[RyanTheTide]

The G16S zero-utilisation-under-load failure is still unresolved. 031c80d1 fixes utilisation admission across core gating, but it is not evidence for the remaining G16S cause. Keep this PR draft until a paired raw-statistics/load capture establishes that cause and a candidate passes the same capture.

The retained core-gating fix separates pstate from utilisation: completed utilisation intervals keep their bounded 16/32 ms age policy, while pstate needs a new observation after gating. Suspend, restart, quarantine and teardown still end both capabilities.

6dc719d3c21c215c6885bdb97f1c6d3dddc13ff5 adds a minimal diagnostic to the shared backend. /sys/kernel/debug/asahi-g17/firmware_reports now identifies its platform device and firmware graph, shows live cursors only for the graph's actual roles, and prints quarantine, suspension, queued-record discard and core-power admission state. Each raw tag has its transport-admitted count and last-admission decision. Unknown tags retain their complete offending copy before the existing quarantine takes effect. Qualified values appear beside their validity status. The fast-forward correction b6465f7ab5c2f8b6ca09c93f62008cd34eb9d52e formats the device CStr as a borrowed UTF-8 string for the core formatter. These additions change no decoder field mapping, channel setup or admission policy; there are no new module parameters or SoC switches.

The source path already gives T6040/G16S one V25G83 firmware instance, initializes its statistics-channel pointers from that instance's status area, and drains the primary role through the common event service. Sysfs retains the same session Arc as the receiver. In the current code, a zero core-power observation gates pstate only. A valid numeric zero must be checked against the admitted tag-0x13 words and the bounded integer mean; a missing or expired sample is unavailable/stale. None of these source checks proves what G16S firmware publishes during the failing workload.

Validation for the diagnostic: the focused host gate passes 11/11 runs. Its receiver control executes 12 production Rust cases, including the raw-report store, plus 512 byte-identical legacy C formatter fixtures. All nine negative controls compile and are killed by their intended assertion/panic. Standalone decoder tests also pass, and rustfmt is clean. The initial kernel directory build caught the CStr/core Display mismatch in the diagnostic identity line; the correction above fixes it without allocating. The corrected head passes the controller buildbox check with defconfig + asahi.config, CONFIG_RUST=y, LLVM 22.1.6, Rust 1.97.1, bindgen 0.73.2, modules_prepare and -j8 W=1 Asahi/PMP directory builds. Both required objects and the configuration were recovered and their SHA256 values verified. The only W=1 warnings are unchanged pmc.c kernel-doc for has_pending. This is a directory build, not an Image link or target boot. No new target runtime pass is claimed.

The existing fast-forward revert 0764ba4630db continues to remove the rejected DRAM fallback and one-off CI changes. PMP parsing remains driven by DT: absent metadata produces no dram/ group. The separate DRAM root cause is an embedded Stage 2 DTB set older than the matching kernel DTS, so an Image/modules or /boot/dtb update cannot supply those properties. aura#28 contains the Stage 2 packaging fix and is unchanged by this diagnostic; its deployment remains with the lead. No kernel fallback is reintroduced.

Exact verification after separate controller approval for the explicitly named device (and, for the new diagnostic, a lead-installed candidate with the existing bootguard fallback):

# J616s / G16S
stats=/sys/devices/platform/28a600000.gpu/agx_stats
dram=/sys/bus/platform/devices/5003c0000.pmp-report/dram
# Neo / G17P: substitute these two paths instead
# stats=/sys/devices/platform/soc/482600000.gpu/agx_stats
# dram=/sys/bus/platform/devices/3003c0000.pmp-report/dram
reports=/sys/kernel/debug/asahi-g17/firmware_reports
uname -r
tr '\0' '\n' < /proc/device-tree/compatible
cat "$reports"  # confirm the platform device matches $stats; not tied to a DRM minor
for file in type data_bus_width state_count max_data_rate peak_bandwidth bandwidth_class; do
    printf '%s ' "$file"; cat "$dram/$file"
done
sample() {
    while :; do
        cat /proc/uptime
        cat "$stats"
        cat "$reports"
        sleep 0.005
    done
}
sample > idle-before.log & sampler=$!
sleep 3
kill "$sampler"; wait "$sampler" || true
sample > load.log & sampler=$!
load_rc=0
timeout --signal=INT --kill-after=5s 20s glmark2-es2-drm --off-screen --run-forever -b terrain > glmark2.log 2>&1 || load_rc=$?
kill "$sampler"; wait "$sampler" || true
printf 'load_rc=%s\n' "$load_rc" # timeout normally returns 124
sample > idle-after.log & sampler=$!
sleep 3
kill "$sampler"; wait "$sampler" || true
cat glmark2.log

Require an Apple hardware renderer and a workload that survives the full interval. Preserve all three phases and the workload exit status. Correlate tag counts/cursor progress, transport admission, raw words 3/4/6 of tag 0x13, and the qualified mean/status. Unknown tags or nonzero alternative fields require firmware-layout evidence before changing the decoder. A last_admitted=false record must be explained by the printed admission state. For the eventual fix, require fresh nonzero device utilisation under load, a higher load distribution than idle, and independent pstate invalidation across gating. Stale/unsupported samples are not zeros; do not change power policy to make the test pass.

After the separately reviewed Stage 2 refresh, CONFIG_APPLE_PMP_REPORT and the bound apple-pmp-report driver should expose these static interface classes (not measured current clocks):

Board	type	data_bus_width	state_count	max_data_rate	peak_bandwidth	bandwidth_class
J616s/T6040	lpddr5	256	6	8533333333	273066666656	273000000000
Neo/T8140	lpddr5	64	5	7500000000	60000000000	60000000000

Rates are transfers/s/pin; bandwidth is decimal bytes/s. This PR does not deploy Stage 2 or modify either target.

[RyanTheTide]

fleet12 hardware result (7.3.0-aurora-fleet12, this PR merged in): on the Neo (G17P), agx_stats reads utilisation_status valid with util_renderer and util_device at 55% under glmark2. On the J616s (G16S), the target of this fix, it still reads valid but 0% renderer and device for the whole 30 s glmark2 terrain run, sampled every 5 s. That's the same as fleet11. No regression anywhere, but the G16S zero is not fixed, so I'm leaving this unmerged. Evidence is in the fleet12 rollout report on the Aurora host.

🤖 Generated with Claude Code