2018-10-18 17:14:30

by Wenwen Wang

[permalink] [raw]
Subject: [PATCH] drm/radeon: fix a missing-check bug

In radeon_read_bios(), the bios rom is firstly mapped to the IO memory
region 'bios' through pci_map_rom(). Then the first two bytes of 'bios' are
copied to 'val1' and 'val2' respectively through readb(). After that,
'val1' and 'val2' are checked to see whether they have expected values,
i.e., 0x55 and 0xaa, respectively. If yes, the whole data in 'bios' is then
copied to 'rdev->bios' through memcpy_fromio(). Obviously, the first two
bytes in 'bios' are copied twice. More importantly, no check is enforced on
the first two bytes of 'rdev->bios' after memcpy_fromio(). Given that the
IO memory region can also be accessed by the device, it is possible that a
malicious device can race to modify these two bytes between the two copies
and thus after memcpy_fromio(), the first two bytes in 'rdev->bios' can
have unexpected values. This can cause undefined behavior of the kernel
and introduce potential security risk, if the device can be controlled by
attackers.

This patch rewrites the first two bytes of 'rdev->bios' after
memcpy_fromio() with expected values. Through this way, the above issue can
be avoided.

Signed-off-by: Wenwen Wang <[email protected]>
---
drivers/gpu/drm/radeon/radeon_bios.c | 2 ++
1 file changed, 2 insertions(+)

diff --git a/drivers/gpu/drm/radeon/radeon_bios.c b/drivers/gpu/drm/radeon/radeon_bios.c
index 04c0ed4..f336719 100644
--- a/drivers/gpu/drm/radeon/radeon_bios.c
+++ b/drivers/gpu/drm/radeon/radeon_bios.c
@@ -98,6 +98,8 @@ static bool radeon_read_bios(struct radeon_device *rdev)
return false;
}
memcpy_fromio(rdev->bios, bios, size);
+ rdev->bios[0] = val1;
+ rdev->bios[1] = val2;
pci_unmap_rom(rdev->pdev, bios);
return true;
}
--
2.7.4



2018-10-19 07:03:04

by Christian König

[permalink] [raw]
Subject: Re: [PATCH] drm/radeon: fix a missing-check bug

Am 18.10.18 um 19:13 schrieb Wenwen Wang:
> In radeon_read_bios(), the bios rom is firstly mapped to the IO memory
> region 'bios' through pci_map_rom(). Then the first two bytes of 'bios' are
> copied to 'val1' and 'val2' respectively through readb(). After that,
> 'val1' and 'val2' are checked to see whether they have expected values,
> i.e., 0x55 and 0xaa, respectively. If yes, the whole data in 'bios' is then
> copied to 'rdev->bios' through memcpy_fromio(). Obviously, the first two
> bytes in 'bios' are copied twice. More importantly, no check is enforced on
> the first two bytes of 'rdev->bios' after memcpy_fromio(). Given that the
> IO memory region can also be accessed by the device, it is possible that a
> malicious device can race to modify these two bytes between the two copies
> and thus after memcpy_fromio(), the first two bytes in 'rdev->bios' can
> have unexpected values. This can cause undefined behavior of the kernel
> and introduce potential security risk, if the device can be controlled by
> attackers.
>
> This patch rewrites the first two bytes of 'rdev->bios' after
> memcpy_fromio() with expected values. Through this way, the above issue can
> be avoided.

Well NAK, that doesn't make any sense to me.

First of all we don't map VRAM, but rather the ROM which is a read only
flash. Writing to the flash is not supported through the BAR as far as I
know.

Then we check the first two bytes to make sure that the ROM is correctly
mapped and not to prevent any malicious attacks.

Regards,
Christian.

>
> Signed-off-by: Wenwen Wang <[email protected]>
> ---
> drivers/gpu/drm/radeon/radeon_bios.c | 2 ++
> 1 file changed, 2 insertions(+)
>
> diff --git a/drivers/gpu/drm/radeon/radeon_bios.c b/drivers/gpu/drm/radeon/radeon_bios.c
> index 04c0ed4..f336719 100644
> --- a/drivers/gpu/drm/radeon/radeon_bios.c
> +++ b/drivers/gpu/drm/radeon/radeon_bios.c
> @@ -98,6 +98,8 @@ static bool radeon_read_bios(struct radeon_device *rdev)
> return false;
> }
> memcpy_fromio(rdev->bios, bios, size);
> + rdev->bios[0] = val1;
> + rdev->bios[1] = val2;
> pci_unmap_rom(rdev->pdev, bios);
> return true;
> }