Re: [PATCH] io_uring: store back buffer in case of failure

[Olivier Langlois]

On Wed, 2021-06-16 at 16:37 +0100, Pavel Begunkov wrote:
> On 6/16/21 3:44 PM, Jens Axboe wrote:
> > On 6/16/21 8:01 AM, Pavel Begunkov wrote:
> > > On 6/16/21 2:42 PM, Olivier Langlois wrote:
> > > > On Tue, 2021-06-15 at 15:51 -0600, Jens Axboe wrote:
> > > > > Ditto for this one, don't see it in my email nor on the list.
> > > > > 
> > > > I can resend you a private copy of this one but as Pavel
> > > > pointed out,
> > > > it contains fatal flaws.
> > > > 
> > > > So unless someone can tell me that the idea is interesting and
> > > > has
> > > > potential and can give me some a hint or 2 about how to address
> > > > the
> > > > challenges to fix the current flaws, it is pretty much a show
> > > > stopper
> > > > to me and I think that I am going to let it go...
> > > 
> > > It'd need to go through some other context, e.g. task context.
> > > task_work_add() + custom handler would work, either buf-select
> > > synchronisation can be reworked, but both would rather be
> > > bulky and not great.
> > 
> > Indeed - that'd solve both the passing around of locking state
> > which
> > I really don't like, and make it much simpler. Just use task work
> > for
> > the re-insert, and you can grab the ring lock unconditionally from
> > there.
> 
> Hmm, it might be much simpler than I thought if we allocate
> a separate struct callback_head, i.e. task_work, queued it
> with exactly task_work_add() but not io_req_task_work_add(),
> and continue with the request handler. 
> 
ok thx a lot for the excellent suggestions! I think that you have
provided me everything that I need to give a shot for a second version
of this patch.