Re: [PATCH] cpufreq/amd-pstate: Don't fail active mode init on read-only auto_sel

[Mario Limonciello]

On 10/7/26 23:07, Akram Boulahia wrote:
> Since commit 9dfd13f80c85 ("cpufreq/amd-pstate: Toggle auto_sel in active
> mode on shared memory systems"), shmem_init_perf() no longer returns
> early in active mode and always calls cppc_set_auto_sel(), which was
> needed for shared memory systems. However, on firmware where the
> Autonomous Selection Enable entry of _CPC is a constant integer rather
> than a register, that write fails with -EOPNOTSUPP. The error is
> returned, every CPU fails to initialize, cpufreq_register_driver() finds
> no policy and returns -ENODEV, and the kernel falls back to
> acpi-cpufreq:
>
>    amd_pstate: failed to set auto_sel, ret: -95
>    amd_pstate: Failed to initialize CPU 0: -95
>    ...
>    amd_pstate: failed to register with return -19
>
> This was seen on an ASUS ROG Zephyrus G14 (GA403UV, Ryzen 9 8945HS, BIOS
> 308). Its _CPC tables are revision 3 with the entry fixed at 1, and
> cppc_get_auto_sel() reads back 1 on all CPUs.
>
> If the write returns -EOPNOTSUPP while active mode is requested and
> auto_sel already reads as enabled, which is what active mode needs, the
> refused write is harmless and initialization can continue. Any other
> failure, including a refusal where the current value does not match the
> requested mode, is still returned as before.
>
> With this change, amd-pstate-epp loads on all 16 CPUs of that machine
> without warnings. amd_pstate=passive behaves as before: it still falls
> back to acpi-cpufreq with the same -95 errors, because auto_sel stays at 1
> and cannot be changed.
>
> Tested on Linux 7.3.0-rc6-00037-g7b63ef2d55f2.
>
> Fixes: 9dfd13f80c85 ("cpufreq/amd-pstate: Toggle auto_sel in active mode on shared memory systems")
> Cc: stable@xxxxxxxxxxxxxxx
> Cc: Marco Scardovi <scardracs@xxxxxxxxxxx>
> Signed-off-by: Akram Boulahia <akram02st@xxxxxxxxx>
> ---

You're in a really weird path if you have an 8945HS that is using shared 
memory.  That system has MSR access.

IE this sounds like a severe BIOS bug.  Let's see what happens after you 
update BIOS.  If it's still failing please share an acpidump somewhere 
(like a kernel bugzilla).

Need to make sure we understand why your system is behaving this way 
before making code changes.

>   drivers/cpufreq/amd-pstate.c | 14 ++++++++++++--
>   1 file changed, 12 insertions(+), 2 deletions(-)
>
> diff --git a/drivers/cpufreq/amd-pstate.c b/drivers/cpufreq/amd-pstate.c
> index 8bfd46d60843..05b496946d13 100644
> --- a/drivers/cpufreq/amd-pstate.c
> +++ b/drivers/cpufreq/amd-pstate.c
> @@ -500,6 +500,7 @@ static int shmem_init_perf(struct amd_cpudata *cpudata)
>   	union perf_cached perf = READ_ONCE(cpudata->perf);
>   	u64 numerator;
>   	bool auto_sel;
> +	bool want;
>   
>   	int ret = cppc_get_perf_caps(cpudata->cpu, &cppc_perf);
>   	if (ret)
> @@ -518,14 +519,23 @@ static int shmem_init_perf(struct amd_cpudata *cpudata)
>   	WRITE_ONCE(cpudata->perf, perf);
>   	WRITE_ONCE(cpudata->prefcore_ranking, cppc_perf.highest_perf);
>   
> +	want = (cppc_state != AMD_PSTATE_PASSIVE);
> +
>   	ret = cppc_get_auto_sel(cpudata->cpu, &auto_sel);
>   	if (ret) {
>   		pr_warn("failed to get auto_sel, ret: %d\n", ret);
>   		return 0;
>   	}
>   
> -	ret = cppc_set_auto_sel(cpudata->cpu,
> -			(cppc_state == AMD_PSTATE_PASSIVE) ? 0 : 1);
> +	ret = cppc_set_auto_sel(cpudata->cpu, want);
> +
> +	/*
> +	 * Firmware may expose Autonomous Selection Enable as a read-only
> +	 * constant. If it already matches what active mode needs, the
> +	 * failed write is harmless.
> +	 */
> +	if (ret == -EOPNOTSUPP && want && auto_sel)
> +		return 0;
>   
>   	if (ret)
>   		pr_warn("failed to set auto_sel, ret: %d\n", ret);