[PATCH 1/1] drm/hyperv: Fix address space leak when Hyper-V DRM device is removed

[Michael Kelley]

From: Saurabh Singh Sengar <ssengar at linux.microsoft.com> Sent: Monday, February 10, 2025 7:33 PM
> 
> On Mon, Feb 10, 2025 at 11:34:41AM -0800, mhkelley58 at gmail.com wrote:
> > From: Michael Kelley <mhklinux at outlook.com>
> >
> > When a Hyper-V DRM device is probed, the driver allocates MMIO space for
> > the vram, and maps it cacheable. If the device removed, or in the error
> > path for device probing, the MMIO space is released but no unmap is done.
> > Consequently the kernel address space for the mapping is leaked.
> >
> > Fix this by adding iounmap() calls in the device removal path, and in the
> > error path during device probing.
> >
> > Fixes: f1f63cbb705d ("drm/hyperv: Fix an error handling path in hyperv_vmbus_probe()")
> > Fixes: a0ab5abced55 ("drm/hyperv : Removing the restruction of VRAM allocation with PCI bar size")
> > Signed-off-by: Michael Kelley <mhklinux at outlook.com>
> > ---
> >  drivers/gpu/drm/hyperv/hyperv_drm_drv.c | 2 ++
> >  1 file changed, 2 insertions(+)
> >
> > diff --git a/drivers/gpu/drm/hyperv/hyperv_drm_drv.c
> b/drivers/gpu/drm/hyperv/hyperv_drm_drv.c
> > index e0953777a206..b491827941f1 100644
> > --- a/drivers/gpu/drm/hyperv/hyperv_drm_drv.c
> > +++ b/drivers/gpu/drm/hyperv/hyperv_drm_drv.c
> > @@ -156,6 +156,7 @@ static int hyperv_vmbus_probe(struct hv_device *hdev,
> >  	return 0;
> >
> >  err_free_mmio:
> > +	iounmap(hv->vram);
> >  	vmbus_free_mmio(hv->mem->start, hv->fb_size);
> >  err_vmbus_close:
> >  	vmbus_close(hdev->channel);
> > @@ -174,6 +175,7 @@ static void hyperv_vmbus_remove(struct hv_device *hdev)
> >  	vmbus_close(hdev->channel);
> >  	hv_set_drvdata(hdev, NULL);
> >
> > +	iounmap(hv->vram);
> >  	vmbus_free_mmio(hv->mem->start, hv->fb_size);
> >  }
> >
> > --
> > 2.25.1
> >
> 
> Thanks for the fix. May I know how do you find such issues ?

I think it was that I was looking at the Hyper-V FB driver for the
vmbus_free_mmio() call sites, and realizing that such call sites
should probably also have an associated iounmap(). Then I was
looking at the same thing in the Hyper-V DRM driver, and
realizing there were no calls to iounmap()!

To confirm, the contents of /proc/vmallocinfo can be filtered
for ioremap calls with size 8 MiB (which actually show up as
8 MiB + 4KiB because the address space allocator adds a guard
page to each allocation). When doing repeated unbind/bind
sequences on the DRM driver, those 8 MiB entries in
/proc/vmallocinfo kept accumulating and were never freed.

Michael

> 
> Reviewed-by: Saurabh Sengar <ssengar at linux.microsoft.com>
> Tested-by: Saurabh Sengar <ssengar at linux.microsoft.com>
> 
> - Saurabh