Re: [PATCH v1 3/4] bus: mhi: core: Add support to pause or resume channel data transfers

[Bhaumik Bhatt <bbhatt@codeaurora.org>]

Hi Loic,

On 2020-11-10 03:14, Loic Poulain wrote:
> Hi Bhaumik,
> 
> On Mon, 9 Nov 2020 at 23:44, Bhaumik Bhatt <bbhatt@codeaurora.org> 
> wrote:
>> 
>> Some MHI clients may want to request for pausing or resuming of the
>> data transfers for their channels. Enable them to do so using the new
>> APIs provided for the same.
>> 
>> Signed-off-by: Bhaumik Bhatt <bbhatt@codeaurora.org>
>> ---
>>  drivers/bus/mhi/core/main.c | 41 
>> +++++++++++++++++++++++++++++++++++++++++
>>  include/linux/mhi.h         | 16 ++++++++++++++++
>>  2 files changed, 57 insertions(+)
>> 
>> diff --git a/drivers/bus/mhi/core/main.c b/drivers/bus/mhi/core/main.c
>> index 1226933..01845c6 100644
>> --- a/drivers/bus/mhi/core/main.c
>> +++ b/drivers/bus/mhi/core/main.c
>> @@ -1560,6 +1560,47 @@ void mhi_unprepare_from_transfer(struct 
>> mhi_device *mhi_dev)
>>  }
>>  EXPORT_SYMBOL_GPL(mhi_unprepare_from_transfer);
>> 
>> +static int mhi_update_transfer_state(struct mhi_device *mhi_dev,
>> +                                    enum mhi_ch_state_type to_state)
>> +{
>> +       struct mhi_controller *mhi_cntrl = mhi_dev->mhi_cntrl;
>> +       struct mhi_chan *mhi_chan;
>> +       int dir, ret;
>> +
>> +       for (dir = 0; dir < 2; dir++) {
>> +               mhi_chan = dir ? mhi_dev->ul_chan : mhi_dev->dl_chan;
>> +
>> +               if (!mhi_chan)
>> +                       continue;
>> +
>> +               /*
>> +                * Bail out if one of the channels fail as client will 
>> reset
>> +                * both upon failure
>> +                */
>> +               mutex_lock(&mhi_chan->mutex);
>> +               ret = mhi_update_channel_state(mhi_cntrl, mhi_chan, 
>> to_state);
>> +               if (ret) {
>> +                       mutex_unlock(&mhi_chan->mutex);
>> +                       return ret;
>> +               }
>> +               mutex_unlock(&mhi_chan->mutex);
>> +       }
>> +
>> +       return 0;
>> +}
>> +
>> +int mhi_pause_transfer(struct mhi_device *mhi_dev)
>> +{
>> +       return mhi_update_transfer_state(mhi_dev, 
>> MHI_CH_STATE_TYPE_STOP);
>> +}
>> +EXPORT_SYMBOL_GPL(mhi_pause_transfer);
>> +
>> +int mhi_resume_transfer(struct mhi_device *mhi_dev)
>> +{
>> +       return mhi_update_transfer_state(mhi_dev, 
>> MHI_CH_STATE_TYPE_START);
>> +}
>> +EXPORT_SYMBOL_GPL(mhi_resume_transfer);
> 
> Look like it is stop and start, not pause and resume?
I wanted to keep it pause and resume because it could get confusing for 
someone
looking at this pair of APIs, that a client driver would also need to 
"start"
channels after "preparing" them. Since that is not that case, and the
mhi_prepare_for_transfer() API itself is supposed to also start the 
channels, it
would be better to keep these as "pause" and "resume" instead IMO.

Any comments in favor or "stop" and "start"?
> 
> TBH maybe we should rework/clarify MHI core and having well-defined
> states, maybe something like that:
> 
> 1. When MHI core detects device for a driver, MHI core resets and
> initializes the channel(s), then call client driver probe function
>     => channel UNKNOWN->DISABLED state
>     => channel DISABLED->ENABLED state
> 2. When driver is ready for sending data, drivers calls 
> mhi_start_transfer
>     => Channel is ENABLED->RUNNING state
> 3. Driver performs normal data transfers
> 4. The driver can suspend/resume transfer, it stops (suspend) the 
> channel, can
>     => Channel is RUNNING->STOP
>     => Channel is STOP->RUNNING
>    ...
> 5. When device is removed, MHI core reset the channel
>     => channel is (RUNNING|STOP) -> DISABLED
> 
> Today mhi_prepare_for_transfer performs both ENABLE and RUNNING
> transition, the idea would be to keep channel enabling/disabling in
> the MHI core (before/after driver probe/remove) and channel start/stop
> managed by the client driver.
> 
> Regards,
> Loic

Your idea is good but it would not have much additional benefits and 
would
involve MHI core "enabling" channels and allocating memory for each 
channel
context when they are only declared as supported by the controller but 
are not
actually being put to use.

mhi_prepare_for_transfer() does both channel context initialization and 
starts
the channels, which is good because it allocates memory when needed. So, 
this
benefits system memory if a controller with support for many channels 
exists but
only a few channels are used.

Regarding the states to track from host:
-> DISABLED (We know channels are not active: in reset state or not 
probed yet)
-> ENABLED (Active and running when needed for data transfers)
-> STOP (Paused: leaves the channel context as is since channels are not 
reset)
-> SUSPENDED (Unload in progress: Entered before resetting 
channels/remove())

BTW, we have the debugfs entry for "channels" that dumps the context to 
show
exactly what the channel states are from device perspective. We can rely 
on it
if needed.

If there are some comments I can add to make things clear, please let me 
know.

Thanks,
Bhaumik
-- 
The Qualcomm Innovation Center, Inc. is a member of the Code Aurora 
Forum,
a Linux Foundation Collaborative Project