device removal seems to be very slow (kernel 4.1.15)

Using btrfs progs 4.3.1 on a Vanilla kernel.org 4.1.15 kernel.

time btrfs device delete /dev/xvdh /backups

real    13936m56.796s
user    0m0.000s
sys     1351m48.280s

(which is about 9 days).

Where :

/dev/xvdh was 120gb in size.

/backups is a single / "raid 0" volume that now looks like :

Label: 'BACKUP_BTRFS_SNAPS' uuid: 6ee08c31-f310-4890-8424-b88bb77186ed

	Total devices 3 FS bytes used 301.09GiB
	devid    1 size 100.00GiB used 90.00GiB path /dev/xvdg
	devid    3 size 220.00GiB used 196.06GiB path /dev/xvdi
	devid    4 size 221.00GiB used 59.06GiB path /dev/xvdj

There are about 400 snapshots on it.

thanks
David.

OK, based on the device names, you're running this inside a Xen instance with para-virtualized storage drivers (or Amazon EC2, which is the same thing at it's core), and that will have at least some impact on performance (although it will be less impact than if you were using full virtualization). If you have administrative access to Domain 0, and can afford to have the VM down, I would suggest checking how long the equivalent operation takes from Domain 0 (note that to properly check this, you would need to re-add the device to the FS, re-balance the FS, and then delete the device). If you get similar results in Domain 0 and in the VM, then that rules out virtualization as the bottleneck (for para-virtualized storage backed by physical block devices on the local system (as opposed to files, or networked block devices), you should see at most a 10% performance gain running it in Domain 0 assuming both the VM and Domain 0 have the same number of VCPU's and same amount of RAM).

This may be part of the issue. Assuming that /dev/xvdh was mostly full like /dev/xvdg and /dev/xvdi are now, then that would mean it would take longer to remove from the filesystem, because all the chunks that are partially on the device being removed need to be moved to another device. On top of that, whenever a chunk moves, metadata needs to be updated, which means a lot of updates if you have a lot of shared extents, which I'm assuming is the case based on the number of snapshots.

That's very slow. Last week with a 4.1.12 kernel I just deleted a 3TB SATA 7200rpm device with ~1.5TB used on a RAID10 filesystem (reduced from 6 3TB devices to 5 devices in the process) in approximately 38 hours. This was without virtualisation though but there were some damaged sectors to handle along the way which should have slowed the delete a bit and it had more than 10 times the data to move than your /dev/xvdh.

Note about the damaged sectors :
we use 7 disks for this BTRFS RAID10 arrays but to reduce the risk of having to restore huge backups (see recent discussion about BTRFS RAID10 not protecting against 2-devices failure at all), as soon as numerous damaged sectors appear on a drive we delete it from the RAID10 and add it to a MD RAID1 array which is one of the devices on the BTRFS RAID10 (right now we have 5 devices in the RAID10 one of them being a 3-way md RAID1 with disks having these numerous reallocated sectors) : so the reads from the deleted device had some errors to handle and the writes on the md RAID1 device triggered some sector relocations too. Note that ideally I would replace at least 2 of the disks in the md RAID1 because I know from experience that they will fail in the short future (my estimate is between right now and 6 months at best given the current rate of reallocated sectors) but replacing a working drive with damaged sectors costs us some downtime and a one time fee (unlike a drive which is either unreadable or doesn't pass SMART tests anymore). We can live with both the occasional slowdowns (SATA errors generated when the drives detect new damaged sectors usually block IOs for a handful of seconds) and the minor risk this causes : until now this worked OK for this server, the md RAID1 array acts as a buffer for disks that are slowly dying (and the monthly BTRFS scrub + md raid check helps getting the worst ones up to the point where they fail fast enough to avoid accumulating too much bad drives in this array for long periods of time).

I'm not sure if the number of snapshots can impact the device delete
operation: the slow part of device delete is relocating block groups which (AFAIK) seems to be one level down in the stack and shouldn't even know about snapshots. If however you create or delete snapshots during the delete operation you could probably slow down the delete.

Best regards,

Lionel