ARM: mmp: fix timer_read delay

[ Upstream commit e348b4014c31041e13ff370669ba3348c4d385e3 ]

timer_read() was using an empty 100-iteration loop to wait for the
TMR_CVWR register to capture the latest timer counter value. The delay
wasn't long enough. This resulted in CPU idle time being extremely
underreported on PXA168 with CONFIG_NO_HZ_IDLE=y.

Switch to the approach used in the vendor kernel, which implements the
capture delay by reading TMR_CVWR a few times instead.

Fixes: 49cbe78637eb ("[ARM] pxa: add base support for Marvell's PXA168 processor line")
Signed-off-by: Doug Brown <doug@schmorgal.com>
Link: https://lore.kernel.org/r/20221204005117.53452-3-doug@schmorgal.com
Signed-off-by: Arnd Bergmann <arnd@arndb.de>
Signed-off-by: Sasha Levin <sashal@kernel.org>

diff --git a/arch/arm/mach-mmp/time.c b/arch/arm/mach-mmp/time.c
index 41b2e8a..708816c 100644 (file)
--- a/arch/arm/mach-mmp/time.c
+++ b/arch/arm/mach-mmp/time.c
@@ -43,18 +43,21 @@
 static void __iomem *mmp_timer_base = TIMERS_VIRT_BASE;
 
 /*
- * FIXME: the timer needs some delay to stablize the counter capture
+ * Read the timer through the CVWR register. Delay is required after requesting
+ * a read. The CR register cannot be directly read due to metastability issues
+ * documented in the PXA168 software manual.
  */
 static inline uint32_t timer_read(void)
 {
-       int delay = 100;
+       uint32_t val;
+       int delay = 3;
 
        __raw_writel(1, mmp_timer_base + TMR_CVWR(1));
 
        while (delay--)
-               cpu_relax();
+               val = __raw_readl(mmp_timer_base + TMR_CVWR(1));
 
-       return __raw_readl(mmp_timer_base + TMR_CVWR(1));
+       return val;
 }
 
 static u64 notrace mmp_read_sched_clock(void)