Re: [PATCH 08/10] dt/bindings: Add binding for BCM2835 mailbox property channel driver

Stephen Warren [off-list ref] writes:

On 03/02/2015 01:54 PM, Eric Anholt wrote:

quoted

I was tempted to have the mailbox property channel support just be in
the 2835 mailbox driver itself, but mbox_request_channel() wants its
device to have the "mboxes" node, and that appears to be only intended
for mailbox clients, not controllers.

This is more of a particular format/protocol of messages you can send
over the mailbox HW than a device. I wonder if it actually makes sense
to represent it in DT as a device at all?

If we do represent this as a device in DT, shouldn't it also look like a
mailbox device so that other drivers (clock, display, ...) can bind to
it and send messages using the mailbox API?

I don't think it makes sense as a mailbox device.  mailbox devices can
only have one client per channel, while there's no reason to have that
limitation on the property channel.  You could imagine having the
individual tags be channels, except that again there's no reason to have
the one-client limitation, but more importantly the indivudual messages
to the firmware are composed of N tags.

I might have expected to just add property support directly into the
basic bcm2835 mailbox driver itself. Perhaps some attempt might be made
to isolate the HW register level access in one file/driver, and expose
both the regular and property mailbox protocols as a higher level that
uses the low-level mailbox functionality? The concept of the lower 4
bits of mailbox data being a channel ID might belong in the higher
protocol level rather than the lower HW layer?

Yes, the lower 4 bits being the channel is firmware protocol.

The reason I didn't have it in the same device as the mailbox is that
you need the channel reference in order to set up a mailbox client, and
I didn't think it made sense to have the mailbox device dt reference
itself.

The higher-level interface I think might make sense would be a send_data
replacement that took your device and mbox channel index, and a u32 of
data (low 4 bits unused), and synchronously returned a u32 of response.
It would do the client setup/send/wait_for_completion/teardown for you.

I'm skeptical of putting much more work into mailboxes on this platform,
though.  All we've got for channels are:

0: power (present in this series)
1: fb (Won't be used in Linux)
2: unused
3: VCHIQ (I've heard skepticism that the kernel community would accept this
          interface)
4: unused
5: unused
6: unused
7: settings (not sure if there's anything here not exposed in properties)
8: property (present in this series)
9-15: unused