You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
aurora-sep: mesa-m3-session-env leaves the M3 Max alone, so Hyprland spins at 100 % CPU when the GPU is not started #190
recipes/mesa-m3/mesa-m3-session-env exports LIBGL_ALWAYS_SOFTWARE=1 (with MESA_M3_FALLBACK=software) when an M3 render node is present but /opt/mesa-m3 is not used.
It does this only on a T6030 or a T8122 (_mm3_soc_of, about :310-318). The header says that
"M1, M2, the M3 Max and anything else are left alone" (:60).
On a J516C boot where the GPU node stays disabled (Issue 1), no G15 render node appears. The
session then runs with the system Mesa and no working renderer: Hyprland spins at about 100 %
CPU and the desktop is unusable. Seen on the boots of 2026-10-23 01:45 and 02:30 (stack on #177, plain reboot with the ESP boot.bin, t6031-gpu=0).
Needed: a software-rendering fallback for the M3 Max (apple,t6031) in the session hook. The
case "no G15 render node" probably needs it as well as "render node present but not usable".
Nothing here was tested on the J514C or the t6034.
Not checked: I have not reproduced the no-render-node case on a current install, so a software-rendering fallback in mesa-m3-session-env may still be worth adding for it.
I would close this unless someone reports Hyprland spinning on a current install.
No /etc/m1n1.conf; GPU node apple,agx-t6034, status = "disabled"; no render node (/dev/dri has only card0, simpledrm)
Idle desktop, top -b -n 2 -d 3, second sample: Hyprland 413.6 %CPU
Hyprland log:
ERR from aquamarine ]: CDRMRenderer(drm): Can't create renderer, no matching devices found
ERR from aquamarine ]: drm: initMgpu: no renderer
ERR from aquamarine ]: drm: Failed to initialize renderer state for /dev/dri/card0
Frequency: every boot so far (one boot). Per-Mac report: #333.
On the GPU node: omacom/m1n1#19 is merged, but it only enables the GPU and display handoff by default for the J516C (T6031). From my reading of m1n1's kboot_gpu.c and kboot_display_handoff.c, there is no handoff path for the J514M (T6034) at all, so the GPU node stays disabled on a normal boot there. That is a separate gap from #19, not something #19 left half-done. I haven't tested it on a J514M.
Your log lines (initMgpu: no renderer, Failed to initialize renderer state for /dev/dri/card0) look like hyprwm/aquamarine#370, where the backend keeps retrying renderer creation when there is no render node. I haven't confirmed that this is what drives the CPU use on your Mac.
Could you tell me how big the Hyprland log is and how often the initMgpu lines repeat? That would show whether the retry loop is what burns the CPU.
Log size:hyprland.log was 3,810,431 bytes after 2 min 13 s of uptime, and 12,963,191 bytes (82,220 lines) after 7 min 11 s. It grows by about 100–243 MiB per hour.
How often initMgpu repeats: 29/s idle (30 s window), 71/s with the trackpad in use (60 s window), 43.5/s and 50/s in two other windows (10 s and 30 s). Each repeat writes these 6 lines (956 bytes):
DEBUG from aquamarine ]: Supported EGL client extensions: (16) ...
ERR from aquamarine ]: [EGL] Command eglQueryDeviceStringEXT errored out with EGL_BAD_PARAMETER (0x12300): eglQueryDeviceStringEXT (x2)
ERR from aquamarine ]: CDRMRenderer(drm): Can't create renderer, no matching devices found
ERR from aquamarine ]: drm: initMgpu: no renderer
ERR from aquamarine ]: drm: Failed to update renderer state for Unknown-1 on applyCommit
The log block you posted looks like the Hyprland/aquamarine retry loop tracked upstream in hyprwm/aquamarine#425 (same issue family as #370): with no GPU render node, aquamarine tries to build a renderer on every commit, fails, and doesn't remember the failure, so it repeats and writes the same lines each time. It is still open there, and as far as I can see no fix is in aquamarine main or Hyprland 0.57.1 yet.
Two things I have not shown yet:
That this loop is what causes the high CPU use. At 29-71 commits a second, Hyprland may simply be drawing frames in software, which could account for much of it. The log growth is certainly from this loop.
That your Mac uses the generic firmware framebuffer driver (simpledrm) with no render node. Could you run these and paste the output?
ls -l /dev/dri
readlink /sys/class/drm/card0/device/driver
pacman -Q aquamarine hyprland
The underlying reason on your Mac is still the missing GPU and display handoff for the J514M (T6034) that I described above; that is a separate gap and not something this issue fixes.
Per-thread CPU of Hyprland (pid 1239), from utime+stime deltas in /proc/1239/task/*/stat
over 30 s, grouped by thread name. 100 % = one core. Two windows, same boot:
window A (02:53:10Z) window B (02:54:13Z)
llvmpipe-0..13 (14 threads) 251.4 % 184.5 %
Hyprland main thread 8.1 % 5.8 %
other 18 threads 0.1 % 0.0 %
whole process (/proc/1239/stat) 259.4 % 190.5 %
"initMgpu: no renderer" lines/s 83.8 49.1
The main thread does all of the process's writes (/proc/1239/task/*/io: in a 10 s window its wchar delta is 585911 bytes, equal to the whole process; the log grew 498477 bytes). So the loop
that writes those lines runs on a thread using 6-8 % of a core. Almost all of the CPU is in
llvmpipe, the software renderer.
Thanks for the per-thread numbers, that settles it: nearly all the CPU is llvmpipe (software rendering), and the repeating initMgpu lines are only the log growth.
To see why the GPU stays off on this T6034, could you paste:
lsmod
dmesg filtered for the GPU and display probe, e.g. dmesg | grep -i -E 'agx|asahi|dcp|drm|simpledrm'
I haven't been able to retest on a J514M. The lab Macs are offline right now, so this is from reading the code only.
One safety note: at ~50 KB/s the log is roughly 100-240 MiB an hour. hyprland.log sits in $XDG_RUNTIME_DIR (tmpfs), so it eats RAM rather than disk, and a full tmpfs breaks other things (a VMware report in aquamarine#425 saw that). A workaround from that thread is a timer that truncates the log in place once it gets big (e.g. every few minutes); it only stops the growth, not the CPU use. I haven't tried it on an M3 myself.
Nothing matches agx, asahi or dcp; the kernel log has no line about the GPU or a display controller at all (also checked the full 851-line buffer for gpu|display|dptx|t603). Since the pattern can't show a node that is never probed, here is what the log doesn't say:
So: the asahi module is in the module tree but not loaded, the GPU node is status = "disabled", and there is no node in the live device tree whose compatible contains dcp. The only display node is chosen/framebuffer@108d2f82880 (apple,simple-framebuffer), which is what simpledrm bound to.
lsmod (133 lines; simpledrm is built in, so it does not appear):
Follow-up: with the GPU running (unreleased test build: kernel from #408, Mesa from omacom/mesa#33, m1n1 from omacom/m1n1#32), Hyprland on this J514M idles at 7-10 % of one core instead of 190-731 %, renders on Apple M3 Max (G15C B1), and logs no initMgpu lines. Results: #333 (comment)
Activity
malik-na commentedon Oct 31, 2026
Update: this should no longer happen on a normal boot, because of omacom/m1n1#19 (see #189).
mesa-m3-session-envmay still be worth adding for it.I would close this unless someone reports Hyprland spinning on a current install.
techgyani commentedon Nov 3, 2026
Still happening on a T6034 (J514M), which m1n1#19 does not cover, so the GPU node stays disabled on a normal boot.
apple,j514m,apple,t6034uname -r7.2.11-14-1-ARCH,linux-aurora 7.2.11.aurora14-1,m1n1-aurora 1.6.1.aurora22-1(stage 2bc79acc4-dirty),mesa-aurora 1:26.3.0_devel.20261031.231483.a1969111475-1,hyprland 0.57.0-4asahi,os-fw-version: 14.7,asahi,system-fw-version: unknown/etc/m1n1.conf; GPU nodeapple,agx-t6034,status = "disabled"; no render node (/dev/drihas onlycard0, simpledrm)top -b -n 2 -d 3, second sample: Hyprland413.6%CPUFrequency: every boot so far (one boot). Per-Mac report: #333.
malik-na commentedon Nov 3, 2026
Thanks for the detailed report, this is useful.
On the GPU node: omacom/m1n1#19 is merged, but it only enables the GPU and display handoff by default for the J516C (T6031). From my reading of m1n1's
kboot_gpu.candkboot_display_handoff.c, there is no handoff path for the J514M (T6034) at all, so the GPU node stays disabled on a normal boot there. That is a separate gap from #19, not something #19 left half-done. I haven't tested it on a J514M.Your log lines (
initMgpu: no renderer,Failed to initialize renderer state for /dev/dri/card0) look like hyprwm/aquamarine#370, where the backend keeps retrying renderer creation when there is no render node. I haven't confirmed that this is what drives the CPU use on your Mac.Could you tell me how big the Hyprland log is and how often the
initMgpulines repeat? That would show whether the retry loop is what burns the CPU.techgyani commentedon Nov 3, 2026
Log size:
hyprland.logwas 3,810,431 bytes after 2 min 13 s of uptime, and 12,963,191 bytes (82,220 lines) after 7 min 11 s. It grows by about 100–243 MiB per hour.How often
initMgpurepeats: 29/s idle (30 s window), 71/s with the trackpad in use (60 s window), 43.5/s and 50/s in two other windows (10 s and 30 s). Each repeat writes these 6 lines (956 bytes):malik-na commentedon Nov 3, 2026
Thanks, this is exactly the data I needed.
The log block you posted looks like the Hyprland/aquamarine retry loop tracked upstream in hyprwm/aquamarine#425 (same issue family as #370): with no GPU render node, aquamarine tries to build a renderer on every commit, fails, and doesn't remember the failure, so it repeats and writes the same lines each time. It is still open there, and as far as I can see no fix is in aquamarine main or Hyprland 0.57.1 yet.
Two things I have not shown yet:
simpledrm) with no render node. Could you run these and paste the output?ls -l /dev/drireadlink /sys/class/drm/card0/device/driverpacman -Q aquamarine hyprlandThe underlying reason on your Mac is still the missing GPU and display handoff for the J514M (T6034) that I described above; that is a separate gap and not something this issue fixes.
techgyani commentedon Nov 4, 2026
simple-framebufferis the platform driver name, whichsimplefbuses as well. This shows the module behind it:Per-thread CPU of Hyprland (pid 1239), from
utime+stimedeltas in/proc/1239/task/*/statover 30 s, grouped by thread name. 100 % = one core. Two windows, same boot:
The main thread does all of the process's writes (
/proc/1239/task/*/io: in a 10 s window itswchardelta is 585911 bytes, equal to the whole process; the log grew 498477 bytes). So the loopthat writes those lines runs on a thread using 6-8 % of a core. Almost all of the CPU is in
llvmpipe, the software renderer.
malik-na commentedon Nov 4, 2026
Thanks for the per-thread numbers, that settles it: nearly all the CPU is llvmpipe (software rendering), and the repeating initMgpu lines are only the log growth.
To see why the GPU stays off on this T6034, could you paste:
lsmoddmesgfiltered for the GPU and display probe, e.g.dmesg | grep -i -E 'agx|asahi|dcp|drm|simpledrm'I haven't been able to retest on a J514M. The lab Macs are offline right now, so this is from reading the code only.
One safety note: at ~50 KB/s the log is roughly 100-240 MiB an hour. hyprland.log sits in $XDG_RUNTIME_DIR (tmpfs), so it eats RAM rather than disk, and a full tmpfs breaks other things (a VMware report in aquamarine#425 saw that). A workaround from that thread is a timer that truncates the log in place once it gets big (e.g. every few minutes); it only stops the growth, not the CPU use. I haven't tried it on an M3 myself.
techgyani commentedon Nov 6, 2026
Both from a fresh boot (captured ~2.5 min after boot, 2026-11-06T04:46Z).
dmesgruns unprivileged here (kernel.dmesg_restrict= 0).dmesg | grep -i -E 'agx|asahi|dcp|drm|simpledrm'— these are all the matches:Nothing matches
agx,asahiordcp; the kernel log has no line about the GPU or a display controller at all (also checked the full 851-line buffer forgpu|display|dptx|t603). Since the pattern can't show a node that is never probed, here is what the log doesn't say:So: the
asahimodule is in the module tree but not loaded, the GPU node isstatus = "disabled", and there is no node in the live device tree whosecompatiblecontainsdcp. The only display node ischosen/framebuffer@108d2f82880(apple,simple-framebuffer), which is whatsimpledrmbound to.lsmod(133 lines;simpledrmis built in, so it does not appear):lsmod
techgyani commentedon Nov 8, 2026
Follow-up: with the GPU running (unreleased test build: kernel from #408, Mesa from omacom/mesa#33, m1n1 from omacom/m1n1#32), Hyprland on this J514M idles at 7-10 % of one core instead of 190-731 %, renders on
Apple M3 Max (G15C B1), and logs noinitMgpulines. Results: #333 (comment)