Corrupted sizes in cache once again

[Tim Evers]

Am 02.02.23 um 16:23 schrieb Aki Tuomi:
>> On 02/02/2023 17:19 EET Stuart Henderson <stu at spacehopper.org> wrote:
>>
>>   
>> On 2023-02-01, Tim Evers <te-ml-ext at artfiles.de> wrote:
>>> I run a fairly large Dovecot Installation (around 100k mailboxes) on
>>> several servers.
>>>
>>> gzip compression is on.
>>>
>>> Every once in a while I get the dreaded "cache corruption" messages in
>>> the log:
>>>
>>> Error: Corrupted record in index cache file
>>> /[redacted]/Maildir/dovecot.index.cache: UID 3868: Broken physical size
>>> in mailbox INBOX:
>>> read(zlib(/[redacted]/Maildir/cur/1674129792.M797543P21755.node2,S=8099,W=8276:2,))
>>> failed: Cached message size smaller than expected (2877 < 8099,
>>> box=INBOX, UID=3868)
>>>
>>> Error: Corrupted record in index cache file
>>> /[redacted]/Maildir/dovecot.index.cache: UID 3875: Broken physical size
>>> in mailbox INBOX:
>>> read(zlib(/[redacted]/Maildir/cur/1674212201.M985809P29112.node2,S=13907,W=14121:2,))
>>> failed: Cached message size smaller than expected (5533 < 8192,
>>> box=INBOX, UID=3875)
>>>
>>> The first entry shows 2877 (size on disk) vs. 8099 (real size unzipped,
>>> also in the filename: S=8099).
>>>
>>> The second entry shows 5533 (size on disk) vs. 8192 - this is not
>>> correct in any way. Size on disk is 13907 as noted in the filename.
>>>
>>> Both mails were delivered trough LMTP and retrieved by the POP3 service.
>>>
>>> Anyone with an idea what might be happening here? I've read all
>>> available info in the doc and in the previous discussions / bug reports,
>>> but nothing seems to match my case. And where does that 8192 come from -
>>> it looks suspicious?
>>>
>>> Version is 2.3.7.2 (Ubuntu 20.04)
>> 2.3.7.2 is rather old now. There were definitely fixes regarding compression
>> around the 2.3.10-2.3.12 timeframe or thereabouts (I forget all the details
>> but it took a release or two before some remaining issues were sorted out
>> after changes in the area). I'd be looking to get it updated to a current
>> version first.
> For bug reports, we do ask that you try to reproduce it with 2.3.20 (current latest), you can get packages from https://repo.dovecot.org/ and would be nice if you can provide steps to reproduce this issue.
>
> Also please see https://www.dovecot.org/bugreport-mail/ if you have not yet.
>
> Aki

This is not a bug report (yet) - I am asking for ideas to narrow down 
the issue myself. I would not expect something this obvious and 
prevalent being a bug in a 10+ years old subsystem (zlib).

Tim