[Dovecot] dovecot servers hanging with fuse/glusterfs errors

[Bill Cole]

At 11:10 AM +0100 1/28/08, Jordi Moles wrote:
>hi.
>
>i've got a clustered mail system with glusterfs and some postfix, 
>squirrelmail and dovecot machines sharing the same storage system 
>through glusterfs.

Why?

I can see the Postfix/Dovecot justification (barely) but it seems 
pointless for SquirrelMail. What storage does SquirrelMail share with 
the other types of machines, and why? Since SquirrelMail normally 
accesses mail only via IMAP, there's no point in making the back end 
of the IMAP server visible to it.


>each server has only postfix or squirrel or dovecot installed on it.
>The thing is... dovecot servers hang very often and the last thing 
>they always log is this:
>
>*************
>
>login: Unable to handle kernel paging request at 0000000000100108 RIP:
>[<ffffffff88020838>] :fuse:request_end+0x45/0x109
>PGD 1f729067 PUD 1faae067 PMD 0
>Oops: 0002 [1] SMP
>CPU 0
>Modules linked in: ipv6 fuse dm_snapshot dm_mirror dm_mod
>Pid: 678, comm: glusterfs Not tainted 2.6.18-xen #1
>RIP: e030:[<ffffffff88020838>]  [<ffffffff88020838>] 
>:fuse:request_end+0x45/0x109
>RSP: e02b:ffff88001f04dd68  EFLAGS: 00010246
>RAX: 0000000000200200 RBX: ffff88001db9fa58 RCX: ffff88001db9fa68
>RDX: 0000000000100100 RSI: ffff88001db9fa58 RDI: ffff88001f676400
>RBP: ffff88001f676400 R08: 00000000204abb00 R09: ffff88001db9fb58
>R10: 0000000000000008 R11: ffff88001f04dcf0 R12: 0000000000000000
>R13: ffff88001db9fa90 R14: ffff88001f04ddf8 R15: 0000000000000001
>FS:  00002b29187c53b0(0000) GS:ffffffff804cd000(0000) knlGS:0000000000000000
>CS:  e033 DS: 0000 ES: 0000
>Process glusterfs (pid: 678, threadinfo ffff88001f04c000, task 
>ffff88001fc5a860)
>Stack:  ffff88001db9fa58 ffff88001f676400 00000000fffffffe ffffffff88021056
>ffff88001f04def8 000000301f04de88 ffffffff8020dd40 ffff88001f04ddb8
>ffffffff80225ca3 ffff88001de23500 ffffffff803ef023 ffff88001f04de98
>Call Trace:
>[<ffffffff88021056>] :fuse:fuse_dev_readv+0x385/0x435

This is almost certainly a bug in the kernel proper, the fuse module, 
or possibly the glusterfs client. If you're using cheap hardware it 
could also be caused by a hardware issue like an intermittent memory 
failure. This failure is at too low a level to be the fault of a 
misbehaving application.

[...]

>these are basic debian etch brand new set ups, they have nothing 
>else than dovecot and glusterfs-client installed on them. The thing 
>is all machines i've got have the very same version of 
>glusterfs-client, but only the dovecot ones hang.
>
>Do you have any idea?

The same failure across multiple machines implies a problem in 
software, not hardware.

The fact that Dovecot is poking this bug where postfix and 
squirrelmail are not is not particularly surprising, and doesn't 
really indicate a Dovecot flaw. A mailstore server like Dovecot makes 
more complex demands for proper behavior from a filesystem than an 
MTA like Postfix.



-- 
Bill Cole                                  
bill at scconsult.com