Re: [PATCH v6 4/5] 9pfs: T_readdir latency optimization

[Christian Schoenebeck]

On Donnerstag, 2. Juli 2020 19:23:35 CEST Christian Schoenebeck wrote:
> > > Back to the actual topic: so what do we do about the mutex then? CoMutex
> > > for 9p2000.u and Mutex for 9p2000.L? I know you find that ugly, but it
> > > would just be a transitional measure.
> > 
> > That would ruin my day...
> > 
> > More seriously, the recent fix for a deadlock condition that was present
> > for years proves that nobody seems to be using silly clients with QEMU.
> > So I think we should just dump the lock and add a one-time warning in
> > the top level handlers when we detect a duplicate readdir request on
> > the same fid. This should be a very simple patch (I can take care of
> > it right away).
> 
> Looks like we have a consensus! Then I wait for your patch removing the
> lock, and for your feedback whether or not you see anything else in this
> patch set?

Please wait before you work on this patch. I really do think your patch should 
be based on/after this optimization patch for one reason: if (and even though 
it's a big if) somebody comes along with a silly client as you named it, then 
your patch can simply be reverted, which would not be possible if it's next.

So I would really suggest I change this patch here to go the ugly way with 2 
mutex types for readdir 9p2000.L vs 9p2000.L, and your patch would get rid of 
that mess by removing the lock entirely, okay?

Best regards,
Christian Schoenebeck