Skip to contentSkip to conversation

dealing with ->d_revalidate() right ways (automounting, atomic opens and the whole mess around that)

topic proposal Al Viro 4 participants 4 messages Jun 9, 2010 – Jun 21, 2010 VFS

lore
Share & tools
lore

Al Viro

author maintainer
#0 lore raw

That's going to be the next big change of VFS architecture; bits
and pieces had been discussed on lists and in private mail, but it would be really useful to sit down and discuss that face to face with everyone involved.

* ->follow_link() is abused for triggering submounts.
* ->d_revalidate() is abused in many, _many_ ways, including
autofs, NFS "reval_dot" stuff, atomic opens, etc. Moreover, it has nasty interplay with crossing mountpoints.
* NFS handling of atomic opens is near the point where we can
switch to methods that would get proper arguments directly, rather than abusing ->lookup/->d_revalidate/whatnot. Again, ->d_revalidate and mountpoint crossing are stumbling blocks. And NFS4 bits around open are just plain scary...
* NFSv4 has rather unpleasant stuff in submount handling - what
it does with ->mnt_devname doesn't bear describing in polite company and it actually doesn't work right with mount --bind (moderately) and mountinfo contents (completely). Both issues are fixable and that gets entangled into the aforementioned ball of barbed wire...

Who might want to be there: nfsd/nfs/cifs/afs/autofs4 and general
VFS-interested folks.

#1 lore raw
Quoted 4 lines · Al Viro

That's going to be the next big change of VFS architecture; bits
and pieces had been discussed on lists and in private mail, but it would be really useful to sit down and discuss that face to face with everyone involved.

I'd want to be in on this from the ocfs2 perspective. I think
that d_revalidate() works for us as-is, but I'd want to be in on proposed changes.

Joel

Signature · 8 lines
"We'd better get back, `cause it'll be dark soon,
 and they mostly come at night.  Mostly."

Joel Becker
Principal Software Developer
Oracle
E-mail: [email protected]
Phone: (650) 506-8127
#2 lore raw
In reply to Al Viro
Quoted 20 lines · Al Viro

That's going to be the next big change of VFS architecture; bits
and pieces had been discussed on lists and in private mail, but it would be really useful to sit down and discuss that face to face with everyone involved.

* ->follow_link() is abused for triggering submounts.
* ->d_revalidate() is abused in many, _many_ ways, including
autofs, NFS "reval_dot" stuff, atomic opens, etc. Moreover, it has nasty interplay with crossing mountpoints.
* NFS handling of atomic opens is near the point where we can
switch to methods that would get proper arguments directly, rather than abusing ->lookup/->d_revalidate/whatnot. Again, ->d_revalidate and mountpoint crossing are stumbling blocks. And NFS4 bits around open are just plain scary...
* NFSv4 has rather unpleasant stuff in submount handling - what
it does with ->mnt_devname doesn't bear describing in polite company and it actually doesn't work right with mount --bind (moderately) and mountinfo contents (completely). Both issues are fixable and that gets entangled into the aforementioned ball of barbed wire...

Quoted 3 lines

Who might want to be there: nfsd/nfs/cifs/afs/autofs4 and general
VFS-interested folks.
--

I second. It would be great to clean all this up and get rid of the intents stuff. Al and I started the discussion last year, but it would be useful to include a larger audience.

Cheers
Trond

#3 lore raw
In reply to Al Viro
Quoted 4 lines · Al Viro

That's going to be the next big change of VFS architecture; bits
and pieces had been discussed on lists and in private mail, but it would be really useful to sit down and discuss that face to face with everyone involved.

Sorry to come in late; I've been on vacation. Are you proposing that this happen at the summit in August? Or at some other time?

Signature · 2 lines
Frank Mayhar <[email protected]>
Google, Inc.