Re: quotas: failure on removing a file via SFTP/SSH

[ST <smntov@gmail.com>]

On Tue, 2017-11-21 at 11:33 -0700, Chris Murphy wrote:
> On Tue, Nov 21, 2017 at 8:29 AM, ST <smntov@gmail.com> wrote:
> >> >>> I'm trying to use quotas for a simple chrooted sftp setup, limiting
> >> >>> space for each user's subvolume (now for testing to 1M).
> >> >>>
> >> >>> I tried to hit the limit by uploading files and once it comes to the
> >> >>> limit I face following problem: if I try to free space by removing a
> >> >>> file via Linux sftp client (or Filezilla) - I get error:
> >> >>> "Couldn't delete file: Failure"
> >> >>>
> >> >>> Sometimes, but not always, if I repeat it for 3-5 times it does removes
> >> >>> the file at the end.
> >> >>> If I login as root and try to remove the file via SSH I get the error:
> >> >>> "rm: cannot remove 'example.txt': Disk quota exceeded"
> >> >>>
> >> >>> What is the problem? And how can I solve it?
> >> >>
> >> >> Kernel version first.
> >> >>
> >> >> If it's possible, please use latest kernel, at least newer than v4.10,
> >> >> since we have a lot of qgroup reservation related fixes in newer kernel.
> >> >>
> >> >> Then, for small quota, due to the nature of btrfs metadata CoW and
> >> >> relative large default node size (16K), it's quite easy to hit disk
> >> >> quota for metadata.
> >> >
> >> > Yes, but why I get the error specifically on REMOVING a file? Even if I
> >> > hit disk quota - if I free up space - it should be possible, isn't it?
> >>
> >> It's only true for fs modifying its metadata in-place (and use journal
> >> to protect it).
> >>
> >> For fs using metadata CoW, even freeing space needs extra space for new
> >> metadata.
> >>
> >
> > Wait, it doesn't sound like a bug, but rather like a flaw in design.
> > This means - each time a user hits his quota limit he will get stuck
> > without being able to free space?!!
> 
> It's a good question if quotas can make it possible for a user to get
> wedged into a situation that will require an admin to temporarily
> raise the quota in order to make file deletion possible.

Why question? It's a fact. That's what I face right now.

>  This is not a
> design flaw, all COW file systems *add* data when deleting. The
> challenge is how to teach the quota system to act like a hard limit
> for data writes that clearly bust the quota, versus a soft limit that
> tolerates some extra amount above the quota for the purpose of
> eventually deleting data. That's maybe non-trivial. It's not that it's
> a design flaw. Metadata can contain inline data, so how exactly to you
> tell what kinds of writes are permitted (deleting a file) and what
> kind of writes are not (append data to a file, or create new file)?
> 
> But for sure the user space tools should prevent setting too low a
> quota limit. If the limit cannot be reasonably expected to work, it
> should be disallowed. So maybe the user space tools need to enforce a
> minimum quota, something like 100MiB, or whatever.
> 

Would you like to open an issue with your enhancement suggestions on the
bug tracker so this case doesn't get forgotten?

Thank you!