drm/i915 Updates Daniel Vetter, Intel OTC FOSDEM 2013 bug - - PowerPoint PPT Presentation

▶

Dec 24, 2022 189 likes •441 views

drm/i915 Updates Daniel Vetter, Intel OTC FOSDEM 2013 bug squashing bugs fixed by the truckload especially regressions! intel-gpu-tools kernel tests no testsuite is worse than having none debugfs knobs for tricky cases

SLIDE 1

drm/i915 Updates

Daniel Vetter, Intel OTC FOSDEM 2013

SLIDE 2

bug squashing

bugs fixed by the truckload
especially regressions!

SLIDE 3

intel-gpu-tools kernel tests

no testsuite is worse than having none
debugfs knobs for tricky cases
started with automake testrunner, now piglit

SLIDE 4

i-g-t kernel tests: future

modeset coverage still really spotty
gpu hang&reset code is tricky
unified kernel tests for other gpus?

SLIDE 5

dear gfx bug reporter

please file duplicates srly

SLIDE 6

haswell display

LVDS, SDVO gone, VGA marginalised
ports on CPU, not PCH

SLIDE 7

ld: PCH-split

CPU PCH PIPE A ports: DP SDVO HDMI LVDS VGA ... PIPE B FDI A FDI B MUX MUX eDP

SLIDE 8

new: DDI ports

CPU PCH PIPE A VGA PIPE B FDI A MUX DDI E DDI A ... eDP DP/HDMI

SLIDE 9

modeset rework

haswell DP DDI like old cpu eDP
„wrong“ enable/disable sequence

SLIDE 10

crtc helpers

too flexible: independant encoders
not flexible enough:

simple encoder enable/disable sequence

SLIDE 11

new i915 modeset code

dpms simplified
modeset sequence driven by crtc

SLIDE 12

new modeset: code tools

output state staging
hw state read-out

SLIDE 13

new modeset: results

simplified modeset state space
paranoid modeset state checker
enabled haswell DP support
drm helpers are in rather good shape

SLIDE 14

new modeset: future

atomic modeset
fastboot

SLIDE 15

minor modeset stuff

EDID improvements
hotplug is a mess

SLIDE 16

drm modeset locking rework

per-crtc locking for pageflip/cursor/fb ops
required proper framebuffer refcounting

SLIDE 17

drm modeset locking rework

per-crtc locking for pageflip/cursor/fb ops
required proper framebuffer refcounting

SLIDE 18

GEM

little bits of tuning all over
hw context support
cachability control

SLIDE 19

GEM: no more flushing list

GEM: per-cache GPU domains
complex and delayed flushing caused stalls
completely ripped out (almost)

SLIDE 20

GEM: tuning

unbound tracking
reduce mappable pressure
no-reloc optimization

SLIDE 21

GEM: the future

real per-process address spaces
userptr, maybe ...
more than One Lock to Rule Them All

SLIDE 22

dma_buf: fences & reservations

Maarten Lankhorst & Rob Clark
essentially ttm, trimmed down
currently just for synchronization

SLIDE 23

dma_buf: opens

needs review of mutex changes
android sync_points, SoC adoption
integration with i915/GEM will be fun
better i915 OOM handling

SLIDE 24

summary

fewer bugs & better QA tools
improved modeset infrastructure
fastboot & atomic modeset in the future
tuned GEM
big changes around dma_buf ahead