while porting the -rt tree to 2.6.18-rc7 i noticed the following
screaming-IRQ scenario on an SMP system:

 2274  0Dn.:1 0.001ms: do_IRQ+0xc/0x103  <= (ret_from_intr+0x0/0xf)
 2274  0Dn.:1 0.010ms: do_IRQ+0xc/0x103  <= (ret_from_intr+0x0/0xf)
 2274  0Dn.:1 0.020ms: do_IRQ+0xc/0x103  <= (ret_from_intr+0x0/0xf)
 2274  0Dn.:1 0.029ms: do_IRQ+0xc/0x103  <= (ret_from_intr+0x0/0xf)
 2274  0Dn.:1 0.039ms: do_IRQ+0xc/0x103  <= (ret_from_intr+0x0/0xf)
 2274  0Dn.:1 0.048ms: do_IRQ+0xc/0x103  <= (ret_from_intr+0x0/0xf)
 2274  0Dn.:1 0.058ms: do_IRQ+0xc/0x103  <= (ret_from_intr+0x0/0xf)
 2274  0Dn.:1 0.068ms: do_IRQ+0xc/0x103  <= (ret_from_intr+0x0/0xf)
 2274  0Dn.:1 0.077ms: do_IRQ+0xc/0x103  <= (ret_from_intr+0x0/0xf)
 2274  0Dn.:1 0.087ms: do_IRQ+0xc/0x103  <= (ret_from_intr+0x0/0xf)
 2274  0Dn.:1 0.097ms: do_IRQ+0xc/0x103  <= (ret_from_intr+0x0/0xf)

as it turns out, the bug is caused by handle_level_irq(), which if it
races with another CPU already handling this IRQ, it _unmasks_ the IRQ
line on the way out. This is not how 2.6.17 works, and we introduced
this bug in one of the early genirq cleanups right before it went into
-mm. (the bug was not in the genirq patchset for a long time, and we
didnt notice the bug due to the lack of -rt rebase to the new genirq
code. -rt, and hardirq-preemption in particular opens up such races much
wider than anything else.)
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Acked-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>