drm/apple-g5p: Bound the kexec handover and fix two lost-GPU races by brkosebay · Pull Request #125 · OrchardLinux/orchard

brkosebay · 2026-11-12T16:13:08Z

Part 11 of 11 of the iPad 5 (J71t) series in this repository; merge after #120.

Problem

A warm restart could jump into the next kernel while the GPU still had DMA in flight, or hang in an unbounded wait for the firmware to stop. Two races could also lose the GPU: with two jobs in flight, the previous job's 3D completion could make a new job's 3D phase look retired, so its doorbell was never rung and the job was lost after a silent second; and when the firmware's ALIVE overtook the GPU cores' power-on request, the busy request was counted as a failed power-on and the boot fell back to software rendering.

Change

Bound and serialise the GPU handover, accept an already completed handover, keep halt and poweroff working after a quiesce failure, and flush the consoles before an emergency cold reset. Ring a render phase's doorbell unless the phase's own stamp shows it retired, as Apple's TA and 3D submit paths ring after every ring entry. When ALIVE overtakes a power-on request that is still in flight, let the first-ALIVE handler wait for the request's busy bit to clear within the power-on bound, as Apple's power-state code does, before any GPU register write.

Testing

Hardware: iPad 5 Wi-Fi (J71t). Hands-on testing on 2026-11-11 with an integrated build of this series, made before the final failure-path fixes, covered warm restart with the GPU in use.
Not covered on hardware: the cold-reset escape when the handover fails. The power-on race is rare (twice in a day of boots before the fix), so the fix is not yet shown on hardware.
2026-11-12, doorbell fix: glmark2 terrain ran 20 of 20 and then 5 of 5 times without a lost job (before the fix, 2 of about 11 runs lost a 3D job). 20+ warm restarts with the GPU in use, several right after a glmark2 run, completed without a freeze.
Host: the Linux test suite passes at this branch's head, and the kernel and m1n1 patch series apply on apple-16k and apple-4k. At the top of the series the applied kernel tree differs from the built release image only where main took earlier PRs of this series in reworked form, by the closed perf: apple: A9 memory port byte counters #123, and by the new G5P sleep refusal (drivers/drm 0045, not yet run on the device).
Other boards that share these drivers or compatibles have not been tested.

Dependencies

Depends on #120.

This branch is stacked on the PRs it depends on and carries their commits; the change of this PR is the commit "drm/apple-g5p: Bound the kexec handover and fix two lost-GPU races", and the merge commits after it bring the branch onto current main.

This update keeps the branch's history: it merges current main and the updated branch below it into this branch. Nothing is forced.

🤖 Generated with Claude Code

Add the A9 (S8003) G5P render driver used by the merged PowerVR/G5P Mesa winsys as kernel patches, enable it on J71t in the device tree, set CONFIG_DRM_APPLE_G5P=y for apple-16k, and add an install-time recipe that extracts the original META LDR (apple/agx-g5p-c0.ldr) from the device's own iOS kernelcache into /lib/firmware. No firmware is committed. The driver provides TYPED_V3 rendering, GEM buffers, VM/arena mappings, PRIME import/export, syncobj, bounded retirement and recovery. It does not expose the old compute or fixed-fixture ABI. Hardware: the exact kernel this derives from (brkosebay/linux a9/gpu-pr1 407b7bf2cbbe) booted on a J71t with the merged Mesa, rendered labwc on the GPU with KMS scanout at 60 Hz, and passed the G5P device rungs with zero GPU events. This Orchard adaptation itself has not yet been built by orchard or run on hardware. Assisted-by: LLM

Assisted-by: LLM

kernel/configs/apple-16k/config: CONFIG_DRM_APPLE_G5P=y These options provide g5p and its Kconfig dependencies on both page sizes. Assisted-by: LLM

kernel/configs/apple-16k/config: # CONFIG_NVME_APPLE_ANS_EARLY_USB is not set These options provide warm-storage and its Kconfig dependencies on both page sizes. kernel/configs/apple-4k/config: # CONFIG_NVME_APPLE_ANS_EARLY_USB is not set These options provide warm-storage and its Kconfig dependencies on both page sizes. Assisted-by: LLM

kernel/configs/apple-16k/config: CONFIG_APPLE_AOP_GYRO=m CONFIG_APPLE_AOP_PRESSURE=m These options provide aop and its Kconfig dependencies on both page sizes. Assisted-by: LLM

Assisted-by: LLM

kernel/configs/apple-16k/config: # CONFIG_NVME_APPLE_ANS_EARLY_USB is not set These options provide warm-storage and its Kconfig dependencies on both page sizes. kernel/configs/apple-4k/config: # CONFIG_NVME_APPLE_ANS_EARLY_USB is not set These options provide warm-storage and its Kconfig dependencies on both page sizes. Assisted-by: LLM

Assisted-by: LLM

kernel/configs/apple-16k/config: CONFIG_APPLE_AOP_GYRO=m CONFIG_APPLE_AOP_PRESSURE=m These options provide aop and its Kconfig dependencies on both page sizes. Assisted-by: LLM

Assisted-by: LLM

kernel/configs/apple-16k/config: # CONFIG_NVME_APPLE_ANS_EARLY_USB is not set These options provide warm-storage and its Kconfig dependencies on both page sizes. kernel/configs/apple-4k/config: # CONFIG_NVME_APPLE_ANS_EARLY_USB is not set These options provide warm-storage and its Kconfig dependencies on both page sizes. Assisted-by: LLM

Assisted-by: LLM

The VXD decoder only exposed H.264, so HEVC video was decoded on the CPU. Prepare the driver for a second coded format, add the HEVC command-stream builder and the stateless HEVC decoder, enable it on the S8000 and T8010 variants, and require a proven firmware stop before kexec. Adds the video_decode_hevc feature (working on J71t, untested on the other boards that have the driver, missing elsewhere). Assisted-by: LLM

Bring this branch onto current main and the series it depends on. The merge keeps the other side's tree; the next commit carries the reworked version of this branch's change.

S8000/S8003 left the SoC and memory performance state at the boot default, which bounds GPU throughput by memory bandwidth. Describe the SoC performance state in the binding and DT, run the SoC and memory at the high state while active, and restore the low state for sleep. Assisted-by: LLM

A Mac grants an iPad more than the standard USB current, but the charger took the gadget's reported budget unchecked and the composite gadget could draw the extra current while the bus was suspended. Cap and log the USB host budget in the TriStar charger, draw no extra Apple current while suspended, and let the initramfs and orchard USB helpers opt J71t into the Mac current profile. Documented in docs/bring-up/charging/j71t-mac-current.md, with tests. Assisted-by: LLM

There was no way to measure memory bandwidth per port on A9, which is needed to tell GPU work from memory stalls. Add a binding, DT node and perf driver for the A9 AMC port monitor, exposing CPU and GFX read/write byte counters, and enable it for apple-16k. Assisted-by: LLM

Bring this branch onto current main and the series it depends on. The merge keeps the other side's tree; the next commit carries the reworked version of this branch's change.

Each render job paid a fixed host cost (serial submission, polling after events, uninterruptible VM binds), which capped frame rates well below the GPU's capacity. Account per-job GPU and host time, reduce serial job overhead, make VM binds killable and tune post-event polling, record joint latency, pipeline two render jobs deep, honour fragment-only waits, start the next job early during the 3D wait, gate TA PM records on dirty state, and default to the measured pipeline configuration. Assisted-by: LLM

The GPU blocked every system sleep, and an idle GPU kept its power on because the firmware was never allowed to park itself. Let the firmware self-park after bring-up and ring a parked firmware without the GPC settle wait. Implement a bounded retained system sleep: register the reservation as a suspend admission, hold clients across it, answer the readiness query, and refuse sleep with a named reason while work, a kexec power hold or pmgr's gfx hold is active. pmdomain ends the warm boot hold at sync_state, the gfx gate is cycled with auto-PM like pmgr-pwrstate, and the other admission users name their refusals too. Assisted-by: LLM

A warm restart could jump into the next kernel while the GPU still had DMA in flight, or hang in an unbounded wait for the firmware to stop. Two races could also lose the GPU: with two jobs in flight, the previous job's 3D completion could make a new job's 3D phase look retired, so its doorbell was never rung and the job was lost after a silent second; and when the firmware's ALIVE overtook the GPU cores' power-on request, the busy request was counted as a failed power-on and the boot fell back to software rendering. Bound and serialise the GPU handover, accept an already completed handover, keep halt and poweroff working after a quiesce failure, and flush the consoles before an emergency cold reset. Ring a render phase's doorbell unless the phase's own stamp shows it retired, as Apple's TA and 3D submit paths ring after every ring entry. When ALIVE overtakes a power-on request that is still in flight, let the first-ALIVE handler wait for the request's busy bit to clear within the power-on bound, as Apple's power-state code does, before any GPU register write. Assisted-by: LLM

Bring this branch onto current main. Main has since merged OrchardLinux#124, OrchardLinux#122, OrchardLinux#118, OrchardLinux#121, OrchardLinux#110, OrchardLinux#127, OrchardLinux#101, OrchardLinux#126, OrchardLinux#108 and OrchardLinux#111, several in part or reworked, and OrchardLinux#123 and OrchardLinux#128 were closed. Where main carries a topic, main's version is kept; the left-out parts that warm restart and suspend need come back in OrchardLinux#107 and OrchardLinux#117.

Bring this branch onto current main. Main has since merged OrchardLinux#124, OrchardLinux#122, OrchardLinux#118, OrchardLinux#121, OrchardLinux#110, OrchardLinux#127, OrchardLinux#101, OrchardLinux#126, OrchardLinux#108 and OrchardLinux#111, several in part or reworked, and OrchardLinux#123 and OrchardLinux#128 were closed. Where main carries a topic, main's version is kept; the left-out parts that warm restart and suspend need come back in OrchardLinux#107 and OrchardLinux#117. This branch's own patches in drivers/usb are renumbered to follow main's.

Bring this branch onto current main. Main has since merged OrchardLinux#124, OrchardLinux#122, OrchardLinux#118, OrchardLinux#121, OrchardLinux#110, OrchardLinux#127, OrchardLinux#101, OrchardLinux#126, OrchardLinux#108 and OrchardLinux#111, several in part or reworked, and OrchardLinux#123 and OrchardLinux#128 were closed. Where main carries a topic, main's version is kept; the left-out parts that warm restart and suspend need come back in OrchardLinux#107 and OrchardLinux#117. This branch's own patches in devices/s8000, drivers/base and drivers/usb are renumbered to follow main's.

Bring this branch onto current main. Main has since merged OrchardLinux#124, OrchardLinux#122, OrchardLinux#118, OrchardLinux#121, OrchardLinux#110, OrchardLinux#127, OrchardLinux#101, OrchardLinux#126, OrchardLinux#108 and OrchardLinux#111, several in part or reworked, and OrchardLinux#123 and OrchardLinux#128 were closed. Where main carries a topic, main's version is kept; the left-out parts that warm restart and suspend need come back in OrchardLinux#107 and OrchardLinux#117. This branch's own patches in devices/s8000, drivers/base, drivers/pmdomain and drivers/usb are renumbered to follow main's.

The A9 AOP driver on main reads apple,aop-target, apple,chip-id and apple,chip-revision, and the motion sensor drivers read their factory calibration, from properties the bootloader fills in. Nothing on main fills them in; going by the driver's probe, the A9 AOP cannot start without them (read from the code, not run). This is the m1n1 patch that fills them in, from OrchardLinux#101, unchanged. Assisted-by: LLM

From OrchardLinux#101, the part that main did not take with the sensors: the retained AOP reload across kexec (checksummed shutdown log, power-state observer, pr_flush export, warm_state attribute). The sensor parts that main carries are left out of these patches; what remains applies on top of main's AOP driver and its fresh-start order. Assisted-by: LLM

From OrchardLinux#110, the part main did not take with the stable USB serial: the platform layer embeds each supported board's pristine DTB and a compatible-to-file map in the initrd. Before switching root, the initrd copies the running board's tree to /run/orchard-kexec/pristine.dtb, or, after a warm restart, the retained tree the previous kernel named with orchard_warm.dtb=, and records the boot ID next to it. Warm restart boots the next kernel with that tree, never with one rebuilt from an updated /boot. No feature table is added: whether warm restart is offered is decided from the mechanisms it uses, not from a feature state. Assisted-by: LLM

Bring this branch onto current main. Main has since merged OrchardLinux#124, OrchardLinux#122, OrchardLinux#118, OrchardLinux#121, OrchardLinux#110, OrchardLinux#127, OrchardLinux#101, OrchardLinux#126, OrchardLinux#108 and OrchardLinux#111, several in part or reworked, and OrchardLinux#123 and OrchardLinux#128 were closed. Where main carries a topic, main's version is kept; the left-out parts that warm restart and suspend need come back in OrchardLinux#107 and OrchardLinux#117. This branch's own patches in devices/s8000, drivers/base, drivers/pmdomain and drivers/usb are renumbered to follow main's.

Bring this branch onto current main. Main has since merged OrchardLinux#124, OrchardLinux#122, OrchardLinux#118, OrchardLinux#121, OrchardLinux#110, OrchardLinux#127, OrchardLinux#101, OrchardLinux#126, OrchardLinux#108 and OrchardLinux#111, several in part or reworked, and OrchardLinux#123 and OrchardLinux#128 were closed. Where main carries a topic, main's version is kept; the left-out parts that warm restart and suspend need come back in OrchardLinux#107 and OrchardLinux#117. This branch's own patches in devices/s8000, devices/t8010, drivers/base, drivers/media, drivers/pmdomain, drivers/power and drivers/usb are renumbered to follow main's. Main's versions apply first now, so 0011 in drivers/media keep their change with a diff regenerated for the new context.

From OrchardLinux#101, the suspend precheck that main did not take with the sensors. It belongs with suspend: it makes the AOP refuse a pre-sync system sleep while its firmware is starting or stopping. Assisted-by: LLM

From OrchardLinux#121, the patch main did not take with HEVC. It uses the VXD sleep stop and sleep gate this branch adds, so it sits here rather than with the earlier warm restart branch: a failed firmware stop keeps the firmware, command buffers and power, and kexec takes the cold-reset path before those pages become the next kernel's RAM. Assisted-by: LLM

Bring this branch onto current main. Main has since merged OrchardLinux#124, OrchardLinux#122, OrchardLinux#118, OrchardLinux#121, OrchardLinux#110, OrchardLinux#127, OrchardLinux#101, OrchardLinux#126, OrchardLinux#108 and OrchardLinux#111, several in part or reworked, and OrchardLinux#123 and OrchardLinux#128 were closed. Where main carries a topic, main's version is kept; the left-out parts that warm restart and suspend need come back in OrchardLinux#107 and OrchardLinux#117. This branch's own patches in devices/s8000, devices/s8003, devices/t8010, drivers/base, drivers/media, drivers/pmdomain, drivers/power and drivers/usb are renumbered to follow main's. Main's versions apply first now, so 0011 in drivers/media keep their change with a diff regenerated for the new context.

Bring this branch onto current main. Main has since merged OrchardLinux#124, OrchardLinux#122, OrchardLinux#118, OrchardLinux#121, OrchardLinux#110, OrchardLinux#127, OrchardLinux#101, OrchardLinux#126, OrchardLinux#108 and OrchardLinux#111, several in part or reworked, and OrchardLinux#123 and OrchardLinux#128 were closed. Where main carries a topic, main's version is kept; the left-out parts that warm restart and suspend need come back in OrchardLinux#107 and OrchardLinux#117. This branch's own patches in devices/s8000, devices/s8003, devices/t8010, drivers/base, drivers/media, drivers/pmdomain, drivers/power, drivers/soc-apple and drivers/usb are renumbered to follow main's. Main's versions apply first now, so 0011 in drivers/media keep their change with a diff regenerated for the new context.

On J71t a completed suspend/resume cycle does not bring the GPU back, yet the G5P sleep admission admits an idle GPU, so /sys/power/suspend_ready reads "ready" and userspace that follows it suspends into a GPU that does not come back. drivers/drm 0045 adds a read-only "suspend" parameter to the G5P driver, off by default. While it is off, the GPU's sleep admission refuses before anything is shut, and suspend_ready reads "not-ready <gpu>:resume-unsupported -95". Booting with a9_gpu_drm.suspend=1 admits sleep exactly as before. Boards without the G5P driver are not affected. The patch takes the next free number in drivers/drm, after the gpu-handover branch's 0042-0044, so every existing patch keeps its number. Assisted-by: LLM

Carry the G5P sleep refusal (drivers/drm 0045) from j71t/g5p-suspend. Assisted-by: LLM

Bringing this branch onto main took main's devices/j71t/firmware.json, which main rewrote with its AOP firmware, and lost the output the GPU needs: lib/firmware/apple/agx-g5p-c0.ldr, extracted from the kernelcache by agx-g5p-ldr-from-kernelcache@1 (src/orchard/recipe.py still has the step). Without it the G5P driver finds no loader and the GPU never starts ("autoboot stopped at validate (-2)"). Put the entry back after main's outputs, unchanged; the file is again main's plus this one output. Assisted-by: LLM

Carry the restored G5P firmware loader entry in devices/j71t/firmware.json.

J71t's agx-g5p-c0.ldr output was lost when main's firmware.json was merged in, while its recipe step stayed, and no test noticed: the GPU simply never started. Assert that J71t ships the loader, and that every step in orchard.recipe.STEPS is used by some board's firmware output, which holds on main as well. Assisted-by: LLM

Carry the firmware recipe test for the G5P loader.

brkosebay added 30 commits November 5, 2026 10:15

Quiet polled cover inputs and exclusive RTC event

c4ef657

Assisted-by: LLM

J71t guarded Mac current profile

b37b4ff

Assisted-by: LLM

NCM retry after busy aggregation flush

16de1ff

Assisted-by: LLM

usb: keep per-device system and RAM serials stable across boots

48be1c6

Assisted-by: LLM

G5P render, autoboot, META71C and boot-origin policy

cfa4c8d

kernel/configs/apple-16k/config: CONFIG_DRM_APPLE_G5P=y These options provide g5p and its Kconfig dependencies on both page sizes. Assisted-by: LLM

AOP sensors and certified warm reload

fd6e7fe

kernel/configs/apple-16k/config: CONFIG_APPLE_AOP_GYRO=m CONFIG_APPLE_AOP_PRESSURE=m These options provide aop and its Kconfig dependencies on both page sizes. Assisted-by: LLM

Optional resident m1n1 EL3 PSCI

06e94b2

Assisted-by: LLM

pinctrl/spi/input: quiesce S8000 peripherals for kexec

8e61475

Assisted-by: LLM

Optional resident m1n1 EL3 PSCI

419fe19

Assisted-by: LLM

Warm Restart evidence, reset guard and pristine DT

df8ccca

Assisted-by: LLM

pinctrl/spi/input: quiesce S8000 peripherals for kexec

a76d376

Assisted-by: LLM

J71t guarded Mac current profile

1ca51ef

Assisted-by: LLM

AOP sensors and certified warm reload

ca3266c

kernel/configs/apple-16k/config: CONFIG_APPLE_AOP_GYRO=m CONFIG_APPLE_AOP_PRESSURE=m These options provide aop and its Kconfig dependencies on both page sizes. Assisted-by: LLM

usb: keep per-device system and RAM serials stable across boots

6fd2737

Assisted-by: LLM

J71t evidence-backed feature registry

c8a723f

Assisted-by: LLM

Optional resident m1n1 EL3 PSCI

6c1601e

Assisted-by: LLM

Quiet polled cover inputs and exclusive RTC event

eb6a0b1

Assisted-by: LLM

pinctrl/spi/input: quiesce S8000 peripherals for kexec

33047d2

Assisted-by: LLM

Warm Restart evidence, reset guard and pristine DT

0886d8c

Assisted-by: LLM

Merge the j71t/warm-peripherals branch into j71t/mac-current

fa1cb2e

Bring this branch onto current main and the series it depends on. The merge keeps the other side's tree; the next commit carries the reworked version of this branch's change.

Merge the j71t/psci branch into j71t/aop

82666fd

Bring this branch onto current main and the series it depends on. The merge keeps the other side's tree; the next commit carries the reworked version of this branch's change.

Merge the j71t/mac-current branch into j71t/usb-stable-serial

502c5ac

Bring this branch onto current main and the series it depends on. The merge keeps the other side's tree; the next commit carries the reworked version of this branch's change.

brkosebay added 3 commits November 12, 2026 09:00

brkosebay mentioned this pull request Nov 12, 2026

kernel: USB host through the Lightning camera adapter on the iPad 5, 6 and 7 #126

Merged

brkosebay added 26 commits November 13, 2026 05:31

kernel: Refuse a system sleep while the AOP firmware is not settled

7e6d459

From OrchardLinux#101, the suspend precheck that main did not take with the sensors. It belongs with suspend: it makes the AOP refuse a pre-sync system sleep while its firmware is starting or stopping. Assisted-by: LLM

Merge the j71t/g5p-suspend branch into j71t/gpu-handover

21b0202

Carry the G5P sleep refusal (drivers/drm 0045) from j71t/g5p-suspend. Assisted-by: LLM

Merge the apple-g5p-render branch into j71t/g5p-pipeline

7f58ae5

Carry the restored G5P firmware loader entry in devices/j71t/firmware.json.

Merge the j71t/g5p-pipeline branch into j71t/g5p-suspend

9769265

Carry the restored G5P firmware loader entry in devices/j71t/firmware.json.

Merge the j71t/g5p-suspend branch into j71t/gpu-handover

daa4bec

Carry the restored G5P firmware loader entry in devices/j71t/firmware.json.

Merge the apple-g5p-render branch into j71t/g5p-pipeline

b32ac15

Carry the firmware recipe test for the G5P loader.

Merge the j71t/g5p-pipeline branch into j71t/g5p-suspend

a21768d

Carry the firmware recipe test for the G5P loader.

Merge the j71t/g5p-suspend branch into j71t/gpu-handover

7f43bad

Carry the firmware recipe test for the G5P loader.

Navigation Menu

Repository navigation

drm/apple-g5p: Bound the kexec handover and fix two lost-GPU races - #125

drm/apple-g5p: Bound the kexec handover and fix two lost-GPU races#125
brkosebay wants to merge 79 commits into
OrchardLinux:mainfrom
brkosebay:j71t/gpu-handover

brkosebay commented Nov 12, 2026 •

edited

Loading

Uh oh!

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

Conversation

brkosebay commented Nov 12, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Problem

Change

Testing

Dependencies

Uh oh!

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

brkosebay commented Nov 12, 2026 •

edited

Loading