diff options
author | Ville Syrjälä <ville.syrjala@linux.intel.com> | 2014-04-15 21:41:34 +0300 |
---|---|---|
committer | Daniel Vetter <daniel.vetter@ffwll.ch> | 2014-05-21 09:55:26 +0200 |
commit | 75f7f3ec600524c9544cc31695155f1a9ddbe1d9 (patch) | |
tree | 6ef900100c747383076e1ad89d8062c31dc86ecc /drivers/gpu/drm/i915/intel_drv.h | |
parent | 1ff74cf15a376a98d5c7c5b6dcb9aebd9734acdd (diff) |
drm/i915: Fix mmio vs. CS flip race on ILK+
Starting from ILK, mmio flips also cause a flip done interrupt to be
signalled. This means if we first do a set_base and follow it
immediately with the CS flip, we might mistake the flip done interrupt
caused by the set_base as the flip done interrupt caused by the CS
flip.
The hardware has a flip counter which increments every time a mmio or
CS flip is issued. It basically counts the number of DSPSURF register
writes. This means we can sample the counter before we put the CS
flip into the ring, and then when we get a flip done interrupt we can
check whether the CS flip has actually performed the surface address
update, or if the interrupt was caused by a previous but yet
unfinished mmio flip.
Even with the flip counter we still have a race condition of the CS flip
base address update happens after the mmio flip done interrupt was
raised but not yet processed by the driver. When the interrupt is
eventually processed, the flip counter will already indicate that the
CS flip has been executed, but it would not actually complete until the
next start of vblank. We can use the DSPSURFLIVE register to check
whether the hardware is actually scanning out of the buffer we expect,
or if we managed hit this race window.
This covers all the cases where the CS flip actually changes the base
address. If the base address remains unchanged, we might still complete
the CS flip before it has actually completed. But since the address
didn't change anyway, the premature flip completion can't result in
userspace overwriting data that's still being scanned out.
CTG already has the flip counter and DSPSURFLIVE registers, and
although the flip done interrupt is still limited to CS flips alone,
the code now also checks the flip counter on CTG as well.
v2: s/dspsurf/gtt_offset/ (Chris)
Testcase: igt/kms_mmio_vs_cs_flip/setcrtc_vs_cs_flip
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=73027
Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Reviewed-by: Rodrigo Vivi <rodrigo.vivi@gmail.com>
[danvet: Add g4x_ prefix to flip_count_after_eq.]
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Diffstat (limited to 'drivers/gpu/drm/i915/intel_drv.h')
-rw-r--r-- | drivers/gpu/drm/i915/intel_drv.h | 2 |
1 files changed, 2 insertions, 0 deletions
diff --git a/drivers/gpu/drm/i915/intel_drv.h b/drivers/gpu/drm/i915/intel_drv.h index 0ef2777514f..df7808cbb11 100644 --- a/drivers/gpu/drm/i915/intel_drv.h +++ b/drivers/gpu/drm/i915/intel_drv.h @@ -615,6 +615,8 @@ struct intel_unpin_work { #define INTEL_FLIP_INACTIVE 0 #define INTEL_FLIP_PENDING 1 #define INTEL_FLIP_COMPLETE 2 + u32 flip_count; + u32 gtt_offset; bool enable_stall_check; }; |