30.01.2022 19:43, Akhil R пишет:
>> 30.01.2022 13:26, Dmitry Osipenko пишет:
>>> 30.01.2022 13:05, Dmitry Osipenko пишет:
>>>> Still nothing prevents interrupt handler to fire during the pause.
>>>>
>>>> What you actually need to do is to disable/enable interrupt. This
>>>> will prevent the interrupt racing and then pause/resume may look like this:
>>>
>>> Although, seems this won't work, unfortunately. I see now that
>>> device_pause() doesn't have might_sleep().
>>>
>>
>> Ah, I see now that the pause/unpause is actually a separate control and doesn't
>> conflict with "start next transfer".
>>
>> So you just need to set/unset the pause under lock. And don't touch
>> tdc->dma_desc. That's it.
>>
>> static int tegra_dma_device_pause(struct dma_chan *dc) {
>> struct tegra_dma_channel *tdc = to_tegra_dma_chan(dc);
>> unsigned long flags;
>> u32 val;
>>
>> if (!tdc->tdma->chip_data->hw_support_pause)
>> return -ENOSYS;
>>
>> spin_lock_irqsave(&tdc->vc.lock, flags);
>>
>> val = tdc_read(tdc, TEGRA_GPCDMA_CHAN_CSRE);
>> val |= TEGRA_GPCDMA_CHAN_CSRE_PAUSE;
>> tdc_write(tdc, TEGRA_GPCDMA_CHAN_CSRE, val);
>>
>> spin_unlock_irqrestore(&tdc->vc.lock, flags);
>>
>> return 0;
>> }
>>
>> static int tegra_dma_device_resume(struct dma_chan *dc) {
>> struct tegra_dma_channel *tdc = to_tegra_dma_chan(dc);
>> unsigned long flags;
>> u32 val;
>>
>> if (!tdc->tdma->chip_data->hw_support_pause)
>> return -ENOSYS;
>>
>> spin_lock_irqsave(&tdc->vc.lock, flags);
>>
>> val = tdc_read(tdc, TEGRA_GPCDMA_CHAN_CSRE);
>> val &= ~TEGRA_GPCDMA_CHAN_CSRE_PAUSE;
>> tdc_write(tdc, TEGRA_GPCDMA_CHAN_CSRE, val);
>>
>> spin_unlock_irqrestore(&tdc->vc.lock, flags);
>>
>> return 0;
>> }
>
> The reason I separated out register writes was to conveniently call those
> in dma_start() and terminate_all(). Do you see any issue there?
> The recommended way of terminating a transfer in between is to pause
> it before disabling the channel.
This is a sample code, feel free to adjust it as needed.
> dma_desc could be NULL while these functions are called. pause() or
> resume() is unneeded if there isn't any transfer going on. Moreover,
> if we are to calculate the xfer_size, the check would be mandatory.
For now looks like it should be better to move the xfer_size updating to
other places, like terminate_all() and tx_status().
I also see now that you have
residue_granularity=DMA_RESIDUE_GRANULARITY_BURST, while tx_status() in
fact uses segment granularity. This needs to be corrected.
You also must add locking to tx_status(), to protect tdc->dma_desc
pointer updates.
