[dm-devel] [PATCH 56/57] multipathd: push down lock in checkerloop()

[Hannes Reinecke]

On 05/04/2016 01:24 AM, Benjamin Marzinski wrote:
> On Tue, May 03, 2016 at 05:17:34PM -0500, Benjamin Marzinski wrote:
>> On Wed, Apr 27, 2016 at 01:10:57PM +0200, Hannes Reinecke wrote:
>>> Instead of grabbing the lock at the start of the checkerloop
>>> and releasing it at the end we should be holding it only
>>> during the time when we actually need it.
>>
>> I'm pretty sure that this can cause crashes if multipathd reconfigures
>> when it's just started servicing a uevent. It goes like this. The
>> uevq_thr thread calls uev_trigger(), which checks the running_state and
>> sees that it's DAEMON_IDLE.  The uxlsnr_thr thread gets a
>> reconfiguration request, which eventually calls set_config_state().
>> set_config_state() sees the state is DAEMON_IDLE, sets it to
>> DAEMON_CONFIGURE, and returns without waiting.  This wakes up the child
>> thread, which runs reconfigure().  reconfigure() sets conf to NULL. Then
>> the uevq_thr thread in uev_trigger() calls filter_devnode()
>> dereferencing conf->blist_devnode and conf->elist_devnode. conf is NULL,
>> and thus it crashes.
>>
>> The short version is that multipathd has been using the vecs lock to
>> protect conf.  With your change to uev_trigger, you access conf outside
>> of the vecs lock. This isn't a problem in checkerloop(), since you can't
>> reconfigure while multipathd is in the DAEMON_RUNNING state, so that is
>> protecting conf. You need to do the same
>> set_config_state(DAEMON_RUNNING) in uev_trigger to make this safe.
> 
> Or not. ev_add_map calls set_config_state(DAEMON_CONFIGURE), so this
> would deadlock if you called  set_config_state(DAEMON_RUNNING) in
> uev_trigger. At any rate, either reconfigures need to be blocked while
> running uev_trigger, or we need some sort of reference counting on the
> config structure so that we don't free it when we do a reconfigure,
> until everyone has switched to the new structure.
> 
See my previous reply. We can easily move the call to
filter_devnode() so that it'll be called with the vector lock held.

Cheers,

Hannes
-- 
Dr. Hannes Reinecke		   Teamlead Storage & Networking
hare at suse.de			               +49 911 74053 688
SUSE LINUX GmbH, Maxfeldstr. 5, 90409 Nürnberg
GF: F. Imendörffer, J. Smithard, J. Guild, D. Upmanyu, G. Norton
HRB 21284 (AG Nürnberg)