Remove an erroneous check #53889
Remove an erroneous check #53889
Conversation
|
Tagging subscribers to this area: @dotnet/gc Issue DetailsFixes #53401 I claim that the check I removed is erroneous. The check is performed during the In the earlier plan phase, if we discovered a non-pinned plug that is too close to the earlier pinned plug, we would like to be able to relocate this plug, but then we have no space to store the relocation data. In that case, we will temporarily overwrite the last plug's end to store the relocation data after backing it up. The backup will be restored later (either after compact or sweep) Therefore, it is possible that a pointer in the swept in plan region points to a location where it was temporarily overwritten by the relocation data at that point in time. In that case, the pointer might not be pointing to a valid method table pointer, and that why the check fails. This explains why the bug does not repro if we do not define I ran the test that repro it in a loop for an hour and nothing bad happened with it, I believe this fix is good to go.
|
ghost
locked as resolved and limited conversation to collaborators
Fixes #53401
I claim that the check I removed is erroneous. The check is performed during the
relocate_advance_to_non_sip, which is part ofrelocate_phase.In the earlier plan phase, if we discovered a non-pinned plug that is too close to the earlier pinned plug, we would like to be able to relocate this plug, but then we have no space to store the relocation data. In that case, we will temporarily overwrite the last plug's end to store the relocation data after backing it up. The backup will be restored later (either after compact or sweep)
Therefore, it is possible that a pointer in the swept in plan region points to a location where it was temporarily overwritten by the relocation data at that point in time. In that case, the pointer might not be pointing to a valid method table pointer, and that why the check fails.
This explains why the bug does not repro if we do not define
STRESS_REGIONS, the chances for that to happen is significantly reduced without it.I ran the test that repro it in a loop for an hour and nothing bad happened with it, I believe this fix is good to go.