Message-ID: <20150327200512.4309377d@thinkpad-w530>
Date:	Fri, 27 Mar 2015 20:05:12 +0100
From:	David Hildenbrand <dahi@...ux.vnet.ibm.com>
To:	Peter Zijlstra <peterz@...radead.org>
Cc:	linux-kernel@...r.kernel.org, linux-arch@...r.kernel.org,
	tglx@...utronix.de, benh@...nel.crashing.org, paulus@...ba.org,
	akpm@...ux-foundation.org, heiko.carstens@...ibm.com,
	schwidefsky@...ibm.com, borntraeger@...ibm.com, mst@...hat.com,
	David.Laight@...LAB.COM, hughd@...gle.com, hocko@...e.cz
Subject: Re: [PATCH v2 0/5] Reenable might_sleep() checks for might_fault()

> On Fri, Mar 27, 2015 at 04:40:50PM +0100, David Hildenbrand wrote:
> > e.g. futex_atomic_op_inuser(): easy to fix, add preempt_enable/disable
> > respectively.
> > 
> > e.g. futex_atomic_cmpxchg_inatomic(): not so easy / nice to fix.
> > 
> > The "inatomic" variants rely on the caller to make sure that preemption is
> > disabled.
> > 
> >         pagefault_disable();
> >         ret = futex_atomic_cmpxchg_inatomic(curval, uaddr, uval, newval);
> >         pagefault_enable();
> 
> Typically the _inatomic() variants of functions have the exception
> tables required for fixups and can return -EFAULT. In that regard the
> futex_atomic_cmpxchg_inatomic() is consistently named.
> 
> In specific the above is taken from cmpxchg_futex_value_locked(), which
> is private to futex.c, so we don't really need to worry about it.
> 
> Furthermore, the futex.c helpers that wrap them in pagefault_disable()
> do so because they want to handle the fault themselves. I don't think we
> need to worry about that.

I totally agree with pagefault_disable() and that -EFAULT logic to handle that
themselves. I'm basically only concerned about implicitly used disabled
preemption.

> 
> > 1. We could simply add preempt_disable/enable to the calling code. However that
> > results in _all_ futex_atomic_cmpxchg_inatomic() running with disabled
> > preemption, although the implementation doesn't really need it. So there is not
> > really a "decoupling", but to counters to set.
> 
> Not really needed, the few callsites where they are not already under a
> lock is where we want to explicitly handle the -EFAULT case ourselves.
> 
> > 2. We could add the preempt_disable/enable to the implementations that only
> > need it, leaving calling code as is. However, then the name
> > "futex_atomic_cmpxchg_inatomic" is misleading, because it has nothing to do
> > with "inatomic" anymore.
> 
> The _inatomic() naming is because it _can_ be called from atomic
> context, like __copy_to_user_inatomic(). It doesn't mean it has to.

Well, they have to be called from an pagefault_disabled environment (for now
atomic). Atomic context is optional, with a few exceptions (see next section).

> These functions work just fine outside of atomic regions.

To make clear what I'm worried about, have a look at the following code taken
from include/asm-generic/futex.h):

static inline int
futex_atomic_cmpxchg_inatomic(u32 *uval, u32 __user *uaddr,
                              u32 oldval, u32 newval)
{
        u32 val;

        if (unlikely(get_user(val, uaddr) != 0))
                return -EFAULT;

        if (val == oldval && unlikely(put_user(newval, uaddr) != 0))
                return -EFAULT;

        *uval = val;

        return 0;
}

This _has to_ be called from an atomic context. Otherwise the logic is broken
(mutual exclusion). Not adding a preempt_disable() somewhere in the calling code
(or the function itself) will not allow this function to work properly. At
least that's my understanding :)

And we have exactly that case when we drop preempt_disable() from pagefault_disable()
in the futex code.

My quick hack for this special case would be to add preempt_disable/enable to
that function body. But maybe I am totally wrong about that given code and
preemption.

> 
> And they still can be used in atomic regions, but now
> pagefault_disable() will also trigger the exception fixup.
> 
> I don't think we should worry too much about this.
> 
> > The same applies to other "inatomic" functions. I think most of these functions
> > rely on pagefaults to be disabled in order to work correctly, not disabled
> > preemption.

I agree. The kmap_atomic stuff is another candidate I identified that needs
additional preempt_disable().

> > 
> > Any idea how to fix this or what would be the way to go?
> 
> I have the feeling you're over thinking this. _inatomic() has exception
> fixups and will return -EFAULT when it cannot do the pagefault in place,
> for whatever reason -- traditionally because of atomic context, now also
> pagefault_disable().

Haha, well I don't want to break things. And places like the futex code look
suspicious. That's why I better double check with an expert.

> 
> And esp. things like futexes have been extensively used under -rt and
> are known good.

Yes, on most configuration, but maybe not all (archs that use asm-generic code
+ !CONFIG_SMP + CONFIG_PREEMPT)

Thanks for your reply.

David

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/

Powered by blists - more mailing lists