Re: [PATCH v4 2/3] i2c: iproc: Add Broadcom iProc I2C Driver

On 1/19/2015 11:28 AM, Russell King - ARM Linux wrote:

On Sat, Jan 17, 2015 at 04:30:33PM -0800, Ray Jui wrote:

quoted

On 1/17/2015 2:40 PM, Russell King - ARM Linux wrote:

quoted

On Sat, Jan 17, 2015 at 01:26:41PM -0800, Ray Jui wrote:

quoted

	time_left = wait_for_completion_timeout(&iproc_i2c->done, time_left);

	/* disable all interrupts */
	writel(0, iproc_i2c->base + IE_OFFSET);

	if (!time_left && !atomic_read(&iproc_i2c->transfer_is_successful)) {

Why are you using atomic_read() here?

transfer_is_successful 1) will be reset to 0 in this function (before
kick start the I2C transfer), 2) will be set to 1 in the ISR (to signal
completion of the I2C transfer), and 3) will be checked in this function
here. I thought that means I should declare it volatile, because it can
be modified in both the process context and interrupt context (and I use
atomic because I remember Linux checkpatch warns against using volatile)?

You don't need volatile or atomic_t for that.

Rather than switching to atomic_t when seeing the checkpatch warning,
you'd do better to read Documentation/volatile-considered-harmful.txt
to understand why checkpatch issues the warning, and realise that you
don't need it for the above.

Note that in the above code, the compiler can't make an assumption
about iproc_i2c->transfer_is_successful because it can't tell whether
a called function (eg, wait_for_completion_timeout()) could modify it.

Got it. Thanks.

Another possible issue with the above code are these lines:

	/* disable all interrupts */
	writel(0, iproc_i2c->base + IE_OFFSET);

It would be nice to think that would hit the hardware immediately, but
that's making assumptions about hardware which are not necessary true.
Your interrupt handler could even be running on another CPU after you've
asked for that register to be written.

Depending on what you're trying to achieve here, you may need:

	/* disable all interrupts */
	writel(0, iproc_i2c->base + IE_OFFSET);
	/* read it back to ensure the write has hit */
	readl(iproc_i2c->base + IE_OFFSET);

	/* make sure the interrupt handler isn't running */
	synchronize_irq(...->irq);

if what you're trying to do is to ensure that the interrupt handler has
finished running.

This will be the most robust way of handling this. Given that we've
added an additional flag to check to make sure there's no interrupt
missed after wait_for_completion_timeout times out, it makes sense to
ensure that by the time when we check the flag there's no pending irq.
I'll add this to the driver and make 'xfer_is_done' an 'int' instead of
'atomic_t'. I will also add the call to readl to flush the write in the
remove function after interrupts are disabled.