2013/2/17 Linus Torvalds : > On Sun, Feb 17, 2013 at 7:11 AM, Frederic Weisbecker wrote: >> >> 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. > > No, that's not it. The value hasn't been written back yet, but it already did: > > + int offset = IRQ_EXIT_OFFSET; > + int count = preempt_count() - offset; > > so the 'count' has the IRQ_EXIT_OFFSET already subtracted. So no, > HARDIRQ_MASK is *not* always set. Another thing. Perhaps we can push the idea of your patch a little further by re-entering HARDIRQ_OFFSET at the end of the softirq processing and do the final sub_preempt_count(HARDIRQ_OFFSET) at the very end of irq_exit(). This way irq_exit() looks a bit more simple to me: everything there becomes considered as in hardirq: (ie: rcu_irq_exit() and tick_nohz_irq_exit() won't appear anymore as weird special cases) and we get rid of that IRQ_EXIT_OFFSET hack that fixes the CONFIG_PREEMPT case. I'm attaching an untested patch that modify yours. It's probably missing some corner cases that rely on in_interrupt() value in rcu_irq_exit() and tick_nohz_irq_exit() and may be other things.