Fix T84637
See T84637#1134707 and T84637#1135324
- Replace allocation tests with direct tests of MEM_size_safe_multiply.
- Disable death tests on macOS with ASan. Release builds will check if abort occurs or not, upon integer overflow.
Details
Details
Diff Detail
Diff Detail
- Repository
- rB Blender
- Branch
- alloc (branched from master)
- Build Status
Buildable 13797 Build 13797: arc lint + arc unit
Event Timeline
Comment Actions
Would be nice to test the actual allocations functions, but I can't think of a batter solution than this.
About the names, perhaps use safe_multiply_pass and safe_multiply_fail.
Comment Actions
Rename
LockFreeAllocatorTest to AllocatorTest,
LockfreeIntegerOverflow to IntegerOverflow, (as lock-free is not a highlight here.)
multiply_pass to safe_multiply_pass,
multiply_fail to safe_multiply_fail,
Remove the base class include since test fixture is not used.
Remove GCC pragma now that there are no allocations.
Update top level comment.
| intern/guardedalloc/tests/guardedalloc_overflow_test.cc | ||
|---|---|---|
| 67 | total_size is size_t, so using boolean-typed EXPECT_TRUE seems meaningless to me. | |
Comment Actions
If the EXPECT_EXIT can not be trusted on some platform/configuration this is to be checked with the upstream gtest library.
Comment Actions
It's not that EXPECT_EXIT can't be trusted, it's that it tests for exit with SIGABRT specifically and that the signal you get from a failing malloc (with asan) is not well-defined.
Maybe the best solution is to add a EXIT_ANY_SIGNAL_PREDICATE that does not test for abort specifically.
Comment Actions
... that the signal you get from a failing malloc (with asan) ..
The malloc is not relevant here.. Even if MallocArray function is edited to be empty (do nothing), the test will fail. AFAIK, it's due to fork(). See T84637#1135324
Also, these tests are the integer overflow ones. The malloc would not be called even if the fork issue is resolved (like in builds without asan). The abort happens before malloc. https://developer.blender.org/diffusion/B/browse/master/intern/guardedalloc/intern/mallocn_guarded_impl.c$480-489
Comment Actions
Ok, I think that may be a bug in asan/ubsan, or some odd behavior.
Perhaps we should then change ABORT_PREDICATE to check for either SIGBART or SIGILL on macOS.
Comment Actions
- Disable tests if apple and asan enabled.
It'd be slightly deceptive to see a test pass without it even running.
Here's a compromise to disable the test on macOS if Asan is enabled.