Re: [PATCH] virtio pmem: fix async flush ordering

[Dan Williams]

On Fri, Nov 22, 2019 at 8:09 AM Jeff Moyer <jmoyer@xxxxxxxxxx> wrote:
>
> Dan Williams <dan.j.williams@xxxxxxxxx> writes:
>
> > On Wed, Nov 20, 2019 at 9:26 AM Jeff Moyer <jmoyer@xxxxxxxxxx> wrote:
> >>
> >> Pankaj Gupta <pagupta@xxxxxxxxxx> writes:
> >>
> >> >  Remove logic to create child bio in the async flush function which
> >> >  causes child bio to get executed after parent bio 'pmem_make_request'
> >> >  completes. This resulted in wrong ordering of REQ_PREFLUSH with the
> >> >  data write request.
> >> >
> >> >  Instead we are performing flush from the parent bio to maintain the
> >> >  correct order. Also, returning from function 'pmem_make_request' if
> >> >  REQ_PREFLUSH returns an error.
> >> >
> >> > Reported-by: Jeff Moyer <jmoyer@xxxxxxxxxx>
> >> > Signed-off-by: Pankaj Gupta <pagupta@xxxxxxxxxx>
> >>
> >> There's a slight change in behavior for the error path in the
> >> virtio_pmem driver.  Previously, all errors from virtio_pmem_flush were
> >> converted to -EIO.  Now, they are reported as-is.  I think this is
> >> actually an improvement.
> >>
> >> I'll also note that the current behavior can result in data corruption,
> >> so this should be tagged for stable.
> >
> > I added that and was about to push this out, but what about the fact
> > that now the guest will synchronously wait for flushing to occur. The
> > goal of the child bio was to allow that to be an I/O wait with
> > overlapping I/O, or at least not blocking the submission thread. Does
> > the block layer synchronously wait for PREFLUSH requests?
>
> You *have* to wait for the preflush to complete before issuing the data
> write.  See the "Explicit cache flushes" section in
> Documentation/block/writeback_cache_control.rst.

I'm not debating the ordering, or that the current implementation is
obviously broken. I'm questioning whether the bio tagged with PREFLUSH
is a barrier for future I/Os. My reading is that it is only a gate for
past writes, and it can be queued. I.e. along the lines of
md_flush_request().