Re: [PATCH 0/4] btrfs: offline dedupe v2

[Gabriel de Perthuis <g2p.code@gmail.com>]

Le 11/06/2013 23:04, Mark Fasheh a écrit :
> On Tue, Jun 11, 2013 at 10:56:59PM +0200, Gabriel de Perthuis wrote:
>>> What I found however is that neither of these is a great idea ;)
>>>
>>> - We want to require that the inode be open for writing so that an
>>>   unprivileged user can't do things like run dedupe on a performance
>>>   sensitive file that they might only have read access to.  In addition I
>>>   could see it as kind of a surprise (non-standard behavior) to an
>>>   administrator that users could alter the layout of files they are only
>>>   allowed to read.
>>>
>>> - Readonly snapshots won't let you open for write anyway (unsuprisingly,
>>>   open() returns -EROFS).  So that kind of kills the idea of them being able
>>>   to open those files for write which we want to dedupe.
>>>
>>> That said, I still think being able to run this against a set of readonly
>>> snapshots makes sense especially if those snapshots are taken for backup
>>> purposes. I'm just not sure how we can sanely enable it.
>>
>> The check could be: if (fmode_write || cap_sys_admin).
>>
>> This isn't incompatible with mnt_want_write, that check is at the
>> level of the superblocks and vfsmount and not the subvolume fsid.
> 
> Oh ok that's certainly better. I think we still have a problem though - how
> does a process gets write access to a file from a ro-snapshot? If I open a
> file (as root) on a ro-snapshot on my test machine here I'll get -EROFS.

Your first series did work in that case.
The process does get a read-only fd, but that's no obstacle for the ioctl.

> I'm a bit confused - how does mnt_want_write factor in here? I think that's
> for a totally seperate kind of accounting, right?

It doesn't, it's just that I had spent a few minutes checking anyway.