Groups | Search | Server Info | Keyboard shortcuts | Login | Register [http] [https] [nntp] [nntps]


Groups > linux.kernel > #1571271 > unrolled thread

Re: [PATCH v20 2/4] mailbox: mediatek: Add Mediatek CMDQ driver

Started byJassi Brar <jassisinghbrar@gmail.com>
First post2017-02-01 06:30 +0100
Last post2017-02-09 13:10 +0100
Articles 3 — 2 participants

Back to article view | Back to linux.kernel

This discussion starts older than the indexed window; earlier articles aren't shown. The article labeled Started by below is the oldest one visible, not the original post.


Contents

  Re: [PATCH v20 2/4] mailbox: mediatek: Add Mediatek CMDQ driver Jassi Brar <jassisinghbrar@gmail.com> - 2017-02-01 06:30 +0100
    Re: [PATCH v20 2/4] mailbox: mediatek: Add Mediatek CMDQ driver Horng-Shyang Liao <hs.liao@mediatek.com> - 2017-02-06 06:40 +0100
      Re: [PATCH v20 2/4] mailbox: mediatek: Add Mediatek CMDQ driver Horng-Shyang Liao <hs.liao@mediatek.com> - 2017-02-09 13:10 +0100

#1571271 — Re: [PATCH v20 2/4] mailbox: mediatek: Add Mediatek CMDQ driver

FromJassi Brar <jassisinghbrar@gmail.com>
Date2017-02-01 06:30 +0100
SubjectRe: [PATCH v20 2/4] mailbox: mediatek: Add Mediatek CMDQ driver
Message-ID<t5VEt-1L5-7@gated-at.bofh.it>
On Thu, Jan 26, 2017 at 2:07 PM, Horng-Shyang Liao <hs.liao@mediatek.com> wrote:
> Hi Jassi,
>
> On Thu, 2017-01-26 at 10:08 +0530, Jassi Brar wrote:
>> On Wed, Jan 4, 2017 at 8:36 AM, HS Liao <hs.liao@mediatek.com> wrote:
>>
>> > diff --git a/drivers/mailbox/mtk-cmdq-mailbox.c b/drivers/mailbox/mtk-cmdq-mailbox.c
>> > new file mode 100644
>> > index 0000000..747bcd3
>> > --- /dev/null
>> > +++ b/drivers/mailbox/mtk-cmdq-mailbox.c
>>
>> ...
>>
>> > +static void cmdq_task_exec(struct cmdq_pkt *pkt, struct cmdq_thread *thread)
>> > +{
>> > +       struct cmdq *cmdq;
>> > +       struct cmdq_task *task;
>> > +       unsigned long curr_pa, end_pa;
>> > +
>> > +       cmdq = dev_get_drvdata(thread->chan->mbox->dev);
>> > +
>> > +       /* Client should not flush new tasks if suspended. */
>> > +       WARN_ON(cmdq->suspended);
>> > +
>> > +       task = kzalloc(sizeof(*task), GFP_ATOMIC);
>> > +       task->cmdq = cmdq;
>> > +       INIT_LIST_HEAD(&task->list_entry);
>> > +       task->pa_base = dma_map_single(cmdq->mbox.dev, pkt->va_base,
>> > +                                      pkt->cmd_buf_size, DMA_TO_DEVICE);
>> >
>> You seem to parse the requests and responses, that should ideally be
>> done in client driver.
>> Also, we are here in atomic context, can you move it in client driver
>> (before the spin_lock)?
>> Maybe by adding a new 'pa_base' member as well in 'cmdq_pkt'.
>
> will do
>
>> ....
>> > +
>> > +       cmdq->mbox.num_chans = CMDQ_THR_MAX_COUNT;
>> > +       cmdq->mbox.ops = &cmdq_mbox_chan_ops;
>> > +       cmdq->mbox.of_xlate = cmdq_xlate;
>> > +
>> > +       /* make use of TXDONE_BY_ACK */
>> > +       cmdq->mbox.txdone_irq = false;
>> > +       cmdq->mbox.txdone_poll = false;
>> > +
>> > +       for (i = 0; i < ARRAY_SIZE(cmdq->thread); i++) {
>> >
>> You mean  i < CMDQ_THR_MAX_COUNT
>
> will do
>
>> > +               cmdq->thread[i].base = cmdq->base + CMDQ_THR_BASE +
>> > +                               CMDQ_THR_SIZE * i;
>> > +               INIT_LIST_HEAD(&cmdq->thread[i].task_busy_list);
>> >
>> You seem the queue mailbox requests in this controller driver? why not
>> use the mailbox api for that?
>>
>> > +               init_timer(&cmdq->thread[i].timeout);
>> > +               cmdq->thread[i].timeout.function = cmdq_thread_handle_timeout;
>> > +               cmdq->thread[i].timeout.data = (unsigned long)&cmdq->thread[i];
>> >
>> Here again... you seem to ignore the polling mechanism provided by the
>> mailbox api, and implement your own.
>
> The queue is used to record the tasks which are flushed into CMDQ
> hardware (GCE). We are handling time critical tasks, so we have to
> queue them in GCE rather than a software queue (e.g. mailbox buffer).
> Let me use display as an example. Many display tasks are flushed into
> CMDQ to wait next vsync event. When vsync event is triggered by display
> hardware, GCE needs to process all flushed tasks "within vblank" to
> prevent garbage on screen. This is all done by GCE (without CPU)
> to fulfill time critical requirement. After GCE finish its work,
> it will generate interrupts, and then CMDQ driver will let clients know
> which tasks are done.
>
Does the GCE provide any 'lock' to prevent modifying (by adding tasks
to) the GCE h/w buffer when it is processing it at vsync?  Otherwise
there maybe race/error. If there is such a 'lock' flag/irq, that could
help here. However, you are supposed to know your h/w better, so I
will accept this implementation assuming it can't be done any better.

Please address other comments and resubmit.

Thanks

[toc] | [next] | [standalone]


#1574367

FromHorng-Shyang Liao <hs.liao@mediatek.com>
Date2017-02-06 06:40 +0100
Message-ID<t7KbT-H2-5@gated-at.bofh.it>
In reply to#1571271
Hi Jassi,

On Wed, 2017-02-01 at 10:52 +0530, Jassi Brar wrote:
> On Thu, Jan 26, 2017 at 2:07 PM, Horng-Shyang Liao <hs.liao@mediatek.com> wrote:
> > Hi Jassi,
> >
> > On Thu, 2017-01-26 at 10:08 +0530, Jassi Brar wrote:
> >> On Wed, Jan 4, 2017 at 8:36 AM, HS Liao <hs.liao@mediatek.com> wrote:
> >>
> >> > diff --git a/drivers/mailbox/mtk-cmdq-mailbox.c b/drivers/mailbox/mtk-cmdq-mailbox.c
> >> > new file mode 100644
> >> > index 0000000..747bcd3
> >> > --- /dev/null
> >> > +++ b/drivers/mailbox/mtk-cmdq-mailbox.c
> >>
> >> ...
> >>
> >> > +static void cmdq_task_exec(struct cmdq_pkt *pkt, struct cmdq_thread *thread)
> >> > +{
> >> > +       struct cmdq *cmdq;
> >> > +       struct cmdq_task *task;
> >> > +       unsigned long curr_pa, end_pa;
> >> > +
> >> > +       cmdq = dev_get_drvdata(thread->chan->mbox->dev);
> >> > +
> >> > +       /* Client should not flush new tasks if suspended. */
> >> > +       WARN_ON(cmdq->suspended);
> >> > +
> >> > +       task = kzalloc(sizeof(*task), GFP_ATOMIC);
> >> > +       task->cmdq = cmdq;
> >> > +       INIT_LIST_HEAD(&task->list_entry);
> >> > +       task->pa_base = dma_map_single(cmdq->mbox.dev, pkt->va_base,
> >> > +                                      pkt->cmd_buf_size, DMA_TO_DEVICE);
> >> >
> >> You seem to parse the requests and responses, that should ideally be
> >> done in client driver.
> >> Also, we are here in atomic context, can you move it in client driver
> >> (before the spin_lock)?
> >> Maybe by adding a new 'pa_base' member as well in 'cmdq_pkt'.
> >
> > will do

I agree with moving dma_map_single out from spin_lock.

However, mailbox clients cannot map virtual memory to mailbox
controller's device for DMA. In our previous discussion, we decided to
remove mailbox_controller.h from clients to restrict their capabilities.

Please take a look at following link from 2016/9/22 to 2016/9/30 about
mailbox_controller.h.
https://patchwork.kernel.org/patch/9312953/

Is there any better place to do dma_map_single?

> >> ....
> >> > +
> >> > +       cmdq->mbox.num_chans = CMDQ_THR_MAX_COUNT;
> >> > +       cmdq->mbox.ops = &cmdq_mbox_chan_ops;
> >> > +       cmdq->mbox.of_xlate = cmdq_xlate;
> >> > +
> >> > +       /* make use of TXDONE_BY_ACK */
> >> > +       cmdq->mbox.txdone_irq = false;
> >> > +       cmdq->mbox.txdone_poll = false;
> >> > +
> >> > +       for (i = 0; i < ARRAY_SIZE(cmdq->thread); i++) {
> >> >
> >> You mean  i < CMDQ_THR_MAX_COUNT
> >
> > will do
> >
> >> > +               cmdq->thread[i].base = cmdq->base + CMDQ_THR_BASE +
> >> > +                               CMDQ_THR_SIZE * i;
> >> > +               INIT_LIST_HEAD(&cmdq->thread[i].task_busy_list);
> >> >
> >> You seem the queue mailbox requests in this controller driver? why not
> >> use the mailbox api for that?
> >>
> >> > +               init_timer(&cmdq->thread[i].timeout);
> >> > +               cmdq->thread[i].timeout.function = cmdq_thread_handle_timeout;
> >> > +               cmdq->thread[i].timeout.data = (unsigned long)&cmdq->thread[i];
> >> >
> >> Here again... you seem to ignore the polling mechanism provided by the
> >> mailbox api, and implement your own.
> >
> > The queue is used to record the tasks which are flushed into CMDQ
> > hardware (GCE). We are handling time critical tasks, so we have to
> > queue them in GCE rather than a software queue (e.g. mailbox buffer).
> > Let me use display as an example. Many display tasks are flushed into
> > CMDQ to wait next vsync event. When vsync event is triggered by display
> > hardware, GCE needs to process all flushed tasks "within vblank" to
> > prevent garbage on screen. This is all done by GCE (without CPU)
> > to fulfill time critical requirement. After GCE finish its work,
> > it will generate interrupts, and then CMDQ driver will let clients know
> > which tasks are done.
> >
> Does the GCE provide any 'lock' to prevent modifying (by adding tasks
> to) the GCE h/w buffer when it is processing it at vsync?  Otherwise

CPU will suspend GCE when adding a task (cmdq_thread_suspend),
and resume GCE after adding task is done (cmdq_thread_resume).
If GCE is processing task(s) at vsync and CPU wants to add a new task
at the same time, CPU will detect this situation
(by cmdq_thread_is_in_wfe), resume GCE immediately, and then add
following task(s) to wait for next vsync event.
All the above logic is implemented at cmdq_task_exec.

> there maybe race/error. If there is such a 'lock' flag/irq, that could
> help here. However, you are supposed to know your h/w better, so I
> will accept this implementation assuming it can't be done any better.
> 
> Please address other comments and resubmit.
> 
> Thanks

After we figure out a better solution for dma_map_single issue, I will
resubmit a new version.

Thanks,
HS

[toc] | [prev] | [next] | [standalone]


#1577512

FromHorng-Shyang Liao <hs.liao@mediatek.com>
Date2017-02-09 13:10 +0100
Message-ID<t8VHX-5Z8-5@gated-at.bofh.it>
In reply to#1574367
On Mon, 2017-02-06 at 13:37 +0800, Horng-Shyang Liao wrote:
> Hi Jassi,
> 
> On Wed, 2017-02-01 at 10:52 +0530, Jassi Brar wrote:
> > On Thu, Jan 26, 2017 at 2:07 PM, Horng-Shyang Liao <hs.liao@mediatek.com> wrote:
> > > Hi Jassi,
> > >
> > > On Thu, 2017-01-26 at 10:08 +0530, Jassi Brar wrote:
> > >> On Wed, Jan 4, 2017 at 8:36 AM, HS Liao <hs.liao@mediatek.com> wrote:
> > >>
> > >> > diff --git a/drivers/mailbox/mtk-cmdq-mailbox.c b/drivers/mailbox/mtk-cmdq-mailbox.c
> > >> > new file mode 100644
> > >> > index 0000000..747bcd3
> > >> > --- /dev/null
> > >> > +++ b/drivers/mailbox/mtk-cmdq-mailbox.c
> > >>
> > >> ...
> > >>
> > >> > +static void cmdq_task_exec(struct cmdq_pkt *pkt, struct cmdq_thread *thread)
> > >> > +{
> > >> > +       struct cmdq *cmdq;
> > >> > +       struct cmdq_task *task;
> > >> > +       unsigned long curr_pa, end_pa;
> > >> > +
> > >> > +       cmdq = dev_get_drvdata(thread->chan->mbox->dev);
> > >> > +
> > >> > +       /* Client should not flush new tasks if suspended. */
> > >> > +       WARN_ON(cmdq->suspended);
> > >> > +
> > >> > +       task = kzalloc(sizeof(*task), GFP_ATOMIC);
> > >> > +       task->cmdq = cmdq;
> > >> > +       INIT_LIST_HEAD(&task->list_entry);
> > >> > +       task->pa_base = dma_map_single(cmdq->mbox.dev, pkt->va_base,
> > >> > +                                      pkt->cmd_buf_size, DMA_TO_DEVICE);
> > >> >
> > >> You seem to parse the requests and responses, that should ideally be
> > >> done in client driver.
> > >> Also, we are here in atomic context, can you move it in client driver
> > >> (before the spin_lock)?
> > >> Maybe by adding a new 'pa_base' member as well in 'cmdq_pkt'.
> > >
> > > will do
> 
> I agree with moving dma_map_single out from spin_lock.
> 
> However, mailbox clients cannot map virtual memory to mailbox
> controller's device for DMA. In our previous discussion, we decided to
> remove mailbox_controller.h from clients to restrict their capabilities.
> 
> Please take a look at following link from 2016/9/22 to 2016/9/30 about
> mailbox_controller.h.
> https://patchwork.kernel.org/patch/9312953/
> 
> Is there any better place to do dma_map_single?

Hi Jassi,

According to previous discussion, we have two requirements:
(1) CMDQ clients should not access mailbox_controller.h;
(2) dma_map_single should not put inside spin_lock.

I think a trade-off solution is to put in mtk-cmdq-helper.c.
Although it is a mailbox client, it is not a CMDQ client.
We can include mailbox_controller.h in mtk-cmdq-helper.c
(instead of mtk-cmdq.h), and then map dma at cmdq_pkt_flush_async
before mbox_send_message.

pkt->pa_base = dma_map_single(client->chan->mbox->dev, pkt->va_base,
                              pkt->cmd_buf_size, DMA_TO_DEVICE);

What do you think?

Thanks,
HS

> > >> ....
> > >> > +
> > >> > +       cmdq->mbox.num_chans = CMDQ_THR_MAX_COUNT;
> > >> > +       cmdq->mbox.ops = &cmdq_mbox_chan_ops;
> > >> > +       cmdq->mbox.of_xlate = cmdq_xlate;
> > >> > +
> > >> > +       /* make use of TXDONE_BY_ACK */
> > >> > +       cmdq->mbox.txdone_irq = false;
> > >> > +       cmdq->mbox.txdone_poll = false;
> > >> > +
> > >> > +       for (i = 0; i < ARRAY_SIZE(cmdq->thread); i++) {
> > >> >
> > >> You mean  i < CMDQ_THR_MAX_COUNT
> > >
> > > will do
> > >
> > >> > +               cmdq->thread[i].base = cmdq->base + CMDQ_THR_BASE +
> > >> > +                               CMDQ_THR_SIZE * i;
> > >> > +               INIT_LIST_HEAD(&cmdq->thread[i].task_busy_list);
> > >> >
> > >> You seem the queue mailbox requests in this controller driver? why not
> > >> use the mailbox api for that?
> > >>
> > >> > +               init_timer(&cmdq->thread[i].timeout);
> > >> > +               cmdq->thread[i].timeout.function = cmdq_thread_handle_timeout;
> > >> > +               cmdq->thread[i].timeout.data = (unsigned long)&cmdq->thread[i];
> > >> >
> > >> Here again... you seem to ignore the polling mechanism provided by the
> > >> mailbox api, and implement your own.
> > >
> > > The queue is used to record the tasks which are flushed into CMDQ
> > > hardware (GCE). We are handling time critical tasks, so we have to
> > > queue them in GCE rather than a software queue (e.g. mailbox buffer).
> > > Let me use display as an example. Many display tasks are flushed into
> > > CMDQ to wait next vsync event. When vsync event is triggered by display
> > > hardware, GCE needs to process all flushed tasks "within vblank" to
> > > prevent garbage on screen. This is all done by GCE (without CPU)
> > > to fulfill time critical requirement. After GCE finish its work,
> > > it will generate interrupts, and then CMDQ driver will let clients know
> > > which tasks are done.
> > >
> > Does the GCE provide any 'lock' to prevent modifying (by adding tasks
> > to) the GCE h/w buffer when it is processing it at vsync?  Otherwise
> 
> CPU will suspend GCE when adding a task (cmdq_thread_suspend),
> and resume GCE after adding task is done (cmdq_thread_resume).
> If GCE is processing task(s) at vsync and CPU wants to add a new task
> at the same time, CPU will detect this situation
> (by cmdq_thread_is_in_wfe), resume GCE immediately, and then add
> following task(s) to wait for next vsync event.
> All the above logic is implemented at cmdq_task_exec.
> 
> > there maybe race/error. If there is such a 'lock' flag/irq, that could
> > help here. However, you are supposed to know your h/w better, so I
> > will accept this implementation assuming it can't be done any better.
> > 
> > Please address other comments and resubmit.
> > 
> > Thanks
> 
> After we figure out a better solution for dma_map_single issue, I will
> resubmit a new version.
> 
> Thanks,
> HS

[toc] | [prev] | [standalone]


Back to top | Article view | linux.kernel


csiph-web