Re: [PATCH] nfsd: Make creates return EEXIST correctly instead of EPERM

[Jeff Layton]

On Fri, 2016-07-08 at 11:14 -0400, Oleg Drokin wrote:
> On Jul 8, 2016, at 7:02 AM, Jeff Layton wrote:
> 
> > On Thu, 2016-07-07 at 21:47 -0400, Oleg Drokin wrote:
> > > It looks like we are bit overzealous about failing mkdir/create/mknod
> > > with permission denied if the parent dir is not writeable.
> > > Need to make sure the name does not exist first, because we need to
> > > return EEXIST in that case.
> > > 
> > > Signed-off-by: Oleg Drokin <green@xxxxxxxxxxxxxx>
> > > ---
> > > A very similar problem exists with symlinks, but the patch is more
> > > involved, so assuming this one is ok, I'll send a symlink one separately.
> > > Âfs/nfsd/nfs4proc.c |ÂÂ6 +++++-
> > > Âfs/nfsd/vfs.cÂÂÂÂÂÂ| 11 ++++++++++-
> > > Â2 files changed, 15 insertions(+), 2 deletions(-)
> > > 
> > 
> > 
> > nit: subject says EPERM, but I think you mean EACCES. The mnemonic I've
> > always used is that EPERM is "permanent". IOW, changing permissions
> > won't ever allow the user to do something. For instance, unprivileged
> > users can never chown a file, so they should get back EPERM there. When
> > a directory isn't writeable on a create they should get EACCES since
> > they could do the create if the directory were writeable.
> 
> Hm, I see, thanks.
> Confusing that you get "Permission denied" from perror ;)
> 

Yes indeed. It's a subtle and confusing distinction.

> > > diff --git a/fs/nfsd/nfs4proc.c b/fs/nfsd/nfs4proc.c
> > > index de1ff1d..0067520 100644
> > > --- a/fs/nfsd/nfs4proc.c
> > > +++ b/fs/nfsd/nfs4proc.c
> > > @@ -605,8 +605,12 @@ nfsd4_create(struct svc_rqst *rqstp, struct nfsd4_compound_state *cstate,
> > > Â
> > > Â	fh_init(&resfh, NFS4_FHSIZE);
> > > Â
> > > +	/*
> > > +	Â* We just check thta parent is accessible here, nfsd_* do their
> > > +	Â* own access permission checks
> > > +	Â*/
> > > Â	status = fh_verify(rqstp, &cstate->current_fh, S_IFDIR,
> > > -			ÂÂÂNFSD_MAY_CREATE);
> > > +			ÂÂÂNFSD_MAY_EXEC);
> > > Â	if (status)
> > > Â		return status;
> > > Â
> > > diff --git a/fs/nfsd/vfs.c b/fs/nfsd/vfs.c
> > > index 6fbd81e..6a45ec6 100644
> > > --- a/fs/nfsd/vfs.c
> > > +++ b/fs/nfsd/vfs.c
> > > @@ -1161,7 +1161,11 @@ nfsd_create(struct svc_rqst *rqstp, struct svc_fh *fhp,
> > > Â	if (isdotent(fname, flen))
> > > Â		goto out;
> > > Â
> > > -	err = fh_verify(rqstp, fhp, S_IFDIR, NFSD_MAY_CREATE);
> > > +	/*
> > > +	Â* Even though it is a create, first we see if we are even allowed
> > > +	Â* to peek inside the parent
> > > +	Â*/
> > > +	err = fh_verify(rqstp, fhp, S_IFDIR, NFSD_MAY_EXEC);
> > > Â	if (err)
> > > Â		goto out;
> > > Â
> > > @@ -1211,6 +1215,11 @@ nfsd_create(struct svc_rqst *rqstp, struct svc_fh *fhp,
> > > Â		goto out;Â
> > > Â	}
> > > Â
> > > +	/* Now let's see if we actually have permissions to create */
> > > +	err = nfsd_permission(rqstp, fhp->fh_export, dentry, NFSD_MAY_CREATE);
> > > +	if (err)
> > > +		goto out;
> > > +
> > > Â	if (!(iap->ia_valid & ATTR_MODE))
> > > Â		iap->ia_mode = 0;
> > > Â	iap->ia_mode = (iap->ia_mode & S_IALLUGO) | type;
> > 
> > 
> > Ouch. This means two nfsd_permission calls per create operation. If
> > it's necessary for correctness then so be it, but is it actually
> > documented anywhere (POSIX perhaps?) that we must prefer EEXIST over
> > EACCES in this situation?
> 
> Opengroup manpage: http://pubs.opengroup.org/onlinepubs/009695399/functions/mkdir.html
> newer version is here:
> http://pubs.opengroup.org/onlinepubs/9699919799/
> 
> They tell us that we absolutely must fail with EEXIST if the name is a symlink
> (so we need to lookup it anyway), and also that EEXIST is the failure code
> if the path exists.
> 

I'm not sure that that verbiage supersedes the fact that you don't have
write permissions on the directory. Does it?

ISTM that it's perfectly valid to shortcut looking up the dentry if the
user doesn't have write permissions on the directory, even when the
target is a symlink.

IOW, I'm not sure I see a bug here.

> Are double permission checks really as bad for nfs? it looked like it would
> call mostly into VFS so even if first call would be expensive, second call should
> be really cheap?
> 

It depends on the underlying fs. In most cases, you're right, but you
can export things that overload the ->permission op, and those can be
as expensive as they like (within reason of course).


> In theory we can do the fh_verify(.., NFSD_MAY_NOP) to skip the first
> nfsd_permission() call, but I suspect then we'd violate the spec in the other way:
> i.e. mkdir in a directory where you cannot read and execute will give you
> EEXIST where as that's when EACCES is desired I imagine.
> 

-- 

Jeff Layton <jlayton@xxxxxxxxxxxxxxx>