On Sun, Apr 21, 2019 at 03:08:22AM -0500, Justin Pryzby wrote:
>On Sun, Apr 21, 2019 at 01:03:50AM -0400, Gunther wrote:
>> On 4/20/2019 21:14, Tomas Vondra wrote:
>> >Maybe. But before wasting any more time on the memory leak investigation,
>> >I suggest you first try the patch moving the BufFile allocations to a
>> >separate context. That'll either confirm or disprove the theory.
>>
>> OK, fair enough. So, first patch 0001-* applied, recompiled and
>>
>> 2019-04-21 04:08:04.364 UTC [11304] LOG: server process (PID 11313) was terminated by signal 11: Segmentation
fault
>...
>> turns out the MemoryContext is NULL:
>>
>> (gdb) p context
>> $1 = (MemoryContext) 0x0
>
>I updated Tomas' patch to unconditionally set the context.
>(Note, oldctx vs oldcxt is fairly subtle but I think deliberate?)
>
I don't follow - is there a typo confusing oldctx vs. oldcxt? I don't
think so, but I might have missed something. (I always struggle with which
spelling is the right one).
I think the bug is actually such simpler - the memory context was created
only in ExecuteIncreaseNumBatches() when starting with nbatch=1. But when
the initial nbatch value was higher (i.e. when starting with 2 or more
batches), it was left NULL. That was OK for testing with the contrived
data set, but it may easily break on other examples.
So here is an updated patch - hopefully this version works. I don't have
time to do much more testing now, though. But it compiles.
regards
--
Tomas Vondra http://www.2ndQuadrant.com
PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services
