Re: [PATCH v2] platform/x86/intel/tpmi: Clean up CPU mappings on offline

From: Ilpo Järvinen

Date: Mon Nov 16 2026 - 06:02:30 EST


On Sat, 14 Nov 2026, lirongqing wrote:

> From: Li RongQing <lirongqing@xxxxxxxxx>
>
> The driver adds each online CPU to tpmi_cpu_hash and to its power domain
> mask. Neither entry is removed when the CPU goes offline, so the next
> online transition calls hash_add() on an hlist node that is still
> linked, corrupting the hash list.
>
> Register an offline callback that removes the hash entry and clears the
> CPU from its power domain mask.
>
> The TPMI feature does not depend on CPU online/offline and can update the
> MMIO area while a CPU is offline. There is currently no consumer that
> reads the mapping data of an offlined CPU; if such a use case appears,
> retaining the offline mapping can be added on top.
>
> Fixes: 17ca2780458c ("platform/x86/intel: TPMI domain id and CPU mapping")
> Signed-off-by: Li RongQing <lirongqing@xxxxxxxxx>
> Acked-by: Srinivas Pandruvada <srinivas.pandruvada@xxxxxxxxxxxxxxx>
> ---
> Changes in v2:
>
> - Add Acked-by from Srinivas Pandruvada
> - Rename the title
> - Add a new line after guard()
>
> drivers/platform/x86/intel/tpmi_power_domains.c | 20 +++++++++++++++++++-
> 1 file changed, 19 insertions(+), 1 deletion(-)
>
> diff --git a/drivers/platform/x86/intel/tpmi_power_domains.c b/drivers/platform/x86/intel/tpmi_power_domains.c
> index ebe052c..457abac 100644
> --- a/drivers/platform/x86/intel/tpmi_power_domains.c
> +++ b/drivers/platform/x86/intel/tpmi_power_domains.c
> @@ -220,6 +220,24 @@ static int tpmi_cpu_online(unsigned int cpu)
> return 0;
> }
>
> +static int tpmi_cpu_offline(unsigned int cpu)
> +{
> + struct tpmi_cpu_info *info = &per_cpu(tpmi_cpu_info, cpu);
> + int index;
> +
> + if (hlist_unhashed(&info->hnode) || !tpmi_domain_is_valid(info))
> + return 0;
> +
> + index = info->pkg_id * MAX_POWER_DOMAINS + info->punit_domain_id;

Hi,

I've taken this patch into the review-ilpo-next branch.

...But could you make another patch on top of it which adds a helper for
this index calculation (from the tpmi_cpu_info struct, the one in
tpmi_get_linux_die_id() seems to not have it so it can be left as is).
It seems to be repeated a few times and is quite long, online func even
calculates it twice.

--
i.

> +
> + guard(mutex)(&tpmi_lock);
> +
> + cpumask_clear_cpu(cpu, &tpmi_power_domain_mask[index]);
> + hash_del(&info->hnode);
> +
> + return 0;
> +}
> +
> static int __init tpmi_init(void)
> {
> const struct x86_cpu_id *id;
> @@ -249,7 +267,7 @@ static int __init tpmi_init(void)
>
> ret = cpuhp_setup_state(CPUHP_AP_ONLINE_DYN,
> "platform/x86/tpmi_power_domains:online",
> - tpmi_cpu_online, NULL);
> + tpmi_cpu_online, tpmi_cpu_offline);
> if (ret < 0)
> goto free_domain_map;
>
>

--
i.