mrhhsg opened a new pull request, #69208:
URL: https://github.com/apache/doris/pull/69208
### What problem does this PR solve?
Issue Number: None
Related PR: None
Problem Summary: `PipelineTask::stop_if_finished()` is called by
`FragmentMgr::cancel_worker()` (through
`PipelineFragmentContext::clear_finished_tasks()`) on the background cancel
thread. It copied the `_sink` shared_ptr without any lock. At the same time a
scheduler worker can run `PipelineTask::finalize()` for the same task, which
resets `_sink` while holding `_dependency_lifecycle_lock`. Copying and
resetting the same shared_ptr instance from two threads without synchronization
is a data race (undefined behavior): the cancel thread can observe a torn
pointer/control block and then call `is_finished()` on an object that is being
released. Every other cross-thread reader of `_sink` (`debug_string()`, the
hybrid scheduler submit path) is already synchronized; `stop_if_finished()` was
the only one left.
Fix: take `_dependency_lifecycle_lock` in `stop_if_finished()`, return if
the task is already finalized, and copy `_sink` under the lock, following the
pattern of `debug_string()`. `is_finished()` and `unblock_all_dependencies()`
are called after the lock is released, because `unblock_all_dependencies()`
acquires the same non-recursive mutex. A debug point inside the critical
section lets the unit test hold the lock deterministically.
### Release note
None
### Check List (For Author)
- Test: Unit Test
- Added `PipelineTaskTest.TEST_STOP_IF_FINISHED_EXCLUDES_FINALIZE`: it
holds `stop_if_finished()` inside the critical section and checks that a
concurrent `finalize()` is blocked and cannot reset `_sink`, and that a
finalized task is skipped. It fails without the lock and passes with it (200
repeated runs). All `PipelineTaskTest` cases pass.
- No regression test: the race window is a few instructions wide and
cannot be triggered reliably from SQL.
- Behavior changed: No
- Does this need documentation: No
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
