On Fri, Feb 28, 2020 at 1:57 AM Gerd Hoffmann <kraxel at redhat.com> wrote:> > On Thu, Feb 27, 2020 at 01:04:54PM -0800, Alistair Francis wrote: > > The QEMU model for the Bochs display has no VGA memory section at offset > > 0x400 [1]. By writing to this register Linux can create a write to > > unassigned memory which depending on machine and architecture can result > > in a store fault. > > > > I don't see any reference to this address at OSDev [2] or in the Bochs > > source code. > > > > Removing this write still allows graphics to work inside QEMU with > > the bochs-display. > > It's not that simple. The driver also handles the qemu stdvga (-device > VGA, -device secondary-vga) which *does* need the vga port write. > There is no way for the guest to figure whenever the device is > secondary-vga or bochs-display. > > So how about fixing things on the host side? Does qemu patch below > help?That patch looks like it will fix the problem, but it doesn't seem like the correct fix. I would rather avoid adding a large chunk of dummy I/O to handle the two devices.> > Maybe another possible approach is to enable/disable vga access per > arch. On x86 this doesn't cause any problems. I guess you are on > risc-v?I would prefer this option. I do see this on RISC-V, but I suspect the issue will appear on other architectures (although how they handle I/O failures in QEMU is a different story). Can I just do the VGA write if x86? Alistair> > cheers, > Gerd > > diff --git a/hw/display/bochs-display.c b/hw/display/bochs-display.c > index 62085f9fc063..e93e838243b8 100644 > --- a/hw/display/bochs-display.c > +++ b/hw/display/bochs-display.c > @@ -151,6 +151,26 @@ static const MemoryRegionOps bochs_display_qext_ops = { > .endianness = DEVICE_LITTLE_ENDIAN, > }; > > +static uint64_t dummy_read(void *ptr, hwaddr addr, unsigned size) > +{ > + return -1; > +} > + > +static void dummy_write(void *ptr, hwaddr addr, > + uint64_t val, unsigned size) > +{ > +} > + > +static const MemoryRegionOps dummy_ops = { > + .read = dummy_read, > + .write = dummy_write, > + .valid.min_access_size = 1, > + .valid.max_access_size = 4, > + .impl.min_access_size = 1, > + .impl.max_access_size = 1, > + .endianness = DEVICE_LITTLE_ENDIAN, > +}; > + > static int bochs_display_get_mode(BochsDisplayState *s, > BochsDisplayMode *mode) > { > @@ -284,8 +304,8 @@ static void bochs_display_realize(PCIDevice *dev, Error **errp) > memory_region_init_io(&s->qext, obj, &bochs_display_qext_ops, s, > "qemu extended regs", PCI_VGA_QEXT_SIZE); > > - memory_region_init(&s->mmio, obj, "bochs-display-mmio", > - PCI_VGA_MMIO_SIZE); > + memory_region_init_io(&s->mmio, obj, &dummy_ops, NULL, > + "bochs-display-mmio", PCI_VGA_MMIO_SIZE); > memory_region_add_subregion(&s->mmio, PCI_VGA_BOCHS_OFFSET, &s->vbe); > memory_region_add_subregion(&s->mmio, PCI_VGA_QEXT_OFFSET, &s->qext); > >
On Mon, Mar 02, 2020 at 02:14:02PM -0800, Alistair Francis wrote:> On Fri, Feb 28, 2020 at 1:57 AM Gerd Hoffmann <kraxel at redhat.com> wrote: > > > > On Thu, Feb 27, 2020 at 01:04:54PM -0800, Alistair Francis wrote: > > > The QEMU model for the Bochs display has no VGA memory section at offset > > > 0x400 [1]. By writing to this register Linux can create a write to > > > unassigned memory which depending on machine and architecture can result > > > in a store fault. > > > > > > I don't see any reference to this address at OSDev [2] or in the Bochs > > > source code. > > > > > > Removing this write still allows graphics to work inside QEMU with > > > the bochs-display. > > > > It's not that simple. The driver also handles the qemu stdvga (-device > > VGA, -device secondary-vga) which *does* need the vga port write. > > There is no way for the guest to figure whenever the device is > > secondary-vga or bochs-display. > > > > So how about fixing things on the host side? Does qemu patch below > > help? > > That patch looks like it will fix the problem, but it doesn't seem > like the correct fix. I would rather avoid adding a large chunk of > dummy I/O to handle the two devices.It's just a single handler for the parent mmio region, so we have a defined default action instead of undefined behavior. Patch just posted to qemu-devel, lets see what others think ...> > Maybe another possible approach is to enable/disable vga access per > > arch. On x86 this doesn't cause any problems. I guess you are on > > risc-v? > > I would prefer this option. I do see this on RISC-V, but I suspect the > issue will appear on other architectures (although how they handle I/O > failures in QEMU is a different story). > > Can I just do the VGA write if x86?I know ppc needs it too. Not sure about other architectures. I'd suggest to do it the other way around: blacklist known-problematic archs. cheers, Gerd
On Mon, Mar 2, 2020 at 10:24 PM Gerd Hoffmann <kraxel at redhat.com> wrote:> > On Mon, Mar 02, 2020 at 02:14:02PM -0800, Alistair Francis wrote: > > On Fri, Feb 28, 2020 at 1:57 AM Gerd Hoffmann <kraxel at redhat.com> wrote: > > > > > > On Thu, Feb 27, 2020 at 01:04:54PM -0800, Alistair Francis wrote: > > > > The QEMU model for the Bochs display has no VGA memory section at offset > > > > 0x400 [1]. By writing to this register Linux can create a write to > > > > unassigned memory which depending on machine and architecture can result > > > > in a store fault. > > > > > > > > I don't see any reference to this address at OSDev [2] or in the Bochs > > > > source code. > > > > > > > > Removing this write still allows graphics to work inside QEMU with > > > > the bochs-display. > > > > > > It's not that simple. The driver also handles the qemu stdvga (-device > > > VGA, -device secondary-vga) which *does* need the vga port write. > > > There is no way for the guest to figure whenever the device is > > > secondary-vga or bochs-display. > > > > > > So how about fixing things on the host side? Does qemu patch below > > > help? > > > > That patch looks like it will fix the problem, but it doesn't seem > > like the correct fix. I would rather avoid adding a large chunk of > > dummy I/O to handle the two devices. > > It's just a single handler for the parent mmio region, so we have a > defined default action instead of undefined behavior. > > Patch just posted to qemu-devel, lets see what others think ...Thanks, let's wait and see what happens.> > > > Maybe another possible approach is to enable/disable vga access per > > > arch. On x86 this doesn't cause any problems. I guess you are on > > > risc-v? > > > > I would prefer this option. I do see this on RISC-V, but I suspect the > > issue will appear on other architectures (although how they handle I/O > > failures in QEMU is a different story). > > > > Can I just do the VGA write if x86? > > I know ppc needs it too. Not sure about other architectures. I'd > suggest to do it the other way around: blacklist known-problematic > archs.Argh, that is a little uglier. Let's circle back after receiving feedback on the QEMU patch. Alistair> > cheers, > Gerd >