RFR: 8366145: Help diagnose ubsan division by zero in computing pause time ratios
Leo Korinth
lkorinth at openjdk.org
Wed Aug 27 10:06:43 UTC 2025
On Wed, 27 Aug 2025 09:17:45 GMT, Thomas Schatzl <tschatzl at openjdk.org> wrote:
>> This pull request is created to help diagnose the bug report https://bugs.openjdk.org/browse/JDK-8354430.
>>
>> The problem is that we get a ubsan runtime error when dividing by zero. Ubsan is not normally running, so I will add an assert making it easier to catch the problem and printing the value of the two variables that when subtracted will create a zero; it will hopefully help with diagnosis.
>>
>> Some comments:
>> 1) the result of `most_recent_gc_end_time_sec()` will not change within the function, thus I will not clutter the code with a new variable, is that okay?
>> 2) I added two calls to `fabs` which are most probably not needed (ieee floating point rules seems to say that positive and negative zero equals). However I think my proposed code makes it clear that I have thought about it.
>>
>> I am now running tier1-3
>
> src/hotspot/share/gc/g1/g1Analytics.cpp line 167:
>
>> 165:
>> 166: double short_interval_ms = (end_time_sec - most_recent_gc_end_time_sec()) * 1000.0;
>> 167: assert(fabs(short_interval_ms) != fabs(0.0), "short_interval_ms should not be zero, calculated from %f and %f", end_time_sec, most_recent_gc_end_time_sec());
>
> Not sure why one would need the `fabs(0.0)`. It simply returns `0.0`.
> Suggestion:
>
> assert(fabs(short_interval_ms) != 0.0, "short_interval_ms should not be zero, calculated from %f and %f", end_time_sec, most_recent_gc_end_time_sec());
>
> Maybe even compare against some very small epsilon, which would also be suspicious given that pauses smaller than a nanosecond or so seem suspicious in any case.
Do you want me to remove the right `fabs` or both `fabs`? I am not interested in an epsilon, I want to see when we get a real division by zero. The code clamps the big value after the computation.
-------------
PR Review Comment: https://git.openjdk.org/jdk/pull/26939#discussion_r2303474578
More information about the hotspot-gc-dev
mailing list