Re: [PATCH V4 2/5] mailbox: Add support for QTI CPUCP mailbox controller

[Konrad Dybcio]

On 4/22/24 18:40, Sibi Sankar wrote:
> Add support for CPUSS Control Processor (CPUCP) mailbox controller,
> this driver enables communication between AP and CPUCP by acting as
> a doorbell between them.
>
> Reviewed-by: Dmitry Baryshkov <dmitry.baryshkov@xxxxxxxxxx>
> Signed-off-by: Sibi Sankar <quic_sibis@xxxxxxxxxxx>
> ---

[...]

> +
> +static int qcom_cpucp_mbox_send_data(struct mbox_chan *chan, void *data)
> +{
> +	struct qcom_cpucp_mbox *cpucp = container_of(chan->mbox, struct qcom_cpucp_mbox, mbox);
> +	unsigned long chan_id = channel_number(chan);
> +	u32 *val = data;
> +
> +	writel(*val, cpucp->tx_base + APSS_CPUCP_TX_MBOX_CMD(chan_id) + APSS_CPUCP_MBOX_CMD_OFF);

Just checking in, is *this access only* supposed to be 32b instead of 64 like others?

[...]

> +
> +	writeq(0, cpucp->rx_base + APSS_CPUCP_RX_MBOX_EN);
> +	writeq(0, cpucp->rx_base + APSS_CPUCP_RX_MBOX_CLEAR);
> +	writeq(0, cpucp->rx_base + APSS_CPUCP_RX_MBOX_MAP);

If these writes are here to prevent a possible interrupt storm type tragedy,
you need to read back these registers to ensure the writes have left the CPU
complex and reached the observer at the other end of the bus (not to be
confused with barriers which only ensure that such accesses are ordered
*when still possibly within the CPU complex*).

Moreover, if the order of them arriving (en/clear/mask) doesn't matter, you
can add _relaxed for a possible nanosecond-order perf gain

> +
> +	irq = platform_get_irq(pdev, 0);
> +	if (irq < 0)
> +		return irq;
> +
> +	ret = devm_request_irq(dev, irq, qcom_cpucp_mbox_irq_fn,
> +			       IRQF_TRIGGER_HIGH, "apss_cpucp_mbox", cpucp);
> +	if (ret < 0)
> +		return dev_err_probe(dev, ret, "Failed to register irq: %d\n", irq);
> +
> +	writeq(APSS_CPUCP_RX_MBOX_CMD_MASK, cpucp->rx_base + APSS_CPUCP_RX_MBOX_MAP);

Similarly here, unless read back, we may potentially miss some interrupts if
e.g. a channel is opened and that write "is decided" (by the silicon) to leave
the internal buffer first


> +
> +	mbox = &cpucp->mbox;
> +	mbox->dev = dev;
> +	mbox->num_chans = APSS_CPUCP_IPC_CHAN_SUPPORTED;
> +	mbox->chans = cpucp->chans;
> +	mbox->ops = &qcom_cpucp_mbox_chan_ops;
> +	mbox->txdone_irq = false;
> +	mbox->txdone_poll = false;

"false" == 0 is the default value (as you're using k*z*alloc)


> +
> +	ret = devm_mbox_controller_register(dev, mbox);
> +	if (ret)
> +		return dev_err_probe(dev, ret, "Failed to create mailbox\n");
> +
> +	return 0;
> +}
> +
> +static const struct of_device_id qcom_cpucp_mbox_of_match[] = {
> +	{ .compatible = "qcom,x1e80100-cpucp-mbox" },
> +	{}
> +};
> +MODULE_DEVICE_TABLE(of, qcom_cpucp_mbox_of_match);
> +
> +static struct platform_driver qcom_cpucp_mbox_driver = {
> +	.probe = qcom_cpucp_mbox_probe,
> +	.driver = {
> +		.name = "qcom_cpucp_mbox",
> +		.of_match_table = qcom_cpucp_mbox_of_match,
> +	},
> +};
> +module_platform_driver(qcom_cpucp_mbox_driver);

That's turbo late. Go core_initcall.

Konrad