On 02/08/2019, 11:54, "Sudeep Holla" <sudeep.holla@xxxxxxx> wrote:
On Thu, Jul 25, 2019 at 12:49:58AM -0500, Jassi Brar wrote:
> On Sun, Jul 21, 2019 at 4:58 PM Jassi Brar <jassisinghbrar@xxxxxxxxx> wrote:
> >
[...]
> > If the mhuv2 instance implements, say, 3 channel windows between
> > sender (linux) and receiver (firmware), and Linux runs two protocols
> > each requiring 1 and 2-word sized messages respectively. The hardware
> > supports that by assigning windows [0] and [1,2] to each protocol.
> > However, I don't think the driver can support that. Or does it?
> >
> Thinking about it, IMO, the mbox-cell should carry a 128 (4x32) bit
> mask specifying the set of windows (corresponding to the bits set in
> the mask) associated with the channel.
> And the controller driver should see any channel as associated with
> variable number of windows 'N', where N is [0,124]
>
> mhu_client1: proto1@2e000000 {
> .....
> mboxes = <&mbox 0x0 0x0 0x0 0x1>
> }
>
> mhu_client2: proto2@2f000000 {
> .....
> mboxes = <&mbox 0x0 0x0 0x0 0x6>
> }
>
This still doesn't address the overhead I mentioned in my arm_mhu_v1
series.
As per you suggestion, we will have one channel with all possible
bit mask value to specify the window. Let's imagine that 2 protocols
share the same channel, then the requests are serialised.
E.g. if bits 0 and 1 are allocated for say protocol#1 and bits 2 and 3
for protocol#2.
At a given time only one protocol can be used by a client. No mix-match
of protocols are handled by the driver currently. Also its not possible to address all
possible scenarios offered by the IP. That's why the current driver design is
based on the implementation in the existing platforms.
Further protocol#1 has higher latency requirements like sched-governor
DVFS and there are 3-4 pending requests on protocol#2, then the incoming
requests for protocol#1 is blocked.
This is definitely overhead and I have seen lots of issue around this
and hence I was requesting that we need to create individual channels
for each of these. Having abstraction on top to multiplex or arbitrate
won't help.
Also the (mbox-cells) approach will not allow us to differentiate between
single-word and doorbell which is required to make the controller driver
aware of the data expected whether it's a pointer to a location or in
register itself.
--Tushar
--
Regards,
Sudeep
IMPORTANT NOTICE: The contents of this email and any attachments are confidential and may also be privileged. If you are not the intended recipient, please notify the sender immediately and do not disclose the contents to any other person, use it for any purpose, or store or copy the information in any medium. Thank you.
