On Thu, Jan 02, 2020 at 05:40:48PM +0300, Dmitry Osipenko wrote:

30.12.2019 23:36, Michał Mirosław пишет:

quoted

On Sat, Dec 28, 2019 at 11:33:55PM +0300, Dmitry Osipenko wrote:

quoted

There is no good reason for disabling of CPU interrupts in order to
protect the utmip_pad_count modification.

Since there are sleeping functions called outside of the locked sections,
this should be a mutex instead. OTOH if the spin_lock is to protect register
write against IRQ handler, then the patch is wrong.

[...]

quoted

-	spin_unlock_irqrestore(&utmip_pad_lock, flags);
+	spin_unlock(&utmip_pad_lock);
 
 	clk_disable_unprepare(phy->pad_clk);

Hello Michał,

This spinlock isn't for protecting from the IRQ handler, it's used
solely to protect modification of the registers that are shared by all
USB controllers.

It's possible to use mutex instead of spinlock here, but it doesn't
bring any benefits because mutex is more useful when protected code
could block for a long time due to sleep or whatever, while spinlock is
much more efficient when protected code doesn't sleep and takes no more
than dozens microseconds to execute (which is the case here).

In this particular case of the Tegra USB PHY driver, the chance of
getting a block on taking the utmip_pad_lock is zero unless USB
controller drivers will start to use asynchronous probing. So this patch
does a very minor clean-up change and nothing more.

I was concerned that this change allows the kernel to switch away to
another task, but I can see now that spin_lock() does disable preemtion.
So it looks OK after all. Would be nice to see the explanation in the
commit message (that the spinlock is only used to serialize probe()s).

Best Regards,
Michał Mirosław