On 5/10/22 00:22, Andrzej Hajda wrote:

[snip]

quoted

  static void drm_fbdev_fb_destroy(struct fb_info *info)
  {
+       if (info->cmap.len)
+               fb_dealloc_cmap(&info->cmap);
+
         drm_fbdev_release(info->par);
+       framebuffer_release(info);

I would put drm_fbdev_release at the beginning - it cancels workers 
which could expect cmap to be still valid.

Indeed, you are correct again. [0] is the final version of the patch I've
but don't have an i915 test machine to give it a try. I'll test tomorrow
on my test systems to verify that it doesn't cause any regressions since
with other DRM drivers.

I think that besides this patch, drivers shouldn't need to call to the
drm_fb_helper_fini() function directly. Since that would be called during
drm_fbdev_fb_destroy() anyways.

We should probably remove that call in all drivers and make this helper
function static and just private to drm_fb_helper functions.

Or am I missing something here ?

[0]:
From 5170cafcf2936da8f1c53231e3baa7d7a2b16c61 Mon Sep 17 00:00:00 2001
From: Javier Martinez Canillas <javierm@redhat.com>
Date: Tue May 10 00:39:55 2022 +0200
Subject: [RFT PATCH] drm/fb-helper: Don't deallocate fb colormap and free fb info
 too early

Currently these are done in drm_fb_helper_fini() but this helper is called
by drivers in their .remove callback, which could lead to a use-after-free
if a process has opened the emulated fbdev node while a driver is removed.

For example, in i915 driver the call chain during remove is the following:

struct pci_driver i915_pci_driver = {
...
        .remove = i915_pci_remove,
...
};

i915_pci_remove
  i915_driver_remove
    intel_modeset_driver_remove_noirq
      intel_fbdev_fini
        intel_fbdev_destroy
          drm_fb_helper_fini
            framebuffer_release

Later the process will close the fbdev node file descriptor leading to the
mentioned use-after-free bug in drm_fbdev_fb_destroy(), due the following:

drm_fbdev_fb_destroy
  drm_fbdev_release(info->par); <-- info was already freed on .remove

To prevent that, let's move the framebuffer_release() call to the end of
the drm_fbdev_fb_destroy() function.

Also, the call to fb_dealloc_cmap() in drm_fb_helper_fini() is too early
and is more correct to do it in drm_fbdev_fb_destroy() as well. After a
call to drm_fbdev_release() has been made.

Reported-by: Andrzej Hajda <andrzej.hajda@intel.com>
Signed-off-by: Javier Martinez Canillas <javierm@redhat.com>
---
 drivers/gpu/drm/drm_fb_helper.c | 9 +++------
 1 file changed, 3 insertions(+), 6 deletions(-)

diff --git a/drivers/gpu/drm/drm_fb_helper.c b/drivers/gpu/drm/drm_fb_helper.c
index d265a73313c9..7288fbd26bcc 100644
--- a/drivers/gpu/drm/drm_fb_helper.c
+++ b/drivers/gpu/drm/drm_fb_helper.c

@@ -627,12 +627,6 @@ void drm_fb_helper_fini(struct drm_fb_helper *fb_helper)
 	cancel_work_sync(&fb_helper->resume_work);
 	cancel_work_sync(&fb_helper->damage_work);
 
-	info = fb_helper->fbdev;
-	if (info) {
-		if (info->cmap.len)
-			fb_dealloc_cmap(&info->cmap);
-		framebuffer_release(info);
-	}
 	fb_helper->fbdev = NULL;
 
 	mutex_lock(&kernel_fb_helper_lock);

@@ -2112,6 +2106,9 @@ 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);
+	if (info->cmap.len)
+		fb_dealloc_cmap(&info->cmap);
+	framebuffer_release(info);
 }
 
 static int drm_fbdev_fb_mmap(struct fb_info *info, struct vm_area_struct *vma)

-- 
2.35.1