[v7,0/9] Update AUX invalidation sequence

Message

Andi Shyti July 20, 2023, 9:07 p.m. UTC

Hi,

as there are new hardware directives, we need a little adaptation
for the AUX invalidation sequence.

In this version we support all the engines affected by this
change.

The stable backport has some challenges because the original
patch that this series fixes has had more changes in between.

This patch is slowly exploding with code refactorings and
features added and fixed.

Thanks a lot Nirmoy, Andrzej and Matt for your review and for the
fruitful discussions!

Thanks,
Andi

Changelog:
=========
v6 -> v7
 - Fix correct sequence applied to the correct engine. A little
   confusion promptly cought by Nirmoy when applying to the VD
   engine the sequence belonging to the compute engines. Thanks a
   lot, Nirmoy!

v5 -> v6
 - Fixed ccs flush in the engines VE and BCS. They are sent as a
   separate command instead of added in the pipe control.
 - Separated the CCS flusing in the pipe control patch with the
   quiescing of the memory. They were meant to be on separate
   patch already in the previous verision, but apparently I
   squashed them by mistake.

v4 -> v5
 - The AUX CCS is added as a device property instead of checking
   against FLAT CCS. This adds the new HAS_AUX_CCS check
   (Patch 2, new).
 - little and trivial refactoring here and there.
 - extended the flags{0,1}/bit_group_{0,1} renaming to other
   functions.
 - Created an intel_emit_pipe_control_cs() wrapper for submitting
   the pipe control.
 - Quiesce memory for all the engines, not just RCS (Patch 6,
   new).
 - The PIPE_CONTROL_CCS_FLUSH is added to all the engines.
 - Remove redundant EMIT_FLUSH_CCS mode flag.
 - Remove unnecessary NOOPs from the command streamer for
   invalidating the CCS table.
 - Use INVALID_MMIO_REG and gen12_get_aux_inv_reg() instad of
   __MMIO(0) and reg.reg.
 - Remove useless wrapper and just use gen12_get_aux_inv_reg().

v3 -> v4
 - A trivial patch 3 is added to rename the flags with
   bit_group_{0,1} to align with the datasheet naming.
 - Patch 4 fixes a confusion I made where the CCS flag was
   applied to the wrong bit group.

v2 -> v3
 - added r-b from Nirmoy in patch 1 and 4.
 - added patch 3 which enables the ccs_flush in the control pipe
   for mtl+ compute and render engines.
 - added redundant checks in patch 2 for enabling the EMIT_FLUSH
   flag.

v1 -> v2
 - add a clean up preliminary patch for the existing registers
 - add support for more engines
 - add the Fixes tag

Andi Shyti (7):
  drm/i915/gt: Cleanup aux invalidation registers
  drm/i915: Add the has_aux_ccs device property
  drm/i915/gt: Rename flags with bit_group_X according to the datasheet
  drm/i915/gt: Enable the CCS_FLUSH bit in the pipe control
  drm/i915/gt: Refactor intel_emit_pipe_control_cs() in a single
    function
  drm/i915/gt: Ensure memory quiesced before invalidation for all
    engines
  drm/i915/gt: Support aux invalidation on all engines

Jonathan Cavitt (2):
  drm/i915/gt: Ensure memory quiesced before invalidation
  drm/i915/gt: Poll aux invalidation register bit on invalidation

 drivers/gpu/drm/i915/gt/gen8_engine_cs.c     | 188 ++++++++++++-------
 drivers/gpu/drm/i915/gt/gen8_engine_cs.h     |  21 ++-
 drivers/gpu/drm/i915/gt/intel_gpu_commands.h |   2 +
 drivers/gpu/drm/i915/gt/intel_gt_regs.h      |  16 +-
 drivers/gpu/drm/i915/gt/intel_lrc.c          |  17 +-
 drivers/gpu/drm/i915/i915_drv.h              |   1 +
 drivers/gpu/drm/i915/i915_pci.c              |   5 +-
 drivers/gpu/drm/i915/intel_device_info.h     |   1 +
 8 files changed, 151 insertions(+), 100 deletions(-)