| Message ID | 20220722215234.129793-1-Yunxiang.Li@amd.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | [1/2] drm: Fix vblank refcount during modeset | expand |
[AMD Official Use Only - General] I found out that elsewhere in the drm code (e.g. in drm_atomic_helper.c) expects drm_vblank_get() to fail as part of normal operation. So this patch doesn't seem appropriate anymore and it seems more appropriate to hunt down all the unchecked calls for drm_vblank_get and fix them instead. I don't know how to best make sure this doesn't reoccur in the future though. Regards, Yunxiang
diff --git a/drivers/gpu/drm/drm_vblank.c b/drivers/gpu/drm/drm_vblank.c index 2ff31717a3de..159d13b5d97b 100644 --- a/drivers/gpu/drm/drm_vblank.c +++ b/drivers/gpu/drm/drm_vblank.c @@ -1174,7 +1174,7 @@ int drm_vblank_get(struct drm_device *dev, unsigned int pipe) if (atomic_add_return(1, &vblank->refcount) == 1) { ret = drm_vblank_enable(dev, pipe); } else { - if (!vblank->enabled) { + if (!vblank->enabled && !vblank->inmodeset) { atomic_dec(&vblank->refcount); ret = -EINVAL; }
drm_crtc_vblank_off increments the refcount to prevent vblank from getting enabled during modeset, but this causes the refcount elsewhere to be off if they call drm_vblank_get without checking the return and do a drm_vblank_put later. This can be reproduced by toggling vrr mode on amdgpu. Since drm_crtc_vblank_on later re-enables vblank if the refcount is not zero, letting drm_vblank_get succeed during modeset should fix the behavior. Closes: https://gitlab.freedesktop.org/drm/amd/-/issues/1380 Signed-off-by: Yunxiang Li <Yunxiang.Li@amd.com> --- drivers/gpu/drm/drm_vblank.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)