Re: [PATCH 09/32] ext4: extended attribute value size limit is enforced by vfs

[Darrick J. Wong]

On Thu, Jun 22, 2017 at 06:27:08PM -0400, Theodore Ts'o wrote:
> On Thu, Jun 22, 2017 at 01:40:25PM -0600, Andreas Dilger wrote:
> > 
> > The EXT4_XATTR_MAX_LARGE_EA_SIZE limit of 1MB was also totally arbitrary,
> > but a reasonable upper limit for the atomic get/set interface used by
> > xattrs.  The underlying disk format could actually store xattrs of any size.
> > 
> > I'd think if we want to get huge xattrs that they should be handled by
> > having separate streams (e.g. open file descriptor, ioctl/syscall to select
> > a different stream number on that file) so that the data doesn't have to be
> > completely rewritten any time it is modified, but streams are frowned upon
> > by many Linux developers for security reasons so will probably be a no-go.
> 
> This has been discussed before.  Most of the arguments (both pro and
> con) can be found on this thread (2016):
> 
> 	https://www.spinics.net/lists/linux-nfs/msg57055.html
> 
> Jeremy Allison used to pester me for alternate data stream support
> many years ago, but he has since completely changed his opinion on the
> matter:
> 
> 	http://www.spinics.net/lists/linux-fsdevel/msg96083.html
> 
> So if jra@ doesn't add support for this interface to Samba, the thing
> I'd ask is who is actually going to used such an interface; because if
> there aren't any user space applications lining up to use it, it's
> probably not worth adding it until we can find a flagship user of the
> proposed new API.

Heh, ok, that's about what I expected to hear. :)

> There are plenty of other things we could add that Jeremy has asked
> for, that *would* improve Samba performance.  (For example, adding a
> mount option which provides case-insensitive lookup; the one thing
> that makes this hard is that while Android doesn't seem to care about
> I18N --- they are currently using an ascii-only tolower() in their
> sdcardfs hack --- Samba/Windows does care about I18N done right, so we
> wouldn't have to get a Unicode case folding implementation into the
> kernel.)  So I'd argue that those are probably more interesting
> projects than O_XATTR support.

--D

> 
> 	       	  	       	    - Ted