> Am 07.09.2015 um 21:34 schrieb Timo Sirainen <t...@iki.fi>:
>
> On 07/16/2015 06:03 PM, Sebastian Wiedenroth wrote:
>> Fix hang in safe_sendfile on SmartOS
>>
>> The call to sendfile on SmartOS can fail with EOPNOTSUPP. This is a valid
>> error
>> code and documented in the man page. This error code needs to be handled or
>> else dovecot will retry the sendfile call endlessly and hang.
>
> Committed .. However, I think a more important bug is that it hangs.
> It's definitely not supposed to hang. Which process was it that was
> hanging How can I reproduce that? I can only get it to disconnect the
> IMAP client.
Thanks!
It was the managesieve process that was hanging.
To trigger it we used sieve-connect [1] like this:
sieve-connect -u d...@example.com mailbox.example.com
To find the issue we used a dtrace script [2].
With an unpatched version it would show:
CPU ID FUNCTION:NAME
0 4623 safe_sendfile:return 7 4 fffffd7fffdff8b8 154 ->
77
0 4623 safe_sendfile:return 7 4 fffffd7fffdff8b8
9223372036854775807 -> 77
0 4623 safe_sendfile:return 7 4 fffffd7fffdff8b8
9223372036854775807 -> 77
and then just repeat the last line as it retried the call forever. This is
where it hangs, spinning on the cpu.
After looking at the code I confirmed the issue with a dtrace one-liner that
tracks the sendfilev syscall:
dtrace -n 'syscall::sendfilev:return {printf("%d %x\n", arg0, errno)}'
This showed that the call returned with EOPNOTSUPP (0x7a):
0 6155 sendfilev:return -1 7a
The man page lists this as a valid error code and handling it the same way as
EAFNOSUPPORT fixed the issue for us.
There are a few more error codes in the man page that currently are not handled
by dovecot.
This might be something to look into in the future.
I hope this answer provides the details you’re looking for.
Best regards,
Sebastian
[1] https://github.com/philpennock/sieve-connect
[2] https://gist.github.com/wiedi/4b4ebe5f92ac5b54951b
