[hishamMounir]

These are three small changes that the T8142 (Apple M5, Mac17,2 / J704) needs to boot Linux with all CPUs.

kboot: match DT CPU nodes to running CPUs by MPIDR

kboot paired DT CPU nodes with live CPUs by position, and bailed on the first MPIDR mismatch. This change looks up the live CPU by the node's MPIDR instead.

• Trees that list CPUs in order behave exactly as before.
• A bring-up DT can now list the CPUs in a different order. We needed that on J704 to bring up one specific E-core first while chasing an SMP problem.

kboot: reserve the second CTRR range on t8142

T8142 has a second CTRR instance. It covers the DeviceTree/TrustCache area just below the bootloader, and iBoot leaves it enabled on the secondary cores. If Linux hands one of those pages to a secondary CPU and writes to it, the CPU faults, just as it does for the first instance.

• main already reserves the first instance on T8142, through the apple_sysregs_unlocked check.
• This adds the second instance: its two registers (3, 0, 11, 0, 2/3, next to the existing pair) and the reservation itself.
• It runs only on T8142 and is skipped when the instance reads back zero. Other SoCs are untouched.

How it was found: a native boot oopsed at exactly the first page of a protected range. A read-only sweep of the CTRR system registers on the boot core then gave the second instance's bounds, which cover the DeviceTree and TrustCache.

smp: log once when a call target hasn't taken a call after 1 s

A secondary that never takes a cross-CPU call leaves smp_call4() spinning silently, which looks exactly like a dead board. This change prints the CPU, the call and the target's state once after a second, flushes the console, and then keeps waiting as before. Apart from the message, behaviour is unchanged.

Testing

The test machine was a J704 with iBoot mBoot-20457.41.3.

• This tree, on current main, was chainloaded into RAM.
• It booted a Linux 7.1 Asahi kernel (16K pages) with root on the internal SSD.

On the M5, this series needs two other PRs:

• dapf: Find each DART's DAPF from the ADT and initialize them on T8132 #14 (DAPF from the ADT instance list). Without it, main takes an SError during DAPF setup on T8142, before Linux starts.
• kboot: disable WFI/WFIT where it loses state (from AsahiLinux/m1n1) #18 (Yureka's WFI bootargs, from AsahiLinux/m1n1). Without it, Linux freezes in idle within minutes. This happened twice.

Result with this series + #14 + #18:

• 10 CPUs online.
• No "has not taken call" message.
• No oops, SError or DART fault.
• Keyboard and trackpad work, and USB networking works.
• The machine stayed up for a 5-minute check.

Limits of this test:

• On this chainloaded path the second CTRR instance reads back zero, so the new reservation was skipped. It is exercised only on a native boot, where it was found and used with the AsahiLinux-based bootloader. That machine has booted Linux from its SSD this way since mid-October.
• The MPIDR change runs on every boot. All 9 secondaries are matched and started, and CPU 6 is the boot CPU here.
• The smp message never fired, as expected.
• Builds and links with clang and the Rust target, with no warnings in the touched files. clang-format was not run.

Provenance

These changes come from boot failures, m1n1's own error output and live system-register reads on the machine. No Apple binary was disassembled for them, and none of them touches the components listed in my disclosure in #tainted-re.

🤖 Generated with Claude Code