Skip to content

arm64: dts: exynos: google: Add the Tensor G4 SoC and the Pixel 9a - #80

Closed
W4HO-ham wants to merge 4 commits into
derekross:omamobilefrom
W4HO-ham:tegu-dt-base
Closed

W4HO-ham wants to merge 4 commits into
derekross:omamobilefrom
W4HO-ham:tegu-dt-base

Conversation

@W4HO-ham

@W4HO-ham W4HO-ham commented Oct 27, 2026 •

Copy link
Copy Markdown
Contributor

Rebased onto omamobile now that the clock series (#73-#79) is merged. Only these four commits remain.

This adds the Tensor G4 SoC device tree and a first-light Pixel 9a board, in the same spirit as zuma-husky.dts:

  • dt-bindings: soc: samsung: exynos-sysreg: the APM, HSI0 and PERIC1 system registers, the blocks the USIs select their mode through. Each lists its clock, as PERIC0 and UFS do.
  • dt-bindings: serial: samsung: google,zumapro-uart, with google,gs101-uart as the fallback, as gs201 does.
  • zumapro.dtsi and zumapro-reserved-memory.dtsi: CPUs, PSCI, timer, GIC, every CMU with a driver, the eight pin controllers, the system registers, the PMU with reboot mode, both watchdogs, the APM mailbox and SRAM for ACPM, UFS and its PHY, USB 2.0 (dwc3, DRD PHY, eUSB2 PHY), the USIs and their I2C/SPI/UART children, DECON and the GPU. All are disabled unless a board turns them on. The soc children are sorted by address. The ect, dpm and dpm-stubs nodes are ABL's overlay targets, as on zuma.
  • zumapro-tegu.dts: ABL's selector nodes for the A0/A1 IPOP parts, the clocks, pins, sysregs and PMU, both watchdogs, ACPM with the S2MPG14's eUSB2 supplies, UFS with its GPIO-switched VCC, USB 2.0 peripheral, and the three keys. Display, touch, battery, sound and the radios come in later changes, one each.

Checks: CHECK_DTBS=y W=1 for zumapro-tegu.dtb is clean, with no dtc warnings and no schema errors. dt_binding_check passes for both schemas. checkpatch --strict on the series shows only the MAINTAINERS question for the new files.

Against the palmtop: I compared this DTB node by node with the one the Pixel 9a boots now (the integration kernel, 7.3.0-rc6-omamobile-tegu-s2mpg15-m1-g991042e2eff3-dirty), using a script that resolves phandles to paths and clock specifiers to their names in each tree's header. Every node and property here exists there with the same value. The only differences are the intended ones:

I haven't booted this tree's own kernel on the palmtop yet. That's next, and I'll post the result here.

@derekross

Copy link
Copy Markdown
Owner

Thank you for this. A first SoC and board file for the Tensor G4 is a big step, and it's careful work: everything disabled in the dtsi, the soc children in address order, the ABL overlay stubs, and no syscon on the PMU. I checked it against Google's GPL trees (kernel/devices/google/zumapro at 8831630eee2a, zuma at 42395e8adcb8, soc/gs at 204078758737):

  • Every USI's base, interrupt and sysreg offset matches zumapro-usi.dtsi with the names resolved through zumapro.h, and so do the GPU, UFS, USB, watchdog, mailbox, DECON, key-bank and gpn0 interrupts and the GIC regions. The A0/A1 IPOP selector values match zumapro-a0-ipop.dts and -a1-ipop.dts, and the TMU channel matches acpm_tmu.
  • Every PMU offset the board reaches through the zuma regmap (reboot mode, the watchdogs, UFS isolation, HSI0 status) is in zuma_pmu_registers.
  • Reserved memory: every reg is 12 bytes, every fixed region of zumapro-rmem.dtsi is there with the same base and size, the ten debug-snapshot regions agree with debug-snapshot-def.h, and the xhci pools come from zuma-usb.dtsi, which zumapro-usb.dtsi includes. The splash size is exactly 1080 × 2424 × 4.
  • Gate on today's omamobile (with clk: samsung: acpm: Add Tensor G4 (zumapro) with a paired GPU clock #84 and thermal: samsung: gs101_acpm_tmu: Add the Tensor G4 #86 merged in, so the whole stack builds): every commit builds, the Pixel 6, 6 Pro, 7 Pro, Fold, 7a and 8 Pro DTBs are byte-identical to the base, and zumapro-tegu.dtb validates with no messages. All 60 compatibles in it are documented now that dt-bindings: Add Tensor G4 USI6 bus identities #39 and dt-bindings: Pixel 9a (tegu) board and zumapro UFS sysreg #60 are in.

Two things before I merge:

  1. mem_sleep_default=s2idle. On pinctrl: samsung: zumapro: accept wake requests on the wake-up EINTs #59 you found that the 9a advertises PSCI SYSTEM_SUSPEND and said you'd put this in the board's bootargs. It's not in chosen/bootargs yet. Without it, a plain systemctl suspend on this DTB enters deep, and Linux has no Tensor SoC sleep sequence for that (on the Pixel 7 Pro the phone wouldn't wake). Could you add it, with a comment line like cheetah's?
  2. The boot you mentioned. As you say, this tree's own kernel hasn't booted on the 9a yet. For the first board DTB, with UFS as the root and USB as the only link, I'd like that boot before merging: uname -r, the commit it was built from, the tools/phone-smoke report, and that the phone stays up past the bootloader watchdogs' 60 seconds. Then a "Tested on" line in the board commit.

Smaller things, none blocking:

  • What force-restarts the 9a under our kernel? On the Pixel 8 Pro Power alone does nothing, so husky's bootargs add softlockup_panic=1 nmi_watchdog=panic sysctl.kernel.panic_on_rcu_stall=1 to let panic=10 reset it. If the 9a is the same, those three are cheap, and the button combination should go into the recovery notes either way.
  • abl@f8800000 and the two gap regions (0x913a0000, 0xfd7ff000) aren't in Google's zumapro tree. Keeping them reserved is the safe choice. Where do they come from, the stock /proc/iomem or the live stock DT? Could you record that comparison in device/tegu/NOTES.md? The dtsi says "until they're shown to be unused", so it would help to name the check that would show it.
  • The SoC commit says the reserved file carries the fixed regions from Google's tree. It also carries the gaps and zuma's xhci pools, so half a sentence would make that exact.
  • Once clk: samsung: zumapro: Add the HSI0 USI3 clocks #79 is rebase-merged, the clock commits under this PR get new ids, so this will need a rebase onto omamobile.

Thanks again. This is the base everything else for the 9a builds on, and it reads very cleanly.

@W4HO-ham
W4HO-ham force-pushed the tegu-dt-base branch 2 times, most recently from fd1ee01 to 87e87a9 Compare October 27, 2026 20:19
@W4HO-ham

Copy link
Copy Markdown
Contributor Author

Thanks for checking it so thoroughly. Pushed, rebased onto today's omamobile (the merged clock series) and onto #84's id-header commit, which the CPU and GPU nodes further up the stack use by name:

  1. mem_sleep_default=s2idle is in chosen/bootargs, with a comment like cheetah's, and the board commit's message says why. (The line is 124 columns; checkpatch notes it, but the string can't be split, and husky's is longer.)
  2. The boot of this tree's own kernel: still to do. It needs Dave at the palmtop's buttons in case it doesn't come up, so I'll run it at the next chance and post uname -r, the commit, the smoke report and the 60-second watchdog check. The "Tested on" line goes into the board commit then.

On the smaller points:

  • Force restart: on the 9a, Power held about 30 s, then Power, does restart under our kernel, unlike the Pixel 8 Pro. It has always come back that way, including from a hung shutdown and a suspend that never woke. I've left husky's three panic options out for now, and the button combination goes into the tegu notes.
  • abl@f8800000 and the two gaps: the ABL region isn't from /proc/iomem. It's predeclared in stock's own factory base DTB (32 MiB, which ABL shrinks to 16 MiB at handoff). The gaps (0x913a0000 + 0x60000 and 0xfd7ff000 + 0x1000) are two ranges the earlier bring-up kept reserved beyond Google's zumapro-rmem.dtsi. They lie next to GXP and the debug tail, in the private DRAM map we built the first boots from. I'll write the comparison and the check that would free them (stock's live /proc/iomem and reserved-memory nodes against ours, on a stock boot of slot A) into device/tegu/NOTES.md in the next tegu notes PR.
  • The SoC commit now says exactly what the reserved file carries: Google's fixed regions, zuma's xhci pools through zumapro-usb.dtsi, and the two gaps.

@derekross

Copy link
Copy Markdown
Owner

Thanks. mem_sleep_default=s2idle and its comment are in, and the SoC commit's description of the reserved file is exact now. The long bootargs line is fine.

Thanks for the other answers too:

  • Force restart: Power held for about 30 s, then Power, is already the default in the contributing guide, so a line in the tegu notes is all it needs.
  • abl@f8800000 and the two gaps: keeping them reserved is the safe direction. The stock /proc/iomem and reserved-memory comparison in device/tegu/NOTES.md, as you describe, will settle it. Since the gaps come from the early bring-up map rather than a public source, please say that in the note.
  • No spi-cs-setup-delay-ns on arm64: dts: exynos: google: zumapro-tegu: Add the touchscreen #81: your reason works for me.

What's left is the boot of this tree, as you said: uname -r, the commit it was built from, the tools/phone-smoke report and the phone still up past the bootloader watchdogs' 60 seconds. Then add the "Tested on" line to the board commit.

On my side the stack checks out on today's omamobile: every commit builds, the other six Google boards' DTBs are byte-identical, and zumapro-tegu.dtb validates with no messages. I'll merge #84 now. After the boot, please rebase the stack onto omamobile and push it to #87 (its branch didn't get your last push; details there). I'll merge #87 on its own, which lands this PR's commits with it.

@W4HO-ham

Copy link
Copy Markdown
Contributor Author

The reserved-memory note is in omarchy-mobile #100 (device/tegu/NOTES.md, "Reserved memory beyond Google's map"): abl@f8800000 from stock's factory base DTB, the two gaps from the early bring-up's private DRAM map and not a public source, and what would free them (a stock /proc/iomem, which this palmtop can't give with Arch on userdata). The force-restart line was already in the tegu notes. The boot of this tree is next when the palmtop is back in the morning; then the Tested-on line and the rebase and push to #87.

@derekross

Copy link
Copy Markdown
Owner

Thanks for the reserved-memory note in omarchy-mobile #100. It says exactly what I asked, including that the two gaps come from the early bring-up map rather than a public source.

Last night's push reached every branch this time. The stack is now one line on today's omamobile: #80, #81, #85, #87, #90, #96. Every commit is the version I checked yesterday, only rebased, and git dropped your copy of the header commit as expected now that #84 is merged. On my side:

So the only thing left is the boot. Since the whole stack can now land at once, please boot its top (#96's head) rather than this PR alone, so the test covers exactly what lands. I'd like to see uname -r, the commit it was built from, the tools/phone-smoke report (its display, touch, GPU and haptics sections cover the rest of the stack), and the phone still up past the bootloader watchdogs' 60 seconds. Then add the "Tested on" line to the board commit and push the stack once more. If the #90 binding wording and the ports move go into the same push, so much the better.

After that I'll rebase-merge #96 on its own, which lands all twelve commits, and close #80, #81, #85, #87 and #90 as landed there.

@W4HO-ham

Copy link
Copy Markdown
Contributor Author

Booted this head on the palmtop this morning:

  • Kernel: 7.3.0-rc6-omamobile-tegu-pr80-gebd7e5e2450e, built from ebd7e5e (the int tree's config, olddefconfig), with this tree's zumapro-tegu.dtb.
  • The DTB: applied as it is. The only overlay was the one that sets /chosen/bootargs (the palmtop's boot policy: masked services and the root guard, no hardware nodes); none of the integration tree's hardware overlays.
  • Up past the watchdogs: up 85 s at the smoke run, past the bootloader watchdogs' 60 s, on slot B. Then a normal reboot back to the integration build.

tools/phone-smoke:

Check Result Detail
boot (required) ok kernel 7.3.0-rc6-omamobile-tegu-pr80-gebd7e5e2450e, slot b, up 85s
storage guard (required) ok 62 partitions; only userdata and slot B's boot images writable
root (required) ok ext4 on userdata, 20G free
kernel health (required) ok no oops or BUG; 0 WARN, 10 error lines in dmesg
systemd units not ok failed: tegu-curated-autostart.service
USB gadget (required) ok bound to 11210000.usb
display not ok no connected DRM connector
GPU not ok no render node
session (required) not ok Hyprland isn't running, the shell isn't running
lock ok PAM file present (the shell didn't answer lock status)
touch not ok no multitouch input device
buttons ok gpio-keys
Wi-Fi skip no wireless interface
Bluetooth skip no controller
battery skip no battery power supply
sound cards skip none
sensors skip no sensor proxy
cameras not ok libcamera lists none (or a camera app holds them)

Every required check passes: boot, storage guard, root, kernel health (no oops or BUG), USB gadget. The "not ok" rows are what this PR doesn't describe yet: display (#90), GPU (#85), touch (#81) and cameras. So this PR's tree on its own gives a headless system that gets past the watchdogs, with storage and USB.

Next I'll add the "Tested on" line to the board commit, then rebase the stack onto omamobile and push it to #87.

@derekross

Copy link
Copy Markdown
Owner

Thank you for the boot report and the "Tested on" paragraph. The tree you booted (ebd7e5e2450e) is exactly this PR's new head (same tree), the paragraph says what was tested, and the phone staying up for 85 s on slot B with the storage guard and USB in place is what this commit needed.

Two things before the stack can land:

  1. A rebase. omamobile moved at about 14:30Z today (the Pixel 7 board and the soc_id binding), so it wasn't 6beaf6d62 any more when you pushed. Every branch from here up to arm64: dts: exynos: google: zumapro-tegu: Add Bluetooth #122 now conflicts in arch/arm64/boot/dts/exynos/google/Makefile, and because of the conflict GitHub ran no CI on any of the new heads. The fix is to keep both new boards:

    	gs201-lynx.dtb \
    	gs201-panther.dtb \
    	zuma-husky.dtb \
    	zumapro-tegu.dtb
    ...
    DTC_FLAGS_gs201-lynx += -@
    DTC_FLAGS_gs201-panther += -@
    DTC_FLAGS_zuma-husky += -@
    DTC_FLAGS_zumapro-tegu += -@
    

    I rebased the whole chain that way here (arm64: dts: exynos: google: Add the Tensor G4 SoC and the Pixel 9a #80 through arm64: dts: exynos: google: zumapro-tegu: Add Bluetooth #122, plus arm64: dts: exynos: google: zumapro-tegu: Add the fuel gauge and USB-C port controller #119 and arm64: dts: exynos: google: zumapro-tegu: Add the eUSB2 repeater #118 on top): everything else applies cleanly, the gate passes, the 9a's DTB validates with no messages against the new bindings, and the other seven Google DTBs (Pixel 7 included) are byte-identical to the base.

  2. The boot of the stack's top. My comment this morning asked for the boot of arm64: dts: exynos: google: zumapro-tegu: Add the haptics and the torch #96's head rather than this PR on its own, since arm64: dts: exynos: google: zumapro-tegu: Add the haptics and the torch #96 is what I'll merge. Your boot covers the board commit, but display, touch, GPU, the CPU DVFS and thermal zones and the haptics haven't run in the form that lands. After the rebase, please boot arm64: dts: exynos: google: zumapro-tegu: Add the haptics and the torch #96's head and post uname -r, the commit, the tools/phone-smoke report and the uptime. Please also use the DTB's own /chosen/bootargs for that boot, or list what your bootargs overlay changes, because otherwise the board's bootargs are the one part of the DT that hasn't run.

One heads-up for the order: kernel #55 (S2MPG14 interrupt and RTC) makes interrupts and google,vgpio2ap-sysreg required on pmic-1, and this stack's pmic-1 has neither (it can't yet: today's binding doesn't allow them, and zumapro has no VGPIO2AP syscon node). If #55 merged as it stands, the PMIC wouldn't probe on the 9a, and its eUSB2 and touch rails with it. So I'll land this stack first, and #55 has to bring the 9a's DT along (the syscon node and the PMIC's interrupt) or keep the interrupt optional. Details are on #55.

After that, I'll rebase-merge #96, which lands all twelve commits, and close #80, #81, #85, #87 and #90 as landed there.

… PERIC1

Add the Tensor G4 (zumapro) system-register blocks the SoC's USIs select
their mode through: APM (one UART), HSI0 (two SPI buses and an I2C
bus) and PERIC1 (one I2C bus). Each block's interface clock is
a gate the Tensor G4 clock driver exports, so they list clocks, as
PERIC0 and UFS do.

Signed-off-by: W4HO <87742609+W4HO-ham@users.noreply.github.com>
The Tensor G4 (zumapro) UART behind the APM's USI1 works as gs101's, with
a 256-byte FIFO, so add it with google,gs101-uart as the fallback, as
gs201 does.

Signed-off-by: W4HO <87742609+W4HO-ham@users.noreply.github.com>
Add zumapro.dtsi, the Tensor G4 SoC as far as Linux drives it today: the
CPUs, PSCI, the timer and GIC, every CMU with a driver (TOP, MISC, APM,
HSI0, HSI1, HSI2, PERIC0, PERIC1, G3D), the eight pin controllers and
their banks, the system registers the USIs and UFS use, the PMU with its
reboot mode, both cluster watchdogs, the APM mailbox and SRAM for ACPM,
UFS and its PHY, the USB 2.0 path (dwc3, the USB 3.2 DRD PHY wrapper and
the eUSB2 PHY), the USIs behind the cameras', battery's, haptics', NFC's,
touchscreen's and GNSS's buses, the Bluetooth UART, DECON and the Mali
GPU. Everything a board doesn't need is left disabled; zumapro-tegu.dts
turns on what the Pixel 9a uses.

zumapro-reserved-memory.dtsi carries the fixed firmware regions from
Google's GPL zumapro tree, the xhci pools of zuma-usb.dtsi (which
zumapro-usb.dtsi includes), and two small gaps next to them that the
earlier bring-up kept reserved; they stay reserved until shown unused.

Addresses, interrupts and bank layouts come from Google's GPL zumapro
device tree; clock ids and parents from the Tensor G4 clock driver.

Link: https://android.googlesource.com/kernel/devices/google/zumapro/+/8831630eee2ad095b85ed452944e6353940966a5/dts/zumapro.dtsi
Signed-off-by: W4HO <87742609+W4HO-ham@users.noreply.github.com>
Add a first-light board file for the Pixel 9a (tegu): ABL's selector
nodes for its A0 and A1 IPOP parts, the clocks, pins, system registers
and PMU, both watchdogs, ACPM with the S2MPG14's two eUSB2 supplies,
UFS behind its GPIO-switched VCC, USB 2.0 peripheral and the three
keys. That is enough to boot the root file system from UFS and reach it
over USB networking. Display, touch, battery, sound and the radios
follow in their own changes. The bootargs pick suspend to idle: the
firmware advertises PSCI SYSTEM_SUSPEND, but Linux has no Tensor SoC
sleep sequence for "deep".

The USB core also holds CMU_HSI0's sub-control clock: it clocks the
link's registers, the PHY turns it off in suspend, and dwc3 touches
those registers before it initialises the PHY again on resume.

Tested on the Pixel 9a: a kernel built from this tree with its
zumapro-tegu.dtb boots to the root file system on UFS, reachable over USB
networking, and stays up past the bootloader watchdogs (85 s at the
check). tools/phone-smoke passes boot, the storage guard, root, kernel
health and USB gadget.

Signed-off-by: W4HO <87742609+W4HO-ham@users.noreply.github.com>
@derekross

Copy link
Copy Markdown
Owner

Closing as landed: these commits went in with #81, thank you.

@derekross derekross closed this Oct 30, 2026
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