Hello, Michael!
> I fail to see how this is related? The original issue was that this
> was impossible to run safely concurrently, but now we have the
> facilities able to do so. There are a few cases where using a wait
> point has limits, for example outside a transaction context for some
> of the processes, but that has not really been an issue up to now.
I encountered this issue while trying to stabilize tests for [0].
Tests were crashing during the installcheck with assertion of that thread.
I have spent time trying to identify the root cause - injection points and
the way it may affect another tests even if them set to be executed locally.
In the spec, the backend performs the following:
SELECT injection_points_set_local();
SELECT injection_points_attach('invalidate_catalog_snapshot_end',
'wait');
That's all - we don't even execute any command that would trigger the wait
condition.
Meanwhile, three different backends are attempting to test SERIALIZABLE
isolation without any injection points.
Initially, this was a separate 'two-ids' test executed in parallel during
installcheck,
but I incorporated it into the reproducer spec for simplicity.
> Isn't that pointing to an actual bug with serializable transactions?
No, let me explain.
> What you are telling here is that there is a race condition where it
> is possible to trigger an assertion failure when finishing a session
> while another one is waiting on an invalidation, if there's in the mix
> a read/write dependency error.
Actually, no backend is waiting for invalidation in this case.
Here's the sequence of events:
* The s4 backend creates a local injection point but performs no further
actions. The injection point is marked as local for that pid.
* Three other backends proceed with their serializable snapshot operations.
* s2 determines it cannot commit and correctly decides to abort the
transaction.
* s2 begins releasing resources:
ResourceOwnerReleaseInternal resowner.c:694 <--- NOTE: After
starting the release process, by calling this function, no new
ResourceOwnerRelease resowner.c:654 resources
can be remembered in the resource owner.
AbortTransaction xact.c:2960
AbortCurrentTransactionInternal xact.c:3531
AbortCurrentTransaction xact.c:3449
PostgresMain postgres.c:4513
BackendMain backend_startup.c:107
postmaster_child_launch launch_backend.c:274
BackendStartup postmaster.c:3377
ServerLoop postmaster.c:1663
PostmasterMain postmaster.c:1361
main main.c:196
* During transaction abort, s2 invalidates its catalog snapshot with this
stack trace:
InvalidateCatalogSnapshot snapmgr.c:430
AtEOXact_Snapshot snapmgr.c:1050
CleanupTransaction xact.c:3016
AbortCurrentTransactionInternal xact.c:3532
AbortCurrentTransaction xact.c:3449
PostgresMain postgres.c:4513
BackendMain backend_startup.c:107
postmaster_child_launch launch_backend.c:274
BackendStartup postmaster.c:3377
ServerLoop postmaster.c:1663
PostmasterMain postmaster.c:1361
main main.c:196
* Consequently, s2 encounters
INJECTION_POINT("invalidate_catalog_snapshot_end");
* Although invalidate_catalog_snapshot_end is set to 'wait' only for s4, s2
enters the injection_wait handler
* Since this is s2's first interaction with injection points (as
injection_points isn't in shared_preload_libraries), it calls
injection_init_shmem
* Here, GetNamedDSMSegment is called - this is new infrastructure for
initializing shared memory for extensions without shared_preload_libraries,
committed by Nathan [1]
* GetNamedDSMSegment attempts to attach to memory and triggers the
assertion:
if (owner->releasing)
elog(ERROR, "ResourceOwnerEnlarge called after release
started");
ResourceOwnerEnlarge resowner.c:449
dsm_create_descriptor dsm.c:1206
dsm_attach dsm.c:696
dsa_attach dsa.c:519
init_dsm_registry dsm_registry.c:115
GetNamedDSMSegment dsm_registry.c:156
injection_init_shmem injection_points.c:185
injection_wait injection_points.c:277
InjectionPointRun injection_point.c:551
InvalidateCatalogSnapshot snapmgr.c:430
AtEOXact_Snapshot snapmgr.c:1050
CleanupTransaction xact.c:3016
AbortCurrentTransactionInternal xact.c:3532
AbortCurrentTransaction xact.c:3449
PostgresMain postgres.c:4513
BackendMain backend_startup.c:107
postmaster_child_launch launch_backend.c:274
BackendStartup postmaster.c:3377
ServerLoop postmaster.c:1663
PostmasterMain postmaster.c:1361
main main.c:196
* This assertion during transaction abort triggers another abort call (this
could be improved):
ProcArrayEndTransaction procarray.c:677
AbortTransaction xact.c:2946
AbortCurrentTransactionInternal xact.c:3531
AbortCurrentTransaction xact.c:3449
PostgresMain postgres.c:4513 <--------------- exception handler here
BackendMain backend_startup.c:107
postmaster_child_launch launch_backend.c:274
BackendStartup postmaster.c:3377
ServerLoop postmaster.c:1663
PostmasterMain postmaster.c:1361
main main.c:196
This isn't the same abort attempt - it's the second one, which triggers
Assert(TransactionIdIsValid(proc->xid));
In summary:
Each code component functions correctly in isolation
However, when an injection point registered as local by one backend causes
another backend to register resources (and potentially other operations),
it can lead to difficult-to-diagnose issues
I see several potential solutions:
* Add injection_points to shared_preload_libraries for all tests
* Implement a mechanism to call prev_shmem_startup_hook for libraries
outside shared_preload_libraries
* Modify GetNamedDSMSegment's behavior
* Run all injection_points tests in an isolated environment
In my opinion, the last option seems most appropriate.
Best regards,
Mikhail.
[0]: https://commitfest.postgresql.org/50/5160/
[1]:
https://github.com/postgres/postgres/commit/8b2bcf3f287c79eaebf724cba57e5ff664b01e06
