On Thu, Mar 12, 2015 at 10:21:47AM +0000, Peter Maydell wrote: > On 12 March 2015 at 10:16, Michael S. Tsirkin <m...@redhat.com> wrote: > > So thinking about this more, by the time kdump tries to reset device, > > linux has probably already disabled the IRQ at the APIC level. > > Isn't that the case? If so, the patch won't help, will it? > > Trying to deassert (or worse, assert) interrupt lines in > device reset functions is slightly bogus, yes. In general > the theory is that the interrupt controller the interrupt line > is connected to should have its own reset handling which > treats the line as going back to deasserted, because there's > no guarantee made about which of the two ends of the line > gets its reset handler called first. > > Things are not really this neat in practice though. (There's > no good way to model a device which comes out of reset with > a line asserted, for instance.) > > -- PMM
This isn't a device reset though. The function that Fam is touching is called when a special "virtio reset" register to poked by the driver. It only resets part of the device, not all of it, and it seems reasonable to ask that it clear the interrupt. So I think the patch is correct, even if just for spec compliance reasons, but I would like to find out and document the workloads that actually benefit. -- MST
