FW: [PATCH v5 0/8] Add MT8173 Video Encoder Driver and VPU Driver

Commit Message

tiffany.lin March 14, 2016, 9:04 a.m. UTC

On Mon, 2016-03-14 at 08:21 +0100, Hans Verkuil wrote:
> On 03/14/2016 08:12 AM, tiffany lin wrote:
> > Hi Hans,
> > 
> > After change to use "v4l-utils.git master branch", "V4l2-compliance
> > -d /dev/video1" fail on "fail: v4l2-test-buffers.cpp(555):
> > check_0(crbufs.reserved, sizeof(crbufs.reserved))".
> > 
> > Check the source code and found
> > 
> > 	memset(&crbufs, 0xff, sizeof(crbufs));   -> crbufs to 0xff
> >         node->g_fmt(crbufs.format, i);
> >         crbufs.count = 0;
> >         crbufs.memory = m;
> >         fail_on_test(doioctl(node, VIDIOC_CREATE_BUFS, &crbufs));   
> >         fail_on_test(check_0(crbufs.reserved, sizeof(crbufs.reserved)));
> >         fail_on_test(crbufs.index != q.g_buffers());
> > 
> > crbufs is initialized to fill with 0xff and after VIDIOC_CREATE_BUFS,
> > crbufs.reserved field should be 0x0. But v4l2_m2m_create_bufs and
> > vb2_create_bufs do not process reserved filed.
> > Do we really need to check reserved filed filled with 0x0? Or we need to
> > change vb2_create_bufs to fix this issue?
> 
> The reserved field is zeroed in v4l_create_bufs() in v4l2-ioctl.c, so even before
> vb2_create_bufs et al is called.
> 
> The fact that it is no longer zeroed afterwards suggests that someone is messing
> with the reserved field.
> 
> You'll have to do a bit more digging, I'm afraid.
> 
Hi Hans,

Thanks for your information.
I found the root cause is in "put_v4l2_create32".
It do not copy reserved field from kernel space to user space.
After modification,"test VIDIOC_REQBUFS/CREATE_BUFS/QUERYBUF: OK"

format)) ||
+           copy_to_user(up->reserved, kp->reserved,
sizeof(kp->reserved)))
                return -EFAULT;
        return __put_v4l2_format32(&kp->format, &up->format);
 }

best regards,
Tiffany

> Regards,
> 
> 	Hans

Comments

Hans Verkuil March 14, 2016, 9:10 a.m. UTC | #1

On 03/14/2016 10:04 AM, tiffany lin wrote:
> On Mon, 2016-03-14 at 08:21 +0100, Hans Verkuil wrote:
>> On 03/14/2016 08:12 AM, tiffany lin wrote:
>>> Hi Hans,
>>>
>>> After change to use "v4l-utils.git master branch", "V4l2-compliance
>>> -d /dev/video1" fail on "fail: v4l2-test-buffers.cpp(555):
>>> check_0(crbufs.reserved, sizeof(crbufs.reserved))".
>>>
>>> Check the source code and found
>>>
>>> 	memset(&crbufs, 0xff, sizeof(crbufs));   -> crbufs to 0xff
>>>         node->g_fmt(crbufs.format, i);
>>>         crbufs.count = 0;
>>>         crbufs.memory = m;
>>>         fail_on_test(doioctl(node, VIDIOC_CREATE_BUFS, &crbufs));   
>>>         fail_on_test(check_0(crbufs.reserved, sizeof(crbufs.reserved)));
>>>         fail_on_test(crbufs.index != q.g_buffers());
>>>
>>> crbufs is initialized to fill with 0xff and after VIDIOC_CREATE_BUFS,
>>> crbufs.reserved field should be 0x0. But v4l2_m2m_create_bufs and
>>> vb2_create_bufs do not process reserved filed.
>>> Do we really need to check reserved filed filled with 0x0? Or we need to
>>> change vb2_create_bufs to fix this issue?
>>
>> The reserved field is zeroed in v4l_create_bufs() in v4l2-ioctl.c, so even before
>> vb2_create_bufs et al is called.
>>
>> The fact that it is no longer zeroed afterwards suggests that someone is messing
>> with the reserved field.
>>
>> You'll have to do a bit more digging, I'm afraid.
>>
> Hi Hans,
> 
> Thanks for your information.
> I found the root cause is in "put_v4l2_create32".
> It do not copy reserved field from kernel space to user space.
> After modification,"test VIDIOC_REQBUFS/CREATE_BUFS/QUERYBUF: OK"
> 
> diff --git a/drivers/media/v4l2-core/v4l2-compat-ioctl32.c
> b/drivers/media/v4l2-core/v4l2-compat-ioctl32.c
> index f38c076..109f687 100644
> --- a/drivers/media/v4l2-core/v4l2-compat-ioctl32.c
> +++ b/drivers/media/v4l2-core/v4l2-compat-ioctl32.c
> @@ -280,7 +280,8 @@ static int put_v4l2_format32(struct v4l2_format *kp,
> struct v4l2_format32 __user
>  static int put_v4l2_create32(struct v4l2_create_buffers *kp, struct
> v4l2_create_buffers32 __user *up)
>  {
>         if (!access_ok(VERIFY_WRITE, up, sizeof(struct
> v4l2_create_buffers32)) ||
> -           copy_to_user(up, kp, offsetof(struct v4l2_create_buffers32,
> format)))
> +           copy_to_user(up, kp, offsetof(struct v4l2_create_buffers32,
> format)) ||
> +           copy_to_user(up->reserved, kp->reserved,
> sizeof(kp->reserved)))
>                 return -EFAULT;
>         return __put_v4l2_format32(&kp->format, &up->format);
>  }

Yup, that's the cause. Can you post this as a 'proper' patch to the mailinglist?

I'll take it for kernel 4.6 (and I'll add a CC to the stable mailinglist to
get it backported as well).

Thanks!

	Hans

Patch

diff --git a/drivers/media/v4l2-core/v4l2-compat-ioctl32.c
b/drivers/media/v4l2-core/v4l2-compat-ioctl32.c
index f38c076..109f687 100644
--- a/drivers/media/v4l2-core/v4l2-compat-ioctl32.c
+++ b/drivers/media/v4l2-core/v4l2-compat-ioctl32.c
@@ -280,7 +280,8 @@  static int put_v4l2_format32(struct v4l2_format *kp,
struct v4l2_format32 __user
 static int put_v4l2_create32(struct v4l2_create_buffers *kp, struct
v4l2_create_buffers32 __user *up)
 {
        if (!access_ok(VERIFY_WRITE, up, sizeof(struct
v4l2_create_buffers32)) ||
-           copy_to_user(up, kp, offsetof(struct v4l2_create_buffers32,
format)))
+           copy_to_user(up, kp, offsetof(struct v4l2_create_buffers32,