On Tue, Apr 02 2019, Al Viro wrote:

> On Tue, Apr 02, 2019 at 09:49:34AM -0600, Jonathan Corbet wrote:
>> On Wed, 27 Mar 2019 16:16:53 +1100
>> "Tobin C. Harding" <tobin@kernel.org> wrote:
>> 
>> > Hi Al,
>> > 
>> > This series converts the VFS file Documentation/filesystems/vfs.txt to
>> > reStructuredText format.  Please consider taking this series through
>> > your tree as apposed to Jon's tree because this set makes a fair amount
>> > of changes to VFS files (and also the VFS tree and docs tree are out of
>> > sync right now with the recent work by Mauro and Neil).
>> 
>> Al, do you have any thoughts on how you want to handle this?  I was about
>> to apply Jeff Layton's vfs.txt update, but would rather not create
>> conflicts unnecessarily.  Let me know if you'd like me to pick this work
>> up.
>
> Frankly, I would rather see that file be eventually replaced by something
> saner, and I'm not talking about the format.  Re Jeff's patch... 
>
> +  d_prune: called prior to pruning (i.e. unhashing and killing) a hashed
> +       dentry from the dcache.
>
> is flat-out misguiding.  First of all, it *is* called for unhashed dentries,
> TYVM.  Furthermore, "prior to" is far too vague.
>
> What really happens: there's a point in state diagram for dentries where
> we commit to destroying a dentry and start taking it apart.  That transition
> happens with ->d_lock of dentry, ->i_lock of its inode (if any) and
> ->d_lock of the parent (again, if any) held; ->d_prune() is the last
> chance for filesystem to see the (now doomed) dentry still intact.
>
> It doesn't matter whether it's hashed or not, etc.  The locks held
> are sufficient to stabilize pretty much everything[1] in dentry and
> nothing is destroyed yet.  The only apparent exception is ->d_count,
> but that's not real - we are guaranteed that there had been no other
> counted references to dentry at the decision point and that none
> could've been added.  So this "oh, it's not 0 now, it's gone negative
> after lockref_mark_dead() the caller has just done" is a red herring.
>
> ->d_prune() must not drop/regain any of the locks held by caller.
> It must _not_ free anything attached to dentry - that belongs
> later in the shutdown sequence.  If anything, I'm tempted to
> make it take const struct dentry * as argument, just to make
> that clear.
>
> No new (counted) references can be acquired by that point;
> lockless dcache lookup might find our dentry a match, but
> result of such lookup is not going to be legitimized - it's
> doomed to be thrown out as stale.
>
> It really makes more sense as part of struct dentry lifecycle
> description...  

I would find it useful if the documentation said something about why
this API exists at all.  As you say, it cannot change the dentry - so
what is it expected to do.

I had a look at the two in-tree users and my guess is that it can be
useful if the filesystem caches some other information which would be
invalidated by a dentry being removed.
I *think* cephfs has a flag which records if "All entries in a directory
are currently in the dcache".  When a dentry is pruned, that flag needs
to be cleared.

i.e. ->d_prune allows a filesystem to maintain summary state about what
   is currently in the dcache.
??

Thanks,
NeilBrown

>
> [1] in theory, ->d_time might be changed by overlapping lockless
> call of ->d_revalidate().  Up to filesystem - VFS doesn't touch
> that field (and AFAICS only NFS uses it these days).