Re: [PATCH 08/16] ublk: remove ublk_commit_and_fetch()

[Ming Lei <ming.lei@redhat.com>]

On Fri, Jul 11, 2025 at 10:05:13AM -0400, Caleb Sander Mateos wrote:
> On Wed, Jul 9, 2025 at 7:57 AM Ming Lei <ming.lei@redhat.com> wrote:
> >
> > On Tue, Jul 08, 2025 at 09:27:57AM -0400, Caleb Sander Mateos wrote:
> > > On Wed, Jul 2, 2025 at 12:04 AM Ming Lei <ming.lei@redhat.com> wrote:
> > > >
> > > > Remove ublk_commit_and_fetch() and open code request completion.
> > > >
> > > > Now request reference is stored in 'ublk_io', which becomes one global
> > > > variable, the motivation is to centralize access 'struct ublk_io' reference,
> > > > then we can introduce lock to protect `ublk_io` in future for supporting
> > > > io batch.
> > >
> > > I didn't follow this. What do you mean by "global variable"?
> >
> > ublk server can send anything to driver with specified tag if batch io
> > extension is added and per-io task is relaxed, then 'ublk_io' instance can be
> > visible to any userpsace command, which needs protection, looks like one
> > global variable.
> >
> > If reference is stored in request pdu, things becomes more like local
> > variable, since the early ublk_io flag check guarantees that concurrent
> > access can't reach 'request'.
> 
> "global variable" means something specific in C, so I would avoid
> using it here to refer to something else. How about something like the
> following?
> 
> Consolidate accesses to struct ublk_io in
> UBLK_IO_COMMIT_AND_FETCH_REQ. When the ublk_io daemon task restriction
> is relaxed in the future, ublk_io will need to be protected by a lock.
> Unregister the auto-registered buffer and complete the request last,
> as these don't need to happen under the lock.

OK.


Thanks,
Ming