clk: ti: clkctrl: use fallback udelay approach if timekeeping is susp…

…ended

In certain cases it is possible that the timekeeping has been suspended
already when attempting to disable/enable a clkctrl clock. This will
happen at least on am43xx platform when attempting to enable / disable
the clockevent source itself, burping out a warning from timekeeping core.

The sequence of events leading to this:
-> timekeeping_suspend()
 -> clockevents_suspend()
  -> omap_clkevt_idle()
   -> omap_hwmod_idle()
    -> _omap4_clkctrl_clk_disable()
     -> _omap4_is_timeout()

Avoid the issue by checking if the timekeeping is suspended and using
the fallback udelay approach for checking timeouts.

Signed-off-by: Tero Kristo <t-kristo@ti.com>
Acked-by: Stephen Boyd <sboyd@codeaurora.org>

drivers/clk/ti/clkctrl.c

@@ -21,6 +21,7 @@
 #include <linux/of_address.h>
 #include <linux/clk/ti.h>
 #include <linux/delay.h>
+#include <linux/timekeeping.h>
 #include "clock.h"
 
 #define NO_IDLEST			0x1
@@ -90,7 +91,18 @@ static bool _omap4_is_ready(u32 val)
 
 static bool _omap4_is_timeout(union omap4_timeout *time, u32 timeout)
 {
-	if (unlikely(_early_timeout)) {
+	/*
+	 * There are two special cases where ktime_to_ns() can't be
+	 * used to track the timeouts. First one is during early boot
+	 * when the timers haven't been initialized yet. The second
+	 * one is during suspend-resume cycle while timekeeping is
+	 * being suspended / resumed. Clocksource for the system
+	 * can be from a timer that requires pm_runtime access, which
+	 * will eventually bring us here with timekeeping_suspended,
+	 * during both suspend entry and resume paths. This happens
+	 * at least on am43xx platform.
+	 */
+	if (unlikely(_early_timeout || timekeeping_suspended)) {
 		if (time->cycles++ < timeout) {
 			udelay(1);
 			return false;