Re: client lookup_revalidate bug

["Benjamin Coddington" <bcodding@redhat.com>]

On 30 May 2019, at 6:35, Benjamin Coddington wrote:

> On 29 May 2019, at 13:19, Trond Myklebust wrote:
>> On Wed, 2019-05-29 at 12:39 -0400, Benjamin Coddington wrote:
>>> On 29 May 2019, at 12:21, Trond Myklebust wrote:
>>
>> Sorry, but that's not the case, because of the abuse of the
>> nosharecache flag. You are manipulating the file on the server and
>> expecting an immediate cache invalidation. That would require
>> information that the client does not have.
>
> Yes, forget about the nosharecache flag.  Let's just assume we move 
> the file
> on the server.  The client does perform a GETATTR and sees the updated
> change_attr for the directory, but it matches the dentry's new d_time.
>
> I don't expect immediate cache invalidation; invalidation will never 
> happen
> as long as B/'s change_attr isn't bumped again.  On the client, B/foo
> dentry's d_time is the same as the change_attr for B/, and even though
> there's no file at B/foo on the server, the client will keep the B/foo
> dentry until something else bumps B/'s change_attr.
>
> Maybe I need to imagine a scenario where this matters more..
>
> But, I think that if we use the method of comparing d_time with the 
> parent's
> change_attr to determine if we need to lookup, this is a case where 
> that's
> not reliable.
>
> I'll try to come up with something as I have time to work on it since 
> I am
> getting pressure about it.  It looks to me like this behavior has been
> around for a long time.  Thanks for the look.

Takayuki Nagata points out that this behavior only happens when we have 
a
read delegation, and it seems obvious that we ought to skip lookup
revalidation if so.

Let me see if I can clarify the order of events so anyone that wishes 
can
better argue about what should happen.

Single client, NFS is mounted on /mnt

Client reads /mnt/A/foo, dentry A/foo is hashed.
Client moves /mnt/A/foo to /mnt/B/foo, dentry B/foo is hashed.
Server moves B/foo back to A/foo.
Client reads A/foo, gets a read delegation.
Client looks up B/foo, skips revalidation because it holds the read 
delegation.

Client will always return a positive lookup for both A/foo and B/foo at 
this
point.

Should this happen?  My position is no, it shouldn't.