Btrfs: don't take the chunk_mutex/dev_list mutex in statfs V2

Our gluster boxes get several thousand statfs() calls per second, which begins to suck hardcore with all of the lock contention on the chunk mutex and dev list mutex. We don't really need to hold these things, if we have transient weirdness with statfs() because of the chunk allocator we don't care, so remove this locking.

We still need the dev_list lock if you mount with -o alloc_start however, which is a good argument for nuking that thing from orbit, but that's a patch for another day. Thanks,

Signed-off-by Josef Bacik <[email protected]>

---
V1->V2: make sure ->alloc_start is set before doing the dev extent lookup logic.

 fs/btrfs/super.c | 72 ++++++++++++++++++++++++++++++++++++--------------------
 1 file changed, 47 insertions(+), 25 deletions(-)

fs/btrfs/super.c +47 -25

index 54bd91e..dc337d1 100644
@@ -1644,8 +1644,20 @@ static int btrfs_calc_avail_data_space(struct btrfs_root *root, u64 *free_bytes)
 	int i = 0, nr_devices;
 	int ret;
 
+	/*
+	 * We aren't under the device list lock, so this is racey-ish, but good
+	 * enough for our purposes.
+	 */
 	nr_devices = fs_info->fs_devices->open_devices;
-	BUG_ON(!nr_devices);
+	if (!nr_devices) {
+		smp_mb();
+		nr_devices = fs_info->fs_devices->open_devices;
+		ASSERT(nr_devices);
+		if (!nr_devices) {
+			*free_bytes = 0;
+			return 0;
+		}
+	}
 
 	devices_info = kmalloc_array(nr_devices, sizeof(*devices_info),
 			       GFP_NOFS);

@@ -1670,11 +1682,17 @@ static int btrfs_calc_avail_data_space(struct btrfs_root *root, u64 *free_bytes)
 	else
 		min_stripe_size = BTRFS_STRIPE_LEN;
 
-	list_for_each_entry(device, &fs_devices->devices, dev_list) {
+	if (fs_info->alloc_start)
+		mutex_lock(&fs_devices->device_list_mutex);
+	rcu_read_lock();
+	list_for_each_entry_rcu(device, &fs_devices->devices, dev_list) {
 		if (!device->in_fs_metadata || !device->bdev ||
 		    device->is_tgtdev_for_dev_replace)
 			continue;
 
+		if (i >= nr_devices)
+			break;
+
 		avail_space = device->total_bytes - device->bytes_used;
 
 		/* align with stripe_len */

@@ -1689,24 +1707,32 @@ static int btrfs_calc_avail_data_space(struct btrfs_root *root, u64 *free_bytes)
 		skip_space = 1024 * 1024;
 
 		/* user can set the offset in fs_info->alloc_start. */
-		if (fs_info->alloc_start + BTRFS_STRIPE_LEN <=
-		    device->total_bytes)
+		if (fs_info->alloc_start &&
+		    fs_info->alloc_start + BTRFS_STRIPE_LEN <=
+		    device->total_bytes) {
+			rcu_read_unlock();
 			skip_space = max(fs_info->alloc_start, skip_space);
 
-		/*
-		 * btrfs can not use the free space in [0, skip_space - 1],
-		 * we must subtract it from the total. In order to implement
-		 * it, we account the used space in this range first.
-		 */
-		ret = btrfs_account_dev_extents_size(device, 0, skip_space - 1,
-						     &used_space);
-		if (ret) {
-			kfree(devices_info);
-			return ret;
-		}
+			/*
+			 * btrfs can not use the free space in
+			 * [0, skip_space - 1], we must subtract it from the
+			 * total. In order to implement it, we account the used
+			 * space in this range first.
+			 */
+			ret = btrfs_account_dev_extents_size(device, 0,
+							     skip_space - 1,
+							     &used_space);
+			if (ret) {
+				kfree(devices_info);
+				mutex_unlock(&fs_devices->device_list_mutex);
+				return ret;
+			}
 
-		/* calc the free space in [0, skip_space - 1] */
-		skip_space -= used_space;
+			rcu_read_lock();
+
+			/* calc the free space in [0, skip_space - 1] */
+			skip_space -= used_space;
+		}
 
 		/*
 		 * we can use the free space in [0, skip_space - 1], subtract

@@ -1725,6 +1751,9 @@ static int btrfs_calc_avail_data_space(struct btrfs_root *root, u64 *free_bytes)
 
 		i++;
 	}
+	rcu_read_unlock();
+	if (fs_info->alloc_start)
+		mutex_unlock(&fs_devices->device_list_mutex);
 
 	nr_devices = i;
 

@@ -1787,8 +1816,6 @@ static int btrfs_statfs(struct dentry *dentry, struct kstatfs *buf)
 	 * holding chunk_muext to avoid allocating new chunks, holding
 	 * device_list_mutex to avoid the device being removed
 	 */
-	mutex_lock(&fs_info->fs_devices->device_list_mutex);
-	mutex_lock(&fs_info->chunk_mutex);
 	rcu_read_lock();
 	list_for_each_entry_rcu(found, head, list) {
 		if (found->flags & BTRFS_BLOCK_GROUP_DATA) {

@@ -1826,15 +1853,10 @@ static int btrfs_statfs(struct dentry *dentry, struct kstatfs *buf)
 
 	buf->f_bavail = total_free_data;
 	ret = btrfs_calc_avail_data_space(fs_info->tree_root, &total_free_data);
-	if (ret) {
-		mutex_unlock(&fs_info->chunk_mutex);
-		mutex_unlock(&fs_info->fs_devices->device_list_mutex);
+	if (ret)
 		return ret;
-	}
 	buf->f_bavail += div_u64(total_free_data, factor);
 	buf->f_bavail = buf->f_bavail >> bits;
-	mutex_unlock(&fs_info->chunk_mutex);
-	mutex_unlock(&fs_info->fs_devices->device_list_mutex);
 
 	buf->f_type = BTRFS_SUPER_MAGIC;
 	buf->f_bsize = dentry->d_sb->s_blocksize;

Quoted 6 lines · Josef Bacik

Our gluster boxes get several thousand statfs() calls per second, which begins to suck hardcore with all of the lock contention on the chunk mutex and dev list mutex. We don't really need to hold these things, if we have transient weirdness with statfs() because of the chunk allocator we don't care, so remove this locking.

Quoted 7 lines

We still need the dev_list lock if you mount with -o alloc_start however, which is a good argument for nuking that thing from orbit, but that's a patch for another day. Thanks,

Signed-off-by Josef Bacik <[email protected]>

---
V1->V2: make sure ->alloc_start is set before doing the dev extent lookup logic.

I am strange that why we need dev_list_lock if we mount with -o alloc_start. AFAIK. ->alloc_start is protected by chunk_mutex.

But I think we needn't care that someone changes ->alloc_start, in other words, we needn't take chunk_mutex during the whole process, the following case can be tolerated by the users, I think.

	Task1						Task2
	statfs
	  mutex_lock(&fs_info->chunk_mutex);
	  tmp = fs_info->alloc_start;
	  mutex_unlock(&fs_info->chunk_mutex);
	  btrfs_calc_avail_data_space(fs_info, tmp)
	    ...
							mount -o remount,alloc_start=xxxx
	    ...

Yeah df if you change alloc_start in the middle of it is going to be slightly racey, which is fine since I'm going to kill alloc_start soon anyway.