RFR: JDK-8298298: NMT: count deltas are printed with 32-bit signed size [v2]

Aleksey Shipilev shade at openjdk.org
Thu Dec 8 10:39:33 UTC 2022


On Thu, 8 Dec 2022 08:56:17 GMT, Thomas Stuefe <stuefe at openjdk.org> wrote:

>> Deltas for counts in NMT differential reports are shown as signed int, restricting them to +-2g. These counts are number of malloc blocks, or number of classes, or number of threads. In all cases number of physically existing somethings.
>> 
>> On large 64-bit machines they can conceivably overflow. Well, at least the malloc block counters. Granted, mallocing even just >2g number of 1 byte blocks costs about 40-50GB due to overhead in libc and NMT (on glibc). But it is conceivable.
>> 
>> Patch changes delta printing from int to ssize_t.
>> 
>> Note that I don't worry about overflowing SSIZE_MIN or SSIZE_MAX here. That would be ridiculous: we just cannot have that many physically existing things. Not on 64bit, not even on 32-bit where ssize_t is int, since in both cases the address space would be the limit. So, if the counters are that large, we have a different problem (probably counter overflow), and so we assert, resp. in release, ignore it.
>> 
>> Note we have the same problem with diff sizes. I expect [8281213](https://bugs.openjdk.org/browse/JDK-8281213) to fix that.
>
> Thomas Stuefe has updated the pull request with a new target base due to a merge or a rebase. The incremental webrev excludes the unrelated changes brought in by the merge/rebase. The pull request contains three additional commits since the last revision:
> 
>  - Factor out delta-ing and assert ranges
>  - Merge branch 'master' into JDK-8298298-nmt-count-deltas-32-bit
>  - JDK-8298298-nmt-count-deltas-32-bit

Looks okay to me.

Windows builds are failing in GHA, though.

-------------

Marked as reviewed by shade (Reviewer).

PR: https://git.openjdk.org/jdk/pull/11568


More information about the hotspot-dev mailing list