Re: [PATCH v4 1/3] hwmon: (pmbus) add spinlock to protect 64-bit timestamp

[Guenter Roeck]

On 3/19/26 16:50, Pradhan, Sanman wrote:
> From: Sanman Pradhan <psanman@xxxxxxxxxxx>
>
> The next_access_backoff variable is a 64-bit ktime_t. On 32-bit
> architectures, accesses to 64-bit variables are not atomic, which can
> result in "torn" reads or writes if accessed concurrently from
> different subsystems (e.g., hwmon sysfs reads vs. regulator updates).
>
> Introduce a spinlock in struct pmbus_data to protect accesses to
> next_access_backoff and prevent torn reads/writes on 32-bit systems.
> This change addresses atomicity of the timestamp field itself as the
> timing helpers become callable from more paths, including PMBus chip
> drivers using the exported helpers.
>
> This does not attempt to serialize the full wait/transfer/update
> sequence across concurrent callers; it only protects the timestamp
> field from torn access.
>
> Signed-off-by: Sanman Pradhan <psanman@xxxxxxxxxxx>
> ---
> v4:
> - New patch in the series. Introduces a spinlock to protect the 64-bit
>    next_access_backoff timestamp from torn reads/writes on 32-bit
>    architectures.

Unfortunately that makes timestamp handling very expensive, even for
architectures not needing it.

We'll have to make sure that all accesses (sysfs, thermal, and regulator)
are serialized instead. I just browsed through the core code. Most of
the accesses are already serialized, but I think the locking is missing
in the regulator operations. We'll need to add the missing locks there.

Thanks,
Guenter