Re: [PATCH 1/2] Input: i8042 - Prevent intermixing i8042 commands

[Raul Rangel]

On Thu, Aug 27, 2020 at 2:12 PM Andy Shevchenko
<andy.shevchenko@xxxxxxxxx> wrote:
>
> On Thu, Aug 27, 2020 at 10:52 PM Raul E Rangel <rrangel@xxxxxxxxxxxx> wrote:
> >
> > The i8042_mutex must be held by writers of the AUX and KBD ports, as
> > well as users of i8042_command. There were a lot of users of
> > i8042_command that were not calling i8042_lock_chip/i8042_unlock_chip.
> > This resulted in i8042_commands being issues in between PS/2
> > transactions.
> >
> > This change moves the mutex lock into i8042_command and removes the
> > burden of locking the mutex from the callers.
>
> Which is wrong according to your very patch. See below.
>
> > It is expected that the i8042_mutex is locked before calling
> > i8042_aux_write or i8042_kbd_write. This is currently done by the PS/2
> > layer via ps2_begin_command and ps2_end_command. Other modules
> > (serio_raw) do not currently lock the mutex, so there is still a
> > possibility for intermixed commands.
>
> ...
>
> > +       mutex_lock(&i8042_mutex);
> > +
> >         spin_lock_irqsave(&i8042_lock, flags);
> >         retval = __i8042_command(param, command);
> >         spin_unlock_irqrestore(&i8042_lock, flags);
> >
> > +        mutex_unlock(&i8042_mutex);
>
> Question 1. Why do you need mutex at all in the above situation? Spin
> lock isn't enough?

No. PS/2 transactions/commands consist of multiple calls to ps2_do_sendbyte.
So the spin lock only helps with sending an individual byte. The mutex
is for the
whole transaction. We don't want i8042_commands being sent in between a PS/2
transaction.

>
> ...
>
> > -       i8042_lock_chip();
> > -
> >         if (value == LED_OFF)
> >                 i8042_command(NULL, CLEVO_MAIL_LED_OFF);
> >         else if (value <= LED_HALF)
> >                 i8042_command(NULL, CLEVO_MAIL_LED_BLINK_0_5HZ);
> >         else
> >                 i8042_command(NULL, CLEVO_MAIL_LED_BLINK_1HZ);
> > -
> > -       i8042_unlock_chip();
> > -
>
> Now, these three commands are not considered as a transaction (no
> atomicity). That's why your patch is wrong.

These are all mutually exclusive. So there is no change in behavior.
>
> >  }
>
> ...
>
> >         int rc;
> >
> > -       i8042_lock_chip();
> >         rc = i8042_command(&param, A1655_WIFI_COMMAND);
> > -       i8042_unlock_chip();
> >         return rc;
>
> rc become redundant.

Good catch. I'll send a v2 with it removed.

>
> --
> With Best Regards,
> Andy Shevchenko