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

[Guenter Roeck]

On Tue, Jan 17, 2023 at 04:18:14PM +0000, Peter Maydell wrote:
> On Tue, 17 Jan 2023 at 16:10, Guenter Roeck <linux at roeck-us.net> wrote:
> >
> > On Mon, Jan 16, 2023 at 09:58:13PM -0700, Keith Busch wrote:
> > > On Mon, Jan 16, 2023 at 10:14:07PM +0100, Klaus Jensen wrote:
> > > > I noticed that the Linux driver does not use the INTMS/INTMC registers
> > > > to mask interrupts on the controller while processing CQEs. While not
> > > > required by the spec, it is *recommended* in setups not using MSI-X to
> > > > reduce the risk of spurious and/or missed interrupts.
> > >
> > > That's assuming completions are deferred to a bottom half. We don't do
> > > that by default in Linux nvme, though you can ask the driver to do that
> > > if you want.
> > >
> > > > With the patch below, running 100 boot iterations, no timeouts were
> > > > observed on QEMU emulated riscv64 or mips64.
> > > >
> > > > No changes are required in the QEMU hw/nvme interrupt logic.
> > >
> > > Yeah, I can see why: it forces the irq line to deassert then assert,
> > > just like we had forced to happen within the device side patches. Still,
> > > none of that is supposed to be necessary, but this idea of using these
> > > registers is probably fine.
> >
> > There is still no answer why this would be necessary in the first place,
> > on either side. In my opinion, unless someone can confirm that the problem
> > is seen with real hardware, we should assume that it happens on the qemu
> > side and address it there.
> 
> Sure, but that means identifying what the divergence
> between QEMU's implementation and the hardware is first. I don't
> want a fudged fix in QEMU's code any more than you want one in
> the kernel's driver code :-)
> 

I actually prefer it in qemu because that means I can test nvme support
on all active LTS releases of the Linux kernel, but that is POV and
secondary. This has been broken ever since I started testing nvme
support with qemu, so it doesn't make much of a difference if fixing
the problem for good takes a bit longer. Plus, I run my own version
of qemu anyway, so carrying the fix (hack) in qemu doesn't make much
of a difference for me.

Anyway - any idea what to do to help figuring out what is happening ?
Add tracing support to pci interrupt handling, maybe ?

Guenter