Re: danger of stats_temp_directory = /dev/shm

Tom Lane wrote:
> Andres Freund <andres@2ndquadrant.com> writes:
> > On 2013-08-19 14:28:28 -0400, Tom Lane wrote:
> >> One possibility is to do the initial check somewhere shortly after
> >> ChangeToDataDir(), and have the GUC check hook only attempt to make a
> >> check in SIGHUP context.  Unfortunately we aren't passing the context to
> >> check hooks, only GucSource which isn't adequate for this.  Not sure if we
> >> want to go so far as to change the check-hook API at this point.  We could
> >> probably think of some other, klugy way to tell if it's initial startup.
>
> > Is it even actually safe to have stats_temp_directory PGC_SIGHUP after
> > the per DB splitup? I haven't audited the code for it, but it seems
> > somewhat likely that we would end up with some files in the old and some
> > in the new directory?
>
> That's a good point.  For the moment we could just change it to
> PGC_POSTMASTER and eliminate this problem.  Reducing it back to SIGHUP
> would be a future feature, which is evidently less than trivial.

Here's a patchset for this.  The first patch is the same as upthread,
with the enum members renamed; the second is the actual pgstats change.

(I considered the idea that checkDirectoryPermissions returned a bitmask
of the failed checks, instead of just returning a code for the first
check that fails.  This might be useful if some caller wants to ignore
certain problems.  But at the moment I didn't see many other places
where this could be used, so it's probably pointless ATM.)

--
Álvaro Herrera                http://www.2ndQuadrant.com/
PostgreSQL Development, 24x7 Support, Training & Services