Re: completion timeouts with pin-based interrupts in QEMU hw/nvme

[Guenter Roeck]

On 1/19/23 00:02, Klaus Jensen wrote:
> On Jan 19 08:28, Klaus Jensen wrote:
> > On Jan 18 21:03, Keith Busch wrote:
> > > On Thu, Jan 19, 2023 at 01:10:57PM +1000, Alistair Francis wrote:
> > > > On Thu, Jan 19, 2023 at 12:44 PM Keith Busch <kbusch@kernel.org> wrote:
> > > > > Further up, it says the "interrupt gateway" is responsible for
> > > > > forwarding new interrupt requests while the level remains asserted, but
> > > > > it doesn't look like anything is handling that, which essentially turns
> > > > > this into an edge interrupt. Am I missing something, or is this really
> > > > > not being handled?
> > > >
> > > > Yeah, that wouldn't be handled. In QEMU the PLIC relies on QEMUs
> > > > internal GPIO lines to trigger an interrupt. So with the current setup
> > > > we only support edge triggered interrupts.
> > >
> > > Thanks for confirming!
> > >
> > > Klaus,
> > > I think we can justify introducing a work-around in the emulated device
> > > now. My previous proposal with pci_irq_pulse() is no good since it does
> > > assert+deassert, but it needs to be the other way around, so please
> > > don't considert that one.
> > >
> > > Also, we ought to revisit the intms/intmc usage in the linux driver for
> > > threaded interrupts.
> >
> > +CC: qemu-riscv
> >
> > Keith,
> >
> > Thanks for digging into this!
> >
> > Yeah, you are probably right that we should only use the intms/intmc
> > changes in the use_threaded_interrupts case, not in general. While my
> > RFC patch does seem to "fix" this, it is just a workaround as your
> > analysis indicate. >
> +CC: Philippe,
>
> I am observing these timeouts/aborts on mips as well, so I guess that
> emulation could suffer from the same issue?

I suspect it does. In my case, I have an Ethernet controller on the same
interrupt line as the NVME controller, and it looks like one of the
interrupts gets lost if both controllers raise an interrupt at the same
time. The suggested workarounds "fix" the problem, but that doesn't seem
to be the correct solution.

Guenter