Skip to content

j615 (MacBook Air 15-inch M3): test results on a patched aurora16; USB keyboard on a monitor hub can't wake from s2idle #492

Open

Description

@freber

Claude Code wrote this up from my tests.

This is the per-Mac issue for my J615. I'll add later results here as comments.

Most results come from my own test kernel, which is aurora16 with three aurora-silicon/linux pull requests and #356 on top. A pass on that kernel doesn't mean stock aurora16 passes.

Machine and build

Mac MacBook Air 15-inch M3 (Mac15,13)
Board / SoC apple,j615 / apple,t8122
Distribution Omarchy on Arch Linux ARM (omarchy-dev 4.0.0.r7840.gac6cb9c-1)
Release tag none (packages from Omarchy's repo)
OS firmware 14.7
System firmware 26.6.2
Running kernel 7.2.11-16-1-j615-ARCH, my test kernel (linux-aurora-j615 7.2.11.aurora16.j615-3)
Stock kernel, kept as fallback linux-aurora 7.2.11.aurora16-1 (bed544b)
Boot chain m1n1-aurora 1.6.1.aurora23-2 (stage 2 ba130ffb-dirty), uboot-aurora 2026.07.aurora6-1, limine 12.9.6-1
Mesa mesa-aurora 1:26.3.0_devel.20261107.231485.b7b03c5b3b8-1
Renderer Apple M3 (G15G C0), OpenGL ES 3.2 Mesa 26.3.0-devel (git-b7b03c5b3b), from Hyprland's log. glxinfo isn't installed
Compositor Hyprland 0.57.2 (hyprland 0.57.2-2, commit 11cbc6c5)
GPU settings asahi.t8122_pstate_cap=5 on the command line, m3_render_batch_override=2 set at boot through tmpfiles
Other kernel arguments user4k=on

What the test kernel adds to aurora16

I made two more changes locally. The floor governor runs with gpu_busy_pct 20 and hold_ms 3000, set at boot through tmpfiles (reasons in aurora-silicon/linux#1486 (comment)). m1n1's boot.bin is rebuilt by hand with the #356 t8122-j615.dtb and the line chosen.asahi,t8122-dcpext=1. While that boot.bin is in place the stock aurora16 entry boots with apple_t8122_display.dcpext=0 (see my question on #356).

Results

"Both" in the kernel column means stock aurora16 and my test kernel.

Feature Result Kernel Notes
Internal panel with GPU rendering in Hyprland pass both 2880x1864 at 60 Hz on the renderer above
External display over USB-C (DP alt mode), both left ports pass test Needs #356. LG Ultra HD at 3840x2160, 60 Hz, 10-bit. Hotplug, replug, coldplug, lid closed and s2idle all passed. Full report in #356 (comment)
Memory bandwidth pass test Needs aurora-silicon/linux#1486 with the J615 enable from aurora-silicon/linux#1985. A 4-thread read went from 19.8 GB/s on stock to 68.2 GB/s with the floor
CPU 0 speed pass test Needs aurora-silicon/linux#1517. A fixed loop on CPU 0 took 302.7 ms on stock and 69 ms with the fix
Chromium scrolling pass both Render batching 2 took dropped frames from 2.2 to 0 percent on stock. On the test kernel, scroll to screen (p50) matches or beats my 13-inch M2 in all four of my test scenes. Numbers in #385 (comment)
s2idle, woken by opening the lid pass test 48 minutes with the monitor connected. Still locked before I authenticated, same Hyprland PID
Wake from s2idle with a USB keyboard on the monitor's hub fail test See below
Panel brightness keys fail both ls /sys/class/backlight lists nothing. #354 should fix it. I've built it into my test kernel but haven't booted that build yet
SoC die temperatures in hwmon pass test With options macsmc_hwmon soc_die_keys= listing the 12 J615 keys. #410 adds them for the J615 (it covers my #464)
Audio over DisplayPort not tested test The driver logs No audio support
Dock or hub with its own DP output not tested I don't have one
Thunderbolt dock or display not tested #356 covers DP alt mode only
Second external monitor not tested #356 enables one external pipe. #127 is J613 work on a second display with the lid closed

For this report I haven't tested Wi-Fi, Bluetooth, speakers, microphone, headphone jack, camera, Touch ID, keyboard backlight, battery life or power draw with the floor on.

USB keyboard on the monitor's hub can't wake the Mac from s2idle

An Apple Magic Keyboard and Magic Trackpad are plugged into the LG monitor's USB hub with cables. The monitor connects to the front left port with one USB-C cable. /sys/power/mem_sleep lists only [s2idle].

  1. Wakeup is enabled by default on the keyboard (1-1.3), xhci-hcd.3.auto and b02280000.usb. I also enabled it on the root hub and the monitor's hub.
    echo enabled | sudo tee /sys/bus/usb/devices/usb1/power/wakeup /sys/bus/usb/devices/1-1/power/wakeup
    
  2. Close the lid and run systemctl suspend.
  3. Wait about 30 seconds and press keys.

I tried this once, after the command above. I expected a key press to wake the Mac, but nothing happened. I opened the lid 35 seconds after the suspend and it woke at once.

On resume the monitor's hub and all devices on it disconnect and enumerate again. That happens about 1.5 seconds after PM: suspend exit and about 50 ms after the DP link to the monitor drops. The USB bus stays registered and the log has no line about the root hub losing power. Both resumes look the same. Here is the second one with logind's lines merged in.

18:12:07.769 systemd-logind: Lid closed.
18:12:13.460 systemd-logind: The system will suspend now!
18:12:14.077 kernel: PM: suspend entry (s2idle)
18:12:49.543 systemd-logind: Lid opened.
18:12:49.549 kernel: PM: suspend exit
18:12:50.991 kernel: apple-dcp 31ac00000.dcp: dcp_hotplug() connected:0 valid_mode:1 nr_modes:16
18:12:51.049 kernel: apple-dcp 31ac00000.dcp: dcp_poweroff() done
18:12:51.060 kernel: apple-dcp 31ac00000.dcp: dcp_dptx_disconnect(port=0)
18:12:51.107 kernel: usb 1-1: USB disconnect, device number 7
18:12:51.107 kernel: usb 1-1.1: USB disconnect, device number 8
18:12:51.109 kernel: usb 1-1.2: USB disconnect, device number 9
18:12:51.137 kernel: usb 1-1.3: USB disconnect, device number 10
18:12:51.205 kernel: usb 1-1.4: USB disconnect, device number 11
18:12:51.456 kernel: usb 1-1: new high-speed USB device number 12 using xhci-hcd

I can't tell from this log whether the controller kept running through s2idle or whether the monitor reset its hub when the DP link dropped. I don't know yet if the disconnect has anything to do with the missing wake.

#476 has the same symptom on a J316C behind a CalDigit TS4, with the XHCI_RUN_IN_S2IDLE code that my kernel also has (dwc3-apple sets it). There the devices stay enumerated across sleep. Mine disconnect on resume. #121 has it on a J516S, filed before that code landed.

I haven't yet tried a keyboard plugged straight into a USB-C port with no hub, which is the test #476 suggests. I also haven't tried the rear left port (by the hinge) or the stock kernel. Dynamic debug for xhci_hcd and usbcore was off.

DCP firmware messages on every boot

These three lines appear on all 10 boots in my journal, on stock aurora14 and aurora16 as well as my test kernel. The internal panel works on all of them, so as far as I can tell they're harmless.

apple-dcp 28ec00000.dcp: RTKit: syslog message: UPPipeDCP_H13P.cpp:1775: virtual bool IOMFB::UPPipeDCP_H13P::late_init_ios(bool): Realtime Bandwidth setup failed
apple-dcp 28ec00000.dcp: RTKit: syslog message: UPPipe.cpp:4406: IOMFB read_pmu_data_sync: pmu ram read error (e00800d8)
apple-dcp 28ec00000.dcp: RTKit: syslog message: FramebufferDCP.cpp:2488: IOMFB:create_default_fb_surface: Creating default FB surface failed - width = 2880, height = 1864

Other threads from this Mac

Activity

  1. freber commented on Nov 17, 2026

    @freber
    Author

    Claude Code wrote this up from my tests.

    Update 2026-11-17. I've moved to a new daily kernel, linux-aurora-j615gpu. It is the test kernel from the first post plus #410, #491 and a T8122 series that lets the GPU use states 6 to 8 (up to 1338 MHz) under the m3_thermal limit. The command line now has asahi.t8122_pstate_cap=8. Since this afternoon it runs a build of that series with adjustable thermal thresholds, set to asahi.m3_thermal_hot=90 asahi.m3_thermal_cool_gap=3. The released form of the series stops at 85 C and has no gap setting, so those two values need that build.

    Before (cap 5, no GPU temperature feedback) Now (cap 8, limit at 90 C, gap 3)
    gpubench throughput, w4, 10 s runs 307 fps 289 fps
    Same, with the default limit at 80 C 218 fps
    gpubench pace, w4, p50 latency 7.6 ms 4.8 ms
    600 s GPU load, w10 test guard stopped it at 16 s (hottest key 106.4 C) 88.1 fps for 600 s, hottest key 93.8 C
    Missed frames 0 0

    The cap 5 kernel wins the 10 s bursts because it has no GPU temperature feedback. In the long run that let Tg0P climb from 56.5 to 90.9 C in the first 15 s, still rising when my 95 C test guard stopped it. Background CPU load from another job may have added to the last few seconds. A GPU limit with temperature feedback is the safer daily setup.

    All six CPU thermal zones from #491 are live. Under a combined CPU and GPU load they held the die at 95 C for 300 s (max 95.8, mean 94.9). Details are in my #491 comment.

    I haven't tested suspend, the external display or hardware video decode in Chromium (#436) on this kernel yet.

  2. freber commented on Nov 18, 2026

    @freber
    Author

    Claude Code wrote this up from my tests.

    Update 2026-11-18. My daily kernel is now linux-aurora-j615next. It is the j615gpu kernel from the update above plus aurora-silicon/linux#1937 (reseed the M3 completion stamps before they wrap), #472 (now aurora-silicon/linux#2161) and #438. The #1937 commit is ported to the aurora16 base. The port leaves out one upstream line that touches peer_stamp, which this tree doesn't have. The boot line is the same as before.

    To test #1937 on the J615 I ran a small GLES program that pushes the completion stamps past every point where the old kernel wraps. Each render pass is one 16x16 draw and a glFlush, so it reaches the kernel as one render command. In compute mode each command is one 1x1x1 dispatch. The counts below are the program's own commands since boot. The desktop adds its own passes, so the firmware count is a little higher.

    Run Our commands since boot Time Points passed
    Render, part 1 16,000,000 99 min reseed 1 at 7,733,248, old TA wrap at 8,781,824, old fragment wrap at 12,779,520 (first reported crash), reseed 2 at 15,466,496
    Render, part 2, same boot 30,000,000 88 min reseed 3 at 23,199,744, old TA wrap at 25,559,039, old fragment wrap at 29,556,735 (second reported crash)
    Compute 6,500,000 57 min reseed 1 at 3,080,192, old compute wrap at 4,128,768, reseed 2 at 6,160,384

    All three runs finished. The kernel log has no "Non Sequential Stamp" line, no firmware crash and no asahi errors. The desktop kept drawing the whole time. The hottest die key stayed under 49 C and the longest wait on a fence was 61 ms.

    The kernel doesn't log a reseed, so the evidence is that the GPU kept working past these counts with a clean log. I haven't reproduced the crash on a kernel without #1937, since that ends in a hard power off.

    The test program is about 600 lines of C. I can share it if someone wants to run it on another M3.

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