Message-ID: <CABb+yY3=yOKfRP=JEuBNG+ssx6rqjuVSYFCLOJJxvTEkAh7H2w@mail.gmail.com>
Date:   Fri, 28 Jul 2017 14:34:13 +0530
From:   Jassi Brar <jassisinghbrar@...il.com>
To:     Anup Patel <anup.patel@...adcom.com>
Cc:     Rob Herring <robh+dt@...nel.org>,
        Mark Rutland <mark.rutland@....com>,
        Catalin Marinas <catalin.marinas@....com>,
        Will Deacon <will.deacon@....com>,
        Florian Fainelli <f.fainelli@...il.com>,
        Scott Branden <sbranden@...adcom.com>,
        Ray Jui <rjui@...adcom.com>,
        Linux Kernel Mailing List <linux-kernel@...r.kernel.org>,
        "linux-arm-kernel@...ts.infradead.org" 
        <linux-arm-kernel@...ts.infradead.org>,
        Devicetree List <devicetree@...r.kernel.org>,
        BCM Kernel Feedback <bcm-kernel-feedback-list@...adcom.com>
Subject: Re: [PATCH v2 6/7] mailbox: bcm-flexrm-mailbox: Set msg_queue_len for
 each channel

On Fri, Jul 28, 2017 at 2:19 PM, Anup Patel <anup.patel@...adcom.com> wrote:
> On Thu, Jul 27, 2017 at 5:23 PM, Jassi Brar <jassisinghbrar@...il.com> wrote:
>> On Thu, Jul 27, 2017 at 11:20 AM, Anup Patel <anup.patel@...adcom.com> wrote:
>>> On Thu, Jul 27, 2017 at 10:29 AM, Jassi Brar <jassisinghbrar@...il.com> wrote:
>>
>>>>>>>>> Sorry for the delayed response...
>>>>>>>>>
>>>>>>>>> On Fri, Jul 21, 2017 at 9:16 PM, Jassi Brar <jassisinghbrar@...il.com> wrote:
>>>>>>>>>> Hi Anup,
>>>>>>>>>>
>>>>>>>>>> On Fri, Jul 21, 2017 at 12:25 PM, Anup Patel <anup.patel@...adcom.com> wrote:
>>>>>>>>>>> The Broadcom FlexRM ring (i.e. mailbox channel) can handle
>>>>>>>>>>> larger number of messages queued in one FlexRM ring hence
>>>>>>>>>>> this patch sets msg_queue_len for each mailbox channel to
>>>>>>>>>>> be same as RING_MAX_REQ_COUNT.
>>>>>>>>>>>
>>>>>>>>>>> Signed-off-by: Anup Patel <anup.patel@...adcom.com>
>>>>>>>>>>> Reviewed-by: Scott Branden <scott.branden@...adcom.com>
>>>>>>>>>>> ---
>>>>>>>>>>>  drivers/mailbox/bcm-flexrm-mailbox.c | 5 ++++-
>>>>>>>>>>>  1 file changed, 4 insertions(+), 1 deletion(-)
>>>>>>>>>>>
>>>>>>>>>>> diff --git a/drivers/mailbox/bcm-flexrm-mailbox.c b/drivers/mailbox/bcm-flexrm-mailbox.c
>>>>>>>>>>> index 9873818..20055a0 100644
>>>>>>>>>>> --- a/drivers/mailbox/bcm-flexrm-mailbox.c
>>>>>>>>>>> +++ b/drivers/mailbox/bcm-flexrm-mailbox.c
>>>>>>>>>>> @@ -1683,8 +1683,11 @@ static int flexrm_mbox_probe(struct platform_device *pdev)
>>>>>>>>>>>                 ret = -ENOMEM;
>>>>>>>>>>>                 goto fail_free_debugfs_root;
>>>>>>>>>>>         }
>>>>>>>>>>> -       for (index = 0; index < mbox->num_rings; index++)
>>>>>>>>>>> +       for (index = 0; index < mbox->num_rings; index++) {
>>>>>>>>>>> +               mbox->controller.chans[index].msg_queue_len =
>>>>>>>>>>> +                                               RING_MAX_REQ_COUNT;
>>>>>>>>>>>                 mbox->controller.chans[index].con_priv = &mbox->rings[index];
>>>>>>>>>>> +       }
>>>>>>>>>>>
>>>>>>>>>> While writing mailbox.c I wasn't unaware that there is the option to
>>>>>>>>>> choose the queue length at runtime.
>>>>>>>>>> The idea was to keep the code as simple as possible. I am open to
>>>>>>>>>> making it a runtime thing, but first, please help me understand how
>>>>>>>>>> that is useful here.
>>>>>>>>>>
>>>>>>>>>> I understand FlexRm has a ring buffer of RING_MAX_REQ_COUNT(1024)
>>>>>>>>>> elements. Any message submitted to mailbox api can be immediately
>>>>>>>>>> written onto the ringbuffer if there is some space.
>>>>>>>>>> Is there any mechanism to report back to a client driver, if its
>>>>>>>>>> message in ringbuffer failed "to be sent"?
>>>>>>>>>> If there isn't any, then I think, in flexrm_last_tx_done() you should
>>>>>>>>>> simply return true if there is some space left in the rung-buffer,
>>>>>>>>>> false otherwise.
>>>>>>>>>
>>>>>>>>> Yes, we have error code in "struct brcm_message" to report back
>>>>>>>>> errors from send_message. In our mailbox clients, we check
>>>>>>>>> return value of mbox_send_message() and also the error code
>>>>>>>>> in "struct brcm_message".
>>>>>>>>>
>>>>>>>> I meant after the message has been accepted in the ringbuffer but the
>>>>>>>> remote failed to receive it.
>>>>>>>
>>>>>>> Yes, even this case is handled.
>>>>>>>
>>>>>>> In case of IO errors after message has been put in ring buffer, we get
>>>>>>> completion message with error code and mailbox client drivers will
>>>>>>> receive back "struct brcm_message" with error set.
>>>>>>>
>>>>>>> You can refer flexrm_process_completions() for more details.
>>>>>>>
>>>> It doesn't seem to be what I suggest. I see two issues in
>>>> flexrm_process_completions()
>>>> 1) It calls mbox_send_message(), which is a big NO for a controller
>>>> driver. Why should you have one more message stored outside of
>>>> ringbuffer?
>>>
>>> The "last_pending_msg" in each FlexRM ring was added to fit FlexRM
>>> in Mailbox framework.
>>>
>>> We don't have any IRQ for TX done so "txdone_irq" out of the question for
>>> FlexRM. We only have completions for both success or failures (IO errors).
>>>
>>> This means we have to use "txdone_poll" for FlexRM. For "txdone_poll",
>>> we have to provide last_tx_done() callback. The last_tx_done() callback
>>> is supposed to return true if last send_data() call succeeded.
>>>
>>> To implement last_tx_done() in FlexRM driver, we added "last_pending_msg".
>>>
>>> When "last_pending_msg" is NULL it means last call to send_data() succeeded
>>> and when "last_pending_msg" is != NULL it means last call to send_data()
>>> did not go through due to lack of space in FlexRM ring.
>>>
>> It could be simpler.
>> Since flexrm_send_data() is essentially about putting the message in
>> the ring-buffer (and not about _transmission_ failures), the
>> last_tx_done() should simply return true if requests_ida has not all
>> ids allocated. False otherwise.
>
> It's not that simple because we have two cases in-which
> send_data() will fail:
> 1. It run-out of IDs in requests_ida
> 2. There is no room in BD queue of FlexRM ring. This because each
> brcm_message can be translated into variable number of descriptors.
> In fact, using SPU2 crypto client we have one brcm_message translating
> into 100's of descriptors. All-in-all few messages (< 1024) can also
> fill-up the BD queue of FlexRM ring.
>
OK let me put it abstractly... return false if "there is no space for
another message in the ringbuffer", true otherwise.


>>>>
>>>> 2) It calls mbox_chan_received_data()  which is for messages received
>>>> from the remote. And not the way to report failed _transmission_, for
>>>> which the api calls back mbox_client.tx_done() .  In your client
>>>> driver please populate mbox_client.tx_done() and see which message is
>>>> reported "sent fine" when.
>>>>
>>>>
>>>>>>>> There seems no such provision. IIANW, then you should be able to
>>>>>>>> consider every message as "sent successfully" once it is in the ring
>>>>>>>> buffer i.e, immediately after mbox_send_message() returns 0.
>>>>>>>> In that case I would think you don't need more than a couple of
>>>>>>>> entries out of MBOX_TX_QUEUE_LEN ?
>>>>>>>
>>>>>>> What I am trying to suggest is that we can take upto 1024 messages
>>>>>>> in a FlexRM ring but the MBOX_TX_QUEUE_LEN limits us queuing
>>>>>>> more messages. This issue manifest easily when multiple CPUs
>>>>>>> queues to same FlexRM ring (i.e. same mailbox channel).
>>>>>>>
>>>>>> OK then, I guess we have to make the queue length a runtime decision.
>>>>>
>>>>> Do you agree with approach taken by PATCH5 and PATCH6 to
>>>>> make queue length runtime?
>>>>>
>>>> I agree that we may have to get the queue length from platform, if
>>>> MBOX_TX_QUEUE_LEN is limiting performance. That will be easier on both
>>>> of us. However I suspect the right fix for _this_ situation is in
>>>> flexrm driver. See above.
>>>
>>> The current implementation is trying to model FlexRM using "txdone_poll"
>>> method and that's why we have dependency on MBOX_TX_QUEUE_LEN
>>>
>>> I think what we really need is new method for "txdone" to model ring
>>> manager HW (such as FlexRM). Let's call it "txdone_none".
>>>
>>> For "txdone_none", it means there is no "txdone" reporting in HW
>>> and mbox_send_data() should simply return value returned by
>>> send_data() callback. The last_tx_done() callback is not required
>>> for "txdone_none" and MBOX_TX_QUEUE_LEN also has no
>>> effect on "txdone_none". Both blocking and non-blocking clients
>>> are treated same for "txdone_none".
>>>
>> That is already supported :)
>
> If you are referring to "txdone_ack" then this cannot be used here
> because for "txdone_ack" we have to call mbox_chan_txdon() API
> after writing descriptors in send_data() callback which will cause
> dead-lock in tx_tick() called by mbox_chan_txdone().
>
Did you read my code snippet below?

It's not mbox_chan_txdone(), but mbox_client_txdone() which is called
by the client.

>>
>> In drivers/dma/bcm-sba-raid.c
>>
>> sba_send_mbox_request(...)
>> {
>>            ......
>>         req->msg.error = 0;
>>         ret = mbox_send_message(sba->mchans[mchans_idx], &req->msg);
>>         if (ret < 0) {
>>                 dev_err(sba->dev, "send message failed with error %d", ret);
>>                 return ret;
>>         }
>>         ret = req->msg.error;
>>         if (ret < 0) {
>>                 dev_err(sba->dev, "message error %d", ret);
>>                 return ret;
>>         }
>>           .....
>> }
>>
>> Here you _do_ assume that as soon as the mbox_send_message() returns,
>> the last_tx_done() is true. In other words, this is a case of client
>> 'knows_txdone'.
>>
>> So ideally you should specify cl->knows_txdone = true during
>> mbox_request_channel() and have ...
>>
>> sba_send_mbox_request(...)
>> {
>>         ret = mbox_send_message(sba->mchans[mchans_idx], &req->msg);
>>         if (ret < 0) {
>>                 dev_err(sba->dev, "send message failed with error %d", ret);
>>                 return ret;
>>         }
>>
>>         ret = req->msg.error;
>>
>>        /* Message successfully placed in the ringbuffer, i.e, done */
>>        mbox_client_txdone(sba->mchans[mchans_idx], ret);
>>
>>        if (ret < 0) {
>>                 dev_err(sba->dev, "message error %d", ret);
>>                 return ret;
>>         }
>>
>>         .....
>> }
>>
>
> I think we need to improve mailbox.c so that
> mbox_chan_txdone() can be called from
> send_data() callback.
>
No please. Other clients call mbox_send_message() followed by
mbox_client_txdone(), and they are right. For example,
drivers/firmware/tegra/bpmp.c

Thanks.

Powered by blists - more mailing lists