glymur: hard reset on the first system-domain idle entry (SS3, 0x0200c354) after a heavy load - Lenovo Yoga Slim 7x Gen 11
Hi,
Lenovo Yoga Slim 7x Gen 11 (Glymur / Snapdragon X2 Elite), linux-next next-20260908 with Konrad's board DTS (now in next) [1], OSI mode ("Initialized CPU PM domain topology using OSI mode"), Ulf's "pmdomain/cpuidle-psci: Fix behaviours for CPU PM domains" v3 applied.
Symptom: after any full-speed all-core load - a kernel build (~85 s here), stress --cpu 18 --timeout 90, rust-analyzer indexing - the machine hard-resets once it goes idle: about one second of freeze, then the firmware splash. Nothing is logged, the APSS watchdog's bootstatus reads 0 afterwards, no pstore. 5 out of 5 attempts, 15 s to 3 min after the load ends. Steady idle without a preceding load never resets, for hours.
What isolates it to SS3:
- Deleting domain-idle-states from power-domain-system in the board DTS (so 0x0200c354 is never requested; CL5 untouched) survives the same trigger 3 out of 3, through tens of thousands of cluster collapses.
- On a quiet machine SS3 is entered hundreds of times right after boot and survives every time. Only the first SS3 entry after a load kills. psci_domain_idle_enter filtered on state==0x200c354 shows no earlier request in the post-load window; the fatal one is the first.
- The state's latencies are not the variable: it resets with glymur.dtsi's entry 2800 / exit 4400 / residency 10150, with the vendor DSDT _LPI figures (entry 0 / exit 5000 / residency 9000, a local change I carry), and with entry 5000 / exit 5000 / residency 9000.
Ruled out by direct test, each with the same trigger:
- NoC QoS programming (reg removed from all 19 interconnect providers so icc-rpmh skips it): still resets.
- cpufreq transitions (all policies pinned with the performance governor): still resets.
- NVMe/PCIe I/O: stress --cpu 18 with no I/O resets too.
- PDC secondary mode: the PDC config register named in the vendor DSDT (\_SB_.GIO0.PDCC = 0x0b220110, mask PDCM = 0x35430) reads 0x4, i.e. 0 under the mask.
So the question: is SS3/CxPC expected to be usable as a runtime idle state on Glymur? x1e80100 only got domain_ss3 in DT once the PDC pass-through configuration landed (95f827ceb21e), while glymur.dtsi has carried it since the base dtsi. Is there a firmware or PDC prerequisite this board does not meet, or should glymur drop domain_ss3 from power-domain-system the way x1e did until then?
What I run meanwhile: min-residency-us = <4000000000> on domain_ss3. cpu_power_down_ok() then never admits SS3 at runtime, while the s2idle path (cpu_system_power_down_ok() checks latency only) still takes it. Three hours, a rebuild, three suspends and a stress cycle: zero runtime SS3 entries, s2idle entered SS3 every time, no reset, and s2idle draw is unchanged at ~300 mW. I am not proposing
that as a patch - a residency value is a hint, not a switch - but it does say SS3 is fine as a suspend state here and only fatal at runtime.
Happy to test patches or collect anything else; the reproducer takes three minutes.
Thanks,
Oleg
[1] https://lore.kernel.org/all/20260731-topic-yoga_submission-v2-0-f1887031da4f@oss.qualcomm.com/
Hi,
A data point from Kaanapali (SM8850), which has the same domain states (cluster 0x01000054, system 0x0200c354): Lenovo Legion Tab Y700 Gen 5, v7.3-rc4, PSCI OSI mode, with the CPU PM domains split in two cluster domains (CPU0-5, CPU6-7) under power-domain-system. With the single cluster domain of upstream kaanapali.dtsi the firmware rejects almost every domain state, so this does not show there. In short: we see silent resets from plain idle, and here they follow the cluster state rather than SS3.
Symptom: a silent reset after minutes to an hour of idle with the display off. Nothing in the printk ring or pstore dmesg; the console ramoops only has "watchdog: CPU7: Watchdog detected hard LOCKUP on cpu 0", and the watchdog bites before the hardlockup panic is printed.
Keeping SS3 out of runtime idle first seemed to help, but the resets came back with SS3 never entered at runtime.
An idle-entry recorder (per-CPU ring, records cleaned to PoC around the PSCI call), read from a RAM dump, shows the lost CPUs entering CPU retention (0x4) around a cluster 0 power-down (0x01000054) and never returning from that PSCI call, with IPIs and expired hrtimers pending. In one case the last CPU, the one requesting the cluster state, did not return either. Most cluster cycles in the same window were fine, so it looks like a race between the cluster state and a CPU entering retention.
Refusing the cluster domain state at runtime: 3 h idle without a reset (562k cluster-off requests refused). Same kernel with the state allowed:
2 resets in 78 min. Screen-off idle power did not change measurably.
Is the cluster state (and SS3) meant to be used from runtime idle on these SoCs, or is there a firmware prerequisite we miss?
Not tested on linux-next with Ulf's CPU PM domain series yet. I can run patches or collect dumps.
Thanks,
Joonhoe Kim
Hi Joonhoe,
Thanks, that is useful. On Glymur (Lenovo Yoga Slim 7x Gen 11) the pattern is different:
- Plain idle has never reset here: 50,000+ cluster (0x01000054) entries and hundreds of SS3 entries without a reset.
- The reset needs a trigger: the first SS3 entry right after a full all-core load (a kernel build or stress), 5 out of 5.
- With SS3 removed from the system domain and the cluster state kept, 3 out of 3 runs of that trigger survived.
- Day to day I run with a cpu_dma_latency constraint of 1500 us, which vetoes both the cluster state and SS3 at runtime, and I have not seen a runtime reset with it in place.
Glymur has a single CPU idle state, 0x4 retention, which is also the CPU domain state, so every cluster entry here has its CPUs in retention. I cannot separate the two the way you did.
A correction to my original report: the min-residency override did not hold. With no timer pending on any CPU the governor's idle estimate is effectively unbounded, so the residency check passes, and I had a reset with it in place. Only removing SS3 from the system domain survived.
Oleg