Hello Peter,
30.04.2023 23:39, Peter Geoghegan wrote:
> On Sun, Apr 30, 2023 at 5:00 AM Alexander Lakhin <exclusion@gmail.com> wrote:
>> Yeah, also the environment changed since 86dc90056, so I couldn't just revert
>> that commit to test the ExecFilterJunk() behavior on master.
>>
>> All the questions you raised require a more thorough investigation.
> Are you aware of the fact that Valgrind has custom instrumentation
> that makes it directly capable of detecting access to no-longer-pinned
> buffers? See commits 7b7ed046 and 1e0dfd16.
Yes, I am. That's why I can't demonstrate real risks now — my repro
triggers a Valgrind warning only, and no other consequences are visible.
> I think that that may be a factor here. If it is, then it's a little
> surprising that you ever found the problem with ASAN, since of course
> we don't have custom ASAN instrumentation that tells ASAN things like
> "until I say otherwise, it is not okay to read from this range of
> memory, which is this backend's memory mapping for an individual
> shared buffer that was just unpinned".
The ASAN detected a consequence of the invalid memory read:
==1912964==ERROR: AddressSanitizer: heap-buffer-overflow on address
0x7f223dd52c78 at pc 0x55c9e846146a bp 0x7fff0c63e560 sp 0x7fff0c63dd30
WRITE of size 42205122 at 0x7f223dd52c78 thread T0
...
#10 0x000055c9e84d09c5 in fill_val (...) at heaptuple.c:270
...
Note the abnormal size 42205122 (the target table contained much shorter
values at the moment). When exploring the coredump and the source code,
I found that that invalid size was gotten as follows:
ExecBRUpdateTriggers(): newtuple = ExecFetchSlotHeapTuple(newslot, true, &should_free_new);
ExecFetchSlotHeapTuple(): slot->tts_ops->materialize(slot)
tts_buffer_heap_materialize(): bslot->base.tuple = heap_form_tuple(...)
heap_form_tuple(): data_len = heap_compute_data_size(tupleDescriptor, values, isnull);
heap_compute_data_size(): here garbage data were read.
So I'm going to start it over and find a repro that will show something
interesting without Valgrind (and let estimate the probability of a
problem in the field).
Best regards,
Alexander
