Re: kobject lifetime issues in blk-mq

[Greg Kroah-Hartman]

On Tue, Nov 20, 2018 at 12:34:40PM +0100, Dmitry Vyukov wrote:
> On Thu, Nov 15, 2018 at 1:56 AM, Greg Kroah-Hartman
> <gregkh@xxxxxxxxxxxxxxxxxxx> wrote:
> > On Thu, Nov 15, 2018 at 08:36:17AM +0800, Ming Lei wrote:
> >> > So even if you think the kernel is not going to do this, remember, you
> >> > have no control over it.  Reference counted objects are done this way
> >> > for a reason, you really do not know who has a reference and you really
> >> > do not care.
> >> >
> >> > You are just papering over the real issue here, see my previous email
> >> > for how to start working on resolving it.
> >>
> >> IMO, there isn't real issue, and the issue is actually in 'delay release'.
> >
> > Nope, sorry, that is not true.
> >
> >> Please look at the code in block/blk-mq-sysfs.c, both q->mq_kobj and all
> >> ctx->kobj share same lifetime with q->kobj, we even don't call get/put
> >> on q->mq_kobj & all ctx->kobj, and all are simply released in q->kobj's
> >> release handler.
> >
> > How do you "know" you are keeping those lifetimes in sync?  The joy of a
> > kobject is that _ANYTHING_ can grab a reference to your object without
> > you knowing about it.  That includes userspace programs.  Yes, sysfs is
> > now much better and it trys to release that reference "quickly" when it
> > determines you are trying to delete a kobject, but it's not perfict,
> > there are still races there.
> >
> > And that is what the delay release code is showing you.  It is showing
> > you that you "think" your reference counting is wrong, but it is not.
> > It is showing you that if someone else grabs a reference, you are not
> > correctly cleaning up for yourself.
> >
> > Never think that you really know the lifetime of a kobject, once you
> > realize that your code gets simpler and you can then just "trust" that
> > the kernel will do the right thing no matter what.
> >
> > Because really, you are using a kobject because you want that correct
> > reference counting logic.  By ignoring that logic, you are ignoring the
> > reason to be using that object at all.  If you don't need reference
> > counting, then don't use it at all.
> >
> > And if you need sysfs files, then you need to use the kobject and then
> > you need to handle it properly, because again, you do NOT have full
> > control over the lifetime of your object.  That's the basis for
> > reference counting in the firstplace.
> >
> > So this code is broken without me evening having to look at it, please
> > fix it to handle release properly.  Again, the kernel tried to tell you
> > this, but you hacked around the kernel core to remove that warning
> > incorrectly.  Please go read the kobject documentation again for even
> > more details about this than what I said here.
> >
> > thanks,
> >
> > greg k-h
> 
> Whoever is the right person to fix this, please prioritize this to the
> degree possible.
> This issue does not allow to use DEBUG_KOBJECT_RELEASE in any
> automated testing (in particular syzbot) on both upstream and stable
> trees. We have to disable it for now, so other bugs won't be noticed
> and will pile up.

Patches for this have already been posted :)