On Thu, Feb 7, 2019 at 12:47 PM Justin Pryzby <pryzby@telsasoft.com> wrote:
> However I *did* reproduce the error in an isolated, non-production postgres
> instance. It's a total empty, untuned v11.1 initdb just for this, running ONLY
> a few simultaneous loops around just one query It looks like the simultaneous
> loops sometimes (but not always) fail together. This has happened a couple
> times.
>
> It looks like one query failed due to "could not attach" in leader, one failed
> due to same in worker, and one failed with "not pinned", which I hadn't seen
> before and appears to be related to DSM, not DSA...
Hmm. I hadn't considered that angle... Some kind of interference
between unrelated DSA areas, or other DSM activity? I will also try
to repro that here...
> I'm also trying to reproduce on other production servers. But so far nothing
> else has shown the bug, including the other server which hit our original
> (other) DSA error with the queued_alters query. So I tentatively think there
> really may be something specific to the server (not the hypervisor so maybe the
> OS, libraries, kernel, scheduler, ??).
Initially I thought these might be two symptoms of the same corruption
but I'm now starting to wonder if there are two bugs here: "could not
allocate %d pages" (rare) might be a logic bug in the computation of
contiguous_pages that requires a particular allocation pattern to hit,
and "dsa_area could not attach to segment" (rarissimo) might be
something else requiring concurrency/a race.
One thing that might be useful would be to add a call to
dsa_dump(area) just before the errors are raised, which will write a
bunch of stuff out to stderr and might give us some clues. And to
print out the variable "index" from get_segment_by_index() when it
fails. I'm also going to try to work up some better assertions.
--
Thomas Munro
http://www.enterprisedb.com