On 2016-05-02 12:44:53 -0400, Robert Haas wrote:
> On Mon, May 2, 2016 at 12:41 PM, Andres Freund <andres@anarazel.de> wrote:
> > On 2016-05-02 12:29:45 -0400, Robert Haas wrote:
> >> On Fri, Apr 29, 2016 at 7:58 PM, Andres Freund <andres@anarazel.de> wrote:
> >> > Basically the reason for the problem is that mdsync() needs to access
> >> > "formally non-existant segments" (as in ones where previous segments are
> >> > < RELSEG_SIZE), because we queue (and the might be preexistant) fsync
> >> > requests via register_dirty_segment() in mdtruncate().
> >>
> >> Shouldn't we just throw those flush requests away?
> >
> > Well, we explicity make them for truncations (register_dirty_segment()
> > calls in mdtruncate()). There's no comment as to why - I suspect the
> > idea is that you want to make sure the truncation sticks in case of
> > crash?
>
> I dunno, I don't understand this well enough yet.
>
> > FWIW, falling back to _mdfd_openseg() fixes the issue.
>
> Can you post a patch?
Sure, attached.
I'm not sure this is the best way to go about this. I can see valid
arguments for *always* using _mdfd_openseg() in mdsync(); and I'm
wondering whether we shouldn't make EXTENSION_* into a bitmask
(extend,extend_recovery,return_null,open_deleted).
Andres