[...]

quoted

+	/*
+	 * If this is the first gpio_request for the bank,
+	 * enable the bank module.
+	 */
+	if (!bank->mod_usage) {
+		if (IS_ERR_VALUE(pm_runtime_get_sync(bank->dev) < 0)) {
+			dev_err(bank->dev, "%s: GPIO bank %d "
+					"pm_runtime_get_sync failed\n",
+					__func__, bank->id);
+			return -EINVAL;

Must spin_unlock_irqrestore() first.

Yes.

quoted

+		}
+
+		/* Initialize the gpio bank registers to init time value */
+		omap_gpio_mod_init(bank);

omap_gpio_mod_init calls mpuio_init calls platform_driver_register
which can't be called in an IRQs off and spinlocked atomic context,
for example, device_private_init calls kzalloc with GFP_KERNEL.

Right! The present form of *_mod_init() is not suitable here.

Concurrency protection for this will need to happen prior to the
spinlock (assuming it really does need to be an IRQ saving spinlock
and not a mutex).   Possibly a new mutex is needed to protect the
check for first usage and init'ing the bank (and blocking a racing
second caller until the init is done).

I looked at the implementation once again with the comments at background.
There are gaps which require fixing. Thanks.

The omap_gpio_mod_init may be unbalanced with the code performed below
on last free of a GPIO for the bank?  If all GPIOs are freed and then
a new GPIO used, does omap_gpio_mod_init do the right thing?  Need a
separate flag to indicate whether one-time init has ever been
performed, vs. needing runtime PM enable/disable?

Yes, there is imbalance. I am re-looking and fixing.
--
Tarun
[...]