On Fri, Apr 6, 2018 at 8:35 PM, Fabien DESSENNE <fabien.dessenne@xxxxxx> wrote:
>
> On 06/04/18 14:56, Jassi Brar wrote:
>> On Fri, Apr 6, 2018 at 5:59 PM, Fabien DESSENNE <fabien.dessenne@xxxxxx> wrote:
>>> Hi
>>>
>>>
>>> On 05/04/18 11:38, Jassi Brar wrote:
>>>> On Mon, Mar 12, 2018 at 4:28 PM, Fabien Dessenne <fabien.dessenne@xxxxxx> wrote:
>>>> ....
>>>>> +
>>>>> + /* irq */
>>>>> + for (i = 0; i < IPCC_IRQ_NUM; i++) {
>>>>> + ipcc->irqs[i] = of_irq_get_byname(dev->of_node, irq_name[i]);
>>>>> + if (ipcc->irqs[i] < 0) {
>>>>> + dev_err(dev, "no IRQ specified %s\n", irq_name[i]);
>>>>> + ret = ipcc->irqs[i];
>>>>> + goto err_clk;
>>>>> + }
>>>>> +
>>>>> + ret = devm_request_threaded_irq(dev, ipcc->irqs[i], NULL,
>>>>> + irq_thread[i], IRQF_ONESHOT,
>>>>> + dev_name(dev), ipcc);
>>>>>
>>>> In your interrupt handlers you don't do anything that could block.
>>>> Threads only adds some delay to your message handling.
>>>> So maybe use devm_request_irq() ?
>>> The interrupt handlers call mbox_chan_received_data() /
>>> mbox_chan_txdone(), which call in turn client's rx_callback() /
>>> tx_done() / tx_prepare() which behavior may be unsafe. Hence, using a
>>> threaded irq here seems to be a good choice.
>>>
>> rx_callback() is supposed to be atomic.
>
> I am worried with this atomic part (and honestly I did not note that the
> callbacks were expected to be)
>
> In my case, remoteproc->virtio->rpmsg is the mailbox client defining the
> rx_callback.
> If I follow your suggestion, I shall make this rx_callback Atomic in
> remoteproc (or in virtio or rpmsg). And this does not seem to be so
> simple (add a worker in the middle of somewhere?). Bjorn, feel free to
> comment this part.
>
> An alternate implementation consists in using a threaded IRQ for the
> mailbox interrupt.
> This option is not only simple, but also ensures to split bottom & half
> parts at the irq level which is IMHO a general good practice.
>
> I can see that some mailbox clients implement callbacks that are NOT
> atomic and I suspect this is the reason why some mailbox drivers use
> threaded_irq (rockchip mailbox splits the bottom & half parts).
>
> Would it be acceptable to consider the "atomic client callback" as a
> non-strict rule ?
>
Of course you can traverse atomic path from sleepable context (but not
vice-versa).
Please send in the final revision.
Thanks.
--
To unsubscribe from this list: send the line "unsubscribe devicetree" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
