Skip to content

drm/asahi: let T8122 thermal control use the full performance table - #317

Draft
DjDeveloperr wants to merge 1 commit into
integration/aurorafrom
fix/t8122-thermal-performance-20261031
Draft

DjDeveloperr wants to merge 1 commit into
integration/aurorafrom
fix/t8122-thermal-performance-20261031

Conversation

@DjDeveloperr

Copy link
Copy Markdown
Collaborator

The J613 boot log reports an eight-state GPU table but clamps both the firmware ceiling and runtime range to state 2 (table label 618 MHz), even while valid SoC die temperatures change. The omitted T8122 experiment cap currently prevents the existing thermal governor from using higher states.

This candidate lets thermal control use the published table when that experiment cap is omitted, retaining state 2 at startup and on critical temperature or sensor failure. Thermal-off boots retain fixed state 2; explicit caps 1–5, hold behavior, global maximum/boot limits and other SoCs retain their policy. Scheduling, batching, Mesa and the firmware ABI are unchanged.

Validation: native ARM64 affected Asahi Rust/C linked object builds; all 27 checked dependency/driver artifacts match buffered/O_DIRECT hashes. The production-policy host suite passes 32 tests, and a temporary restoration of the old ceiling fails the regression test. The Air tool suite passes 86 tests with four skips; checkpatch reports no findings.

Draft pending hardware qualification. State-8 clock/throughput, correctness, sustained thermal response and recovery have not been tested. The full local kernel build did not complete: recurring build-VM cache corruption affected Rust metadata and tools. Evidence was preserved, the idle VM shut down gracefully, and its disk cloned before restart; the affected-driver compile passed afterward. This does not establish a fix for that VM fault. Full CI is requested separately. No candidate is installed on the laptop; auto-unlock and native Limine one-shot recovery must be verified before testing.

The omitted T8122 experiment cap resolves to state 2 and is then applied
as a firmware ceiling, so the thermal governor cannot raise the runtime
cap even with valid SoC temperature readings. On J613 the captured boot
reports eight table states but a firmware/runtime range of 2..2.

Track whether the experiment cap was explicitly supplied. With thermal
control enabled, let omission use the table ceiling while retaining
state 2 for startup and sensor/critical-temperature fallback. Keep the
fixed state-2 policy with thermal control off, and preserve all explicit
caps 1..5 and other SoCs' policy. Hold mode exposes the table but keeps
the runtime cap at state 2 for qualification.

Exercise the production parser and thermal decisions in the existing
host suite, including sensor failures, heat, pacing and cap enforcement.
Correct the experiment help and update its installer checksum together.

Native affected-driver compilation and host tests pass. Higher-state
hardware correctness, thermal behavior and performance remain untested.

Signed-off-by: DjDeveloperr <hey@djdev.me>
@freber

freber commented Nov 17, 2026

Copy link
Copy Markdown

Claude Code wrote this up from my tests.

I've been running states 6 to 8 on a J615 (15-inch M3 Air) and opened #500 with what came out of it. Where this PR lifts the ceiling when the cap is omitted, #500 keeps today's default. An explicit asahi.t8122_pstate_cap of 6 to 8 takes effect only with asahi.m3_thermal=on, while hold and off lower it to 5. It's stacked on #410.

#500's second commit applies here too. When the ceiling is above the safe cap, one client that keeps the pipeline full can stop the after-job thermal check from running, so the cap stays high while the die heats up. The commit makes the pipeline drain once the last check is 500 ms old, so the check runs again. It would fit on top of this PR as well. If you'd rather keep this approach, I can rebase it onto your branch.

On the J615, short GPU runs went from 218 fps at the default 80 C limit to 289 fps at 90 C with a 3 C gap. A 600 s load went from 80 to 88 fps, with the hottest key at 93.8 C. With the default 80 C and gap of 10, a long load never got above state 2 again after the first seconds. The full table is in #500.

freber added a commit to freber/linux that referenced this pull request Nov 22, 2026
… range

Two gaps in the host tests that the rest of this series leans on.

The nine unit tests in m3/t8122_knobs.rs have never run. No file in
tools/asahi includes the module. It does not compile on its own either,
because it takes its default InitData version from m3_firmware, which
needs the kernel crate. pstate_cap_accepts_one_to_five_only, the test
that pins the range of asahi.t8122_pstate_cap, is one of them. Include
the module in tools/asahi/m3-runtime-tests.rs, the suite the kernel CI
already runs, with a stand-in for that one constant. The stand-in's
value is checked by defaults_are_the_plan_baseline, so it cannot drift
from m3_firmware without a failing test.

Policy in m3_thermal_policy.rs moves the runtime cap between the safe
cap and the ceiling. The cap_tests from aurora-silicon#410 drive it with a ceiling of
5 over a safe cap of 2. Nothing covers ceilings 6 to 8, or the check
that finds a firmware ignoring the cap. Add tests for the rules as
they are, for ceilings 6 to 8 above a safe cap of 5 and for ceilings 1
to 5:

- the cap starts at the safe cap and rises one state per 2 s on cool
  readings, never above the ceiling;
- between cool and hot it holds;
- hot lowers one state per fresh reading, at most every 500 ms, never
  below the safe cap;
- critical, no reading, a failed, stale, implausible (including the
  -2.849 C an unpowered SMC GPU key reads) or stuck reading returns to
  the safe cap at once, with the previous cap tolerated for 1 s;
- after a fallback the cap climbs again at the same pace;
- hold never leaves the safe cap;
- a ceiling equal to the safe cap never moves;
- a firmware that stays above a lowered cap is found after 20 checks
  spanning at least 3 s.

Three mutations of Policy (critical keeps the cap, the 3 s span dropped,
the raise not bounded by the ceiling) each fail one to three of the new
tests.

No driver code changes.

Link: omacom/linux-aurora#317
Link: omacom/linux-aurora#410
Signed-off-by: Fredrik Silfver <1684291+freber@users.noreply.github.com>
freber added a commit to freber/linux that referenced this pull request Nov 22, 2026
The thermal limit moves the runtime cap only inside
Config::check_pstate. That runs after a job only in two conditions,
which a ceiling above the safe cap makes matter.

First, the job thread runs the check only when asahi.m3_retire_mmio has
bit 2 (the performance-state register read). The parameter is
root-writable at runtime, so writing 0 to 3 stops both the thermal limit
and the check that the firmware stays within the runtime cap. Once the
ceiling is above the safe cap, the runtime cap is the only thing that
keeps the firmware below the ceiling. A frozen governor would leave it
at whatever state it last allowed. Add Policy::above_safe()
(host-testable), Governor::enforces() (the limit is on or holding and
the ceiling is above the safe cap) and Config::pstate_check_required().
Run the after-job check when bit 2 is set or the check is required. Bit
2 is the T8122 default, so this forces only what runs by default. The
first retirement of a module load at which the check runs although bit 2
is clear logs one line, so a write that leaves bit 2 out is not silently
overridden.

Second, the check runs only at a retirement with no later batch in
flight. The T8122 default pipeline depth is 2. A render batch of the
same VM is published behind the running one while its passes fit the
4000 us budget. So one client that keeps submitting short passes, with
no other client rendering (a still or blanked screen, offscreen work),
can keep the check from running at all, at 90 C too. While the ceiling
is above the safe cap, pipeline_slot() now refuses a slot behind batches
in flight once the last check is 500 ms or more old
(Policy::overlap_allowed, CHECK_EVERY_NS), so the pipeline drains and
the existing checked retirement runs. The sensor is read every 500 ms
while busy, so more frequent checks would see no newer reading. The
longest time between checks is then 500 ms plus the run time of the
batches in flight. Every firmware-visible action stays where it runs
today, after the last batch has retired and the pipes are idle. The
other way, moving the cap on a pipelined retirement, would be a Globals
store with a later batch running on the GPU, which has never been done
on this firmware.

With aurora-silicon#410 the T8122 safe cap is 2, so both already apply to boots that
set asahi.t8122_pstate_cap to 3 to 5 with the thermal limit on or
holding. Those boots now drain the pipeline once per 500 ms under
sustained pipelined load. With the ceiling at the safe cap nothing
changes, since overlap_allowed() returns true before reading the clock.

New tests in m3_thermal_policy.rs simulate a client that keeps the
pipeline full (depth 1 to 4, batches of 0.1 to 50 ms). Without the bound
it is checked once, at boot, so 90 C leaves the cap at 8. With it the
longest gap between checks is within 500 ms plus the batches in flight,
the cap falls from 8 to the safe cap within that time of a critical
reading, a firmware that ignores the cap in hold mode is found. With the
ceiling at the safe cap, the publication order is identical to today's.
Three mutations of the bound each fail two to four of them.

Link: omacom/linux-aurora#317
Link: omacom/linux-aurora#410
Signed-off-by: Fredrik Silfver <1684291+freber@users.noreply.github.com>
freber added a commit to freber/linux that referenced this pull request Nov 22, 2026
… limit

asahi.t8122_pstate_cap refuses anything above 5 (PSTATE_CAP_MAX in
t8122_knobs.rs). The stated reason is that J613 and J615 had no SoC die
temperature zone, so there was no temperature feedback above the
driver's no-feedback limit. Both boards have one now (macsmc_soc_die).
The rest of the path already exists. The firmware always gets the whole
eight-state table. With the thermal limit on, PstatePolicy::new already
makes the cap the ceiling, so only the knob range keeps the governor
below 6.

Accept 6 to 8, as the thermal limit's ceiling only. With aurora-silicon#410 below
this, the runtime cap of every T8122 boot starts at state 2 and falls
back to it, whatever the ceiling.

- asahi.m3_thermal=on (the default). The ceiling is the cap. The runtime
  cap starts at 2 and rises one state per 2 s only on usable, cool
  readings of the SoC die zone. No zone, or a failed, stale,
  implausible, stuck or critical reading, puts it back to 2 at once.
  These are the governor's existing rules.
- asahi.m3_thermal=hold. 6 to 8 are lowered to 5, with one boot line.
  The firmware gets 5 and the runtime cap stays at 2.
- asahi.m3_thermal=off. 6 to 8 are lowered to 5, with one boot line.
  aurora-silicon#410's fixed cap without the limit then gives the firmware 2. These
  values were refused before in every mode (ENODEV, no GPU for that
  boot). A cap left on the command line from a ranging boot while the
  thermal limit is switched off should not cost the GPU.

Values of 0, 9 and up are refused as before.

Whether the zone is usable can only be decided at runtime. On a J615 the
GPU firmware starts about 1.6 s into boot from the initramfs. The
macsmc_soc_die zone appears at about 7.6 s, when macsmc_hwmon registers
it. A probe-time zone check would refuse every boot, so the zone
condition is the governor's existing rule that only a usable reading
raises the runtime cap. The line the governor logs for a usable reading
now ends in "the runtime cap may rise above 2 up to N (MHz)" while the
ceiling is above the safe cap, so the boot log shows when the higher
states became possible. It is one line per change of the reading's
source, as before.

pstate_cap_use() in t8122_knobs.rs states the rules above in plain code
that the host tests drive. PstatePolicy::new gets the lowered ceiling
from admitted_cap(). Contents::select refuses the start if what
PstatePolicy::new computed is ever more than pstate_cap_use() allows, so
the two cannot drift apart silently. A host test checks that they are
equal for every cap and mode. Const assertions tie the no-feedback limit
(5) and the fallback (2) to the two constants of m3_thermal_policy.rs
that PstatePolicy::new uses. The boot lines name the case (lowered,
ranging or held) and the knob's value, which asahi.m3_max_pstate can
still lower later. When that parameter is given they say so and point to
the "firmware highest state" line, which has the result.
tools/aurora-sep/M3-GPU.md gets a paragraph on caps 6 to 8.

Tested on a J615 (15-inch M3 MacBook Air) with this series and aurora-silicon#410 on
bed544b (the 7.2 line), asahi.t8122_pstate_cap=8 and the thermal limit
at its defaults. On a cool die the runtime cap rose from 2 to 8 within
10 s of boot. Under GPU load the die passed 80 C within about 2 s at
state 8 and the cap fell back to 2 one state per 500 ms. A 600 s GPU
load and a 300 s combined CPU and GPU load ran without asahi errors,
faults or timeouts.

Link: omacom/linux-aurora#317
Link: omacom/linux-aurora#410
Signed-off-by: Fredrik Silfver <1684291+freber@users.noreply.github.com>
freber added a commit to freber/linux that referenced this pull request Nov 22, 2026
The M3 thermal limit lowers the runtime cap at asahi.m3_thermal_hot
(default 80 C, at most 85 C) and raises it again only 10 C below that.
On a J615 under a sustained GPU load the die then settles between the
two (83 to 84 C with hot at 85). The cap stays at the safe state 2 for
the whole load, since nothing ever reaches the cool point again.

Let both be set further for boards and users that want it. The defaults
do not change.

- asahi.m3_thermal_hot goes up to 92 C. Critical stays 8 C above it, so
  at most 100 C, below the 110 C critical trip of the J613 and J615 CPU
  thermal zones.
- asahi.m3_thermal_cool_gap (default 10 C, 3 to 10 C) sets how far below
  hot the die must be before the cap rises again.
- Both are writable at runtime in /sys/module/asahi/parameters. The
  governor reads them at each update, applies the clamped values from
  the next reading with the usual pacing and logs the new thresholds.

The bounds and the clamping move into m3_thermal_policy so the host
tests cover them. New tests cover the clamping, the default gap, a
narrower gap, a higher hot and thresholds changed at runtime.

Tested on a J615 with asahi.t8122_pstate_cap=8, stepping the two values
at runtime. A 600 s GPU load (gpubench w10, depth 2) ran at 80.0 fps
with the defaults, 85.8 fps with hot 88 and gap 3, 88.1 fps with hot 90
and gap 3 and 95.2 fps with hot 92 and gap 3. A 300 s combined CPU and
GPU load stayed at the safe state 2 at every setting. The CPU thermal
zones kept the hottest die sensor at 92 to 94 C on average, with peaks
of 93.9 to 97.6 C.

Link: omacom/linux-aurora#317
Link: omacom/linux-aurora#410
Signed-off-by: Fredrik Silfver <1684291+freber@users.noreply.github.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

ci:build Build, test and boot the kernel for this PR now

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants