Skip to contentSkip to conversation

Fwd [[email protected]: Re: [PATCH] btrfs: zoned: handle RAID profiles in btrfs_can_activate_zone()]

DISCUSSION Johannes Thumshirn 1 participant 1 message Aug 25 1 message not in the archive Btrfs file system

lore
Share & tools
lore
#1 lore raw
In reply to a message not in the archive
Quoted 10 lines · David Sterba

Review comments don't sound unreasonable:

- the switch allows to continue for multi-device filesystems

----- Forwarded message from [email protected] -----

Thank you for your contribution! Sashiko AI review found 2 potential issue(s) to consider:

New issues:
- [High] btrfs_can_activate_zone() incorrectly returns true for multi-device RAID profiles when only a single device has available active zones, leading to chunk activation failures, infinite allocation loops, and premature ENOSPC.

This situation must never happen, if this happens, we have a bug elsewhere. If only one zone in a block-group is sequential only, active zone tracking is done and the total number of max_active_zones of the FS should be the min() of all maz of the devices in the FS.

Quoted 3 lines

Pre-existing issues:
- [High] btrfs_zone_activate() permanently leaks reserved_active_zones and temporarily leaks active_zones_left if a multi-stripe block group fails to activate midway through its stripes. --

That one I need to look into.

Quoted 84 lines

--- Patch [1]: [PATCH] btrfs: zoned: handle RAID profiles in btrfs_can_activate_zone() --- commit d42887c3b92e69264b1963962b3b0d4b608afc5e
Author: Johannes Thumshirn <[email protected]>

btrfs: zoned: handle RAID profiles in btrfs_can_activate_zone()

This commit adds support for RAID profiles in btrfs_can_activate_zone(), evaluating multi-device profiles to check if they can be activated based on available active zones.

Quoted 2 lines
fs/btrfs/zoned.c +0 -0

[ ... ]

Quoted 12 lines
fs/btrfs/zoned.c +5 -0
@@ -2691,6 +2691,11 @@ bool btrfs_can_activate_zone(struct btrfs_fs_devices *fs_devices, u64 flags)
 
  		switch (flags & BTRFS_BLOCK_GROUP_PROFILE_MASK) {
  		case 0: /* single */
+		case BTRFS_BLOCK_GROUP_RAID0:
+		case BTRFS_BLOCK_GROUP_RAID1:
+		case BTRFS_BLOCK_GROUP_RAID1C3:
+		case BTRFS_BLOCK_GROUP_RAID1C4:
+		case BTRFS_BLOCK_GROUP_RAID10:
  			ret = (atomic_read(&zinfo->active_zones_left) >= (1 + reserved));
  			break;
  		case BTRFS_BLOCK_GROUP_DUP:
Quoted 5 lines
  			ret = (atomic_read(&zinfo->active_zones_left) >= (2 + reserved));
  			break;
  		}
  		if (ret)
  			break;

[Severity: High]
Does this code incorrectly return true for multi-device RAID profiles when only a single device has available active zones?

Since the loop breaks early after finding just one device that evaluates to true, and RAID profiles span multiple devices (e.g., RAID1 requires 2 devices), a single device with available active zones bypasses setting BTRFS_FS_NEED_ZONE_FINISH.

Will this allow can_allocate_chunk_zoned() to proceed and allocate chunks that cannot actually be activated, potentially creating an infinite chunk allocation loop when find_free_extent() tries to drop and retry?

[Severity: High]
This is a pre-existing issue, but it interacts with the newly added multi-device RAID logic. Does btrfs_zone_activate() permanently leak reserved_active_zones if a multi-stripe block group fails to activate midway through its stripes?

If btrfs_can_activate_zone() allows a multi-device chunk to be allocated but only some devices have active zones available, btrfs_zone_activate() will succeed on the first stripe but fail on a subsequent one:

fs/btrfs/zoned.c:btrfs_zone_activate() {
...

		if (!btrfs_dev_set_active_zone(device, physical)) {
			/* Cannot activate the zone */
			ret = false;
			goto out_unlock;
		}
		if (!is_data)
			zinfo->reserved_active_zones--;

...
}

The error path fails to roll back the decremented reserved_active_zones. Because BLOCK_GROUP_FLAG_ZONE_IS_ACTIVE is never set, do_zone_finish() will silently skip the cleanup:

fs/btrfs/zoned.c:do_zone_finish() {
	spin_lock(&block_group->lock);
	if (!test_bit(BLOCK_GROUP_FLAG_ZONE_IS_ACTIVE, &block_group->runtime_flags)) {
		spin_unlock(&block_group->lock);
		return 0;
	}

...
}

Can this permanent leak of reserved_active_zones eventually prevent metadata block groups from activating and cause filesystem deadlocks under heavy load?