From: Vitaly Kuznetsov <vkuznets@xxxxxxxxxx> Sent: Thursday, August 25, 2022 2:00 AM > > Passed through PCI device sometimes misbehave on Gen1 VMs when Hyper-V > DRM driver is also loaded. Looking at IOMEM assignment, we can see e.g. > > $ cat /proc/iomem > ... > f8000000-fffbffff : PCI Bus 0000:00 > f8000000-fbffffff : 0000:00:08.0 > f8000000-f8001fff : bb8c4f33-2ba2-4808-9f7f-02f3b4da22fe > ... > fe0000000-fffffffff : PCI Bus 0000:00 > fe0000000-fe07fffff : bb8c4f33-2ba2-4808-9f7f-02f3b4da22fe > fe0000000-fe07fffff : 2ba2:00:02.0 > fe0000000-fe07fffff : mlx4_core > > the interesting part is the 'f8000000' region as it is actually the > VM's framebuffer: > > $ lspci -v > ... > 0000:00:08.0 VGA compatible controller: Microsoft Corporation Hyper-V virtual VGA > (prog-if 00 [VGA controller]) > Flags: bus master, fast devsel, latency 0, IRQ 11 > Memory at f8000000 (32-bit, non-prefetchable) [size=64M] > ... > > hv_vmbus: registering driver hyperv_drm > hyperv_drm 5620e0c7-8062-4dce-aeb7-520c7ef76171: [drm] Synthvid Version major 3, minor 5 > hyperv_drm 0000:00:08.0: vgaarb: deactivate vga console > hyperv_drm 0000:00:08.0: BAR 0: can't reserve [mem 0xf8000000-0xfbffffff] > hyperv_drm 5620e0c7-8062-4dce-aeb7-520c7ef76171: [drm] Cannot request framebuffer, boot fb still active? > > Note: "Cannot request framebuffer" is not a fatal error in > hyperv_setup_gen1() as the code assumes there's some other framebuffer > device there but we actually have some other PCI device (mlx4 in this > case) config space there! My apologies for not getting around to commenting on the previous version of this patch. The function hyperv_setup_gen1() and the "Cannot request framebuffer" message have gone away as of commit a0ab5abced55. > > The problem appears to be that vmbus_allocate_mmio() can allocate from > the reserved framebuffer region (fb_overlap_ok), however, if the > request to allocate MMIO comes from some other device before > framebuffer region is taken, it can happily use framebuffer region for > it. Interesting. I had never looked at the details of vmbus_allocate_mmio(). The semantics one might assume of a parameter named "fb_overlap_ok" aren't implemented because !fb_overlap_ok essentially has no effect. The existing semantics are really "prefer_fb_overlap". This patch implements the expected and needed semantics, which is to not allocate from the frame buffer space when !fb_overlap_ok. If that's an accurate high level summary, maybe this commit message could describe it that way? The other details you provide about what can go wrong should still be included as well. > Note, Gen2 VMs are usually unaffected by the issue because > framebuffer region is already taken by EFI fb (in case kernel supports > it) but Gen1 VMs may have this region unclaimed by the time Hyper-V PCI > pass-through driver tries allocating MMIO space if Hyper-V DRM/FB drivers > load after it. Devices can be brought up in any sequence so let's > resolve the issue by always ignoring 'fb_mmio' region for non-FB > requests, even if the region is unclaimed. > > Signed-off-by: Vitaly Kuznetsov <vkuznets@xxxxxxxxxx> > --- > drivers/hv/vmbus_drv.c | 10 +++++++++- > 1 file changed, 9 insertions(+), 1 deletion(-) > > diff --git a/drivers/hv/vmbus_drv.c b/drivers/hv/vmbus_drv.c > index 536f68e563c6..3c833ea60db6 100644 > --- a/drivers/hv/vmbus_drv.c > +++ b/drivers/hv/vmbus_drv.c > @@ -2331,7 +2331,7 @@ int vmbus_allocate_mmio(struct resource **new, struct > hv_device *device_obj, > bool fb_overlap_ok) > { > struct resource *iter, *shadow; > - resource_size_t range_min, range_max, start; > + resource_size_t range_min, range_max, start, end; > const char *dev_n = dev_name(&device_obj->device); > int retval; > > @@ -2366,6 +2366,14 @@ int vmbus_allocate_mmio(struct resource **new, struct > hv_device *device_obj, > range_max = iter->end; > start = (range_min + align - 1) & ~(align - 1); > for (; start + size - 1 <= range_max; start += align) { > + end = start + size - 1; > + > + /* Skip the whole fb_mmio region if not fb_overlap_ok */ > + if (!fb_overlap_ok && fb_mmio && > + (((start >= fb_mmio->start) && (start <= fb_mmio->end)) || > + ((end >= fb_mmio->start) && (end <= fb_mmio->end)))) > + continue; > + > shadow = __request_region(iter, start, size, NULL, > IORESOURCE_BUSY); > if (!shadow) > -- > 2.37.1 Other than my musings on the commit message, Reviewed-by: Michael Kelley <mikelley@xxxxxxxxxxxxx>