btrfs: generic readeahead interface

Add state information for readahead to btrfs_fs_info and btrfs_device

Signed-off-by Arne Jansen <[email protected]>

---
 fs/btrfs/ctree.h   |    4 ++++
 fs/btrfs/disk-io.c |    4 ++++
 fs/btrfs/volumes.c |    8 ++++++++
 fs/btrfs/volumes.h |    8 ++++++++
 4 files changed, 24 insertions(+), 0 deletions(-)

fs/btrfs/ctree.h +4 -0

index 2e61fe1..4a33e30 100644
@@ -1079,6 +1079,10 @@ struct btrfs_fs_info {
 
 	/* filesystem state */
 	u64 fs_state;
+
+	/* readahead tree */
+	spinlock_t reada_lock;
+	struct radix_tree_root reada_tree;
 };
 
 /*

fs/btrfs/disk-io.c +4 -0

index 7753eb9..3d4f9c5 100644
@@ -1803,6 +1803,10 @@ struct btrfs_root *open_ctree(struct super_block *sb,
 	fs_info->max_inline = 8192 * 1024;
 	fs_info->metadata_ratio = 0;
 
+	/* readahead state */
+	INIT_RADIX_TREE(&fs_info->reada_tree, GFP_NOFS);
+	spin_lock_init(&fs_info->reada_lock);
+
 	fs_info->thread_pool_size = min_t(unsigned long,
 					  num_online_cpus() + 2, 8);
 

fs/btrfs/volumes.c +8 -0

index 8b9fb8c..800e670 100644
@@ -396,6 +396,14 @@ static noinline int device_list_add(const char *path,
 		}
 		INIT_LIST_HEAD(&device->dev_alloc_list);
 
+		/* init readahead state */
+		spin_lock_init(&device->reada_lock);
+		device->reada_curr_zone = NULL;
+		atomic_set(&device->reada_in_flight, 0);
+		device->reada_next = 0;
+		INIT_RADIX_TREE(&device->reada_zones, GFP_NOFS);
+		INIT_RADIX_TREE(&device->reada_extents, GFP_NOFS);
+
 		mutex_lock(&fs_devices->device_list_mutex);
 		list_add(&device->dev_list, &fs_devices->devices);
 		mutex_unlock(&fs_devices->device_list_mutex);

fs/btrfs/volumes.h +0 -0

index cc2eada..33acd4e 100644

fs/btrfs/volumes.h +2 -0

+	struct radix_tree_root reada_zones;
+	struct radix_tree_root reada_extents;
 };
 
 struct btrfs_fs_devices {

Quoted 66 lines · Arne Jansen

Add state information for readahead to btrfs_fs_info and btrfs_device

Signed-off-by Arne Jansen <[email protected]>

---
 fs/btrfs/ctree.h   |    4 ++++
 fs/btrfs/disk-io.c |    4 ++++
 fs/btrfs/volumes.c |    8 ++++++++
 fs/btrfs/volumes.h |    8 ++++++++
 4 files changed, 24 insertions(+), 0 deletions(-)

fs/btrfs/ctree.h +4 -0

index 2e61fe1..4a33e30 100644
@@ -1079,6 +1079,10 @@ struct btrfs_fs_info {
 
 	/* filesystem state */
 	u64 fs_state;
+
+	/* readahead tree */
+	spinlock_t reada_lock;
+	struct radix_tree_root reada_tree;
 };
 
 /*

fs/btrfs/disk-io.c +4 -0

index 7753eb9..3d4f9c5 100644
@@ -1803,6 +1803,10 @@ struct btrfs_root *open_ctree(struct super_block *sb,
 	fs_info->max_inline = 8192 * 1024;
 	fs_info->metadata_ratio = 0;
 
+	/* readahead state */
+	INIT_RADIX_TREE(&fs_info->reada_tree, GFP_NOFS);
+	spin_lock_init(&fs_info->reada_lock);
+
 	fs_info->thread_pool_size = min_t(unsigned long,
 					  num_online_cpus() + 2, 8);
 

fs/btrfs/volumes.c +8 -0

index 8b9fb8c..800e670 100644
@@ -396,6 +396,14 @@ static noinline int device_list_add(const char *path,
 		}
 		INIT_LIST_HEAD(&device->dev_alloc_list);
 
+		/* init readahead state */
+		spin_lock_init(&device->reada_lock);
+		device->reada_curr_zone = NULL;
+		atomic_set(&device->reada_in_flight, 0);
+		device->reada_next = 0;
+		INIT_RADIX_TREE(&device->reada_zones, GFP_NOFS);
+		INIT_RADIX_TREE(&device->reada_extents, GFP_NOFS);
+
 		mutex_lock(&fs_devices->device_list_mutex);
 		list_add(&device->dev_list, &fs_devices->devices);
 		mutex_unlock(&fs_devices->device_list_mutex);

fs/btrfs/volumes.h +0 -0

index cc2eada..33acd4e 100644
@@ -86,6 +86,14 @@ struct btrfs_device {
 	u8 uuid[BTRFS_UUID_SIZE];
 

fs/btrfs/volumes.h struct btrfs_device { · line 94

Quoted 7 lines

fs/btrfs/volumes.h +6 -0

 	struct btrfs_work work;
+
+	/* readahead state */
+	spinlock_t reada_lock;
+	atomic_t reada_in_flight;
+	u64 reada_next;
+	struct reada_zone *reada_curr_zone;

struct reada_zone has not been defined yet...

thanks,
liubo

Quoted 5 lines

fs/btrfs/volumes.h +2 -0

+	struct radix_tree_root reada_zones;
+	struct radix_tree_root reada_extents;
 };
 
 struct btrfs_fs_devices {

fs/btrfs/reada.c new file +52 -0

+static struct reada_zone *reada_find_zone(struct btrfs_fs_info *fs_info,
+					  struct btrfs_device *dev, u64 logical,
+					  struct btrfs_multi_bio *multi)
+{
+	int ret;
+	int looped = 0;
+	struct reada_zone *zone;
+	struct btrfs_block_group_cache *cache = NULL;
+	u64 start;
+	u64 end;
+	int i;
+
+again:
+	zone = NULL;
+	spin_lock(&fs_info->reada_lock);
+	ret = radix_tree_gang_lookup(&dev->reada_zones, (void **)&zone,
+				     logical >> PAGE_CACHE_SHIFT, 1);
+	if (ret == 1)
+		kref_get(&zone->refcnt);
+	spin_unlock(&fs_info->reada_lock);
+
+	if (ret == 1) {
+		if (logical >= zone->start && logical < zone->end)
+			return zone;
+		spin_lock(&fs_info->reada_lock);
+		reada_zone_put(zone);
+		spin_unlock(&fs_info->reada_lock);
+	}
+
+	if (looped)
+		return NULL;
+
+	cache = btrfs_lookup_block_group(fs_info, logical);
+	if (!cache)
+		return NULL;
+
+	start = cache->key.objectid;
+	end = start + cache->key.offset - 1;
+	btrfs_put_block_group(cache);
+
+	zone = kzalloc(sizeof(*zone), GFP_NOFS);
+	if (!zone)
+		return NULL;
+
+	zone->start = start;
+	zone->end = end;
+	INIT_LIST_HEAD(&zone->list);
+	spin_lock_init(&zone->lock);
+	zone->locked = 0;
+	kref_init(&zone->refcnt);
+	zone->elems = 0;
+	zone->device = dev; /* our device always sits at index 0 */

fs/btrfs/reada.c new file +10 -0

+	spin_unlock(&fs_info->reada_lock);
+
+	if (ret) {
+		kfree(zone);
+		looped = 1;
+		goto again;
+	}
+
+	return zone;
+}

Hi,

Quoted 52 lines · Arne Jansen

fs/btrfs/reada.c new file +52 -0

+static struct reada_zone *reada_find_zone(struct btrfs_fs_info *fs_info,
+					  struct btrfs_device *dev, u64 logical,
+					  struct btrfs_multi_bio *multi)
+{
+	int ret;
+	int looped = 0;
+	struct reada_zone *zone;
+	struct btrfs_block_group_cache *cache = NULL;
+	u64 start;
+	u64 end;
+	int i;
+
+again:
+	zone = NULL;
+	spin_lock(&fs_info->reada_lock);
+	ret = radix_tree_gang_lookup(&dev->reada_zones, (void **)&zone,
+				     logical >> PAGE_CACHE_SHIFT, 1);
+	if (ret == 1)
+		kref_get(&zone->refcnt);
+	spin_unlock(&fs_info->reada_lock);
+
+	if (ret == 1) {
+		if (logical >= zone->start && logical < zone->end)
+			return zone;
+		spin_lock(&fs_info->reada_lock);
+		reada_zone_put(zone);
+		spin_unlock(&fs_info->reada_lock);
+	}
+
+	if (looped)
+		return NULL;
+
+	cache = btrfs_lookup_block_group(fs_info, logical);
+	if (!cache)
+		return NULL;
+
+	start = cache->key.objectid;
+	end = start + cache->key.offset - 1;
+	btrfs_put_block_group(cache);
+
+	zone = kzalloc(sizeof(*zone), GFP_NOFS);
+	if (!zone)
+		return NULL;
+
+	zone->start = start;
+	zone->end = end;
+	INIT_LIST_HEAD(&zone->list);
+	spin_lock_init(&zone->lock);
+	zone->locked = 0;
+	kref_init(&zone->refcnt);
+	zone->elems = 0;
+	zone->device = dev; /* our device always sits at index 0 */

fs/btrfs/reada.c line 323

Quoted 10 lines

fs/btrfs/reada.c new file +10 -0

+	for (i = 0; i < multi->num_stripes; ++i) {
+		/* bounds have already been checked */
+		zone->devs[i] = multi->stripes[i].dev;
+	}
+	zone->ndevs = multi->num_stripes;
+
+	spin_lock(&fs_info->reada_lock);
+	ret = radix_tree_insert(&dev->reada_zones,
+				(unsigned long)zone->end >> PAGE_CACHE_SHIFT,
+				zone);

this can sleep inside a spinlock, you initialize the radix tree with GFP_NOFS, which allows __GFP_WAIT.

Options:
1) use GFP_ATOMIC in radix tree init flags
2) do the radix_tree_preload/radix_tree_preload_end, GFP_NOFS outside of the locked section is ok but __GFP_WAIT has to be masked out (else radix tree insert will not use the preloaded node)
3) unmask __GFP_WAIT from radix tree init flags

I'd go for 3, as the atomic context is not required, and is easier than 2 to implement.

fs/btrfs/reada.c line 333

Quoted 11 lines

fs/btrfs/reada.c new file +10 -0

+	spin_unlock(&fs_info->reada_lock);
+
+	if (ret) {
+		kfree(zone);
+		looped = 1;
+		goto again;
+	}
+
+	return zone;
+}

--
To unsubscribe from this list: send the line "unsubscribe linux-btrfs" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html

Quoted 76 lines · David Sterba

Hi,

Quoted 52 lines · Arne Jansen

fs/btrfs/reada.c new file +52 -0

+static struct reada_zone *reada_find_zone(struct btrfs_fs_info *fs_info,
+					  struct btrfs_device *dev, u64 logical,
+					  struct btrfs_multi_bio *multi)
+{
+	int ret;
+	int looped = 0;
+	struct reada_zone *zone;
+	struct btrfs_block_group_cache *cache = NULL;
+	u64 start;
+	u64 end;
+	int i;
+
+again:
+	zone = NULL;
+	spin_lock(&fs_info->reada_lock);
+	ret = radix_tree_gang_lookup(&dev->reada_zones, (void **)&zone,
+				     logical >> PAGE_CACHE_SHIFT, 1);
+	if (ret == 1)
+		kref_get(&zone->refcnt);
+	spin_unlock(&fs_info->reada_lock);
+
+	if (ret == 1) {
+		if (logical >= zone->start && logical < zone->end)
+			return zone;
+		spin_lock(&fs_info->reada_lock);
+		reada_zone_put(zone);
+		spin_unlock(&fs_info->reada_lock);
+	}
+
+	if (looped)
+		return NULL;
+
+	cache = btrfs_lookup_block_group(fs_info, logical);
+	if (!cache)
+		return NULL;
+
+	start = cache->key.objectid;
+	end = start + cache->key.offset - 1;
+	btrfs_put_block_group(cache);
+
+	zone = kzalloc(sizeof(*zone), GFP_NOFS);
+	if (!zone)
+		return NULL;
+
+	zone->start = start;
+	zone->end = end;
+	INIT_LIST_HEAD(&zone->list);
+	spin_lock_init(&zone->lock);
+	zone->locked = 0;
+	kref_init(&zone->refcnt);
+	zone->elems = 0;
+	zone->device = dev; /* our device always sits at index 0 */

fs/btrfs/reada.c line 323

Quoted 10 lines

fs/btrfs/reada.c new file +10 -0

+	for (i = 0; i < multi->num_stripes; ++i) {
+		/* bounds have already been checked */
+		zone->devs[i] = multi->stripes[i].dev;
+	}
+	zone->ndevs = multi->num_stripes;
+
+	spin_lock(&fs_info->reada_lock);
+	ret = radix_tree_insert(&dev->reada_zones,
+				(unsigned long)zone->end >> PAGE_CACHE_SHIFT,
+				zone);

this can sleep inside a spinlock, you initialize the radix tree with GFP_NOFS, which allows __GFP_WAIT.

Options:
1) use GFP_ATOMIC in radix tree init flags
2) do the radix_tree_preload/radix_tree_preload_end, GFP_NOFS outside of the locked section is ok but __GFP_WAIT has to be masked out (else radix tree insert will not use the preloaded node)
3) unmask __GFP_WAIT from radix tree init flags

Quoted 2 lines

I'd go for 3, as the atomic context is not required, and is easier than 2 to implement.

I like the second one, because it is the general way to fix this problem.

BTW: I think we can use RCU to protect the radix tree on the read side. Arne, how do you think about?

Thanks
Miao

Quoted 16 lines

fs/btrfs/reada.c line 333

Quoted 11 lines

fs/btrfs/reada.c new file +10 -0

+	spin_unlock(&fs_info->reada_lock);
+
+	if (ret) {
+		kfree(zone);
+		looped = 1;
+		goto again;
+	}
+
+	return zone;
+}

--
To unsubscribe from this list: send the line "unsubscribe linux-btrfs" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html