Incorrect(?) use of open_state->stateid in pnfs

[NeilBrown <neilb@suse.com>]

[-- Attachment #1: Type: text/plain, Size: 1613 bytes --]


Suppose I open/create a file over pNFS and am given a WRITE delegation.
I then chmod the file (as rsync is wont to do) so the delegation is
immediately returned.
I then proceed to write, which triggers a LAYOUT_GET request.  The
stateid for that request it taken from state->stateid, which is still
the delegation stateid.  Naturally it gets NFS4ERR_BAD_STATEID.

When an OPEN is given a delegation, the delegation stateid gets copied
into ->stateid and the open stateid is left in ->open_stateid.
When the delegation is returned, this it done by inode, not open_state,
so it doesn't have easy access to reset the ->stateid.   It could find
all the open_states and fix them up I guess it ....

Other than this usage in pnfs (which dates back to commit b1f69b754e and
is, I believe, incorrect) the ->stateid is *only* used to help find the
write state to recover when an error is reported and the state needs to
be recovered (... though the usage in nfs4_do_handle_exception()
introduced by 272289a3df is a bit different).

Anyway, no other code uses it when choosing a stateid to send in a
request.  So no other code inadvertently uses the delegation stateid
after it has been returned.

How should this be fixed?

- have nfs_start_delegation_return() iterated over all open states
  and copy the open_stateid over the stateid??

- have pnfs_update_layout() use ->open_stateid rather than ->stateid ??
  I suspect that would be wrong.

- have pnfs_update_layout() use ->open_stateid if NFS_I()->delegation
  is NULL ??

Something else?

The last seems easiest, but I'm not certain it is best.

Thanks,
NeilBrown

[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 800 bytes --]