Busy-waiting in drm_crtc_vblank_start_timer() until a previous vblank
timer has completed can deadlock. Rework how the vblank timer's start
and cancel helpers interact with the timer function.

Patch 1 fixes the possible deadlock by removing the busy-wait loop from
drm_crtc_vblank_start_timer(). To avoid possible interference between
_start_timer()'s hrtimer_start() and hrtimer_forward_now() in the timer
function, extend the interval_lock around both. Both, _start_timer() and
the timer function, can still overlap, but hrtimer initialization should
work reliably. At most, the first vblank event after changing the interval
might be off.

Fix the possible overlap by canceling the vblank timer's internal hrtimer.
Do this outside of the vblank code's critical section after switching
the vblank timer off. With the hrtimer being canceled, the next vblank
event will be sent at the correct time.

Tested with bochs on qemu.

Additional notes: there have been reports about vblank timers not happening
in time, such as [1]. We suspected that might happen due to guest systems
running under high CPU load. But the problem could also be caused by the
deadlock fixed in patch 1.

v2:
- handle uninitialized timer in drm_crtc_vblank_cancel_timer()

[1] 
https://lore.kernel.org/dri-devel/[email protected]/

Thomas Zimmermann (2):
  drm/vblank: timer: Do not block within critical section
  drm/vblank: Cancel vblank timer outside of critical section

 drivers/gpu/drm/drm_vblank.c         | 90 ++++++++++++++++++----------
 drivers/gpu/drm/drm_vblank_helper.c  | 16 +++++
 drivers/gpu/drm/vmwgfx/vmwgfx_ldu.c  |  3 +-
 drivers/gpu/drm/vmwgfx/vmwgfx_scrn.c |  2 +
 drivers/gpu/drm/vmwgfx/vmwgfx_stdu.c |  2 +
 drivers/gpu/drm/vmwgfx/vmwgfx_vkms.c |  2 +-
 include/drm/drm_crtc.h               |  9 +++
 include/drm/drm_vblank.h             | 16 +++++
 include/drm/drm_vblank_helper.h      |  2 +
 9 files changed, 108 insertions(+), 34 deletions(-)


base-commit: 7d502136b31169f6b5575f7f89c3da60cdf7891d
-- 
2.56.0

Reply via email to