Skip to content

drm: google: add the Pixel 6 Pro's display (gs101 DECON and DSIM, S6E3HC3 panel) - #116

Merged
derekross merged 9 commits into
derekross:omamobilefrom
robotosaith:omamobile-gs101-display-pr
Nov 10, 2026
Merged

derekross merged 9 commits into
derekross:omamobilefrom
robotosaith:omamobile-gs101-display-pr

Conversation

@robotosaith

@robotosaith robotosaith commented Oct 28, 2026 •

Copy link
Copy Markdown
Contributor

Raven's display on Derek's gs-decon/gs-dsim driver: gs101 has gs201's DECON and DSIM register layout, so the work is the SoC's nodes, a panel driver and three gs-decon fixes found on raven.

  • gs101 in the DECON and DSIM bindings: "google,gs101-decon", "google,gs201-decon" and "google,gs101-dsim", "google,gs201-dsim"; only DECON0's address and the interrupts differ, in the DT. gs-dsim-cmd is unchanged.
  • gs-decon: google,gs101-decon gets its own match data, gs201's plus two flags that stay off for gs201:
    • hold_boot_fb: hold the boot framebuffer's aperture after evicting simpledrm, so simpledrm can't probe late and stay as a second output;
    • vt_switch_fbdev_only: no VT switch around system sleep without the fbdev console.
  • gs-decon, all SoCs: put the memory bus (MIF) back to its probe rate after a 120 Hz stint. On the Pixel 7 Pro the bootloader leaves MIF at its top level, so nothing is written there.
  • S6E3HC3 panel: a binding entry next to the S6E3HC4 and a small driver for brightness (backlight) and display on/off. The panel runs as the bootloader left it; the modes come from gs-decon.
  • DT: DECON0 and DSIM0 in gs101.dtsi (disabled), and raven's panel with its rails, which stay on since the panel driver doesn't switch them.

Tested on raven as 7.3.0-rc6-omamobile-push1-g5a144ff68cc6 (the head before this rebase, plus raven's not-for-merge USB 2 peripheral override): gs-decon and the panel driver take over from the bootloader, one output (DSI-1), 60 Hz at probe and 120 Hz from the shell, DPMS off and on, no flip timeouts with the backlight at 0, and three s2idle cycles each woken by one Power press. The smoke report is in the thread.

tools/kernel-pr-check --build --merge passes: W=1 builds clean, nine device trees with no new dtbs_check warnings, and dt_binding_check on the three bindings. The gs201, zuma and zumapro DTBs are byte-identical to the base; the Pixel 6 and 6a gain only the two disabled nodes from gs101.dtsi. checkpatch --strict is clean: the trailers are Assisted-by: Claude Opus 5.5 now, and the panel driver has a MAINTAINERS entry.

🤖 Generated with Claude Code

@derekross

Copy link
Copy Markdown
Owner

Thanks, this is a lot of careful work, and good news for raven. I checked the hardware side against Google's gs101-drm-dpu.dtsi and gs101.h: DECON0 at 0x1c300000 (which matches raven's SIMPLEFB_TRIG_CON), DPP0's DMA at 0x1c0b0000, DSIM0 and its PHY ranges, and SPI 338/337/336 and 345 all match. The pool's alignment and alloc-ranges have the right cell counts. The panel's 4..2047 range and its 60/120 Hz manual-mode bytes match Google's S6E3HC3 driver. Every commit builds, and the Pixel 7 Pro's DTB is byte-identical.

Since gs-decon and gs-dsim-cmd run on my Pixel 7 Pro every day, I compared their objects function by function against today's base. Three things need sorting out before merging:

  1. Two of the gs-decon fixes change the Pixel 7 Pro. gs101 uses gs201's DECON data, so none of the new behaviour is behind gs101's match data:

    • pm_vt_switch_required(dev, fbdev): cheetah runs with fbdev=0. The PM core does the VT switch while its list is empty or someone requires it, and simpledrm's entry goes away at the eviction. So from this commit on, cheetah stops switching VTs around suspend. That may well fix the same lost wake press there, but it changes the s2idle path I've validated.
    • Holding the aperture adds a new failure path to cheetah's probe.
    • The MIF reset is a no-op on cheetah: its bootloader leaves MIF at 3172 MHz, so the rate is never above the probe rate.

    Either put the VT-switch and aperture changes behind gs101 data, or keep them general and they wait until I've run them on my phone (boot, 60/120 Hz, and some s2idle cycles woken by the power key). Both are fine with me; say which you prefer.

  2. Two drivers writing DECON's TRIG_CON on raven. Raven's &dsi0 gets google,decon = <&decon0> in the same commit that enables gs-decon's KMS on that DECON, while the binding says the phandle is for when no KMS driver runs it. With both, a blanked backlight sends display off and the DSIM sets the trigger mask behind gs-decon, which keeps flipping. gs-decon's own trigger handling and its trigger_masked state do read-modify-writes on the same register without a shared lock, and a later unblank (or the DSIM's unbind) unmasks it under gs-decon. Your "flip timeouts only with DPMS on and the backlight off" is exactly that state. Could you drop the phandle from raven's DT now that gs-decon runs it (its screen_off handles frames with the CRTC off), or have the DSIM skip the masking when DECON has a KMS driver, and then check whether the flip timeouts go away?

  3. Compatibles: gs101's DSIM is gs201's unchanged, and its DECON uses gs201's data. Tegu's PR arm64: dts: exynos: google: zumapro-tegu: Enable the display #90 describes its DSIM as "google,zumapro-dsim", "google,gs201-dsim" with no driver change, and I'd like gs101 the same way: "google,gs101-dsim", "google,gs201-dsim" and "google,gs101-decon", "google,gs201-decon", without the driver match commit. If you go for gs101 data in point 1, a gs101 DECON entry has its reason, and the DSIM can still fall back. After arm64: dts: exynos: google: zumapro-tegu: Enable the display #90 the DSIM binding's compatible would be:

      compatible:
        oneOf:
          - const: google,gs201-dsim
          - items:
              - enum:
                  - google,gs101-dsim
                  - google,zumapro-dsim
              - const: google,gs201-dsim

Smaller:

  • Kernel PRs land commit by commit, and "enable the display" currently comes before the two fixes for problems it causes (the phantom simpledrm output and the lost wake press). Please move those two fixes before the enable commit.
  • This conflicts with drm: google: 120 Hz on the Pixel 9a (per-panel DSIM timers, gs-decon on zumapro) #88 (struct gs_dsim: keep both rates and decon) and with arm64: dts: exynos: google: zumapro-tegu: Enable the display #90 (the binding above). Those two are older, so please rebase onto them once they're in, and onto the raven DT PRs as they land.
  • The LDO22M/BUCKA comment says "until the panel driver owns them", but the S6E3HC3 driver takes no supplies. Either give it the supplies, or say they stay on.
  • Please add the uname -r of the run, and a tools/phone-smoke report from that build.

@robotosaith

Copy link
Copy Markdown
Contributor Author

Smoke report from the build that runs this: #112 (comment).

@robotosaith
robotosaith marked this pull request as draft October 28, 2026 15:33
@derekross

Copy link
Copy Markdown
Owner

Thanks for the smoke report, and I see you've made this a draft while you work on it. The three points from my earlier comment still apply to this head (it hasn't changed since), so here's just where things stand around it:

@robotosaith

Copy link
Copy Markdown
Contributor Author

Reworked along your three points, rebased onto omamobile with #114 under it (the raven DT blocks stack in merge order: battery, torch, haptics, display). Until #114 lands its three commits show here too.

  1. gs101 data: "google,gs101-decon" gets its own match data, gs201's plus hold_boot_fb and vt_switch_fbdev_only. Both stay off for gs201, so cheetah's probe and s2idle paths are unchanged. The MIF reset stays general, since it's a no-op on cheetah as you found.
  2. TRIG_CON: raven's &dsi0 has no google,decon phandle any more, and both DSIM commits (the binding property and the masking) are gone. With the backlight at 0 and the display on for 30 s there were no flip timeouts; the old series gave about 60.
  3. Compatibles: "google,gs101-dsim", "google,gs201-dsim" and "google,gs101-decon", "google,gs201-decon", with the DSIM binding in the oneOf shape you gave (gs101 only in the enum until arm64: dts: exynos: google: zumapro-tegu: Enable the display #90 adds zumapro). There's no DSIM match commit: gs-dsim-cmd.c is identical to the base.

The smaller ones: the aperture and VT-switch fixes now come before the DT commits, and the LDO22M/BUCKA comment says the rails stay on, since the S6E3HC3 driver doesn't switch them.

We know this waits for #88 and #90: it still conflicts with both (struct gs_dsim and the DSIM binding), and I'll rebase onto them once they're in.

Gate (tools/kernel-pr-check --build) clean, the three bindings pass dt-doc-validate, raven's DTB has no dt-validate findings, and cheetah's, felix's, lynx's, panther's and husky's DTBs are byte-identical to the base.

Booted on raven as 7.3.0-rc6-omamobile-push1-g5a144ff68cc6: this head plus one not-for-merge commit (raven's USB 2 peripheral override, DT only). gs-decon.c is blob d60c5ddee877 and panel-samsung-s6e3hc3.c is 0aaaf56eff44 in both. One output (DSI-1), 60 Hz at probe and 120 Hz from the shell, DPMS off and on, and three s2idle cycles (14, 62 and 28 s), each woken by a single Power press that also reached the shell.

tools/phone-smoke report (Wi-Fi, sound, sensors and the modem units fail or skip: this kernel doesn't build them)

phone-smoke: unknown

Kernel 7.3.0-rc6-omamobile-push1-g5a144ff68cc6, tools at 80e18e3, 2026-10-29T06:27Z.

Check Result Detail
boot (required) skip kernel 7.3.0-rc6-omamobile-push1-g5a144ff68cc6, no slot on the command line
storage guard (required) ok 49 partitions; only userdata and slot B's boot images writable
root (required) ok ext4 on userdata, 97G free
kernel health (required) ok no oops or BUG; 0 WARN, 50 error lines in dmesg
systemd units not ok failed: omamobile-esim-boot.service omamobile-modem.service omamobile-sitd-2.service omamobile-sitd.service
USB gadget (required) ok bound to 11110000.usb
display ok card2-DSI-1 connected, 1440x3120
GPU ok render node on panfrost
session (required) ok Hyprland and the Omarchy shell running
touch ok s6sy761
buttons ok gpio-keys
Wi-Fi skip no wireless interface
Bluetooth ok hci0, powered true
battery ok 96% Discharging, 22 °C
sound cards skip none
sensors skip no sensor proxy
cameras skip no libcamera tools
thermal ok hottest zone 43 °C

Packages: hyprland 0.56.2-7, libcamera 0.7.2-4, mesa 1:26.2.4-1.2, omacam 0.1.9-1, omamessages 0.2.0-1, omamobile-apps 0.2.0-1, omamobile-audio 0.4.4-1, omamobile-device-raven 0.1.23.6-1, omamobile-firstboot 0.1.0-1, omamobile-gnss 0.2.1-1, omamobile-keyring 20261013-1, omamobile-mirror 0.1.1-1, omamobile-mmd 0.1.2-1, omamobile-modem 0.1.9-1, omamobile-modem-sit 0.3.57-1, omamobile-omarchy 0.1.1-1, omamobile-owl-widgets 0.1.1-1, omamobile-sensors 0.2.0-1, omamobile-shell 0.2.21-1, omaphone 0.2.0-1, omarchy 4.0.4-2, omatel-agent 0.2.0-1, quickshell 0.3.2.r7.g2b4501a-1

@robotosaith
robotosaith marked this pull request as ready for review October 29, 2026 06:50
@derekross

Copy link
Copy Markdown
Owner

Thanks for reworking this along all three points, and with a fresh boot and smoke report. It reads much more cleanly now. I went through the new head point by point:

  1. The Pixel 7 Pro: google,gs101-decon has its own data, and the aperture hold, the VT-switch opt-out and its unregister are each behind it, so gs201 leaves them off. I compared gs-decon.o against today's base with my config. Apart from functions that only moved by the new field's 8 bytes, the changes are in probe, remove and gs_decon_set_floors (which set_rate and get_dsim no longer inline). The gs201, felix, lynx, panther and husky DTBs are byte-identical.
    On the MIF restore staying general: I'm fine with that. On cheetah, the only difference is one read of the MIF rate from ACPM when leaving 120 Hz. It never writes, because the bootloader leaves MIF at its top level (3172 MHz), so the rate can't exceed what probe saw. Could you add a sentence to that commit's message saying so? Something like "On the Pixel 7 Pro the bootloader leaves MIF at its top level, so the rate never exceeds the probe rate and nothing is written there." That saves the next reader working it out.
  2. TRIG_CON: raven's &dsi0 has no google,decon any more, the DSIM commits are gone, and gs-dsim-cmd.c is identical to the base. Zero flip timeouts in the dark-backlight state, against about 60 before, is a nice confirmation.
  3. Compatibles: the DSIM and DECON fallbacks and the binding shapes are as I hoped.

The smaller ones are all done too: the fixes come before the DT commits, the rail comment matches what the driver does, and the booted gs-decon.c and panel driver are this head's blobs. Every C commit builds on its own here. The gate passes, and CI's dtbs_check and dt_binding_check are green.

What's left is the order. This merges after #114 (its three commits are under this one) and after tegu's #88 and #90, and then it needs a rebase. To save you some time, the two conflicts resolve like this:

A few small things you can fold into that rebase if you like:

  • The PR description still mentions the gs-dsim-cmd change, "gs201's data table" and a new W=1 device-tree warning. None of those apply any more, and it's the record of what landed.
  • In gs_decon_probe, the new mif_rest line and the existing dev_info both sit under if (dc->mif), so they can share one block.
  • checkpatch's one CHECK in the aperture commit: the devm_aperture_acquire_for_platform_device( continuation isn't aligned with the parenthesis.

Thanks again. It'll be good to have raven's display on the published branch.

@robotosaith
robotosaith force-pushed the omamobile-gs101-display-pr branch 3 times, most recently from fb1bb4f to c9ca976 Compare November 2, 2026 13:05
@robotosaith

Copy link
Copy Markdown
Contributor Author

Thanks for the careful review, and for spelling out both conflict resolutions. That made the rebase quick.

Rebased onto today's omamobile: #114's commits are gone, and the two conflicts are resolved exactly as you wrote (all three flags in struct gs_decon_soc_data, and gs101-dsim and zumapro-dsim in the DSIM enum). The small ones are folded in too:

  • the MIF commit's message has your sentence about the Pixel 7 Pro;
  • mif_rest and the dev_info share one if (dc->mif) block;
  • the aperture call's continuation is aligned with the parenthesis, and checkpatch's CHECK is gone;
  • the PR description now matches the series;
  • checkpatch is clean now: Assisted-by: Claude Opus 5.5 replaces the Co-Authored-By trailers, and the S6E3HC3 driver has its own MAINTAINERS entry, like the TG4C's.

With #88's per-panel DSIM timers, raven's bootloader timers match the Pixel 7 Pro's S6E3HC4 set (the same values the shared table had before), so raven keeps its 60/120 Hz switching with no change here.

tools/kernel-pr-check --build --merge passes. The gs201, zuma and zumapro DTBs are byte-identical to the base. The Pixel 6 and 6a (bluejay is new since the last round) gain only the two disabled nodes from gs101.dtsi.

I haven't booted this head yet. Since the booted one, the code changes are the merged if block and the rewrapped call, and the rest is the rebase. I'll post a smoke report from this head once it has run on raven.

@derekross

Copy link
Copy Markdown
Owner

Thank you, @robotosaith, and nicely done on the rebase. Both conflicts are resolved exactly as I'd hoped (all three flags in struct gs_decon_soc_data, and gs101-dsim next to zumapro-dsim in the DSIM enum), the MIF commit has the sentence about the Pixel 7 Pro, and the description matches the series. Our gate passes on today's omamobile (clean merge, all nine signed off, checkpatch --strict and W=1 clean, no new device-tree warnings), each C commit builds on its own, and the 7 Pro's, 9a's and 8 Pro's DTBs are byte-identical to the base.

The one thing left before I merge is the raven run. Since the boot you reported, gs-decon.c changed (the rebase onto #88) and #88's per-panel DSIM timers now pick their set from what the bootloader left. So in the report, please include uname -r, the gs_dsim line that says "link timers for 60 Hz" (rather than "unknown link timers"), a switch to 120 Hz and back, DPMS off and on, a few s2idle cycles woken by Power, and a tools/phone-smoke report.

Optional: in gs_decon_set_floors(), checking dc->mif_rest < GS_DECON_MIF_FLOOR_HZ before reading the clock would skip the restore when the bootloader left MIF above the floor. On the 7 Pro that saves an ACPM read at probe and at every drop from 120 Hz; on raven nothing changes.

@robotosaith

Copy link
Copy Markdown
Contributor Author

Ran this head on raven today: c9ca9760a plus raven's NOT FOR MERGE USB peripheral override (no TCPC driver in the published branch), on Omarchy Mobile freshly installed from main.

  • uname -r: 7.3.0-rc6-omamobile-push1-gb0a658e17f21
  • Link timers: gs-dsim-cmd 1c2c0000.dsi: link timers for 60 Hz, then the panel is taken over from the bootloader and gs-decon comes up on card2, one output DSI-1, 1440x3120.
  • 120 Hz and back: Hyprland moves to 120 Hz at session start (gs-decon: 120 Hz), then 60 Hz and back to 120 Hz through hl.monitor(), each logged by gs-decon. No flip timeouts, no underruns.
  • DPMS: off for 5 s (dpmsStatus 0, CRTC active=0), then on at 120 Hz. No flip timeouts.
  • s2idle: three systemctl suspend cycles, each woken by Power (19 s, 86 s and 25 s asleep), suspend_stats 3 success, 0 fail, display back at 120 Hz with DPMS on every time.

The smoke report flags one thing: new dmesg errors, all UFS. They're the published branch's UFS spm_lvl 3 on gs101: hibern8 exit fails on resume and a host reset recovers it, with no I/O errors (host_reset: total cnt=2, then Power mode changed to : FAST series_B G_4 L_2). Raven's kernel with the gs101 spm_lvl 5 default doesn't do this, so it's not from this PR.

phone-smoke

phone-smoke: raven

Kernel 7.3.0-rc6-omamobile-push1-gb0a658e17f21, tools at c0d0c07, 2026-11-03T10:50Z.

Check Result Detail
boot (required) ok kernel 7.3.0-rc6-omamobile-push1-gb0a658e17f21, slot b, up 336s
storage guard (required) ok 49 partitions; only userdata and slot B's boot images writable
root (required) ok ext4 on userdata, 97G free
kernel health (required) ok no oops or BUG; 0 WARN, 145 error lines in dmesg (17 known, of them UFS dme-set 0x4021 (s2idle) x3, 128 new)
new dmesg errors not ok host_regs: N: N N N N x27; host_regs: N: N ffffffff N N x3; exynos-ufshc N: saved_err=N, saved_uic_err=N x3
systemd units ok none failed
USB gadget (required) ok bound to 11110000.usb
display ok card2-DSI-1 connected, 1440x3120
GPU ok render node on panfrost
session (required) ok Hyprland and the Omarchy shell running
lock ok PAM file present, the shell can lock (passwordPam true, locked false)
touch ok s6sy761
buttons ok gpio-keys
Wi-Fi skip no wireless interface
Bluetooth ok hci0, powered true
battery ok 96% Discharging, 31 °C
sound cards skip none
sensors skip no sensor proxy
cameras skip no libcamera tools
thermal ok hottest zone 43 °C

Packages: hyprland 0.57.0-3.1, libcamera 0.7.2-4.22, mesa 1:26.2.4-1.4, omacam 0.3.3-1, omamessages 0.2.1-1, omamobile-apps 0.2.2-1, omamobile-device-raven 0.1.30-1, omamobile-firstboot 0.1.3-1, omamobile-keyring 20261026-1, omamobile-mirror 0.1.1-1, omamobile-omarchy 0.1.31-1, omamobile-owl-widgets 0.1.2-1, omamobile-shell 0.2.94-1, omamobile-voice 0.1.0-1, omaphone 0.2.2-1, omarchy 4.0.6-1, omatel-agent 0.2.4-2, quickshell 0.3.2.r7.g2b4501a-1

I've left the optional mif_rest < GS_DECON_MIF_FLOOR_HZ check out of this head so the run above matches it exactly; happy to add it as a follow-up if you'd like it.

@derekross

Copy link
Copy Markdown
Owner

Thank you, @robotosaith, for the full run. It covers everything I asked for: the release, the "link timers for
60 Hz" line, 120 Hz and back, DPMS, three s2idle cycles woken by Power with the display back at 120 Hz each time,
and the smoke report. The UFS lines it flags are the spm_lvl 3 resume errors your #188 fixes, so they're not from
this series.

Our gate passes on today's omamobile (a clean merge, all nine signed off, checkpatch --strict and W=1 clean, no
new device-tree warnings), and CI was green on this head.

I'm happy to take the mif_rest < GS_DECON_MIF_FLOOR_HZ check as a small follow-up, as you offered: it makes the
7 Pro's drop from 120 Hz free of the firmware read as well. One heads-up for later: #160, the Pixel 8 Pro display,
touches the same two bindings and gs-decon.c, so whichever lands second will need a rebase. I'll send the
resolved hunks to whoever that is.

Thanks again for seeing this through so many rounds. It's good to have raven's display on the published branch.

@derekross

Copy link
Copy Markdown
Owner

Thanks again, @robotosaith. Before merging I re-ran everything on today's omamobile, which has moved 26 commits since your rebase. The series still merges cleanly and rebases commit by commit without a conflict. Our gate passes (all nine signed off, checkpatch --strict and W=1 clean, no new device-tree warnings). Raven, oriole and bluejay validate with no messages against the series' bindings. The 7 Pro's, Fold's, 7a's, 7's, 8 Pro's, 8a's and 9a's DTBs are byte-identical to the base. On the 7 Pro the only change is still the one ACPM read of the MIF rate when leaving 120 Hz.

One correction, and it's mine. The sentence I asked you to add to "put MIF back after a 120 Hz stint" is wrong. On the Pixel 7 Pro, MIF isn't at its top level: every gs-decon probe in my logs reads "MIF at 2028000000 Hz", and the firmware's own rate read in my ACPM notes says the same. The conclusion still holds, because 2028 MHz is above the 1352 MHz floor and nothing else votes on MIF there, so nothing is written. But the reason I gave you isn't true, and kernel commits land as they are. Could you swap that paragraph for something like this?

On the Pixel 7 Pro MIF reads 2028 MHz at probe, above the 1352 MHz floor,
and nothing else votes on it, so this costs one ACPM read when leaving
120 Hz and writes nothing there.

That's the only change. Once it's pushed I'll merge right away. Your mif_rest < GS_DECON_MIF_FLOOR_HZ follow-up would now skip even that read on the 7 Pro, so it's still welcome. Sorry for handing you a wrong fact.

gs101 (Tensor) has the same DECON and DSIM blocks as gs201 (Google's
cal_9845 register layout); only DECON0's address and the interrupt
numbers differ. Describe both with gs201's compatible as the fallback.

Assisted-by: Claude Opus 5.5
Signed-off-by: saiththerobo <the.tom.sharp@gmail.com>
gs101's DECON is gs201's (Google's cal_9845 register layout), and it
falls back to gs201's compatible. Give "google,gs101-decon" its own match
data, gs201's for now, so that behaviour found and tested on the Pixel 6
Pro only can be turned on for gs101 without changing the Pixel 7 Pro. The
DSIM needs nothing of its own and matches through the gs201 fallback.

Assisted-by: Claude Opus 5.5
Signed-off-by: saiththerobo <the.tom.sharp@gmail.com>
The S6E3HC3 in the Pixel 6 Pro is the S6E3HC4's predecessor, with the
same supplies and reset line. The bootloader leaves it running and its
driver takes it over without a reset, so reset-gpios is required only
for the S6E3HC4.

Assisted-by: Claude Opus 5.5
Signed-off-by: saiththerobo <the.tom.sharp@gmail.com>
The Pixel 6 Pro's bootloader leaves the S6E3HC3 running in DSI command
mode. Take it over as it is, like the S6E3HC4 driver: no reset or rail
changes, nothing sent until the backlight changes.

Brightness is DCS 0x51 in the normal range 4..2047 (Google's GPL
driver). Level 0 is the dimmest level, not off: Omarchy darkens with
level 0 before locking and puts the level back afterwards, and when the
put-back did not happen the panel stayed dark with nothing to wake it.
A blanked backlight turns the display off (0x28, then 20 ms), and the
next unblanked update turns it back on. Since a reloaded driver cannot
know whether the panel was left with its display off, its first
unblanked update sends display on.

The refresh rate switches between 60 and 120 Hz with the vendor
driver's manual mode: unlock F0, BD 21 (manual), 60 00 for 120 Hz or
60 01 for 60 Hz, F7 0F (frequency update), lock F0. The panel takes the
new rate at its next frame without a power cycle; gs-dsim-cmd switches
its link timers around it.

Assisted-by: Claude Opus 5.5
Signed-off-by: saiththerobo <the.tom.sharp@gmail.com>
Dropping the 120 Hz floor left MIF at 1352 MHz, screen off included,
since nothing else votes on it. Remember the rate at probe and set it
back when the floor goes (the CRTC disable drops to 60 Hz first).

On the Pixel 7 Pro MIF reads 2028 MHz at probe, above the 1352 MHz floor,
and nothing else votes on it, so this costs one ACPM read when leaving
120 Hz and writes nothing there.

Assisted-by: Claude Opus 5.5
Signed-off-by: saiththerobo <the.tom.sharp@gmail.com>
gs-decon evicts simpledrm from the boot framebuffer, but on raven
simpledrm probed 30 ms after gs-decon registered, found the aperture
free and stayed: Hyprland drove it as a second, invisible monitor
(Unknown-1) next to the panel. After the eviction, acquire the range
too, so a later simpledrm fails its own acquire with -EBUSY.

Only gs101 holds it: it adds a failure path to probe, and the Pixel 7
Pro (gs201) has not shown the race.

Assisted-by: Claude Opus 5.5
Signed-off-by: saiththerobo <the.tom.sharp@gmail.com>
With no fbdev console the suspend VT switch has nothing to show, but it
still takes the session away from the compositor: logind deactivates and
reactivates it, Hyprland (libseat) drops its input devices and reopens
them, and the power-key press that woke the Pixel 6 Pro from s2idle is
lost with the old device. Every wake needed a second press.

Tell the PM core a VT switch is only needed with the fbdev console, and
stop tracking the device on remove. Only on gs101 for now: the Pixel 7
Pro (gs201) runs without fbdev too, and its validated s2idle path keeps
its VT switch until this has been tested there.

Assisted-by: Claude Opus 5.5
Signed-off-by: saiththerobo <the.tom.sharp@gmail.com>
DECON0 at 0x1c300000 (frame start/done/extra on SPI 338/337/336), the
DPP0 read DMA at 0x1c0b0000 and DSIM0 at 0x1c2c0000 (SPI 345), from
Google's gs101-drm-dpu.dtsi, disabled. No power-domains: nothing may
power the DPU down under DECON while the bootloader's pipeline runs.

Assisted-by: Claude Opus 5.5
Signed-off-by: saiththerobo <the.tom.sharp@gmail.com>
gs-decon takes the bootloader's pipeline over from simpledrm: DECON0
into DSIM0, with the S6E3HC3 panel for brightness, display on/off and
60/120 Hz. gs-decon also stops the frames with the screen off
(screen_off), so the DSIM gets no google,decon phandle.

Scanout buffers come from a reusable 92 MiB pool between 2.5 and
3.5 GiB: DPP0 takes 32-bit addresses with the DPU SysMMU bypassed, and
its reads from 0xf2c00000 and up hung silently.

The framebuffer node holds the panel's vci (LDO22M) and vddi (BUCKA)
rails. When gs-decon takes over, simpledrm goes away and releases them;
the panel lost power and DECON's first shadow update never completed.
Keep both rails on: the S6E3HC3 driver doesn't switch them.

Assisted-by: Claude Opus 5.5
Signed-off-by: saiththerobo <the.tom.sharp@gmail.com>
@robotosaith
robotosaith force-pushed the omamobile-gs101-display-pr branch from c9ca976 to 78d4ce6 Compare November 10, 2026 07:35
@robotosaith

Copy link
Copy Markdown
Contributor Author

Thanks, and no worries about the MIF sentence. The paragraph is now yours word for word in "drm: google: gs-decon: put MIF back after a 120 Hz stint".

omamobile had picked up the Tensor G3 display in the meantime (c55bd90, 1483365, a87146c), so I rebased onto it too. The three conflicts kept both sides:

  • the DECON binding lists google,zuma-decon in the enum next to gs201 and zumapro, with gs101's items pair as before;
  • the DSIM binding has gs101-dsim, zuma-dsim and zumapro-dsim in the fallback enum;
  • gs-decon's header comment now says G3 and G4 both follow cal_9865. zuma_decon_data leaves hold_boot_fb and vt_switch_fbdev_only off, so the G3 runs as it did. zuma's DECON node has no mif clock either, so the MIF restore doesn't run there.

On 78d4ce6, tools/kernel-pr-check --build --merge against today's omamobile (2f890a5): it merges cleanly, all nine commits are signed off, checkpatch --strict is clean, both touched C files build with W=1, and the 10 device trees show no new warnings. With dtschema installed, dt_binding_check on the DECON, DSIM and S6E3HC4 bindings is clean, and CHECK_DTBS for raven, oriole, husky and cheetah reports nothing against them. This is a rebase with no code change, so I didn't boot it again. Raven's run from 11-03 was on c9ca9760a.

@derekross

Copy link
Copy Markdown
Owner

Thank you, @robotosaith. The paragraph is in, and the rebase over the Tensor G3 display is exactly right: zuma-decon and zuma-dsim kept next to gs101's entries in both bindings, the header comment covering G3 and G4, and no code change (a range-diff against c9ca9760a shows six commits identical and the other three differing only in the bindings' context, the comment and the commit message). So your raven run from 11-03 stands for this head.

Our gate passes on today's omamobile (clean merge, all nine signed off, checkpatch --strict and W=1 clean, no new device-tree warnings) and CI is green. The 7 Pro's, Fold's, 7a's, 7's, 8 Pro's, 8a's and 9a's DTBs are byte-identical to the base. I compared gs-decon.o function by function once more: the aperture hold and the VT-switch change are behind gs101's match data, and everything else is the 8-byte shift from the new field, apart from the MIF restore.

That last one is the only thing between this and the merge, and it's on me, not you. The restore isn't gated, so on my 7 Pro it adds one firmware read of the MIF rate at every drop from 120 Hz, which there means every screen-off. I said earlier I was fine taking that by reading, but our own rule for shared drivers says a change that reaches the 7 Pro runs on it first, and I should hold myself to it. So I'll boot this series on my phone (rates, DPMS, a few sleeps), and merge as soon as that's clean. There's nothing for you to push.

Your mif_rest < GS_DECON_MIF_FLOOR_HZ follow-up is still very welcome afterwards: with it the 7 Pro makes no new call at all. And when you next edit the description, the MIF sentence there still has my old, wrong reason; it doesn't land, so no hurry.

Thanks for your patience through all of this.

@derekross

Copy link
Copy Markdown
Owner

The 7 Pro run is done: published omamobile (253208c11b25) with this series merged, 7.3.0-rc6-omamobile-push1-g42be9bc424d4, 2026-11-10.

Step Result
Boot clean, no failed units; gs-decon … MIF at 2028000000 Hz at probe, as the commit message says
120 → 60 → 120 Hz each mode taken (hyprctl monitors read back 60.00 and 120.00)
DPMS off for 6 s, then on back at 120 Hz
Three s2idle cycles (RTC alarm, systemctl suspend, modem off for that boot) 3 of 3 woke; suspend_stats: 3 ok, 0 failed
Kernel log across all of it no "MIF back to", no MIF floor line, no ACPM timeout or error, no flip timeout, no UFS error-handler pass

So on the 7 Pro the restore reads and never writes, as you said. Thank you for waiting for this; it should have been run days ago. Merging.

@derekross
derekross merged commit df3a12b into derekross:omamobile Nov 10, 2026
1 check passed
@robotosaith
robotosaith deleted the omamobile-gs101-display-pr branch November 12, 2026 17:19
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants