RFR: 8264413: Data is written to file header even if its CRC32 was calculated

Yi Yang yyang at openjdk.java.net
Tue Mar 30 09:49:44 UTC 2021


On Tue, 30 Mar 2021 09:25:20 GMT, Yi Yang <yyang at openjdk.org> wrote:

> Many tests under `runtime/cds/appcds/dynamicArchive` are crashed when turning on VerifySharedSpaces, VM reports inconsistent crc32 between dumptime and runtime(See detailed content on JBS attachments):
> 
> $ diff dump.log run.log
> 17c17
> < - base_archive_is_default:        0
> ---
>> - base_archive_is_default:        1
> 19c19
> < - base_archive_name_size:         0
> ---
>> - base_archive_name_size:         113
> 
> The root cause is that even if the CRC32 is calculated(Line 1902), ArchiveBuilder is still writing data to FileMapInfo(Line 1903):
> 
> https://github.com/openjdk/jdk/blob/4ea6abfbd1818fa7049bc4f6e46d568e842acb34/src/hotspot/share/memory/archiveBuilder.cpp#L1091-L1094
> 
> https://github.com/openjdk/jdk/blob/4ea6abfbd1818fa7049bc4f6e46d568e842acb34/src/hotspot/share/memory/filemap.cpp#L1236-L1251
> 
> which results in the expected CRC32(serialized in CDS archive) is different from the calculated ones at runtime. This patch addresses this problem, it writes base_archive_name and size before calculating CRC32 if needed.
> 
> Thanks~
> Yang

We can freeze(just like Metaspace::freeze) the FileMapInfo after CRC32 was calculated, any subsequent writing actions would be reported as illegal. Do you think this is worth doing? Looking forward to your comments.

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

PR: https://git.openjdk.java.net/jdk/pull/3261


More information about the hotspot-runtime-dev mailing list