| Message ID | YOfu5Zb2kpN/0Prn@mwanda (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | media: rockchip/rga: fix a use after free in rga_probe() | expand |
On Fri, Jul 9, 2021 at 2:38 PM Dan Carpenter <dan.carpenter@oracle.com> wrote: > > The video_device_release() frees the "vfd" pointer so passing it to > video_unregister_device() on the next line results in a use after free. > Calling video_unregister_device() on a device that hasn't been > registered is supposed to be a no-op so that can be removed. The paths > with to goto unreg_video_dev have a memory leak and should be updated to > goto rel_vdev instead. > > Fixes: f7e7b48e6d79 ("[media] rockchip/rga: v4l2 m2m support") > Signed-off-by: Dan Carpenter <dan.carpenter@oracle.com> > --- > drivers/media/platform/rockchip/rga/rga.c | 6 ++---- > 1 file changed, 2 insertions(+), 4 deletions(-) > > diff --git a/drivers/media/platform/rockchip/rga/rga.c b/drivers/media/platform/rockchip/rga/rga.c > index bf3fd71ec3af..37f7fd060c38 100644 > --- a/drivers/media/platform/rockchip/rga/rga.c > +++ b/drivers/media/platform/rockchip/rga/rga.c > @@ -863,12 +863,12 @@ static int rga_probe(struct platform_device *pdev) > if (IS_ERR(rga->m2m_dev)) { > v4l2_err(&rga->v4l2_dev, "Failed to init mem2mem device\n"); > ret = PTR_ERR(rga->m2m_dev); > - goto unreg_video_dev; > + goto rel_vdev; > } > > ret = pm_runtime_resume_and_get(rga->dev); > if (ret < 0) > - goto unreg_video_dev; > + goto rel_vdev; > > rga->version.major = (rga_read(rga, RGA_VERSION_INFO) >> 24) & 0xFF; > rga->version.minor = (rga_read(rga, RGA_VERSION_INFO) >> 20) & 0x0F; > @@ -904,8 +904,6 @@ static int rga_probe(struct platform_device *pdev) > > rel_vdev: > video_device_release(vfd); > -unreg_video_dev: > - video_unregister_device(rga->vfd); > unreg_v4l2_dev: > v4l2_device_unregister(&rga->v4l2_dev); > err_put_clk: From the analysis of rga_probe and rga_remove function, the init and cleanup functions are in pairs as follows: pm_runtime_enable(rga->dev); ret = v4l2_device_register(&pdev->dev, &rga->v4l2_dev); vfd = video_device_alloc(); rga->m2m_dev = v4l2_m2m_init(&rga_m2m_ops); rga->cmdbuf_virt = dma_alloc_attrs(rga->dev, RGA_CMDBUF_SIZE, &rga->cmdbuf_phy, GFP_KERNEL, DMA_ATTR_WRITE_COMBINE); rga->src_mmu_pages = (unsigned int *)__get_free_pages(GFP_KERNEL | __GFP_ZERO, 3); rga->dst_mmu_pages = (unsigned int *)__get_free_pages(GFP_KERNEL | __GFP_ZERO, 3); ret = video_register_device(vfd, VFL_TYPE_VIDEO, -1); ...... video_unregister_device(rga->vfd); free_pages((unsigned long)rga->src_mmu_pages, 3); free_pages((unsigned long)rga->dst_mmu_pages, 3); dma_free_attrs(rga->dev, RGA_CMDBUF_SIZE, rga->cmdbuf_virt, rga->cmdbuf_phy, DMA_ATTR_WRITE_COMBINE); v4l2_m2m_release(rga->m2m_dev); video_device_release(rga->vfd); v4l2_device_unregister(&rga->v4l2_dev); pm_runtime_disable(rga->dev); So I think the error handling code is more complex than this patch. ;( > -- > 2.30.2 >
On Fri, Jul 09, 2021 at 05:49:26PM +0800, Dongliang Mu wrote: > On Fri, Jul 9, 2021 at 2:38 PM Dan Carpenter <dan.carpenter@oracle.com> wrote: > > > > The video_device_release() frees the "vfd" pointer so passing it to > > video_unregister_device() on the next line results in a use after free. > > Calling video_unregister_device() on a device that hasn't been > > registered is supposed to be a no-op so that can be removed. The paths > > with to goto unreg_video_dev have a memory leak and should be updated to > > goto rel_vdev instead. > > > > Fixes: f7e7b48e6d79 ("[media] rockchip/rga: v4l2 m2m support") > > Signed-off-by: Dan Carpenter <dan.carpenter@oracle.com> > > --- > > drivers/media/platform/rockchip/rga/rga.c | 6 ++---- > > 1 file changed, 2 insertions(+), 4 deletions(-) > > > > diff --git a/drivers/media/platform/rockchip/rga/rga.c b/drivers/media/platform/rockchip/rga/rga.c > > index bf3fd71ec3af..37f7fd060c38 100644 > > --- a/drivers/media/platform/rockchip/rga/rga.c > > +++ b/drivers/media/platform/rockchip/rga/rga.c > > @@ -863,12 +863,12 @@ static int rga_probe(struct platform_device *pdev) > > if (IS_ERR(rga->m2m_dev)) { > > v4l2_err(&rga->v4l2_dev, "Failed to init mem2mem device\n"); > > ret = PTR_ERR(rga->m2m_dev); > > - goto unreg_video_dev; > > + goto rel_vdev; > > } > > > > ret = pm_runtime_resume_and_get(rga->dev); > > if (ret < 0) > > - goto unreg_video_dev; > > + goto rel_vdev; > > > > rga->version.major = (rga_read(rga, RGA_VERSION_INFO) >> 24) & 0xFF; > > rga->version.minor = (rga_read(rga, RGA_VERSION_INFO) >> 20) & 0x0F; > > @@ -904,8 +904,6 @@ static int rga_probe(struct platform_device *pdev) > > > > rel_vdev: > > video_device_release(vfd); > > -unreg_video_dev: > > - video_unregister_device(rga->vfd); > > unreg_v4l2_dev: > > v4l2_device_unregister(&rga->v4l2_dev); > > err_put_clk: > > >From the analysis of rga_probe and rga_remove function, the init and > cleanup functions are in pairs as follows: > Yeah. You're right. It's leaky. This is like the stuff Christophe Jaillet has been working on. This is also slightly complicated because what about if the call to device_register() fails inside video_register_device()? In that case we're not allowed to call video_device_release(). I think the way people deal with that is because device_register() never fails in real life so let's ignore that possibility. I sort of hate device_register() because it introduces a lot of bugs like this which don't affect real life. Inside the __video_register_device() function itself we just ignore errors from video_register_media_controller() because there is no possible way to handle it in a bug free way. Again, it's not a bug which is going to affect real life, but it's just frustrating to not be able to write code which is correct in a technical sense. Ah well... Fine. I'll resend. regards, dan carpenter
diff --git a/drivers/media/platform/rockchip/rga/rga.c b/drivers/media/platform/rockchip/rga/rga.c index bf3fd71ec3af..37f7fd060c38 100644 --- a/drivers/media/platform/rockchip/rga/rga.c +++ b/drivers/media/platform/rockchip/rga/rga.c @@ -863,12 +863,12 @@ static int rga_probe(struct platform_device *pdev) if (IS_ERR(rga->m2m_dev)) { v4l2_err(&rga->v4l2_dev, "Failed to init mem2mem device\n"); ret = PTR_ERR(rga->m2m_dev); - goto unreg_video_dev; + goto rel_vdev; } ret = pm_runtime_resume_and_get(rga->dev); if (ret < 0) - goto unreg_video_dev; + goto rel_vdev; rga->version.major = (rga_read(rga, RGA_VERSION_INFO) >> 24) & 0xFF; rga->version.minor = (rga_read(rga, RGA_VERSION_INFO) >> 20) & 0x0F; @@ -904,8 +904,6 @@ static int rga_probe(struct platform_device *pdev) rel_vdev: video_device_release(vfd); -unreg_video_dev: - video_unregister_device(rga->vfd); unreg_v4l2_dev: v4l2_device_unregister(&rga->v4l2_dev); err_put_clk:
The video_device_release() frees the "vfd" pointer so passing it to video_unregister_device() on the next line results in a use after free. Calling video_unregister_device() on a device that hasn't been registered is supposed to be a no-op so that can be removed. The paths with to goto unreg_video_dev have a memory leak and should be updated to goto rel_vdev instead. Fixes: f7e7b48e6d79 ("[media] rockchip/rga: v4l2 m2m support") Signed-off-by: Dan Carpenter <dan.carpenter@oracle.com> --- drivers/media/platform/rockchip/rga/rga.c | 6 ++---- 1 file changed, 2 insertions(+), 4 deletions(-)