Re: [PATCH 7/9] videobuf2/videobuf2-dma-sg.c: Convert to use vm_insert_range_buggy

[Marek Szyprowski]

Hi Souptick,

On 2019-01-25 05:55, Souptick Joarder wrote:
> On Tue, Jan 22, 2019 at 8:37 PM Marek Szyprowski
> <m.szyprowski@xxxxxxxxxxx> wrote:
>> On 2019-01-11 16:11, Souptick Joarder wrote:
>>> Convert to use vm_insert_range_buggy to map range of kernel memory
>>> to user vma.
>>>
>>> This driver has ignored vm_pgoff. We could later "fix" these drivers
>>> to behave according to the normal vm_pgoff offsetting simply by
>>> removing the _buggy suffix on the function name and if that causes
>>> regressions, it gives us an easy way to revert.
>> Just a generic note about videobuf2: videobuf2-dma-sg is ignoring vm_pgoff by design. vm_pgoff is used as a 'cookie' to select a buffer to mmap and videobuf2-core already checks that. If userspace provides an offset, which doesn't match any of the registered 'cookies' (reported to userspace via separate v4l2 ioctl), an error is returned.
> Ok, it means once the buf is selected, videobuf2-dma-sg should always
> mapped buf->pages[i]
> from index 0 ( irrespective of vm_pgoff value). So although we are
> replacing the code with
> vm_insert_range_buggy(), *_buggy* suffix will mislead others and
> should not be used.
> And if we replace this code with  vm_insert_range(), this will
> introduce bug for *non zero*
> value of vm_pgoff.
>
> Please correct me if my understanding is wrong.

You are correct. IMHO the best solution in this case would be to add
following fix:


diff --git a/drivers/media/common/videobuf2/videobuf2-core.c
b/drivers/media/common/videobuf2/videobuf2-core.c
index 70e8c3366f9c..ca4577a7d28a 100644
--- a/drivers/media/common/videobuf2/videobuf2-core.c
+++ b/drivers/media/common/videobuf2/videobuf2-core.c
@@ -2175,6 +2175,13 @@ int vb2_mmap(struct vb2_queue *q, struct
vm_area_struct *vma)
ÂÂÂÂ ÂÂÂ goto unlock;
ÂÂÂÂ }
Â
+ÂÂÂ /*
+ÂÂÂ Â* vm_pgoff is treated in V4L2 API as a 'cookie' to select a buffer,
+ÂÂÂ Â* not as a in-buffer offset. We always want to mmap a whole buffer
+ÂÂÂ Â* from its beginning.
+ÂÂÂ Â*/
+ÂÂÂ vma->vm_pgoff = 0;
+
ÂÂÂÂ ret = call_memop(vb, mmap, vb->planes[plane].mem_priv, vma);
Â
Âunlock:
diff --git a/drivers/media/common/videobuf2/videobuf2-dma-contig.c
b/drivers/media/common/videobuf2/videobuf2-dma-contig.c
index aff0ab7bf83d..46245c598a18 100644
--- a/drivers/media/common/videobuf2/videobuf2-dma-contig.c
+++ b/drivers/media/common/videobuf2/videobuf2-dma-contig.c
@@ -186,12 +186,6 @@ static int vb2_dc_mmap(void *buf_priv, struct
vm_area_struct *vma)
ÂÂÂÂ ÂÂÂ return -EINVAL;
ÂÂÂÂ }
Â
-ÂÂÂ /*
-ÂÂÂ Â* dma_mmap_* uses vm_pgoff as in-buffer offset, but we want to
-ÂÂÂ Â* map whole buffer
-ÂÂÂ Â*/
-ÂÂÂ vma->vm_pgoff = 0;
-
ÂÂÂÂ ret = dma_mmap_attrs(buf->dev, vma, buf->cookie,
ÂÂÂÂ ÂÂÂ buf->dma_addr, buf->size, buf->attrs);
Â
-- 

Then you can simply use non-buggy version of your function in
drivers/media/common/videobuf2/videobuf2-dma-sg.c.

I can send above as a formal patch if you want.

> So what your opinion about this patch ? Shall I drop this patch from
> current series ?
> or,
> There is any better way to handle this scenario ?
>
>
>>> There is an existing bug inside gem_mmap_obj(), where user passed
>>> length is not checked against buf->num_pages. For any value of
>>> length > buf->num_pages it will end up overrun buf->pages[i],
>>> which could lead to a potential bug.
> It is not gem_mmap_obj(), it should be vb2_dma_sg_mmap().
> Sorry about it.
>
> What about this issue ? Does it looks like a valid issue ?

It is already handled in vb2_mmap(). Such call will be rejected.


> ...

Best regards
-- 
Marek Szyprowski, PhD
Samsung R&D Institute Poland