[PATCH 0/3] Convert acked !GENERIC_TIME architectures to use the generic timekeeping core.

[PATCH 0/3] Convert acked !GENERIC_TIME architectures to use the generic timekeeping core.

[john stultz]

Andrew, All,

	Its been two and a half years since the GENERIC_TIME infrastructure
landed. It allowed architectures that have free running counters to only
have to provide a clocksource driver and leave the rest of the
timekeeping management to generic code.

So far out of the 23 architectures we've fully converted 15 of them over
to GENERIC_TIME. The remaining 8 have in some cases been able to
convert, but not fully. The issue that keeps most architectures from
converting is that they do not have free running counters. They instead
use the counter on their tick timer to determine their inter-tick time.
This counter usually wraps every interrupt, so its not really efficient
to use as a clocksource.

This has created some difficulty as the generic timekeeping core has had
to deal with both GENERIC_TIME arches as well as !GENERIC_TIME arches,
and since it didn't manage 100% of the calculation in the !GENERIC_TIME
case, we've been limited in some of the changes we could make.

So since these last remaining architectures are unlikely to be able to
fully convert to GENERIC_TIME as it stands, the generic core should try
to adopt to their needs so we can bring them into the fold and reduce
code duplication.

The solution is to provide a arch callout from the generic timekeeping
core: arch_gettimeoffset().  On architectures that do not provide a
clocksource, the jiffies clocksource is used, and arch_gettimeoffset is
added in to create finer grained inter-tick time.


The patchset I'm submitting does the following:
1) Introduces the arch_gettimeoffset infrastructure to the timekeeping
core.
2) Converts two Acked arches (sh, sparc) to use the arch_gettimeoffset()
hook.

I still have patches for the 6 other arches that need to be converted,
but have not received acks for them yet. Instead of stalling the
patchset for all of those arches, I'm submitting these, so overloaded
arch maintainers can convert their arches or take my patches on their
schedule without blocking arches that are ready to go.

Once all 8 arches are converted, I'll send a final cleanup patch that
removes all conditionals and references to GENERIC_TIME.

Credits to Roman for initially suggesting the idea quite awhile back.

thanks
-john

[PATCH 1/3] Create arch_gettimeoffset infrastructure for use in the generic timekeeping core.

[john stultz]

Andrew, please consider for inclusion in -mm in preparation for 2.6.30.

Some arches don't supply their own clocksource. This is mainly the case
in architectures that get their inter-tick times by reading the counter
on their interval timer. Since these timers wrap every tick, they're not
really useful as clocksources. Wrapping them to act like one is possible
but not very efficient. So we provide a callout these arches  can
implement for use with the jiffies clocksource to provide finer then
tick granular time.

Signed-off-by: John Stultz <johnstul@us.ibm.com>


diff --git a/include/linux/time.h b/include/linux/time.h
index fbbd2a1..4768f16 100644
--- a/include/linux/time.h
+++ b/include/linux/time.h
@@ -113,6 +113,21 @@ struct timespec current_kernel_time(void);
 #define CURRENT_TIME		(current_kernel_time())
 #define CURRENT_TIME_SEC	((struct timespec) { get_seconds(), 0 })
 
+/* Some architectures do not supply their own clocksource.
+ * This is mainly the case in architectures that get their
+ * inter-tick times by reading the counter on their interval
+ * timer. Since these timers wrap every tick, they're not really
+ * useful as clocksources. Wrapping them to act like one is possible
+ * but not very efficient. So we provide a callout these arches
+ * can implement for use with the jiffies clocksource to provide
+ * finer then tick granular time.
+ */
+#ifdef CONFIG_ARCH_USES_GETTIMEOFFSET
+extern u32 arch_gettimeoffset(void);
+#else
+#define arch_gettimeoffset() (0)
+#endif
+
 extern void do_gettimeofday(struct timeval *tv);
 extern int do_settimeofday(struct timespec *tv);
 extern int do_sys_settimeofday(struct timespec *tv, struct timezone *tz);
diff --git a/kernel/time/timekeeping.c b/kernel/time/timekeeping.c
index 900f1b6..45d777f 100644
--- a/kernel/time/timekeeping.c
+++ b/kernel/time/timekeeping.c
@@ -77,6 +77,10 @@ static void clocksource_forward_now(void)
 	clock->cycle_last = cycle_now;
 
 	nsec = cyc2ns(clock, cycle_delta);
+
+	/* If arch requires, add in gettimeoffset() */
+	nsec += arch_gettimeoffset();
+
 	timespec_add_ns(&xtime, nsec);
 
 	nsec = ((s64)cycle_delta * clock->mult_orig) >> clock->shift;
@@ -111,6 +115,9 @@ void getnstimeofday(struct timespec *ts)
 		/* convert to nanoseconds: */
 		nsecs = cyc2ns(clock, cycle_delta);
 
+		/* If arch requires, add in gettimeoffset() */
+		nsecs += arch_gettimeoffset();
+
 	} while (read_seqretry(&xtime_lock, seq));
 
 	timespec_add_ns(ts, nsecs);

[PATCH 2/3] Convert sh to use arch_getoffset() infrastructure

[john stultz]

Andrew, please consider for inclusion into -mm in preparation for 2.6.30

This patch converts sh to use GENERIC_TIME via the arch_getoffset()
infrastructure 

I do not have cross compilers for these architectures, and in some cases
the architectures can be compiles both with and without clocksources. So
I've taken my best swing at converting this, but I'm not confident I got
it right. Any assistance from arch maintainers or testers would be
great.

Signed-off-by: John Stultz <johnstul@us.ibm.com>
Acked-by: Paul Mundt <lethal@linux-sh.org>

diff --git a/arch/sh/Kconfig b/arch/sh/Kconfig
index ebabe51..c48ab20 100644
--- a/arch/sh/Kconfig
+++ b/arch/sh/Kconfig
@@ -72,7 +72,7 @@ config GENERIC_IOMAP
 	bool
 
 config GENERIC_TIME
-	def_bool n
+	def_bool y
 
 config GENERIC_CLOCKEVENTS
 	def_bool n
@@ -406,6 +406,10 @@ config SH_TMU
 	help
 	  This enables the use of the TMU as the system timer.
 
+config ARCH_USES_GETTIMEOFFSET
+	def_bool y
+	depends on !SH_TMU
+
 config SH_CMT
 	def_bool y
 	prompt "CMT timer support"
diff --git a/arch/sh/include/asm/timer.h b/arch/sh/include/asm/timer.h
index a7ca3a1..da122a3 100644
--- a/arch/sh/include/asm/timer.h
+++ b/arch/sh/include/asm/timer.h
@@ -10,7 +10,7 @@ struct sys_timer_ops {
 	int (*start)(void);
 	int (*stop)(void);
 	cycle_t (*read)(void);
-#ifndef CONFIG_GENERIC_TIME
+#ifdef CONFIG_ARCH_USES_GETTIMEOFFSET
 	unsigned long (*get_offset)(void);
 #endif
 };
@@ -27,7 +27,7 @@ struct sys_timer {
 extern struct sys_timer tmu_timer, cmt_timer, mtu2_timer;
 extern struct sys_timer *sys_timer;
 
-#ifndef CONFIG_GENERIC_TIME
+#ifdef CONFIG_ARCH_USES_GETTIMEOFFSET
 static inline unsigned long get_timer_offset(void)
 {
 	return sys_timer->ops->get_offset();
diff --git a/arch/sh/kernel/time_32.c b/arch/sh/kernel/time_32.c
index 8457f83..8e9827a 100644
--- a/arch/sh/kernel/time_32.c
+++ b/arch/sh/kernel/time_32.c
@@ -52,65 +52,12 @@ static cycle_t null_hpt_read(void)
 void (*rtc_sh_get_time)(struct timespec *) = null_rtc_get_time;
 int (*rtc_sh_set_time)(const time_t) = null_rtc_set_time;
 
-#ifndef CONFIG_GENERIC_TIME
-void do_gettimeofday(struct timeval *tv)
+#ifdef CONFIG_ARCH_USES_GETTIMEOFFSET
+u32 arch_gettimeoffset(void)
 {
-	unsigned long flags;
-	unsigned long seq;
-	unsigned long usec, sec;
-
-	do {
-		/*
-		 * Turn off IRQs when grabbing xtime_lock, so that
-		 * the sys_timer get_offset code doesn't have to handle it.
-		 */
-		seq = read_seqbegin_irqsave(&xtime_lock, flags);
-		usec = get_timer_offset();
-		sec = xtime.tv_sec;
-		usec += xtime.tv_nsec / NSEC_PER_USEC;
-	} while (read_seqretry_irqrestore(&xtime_lock, seq, flags));
-
-	while (usec >= 1000000) {
-		usec -= 1000000;
-		sec++;
-	}
-
-	tv->tv_sec = sec;
-	tv->tv_usec = usec;
-}
-EXPORT_SYMBOL(do_gettimeofday);
-
-int do_settimeofday(struct timespec *tv)
-{
-	time_t wtm_sec, sec = tv->tv_sec;
-	long wtm_nsec, nsec = tv->tv_nsec;
-
-	if ((unsigned long)tv->tv_nsec >= NSEC_PER_SEC)
-		return -EINVAL;
-
-	write_seqlock_irq(&xtime_lock);
-	/*
-	 * This is revolting. We need to set "xtime" correctly. However, the
-	 * value in this location is the value at the most recent update of
-	 * wall time.  Discover what correction gettimeofday() would have
-	 * made, and then undo it!
-	 */
-	nsec -= get_timer_offset() * NSEC_PER_USEC;
-
-	wtm_sec  = wall_to_monotonic.tv_sec + (xtime.tv_sec - sec);
-	wtm_nsec = wall_to_monotonic.tv_nsec + (xtime.tv_nsec - nsec);
-
-	set_normalized_timespec(&xtime, sec, nsec);
-	set_normalized_timespec(&wall_to_monotonic, wtm_sec, wtm_nsec);
-
-	ntp_clear();
-	write_sequnlock_irq(&xtime_lock);
-	clock_was_set();
-
-	return 0;
+	return get_timer_offset() * 1000;
 }
-EXPORT_SYMBOL(do_settimeofday);
-#endif /* !CONFIG_GENERIC_TIME */
+#endif /* CONFIG_ARCH_USES_GETTIMEOFFSET */
 
 #ifndef CONFIG_GENERIC_CLOCKEVENTS
 /* last time the RTC clock got updated */
@@ -231,7 +178,7 @@ static void __init init_sh_clocksource(void)
 	clocksource_register(&clocksource_sh);
 }
 
-#ifdef CONFIG_GENERIC_TIME
+#ifndef CONFIG_ARCH_USES_GETTIMEOFFSET
 unsigned long long sched_clock(void)
 {
 	unsigned long long ticks = clocksource_sh.read();
diff --git a/arch/sh/kernel/time_64.c b/arch/sh/kernel/time_64.c
index 59d2a03..cfd75d8 100644
--- a/arch/sh/kernel/time_64.c
+++ b/arch/sh/kernel/time_64.c
@@ -144,59 +144,10 @@ static unsigned long usecs_since_tick(void)
 	return result;
 }
 
-void do_gettimeofday(struct timeval *tv)
+u32 arch_gettimeoffset(void)
 {
-	unsigned long flags;
-	unsigned long seq;
-	unsigned long usec, sec;
-
-	do {
-		seq = read_seqbegin_irqsave(&xtime_lock, flags);
-		usec = usecs_since_tick();
-		sec = xtime.tv_sec;
-		usec += xtime.tv_nsec / 1000;
-	} while (read_seqretry_irqrestore(&xtime_lock, seq, flags));
-
-	while (usec >= 1000000) {
-		usec -= 1000000;
-		sec++;
-	}
-
-	tv->tv_sec = sec;
-	tv->tv_usec = usec;
-}
-EXPORT_SYMBOL(do_gettimeofday);
-
-int do_settimeofday(struct timespec *tv)
-{
-	time_t wtm_sec, sec = tv->tv_sec;
-	long wtm_nsec, nsec = tv->tv_nsec;
-
-	if ((unsigned long)tv->tv_nsec >= NSEC_PER_SEC)
-		return -EINVAL;
-
-	write_seqlock_irq(&xtime_lock);
-	/*
-	 * This is revolting. We need to set "xtime" correctly. However, the
-	 * value in this location is the value at the most recent update of
-	 * wall time.  Discover what correction gettimeofday() would have
-	 * made, and then undo it!
-	 */
-	nsec -= 1000 * usecs_since_tick();
-
-	wtm_sec  = wall_to_monotonic.tv_sec + (xtime.tv_sec - sec);
-	wtm_nsec = wall_to_monotonic.tv_nsec + (xtime.tv_nsec - nsec);
-
-	set_normalized_timespec(&xtime, sec, nsec);
-	set_normalized_timespec(&wall_to_monotonic, wtm_sec, wtm_nsec);
-
-	ntp_clear();
-	write_sequnlock_irq(&xtime_lock);
-	clock_was_set();
-
-	return 0;
+	return usecs_since_tick() * 1000;
 }
-EXPORT_SYMBOL(do_settimeofday);
 
 /* Dummy RTC ops */
 static void null_rtc_get_time(struct timespec *tv)
diff --git a/arch/sh/kernel/timers/timer-cmt.c b/arch/sh/kernel/timers/timer-cmt.c
index c127293..6c729db 100644
--- a/arch/sh/kernel/timers/timer-cmt.c
+++ b/arch/sh/kernel/timers/timer-cmt.c
@@ -178,7 +178,7 @@ static struct sys_timer_ops cmt_timer_ops = {
 	.init		= cmt_timer_init,
 	.start		= cmt_timer_start,
 	.stop		= cmt_timer_stop,
-#ifndef CONFIG_GENERIC_TIME
+#ifdef CONFIG_ARCH_USES_GETTIMEOFFSET
 	.get_offset	= cmt_timer_get_offset,
 #endif
 };
diff --git a/arch/sh/kernel/timers/timer-mtu2.c b/arch/sh/kernel/timers/timer-mtu2.c
index c3d237e..acf6a96 100644
--- a/arch/sh/kernel/timers/timer-mtu2.c
+++ b/arch/sh/kernel/timers/timer-mtu2.c
@@ -191,7 +191,7 @@ struct sys_timer_ops mtu2_timer_ops = {
 	.init		= mtu2_timer_init,
 	.start		= mtu2_timer_start,
 	.stop		= mtu2_timer_stop,
-#ifndef CONFIG_GENERIC_TIME
+#ifdef CONFIG_ARCH_USES_GETTIMEOFFSET
 	.get_offset	= mtu2_timer_get_offset,
 #endif
 };

[PATCH 3/3] Convert sparc to use arch_getoffset() infrastructure

[john stultz]

Andrew, please consider for inclusion into -mm in preparation for 2.6.30

This patch converts sparc to use GENERIC_TIME via the arch_getoffset()
infrastructure 

I do not have cross compilers for these architectures, and in some cases
the architectures can be compiles both with and without clocksources. So
I've taken my best swing at converting this, but I'm not confident I got
it right. Any assistance from arch maintainers or testers would be
great.

Signed-off-by: John Stultz <johnstul@us.ibm.com>
Acked-by: David S. Miller <davem@davemloft.net>

diff --git a/arch/sparc/Kconfig b/arch/sparc/Kconfig
index c3ea215..a5b465f 100644
--- a/arch/sparc/Kconfig
+++ b/arch/sparc/Kconfig
@@ -55,8 +55,11 @@ config BITS
 	default 64 if SPARC64
 
 config GENERIC_TIME
+	def_bool y
+
+config ARCH_USES_GETTIMEOFFSET
 	bool
-	default y if SPARC64
+	default y if SPARC32
 
 config GENERIC_CMOS_UPDATE
 	bool
diff --git a/arch/sparc/kernel/time_32.c b/arch/sparc/kernel/time_32.c
index 614ac7b..e946b55 100644
--- a/arch/sparc/kernel/time_32.c
+++ b/arch/sparc/kernel/time_32.c
@@ -227,7 +227,7 @@ void __init time_init(void)
 	sbus_time_init();
 }
 
-static inline unsigned long do_gettimeoffset(void)
+u32 arch_gettimeoffset(void)
 {
 	unsigned long val = *master_l10_counter;
 	unsigned long usec = (val >> 10) & 0x1fffff;
@@ -236,84 +236,7 @@ static inline unsigned long do_gettimeoffset(void)
 	if (val & 0x80000000)
 		usec += 1000000 / HZ;
 
-	return usec;
-}
-
-/* Ok, my cute asm atomicity trick doesn't work anymore.
- * There are just too many variables that need to be protected
- * now (both members of xtime, et al.)
- */
-void do_gettimeofday(struct timeval *tv)
-{
-	unsigned long flags;
-	unsigned long seq;
-	unsigned long usec, sec;
-	unsigned long max_ntp_tick = tick_usec - tickadj;
-
-	do {
-		seq = read_seqbegin_irqsave(&xtime_lock, flags);
-		usec = do_gettimeoffset();
-
-		/*
-		 * If time_adjust is negative then NTP is slowing the clock
-		 * so make sure not to go into next possible interval.
-		 * Better to lose some accuracy than have time go backwards..
-		 */
-		if (unlikely(time_adjust < 0))
-			usec = min(usec, max_ntp_tick);
-
-		sec = xtime.tv_sec;
-		usec += (xtime.tv_nsec / 1000);
-	} while (read_seqretry_irqrestore(&xtime_lock, seq, flags));
-
-	while (usec >= 1000000) {
-		usec -= 1000000;
-		sec++;
-	}
-
-	tv->tv_sec = sec;
-	tv->tv_usec = usec;
-}
-
-EXPORT_SYMBOL(do_gettimeofday);
-
-int do_settimeofday(struct timespec *tv)
-{
-	int ret;
-
-	write_seqlock_irq(&xtime_lock);
-	ret = bus_do_settimeofday(tv);
-	write_sequnlock_irq(&xtime_lock);
-	clock_was_set();
-	return ret;
-}
-
-EXPORT_SYMBOL(do_settimeofday);
-
-static int sbus_do_settimeofday(struct timespec *tv)
-{
-	time_t wtm_sec, sec = tv->tv_sec;
-	long wtm_nsec, nsec = tv->tv_nsec;
-
-	if ((unsigned long)tv->tv_nsec >= NSEC_PER_SEC)
-		return -EINVAL;
-
-	/*
-	 * This is revolting. We need to set "xtime" correctly. However, the
-	 * value in this location is the value at the most recent update of
-	 * wall time.  Discover what correction gettimeofday() would have
-	 * made, and then undo it!
-	 */
-	nsec -= 1000 * do_gettimeoffset();
-
-	wtm_sec  = wall_to_monotonic.tv_sec + (xtime.tv_sec - sec);
-	wtm_nsec = wall_to_monotonic.tv_nsec + (xtime.tv_nsec - nsec);
-
-	set_normalized_timespec(&xtime, sec, nsec);
-	set_normalized_timespec(&wall_to_monotonic, wtm_sec, wtm_nsec);
-
-	ntp_clear();
-	return 0;
+	return usec * 1000;
 }
 
 static int set_rtc_mmss(unsigned long secs)