Re: [PATCH] thermal: sysfs: take the zone lock in the tzp attribute handlers

From: Rafael J. Wysocki (Intel)

Date: Fri Nov 20 2026 - 11:51:24 EST


On Fri, Nov 13, 2026 at 5:55 AM Yu Zhang <yuz08559@xxxxxxxxx> wrote:
>
> The thermal zone parameter (tz->tzp) sysfs attributes - sustainable_power,
> k_po, k_pu, k_i, k_d, integral_cutoff, slope and offset - dereference
> tz->tzp without holding the zone lock:
>
> if (!tz->tzp)
> return -EIO;
> ...
> tz->tzp->name = value;
>
> For a zone registered without thermal zone parameters (tzp == NULL at
> registration - the thermal testing zones, ACPI thermal and every driver
> using thermal_zone_device_register_with_trips(..., NULL, ...)),
> power_allocator_bind() allocates tz->tzp and power_allocator_unbind()
> frees it and sets it back to NULL, both under the zone lock:
>
> tz->tzp = kzalloc(sizeof(*tz->tzp), GFP_KERNEL);
> ...
> params->allocated_tzp = true;
> /* unbind */
> if (params->allocated_tzp) {
> kfree(tz->tzp);
> tz->tzp = NULL;
> }
>
> bind/unbind run from policy_store() -> thermal_zone_device_set_policy()
> under the zone lock when the governor is switched to or away from
> power_allocator. The tzp attribute handlers take no lock, so a read or
> write of one of these files racing a governor switch dereferences a
> freed tz->tzp or a NULL pointer. The read side (0644) is reachable by
> an unprivileged user.
>
> KASAN, one test thermal zone, one loop switching the policy between
> power_allocator and step_wise and one loop writing k_pu, panics in
> seconds:
>
> BUG: KASAN: slab-use-after-free in k_pu_store+0xd7/0x130
> Write of size 4 at addr ffff8880064e5c94 by task init
> k_pu_store+0xd7/0x130
> dev_attr_store+0x46/0x70
> sysfs_kf_write+0xed/0x110
> kernfs_fop_write_iter+0x29f/0x380
> vfs_write+0x5df/0x950
> ksys_write+0xfd/0x1e0
> Allocated by task:
> power_allocator_bind+0x51a/0x600
> thermal_set_governor+0x86/0x170
> thermal_zone_device_set_policy+0x58/0xb0
> Freed by task:
> power_allocator_unbind+0xad/0xd0
> thermal_set_governor+0x57/0x170
> thermal_zone_device_set_policy+0x58/0xb0
>
> Take the zone lock in the tzp attribute handlers, as mode_show() and
> weight_store() already do ("Don't race with governors using the
> 'weight' value"). The governor already reads these fields under the
> zone lock, so this also serialises the PID constant updates against a
> running power_allocator.
>
> Fixes: f5cbb182586e ("thermal: power_allocator: don't require tzp to be present for the thermal zone")
> Cc: stable@xxxxxxxxxxxxxxx
> Signed-off-by: Yu Zhang <yuz08559@xxxxxxxxx>
> ---
> drivers/thermal/thermal_sysfs.c | 8 ++++++++
> 1 file changed, 8 insertions(+)
>
> diff --git a/drivers/thermal/thermal_sysfs.c b/drivers/thermal/thermal_sysfs.c
> --- a/drivers/thermal/thermal_sysfs.c
> +++ b/drivers/thermal/thermal_sysfs.c
> @@ -258,6 +258,8 @@ sustainable_power_show(struct device *de
> {
> struct thermal_zone_device *tz = to_thermal_zone(dev);
>
> + guard(thermal_zone)(tz);
> +
> if (tz->tzp)
> return sysfs_emit(buf, "%u\n", tz->tzp->sustainable_power);
> else
> @@ -271,6 +273,8 @@ sustainable_power_store(struct device *d
> struct thermal_zone_device *tz = to_thermal_zone(dev);
> u32 sustainable_power;
>
> + guard(thermal_zone)(tz);
> +
> if (!tz->tzp)
> return -EIO;
>
> @@ -289,6 +293,8 @@ sustainable_power_store(struct device *d
> { \
> struct thermal_zone_device *tz = to_thermal_zone(dev); \
> \
> + guard(thermal_zone)(tz); \
> + \
> if (!tz->tzp) \
> return -EIO; \
> \
> @@ -302,6 +308,8 @@ sustainable_power_store(struct device *d
> struct thermal_zone_device *tz = to_thermal_zone(dev); \
> s32 value; \
> \
> + guard(thermal_zone)(tz); \
> + \
> if (!tz->tzp) \
> return -EIO; \
> \

Applied as 7.5 material, thanks!