Re: [Qemu-devel] [PATCH 11/22] i8259: Update IRQ state after reset

[Peter Maydell]

On 30 September 2011 21:47, Blue Swirl <address@hidden> wrote:
> If reliable post reset effects are needed, reset should be divided
> into two separate phases. So far no interesting cases have been
> demonstrated.

I've already cited one in this thread [realview mmc card present]
and that's just the one I happen to know about because I made some
patches in that area. I bet there are more in the codebase that
we're not handling properly; they're not easy to find because they
depend on an interaction between how two devices happen to be wired
up rather than being a bug in a particular device.

Another random example: lan9118_reset() calls phy_reset() calls
phy_update_link() calls phy_update_irq() calls lan9118_update()
calls qemu_set_irq(). That's a bug at the moment (ie I agree
that with qdev as it is currently trying to assert outgoing
gpio lines in the reset callback is broken), but we do somehow
need to drive the irq line high on reset (both of the
"write-to-register" triggered kind and the plain old power-on
kind) because the signal defaults to active-low on reset.
(we can't just say our model of the irq line tracks the logical
state of the line, because there's a device config register
that lets you flip it from active-low to active-high, so
our model of the irq line has to have the same sense that the
interrupt controller expects.)

-- PMM