Re: [PATCH v3 3/3] leds: tlc5925: Add support for non blocking operations

[Jean-Jacques Hiblot]

On 09/06/2022 18:43, Andy Shevchenko wrote:
> On Thu, Jun 9, 2022 at 6:29 PM Jean-Jacques Hiblot
> <jjhiblot@xxxxxxxxxxxxxxx> wrote:
> > Settings multiple LEDs in a row can be a slow operation because of the
> > time required to acquire the bus and prepare the transfer.
> > And, in most cases, it is not required that the operation is synchronous.
> >
> > Implementing the non-blocking brightness_set() for such cases.
> > A work queue is used to perform the actual SPI transfer.
> >
> > The blocking method is still available in case someone needs to perform
> > this operation synchronously (ie by calling led_set_brightness_sync()).
> i.e.
>
> > +#define BITS_PER_ATOMIC (sizeof(atomic_t) * 8)
> We have BITS_PER_TYPE(). Use it directly in the code, no need for a
> whole new macro.
>
> ...
>
> > +static int xmit(struct tlc5925_leds_priv *priv)
> > +{
> > +       int i;
> > +
> > +       spin_lock(&priv->lock);
> This can't be called during IRQ?
>
> > +       for (i = 0; i < priv->max_state / (sizeof(atomic_t) * 8) ; i++)
> BITS_PER_TYPE() ?
>
> > +               priv->spi_buffer[i] = atomic_read(&priv->state[i]);
> > +       spin_unlock(&priv->lock);
> > +
> > +       return spi_write(priv->spi, priv->spi_buffer, priv->max_num_leds / 8);
> > +}
> ...
>
> > +static void xmit_work(struct work_struct *ws)
> > +{
> > +       struct tlc5925_leds_priv *priv =
> > +               container_of(ws, struct tlc5925_leds_priv, xmit_work);
> One line?
>
> Missed blank line here.
>
> > +       xmit(priv);
> > +};
> ...
>
> >          if (brightness)
> > -               priv->state[index / 8] |= (1 << (index % 8));
> > +               atomic_or(1 << (index % BITS_PER_ATOMIC),
> > +                         &priv->state[index / BITS_PER_ATOMIC]);
> >          else
> > -               priv->state[index / 8] &= ~(1 << (index % 8));
> > -       spin_unlock(&priv->lock);
> > +               atomic_and(~(1 << (index % BITS_PER_ATOMIC)),
> > +                          &priv->state[index / BITS_PER_ATOMIC]);
> The whole bunch looks like reinventing the bitmap / bitops.
> Use unsigned long (or DECLARE_BITMAP() if it can be higher than 32)
> for state and set_bit() / clear_bit() / assign_bit() that are atomic.

Thanks for pointing this out.

It will drastically simplify the code.

> ...
>
> > +       if (brightness)
> > +               atomic_or(1 << (index % BITS_PER_ATOMIC),
> > +                         &priv->state[index / BITS_PER_ATOMIC]);
> > +       else
> > +               atomic_and(~(1 << (index % BITS_PER_ATOMIC)),
> > +                          &priv->state[index / BITS_PER_ATOMIC]);
> assign_bit()
>
> ...
>
> > +       // Allocate the buffer used to hold the state of each LED
> > +       priv->max_state = round_up(max_num_leds, BITS_PER_ATOMIC);
> > +       priv->state = devm_kzalloc(dev,
> > +                                  priv->max_state / 8,
> > +                                  GFP_KERNEL);
> >          if (!priv->state)
> >                  return -ENOMEM;
> devm_bitmap_zalloc() ?
>
> ...
>
> > +       // Allocate a second buffer for the communication on the SPI bus
> > +       priv->spi_buffer = devm_kzalloc(dev,
> > +                                  priv->max_state / 8,
> > +                                  GFP_KERNEL);
> Not sure I understand the output, but perhaps here the BITS_TO_BYTES()
> should be used.