Re: [PATCH v3 1/4] fbdev: Prevent possible use-after-free in fb_release()
From: Javier Martinez Canillas <javierm@redhat.com>
Date: 2022-05-09 15:30:40
Also in:
dri-devel, lkml
Subsystem:
drm drivers, drm drivers and misc gpu patches, the rest · Maintainers:
David Airlie, Simona Vetter, Maarten Lankhorst, Maxime Ripard, Thomas Zimmermann, Linus Torvalds
Hello Andrzej, On 5/9/22 16:56, Andrzej Hajda wrote:
On 06.05.2022 00:04, Javier Martinez Canillas wrote:quoted
From: Daniel Vetter <redacted> Most fbdev drivers have issues with the fb_info lifetime, because call to framebuffer_release() from their driver's .remove callback, rather than doing from fbops.fb_destroy callback. Doing that will destroy the fb_info too early, while references to it may still exist, leading to a use-after-free error. To prevent this, check the fb_info reference counter when attempting to kfree the data structure in framebuffer_release(). That will leak it but at least will prevent the mentioned error. Signed-off-by: Daniel Vetter <redacted> Signed-off-by: Javier Martinez Canillas <javierm@redhat.com> Reviewed-by: Thomas Zimmermann <tzimmermann@suse.de> --- (no changes since v1) drivers/video/fbdev/core/fbsysfs.c | 4 ++++ 1 file changed, 4 insertions(+)diff --git a/drivers/video/fbdev/core/fbsysfs.c b/drivers/video/fbdev/core/fbsysfs.c index 8c1ee9ecec3d..c2a60b187467 100644 --- a/drivers/video/fbdev/core/fbsysfs.c +++ b/drivers/video/fbdev/core/fbsysfs.c@@ -80,6 +80,10 @@ void framebuffer_release(struct fb_info *info) { if (!info) return; + + if (WARN_ON(refcount_read(&info->count))) + return; +Regarding drm: What about drm_fb_helper_fini? It calls also framebuffer_release and is called often from _remove paths (checked intel/radeon/nouveau). I guess it should be fixed as well. Do you plan to fix it?
I think you are correct. Maybe we need something like the following?
diff --git a/drivers/gpu/drm/drm_fb_helper.c b/drivers/gpu/drm/drm_fb_helper.c
index d265a73313c9..b09598f7af28 100644
--- a/drivers/gpu/drm/drm_fb_helper.c
+++ b/drivers/gpu/drm/drm_fb_helper.c@@ -631,7 +631,6 @@ void drm_fb_helper_fini(struct drm_fb_helper *fb_helper) if (info) { if (info->cmap.len) fb_dealloc_cmap(&info->cmap); - framebuffer_release(info); } fb_helper->fbdev = NULL;
@@ -2112,6 +2111,7 @@ static void drm_fbdev_release(struct drm_fb_helper *fb_helper) static void drm_fbdev_fb_destroy(struct fb_info *info) { drm_fbdev_release(info->par); + framebuffer_release(info); } static int drm_fbdev_fb_mmap(struct fb_info *info, struct vm_area_struct *vma)
Regarding fb drivers, just for stats: git grep -p framebuffer_release | grep _remove | wc -l Suggests there is at least 70 incorrect users of this :)
Yes, Daniel already mentioned that most of them get this wrong but I was
mostly interested in {simple,efi,vesa}fb since are used with "nomodeset".
But given that I only touched those tree and still managed to introduce
a regression, I won't attempt to fix the others.
--
Best regards,
Javier Martinez Canillas
Linux Engineering
Red Hat