Re: [RFC PATCH 2/2] imx: mailbox: Introduce TX doorbell with ACK

From: Oleksij Rempel
Date: Tue Jun 11 2019 - 04:44:47 EST


On Tue, Jun 11, 2019 at 10:52:47AM +0300, Daniel Baluta wrote:
> Hi Oleksij,
>
> On Tue, Jun 11, 2019 at 8:56 AM Oleksij Rempel <o.rempel@xxxxxxxxxxxxxx> wrote:
> >
> > Hi Daniel,
> >
> > On Mon, Jun 10, 2019 at 10:16:09PM +0800, daniel.baluta@xxxxxxx wrote:
> > > From: Daniel Baluta <daniel.baluta@xxxxxxx>
> > >
> > > TX doorbell with ACK will allow us to push the doorbell ring button
> > > (trigger GIR) and also will allow us to handle the response from DSP.
> > >
> > > DSP firmware found on i.MX8 boards implements a duplex
> > > communication protocol over MU channels.
> > >
> > > On the host side (Linux) we need to plugin into Sound Open Firmware IPC
> > > communication infrastructure which handles all the details (e.g message
> > > queuing, tx/rx logic) [1] and the users are only required to provide the
> > > following callbacks:
> > >
> > > - send_msg (for Tx)
> > > - irq_handler (Ack of Tx, request from DSP)
> > >
> > > In order to implement send_msg and irq_handler we will use two MU
> > > channels:
> > > * channel #0, TX doorbell with ACK
> > > * channel #1, RX doorbell
> > >
> > > Sending a request Host -> DSP (channel #0)
> > > - send_msg callback
> > > - write data into SHMEM
> > > - push doorbell ring button (trigger GIR)
> > > - irq handler
> > > - handle DSP request (channel #1)
> > > - read SHMEM and trigger SOF IPC state machine
> > > - send ACK (push doorbell ring button for channel #1)
> > > - handle DSP response (ACK) (channel #0)
> > > - read SHMEM and trigger IPC state machine
> > >
> > > The easisest way to implement this is to directly access the MU
> > > registers but since the MU is abstracted using the mailbox interface
> > > we need to use that instead.
> > >
> > > [1] https://elixir.bootlin.com/linux/v5.2-rc4/source/sound/soc/sof/ipc.c
> > >
> > > Signed-off-by: Daniel Baluta <daniel.baluta@xxxxxxx>
> > > ---
> > > drivers/mailbox/imx-mailbox.c | 3 +++
> > > 1 file changed, 3 insertions(+)
> > >
> > > diff --git a/drivers/mailbox/imx-mailbox.c b/drivers/mailbox/imx-mailbox.c
> > > index 9f74dee1a58c..3a91611e17d2 100644
> > > --- a/drivers/mailbox/imx-mailbox.c
> > > +++ b/drivers/mailbox/imx-mailbox.c
> > > @@ -42,6 +42,7 @@ enum imx_mu_chan_type {
> > > IMX_MU_TYPE_RX, /* Rx */
> > > IMX_MU_TYPE_TXDB, /* Tx doorbell */
> > > IMX_MU_TYPE_RXDB, /* Rx doorbell */
> > > + IMX_MU_TYPE_TXDB_ACK /* Tx doorbell with Ack */
> > > };
> > >
> > > struct imx_mu_con_priv {
> > > @@ -124,6 +125,7 @@ static irqreturn_t imx_mu_isr(int irq, void *p)
> > > (ctrl & IMX_MU_xCR_RIEn(cp->idx));
> > > break;
> > > case IMX_MU_TYPE_RXDB:
> > > + case IMX_MU_TYPE_TXDB_ACK:
> > > val &= IMX_MU_xSR_GIPn(cp->idx) &
> > > (ctrl & IMX_MU_xCR_GIEn(cp->idx));
> > > break;
> > > @@ -200,6 +202,7 @@ static int imx_mu_startup(struct mbox_chan *chan)
> > > imx_mu_xcr_rmw(priv, IMX_MU_xCR_RIEn(cp->idx), 0);
> > > break;
> > > case IMX_MU_TYPE_RXDB:
> > > + case IMX_MU_TYPE_TXDB_ACK:
> > > imx_mu_xcr_rmw(priv, IMX_MU_xCR_GIEn(cp->idx), 0);
> > > break;
> > > default:
> > > --
> > > 2.17.1
> >
> > If I see it correctly, with your implementation the mbox client
> > communication on channel 0 will look as follow:
> > mbox_client -> send_msg()
> > /* sheduling of mbox_chan_txdone tasklet is avoided */
> > mbox_client <- cl->rx_callback()
> > mbox_client -> mbox_client_txdone()
> > mbox_client -> send_msg()
> >
>
> Correct! But what I thought it is most important here is that
> send_msg and rx_callback are called on the same "connection"
> so that inside rx_callback I can figure out if I got an ACK or I got
> a request from the other side.
>
> > Without your patch you will need to register tx and rx doorbell
> > channels and the communication will looks like this:
> > mbox_client -> send_msg()
> > mbox_client <- mbox_chan_txdone() /* dummy notification, can be ignored */
> > mbox_client <- cl->rx_callback()
> > mbox_client -> send_msg()
> >
> > I assume, you are trying to optimize it and avoid dummy
> > mbox_chan_txdone() notification. Correct?
>
> TBH I am little bit confused about the role of txdb_tasklet. Is it there
> just to trigger the sending of the next message in the buffer?

ACK. The TX type channel is acknowledged as soon as opposite side will
read out the TX/RX FIFO. In this case we will get an interrupt and
call mbox_chan_txdone(). With the TXDB type channels we have no this
kind of interrupts, so we need to some how to tell the mailbox framework
that this transmission is kind of done.

Beside, are there any reason why do you use TXDB instead of TX channel?
In this case you will get your ACK for free.

> Using IMX mailbox with SOF IPC mechanism has a lot of overhead because
> I think they are both doing the same thing.

Sorry, I fail to follow this argumentation. If it important, please
describe your issue.

>
> E.g: On TX queuing the messages and then when ACK arrives it will send the
> next message.
>
> >
> > The problem is, that current mailbox-framework will set txdone_method
> > inside of mbox_controller_register() for all channels even if
> > imx-mailbox has different types of channels.
> >
> > The problem with your patch is, that it will silently merge two channels
> > (TXDB and RXDB) and not setting actual ACK by controller - mbox_chan_txdone().
> > Not sure, why we need to merge it in this case.
> >
> > So, with current imx_mailbox implementation your firmware should work as
> > is. You will need to register two separate channels for TXDB and
> > RXDB. It will run with some overhead by triggering txdone tasklet in
> > imx-mailbox driver.
>
> I see your point. I think with the current imx_mailbox implementation SOF IPC
> should work as it is with some overhead (at least when Host (Linux) initiates
> the communication).
>
> The keypoint here (I figured it out now) is that the two channels
> TXDB and RXDB share the same IDX right?

ACK. Same idx with different types.

> So, with
> this idx we can figure out that we have an ACK from DSP or a request from DSP.

Actually you will get it no by idx, you will get it by proper device
tree configuration and assigned cl->rx_callback(). You assign rx_callback() per
each rx-able channel.

>
> So we are set for communication when Host (Linux) starts the communication. It
> should just work.
>
> How about the case when DSP initiates the communication and Host needs to
> send an ACK.
>
> In this case we have:
>
> mbox_client <- cl->rx_callback()
> mbox_client -> send_msg()

The call chain of rx_callback() is:
imx_mu_isr()
mbox_chan_received_data()
cl->rx_callback()

Since rx_callback() is in interrupt context, send_msg() should be non
blocking. You can use workqueue or some thing like this to mbox_send_message()
out of interrupt context.

> The problem here is that the Host needs to call send_msg() from inside
> the rx_callback().
> Do you see any problems with that:
>
> Basically the pseudocode will be something like this:
>
> Create connection #0 (RXDB0, TXDB0) Host -> DSP (Host initiates
> communication, DSP sends ACK)
> Create connection #1 (RXDB1, TXDB1) DSP -> Host (DSP initiates
> communication, Host sends ACK).
>
> rx_callback:
> -> if (chan_id == 0) // this has to be an ACK from DSP
> get_reply_from_shmem
> trigger SOF IPC state machine (let them know we got a reply)
> else if (chan_id == 1) // this has to be a request from DSP
> get_request()
> mbox_send_message(sc_chan->ch, NULL); //send back ACK
>

If properly implemented, you will have separate rx_callback()s for each
channel.

I assume at some later point you will need to extend the ACK to some
thing like:
- ACK - message is ok and processed
- BUSY - endpoint is still processing
- ERR - message is corrupt or wrong formatting
... and so on.

To be able to transmit different types of ACK you will probably
use TX/RX type channels + shmem instead of the doorbell.

>
> >
> > If this overhead is a problem, then this should be fixed.
>
> I need to have a working version with the current implementation
> and then tell if overhead is a problem.
>
> > Merging two doorbell channels in to one with ACK support is nice,
> > but will introduce more issues if we need other doorbell channels
> > without ACK support on same controller
> >
> > I personally would prefer to to extend mailbox framework to support
> > controllers with mixed channel types and remove dummy txdone tasklet
> > from imx-mailbox.
> >
> > Since we already initialize part of &mbox->chans[i] by imx-mailbox driver,
> > we can set proper chan->txdone_method as well. So we need only to
> > prevent mbox_controller_register() to overwrite it.
>
> Thanks for your explanations! Thinks are much clearer now!
>
> Will try to work on implementing what I need with the current IMX mailbox
> support and send patches for review.
>
> thanks,
> Daniel.
>
>

--
Pengutronix e.K. | |
Industrial Linux Solutions | http://www.pengutronix.de/ |
Peiner Str. 6-8, 31137 Hildesheim, Germany | Phone: +49-5121-206917-0 |
Amtsgericht Hildesheim, HRA 2686 | Fax: +49-5121-206917-5555 |