On Thu, Jul 23, 2020 at 09:10:15PM +0000, Mazin Rezk wrote:
When amdgpu_dm_atomic_commit_tail is running in the workqueue,
drm_atomic_state_put will get called while amdgpu_dm_atomic_commit_tail is
running, causing a race condition where state (and then dm_state) is
sometimes freed while amdgpu_dm_atomic_commit_tail is running. This bug has
occurred since 5.7-rc1 and is well documented among polaris11 users .
Prior to 5.7, this was not a noticeable issue since the freelist pointer
was stored at the beginning of dm_state (base), which was unused. After
changing the freelist pointer to be stored in the middle of the struct, the
freelist pointer overwrote the context, causing dc_state to become garbage
data and made the call to dm_enable_per_frame_crtc_master_sync dereference
a freelist pointer.
This patch fixes the aforementioned issue by calling drm_atomic_state_get
in amdgpu_dm_atomic_commit before drm_atomic_helper_commit is called and
drm_atomic_state_put after amdgpu_dm_atomic_commit_tail is complete.
According to my testing on 5.8.0-rc6, this should fix bug 207383 on
Nice work tracking this down!
Fixes: 3202fa62f ("slub: relocate freelist pointer to middle of object")
I do, however, object to this Fixes tag. :) The flaw appears to have
been with amdgpu_dm's reference tracking of "state" in the nonblocking
case. (How this reference counting is supposed to work correctly, though,
I'm not sure.) If I look at where the drm helper was split from being
the default callback, it looks like this was what introduced the bug:
da5c47f682ab ("drm/amd/display: Remove acrtc->stream")
? 3202fa62f certainly exposed it much more quickly, but there was a race
even without 3202fa62f where something could have realloced the memory
and written over it.