[<prev] [next>] [thread-next>] [day] [month] [year] [list]
Message-Id: <20190827213421.21917-1-robdclark@gmail.com>
Date: Tue, 27 Aug 2019 14:33:30 -0700
From: Rob Clark <robdclark@...il.com>
To: dri-devel@...ts.freedesktop.org
Cc: Rob Clark <robdclark@...omium.org>,
Abhinav Kumar <abhinavk@...eaurora.org>,
Alexios Zavras <alexios.zavras@...el.com>,
Allison Randal <allison@...utok.net>,
Boris Brezillon <bbrezillon@...nel.org>,
Bruce Wang <bzwang@...omium.org>,
Daniel Vetter <daniel.vetter@...ll.ch>,
Enrico Weigelt <info@...ux.net>,
freedreno@...ts.freedesktop.org (open list:DRM DRIVER FOR MSM ADRENO
GPU), Fritz Koenig <frkoenig@...gle.com>,
Georgi Djakov <georgi.djakov@...aro.org>,
Greg Kroah-Hartman <gregkh@...uxfoundation.org>,
Jeykumar Sankaran <jsanka@...eaurora.org>,
Jonathan Marek <jonathan@...ek.ca>,
Jordan Crouse <jcrouse@...eaurora.org>,
linux-arm-msm@...r.kernel.org (open list:DRM DRIVER FOR MSM ADRENO GPU),
linux-kernel@...r.kernel.org (open list),
Mamta Shukla <mamtashukla555@...il.com>,
Sean Paul <seanpaul@...omium.org>,
Sravanthi Kollukuduru <skolluku@...eaurora.org>,
Thomas Gleixner <tglx@...utronix.de>
Subject: [PATCH 0/9] drm/msm: async commit support
From: Rob Clark <robdclark@...omium.org>
Currently the dpu backend attempts to handle async commits. But it
is racey and could result in flushing multiple times in a frame, or
modifying hw state (such as scanout address or cursor position) after
the previous flush, but before vblank, causing underflows (which
manifest as brief black flashes).
This patchset removes the previous dpu async commit handling, and
reworks the internal kms backend API to decouple flushing. And in
the end introduces an hrtimer to flush async updates. The overall
approach is:
1) Move flushing various hw state out of encoder/crtc atomic commit
(which is anyways an improvement over the current state, where
we either flush from crtc or encoder, depending on whether it is
a full modeset)
2) Switch to crtc_mask for anything that completes after atomic
_commit_tail(), so we do not need to keep the atomic state
around. Ie. from drm core's perspective, an async commit
completes immediately.
This avoids fighting with drm core about the lifecycle of an
atomic state object.
3) Track a bitmask of crtcs w/ pending async flush, and setup
an hrtimer with expiration 1ms before vblank, to trigger
the flush. For async commits, we push the state down to
the double buffered hw registers immediately, and only
defer writing the flush registers.
Current patchset only includes the dpu backend support for async
commits.. mdp4 and mdp5 should be relatively trivial (less layers
of indirection involved). But I won't have access to any mdp4 hw
for a few more weeks, so at least that part I might punt on for
now.
Rob Clark (9):
drm/msm/dpu: unwind async commit handling
drm/msm/dpu: add real wait_for_commit_done()
drm/msm/dpu: handle_frame_done() from vblank irq
drm/msm: add kms->wait_flush()
drm/msm: convert kms->complete_commit() to crtc_mask
drm/msm: add kms->flush_commit()
drm/msm: split power control from prepare/complete_commit
drm/msm: async commit support
drm/msm/dpu: async commit support
drivers/gpu/drm/msm/disp/dpu1/dpu_crtc.c | 48 +---
drivers/gpu/drm/msm/disp/dpu1/dpu_crtc.h | 7 +-
drivers/gpu/drm/msm/disp/dpu1/dpu_encoder.c | 46 ++--
drivers/gpu/drm/msm/disp/dpu1/dpu_encoder.h | 8 +-
.../drm/msm/disp/dpu1/dpu_encoder_phys_vid.c | 39 ++--
drivers/gpu/drm/msm/disp/dpu1/dpu_kms.c | 99 +++++----
drivers/gpu/drm/msm/disp/mdp4/mdp4_kms.c | 48 ++--
drivers/gpu/drm/msm/disp/mdp5/mdp5_kms.c | 47 ++--
drivers/gpu/drm/msm/msm_atomic.c | 210 +++++++++++++++---
drivers/gpu/drm/msm/msm_drv.c | 1 +
drivers/gpu/drm/msm/msm_drv.h | 4 +
drivers/gpu/drm/msm/msm_kms.h | 108 ++++++++-
12 files changed, 466 insertions(+), 199 deletions(-)
--
2.21.0
Powered by blists - more mailing lists