Using the mailbox subsystem for plain doorbells?

Using the mailbox subsystem for plain doorbells?

[Stephen Warren]

Jassi,

Does the HW described below sound like something that should be 
represented using the Linux kernel's mailbox subsystem, and related DT 
bindings? I think the existing drivers/mailbox/pcc.c is similar, but 
wanted to double-check.

We have some HW that literally just allows a SW-generated interrupt to 
be generated by our main CPU complex to an auxiliary CPU, and likewise a 
different interrupt can be generated in the opposite direction. There's 
no ability to transfer any data; just an IRQ is generated. Our current 
mailbox implementation just handles IRQ generation/reception so struct 
mbox_chan_ops.send_data completely ignores the data parameter, and our 
IRQ handler "receives" hard-coded NULL messages when the IRQ fires. 
Higher level protocol code (using shared memory along with the plain-IRQ 
mbox channels) is outside the mailbox driver.

Does that fit the mailbox subsystem?

Re: Using the mailbox subsystem for plain doorbells?

[Jassi Brar]

On Tue, May 10, 2016 at 5:15 AM, Stephen Warren <swarren@wwwdotorg.org> wrote:
> Jassi,
>
> Does the HW described below sound like something that should be represented
> using the Linux kernel's mailbox subsystem, and related DT bindings? I think
> the existing drivers/mailbox/pcc.c is similar, but wanted to double-check.
>
> We have some HW that literally just allows a SW-generated interrupt to be
> generated by our main CPU complex to an auxiliary CPU, and likewise a
> different interrupt can be generated in the opposite direction. There's no
> ability to transfer any data; just an IRQ is generated. Our current mailbox
> implementation just handles IRQ generation/reception so struct
> mbox_chan_ops.send_data completely ignores the data parameter, and our IRQ
> handler "receives" hard-coded NULL messages when the IRQ fires. Higher level
> protocol code (using shared memory along with the plain-IRQ mbox channels)
> is outside the mailbox driver.
>
> Does that fit the mailbox subsystem?
>
>From the sound of it, yes.

Some controllers need a mask/list of destination cpus, to which the
irq is raised, written to some 'data' register. You too probably need
to program the destination "id" in the controller? Maybe that should
be done in send_data().

Re: Using the mailbox subsystem for plain doorbells?

[Stephen Warren]

On 05/09/2016 09:29 PM, Jassi Brar wrote:
> On Tue, May 10, 2016 at 5:15 AM, Stephen Warren <swarren@wwwdotorg.org> wrote:
>> Jassi,
>>
>> Does the HW described below sound like something that should be represented
>> using the Linux kernel's mailbox subsystem, and related DT bindings? I think
>> the existing drivers/mailbox/pcc.c is similar, but wanted to double-check.
>>
>> We have some HW that literally just allows a SW-generated interrupt to be
>> generated by our main CPU complex to an auxiliary CPU, and likewise a
>> different interrupt can be generated in the opposite direction. There's no
>> ability to transfer any data; just an IRQ is generated. Our current mailbox
>> implementation just handles IRQ generation/reception so struct
>> mbox_chan_ops.send_data completely ignores the data parameter, and our IRQ
>> handler "receives" hard-coded NULL messages when the IRQ fires. Higher level
>> protocol code (using shared memory along with the plain-IRQ mbox channels)
>> is outside the mailbox driver.
>>
>> Does that fit the mailbox subsystem?
>>
>  From the sound of it, yes.

Good to know, thanks.

> Some controllers need a mask/list of destination cpus, to which the
> irq is raised, written to some 'data' register. You too probably need
> to program the destination "id" in the controller? Maybe that should
> be done in send_data().

In this case, each mailbox communicates with a different remote CPU, and 
there's a separate register to communicate with each remote CPU. So, 
send_data() completely ignores the data parameter since everything is 
derived from the mailbox's identity.

Re: Using the mailbox subsystem for plain doorbells?

[Jassi Brar]

On Tue, May 10, 2016 at 9:26 PM, Stephen Warren <swarren@wwwdotorg.org> wrote:
> On 05/09/2016 09:29 PM, Jassi Brar wrote:
>
>> Some controllers need a mask/list of destination cpus, to which the
>> irq is raised, written to some 'data' register. You too probably need
>> to program the destination "id" in the controller? Maybe that should
>> be done in send_data().
>
>
> In this case, each mailbox communicates with a different remote CPU, and
> there's a separate register to communicate with each remote CPU. So,
> send_data() completely ignores the data parameter since everything is
> derived from the mailbox's identity.
>
Yes, that is ok. For example, mailbox-sti.c does that.

Cheers!