Re: [PATCH V2 4/6] slim: qcom: Add Qualcomm Slimbus controller driver

From: Sagar Dharia <sdharia@codeaurora.org>
To: Mark Brown <broonie@kernel.org>
Cc: gregkh@linuxfoundation.org, bp@suse.de, poeschel@lemonage.de,
	treding@nvidia.com, gong.chen@linux.intel.com,
	andreas.noever@gmail.com, alan@linux.intel.com,
	mathieu.poirier@linaro.org, daniel@ffwll.ch, oded.gabbay@amd.com,
	jkosina@suse.cz, sharon.dvir1@mail.huji.ac.il, joe@perches.com,
	davem@davemloft.net, james.hogan@imgtec.com,
	michael.opdenacker@free-electrons.com,
	daniel.thompson@linaro.org, linux-kernel@vger.kernel.org,
	nkaje@codeaurora.org, kheitke@audience.com,
	mlocke@codeaurora.org, agross@codeaurora.org,
	linux-arm-msm@vger.kernel.org
Subject: Re: [PATCH V2 4/6] slim: qcom: Add Qualcomm Slimbus controller driver
Date: Sun, 28 Jun 2015 13:13:41 -0600	[thread overview]
Message-ID: <55904765.4070108@codeaurora.org> (raw)
In-Reply-To: <20150617135313.GH3214@sirena.org.uk>

On 6/17/2015 7:53 AM, Mark Brown wrote:
> On Tue, Jun 16, 2015 at 07:46:02PM -0600, Sagar Dharia wrote:
>
>> + - dmaengine, and pipes used to communicate between controller and memory if
>> +	sps-BAM HW is used
> This needs more detail.
>> +		 */
>> +		mb();
>> +		if (notify_rx)
>> +			complete(&dev->rx_msgq_notify);
>> +	}
>> +	return IRQ_HANDLED;
> This interrupt handler unconditionally returns IRQ_HANDLED regardless of
> if that's true or not.

Some interrupt handling is done in ISR itself. RX-msgq thread 
notifier/workQ is only used for items that can sleep/operations that can 
take longer. (e.g. assignment of logical address based on report_present 
message)
That's the reason IRQ_HANDLED is always true here.
>
>> +static void msm_slim_wait_retry(struct msm_slim_ctrl *dev)
>> +{
>> +	int msec_per_frm = 0;
>> +	int sfr_per_sec;
>> +
>> +	/* Wait for 1 superframe, or default time and then retry */
>> +	sfr_per_sec = dev->framer.superfreq /
>> +			(1 << (SLIM_MAX_CLK_GEAR - dev->ctrl.clkgear));
>> +	if (sfr_per_sec)
>> +		msec_per_frm = MSEC_PER_SEC / sfr_per_sec;
>> +	if (msec_per_frm < DEF_RETRY_MS)
>> +		msec_per_frm = DEF_RETRY_MS;
>> +	msleep(msec_per_frm);
>> +}
> Is this device specific?

This is seen on most of the Qualcomm-controllers I've worked with, where 
retries in early initialization were needed to avoid issues due to noise.
>
>> +static void msm_slim_cb(void *ctx, int err)
>> +{
>> +	if (err)
>> +		pr_err("MSM-slim completion reported err:%d\n", err);
> dev_err()?
>
>> +	else if (ctx)
>> +		complete(ctx);
>> +}
> That's weird, if we get an error we don't signal whatever's waiting -
> won't it just time out at best then?  Also, what happens if we get
> neither an error nor context?
>
>> +	if (txn->msg->wbuf)
>> +		memcpy(puc, txn->msg->wbuf, txn->msg->num_bytes);
>> +	msm_slim_queue_tx(dev, pbuf, txn->rl, MGR_TX_MSG);
>> +	timeout = wait_for_completion_timeout(&done,
>> +					      msecs_to_jiffies(txn->rl + 100));
>> +
>> +	if (!timeout)
>> +		dev_err(dev->dev, "TX timed out:MC:0x%x,mt:0x%x\n", txn->mc,
>> +			txn->mt);
>> +
>> +	mutex_unlock(&dev->txn_lock);
>> +	return timeout ? 0 : -ETIMEDOUT;
>> +}
> Shouldn't we provide a route for error reports here (and might some of
> this timeout stuff go into the core)?

Sure, good point. If I move it to core, all controllers will be expected 
to have 'write-done' notification. If that's a fair assumption, I will 
move it to core.
That will also mean the above callback/completion will move to framework.
>
>> +
>> +	if ((msm_slim_put_rx(dev, (u8 *)buf)) != -ENODATA) {
>> +		len = buf[0] & 0x1F;
>> +		mt = (buf[0] >> 5) & 0x7;
>> +		mc = buf[1];
>> +		if (mt == SLIM_MSG_MT_CORE &&
>> +			mc == SLIM_MSG_MC_REPORT_PRESENT) {
> Looks like a switch statement.
>
>> +static int msm_slim_rx_msgq_thread(void *data)
>> +{
>> +	struct msm_slim_ctrl *dev = (struct msm_slim_ctrl *)data;
>> +	struct completion *notify = &dev->rx_msgq_notify;
>> +	int ret;
>> +
>> +	while (!kthread_should_stop()) {
>> +		set_current_state(TASK_INTERRUPTIBLE);
>> +		ret = wait_for_completion_interruptible(notify);
>> +
>> +		if (ret)
>> +			dev_err(dev->dev, "rx thread wait error:%d\n", ret);
>> +		else
>> +			msm_slim_rxwq(dev);
>> +	}
>> +
>> +	return 0;
>> +}
> It's not entirely clear to me why this is a kthread rather than a
> workqueue or something.  I'm also unclear what happens if more than one
> piece of work is queued prior to msm_slim_rxwq() running, it looks like
> it only handles a single operation.

Sure, I will change this to workqueue. That should not be a problem.
>
>> +	/* SLEW RATE register for this slimbus */
>> +	dev->slew_mem = platform_get_resource_byname(pdev, IORESOURCE_MEM,
>> +						"slimbus_slew_reg");
>> +	if (!dev->slew_mem) {
>> +		dev_err(&pdev->dev, "no slimbus slew resource\n");
>> +		return;
>> +	}
> Warning not an error isn't it?
>
>> +	hclk = clk_get(&pdev->dev, "iface_clk");
>> +	if (IS_ERR(hclk))
>> +		return PTR_ERR(hclk);
> devm_clk_get()
>
>> +	ret = clk_set_rate(rclk, SLIM_ROOT_FREQ);
> You're ignoring this error in spite of assigning to ret.
>
>> +	slim_mem = platform_get_resource_byname(pdev, IORESOURCE_MEM,
>> +						"slimbus_physical");
>> +	if (!slim_mem) {
>> +		dev_err(&pdev->dev, "no slimbus physical memory resource\n");
>> +		ret = ENODEV;
>> +		goto err_get_mem_failed;
>> +	}
>> +	slim_io = request_mem_region(slim_mem->start, resource_size(slim_mem),
>> +					pdev->name);
>> +	if (!slim_io) {
>> +		dev_err(&pdev->dev, "slimbus memory already claimed\n");
>> +		ret = EBUSY;
>> +		goto err_get_mem_failed;
>> +	}
> devm_ioremap_resource() and a lot of this starts looking simpler.
>
>> +	dev = kzalloc(sizeof(struct msm_slim_ctrl), GFP_KERNEL);
> devm_kzalloc()
>
>> +	if (pdev->dev.of_node) {
>> +
>> +		ret = of_property_read_u32(pdev->dev.of_node, "cell-index",
>> +					&dev->ctrl.nr);
> Was that in the binding?
>
>> +	/* Register with framework before enabling frame, clock */
>> +	ret = slim_add_numbered_controller(&dev->ctrl);
>> +	if (ret) {
>> +		dev_err(dev->dev, "error adding controller\n");
>> +		goto err_ctrl_failed;
>> +	}
> This is suspicious - why are you adding a numbered controller with DT?
> And looking at this interface why is it a separate call to add a
> numbered controller, why not just register the controller using the same
> API and handle any number that was set when doing that?

Agreed, I will change add_numbered_controller to register_controller. 
That will also mean the 'cell-index' property above is not needed.
>
>> +	dev->ver = readl_relaxed(dev->base);
>> +	/* Version info in 16 MSbits */
>> +	dev->ver >>= 16;
>> +	/* Component register initialization */
>> +	writel_relaxed(1, dev->base + CFG_PORT(COMP_CFG, dev->ver));
>> +	writel_relaxed((EE_MGR_RSC_GRP | EE_NGD_2 | EE_NGD_1),
>> +				dev->base + CFG_PORT(COMP_TRUST_CFG, dev->ver));
> You've registered the device with the core prior to initialising the
> hardware - I'd expect this means the generic code will start trying to
> register slaves immediately.  The normal pattern would be to initialise
> the hardware then register it.

This is for handling the logical-address assignment. That's only 
possible after registering controller.
The internal controller has 2 in-built slimbus devices (interface, 
framer) which will start to report-present right-away if I initialize 
the HW. Now if register_controller is not done, the logical-address 
assignment will not happen and as per spec, the devices will keep 
sending report_present until they get an assignment.

Registering slaves immediately is not an issue since that part will not 
be waiting on logical-address assignment (due to device_up callback)
>
>> +#ifdef CONFIG_PM_SLEEP
>> +static int msm_slim_suspend(struct device *dev)
>> +{
>> +	return 0;
>> +}
>> +
>> +static int msm_slim_resume(struct device *dev)
>> +{
>> +	return 0;
>> +}
>> +#endif /* CONFIG_PM_SLEEP */
> Omit empty functions.
>
>> +static int msm_slim_init(void)
>> +{
>> +	return platform_driver_register(&msm_slim_driver);
>> +}
>> +module_init(msm_slim_init);
> module_platform_driver().
>
>> +u8 *msm_slim_get_tx(struct msm_slim_ctrl *dev, struct msm_wr_cb *cur)
>> +{
>> +	unsigned long flags;
>> +	int idx;
>> +
>> +	spin_lock_irqsave(&dev->tx.lock, flags);
>> +	if (((dev->tx.head + 1) % MSM_TX_MSGS) == dev->tx.tail) {
>> +		spin_unlock_irqrestore(&dev->tx.lock, flags);
>> +		return NULL;
>> +	}
>> +	idx = dev->tx.tail;
>> +	dev->tx.tail = (dev->tx.tail + 1) % MSM_TX_MSGS;
>> +	spin_unlock_irqrestore(&dev->tx.lock, flags);
>> +	dev->pending_wr[idx] = cur;
>> +	return dev->tx.base + (idx * SLIM_MSGQ_BUF_LEN);
>> +}
> Would just using a regular list hurt?

This provides O(1) circular buffer access rather than a list with O(n).
I will address other comments and upload patch later today/tomorrow.

Thanks for your comments
Sagar

-- 
Qualcomm Innovation Center, Inc. is a member of Code Aurora Forum,
hosted by The Linux Foundation