Re: Problem with btrfs send/receive

[Stefan Behrens <sbehrens@giantdisaster.de>]

On Thu, 23 May 2013 13:51:59 +0200, Felix Blanke wrote:
> Hi everyone,
> 
> I was trying the new send/receive feature today but can't make it work.
> 
> These are the commands I was using:
> 
> btrfs subvol snap -r /mnt/data1/@downloads/ /mnt/data1/snapshots/testsnap
> btrfs send /mnt/data1/snapshots/testsnap | btrfs receive
> /mnt/data1/snapshots/testreceive/
> 
> This command never finishes. A 'ls /mnt/data1/snapshots/testreceive/ '
> never finishes too. After killing the send/receive process one can see
> that the target subvolume was created but is empty.
> 
> 
> Sending the snapshot to a file and use this for receiving does work.
> What am I missing?
> 

That's a known design flaw in the btrfs send code.
'btrfs send /subvol | sleep 666'
is an easier way to block the system. But it's interruptable if you
terminate the sleep task in this case.

btrfs send blocks on the pipe while being in the kernel, the call chain
is like this:
btrfs_ioctl_send -> send_subvol -> full_send_tree -> changed_cb ->
send_cmd -> write_buf -> vfs_write -> pipe_write -> pipe_wait

And full_send_tree() has called btrfs_join_transaction() before, thus
the whole file system and afterwards the system is blocked.

You can avoid it if you receive to a different filesystem or if you
redirect the output of btrfs send into a file. In general, the output of
btrfs send must never be blocked or you are lost.