Re: Debugging Thinkpad T430s occasional suspend failure.

[Frederic Weisbecker <fweisbec@gmail.com>]

2013/2/15 Linus Torvalds <torvalds@linux-foundation.org>:
> On Fri, Feb 15, 2013 at 9:44 AM, Paul E. McKenney
> <paulmck@linux.vnet.ibm.com> wrote:
>>
>> This commit was designed to increase the probability of hitting the
>> races described in http://lwn.net/Articles/453002/.  These races result
>> in deadlocks involving the runqueue lock (and perhaps also the priority
>> inheritance locks).  And yes, I most certainly should have described
>> this in the commit message.  :-(
>
> Ugh. That particular race seems to be because the softirq handling is
> just crazy, and does the "wakeup_softirqd()" form interrupt context,
> BUT HAS SPECIFICALLY BROKEN THE IRQ COUNTING!
>
> Because it claims to do it from softirq context, which is pure
> garbage. It's not actually in softirq context.
>
> The whole hardirq -> softirq transition seems stupid. I'm sure I made
> some serious mistake in cleaning it up, and there's probably some
> missed tracepoint (or perhaps screwed-up lockdep annotation), but I
> think the hardirq -> softirq preempt thing shoudl be done as an atomic
> preempt downgrade, so that we never have a window of "uhhuh, another
> interrupt can come in between and see us as being in neither). And the
> wakeup_softirqd should be done without playing with preempt count at
> all.
>
> Something like this ENTIRELY UNTESTED patch.
>
> Note: I doubt this patch affects Dave's issue at all, I just started
> looking at that do_softirq code when I read your bug explanation.
>
> Adding random people for kernel/softirq.c to the participants list.
> Comments about the patch? Do note that it's entirely untested, so
> consider it more a RFD than a real patch.. It looks like it adds a lot
> of lines, but most of it is for comments and simplification of the
> logic.

preempt_value_in_interrupt() looks buggy in your patch: it makes
invoke_softirq() returning if (val & HARDIRQ_MASK). But that's always
true since you have moved further the sub_preempt_count(IRQ_EXIT)
further.