| Commit message (Collapse) | Author | Age | Files | Lines |
|---|
| | |
|
| |
|
|
| |
It will help keep more generic reports in "net".
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Let's just accept that we cannot fully trust guilty paths and try to
increase the weight of subsystems extracted from reproducers.
Instead of taking all subsystems that have received the highest number
of votes, take all which have received >= 33%. This will reduce noise
and in almost all cases limit the number of assigned subsystems to 2.
If there are >= 3 reproducers that point to exactly the same set of
subsystems, give them a preference. But still take one subsystem from
guilty paths if there's one that's mentioned >= 66% times.
The numbers themselves are somewhat arbitrary, but hopefully this will
improve the quality of subsystem inference.
Add some more tests.
|
| |
|
|
|
|
|
|
|
|
|
| |
We currently prioritize a subsystem if it's present in all reproducers
and that is supported by at least one guilty path among the considered
crashes.
Due to a small bug in the code we only considered it to be supported if
the guilty path belonged to one of parent subsystems of the one
mentioned in the reproducer. It's fair to also consider full overlap
between them.
|
| |
|
|
|
|
|
| |
There was a small bug and, as a result, subsystems from reproducers
always superceded all other ones. That was not the desired side-effect.
Fix the logic and add a test to linux_test.go.
|
| |
|
|
|
|
|
|
|
| |
There are some minor subsystems (e.g. PAGE CACHE in Linux) that are
parts of several big subsystems. At the same time, a reproducer can
clearly disambiguate such case.
If subsystems from reproducers and subsystems from guilty files
intersect, only proceed with the results of the intersection.
|
| |
|
