[Qemu-devel] SCSI bus failures with qemu-arm in kernel 3.8+

[Guenter Roeck]

On Thu, Aug 15, 2013 at 07:05:22PM +0100, Peter Maydell wrote:
> On 15 August 2013 18:54, Guenter Roeck <linux at roeck-us.net> wrote:
> > On Thu, Aug 15, 2013 at 05:45:42PM +0100, Peter Maydell wrote:
> >> On 13 August 2013 04:40, Guenter Roeck <linux at roeck-us.net> wrote:
> >> > Patch tested and working with qemu 1.5.2, using the configuration file
> >> > from the yocto project. Patch applied on top of kernel version 3.11-rc5.
> >>
> >> OK, I tested this on PB926+PCI backplane hardware, and it is
> >> definitely better than current mainline, in that the test USB
> >> card that I have no longer causes the kernel to generate this sort of
> >> backtrace:
> >>
> > Do you mean my patch fixes the traceback below as a side effect ?
> > Would be great ... it would be one more reason to get it applied.
> 
> Yes, exactly -- the kernel currently has the wrong irq mapping,
> which causes the traceback (ie h/w asserts irq 93 but the kernel
> is listening on something else). That the patch fixes this confirms
> that it is the behaviour of hardware as well as of QEMU.
> 
> >> However it still doesn't seem to reliably detect the USB harddisk
> >> plugged into the card, so I think there may be further issues, possibly
> >> some subset of those Arnd identified and fixed with this patch:
> >> http://permalink.gmane.org/gmane.linux.ports.arm.kernel/93397
> >>
> > Does it get better if you apply Arnd's patch ?
> 
> Arnd's patch is ancient and won't apply as is (due to intervening
> changes and also because some of the things it fixes were fixed
> in later patches); I'm currently trying to extract the relevant parts.
> 
> If you want you can confirm that I/O port PCI access is broken on
> QEMU too -- disable CONFIG_SCSI_SYM53C8XX_MMIO so
> the driver uses PCI IO rather than MMIO and you'll see QEMU's
> SCSI device doesn't work any more.
> 
> >> so I'd like to continue testing.
> >>
> >> The other thing this patch should (IMHO) have is the
> >> line in pci_versatile_setup() which tells QEMU that the
> >> kernel really does expect hardware-like behaviour:
> 
> >> (Without this line QEMU will guess whether the kernel is broken
> >> or not and will get it right most but not necessarily all of the time.)
> >>
> > Might make sense, but I think it should be a separate patch.
> 
> It needs to go in the same patch, because a kernel with the fixed
> irq remapping must also tell QEMU it is fixed; if you split the
> two then at the point between the two patches the kernel is
> broken for bisection purposes.
> 
Thinking about it - is that really true ? My image with the patch applied
works just fine under qemu 1.5.2, and unless I am missing something it won't
work with qemu 1.4 anyway. So what exactly is broken ?

Thanks,
Guenter