Skip to content

device/dipper: add a RAM-only first-light build path - #505

Merged
derekross merged 22 commits into
derekross:mainfrom
quatrixone:dipper-device-support
Nov 12, 2026
Merged

derekross merged 22 commits into
derekross:mainfrom
quatrixone:dipper-device-support

Conversation

@quatrixone

@quatrixone quatrixone commented Nov 7, 2026 •

Copy link
Copy Markdown
Contributor

Adds the Xiaomi Mi 8 (dipper) as a WIP, RAM-only first-light target. The descriptor refuses generic install, flash and push paths. The image boots with fastboot boot; its empty UFS write allowlist keeps every partition read-only. There is no persistent install or update path.

The runbook gates the verified backups, serial-pinned fastboot commands, forced restart and RAM boot. #330 is merged; the device row links the bring-up notes.

The builder uses the Dipper DTS from linux-omarchy-mobile #220. The EA8074 panel is disabled in this first-light DTS; its separate driver and attended panel test are in kernel PR #218. Touch changes from #219 are in the kernel tree but touch remains disabled pending a hardware test.

Validation on head 5d917b4b198f454123f2ac029a665325e2262288: tools/pr-check origin/main, tests/dipper-probe, and git diff --check pass. Both CI checks pass, including tests/all.

Attended RAM boot on 2026-11-11 used kernel/DTS #220 head 26aa77c317cd029d16e58890c532562ca275c9f0. boot.img SHA-256: 665fcbd682cd8e64b3ffec1ff7f4ab9ca33a26b9b47b11073be7a3c267abf2b1; uname -r: 7.3.0-rc6-omamobile-push1. /dev/sda was absent, every UFS partition read ro=1, and sdb–sdf read ro=1. The watchdog driver and feeder, simpledrm framebuffer, USB/SSH, and the reserved 0x85fc0000–0x8a8fffff memory range were verified. No UFS partition was written.

Physical screen state, keys, touch, GPU rendering, Wi-Fi and NFS link-loss recovery have not been tested in this run. Detailed boot records are in device/dipper/NOTES.md.

@derekross

Copy link
Copy Markdown
Owner

Thank you for this. A RAM-only first-light path is the right way to start a phone with no second slot, and the descriptor does what you say. BOOT_METHOD=ram-only and the empty UFS_RW make bin/omamobile and the installer refuse it before any phone command, as they do for the Pixel 4. The image packs an empty allowlist, so the shared init leaves every partition read-only and never hands over to userdata. The guard code goes into the image unchanged, the watchdog is taken over and fed, panic=10 is on the command line, and nothing writes the phone. The identifier and blob checks pass, and the device-related suites pass on the branch (CI is waiting for my approval).

A few things before it can merge:

1. The matrix row. DEVICES.md's matrix is for devices with merged notes, and the notes are still in review in #330. Until #330 merges, the Mi 8 belongs in the "Wanted" table, with a link to this PR and #330. Once #330 is in, a matrix row is welcome, with these changes:

  • the cells should say what they rest on. The 10-29 boot used an unpublished kernel (…-dirty) and a bring-up init that left logdump writable, so "testing (UFS read-only guard)" can't cite it. Until this builder's image has run, Storage, Display, Buttons and USB net are WIP, with the kernel PRs linked as other rows do ([kernel #218](https://github.com/derekross/linux-omarchy-mobile/pull/218)). A bare "device-dodge: follow-ups from #111 (fastboot refuse, install bind trap, park deliberate panic) #218" reads as a PR in this repository;
  • the row has 24 cells and the table has 25. The AoC column should be "-" (the phone has none), as on the Pixel 4;
  • please put it after the Pixel 4 and 4 XL pair rather than between them.

2. One run of this image. The description says no kernel build or phone probe ran for this PR. A device change needs a report from the phone. tools/phone-smoke is built around Pixels, so read-only checks will do. Please boot the image this builder makes once, attended, and record in the notes:

  • uname -r and the kernel PR heads it was built from;
  • the sha256 of boot.img;
  • grep . /sys/block/sd*/ro /sys/block/sd*/sd*/ro, every line 1, and the ufs-ro lines from dmesg;
  • the watchdog taking over and USB/SSH coming up;
  • what you didn't test.

3. The runbook in gated form. BRINGUP-M1 explains the recovery well. The phone steps need the usual shape, though: Expect/Pass, a stop condition, a rollback and a risk level for each step, read-only checks first. Also please pass -s <serial> on every fastboot command (getvar, boot), so nothing reaches another phone on the same computer. Please also add a step that confirms your backup's checksums before the boot.

Smaller things:

  • OUT defaults to ../out/dipper, outside the repository. Our builders write to the repository's gitignored out/<codename>/, so please use that.
  • modules_install runs on the host with LLVM=1, so it needs the host's llvm-strip. Everything else in the project builds in the container, and bin/omamobile modules already makes the modules tarball, so it could do this step too.
  • initramfs-inside.sh says "Adapted from tools/alioth/build-initramfs.sh", which is in no tracked file (it's in the open device-alioth: add experimental POCO F3 source build and guarded installer #25). Please drop it, or name the PR.
  • Patching the shared init by string replacement works, and the guard goes in untouched. The shared init already has a descriptor-driven module step (INITRAMFS_MODULES), though, and a watchdog step could be added the same way, so the image would run the checked-in init as it is. That doesn't have to happen now. In the meantime, a small test like tests/flame-probe would help: run build-boot.py on dummy inputs to check the header and the size limits, and run the patch step on initramfs/init to show the anchors still match.
  • FIRMWARE_DIR staging has no user in a RAM-only build. It could wait for the root filesystem PR.
  • device/dipper/kernel.config is right for bring-up. Most of what's in it (built-in PHYs, UFS glue) would break no other phone, though, and the release kernel is one build from the shared fragment. Given this ABL's 16 MiB kernel limit, I'd like the plan to say how a release kernel could ever fit. That's for later, not for this PR.
  • The PR description shows literal \n\n.

On order: this goes after #330, since its links and the matrix row need the notes. It builds only on linux-omarchy-mobile #218, because prune-modules.py and the vermagic check stop without the EA8074 module, and on #220 for the DTB. Its root compatible is xiaomi,dipper, which matches the descriptor. #219 adds touch, and #221 isn't needed for a RAM probe. The title is fine for the squash.

@derekross

Copy link
Copy Markdown
Owner

Thanks for working through the list. The Mi 8 now sits in the Wanted table with the right links. The runbook has a backup gate, a fastboot gate and a RAM boot gate, with -s on every fastboot command. OUT is out/dipper, the modules come from bin/omamobile modules, the alioth reference and FIRMWARE_DIR are gone, and there's a tests/dipper-probe. I ran it, tests/storage-guard and tests/device-overlays on the branch, and they pass.

What's still needed before merge:

1. One run of this image. This is the same request as last time, and the main one. Please boot the image this builder makes once, attended, and record in the notes:

  • uname -r and the kernel PR heads it was built from;
  • the sha256 of boot.img;
  • grep . /sys/block/sd*/ro /sys/block/sd*/sd*/ro, every line 1, and the ufs-ro lines from dmesg;
  • the watchdog taken over (qcom_wdt bound, the feeder running) and USB/SSH up;
  • grep -i -E 'reserved|System RAM' /proc/iomem (see the kernel [Device support request] Samsung Galaxy S22 Ultra #220 thread for why);
  • what you didn't test.

One expectation to put in that step: with kernel #220 as it stands, the screen will most likely go dark about 30 seconds after boot. I've explained why on #220. If the screen does go dark there, that's the regulator cleanup, not a hang.

2. The RAM boot gate's Pass. "USB/SSH checks succeed" should name the checks above, the read-only flags first, with "any partition reads 0" as a stop condition. The rollback ("power-cycle to the installed boot image") also needs the forced-restart combination, which I've asked for on #499.

3. Order. This branch now also carries #503's init change (your 4f7bc1f3 is #503's tree). That's fine, but it means #505 merges after #503. If #503 changes again, after #484, please carry the same version here, or rebase onto main once #503 is in.

Smaller things:

  • The builder's closing hint still says fastboot boot boot.img. Please make it fastboot -s <serial> boot boot.img, as the runbook has it.
  • The initramfs loads panel-samsung-ea8074, but [Device support request] Samsung Galaxy S22 Ultra #220's DTB keeps the panel disabled and gives its node no compatible, so the module can't bind. Leaving it out until the panel is enabled would also drop this PR's build dependency on kernel device-dodge: follow-ups from #111 (fastboot refuse, install bind trap, park deliberate panic) #218.
  • tests/dipper-probe checks that the patcher contains its anchors, not that initramfs/init does. The patcher's own asserts stop the build if they drift, so nothing breaks silently. Running the patch step on initramfs/init in the test would show the drift before a build, though.

The title is fine for the squash.

@quatrixone

quatrixone commented Nov 9, 2026 •

Copy link
Copy Markdown
Contributor Author

2026-11-09 device check update:

  • Local checks passed: tools/pr-check origin/main, tests/dipper-probe, tests/storage-guard, and tests/device-overlays.
  • The phone is reachable over USB SSH before entering fastboot. Its current recovery partition SHA-256 matches the recovery image in the checksum-verified 2026-10-29 19:01 snapshot. The historical order of that snapshot and the two diagnostic recovery writes is still unknown; device/dipper: record stage-0 facts and first light #330 records this limit.
  • A private live file-level userdata archive is complete (6.67 GiB compressed); zstd integrity and tar listing checks passed, and its SHA-256 matched the phone-side archive. This is not a raw partition image or atomic snapshot.
  • Attended physical test: Power alone held for more than 10 seconds restarted this Mi 8; Volume Down + Power entered fastboot. Read-only fastboot variables reported product dipper and unlocked=yes. No serial is recorded. The restart was tested from the running system, not from a simulated kernel hang.
  • The existing 15-partition checksum manifest verifies but is not a whole-GPT backup. No flash has been attempted.
  • The installed kernel is still the dirty October build, not this PR image. Its root is mounted read-write; the kernel log says setting the qcom watchdog timeout failed, and the feeder was not verified.
  • The exact PR image boot, its SHA, reserved-memory check, read-only storage values, watchdog feeder and USB/SSH checks remain pending. PR initramfs, device-cheetah, device-raven, device-lynx (versions at merge): guard the boot LUNs as the main LUN (#252) #484 and initramfs: add guarded NFS root for bring-up #503 remain open; this branch is stacked after initramfs: add guarded NFS root for bring-up #503.

@derekross

Copy link
Copy Markdown
Owner

Thank you for the new recovery gate. Putting "find out how to force a restart" ahead of the first RAM boot is the right order on a phone with no second slot, and I'd rather wait for that than have you guess. The rest of my last list is done: the RAM boot gate's Pass now names every check, any ro of 0 is a stop, the closing hint has -s, the panel module is out of the initramfs (so this no longer depends on kernel #218 to build), and tests/dipper-probe now runs the patch step on the checked-in initramfs/init. CI passes on this head.

One thing I missed last time: tests/dipper-probe isn't called from tests/all, so CI has never run it. I ran it by hand and didn't notice. Please add it to the suite loop in tests/all next to tests/flame-probe, and add bash -n tools/dipper/build tools/dipper/initramfs-inside.sh next to the flame builder's syntax check, so the anchor check you added runs on every pull request. If another PR adds a suite to that line first, it's a one-line conflict.

One correction, also mine. Last time I wrote that the screen would most likely go dark about 30 seconds after boot. That was true of kernel #220 when I wrote it, but you fixed #220 the same evening: the panel supplies are back on the framebuffer node, and simpledrm, which is built in, keeps them on. So step 4 should now expect the screen to stay lit. Something like: "The screen should stay lit: simpledrm holds the panel's VCI and VDDIO rails. If it goes dark about 30 s in, record dmesg | grep -i -E 'regulator|simple' and stop." The current reason, "the panel regulator is not enabled", can go.

Two smaller points:

  • In step 3, please say the forced restart is tried from stock, with Android or the stock boot image running, so that learning the buttons never depends on the image we're testing.
  • The NFS-root note relies on RuntimeWatchdogSec=60s, and your 11-09 notes say setting the qcom_wdt timeout to one minute failed. That path isn't part of this PR's RAM boot, so it's not a condition here, but the note should say it hasn't worked yet.

Nits, only if you're in the files anyway: device/dipper/kernel.config still explains CONFIG_DRM_PANEL_SAMSUNG_EA8074=m as "so the probe can try it" and CONFIG_TOUCHSCREEN_STMFTS=y as the touchscreen on i2c14, but neither node is in #220 now, so those comments are out of date.

The order is unchanged: this merges after #503. The title is fine for the squash.

@derekross

Copy link
Copy Markdown
Owner

Thank you for running it. This is what I hoped for: the image this PR builds, booted by RAM, attended, with /dev/sda removed, every partition read-only, the watchdog fed, USB and SSH up and the console on screen, and you found and fixed the missing PMIC keys along the way. Everything from my last comment is in as well: tests/dipper-probe and the builder's syntax check are in tests/all, step 4 expects a lit screen, step 3 says the restart was tried from Android, the NFS note says it's unverified, and the kernel.config comments match #220.

What's left:

1. A rebase. Main has added four suites to the same tests/all line since you branched, so the PR now conflicts and CI can't run. Main's line with tests/dipper-probe after tests/flame-probe resolves it. The branch also still carries #503 as it was before your last push there: the allowlist line in the NFS test and the reworded comments are missing here. The simplest way is to wait for #503 to land and then rebase onto main with git rebase --onto origin/main <#503's head>, so only this PR's own commits remain.

2. Step 4's whole-disk check. It now asks for every whole-disk ro except sda to read 1, and stops if another disk reads 0. That's right for main's guard today, but #484, which lands before #503 and this PR, guards every LUN the way it guards sda: it removes the whole-disk node and leaves the disk's own flag clear. After #484, sdb and the others read 0 with no node, and the step would stop on a guard that's working. A version that holds on both: "every whole-disk node (/dev/sda to /dev/sdf) is absent and every partition's ro is 1; record the whole-disk ro values, but don't check them."

3. The status line. It credits Volume Down to this PR's image, but the notes on #330 say that result came from the second boot and the current image hasn't been rechecked. "Volume Down produced KEY_VOLUMEDOWN on the second image; not yet rechecked on the current one" keeps it exact. And #219 was merged into this project's kernel, not upstream Linux.

4. One line about reboot-to recovery. The rescue image offers it, and on #330 I've asked what recovery holds now. Until that's answered, please add to the runbook that only reboot-to bootloader is used.

Not conditions: CONFIG_POWER_RESET_QCOM_PON=y in the Mi 8's fragment is fine for bring-up. It will move to the shared fragment when the Mi 8 gets a release kernel, since the release kernel is one build for every phone. And the two RuntimeWatchdogSec=60s bullets under "What remains" say the same thing, so one can go.

The order stays #484, #503, then this. The title is fine for the squash.

@quatrixone

Copy link
Copy Markdown
Contributor Author

I addressed the four follow-ups in e79fd369d275413298b4e2ed81c13257cd12a404 and ran tests/dipper-probe successfully.

  • Step 4 now checks that every whole-disk node /dev/sda–/dev/sdf is absent and every partition is read-only, while recording (not gating on) whole-disk ro values.
  • The status line attributes the key events to the second image and says they have not been rechecked on the current image. It also says omamobile-audio 0.5.6: earpiece mode, a Voice Call verb, and no stored Voice Call profile at startup #219 merged into linux-omarchy-mobile.
  • The runbook restricts the rescue return path to reboot-to bootloader until the recovery contents are known. The stock recovery image is now identified in device/dipper: record stage-0 facts and first light #330 notes.
  • The duplicate watchdog note is removed. REFGEN and EA8074 modules are included for the attended DSI test; a module-prune check retained them with dependencies.

The panel overlay test is recorded with the visible console, brightness and DPMS results. That test predates #484: /dev/sda was absent, all 85 partitions were ro=1, but /dev/sdb–/dev/sdf still existed with whole-disk ro=1; no UFS was written. #484 and #503 remain open, so I have not rebased yet. After they land, the all-whole-disk-node gate still needs a fresh boot on the rebased image.

@quatrixone

Copy link
Copy Markdown
Contributor Author

The latest #505 runbook update records the actual Power-hold result from installed Android (restart back into Android), and explicitly says it has not been tested under the RAM kernel. The current recovery image is identified as stock from its 4.9.337 kernel header; the rescue path still uses the tested reboot-to bootloader, since recovery-mode reboot itself has not been tested.

@quatrixone

Copy link
Copy Markdown
Contributor Author

The recovery gate now records the 2026-11-09 attended result: a 10-second Power-only hold restarted back into Android from both installed Android and the Omarchy Mobile RAM kernel. Volume Down + Power then entered fastboot from Android. A simulated hard kernel hang remains untested.

@derekross

Copy link
Copy Markdown
Owner

Thank you for the quick turn on all four points, and for two things beyond them: the full backup read through the guarded RAM boot, and the Power-hold restart tried under our kernel, which is the result I most wanted to see before anything gets flashed. Step 3 now records both.

Three things before this can merge. The second one is my mistake.

1. The panel pieces make the builder depend on kernel #218. e79fd369 adds CONFIG_DRM_PANEL_SAMSUNG_EA8074=m to the config and msm panel-samsung-ea8074 qcom-refgen-regulator to the module list in initramfs-inside.sh. The EA8074 driver isn't on omamobile yet; it's #218, still open. On omamobile plus #220, which is what the runbook builds, Kconfig drops the unknown symbol, no panel-samsung-ea8074.ko exists, and prune-modules.py stops with "not in modules.dep". So this head no longer builds the image you booted three times. The panel test is a good result and I'd like it in the tree, but as its own step: could you take the two config lines, the three extra module names and the two new test lines out of this PR, and bring them back as a follow-up on top of #218, with the overlay you used under device/dipper/overlays/ and a runbook step for the attended test (what you expect, when to stop, how to get back)? The result itself is already in the notes on #330, and that's the right place for it.

2. Step 4, where I gave you the wrong wording. I suggested "every whole-disk node (/dev/sda to /dev/sdf) is absent" as a version that holds before and after #484. It doesn't, and your panel boot showed why: on today's guard only /dev/sda is removed, and sdb to sdf stay, read-only at the disk level. As written, the step can only pass after #484. This holds on today's main, after #484, and in #503's NFS mode, where sda keeps its node but is read-only:

/dev/sda is absent, or present with /sys/block/sda/ro 1 (NFS mode). Each other whole-disk node is absent, or present with its ro 1. Every partition's ro is 1. Stop if any partition reads 0, or any whole-disk node exists with ro 0.

The sentence "Repeat the whole-disk node check after #484 is included" can then go.

3. The rebase, and no more waiting for #484. On 2026-11-09 I merged #530, which changed the same guard lines that #503 edits, so #503 and this PR now conflict with main itself. I've posted the resolution on #503. Once #503 has landed, git rebase --onto origin/main <#503's head> leaves only this PR's own commits; in tests/all, take main's loop line and put tests/dipper-probe after tests/flame-probe, and keep your bash -n line for the two builder scripts. #484 is mine to redo afterwards.

Two wording points:

  • "Stock". The notes on device/dipper: record stage-0 facts and first light #330 say the phone runs a custom Android 14 ROM on a 4.9.337 kernel, so the header check shows that recovery holds that ROM's recovery, not Xiaomi's. That still answers my question (it isn't the diagnostic image), but "the installed ROM's recovery" and "the backup of the installed system" say it exactly.
  • Which boot was which. The status section now mentions a second image, a panel image and a backup boot. I've asked for one list of the day's boots on device/dipper: record stage-0 facts and first light #330; here a pointer to it is enough.

The title is fine for the squash.

@quatrixone

Copy link
Copy Markdown
Contributor Author

Owner follow-ups are implemented in the updated head c4004c650de24b3a1ec811c78904753762964489: removed the EA8074 panel config/module/test pieces for the follow-up based on kernel #218, and aligned the runbook guard criteria with the current whole-disk-node check. tests/all, shell syntax checks, and git diff --check passed. I have not rebased onto #503 because it remains open, per your requested sequence.\n\nThe exact rebuilt image (0ee4f11d1973085c434e741a2e9dd73af5a9cd157296c25801ed5a92a6b9e9d9) booted via RAM on dipper with #220 at 055a4019998f22cc0e2a8afef7650274a6ce0657. All 85 partition flags were ro=1; sda had sysfs ro=0 but no /dev/sda node, while sdb–sdf were ro=1 and present. The guard confirmed this. qcom_wdt and feeder, simpledrm/fb0, USB/SSH, and the reserved TZ memory range were verified. reboot-to bootloader returned the phone to fastboot. No UFS write was attempted. Button events were not captured on this exact image.

@quatrixone

Copy link
Copy Markdown
Contributor Author

The owner follow-ups are implemented in the updated head b39f7ebc10d172e1c28f538633a9d4bcab1b7bf5. I also fixed the rescue SSH host-key mismatch: the init now passes the builder’s pinned Ed25519 key explicitly to Dropbear instead of asking it to generate a runtime key. tests/all, tests/dipper-probe, shell syntax checks and git diff --check pass.

The combined #503/#505 non-NFS image booted by RAM on the Mi 8 (SHA-256 540b5d31ea73a31ade9a8d8cb5152cf701dc3c2ad8256734790b0ef2db6dbdd0, kernel 7.3.0-rc6-omamobile-push1). The initramfs root was active; all 85 partitions were ro=1; sda had sysfs ro=0 but its /dev/sda node was absent, and sdb–sdf were ro=1. qcom_wdt and its feeder ran, simpledrm registered fb0, USB/SSH worked with the pinned key, the TZ range was reserved, and reboot-to returned to fastboot. No UFS write was attempted. Physical button events were not captured on this exact image. The NFS run is recorded in the updated runbook and #330 notes.

I have not rebased #505 yet because #503 has not landed; the PR order you requested is preserved.

@quatrixone
quatrixone force-pushed the dipper-device-support branch from b39f7eb to a952d38 Compare November 10, 2026 08:35
@quatrixone

Copy link
Copy Markdown
Contributor Author

Resolved the base conflict by rebasing onto current main. Dropped the stale shared #503 commit; #505 now contains only its Dipper changes on top of main. tests/all, tools/pr-check origin/main, and git diff --check pass. GitHub reports this PR as CLEAN/MERGEABLE. Keeping the requested #503-before-#505 merge order.

@derekross

Copy link
Copy Markdown
Owner

Thank you, this is nearly there. The panel pieces are out, so the builder needs only what #220 provides again. Step 4 now holds on today's guard, after #484 and in NFS mode. The rebase is clean on main, and I also merged it together with #503 and ran tests/dipper-probe, tests/storage-guard and the grep -q lint on the result: all pass, and the patched init parses. Pinning the rescue host key is a good fix for what cost you the first NFS session, and the two new boots (0ee4f11d… and the combined 540b5d31…) are recorded with their limits.

Three small text changes are left, none needing a boot. If you'd rather, I can make them when I merge:

  1. "Stock" in the runbook. It still says "stock backup", "restore the verified stock backup" and "the stock recovery image". Your notes say the phone runs a custom Android 14 ROM, so "the backup of the installed system" and "the installed ROM's recovery" are exact. In the same places, the rollback "restore the verified stock backup using the device recovery procedure" names a procedure we don't have yet. Since this runbook writes nothing, "there is no reviewed restore procedure yet; nothing here writes UFS" says it as it is.
  2. The DEVICES row. "Under review: device/dipper: add a RAM-only first-light build path #505; notes await device/dipper: record stage-0 facts and first light #330" is stale the moment this merges. Something like: "WIP, community port: the RAM-only first-light probe is merged (BRINGUP-M1, NOTES); single-slot, no install path yet. Kernel: touch omamobile-audio 0.5.6: earpiece mode, a Voice Call verb, and no stored Voice Call profile at startup #219 merged, board DTS [Device support request] Samsung Galaxy S22 Ultra #220 and panel device-dodge: follow-ups from #111 (fastboot refuse, install bind trap, park deliberate panic) #218 in review."
  3. Step 2 points to PR CONTRIBUTING: add Dipper single-slot recovery entry #499 for the §3.1 entry. CONTRIBUTING: add Dipper single-slot recovery entry #499 is merged, so it can cite CONTRIBUTING §3.1 directly. Likewise "[the notes PR device/dipper: record stage-0 facts and first light #330]" can point at NOTES.md once device/dipper: record stage-0 facts and first light #330 is in.

Suggestions, not conditions: the builder mounts the whole repository into the container, and initramfs/, tools/dipper/ and reboot-to.c would be enough; the patched init prints "ssh: root@172.16.42.1" even when usb0 had no address and dropbear wasn't started; and tests/dipper-probe could run sh -n on the patched init.

The order stays #503, then this. #503 is waiting on one boot of my own phone, not on you. The title is fine for the squash.

@quatrixone
quatrixone force-pushed the dipper-device-support branch from a952d38 to 4edbe44 Compare November 11, 2026 07:32
@quatrixone

Copy link
Copy Markdown
Contributor Author

Addressed the remaining review notes in 4edbe44: the runbook now names the installed-system backup and says no reviewed restore procedure exists; it links CONTRIBUTING.md §3.1 and calls the recovery image the installed ROM’s. The device row describes the RAM-only WIP and current kernel status. The container now mounts only the four build inputs, SSH is printed only after Dropbear starts, and dipper-probe syntax-checks the patched init. Rebased onto current main after #503 merged. tests/all, tests/dipper-probe, tools/pr-check origin/main, shell syntax checks, and git diff --check pass. #330 remains open, so the row still links its notes PR. GitHub reports CLEAN/MERGEABLE; no CI checks are currently reported.

@derekross

Copy link
Copy Markdown
Owner

Thank you, @quatrixone. The runbook now says "the backup of the installed system" and "the installed ROM's recovery", the rollback says plainly that there is no reviewed restore procedure yet, step 2 points at CONTRIBUTING §3.1, and the DEVICES row reads as I suggested. The gate and CI pass on this head.

One of my own optional suggestions has broken the builder, so one more small push, please:

tools/dipper/build runs a script the container can't see. The whole-repository mount is gone, but line 56 still runs sh /src/tools/dipper/initramfs-inside.sh, and nothing mounts /src any more: the four new mounts are init, prune-modules.py, patch-init.py and reboot-to.c. Mounting initramfs-inside.sh under /input as well and running it from there fixes it. tests/dipper-probe only syntax-checks the builder, which is why CI is green; a check there that every /input/… path the two scripts use has its -v line would catch this next time. Please run the builder once at the new head and post the boot.img sha256. No boot is needed: the booted images stay the hardware record.

Two small things for the same push:

The order is #330, then kernel #220, then this. The title is fine for the squash.

The Mi 8 has one OS boot partition, so a failed persistent update has no
known-good OS slot to return to. Keep this first integration RAM-only:
the descriptor is non-installable, the builder emits an Android v1 image
for fastboot boot, and its storage allowlist is empty.

Add the Dipper kernel fragment, a bounded image packer, a read-only
initramfs build, recovery instructions, and a WIP device matrix row. Wi-Fi
identity, firmware and any persistent install path stay out until their
own safety gates are met.

Checks: git diff --check and bash -n passed. Kernel build and hardware
probe not run in this session.
@quatrixone
quatrixone force-pushed the dipper-device-support branch from 4edbe44 to c4238ab Compare November 11, 2026 17:41
Signed-off-by: quatrixone <quatrix1@gmail.com>
Signed-off-by: quatrixone <quatrix1@gmail.com>
@quatrixone

Copy link
Copy Markdown
Contributor Author

Completed the follow-up on head 2d67708e07728b13b5d2a9eda299ace8a1608163.

  • The builder now mounts initramfs-inside.sh at /input and runs it from there. tests/dipper-probe checks every /input file has a read-only mount.
  • The init patcher replaces the complete shared Dropbear line, so SSH is announced only after Dropbear starts with the USB address.
  • tools/pr-check origin/main and tests/dipper-probe pass; both CI checks pass.
  • Rebuilt from kernel/DTS [Device support request] Samsung Galaxy S22 Ultra #220 head 26aa77c317cd029d16e58890c532562ca275c9f0. boot.img SHA-256: 665fcbd682cd8e64b3ffec1ff7f4ab9ca33a26b9b47b11073be7a3c267abf2b1. This image was booted with fastboot boot only.
  • The phone reported 7.3.0-rc6-omamobile-push1; /dev/sda was absent, all 85 partitions read ro=1, and sdb–sdf read ro=1. The guard logged every partition. qcom_wdt and its feeder, simpledrm fb0, authenticated USB/SSH, and the reserved 0x85fc0000–0x8a8fffff range were verified.
  • No UFS partition was written. Physical screen state, keys, touch, GPU rendering, Wi-Fi, and NFS link-loss recovery were not tested in this run.

The new run is recorded in device/dipper/NOTES.md.

@derekross

Copy link
Copy Markdown
Owner

Thank you, @quatrixone. Everything from last time is done, and a bit more:

  • The builder mounts initramfs-inside.sh under /input and runs it from there, and tests/dipper-probe now checks that every /input/… file has its read-only mount. That is the check that would have caught my own bad suggestion, so thank you for adding it.
  • The patcher takes the whole shared Dropbear line, so SSH is announced once, and only when Dropbear has started on usb0's address.
  • You rebuilt at the new head, posted the image's sha256, and then booted it as well, on kernel [Device support request] Samsung Galaxy S22 Ultra #220's final head, with the read-only state, the watchdog and the reserved range recorded in NOTES along with what you didn't test. I only asked for the build; the boot is welcome.

Here: tools/pr-check passes, CI is green, it merges cleanly onto today's main, and on the merged result tests/dipper-probe passes against the shared init as it is now.

I'm merging this after linux-omarchy-mobile#220. Three lines will be slightly off once both are in, and I'll fix them on main myself rather than ask you for another push (say so if you'd rather do it):

  • in BRINGUP-M1.md, [NOTES.md] has no link definition at the end of the file (the #330 one was removed), so it shows as plain bracketed text; it needs [NOTES.md]: NOTES.md;
  • the DEVICES row and BRINGUP-M1 say the board DTS [Device support request] Samsung Galaxy S22 Ultra #220 is "in review";
  • BRINGUP-M1's status line is dated 2026-11-10 and names the 540b5d31… image, while the newest boot in NOTES is the 2026-11-11 one, 665fcbd6….

Thank you for your patience through all these rounds. The Mi 8 now has a reviewed, read-only way to first light, and the panel (#218) and touch tests can build on it.

@derekross
derekross merged commit 92074ca into derekross:main Nov 12, 2026
2 checks passed
derekross added a commit that referenced this pull request Nov 12, 2026
…i 8 lines #505 left (#680)

The Pixel 8a's row still described first light. Since then Omarchy's
image runs from userdata (BRINGUP-M2 gate e, #456 and kernel #206
merged), and the buttons and the touchscreen passed gates f and g
(kernel #238 and #241 merged): Boot, Storage, Touch and Buttons say so.

For the Mi 8, three lines promised on #505: BRINGUP-M1's [NOTES.md] link
had no definition, the board DTS (kernel #220) is merged, not in review,
and the status line names the 2026-11-11 boot NOTES records.

The POCO F3 row said #25 was in review; it was closed as paused on
2026-11-11.

Co-authored-by: Claude Fable 5.2 <noreply@anthropic.com>
robertzas pushed a commit to robertzas/omarchy-mobile that referenced this pull request Nov 14, 2026
Adds the Xiaomi Mi 8's descriptor (`BOOT_METHOD=ram-only`, no boot images, an empty UFS write allowlist), its kernel config fragment, a builder for a `fastboot boot` image (`tools/dipper/build`: the shared init patched for the Qualcomm watchdog and a USB-only SSH server, in an Android v1 boot image with the DTB appended) and a gated runbook in BRINGUP-M1. Nothing here flashes, pushes or installs: the phone has one OS slot and no reviewed restore procedure yet. Booted by RAM on 2026-11-11 with linux-omarchy-mobile derekross#220's device tree (image sha256 `665fcbd6…abf2b1`): every UFS partition read-only, watchdog fed, simpledrm and USB SSH working; keys, touch, GPU and Wi-Fi untested.
robertzas pushed a commit to robertzas/omarchy-mobile that referenced this pull request Nov 14, 2026
…i 8 lines derekross#505 left (derekross#680)

The Pixel 8a's row still described first light. Since then Omarchy's
image runs from userdata (BRINGUP-M2 gate e, derekross#456 and kernel derekross#206
merged), and the buttons and the touchscreen passed gates f and g
(kernel derekross#238 and derekross#241 merged): Boot, Storage, Touch and Buttons say so.

For the Mi 8, three lines promised on derekross#505: BRINGUP-M1's [NOTES.md] link
had no definition, the board DTS (kernel derekross#220) is merged, not in review,
and the status line names the 2026-11-11 boot NOTES records.

The POCO F3 row said derekross#25 was in review; it was closed as paused on
2026-11-11.

Co-authored-by: Claude Fable 5.2 <noreply@anthropic.com>
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