Re: [RFC PATCH 8/9] livepatch: allow patch modules to be removed

[Josh Poimboeuf <jpoimboe@redhat.com>]

On Mon, Feb 16, 2015 at 05:06:15PM +0100, Miroslav Benes wrote:
> On Fri, 13 Feb 2015, Josh Poimboeuf wrote:
> 
> > On Fri, Feb 13, 2015 at 05:17:10PM +0100, Miroslav Benes wrote:
> > > On Fri, 13 Feb 2015, Josh Poimboeuf wrote:
> > > > Hm, even with Jiri Slaby's suggested fix to add the completion to the
> > > > unregister path, I still get a lockdep warning.  This looks more insidious,
> > > > related to the locking order of a kernfs lock and the klp lock.  I'll need to
> > > > look at this some more...
> > > 
> > > Yes, I was afraid of this. Lockdep warning is a separate bug. It is caused 
> > > by taking klp_mutex in enabled_store. During rmmod klp_unregister_patch 
> > > takes klp_mutex and destroys the sysfs structure. If somebody writes to 
> > > enabled just after unregister takes the mutex and before the sysfs 
> > > removal, he would cause the deadlock, because enabled_store takes the 
> > > "sysfs lock" and then klp_mutex. That is exactly what the lockdep tells us 
> > > below.
> > > 
> > > We can look for inspiration elsewhere. Grep for s_active through git log 
> > > of the mainline offers several commits which dealt exactly with this. Will 
> > > browse through that...
> > 
> > Thanks Miroslav, please let me know what you find.  It wouldn't surprise
> > me if this were a very common problem.
> > 
> > One option would be to move the enabled_store() work out to a workqueue
> > or something.
> 
> Yes, that is one possibility. It is not the only one.
> 
> 1. we could replace mutex_lock in enabled_store with mutex_trylock. If the 
> lock was not acquired we would return -EBUSY. Or could we 'return 
> restart_syscall' (maybe after some tiny msleep)?

Hm, doesn't that still violate the locking order rules?  I thought locks
always had to be taken in the same order -- always sysfs before klp, or
klp before sysfs.  Not sure if there would still be any deadlocks
lurking, but lockdep might still complain.

> 2. we could reorganize klp_unregister_patch somehow and move sysfs removal 
> out of mutex protection.

Yeah, I was thinking about this too.  Pretty sure we'd have to remove
both the sysfs add and the sysfs removal from mutex protection.  I like
this option if we can get it to work.

-- 
Josh