ARM: sched_clock: provide common infrastructure for sched_clock()

Provide common sched_clock() infrastructure for platforms to use to
create a 64-bit ns based sched_clock() implementation from a counter
running at a non-variable clock rate.

This implementation is based upon maintaining an epoch for the counter
and an epoch for the nanosecond time.  When we desire a sched_clock()
time, we calculate the number of counter ticks since the last epoch
update, convert this to nanoseconds and add to the epoch nanoseconds.

We regularly refresh these epochs within the counter wrap interval.
We perform a similar calculation as above, and store the new epochs.

We read and write the epochs in such a way that sched_clock() can easily
(and locklessly) detect when an update is in progress, and repeat the
loading of these constants when they're known not to be stable.  The
one caveat is that sched_clock() is not called in the middle of an
update.  We achieve that by disabling IRQs.

Finally, if the clock rate is known at compile time, the counter to ns
conversion factors can be specified, allowing sched_clock() to be tightly
optimized.  We ensure that these factors are correct by providing an
initialization function which performs a run-time check.

Acked-by: Peter Zijlstra <peterz@infradead.org>
Tested-by: Santosh Shilimkar <santosh.shilimkar@ti.com>
Tested-by: Will Deacon <will.deacon@arm.com>
Tested-by: Mikael Pettersson <mikpe@it.uu.se>
Tested-by: Eric Miao <eric.y.miao@gmail.com>
Tested-by: Olof Johansson <olof@lixom.net>
Tested-by: Jamie Iles <jamie@jamieiles.com>
Reviewed-by: Nicolas Pitre <nicolas.pitre@linaro.org>
Signed-off-by: Russell King <rmk+kernel@arm.linux.org.uk>

arch/arm/Kconfig

@@ -38,6 +38,9 @@ config HAVE_PWM
 config SYS_SUPPORTS_APM_EMULATION
 	bool
 
+config HAVE_SCHED_CLOCK
+	bool
+
 config GENERIC_GPIO
 	bool
 

arch/arm/include/asm/sched_clock.h

@@ -0,0 +1,118 @@
+/*
+ * sched_clock.h: support for extending counters to full 64-bit ns counter
+ *
+ * This program is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License version 2 as
+ * published by the Free Software Foundation.
+ */
+#ifndef ASM_SCHED_CLOCK
+#define ASM_SCHED_CLOCK
+
+#include <linux/kernel.h>
+#include <linux/types.h>
+
+struct clock_data {
+	u64 epoch_ns;
+	u32 epoch_cyc;
+	u32 epoch_cyc_copy;
+	u32 mult;
+	u32 shift;
+};
+
+#define DEFINE_CLOCK_DATA(name)	struct clock_data name
+
+static inline u64 cyc_to_ns(u64 cyc, u32 mult, u32 shift)
+{
+	return (cyc * mult) >> shift;
+}
+
+/*
+ * Atomically update the sched_clock epoch.  Your update callback will
+ * be called from a timer before the counter wraps - read the current
+ * counter value, and call this function to safely move the epochs
+ * forward.  Only use this from the update callback.
+ */
+static inline void update_sched_clock(struct clock_data *cd, u32 cyc, u32 mask)
+{
+	unsigned long flags;
+	u64 ns = cd->epoch_ns +
+		cyc_to_ns((cyc - cd->epoch_cyc) & mask, cd->mult, cd->shift);
+
+	/*
+	 * Write epoch_cyc and epoch_ns in a way that the update is
+	 * detectable in cyc_to_fixed_sched_clock().
+	 */
+	raw_local_irq_save(flags);
+	cd->epoch_cyc = cyc;
+	smp_wmb();
+	cd->epoch_ns = ns;
+	smp_wmb();
+	cd->epoch_cyc_copy = cyc;
+	raw_local_irq_restore(flags);
+}
+
+/*
+ * If your clock rate is known at compile time, using this will allow
+ * you to optimize the mult/shift loads away.  This is paired with
+ * init_fixed_sched_clock() to ensure that your mult/shift are correct.
+ */
+static inline unsigned long long cyc_to_fixed_sched_clock(struct clock_data *cd,
+	u32 cyc, u32 mask, u32 mult, u32 shift)
+{
+	u64 epoch_ns;
+	u32 epoch_cyc;
+
+	/*
+	 * Load the epoch_cyc and epoch_ns atomically.  We do this by
+	 * ensuring that we always write epoch_cyc, epoch_ns and
+	 * epoch_cyc_copy in strict order, and read them in strict order.
+	 * If epoch_cyc and epoch_cyc_copy are not equal, then we're in
+	 * the middle of an update, and we should repeat the load.
+	 */
+	do {
+		epoch_cyc = cd->epoch_cyc;
+		smp_rmb();
+		epoch_ns = cd->epoch_ns;
+		smp_rmb();
+	} while (epoch_cyc != cd->epoch_cyc_copy);
+
+	return epoch_ns + cyc_to_ns((cyc - epoch_cyc) & mask, mult, shift);
+}
+
+/*
+ * Otherwise, you need to use this, which will obtain the mult/shift
+ * from the clock_data structure.  Use init_sched_clock() with this.
+ */
+static inline unsigned long long cyc_to_sched_clock(struct clock_data *cd,
+	u32 cyc, u32 mask)
+{
+	return cyc_to_fixed_sched_clock(cd, cyc, mask, cd->mult, cd->shift);
+}
+
+/*
+ * Initialize the clock data - calculate the appropriate multiplier
+ * and shift.  Also setup a timer to ensure that the epoch is refreshed
+ * at the appropriate time interval, which will call your update
+ * handler.
+ */
+void init_sched_clock(struct clock_data *, void (*)(void),
+	unsigned int, unsigned long);
+
+/*
+ * Use this initialization function rather than init_sched_clock() if
+ * you're using cyc_to_fixed_sched_clock, which will warn if your
+ * constants are incorrect.
+ */
+static inline void init_fixed_sched_clock(struct clock_data *cd,
+	void (*update)(void), unsigned int bits, unsigned long rate,
+	u32 mult, u32 shift)
+{
+	init_sched_clock(cd, update, bits, rate);
+	if (cd->mult != mult || cd->shift != shift) {
+		pr_crit("sched_clock: wrong multiply/shift: %u>>%u vs calculated %u>>%u\n"
+			"sched_clock: fix multiply/shift to avoid scheduler hiccups\n",
+			mult, shift, cd->mult, cd->shift);
+	}
+}
+
+#endif

arch/arm/kernel/Makefile

@@ -29,6 +29,7 @@ obj-$(CONFIG_MODULES)		+= armksyms.o module.o
 obj-$(CONFIG_ARTHUR)		+= arthur.o
 obj-$(CONFIG_ISA_DMA)		+= dma-isa.o
 obj-$(CONFIG_PCI)		+= bios32.o isa.o
+obj-$(CONFIG_HAVE_SCHED_CLOCK)	+= sched_clock.o
 obj-$(CONFIG_SMP)		+= smp.o
 obj-$(CONFIG_HAVE_ARM_SCU)	+= smp_scu.o
 obj-$(CONFIG_HAVE_ARM_TWD)	+= smp_twd.o

arch/arm/kernel/sched_clock.c

@@ -0,0 +1,69 @@
+/*
+ * sched_clock.c: support for extending counters to full 64-bit ns counter
+ *
+ * This program is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License version 2 as
+ * published by the Free Software Foundation.
+ */
+#include <linux/clocksource.h>
+#include <linux/init.h>
+#include <linux/jiffies.h>
+#include <linux/kernel.h>
+#include <linux/sched.h>
+#include <linux/timer.h>
+
+#include <asm/sched_clock.h>
+
+static void sched_clock_poll(unsigned long wrap_ticks);
+static DEFINE_TIMER(sched_clock_timer, sched_clock_poll, 0, 0);
+static void (*sched_clock_update_fn)(void);
+
+static void sched_clock_poll(unsigned long wrap_ticks)
+{
+	mod_timer(&sched_clock_timer, round_jiffies(jiffies + wrap_ticks));
+	sched_clock_update_fn();
+}
+
+void __init init_sched_clock(struct clock_data *cd, void (*update)(void),
+	unsigned int clock_bits, unsigned long rate)
+{
+	unsigned long r, w;
+	u64 res, wrap;
+	char r_unit;
+
+	sched_clock_update_fn = update;
+
+	/* calculate the mult/shift to convert counter ticks to ns. */
+	clocks_calc_mult_shift(&cd->mult, &cd->shift, rate, NSEC_PER_SEC, 60);
+
+	r = rate;
+	if (r >= 4000000) {
+		r /= 1000000;
+		r_unit = 'M';
+	} else {
+		r /= 1000;
+		r_unit = 'k';
+	}
+
+	/* calculate how many ns until we wrap */
+	wrap = cyc_to_ns((1ULL << clock_bits) - 1, cd->mult, cd->shift);
+	do_div(wrap, NSEC_PER_MSEC);
+	w = wrap;
+
+	/* calculate the ns resolution of this counter */
+	res = cyc_to_ns(1ULL, cd->mult, cd->shift);
+	pr_info("sched_clock: %u bits at %lu%cHz, resolution %lluns, wraps every %lums\n",
+		clock_bits, r, r_unit, res, w);
+
+	/*
+	 * Start the timer to keep sched_clock() properly updated and
+	 * sets the initial epoch.
+	 */
+	sched_clock_timer.data = msecs_to_jiffies(w - (w / 10));
+	sched_clock_poll(sched_clock_timer.data);
+
+	/*
+	 * Ensure that sched_clock() starts off at 0ns
+	 */
+	cd->epoch_ns = 0;
+}