RFR: 8339711: ZipFile.Source.initCEN needlessly reads END header [v2]

[Claes Redestad]

On Wed, 25 Sep 2024 16:22:55 GMT, Eirik Bjørsnøs <eirbjo at openjdk.org> wrote:

>> Please review this cleanup PR which makes `ZipFile.Source.initCEN` not include the 22-byte trailing`END` header when reading the `CEN` section of the ZIP file.
>> 
>> The reading of the END header was probably brought over from native code with the transition to Java in JDK 9.
>> 
>> In the current JDK, the END header is unused. This needlessly complicates multiple code paths accessing the array since they must account for the trailing END record when calculating the end of CEN position.
>> 
>> Additionally, the enforcement of the maximum CEN size limit is currently off by one. It allows the construction of a byte array of size `Integer.MAX_VALUE - 1`, but this size is not supported by OpenJDK. Instead, the maximum CEN limit should be such that is does not exceed  `Integer.MAX_VALUE - 2`.
>> 
>> Testing:
>> 
>> The `EndOfCenValidation` test is updated to test the rejection of a CEN of size `Integer.MAX_VALUE - 1` as the new minumum rejected CEN size.
>> 
>> The `ZipFileOpen` benchmark seems neutral to this change.
>
> Eirik Bjørsnøs 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:
> 
>  - Merge branch 'master' into zipfile-endhdr
>  - Merge branch 'master' into zipfile-endhdr
>  - Do not include the END header when reading the CEN section

LGTM. Simplifies a bit and allows for another 22 bytes of CEN.

src/java.base/share/classes/java/util/zip/ZipFile.java line 1740:

> 1738:                 }
> 1739:                 // read in the CEN
> 1740:                 if (end.cenlen > MAX_CEN_SIZE) {

We'd be throwing an OOMError today if we soared too close to the limit (`Integer.MAX_VALUE - ENDHDR - 2` and above), then throw `zerror` if we go beyond the limit. With these changes we change to throw `zerror` exclusively. An OOME can of course still be thrown from any subsequent allocation on a _real_ OOM, so I'm not sure how valuable this edge case trimming is.

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

Marked as reviewed by redestad (Reviewer).

PR Review: https://git.openjdk.org/jdk/pull/20905#pullrequestreview-2329393055
PR Review Comment: https://git.openjdk.org/jdk/pull/20905#discussion_r1775907509