Re: [PATCH 2/2] src/attr_replace_test: dynamically adjust the max xattr size

["Luís Henriques" <lhenriques@suse.de>]

On Thu, Jun 09, 2022 at 07:59:50AM +1000, Dave Chinner wrote:
> On Wed, Jun 08, 2022 at 10:57:22AM +0100, Luís Henriques wrote:
> > On Wed, Jun 08, 2022 at 10:23:15AM +1000, Dave Chinner wrote:
> > > On Tue, Jun 07, 2022 at 04:15:13PM +0100, Luís Henriques wrote:
> > > > CephFS doesn't had a maximum xattr size.  Instead, it imposes a maximum
> > > > size for the full set of an inode's xattrs names+values, which by default
> > > > is 64K but it can be changed by a cluster admin.
> > > > 
> > > > Test generic/486 started to fail after fixing a ceph bug where this limit
> > > > wasn't being imposed.  Adjust dynamically the size of the xattr being set
> > > > if the error returned is -ENOSPC.
> > > 
> > > Ah, this shouldn't be getting anywhere near the 64kB limit unless
> > > ceph is telling userspace it's block size is > 64kB:
> > > 
> > > size = sbuf.st_blksize * 3 / 4;
> > > .....
> > > size = MIN(size, XATTR_SIZE_MAX);
> > 
> > Yep, that's exactly what is happening.  The cephfs kernel client reports
> > here the value that is being used for ceph "object size", which defaults
> > to 4M.  Hence, we'll set size to XATTR_SIZE_MAX.
> 
> Yikes. This is known to break random applications that size buffers
> based on a multiple of sbuf.st_blksize and assume that it is going
> to be roughly 4kB. e.g. size a buffer at 1024 * sbuf.st_blksize,
> expecting to get a ~4MB buffer, and instead it tries to allocate
> a 4GB buffer....
> 
> > > Regardless, the correct thing to do here is pass the max supported
> > > xattr size from the command line (because fstests knows what that it
> > > for each filesystem type) rather than hard coding
> > > XATTR_SIZE_MAX in the test.
> > 
> > OK, makes sense.  But then, for the ceph case, it becomes messy because we
> > also need to know the attribute name to compute the maximum size.  I guess
> > we'll need an extra argument for that too.
> 
> Just pass in a size for ceph that has enough spare space for the
> attribute names in it, like for g/020. Don't make it more
> complex than it needs to be.

Well, in that case it's just easier for attr_replace_test.c to simply set
the ceiling to (XATTR_SIZE_MAX - strlen(name+value)).  This is because the
fstests don't really know anymore the max xattr size for each filesystem
type; that knowledge is local to generic/020.  The other option is to move
that code (back) to common/attr.

Cheers,
--
Luís