *pool,
 	pp_alloc_add(pool, ind);
 	pp_hit_add(pool, ind);
 	pp_miss_add(pool, nr - ind);
-	nvmap_pp_wake_up_allocator();
 
 	return ind;
 }
@@ -313,33 +258,6 @@ int nvmap_page_pool_alloc_lots(struct nvmap_page_pool *pool,
 }
 
 /*
- * This adds a page to the pool. Returns true if the passed page is added.
- * That means if the pool is full this operation will fail.
- */
-static bool nvmap_page_pool_fill_locked(struct nvmap_page_pool *pool,
-					struct page *page)
-{
-	if (!enable_pp)
-		return false;
-
-	if (pool->count >= pool->max)
-		return false;
-
-	/* Sanity check. */
-	if (IS_ENABLED(CONFIG_NVMAP_PAGE_POOL_DEBUG)) {
-		atomic_inc(&page->_count);
-		BUG_ON(atomic_read(&page->_count) != 2);
-		BUG_ON(pool->count > pool->max);
-	}
-
-	list_add_tail(&page->lru, &pool->page_list);
-	pool->count++;
-	pp_fill_add(pool, 1);
-
-	return true;
-}
-
-/*
  * Fill a bunch of pages into the page pool. This will fill as many as it can
  * and return the number of pages filled. Pages are used from the start of the
  * passed page pointer array in a linear fashion.
@@ -379,21 +297,23 @@ int nvmap_page_pool_fill_lots(struct nvmap_page_pool *pool,
 	int ret;
 
 	mutex_lock(&pool->lock);
-	ret = __nvmap_page_pool_fill_lots_locked(pool, pages, nr);
-	mutex_unlock(&pool->lock);
+	if (zero_memory) {
+		int i;
 
-	return ret;
-}
+		nr = min(nr, pool->max - pool->count - pool->to_zero);
 
-bool nvmap_page_pool_fill(struct nvmap_page_pool *pool, struct page *page)
-{
-	bool ret = false;
+		for (i = 0; i < nr; i++) {
+			list_add_tail(&pages[i]->lru, &pool->zero_list);
+			pool->to_zero++;
+		}
 
-	if (pool) {
-		mutex_lock(&pool->lock);
-		ret = nvmap_page_pool_fill_locked(pool, page);
-		mutex_unlock(&pool->lock);
+		wake_up_interruptible(&nvmap_bg_wait);
+
+		ret = i;
+	} else {
+		ret = __nvmap_page_pool_fill_lots_locked(pool, pages, nr);
 	}
+	mutex_unlock(&pool->lock);
 
 	return ret;
 }
@@ -413,7 +333,9 @@ static int nvmap_page_pool_free(struct nvmap_page_pool *pool, int nr_free)
 
 	mutex_lock(&pool->lock);
 	while (i) {
-		page = nvmap_page_pool_alloc_locked(pool, 1);
+		page = get_zero_list_page(pool);
+		if (!page)
+			page = nvmap_page_pool_alloc_locked(pool, 1);
 		if (!page)
 			break;
 		__free_page(page);
@@ -431,7 +353,7 @@ ulong nvmap_page_pool_get_unused_pages(void)
 	if (!nvmap_dev)
 		return 0;
 
-	total = nvmap_dev->pool.count;
+	total = nvmap_dev->pool.count + nvmap_dev->pool.to_zero;
 
 	return total;
 }
@@ -450,14 +372,18 @@ int nvmap_page_pool_clear(void)
 	while ((page = nvmap_page_pool_alloc_locked(pool, 1)) != NULL)
 		__free_page(page);
 
+	while (!list_empty(&pool->zero_list)) {
+		page = get_zero_list_page(pool);
+		__free_page(page);
+	}
+
 	/* For some reason, if an error occured... */
-	if (!list_empty(&pool->page_list)) {
+	if (!list_empty(&pool->page_list) || !list_empty(&pool->zero_list)) {
 		mutex_unlock(&pool->lock);
 		return -ENOMEM;
 	}
 
 	mutex_unlock(&pool->lock);
-	nvmap_pp_wake_up_allocator();
 
 	return 0;
 }
@@ -610,6 +536,9 @@ int nvmap_page_pool_debugfs_init(struct dentry *nvmap_root)
 	debugfs_create_u32("page_pool_available_pages",
 			   S_IRUGO, pp_root,
 			   &nvmap_dev->pool.count);
+	debugfs_create_u32("page_pool_pages_to_zero",
+			   S_IRUGO, pp_root,
+			   &nvmap_dev->pool.to_zero);
 #ifdef CONFIG_NVMAP_PAGE_POOL_DEBUG
 	debugfs_create_u64("page_pool_allocs",
 			   S_IRUGO, pp_root,
@@ -630,7 +559,6 @@ int nvmap_page_pool_debugfs_init(struct dentry *nvmap_root)
 
 int nvmap_page_pool_init(struct nvmap_device *dev)
 {
-	static int reg = 1;
 	unsigned long totalram_mb;
 	struct sysinfo info;
 	struct nvmap_page_pool *pool = &dev->pool;
@@ -638,6 +566,7 @@ int nvmap_page_pool_init(struct nvmap_device *dev)
 	memset(pool, 0x0, sizeof(*pool));
 	mutex_init(&pool->lock);
 	INIT_LIST_HEAD(&pool->page_list);
+	INIT_LIST_HEAD(&pool->zero_list);
 
 	si_meminfo(&info);
 	totalram_mb = (info.totalram * info.mem_unit) >> 20;
@@ -656,16 +585,13 @@ int nvmap_page_pool_init(struct nvmap_device *dev)
 	pr_info("nvmap page pool size: %u pages (%u MB)\n", pool->max,
 		(pool->max * info.mem_unit) >> 20);
 
-	if (reg) {
-		reg = 0;
-		register_shrinker(&nvmap_page_pool_shrinker);
-	}
-
-	background_allocator = kthread_create(nvmap_background_zero_allocator,
+	background_allocator = kthread_run(nvmap_background_zero_thread,
 					    NULL, "nvmap-bz");
 	if (IS_ERR_OR_NULL(background_allocator))
 		goto fail;
 
+	register_shrinker(&nvmap_page_pool_shrinker);
+
 	return 0;
 fail:
 	nvmap_page_pool_fini(dev);
diff --git a/drivers/video/tegra/nvmap/nvmap_priv.h b/drivers/video/tegra/nvmap/nvmap_priv.h
index 8b715b42faac..0eeaee4d7e69 100644
--- a/drivers/video/tegra/nvmap/nvmap_priv.h
+++ b/drivers/video/tegra/nvmap/nvmap_priv.h
@@ -188,7 +188,6 @@ struct nvmap_page_pool {
 int nvmap_page_pool_init(struct nvmap_device *dev);
 int nvmap_page_pool_fini(struct nvmap_device *dev);
 struct page *nvmap_page_pool_alloc(struct nvmap_page_pool *pool);
-bool nvmap_page_pool_fill(struct nvmap_page_pool *pool, struct page *page);
 int nvmap_page_pool_alloc_lots(struct nvmap_page_pool *pool,
 					struct page **pages, u32 nr);
 int nvmap_page_pool_fill_lots(struct nvmap_page_pool *pool,