[Dovecot] Re: suggestion on crash analysis

[Mark Plaksin]

Timo Sirainen <tss at iki.fi> writes:

> On 22.12.2005, at 21:02, Jeff A. Earickson wrote:
>
>> Another idea...  I know there is the rawlog feature, but I don't
>> want to turn it on for everybody, since asserts are very rare now.
>> Any chance some kind of memory buffer for recent user actions could
>> be kept in memory, then dumped to a file in the person's homedir
>> if an assert occurs?
>
>> This would answer the two unknown questions: 1) What was the user
>> doing?, 2) What was the state of the file he was doing it to?
>
> Hmm. I think I could do that as a plugin, but I'm not sure how useful it
> would be. The commands might tell why the mailbox got into such a state
> that corruption occurred, but might as well not..

*Something* along these lines would be very helpful.  Ever since I started
using the alpha versions I have had Dovecot hang on me at various times.
Try as I might, I can't reliably reproduce the hang.

Looking at truss output, it seems that that Dovecot thinks it has finished
sending a response to the client and has re-entered the poll loop.  But the
rawlogs and the log on my client say that Dovecot has *not* finished
sending the response.

I've asked about this before and you said to try a patch which added some
debug output to the logs (http://dovecot.org/patches/hang-debug.diff).  I
posted the output of that but didn't hear back--I was hoping you found the
cause and fixed it anyhow :)

As you say it would be best if I could show you exactly how to reproduce
the problem.  I'll keep trying to find a recipe for reproducing the problem
but I'm not sure what else to do.  Today it hung while I tried to read two
different large messages (~4k lines of text).  Immediately after the second
hang I copied the problematic folder, inserted a new message just like the
ones that caused the hangs, etc.  But it never hung again.