Skip to content

Latest commit

 

History

History
532 lines (471 loc) · 33.9 KB

mcu.md

File metadata and controls

532 lines (471 loc) · 33.9 KB

The 3DS system controller (MCU) in iOS

Status: implemented and verified in QEMU (-M nintendo3ds), not yet run on hardware. Section 7 lists what a hardware run will show and which assumptions it checks.

Feature 3DS side iOS side QEMU
RTC → date and time MCU regs 0x30–0x36 (BCD) PEGetGMTTimeOfDay → shim; no IORTC wait see §6
Battery level, charging regs 0x0A–0x0F an IOPMPowerSource created by the shim see §6
HOME, POWER IRQ regs 0x10–0x13 input mailbox INBTN_HOME / INBTN_POWER → ios3ds-input → GSEvents see §6
Shell close / open IRQ bits 5/6, reg 0x0F bit 1 POWER tap (lock) / HOME tap (wake), from the ARM9 see §6
Volume slider reg 0x09 (0..63) ios3ds-mcud → AVSystemController setActiveCategoryVolumeTo: see §6
Brightness LCD backlight PWM 0x10202240 / 0x10202A40 (ARM11) SpringBoard's DisplayBrightness × DisplayBrightnessFactor → ios3ds-mcud → sysctl → shim see §6
Accelerometer reg 0x40 (on), 0x45–0x4A (x, y, z) Apple's AppleLIS302DL driver on an emulated LIS302DL (shim I2C register file) → SpringBoard orientation, UIAccelerometer see §6.2
Gyroscope ITG-3270 class chip on I2C bus 3 (not the MCU) Apple's AppleLIS331DLH driver on a new gyro DT node, two vtable words patched → IOHID gyro events → CoreMotion (CMMotionManager gyro / deviceMotion) see §11
iOS shut down / restart reg 0x20 bit 0 (power off) PEHaltRestart → shim → ARM9 see §6

1. Data path

 MCU (I2C bus 2, dev 0x4A)
   │  polled by the ARM9 monitor (boot/arm9/mcu.c, i2c.c): IRQ events 30 ms,
   │  sliders 100 ms, RTC 1 s, battery 2 s, accelerometer 50 ms while wanted
   ▼
 mcubox_t  shim region +0x6400 (boot/common/mcu_proto.h)       input mailbox +0x6200
   │                                                          (INBTN_HOME / INBTN_POWER)
   ├── kernel shim (tools/ios3ds-mcu/shim/mcu.c)                         │
   │     PEGetGMTTimeOfDay / PESetGMTTimeOfDay / IORTC wait / PEHaltRestart
   │     IOPMPowerSource ("battery"), refreshed on every sysctl read      │
   │     sysctl kern.ios3ds_mcu (RD), kern.ios3ds_mcuctl (RW)             │
   ▼                                                                      ▼
 /usr/libexec/ios3ds-mcud (tools/ios3ds-mcu/daemon)             /usr/libexec/ios3ds-input
   volume, brightness, accelerometer                              Home / lock GSEvents

The ARM9 owns the MCU, as it owns SPI (touch, CODEC) and the SD card. The ARM11 only touches the LCD backlight registers, which are ARM11-only IO.

2. Shared memory

mcubox_t (256 bytes) at shim + 0x6400 (PA 0x20846400), between the input mailbox (+0x6200..+0x63FF) and the free space below the GPU mirror mailbox (+0x7F00). Single writer per 32-byte line, as for the other mailboxes:

Line Writer Contents
0 ARM9 magic MCU0, version, state (OFF/OK/FAILED/DISABLED), MCU version, seq, polls, I2C errors, valid bits
1 ARM9 rtc_utc (Unix time = RTC + rtc_offset), rtc_offset, raw BCD, read count
2 ARM9 battery %, temperature, voltage, flags (reg 0x0F), volume and 3D sliders, event counters
3 ARM9 accelerometer x/y/z (raw s16), sample count; gyro x/y/z (deg/s, 16.16, iOS axes), gyro sample count, gyro_state/gyro_dev/gyro_id (§11)
4 ARM11 req_seq, req_power (power off), req_acc (accelerometer wanted), backlight (1 + level last set; 1 = screen off), req_gyro (gyro wanted)
5 ARM9 ack_seq

3. ARM9 driver (boot/arm9/i2c.c, mcu.c)

  • I2C: GodMode9's ARM9 transfer sequence (also libn3ds'): start + device, register, repeated start + device|1 for reads; ACK all bytes but the last, last with STOP; NACK → EN|IRQ|ERROR|STOP and up to 8 retries. Bus setup as libn3ds I2C_init (CNTEX = clock stretching, SCL = delays 5/0). Every busy wait is bounded (~6 ms): a dead bus never hangs the monitor.
  • Boot (mcu_init, before XNU starts): read the version (regs 0/1); no answer → state FAILED, one log line, nothing else is touched. Else write the IRQ mask (regs 0x18–0x1B, 1 = masked) to unmask POWER press/held, HOME press/release and shell close/open, read the IRQ register once to drop stale events, then read RTC, battery and sliders, so the kernel finds the time when it initialises its calendar.
  • HOME: press/release IRQs → a level; press and release in the same 30 ms poll → one input sample down, then up. A lost release ends after 10 s.
  • POWER: the MCU has no release event. A press → a 250 ms tap (iOS: lock or wake). The MCU's "held" IRQ → POWER down for 6 s, long enough for the input daemon's 0.8 s hold detection plus SpringBoard's 2.5 s timer with a lagging run loop → the "slide to power off" sheet (a late release is ignored once the sheet is up).
  • Shell: close → POWER tap unless iOS' screen is already off; open → HOME tap if it is off. "Off" = the shim's last backlight level is 0 (backlight == 1), so it follows iOS' own screen state. In both cases the two LCD backlights are also switched through MCU reg 0x22 (0x14 off, 0x28 on), so the screens follow the shell even if iOS ignored the press.
  • Power off: on req_power, wait 300 ms (the monitor loop drains the kernel's last log lines), sync the log, write reg 0x20 = 1.
  • boot.cfg: mcu=0 disables all of it; rtc_offset=N is added to the RTC to get UTC (the 3DS RTC usually runs on local time: e.g. -3600 for CET). iOS' own time zone (Settings → General → Date & Time) is separate.

4. Kernel side (tools/ios3ds-mcu/shim/mcu.c, kernel/patches-mcu.json)

RE details: re-log.md §33–34.

  • RTC. PEGetGMTTimeOfDay (0x801c84d8) jumps to shim_mcu_get_gmt, which returns rtc_utc (+ a settimeofday delta, below). Without a valid RTC it runs the original body (gIOPlatform->getGMTTimeOfDay()), so the fallback is today's behaviour.
  • No IORTC wait. IOKitInitializeTime (0x801a8004) waits for resourceMatching("IORTC") before clock_initialize_calendar(); the literal of its waitForService call (0x801a8030) now points at shim_mcu_wait_rtc. With an RTC it returns at once (the matching dictionary is released as waitForService would): the calendar no longer depends on AppleARMPE being registered, which was the reason for the 2 s wait (re-log §26). Without an RTC it calls the original waitForService with iortc-wait-2s' 2 s.
  • Setting the time. PESetGMTTimeOfDay (0x801c84f4) → shim_mcu_set_gmt: AppleARMPE's version waits 30 s for an IORTC. The 3DS RTC is the console's clock and stays untouched; the difference is kept until reboot.
  • Halt. In PEHaltRestart (0x801c98b0) the call gIOPlatform->haltRestart(type) (0x801c98f6) goes to shim_mcu_halt_restart instead. It runs after IOKit's shutdown handlers and after the file systems were unmounted. Halt (0, 3) and restart (1): request MCU_PWR_OFF/MCU_PWR_REBOOT from the ARM9 and wait; the ARM9 powers the console off in both cases (a reboot would end in Luma / the HOME Menu, not in iOS). Panic restart, hang and the rest go to the platform as before. pe-halt-skip-pm skips IOPMrootDomain's halt broadcast just before that call: it hung every shutdown in AppleM2CLCD (§6.1).
  • gettimeofday. commpage-gtod-off stops the kernel from publishing the commpage time stamp, whose fast path never advanced with our timebase (§8).
  • Battery. On the first kern.ios3ds_mcu read with valid battery data the shim creates an IOPMPowerSource (IOPMPowerSource::powerSource()), attaches it to the platform expert and registers it. Every later read updates CurrentCapacity, MaxCapacity (100), IsCharging, ExternalConnected, ExternalChargeCapable, FullyCharged, Voltage and AppleRawBatteryVoltage and calls updateStatus(), which copies them into the registry and messages interested clients (kIOPMMessageBatteryStatusHasChanged). These are the keys SpringBoard reads (IOServiceMatching("IOPMPowerSource")). AtCriticalLevel is never set: SpringBoard would shut down immediately; the MCU itself switches the console off when the battery is empty.
  • Backlight. kern.ios3ds_mcuctl carries {acc, backlight} from the daemon. The shim maps the LCD registers with ml_io_map(0x10202000), logs the values Luma left, and writes 0x10 + level × (0x8F − 0x10) to both screens' backlight PWM (0 when iOS' screen is off).

5. Userland (tools/ios3ds-mcu/daemon/mcud.c)

/usr/libexec/ios3ds-mcud (launchd job com.ios3ds.mcud, argument debug) polls kern.ios3ds_mcu every 100 ms:

  • Volume: when the slider moved (2 of 63 steps, or to an end): [[AVSystemController sharedAVSystemController] setActiveCategoryVolumeTo:v/63] (Celestial.framework). In between, iOS' own volume buttons (L/R) still work.
  • Brightness: twice a second it reads DisplayBrightness and DisplayBrightnessFactor from SpringBoard's HID event system (IOHIDEventSystemClientCopyProperty; SpringBoard keeps its backlight state there) and sends brightness × factor to the shim.
  • Accelerometer, old route (only with the acc argument; superseded by the kernel driver, §6.2): while the screen is on it asks the ARM9 for samples and sends IOHIDEventCreateAccelerometerEvent(x, y, z) (type 13) into SpringBoard's HID event system, normalised so the running mean of |a| is 1 g (the raw unit does not matter). Axis map acc=-y,-x,-z (default, an argument): iOS x, y, z from the 3DS axes, see §7.
  • Clock check: logs the same interval measured by the RTC, gettimeofday and mach_absolute_time at ~60 s and ~300 s (§8).

6. QEMU model and verification

emu/qemu/hw/arm/nintendo3ds.c already had the I2C controllers and an MCU (IRQs, RTC from the host clock, battery, HOME = h, POWER = p). Added:

  • accelerometer: reg 0x40 bit 0 on, 0x45–0x4A data (1 g = 0x4000), four poses;
  • keys: - / = volume slider down/up by 8, c shell close/open, [ / ] turn the console by 90°;
  • runtime state through QOM properties on /machine (HMP qom-set / qom-get, emu/run.py --hmp "t=command;…"): mcu-battery, mcu-charging, mcu-volume, mcu-shell-open, mcu-orient, and the read-only lcd-backlight;
  • LCD backlight registers 0x10202240 / 0x10202A40 (reset 0x5F); once the guest writes them, the emulated screens are dimmed accordingly (0x8F = full).

QEMU results (2026-11-18)

Runs with emu/run.py / emu/tap.py; artifacts in build/emu-r* (not committed: the screenshots show Apple's UI).

Driver Evidence
RTC [mcu] version 13/38: RTC 2026-11-17 23:21:31 Tue (offset 0 s) = Unix 1794957691, ios3ds-mcu: 3DS RTC present: not waiting for IORTC, ios3ds-mcu: calendar from the 3DS RTC: 1794957691; the "RTC did not show up" line is gone. Status bar 3:26 PM, calendar icon "Tuesday 17" (host 23:26 UTC; iOS zone US/Pacific). Same in --icount 2
gettimeofday before: tv 1794957716.452931 -> 1794957716.452931; mach_absolute_time … = 212 ms; after commpage-gtod-off: 181 ms vs 182 ms (plain), 2254 vs 2254 ms (icount); [mcud] clocks after ~60 s: RTC +153 s, gettimeofday +154086 ms, mach_absolute_time +154084 ms
Battery ios3ds-mcu: battery 100%, charger, 4200 mV: IOPMPowerSource registered; full battery with bolt → qom-set mcu-charging off, mcu-battery 50 → half battery without bolt; mcu-battery 15 → SpringBoard's "Low Battery, 20% of battery remaining" alert
HOME key h → [input] key 40 down/up → Spotlight (Home on page 1)
POWER p tap → key 30 (lock); p held 5 s → [mcu] POWER held: iOS power-off slider → "slide to power off" sheet
Shutdown sliding the sheet → launchd: System: About to call: reboot(RB_HALT). → [mcu] iOS shut down: log synced, powering off through the MCU → the QEMU model's reg 0x20 write ends QEMU (§6.1)
Shell c → [mcu] shell closed: locking iOS (POWER tap); LCD backlights off (both screens black); c → [mcu] shell opened; LCD backlights on
Volume = twice → [mcud] volume slider 40/63 -> iOS volume 0.63, 48/63 -> 0.76
Brightness [mcud] iOS brightness 0.50 x factor 1.00 -> LCD backlight 500/1000, ios3ds-mcu: LCD backlight as Luma left it: top 0x5f, bottom 0x5f, qom-get lcd-backlight = 79 (0x4F); the screenshots are dimmed accordingly
Accelerometer [mcud] accelerometer raw -16384 0 0 -> iOS -0.00 -1.00 -0.00 g, after ] ]: 16384 0 0 -> iOS 0 1.00 0; SpringBoard's state (§6.2)

6.1 Shutdown hung in AppleM2CLCD before reaching the platform

The first shutdown in QEMU never got to PEHaltRestart's platform call: the CPU idled after reboot(RB_HALT). A SELECT+START thread dump showed launchd's thread in PEHaltRestart+0x3e → IOPMrootDomain halt/restart broadcast → IOMobileGraphicsFamily → AppleM2CLCD (disable_backlight, "Where is my backlight?") → AppleARMPlatform function → AppleS5L8720X+0x37cc IOCommandGate::commandSleep, waiting for an interrupt from S5L backlight hardware the 3DS does not have. So every iOS shutdown used to hang forever. pe-halt-skip-pm skips that broadcast (file systems are unmounted by then and md0 writes are synchronous); the platform call then reaches the shim. The ARM9 waits 300 ms after the request so the kernel's last lines reach the log, then writes MCU reg 0x20.

6.2 Accelerometer: Apple's driver on an emulated LIS302DL

First attempt (still in ios3ds-mcud as acc): IOHID accelerometer events sent by a client. SpringBoard ignores them, as it ignores client keyboard and touch events (re-log §37: hardware breakpoints on its type-13 handler never hit). SpringBoard only takes accelerometer data from a kernel HID service.

Now the n72's own accelerometer node /arm-io/i2c0/accelerometer stays in the device tree (tools/ios3ds-mcu/dt_mcu.py). Apple's AppleEmbeddedAccelerometer (class AppleLIS302DL) starts on it, and the shim answers its I2C transfers (shim_mcu_iic, called first by kernel/shim/audio.c's shim_iic_transfer). Details in re-log §40:

  • WHO_AM_I (0x0F) reads 0x3B.
  • CTRL_REG1 power-up bit → the shim asks the ARM9 for samples (req_acc).
  • The CTRL_REG2 BOOT bit reads back cleared (the driver panics after 500 tries otherwise).
  • OUT_X/Y/Z (0x29/0x2B/0x2D) are the newest MCU sample in iOS axes. They are scaled so the running mean of |a| is 55 counts (1 g at the driver's 18 mg/count), with z negated for the DT's orientation 4.
  • The data-ready interrupt (GPIO 0x63) never fires. The driver polls with its own timer at the ReportInterval that SpringBoard sets while an app that rotates is in front.

QEMU (build/emu-a2, Safari opened by a tap, then qom-set /machine mcu-orient 1 / 3 / 0):

ios3ds-mcu: accelerometer read reg 0x0f len 1                       (probe at boot)
ios3ds-mcu: accelerometer powered on by AppleLIS302DL (CTRL_REG1 0x47)
launchd[31] Builtin profile: MobileSafari (sandbox)
ios3ds-mcu: accelerometer sample #1: 0 -55 0 counts (|a| = 16384 MCU units)
[mcud] SpringBoard device orientation 0 -> 1
[mcud] SpringBoard device orientation 1 -> 4      (pose 1: landscape, Home button left)
[mcud] SpringBoard device orientation 4 -> 3      (pose 3: landscape, Home button right)
[mcud] SpringBoard device orientation 3 -> 1      (pose 0: portrait)

Safari's bookmarks view turns to landscape, to the other landscape, and back to portrait in the screenshots. SpringBoard and the home screen behave as before (the home screen does not rotate on an iPod touch).

Kill switch: mcu_acc=0 in boot.cfg. The shim then answers the driver's probe with kIOReturnNoDevice, so no accelerometer service exists. Without a working MCU it does the same.

6.3 Lock = screen off (SBDontLockEver removed)

The rootfs used to set SBDontLockEver in com.apple.springboard.plist (an early guard against the keybag/lock-state problem). With it set, SpringBoard ignored the lock button and the screen stayed on (re-log §21).

Removing it alone was not enough. The first lock blanked the screen, but after waking the picture stayed frozen on the old frame (status-bar clock stuck). The fake CLCD page read +0x00 = 2 (controller off) after the wake. Cause: on display sleep the system lowers the CPU performance state, and AppleS5L8720XClockController then sleeps on its command gate waiting for a "change done" interrupt that the fake clock page never raises. It is the same wait that hung every shutdown (§6.1). Patch s5l-perf-nowait marks the change as done at once (re-log §41).

With both changes, in QEMU (build/emu-l6):

[input] key 30 down/up                                   (SELECT = lock)
[mcud] iOS brightness 0.50 x factor 0.00 -> LCD backlight 0/1000      (both screens black)
[input] key 40 down/up                                   (3DS HOME)
[mcud] iOS brightness 0.50 x factor 1.00 -> LCD backlight 500/1000
  • After the wake the lock screen appears: "1:32, Wednesday, November 18", the charging battery, "slide to unlock".
  • Sliding it unlocks back to the home screen, whose clock keeps running (1:33, 1:34). The CLCD reads +0x00 = 3 again.
  • Left alone on the lock screen, iOS turns the screen off again after about 10 s, as on an iPod.
  • Passcode-less unlock works; the stubbed keybag is not in the way.

Trade-offs:

  • After a lock the user has to slide to unlock (touch works on the lock screen).
  • With the screen off, iOS ignores touches; HOME, POWER, START or opening the shell wake it.
  • If a hardware run shows trouble, put SBDontLockEver back (rootfs/recipe.json). The lock press is then ignored, and closing the shell still darkens the screens through MCU reg 0x22.

7. Hardware assumptions (unverified) and the log lines that check them

Nothing here has run on a 3DS. Each assumption, and the log line that settles it on the first hardware run:

Assumption Source Log line to read
The ARM9 can drive I2C bus 2 (0x10144000) and the MCU answers at 0x4A Luma/GodMode9 talk to the MCU from the ARM9 [mcu] version xx/yy: RTC … (good) or [mcu] no answer on I2C bus 2 … (then everything behaves as before)
RTC regs 0x30-0x36 are BCD sec/min/hour/weekday/day/month/year-2000 libn3ds, GodMode9 the RTC date in that line; the iOS clock. rtc_offset in boot.cfg if the RTC runs on local time
Battery: 0x0B percent, 0x0D voltage in 20 mV, 0x0A °C; 0x0F bit 3 charger, bit 4 charging, bit 1 shell open 3dbrew, GodMode9 (charging bit) [mcu] battery 85% 3950 mV 27 C, charger, charging, …; the battery icon
Writing the IRQ mask (0x18) unmasks HOME/POWER/shell events; reading 0x10 clears them libn3ds MCU_init IRQ mask set; [mcud] … irq 0000000c home 1 … after pressing HOME
POWER "held" (bit 1) comes while still pressed, early enough before the MCU's own forced power-off 3dbrew [mcu] POWER held: iOS power-off slider and the slider sheet
Volume slider 0x09 is 0..63 libn3ds [mcud] volume slider N/63 -> iOS volume … while moving the slider
3DS LCD backlight PWM at 0x10202240 / 0x10202A40, 0x10..0x8F a safe range libn3ds lcd.h, Luma's presets (0x26..0x5F) ios3ds-mcu: LCD backlight as Luma left it: top 0x.., bottom 0x..; does the iOS brightness slider change the screens?
MCU reg 0x22 bits 2-5 switch the backlights libn3ds MCU_LCD_PWR_* closing the shell darkens both screens: [mcu] shell closed: …; LCD backlights off
Accelerometer: reg 0x40 bit 0 enables it, 0x45-0x4A are x/y/z s16; axes as in the QEMU model (x right, y hinge, z out of the lower screen) 3dbrew, libn3ds [mcu] accelerometer on; `ios3ds-mcu: accelerometer sample #1: x y z counts (
Gyro: an ITG-3200-family chip at bus 3 / 0xD0 (Old 3DS, libn3ds "ITG-3270"), or an ST L3G at 0xD6, or something at 0xD2; the ARM9 can drive bus 3 (0x10148000) libn3ds device table, Horizon's 14.375 LSB/(deg/s) [gyro] ITG-3200 family at I2C 3:d0 (WHO_AM_I 68), asleep, … (good), [gyro] no gyroscope on I2C bus 3 (then CoreMotion gets no rates); gyro=0 in boot.cfg if bus 3 misbehaves
Gyro axes = console axes (x right, y to the hinge, z out of the lower screen) guess turn the console in a gyro app (or diag=1: [gyrotest] lines): a turn about the screen's vertical axis should show in iOS y. Fix with gyro_axes= (e.g. x,-y,z)
Reg 0x20 bit 0 powers the console off Luma mcuPowerOff [mcu] iOS shut down: log synced, powering off through the MCU as the last log line, then the console is off

Kill switch: mcu=0 in sd:/ios3ds/boot.cfg (the MCU is never touched; clock at 1970, no battery, no HOME/POWER).

8. gettimeofday vs mach_absolute_time

The open question from hardware runs 10-13 (gltest: gettimeofday "0 ms", then "1063 ms" while mach_absolute_time said 412 ms) is answered (re-log §34):

  • mach_absolute_time is right; the timebase is 10^9 / 1047311.
  • Userland gettimeofday() first tries the commpage. It needs ticks per microsecond, which is 1047311 / 10^6 = 1 here, and the kernel's reciprocal for a divisor of 1 overflows (M 0, add 1, shift 0, read back from the commpage in QEMU). Elapsed time then always reads 0: the result stays at the last syscall refresh for up to a second. Short intervals read 0, long ones are off by up to a second, and whether the 2 s IORTC wait "helped" was just where a measurement fell.
  • commpage-gtod-off stops publishing the timestamp, so userland always makes the exact syscall. Verified in plain QEMU (181 vs 182 ms) and with --icount 2 (2254 vs 2254 ms, and over 154 s against the RTC).
  • Still approximate: inside one second the kernel's own microtime path adds raw ticks as microseconds (up to 4.7 % fast until the next full conversion). A timebase that is a multiple of 1 MHz (e.g. PERIPHCLK/67 = 2.0008 MHz) would remove both effects, but would change the clock code the whole system runs on; not done.
  • On hardware, ios3ds-mcud logs [mcud] clocks after ~60 s: RTC +N s, gettimeofday +… ms, mach_absolute_time +… ms, and gltest's timing line should show equal values.

9. Power saving while iOS' screen is off

"Off" = iOS has shown its display and AppleM2CLCD has switched the (fake) CLCD controller off again: lock, auto-lock, or the shell closed (which locks). boot/arm9/powersave.c and shim_mcu_irq_hook (tools/ios3ds-mcu/shim/mcu.c) then:

  • pause the screen copy on both paths. The ARM9 CPU copy already follows the CLCD enable bit (display_active()). The GPU mirror gets want = 0 in its mailbox (gm_want(), gpu_mirror.h); the shim checks it before every frame. No GPU code changed.
  • stop reading the touch ADC. Buttons, HOME/POWER and the shell are still read every loop.
  • let the ARM9 sleep. Once md0 has been quiet for 200 ms, the monitor loop waits for an interrupt between iterations. ARM9 timer 1 wakes it every 2 ms; IRQs stay masked in CPSR, so the pending IRQ only ends the wait. A block request is served within 2 ms, then the loop stays busy for the next 200 ms.
  • mask the VBlank interrupt on the ARM11. GIC 42 is masked at the distributor on the next VBlank. It is unmasked from the timer interrupt or a kern.ios3ds_mcu read as soon as the CLCD is on again. The CLCD frame IRQ is not injected while the controller is off anyway.
  • wake: everything is restored, and the next loop copies a fresh frame (last_blit is reset).

The ARM9 logs both transitions with the rates. powersave=0 in boot.cfg turns all of it off; the rates are then still logged, for comparison.

No system sleep. Without SBDontLockEver, iOS put the whole system to sleep about a minute after a lock ("System Sleep", "vic[0] ena 00010000" in the log). Nothing on the 3DS can wake it: on an iPod the PMU interrupt does. ios3ds-mcud therefore holds a NoIdleSleepAssertion (IOPMAssertionCreateWithName; argument sleepok disables it). The screen still goes off and the CPUs idle.

Fewer wakeups while off.

  • Input daemon: ios3ds-input polls its mailbox at 10 Hz instead of 60 Hz while the screen is off. It reads screen_off from kern.ios3ds_mcu; button samples queue in the 30-deep ring, so HOME and POWER still arrive within 100 ms. Its hold timing now counts real sleep time.
  • ios3ds-mcud: runs at 4 Hz instead of 10 Hz while the screen is off, reading the brightness every loop so a wake turns the backlight on at once.
  • Kernel: the boot-args now say serial=1 instead of serial=3. Bit 1 started XNU's serial keyboard thread, which polled the (never-ready) UART every 16 ms, at 62.5 Hz with the screen on or off (re-log 48). Console output is unchanged.

QEMU (plain TCG; locked about 80 s; shim counters read with xp /24wx 0x20840040; host CPU from the vCPU threads):

screen on off, before (powersave only) off, now
ARM9 monitor loops/s 1837 520 (493 in WFI) 530 (495 in WFI)
ARM9 vCPU thread, host CPU 42 % 2 % 2 %
ARM11 IRQs/s 127 (31 VBlank) 245 (0 VBlank) 146 (0 VBlank)
ARM11 timer IRQs/s ~96 243 143
ARM11 set_decrementer calls/s 296 — 416
ARM11 vCPU thread, host CPU 7 % 14 % 10 %

For comparison, powersave=0 while off: 4371 ARM9 loops/s, 304 ARM11 IRQs/s (59 of them VBlank).

Where the remaining ARM11 timer interrupts come from (re-log 48):

  • Method: breakpoints in QEMU's gdbstub on the shim's set_decrementer (backtraces including user frames) and on the thread_call worker's function call.
  • What arms timers while the screen is off: usleep in ios3ds-mcud, in ios3ds-input and (1 Hz) in ios3ds-wifid; kernel thread_call timers, which are BSD networking callouts (resolved callees: ip_output+0x86f4, net_del_domain+0x5c/+0xb8, and others); and the serial keyboard poll, now gone.
  • The CoreAnimation guess was wrong: no SpringBoard or QuartzCore frames show up among the timer arms.
  • The networking callouts' share of the timer arms doubles when the screen goes off, with the same mix of functions. My guess: with the screen on, the busy (emulated) CPU makes them fire less often. They are left alone (they are XNU's own protocol timers); about 140 timer interrupts per second on an idle ARM11 cost well under 1 % CPU.

"Where is my backlight?" ~3x/s after a wake: AppleM2CLCD's safe_enable_backlight (0x807a95d0) finds no backlight service, logs, and keeps its pending flag (+0xea0). So it retries on every frame until the next sleep. Patch clcd-backlight-retry-off clears the flag instead: after a wake there are 0 such lines (build/emu-p3); one remains at the lock (disable_backlight).

Run 17 (button presses lost after many locks). The ARM9 loop samples the buttons and the MCU, so anything that slows the loop delays or loses them. With the log console visible (screen mode "bottom" or "top") the console painting did that: 1-15 passes/s. Consoles are now painted a few rows per pass, and CPU screen copies poll the buttons as well (re-log 57).

10. Notification LED (opt-in)

notify_led=1 in boot.cfg: at ≤ 10 % battery, not charging, the ARM9 writes a red blink pattern to MCU register 0x2D. Otherwise it writes all zeros (off), and it switches the LED off before a power-off.

  • Pattern (3dbrew "MCU Services" / libn3ds McuInfoLedPattern): step time 0x2F, smoothing 0x2F, loop delay 0, then 32 steps each of red (0xA0 for 8 steps, 0 for 8), green (0) and blue (0).
  • Default is off: the timing values are not verified on hardware.
  • QEMU model: captures the 100-byte stream (qom-get /machine mcu-led-rgb0, mcu-led-steps). build/emu-p3: mcu-battery 8 + charger off → [mcu] notification LED blinking red: battery low, steps 16, step-0 colour 0xA00000; charger on → [mcu] notification LED off, steps 0.
  • Hardware check: whether the LED blinks red at ~1 Hz and goes dark again. If the timing is wrong, adjust p[0]/p[1] in led_update() (boot/arm9/mcu.c).

11. Gyroscope → CoreMotion

iPhone 4 apps read the gyro through CMMotionManager (gyroAvailable, startGyroUpdates, gyroData.rotationRate, deviceMotion). On 4.2.1 that works in the app's own process: CoreMotion finds an IOHID service with usage page 0xff00 / usage 9 in SpringBoard's HID event system, sets its ReportInterval, and takes IOHIDEvent type 20 from it (x/y/z in deg/s; CoreMotion converts to rad/s). RE details: re-log.md §64.

 gyro chip (I2C bus 3: ITG-3270 at 0xD0 on the Old 3DS, per libn3ds; 0xD6 / 0xD2 also probed)
   │  ARM9 (boot/arm9/gyro.c): asleep until iOS wants it, then 100 Hz;
   │  counts -> deg/s (16.16) -> iOS axes (boot.cfg gyro_axes)
   ▼
 mcubox line 3: gyro[3], gyro_samples, gyro_state        line 4: req_gyro (ARM11 -> ARM9)
   ▲                                                        ▲
   │ shim_gyro_dispatch (shim +0x2C)                        │ CTRL_REG1 power bits (shim_mcu_iic)
 Apple's AppleLIS331DLH on /arm-io/i2c0/gyro (AppleEmbeddedAccelerometer kext)
   │  vtable +0x3b4 getPrimaryUsage -> 9, +0x3c0 dispatch -> shim (kernel/patches-mcu.json)
   ▼
 IOHIDEventService (0xff00/9) -> SpringBoard's IOHIDEventSystem -> CoreMotion in the app
  • Kernel. The n72 kernelcache has no gyro driver, but the accelerometer kext also contains AppleLIS331DLH, a class nothing on the n72 uses. A second I2C node, /arm-io/i2c0/gyro (accelerometer,lis331dlh, address 0x6b, tools/ios3ds-mcu/dt_mcu.py), starts it. Two words of its vtable are patched: getPrimaryUsage returns 9 (gyro) instead of 3, and dispatchAccelerometerEvent goes to shim_gyro_dispatch, which calls IOHIDFamily's own dispatchGyroEvent (the next vtable slot) with the ARM9's newest sample. The driver itself is unchanged: it registers the HID service, runs its timer at the ReportInterval CoreMotion sets, and switches the "sensor" on and off through CTRL_REG1. The shim answers its I2C transfers (WHO_AM_I 0x32, BOOT cleared) and turns the power bits into req_gyro.
  • CoreMotion's gate. gyroAvailable comes from a per-model table in CoreMotion (CLPlatformInfo::getProviders), not from IOKit: N72 says "no gyro". The rootfs recipe sets the gyro bit in N72's word (0x2012 → 0x2052, N81's value), so gyroAvailable and deviceMotionAvailable are YES.
  • Power. The chip sleeps (ITG PWR_MGM SLEEP / ST power-down) from boot until a CoreMotion client starts updates, and again when the last one stops (the driver powers its "sensor" down when its ReportInterval is cancelled). While awake the ARM9 reads 7 bytes every 10 ms.
  • Units and axes. ITG-3200 family at FS_SEL 3: 14.375 LSB per deg/s (the coefficient Horizon's HID reports for the Old 3DS); ST L3G at 2000 deg/s: 70 mdeg/s per LSB. Converted to deg/s × 65536 (mcu_gyro_fixed, boot/common/mcu_proto.h). gyro_axes=-y,x,z (default) gives iOS x, y, z from the chip's x, y, z, assuming the chip's axes are the console's (x right, y towards the hinge, z out of the lower screen); with iOS shown in portrait on the lower screen, iOS y = console x and iOS x = -console y, as for the accelerometer (§5).
  • Kill switch: gyro=0 in boot.cfg. The ARM9 never touches bus 3 and the shim answers the driver's probe with kIOReturnNoDevice, so there is no gyro service. gyroAvailable still says YES (the CoreMotion word is in the rootfs), but no rotation rates arrive. mcu=0 (or a dead MCU) has the same effect, since the gyro shares the MCU mailbox.
  • Test tool: /usr/libexec/ios3ds-gyrotest (tools/ios3ds-mcu/gyrotest), a diag job (diag=1): uses CMMotionManager like an app for 150 s and logs [gyrotest] lines.
  • QEMU: an ITG-3200 at bus 3 / 0xD0. QOM properties gyro-x, gyro-y, gyro-z (millidegrees per second about the console's axes, read/write), gyro-awake and gyro-reads (read only).

11.1 QEMU results (2026-11-20, --icount 2, diag=1)

emu/run.py --icount 2 --time 560 --hmp "3=qom-set /machine gyro-x 10000;…" (build/emu-g2, not committed). Model rates in deg/s about the console's axes: (10, 30, −45), then y −90 at 280 s, all 0 at 330 s, z 180 at 380 s. Expected in iOS axes (default map −y,x,z): (−30, 10, −45) °/s = (−0.524, 0.175, −0.785) rad/s, then x +1.571, then 0, then z 3.142.

[gyro] ITG-3200 family at I2C 3:d0 (WHO_AM_I 68), asleep, axes -y,x,z
[gyrotest] gyroAvailable 1 deviceMotionAvailable 1 accelerometerAvailable 1
ios3ds-mcu: gyro on
[gyro] on by iOS (0 samples so far)
ios3ds-mcu: gyro event #1: -1964940 656499 -2949690 deg/s/65536 (ARM9 sample 13)
[gyrotest] started (interval 20 ms): gyroActive 1 deviceMotionActive 1
[gyrotest] t=1 gyro -0.523 0.175 -0.786 rad/s | deviceMotion rate -0.518 0.175 -0.778 | attitude roll 1.02 pitch 1.25 yaw -0.93
[gyrotest] t=90 gyro 1.571 0.175 -0.786 rad/s | deviceMotion rate 1.916 0.030 0.031 | attitude roll -2.55 pitch -0.66 yaw 2.15
[gyrotest] t=100 gyro 0.000 0.000 0.000 rad/s | deviceMotion rate 0.339 -0.165 0.769 | …
[gyrotest] t=130 gyro 0.000 0.000 3.142 rad/s | deviceMotion rate 0.299 -0.157 3.753 | attitude roll 2.28 pitch -0.20 yaw 1.82
ios3ds-mcu: gyro off
[gyrotest] done: gyro value changed 6 times, 1305 non-zero reads, last 0.000 0.000 3.142; updates stopped
[gyro] off by iOS (14782 samples so far)
  • gyroData.rotationRate equals the model's rates in iOS axes to 3 decimals. deviceMotion runs CoreMotion's sensor fusion on top: its rotation rate follows with its own bias estimation (it converges to the gyro while the accelerometer is still; steps show the filter catching up), and the attitude turns.
  • Power: the chip is asleep from boot until the first client and again after stopGyroUpdates (qom-get /machine gyro-awake: false at the end; the ARM9 sampled 14782 times, ~100 Hz while on).
  • The first run sent rad/s and CoreMotion reported 1/57.3 of the rates: its HID unit is deg/s (re-log §64). Without gyro-interrupt-calibration it also logged CLDeviceMotion::startSensorFusion … Could not get gyro dt; with the property that line is gone.