Re: [PATCH 4/4] leds: add driver for synology atmega1608 controlled LEDs

["Alexandre Courbot" <acourbot@nvidia.com>]

On Thu Oct 9, 2025 at 9:41 PM JST, Danilo Krummrich wrote:
> (Cc: Alex)
>
> On Thu Oct 9, 2025 at 2:30 PM CEST, Markus Probst wrote:
>> On Thu, 2025-10-09 at 14:20 +0200, Danilo Krummrich wrote:
>>> (Not a full review (let's work out the dependencies first), but
>>> there's one
>>> thing that stood out to me.)
>>> 
>>> On Wed Oct 8, 2025 at 8:10 PM CEST, Markus Probst wrote:
>>> > +struct Atmega1608Led {
>>> > +    addr: Atmega1608LedAddress,
>>> > +    id: Atmega1608LedId,
>>> > +
>>> > +    client: ARef<I2cClient>,
>>> > +
>>> > +    mode_lock: Arc<Mutex<()>>,
>>> 
>>> Mutex<()> raises an eyebrow, since a mutex that doesn't protect
>>> anything is
>>> pointless. So, I assume it is protecting some data, but in an unsound
>>> way.
>>> 
>>> > +impl Atmega1608Led {
>>> > +    fn update_mode(&self, mode: Atmega1608LedMode) ->
>>> > Result<Atmega1608LedMode> {
>>> > +        let _guard = self.mode_lock.lock();
>>> 
>>> What exactly does the mutex protect in the code below?
>> Otherwise there would be a race condition. Each register has 8 bits,
>> each led has 2 bits. If the led mode is updated at the same time with
>> another one in the same register, it could lead to the first action
>> being overwritten by the second.
>> Meaning if two actions run at the same time:
>> - led0 reads from the register
>> - led1 reads from the register
>> - led0 writes to the register
>> - led1 writes to the register (the changes for led0 have been
>> overwritten here, as it did read the register before led0 has written
>> to it)
>
> Ok, so you need exclusive access to a register. I think this is something the
> register abstraction I also mentioned in [1] could support.
>
> @Alex: Have you thought about this already?
>
> [1] https://lore.kernel.org/rust-for-linux/DDDS2V0V2NVJ.16ZKXCKUA1HUV@kernel.org/

I haven't at all. :) And indeed intuitively I couldn't say what would be
the right level to provide such concurrent access protection. When
someone reads from a register, there is no guarantee that they did so
with the intent of writing an updated value back to it, so I don't think we can e.g. hold a
guard for as long as the read value exists.

Even the `alter` accessor does not take any lock for the duration of its
closure, so I don't think it protects us. OTOH, `led0` would hold the IO
resource for as long as `alter` runs, which should prevent `led1` from
accessing it? If so, I'd say that's the correct way of handling such
cases.

(drive-by comment: I will probably rename `alter` into `update` to match
the regmap API once the dust settles with all the code moves related to
register!())