Re: [RFC PATCH] ceph: fix cross quota realms renames with new truncated files

[Jeff Layton]

On Thu, 2020-11-12 at 10:40 +0000, Luis Henriques wrote:
> Jeff Layton <jlayton@xxxxxxxxxx> writes:
> 
> > On Wed, 2020-11-11 at 18:28 +0000, Luis Henriques wrote:
> > > Jeff Layton <jlayton@xxxxxxxxxx> writes:
> > > 
> > > > On Wed, 2020-11-11 at 15:39 +0000, Luis Henriques wrote:
> > > > > When doing a rename across quota realms, there's a corner case that isn't
> > > > > handled correctly.  Here's a testcase:
> > > > > 
> > > > >   mkdir files limit
> > > > >   truncate files/file -s 10G
> > > > >   setfattr limit -n ceph.quota.max_bytes -v 1000000
> > > > >   mv files limit/
> > > > > 
> > > > > The above will succeed because ftruncate(2) won't result in an immediate
> > > > > notification of the MDSs with the new file size, and thus the quota realms
> > > > > stats won't be updated.
> > > > > 
> > > > > This patch forces a sync with the MDS every time there's an ATTR_SIZE that
> > > > > sets a new i_size, even if we have Fx caps.
> > > > > 
> > > > > Cc: stable@xxxxxxxxxxxxxxx
> > > > > Fixes: dffdcd71458e ("ceph: allow rename operation under different quota realms")
> > > > > URL: https://tracker.ceph.com/issues/36593
> > > > > Signed-off-by: Luis Henriques <lhenriques@xxxxxxx>
> > > > > ---
> > > > >  fs/ceph/inode.c | 11 ++---------
> > > > >  1 file changed, 2 insertions(+), 9 deletions(-)
> > > > > 
> > > > > diff --git a/fs/ceph/inode.c b/fs/ceph/inode.c
> > > > > index 526faf4778ce..30e3f240ac96 100644
> > > > > --- a/fs/ceph/inode.c
> > > > > +++ b/fs/ceph/inode.c
> > > > > @@ -2136,15 +2136,8 @@ int __ceph_setattr(struct inode *inode, struct iattr *attr)
> > > > >  	if (ia_valid & ATTR_SIZE) {
> > > > >  		dout("setattr %p size %lld -> %lld\n", inode,
> > > > >  		     inode->i_size, attr->ia_size);
> > > > > -		if ((issued & CEPH_CAP_FILE_EXCL) &&
> > > > > -		    attr->ia_size > inode->i_size) {
> > > > > -			i_size_write(inode, attr->ia_size);
> > > > > -			inode->i_blocks = calc_inode_blocks(attr->ia_size);
> > > > > -			ci->i_reported_size = attr->ia_size;
> > > > > -			dirtied |= CEPH_CAP_FILE_EXCL;
> > > > > -			ia_valid |= ATTR_MTIME;
> > > > > -		} else if ((issued & CEPH_CAP_FILE_SHARED) == 0 ||
> > > > > -			   attr->ia_size != inode->i_size) {
> > > > > +		if ((issued & (CEPH_CAP_FILE_EXCL|CEPH_CAP_FILE_SHARED)) ||
> > > > > +		    (attr->ia_size != inode->i_size)) {
> > > > >  			req->r_args.setattr.size = cpu_to_le64(attr->ia_size);
> > > > >  			req->r_args.setattr.old_size =
> > > > >  				cpu_to_le64(inode->i_size);
> > > > 
> > > > Hmm...this makes truncates more expensive when we have caps. I'd rather
> > > > not do that if we can help it.
> > > 
> > > Yeah, as I mentioned in the tracker, there's indeed a performance impact
> > > with this fix.  That's what made me add the RFC in the subject ;-)
> > > 
> > > > What about instead having the client mimic a fsync when there is a
> > > > rename across quota realms? If we can't tell that reliably then we could
> > > > also just do an effective fsync ahead of any cross-directory rename?
> > > 
> > > Ok, thanks for the suggestion.  That may actually work, although it will
> > > make the rename more expensive of course.  I'll test that tomorrow and
> > > eventually follow-up with a patch.
> > > 
> > 
> > Patrick pointed out to me on IRC that since you're moving the parent
> > directory of the truncated file, flushing the caps on the directory
> > won't really help. You'd need to walk the entire subtree and try to
> > flush every dirty inode, or basically do a syncfs() prior to renaming
> > the directory across quotarealms.
> > 
> > I think we probably will need to revert the change to allow cross-
> > quotarealm renames of directories and make those return EXDEV again.
> > Anything else sounds like it's probably going to be too expensive.
> 
> Hmm... that sounds a bit drastic and it would make the kernel client
> behave differently from the fuse client -- from what I could understand
> the fuse client does the sync ATTR_SIZE and thus doesn't have this issue.
> 

True. I'll note that the fuse client is not exactly built for speed,
however.

> Obviously, I agree with you that the performance penalty is too high for
> such a common operation.  But maybe renames across quotarealms aren't that
> common and paying the penalty of doing a full ceph_flush_dirty_caps() is
> acceptable for such cases?
> 

I wouldn't even do that. If someone is renaming a directory across
quotarealms, just return EXDEV. Saying "sorry, you have to copy/unlink"
in this situation seems like it should be acceptable. Are you aware of
any specific use-cases where people are renaming large directories
across quotarealms?
-- 
Jeff Layton <jlayton@xxxxxxxxxx>