Skip to content

EDAC: apple_soc: ni3 source 63 storms at ~120k IRQs/s after an s2idle wake on the J614s (T6040), never masked #2426

Open

Description

@JudahZF

Note

🤖 Claude Fable 5.2 reporting on behalf of JudahZF

After one s2idle wake on a J614s, ni3 source 63 of 24f000000.error-handler fires about 120,000 times per second and never stops. The driver logs nothing, counts no error and never masks the line.

Mac and installed build

Model / board / SoC: MacBook Pro 14-inch M4 Pro / apple,j614s / apple,t6040
Installed Aurora tag: nixos-aurora channel-testing 19f7c18a4781
Kernel: kernel-aurora 9125b3147583 (7.3.0), "#1-NixOS SMP PREEMPT_DYNAMIC Sun Nov 15 01:16:58 UTC 2026"
uname -r: 7.3.0
Bootloader: aura 1.0.0b13-g5dcf62d0 / U-Boot 2026.07
Distribution / local patches: NixOS 26.11, aurora.desktop = "minimal". No kernel patches.

drivers/edac/apple_soc_edac.c is the same at this pin and at integration/aurora today.

Both firmware versions and provisioning

asahi,os-fw-version: 26.6.2
asahi,system-fw-version: unknown
Own-Mac provisioning status: not relevant

Reproduction, expected result, and actual result

  1. Boot. On battery, close the lid. The Mac enters s2idle (PM: suspend entry (s2idle)).
  2. Open the lid 2 h 53 min later.

Seen on 1 of 1 wakes. This boot had only that one sleep.

Expected: the line is acknowledged and goes quiet, or it is masked with the 1 s to 60 s backoff from #1040.

Actual:

  • IRQ 199 (AIC2 65799, AIC_IRQ 263 in t604x.dtsi, interrupt index 63, range ni3) fires continuously, all on CPU0. 84 samples of 2 s: 63k to 175k per second, median 122k.
  • No message from the driver. ce_count and ue_count are 0.
  • The other 37 claimed lines stay at 0.
  • 8 unbind and bind cycles of 24f000000.error-handler did not stop it. The storm is back as soon as the driver binds.
  • Cost: irq/199 uses about 9% of a core, and hardirq time is about 2.3% of all 14 CPUs. Battery draw rose by less than about 1 W, which I could not separate from workload noise.

The counter says the storm started at the wake: 69.2 M interrupts 586 s after it, which is 118k per second. I have no sample from before the sleep.

Likely cause (from the code, not measured): in apple_soc_edac_irq_thread(), the T6041 path returns IRQ_HANDLED without masking whenever apple_soc_edac_t6041_status() reads a non-zero word. If only bit 3 is set, error is false and it returns with no log and no count, which matches what I see. If the write-back does not clear the word, the level interrupt fires again at once. Since 5e3111b the mask and backoff only run when the word reads 0. A re-read after the write-back, falling through to the backoff when the word is still set, would bound it.

Relevant error lines

$ grep -E '^ *199:' /proc/interrupts
 199:  172630643  0  0  0  0  0  0  0  0  0  0  0  0  0  AIC2  65799 Level  24f000000.error-handler

kernel: apple_soc_edac 24f000000.error-handler: monitoring 38 error sources
kernel: PM: suspend entry (s2idle)
kernel: macsmc-input macsmc-input: lid open 72030001 at +10372564 ms in sleep, MSLD=0: wakeup
kernel: PM: suspend exit

$ cat /sys/devices/system/edac/24f000000.error-handler/24f000000.error-handler0/{ce,ue}_count
0
0

The driver printed nothing else in this boot.

Related issues, pull requests, or code examples

What was not tested

  • The status word at 0x274000000 + 0x416c was not read, so the bit 3 explanation is not confirmed.
  • Only one sleep and wake, on battery, with the lid. No wake on AC, no deep sleep.
  • No other kernel build, and no restart yet to confirm that the line is quiet again after a boot.
  • No other Mac.

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions