[RFC][PATCH 0/4] new human-time soft-timer subsystem

[RFC][PATCH 0/4] new human-time soft-timer subsystem

[Nishanth Aravamudan]

On 14.07.2005 [12:18:41 -0700], john stultz wrote:

<snip>

> Nish has some code, which I hope he'll be sending out shortly that
> does just this, converting the soft-timer subsystem to use absolute
> time instead of ticks for expiration. I feel it both simplifies the
> code and makes it easier to changing the timer interrupt frequency
> while the system is running.

Here's the set of patches John promised :)

1/4: add jiffies conversion helper functions

2/4: core human-time modifications to soft-timer subsystem

3/4: add new human-time schedule_timeout() functions

4/4: rework sys_nanosleep() to use schedule_timeout_nsecs()

The individual patches have more details, but the gist is this:

We no longer use jiffies (the variable) as the basis for determining
what "time" a timer should expire or when it should be added. Instead,
we use a new function, do_monotonic_clock(), which is simply a wrapper
for getnstimeofday(). That is to say, we use uptime in nanoseconds. But,
to avoid modifying the existing soft-timer algorithm, we convert the
64-bit nanosecond value to "timerinterval" units. These units are simply
2^TIMEINTERVAL_BITS nanoseconds in length (thus determined at compile
time).

To sum up, soft-timers now use time (as defined by the
timeofday-subsystem) not ticks. Hopefully, the individual
e-mails/patches make this change clear.

Thanks,
Nish

[RFC][PATCH 1/4] add jiffies_to_nsecs() helper and fix up size of usecs

[Nishanth Aravamudan]

From: Nishanth Aravamudan <nacc@us.ibm.com>

Description: Add a jiffies_to_nsecs() helper function. Make consistent
the size of microseconds (unsigned long) throughout the conversion
functions.

Signed-off-by: Nishanth Aravamudan <nacc@us.ibm.com>

---

 jiffies.h |   15 +++++++++++++--
 1 files changed, 13 insertions(+), 2 deletions(-)

diff -urpN 2.6.13-rc3-base/include/linux/jiffies.h 2.6.13-rc3-dev/include/linux/jiffies.h
--- 2.6.13-rc3-base/include/linux/jiffies.h	2005-03-01 23:37:31.000000000 -0800
+++ 2.6.13-rc3-dev/include/linux/jiffies.h	2005-07-14 12:43:44.000000000 -0700
@@ -263,7 +263,7 @@ static inline unsigned int jiffies_to_ms
 #endif
 }
 
-static inline unsigned int jiffies_to_usecs(const unsigned long j)
+static inline unsigned long jiffies_to_usecs(const unsigned long j)
 {
 #if HZ <= 1000000 && !(1000000 % HZ)
 	return (1000000 / HZ) * j;
@@ -274,6 +274,17 @@ static inline unsigned int jiffies_to_us
 #endif
 }
 
+static inline u64 jiffies_to_nsecs(const unsigned long j)
+{
+#if HZ <= NSEC_PER_SEC && !(NSEC_PER_SEC % HZ)
+	return (NSEC_PER_SEC / HZ) * (u64)j;
+#elif HZ > NSEC_PER_SEC && !(HZ % NSEC_PER_SEC)
+	return ((u64)j + (HZ / NSEC_PER_SEC) - 1)/(HZ / NSEC_PER_SEC);
+#else
+	return ((u64)j * NSEC_PER_SEC) / HZ;
+#endif
+}
+
 static inline unsigned long msecs_to_jiffies(const unsigned int m)
 {
 	if (m > jiffies_to_msecs(MAX_JIFFY_OFFSET))
@@ -287,7 +298,7 @@ static inline unsigned long msecs_to_jif
 #endif
 }
 
-static inline unsigned long usecs_to_jiffies(const unsigned int u)
+static inline unsigned long usecs_to_jiffies(const unsigned long u)
 {
 	if (u > jiffies_to_usecs(MAX_JIFFY_OFFSET))
 		return MAX_JIFFY_OFFSET;

[RFC][PATCH 2/4] human-time soft-timer core changes

[Nishanth Aravamudan]

From: Nishanth Aravamudan <nacc@us.ibm.com>

Description: The core revision to the soft-timer subsystem to divorce it
from the timer interrupt in software, i.e. jiffies. Instead, use
getnstimeofday() (via do_monotonic_clock()) as the basis for addition
and expiration of timers. Add a new unit, the timerinterval, which is
a 2^TIMERINTERVAL_BITS nanoseconds in length. The converted value in
timerintervals is used where we would have used the timer's expires
member before. Add set_timer_nsecs() and set_timer_nsecs_on() functions
to directly request nanosecond delays. These functions replace
add_timer(), mod_timer() and add_timer_on().

Signed-off-by: Nishanth Aravamudan <nacc@us.ibm.com>

---

 include/linux/time.h  |    1 
 include/linux/timer.h |   27 +-----
 kernel/time.c         |   18 ++++
 kernel/timer.c        |  215 +++++++++++++++++++++++++++++++++++++++++++++-----
 4 files changed, 220 insertions(+), 41 deletions(-)

diff -urpN 2.6.13-rc3-base/include/linux/time.h 2.6.13-rc3-dev/include/linux/time.h
--- 2.6.13-rc3-base/include/linux/time.h	2005-03-01 23:38:12.000000000 -0800
+++ 2.6.13-rc3-dev/include/linux/time.h	2005-07-14 12:44:40.000000000 -0700
@@ -103,6 +103,7 @@ struct itimerval;
 extern int do_setitimer(int which, struct itimerval *value, struct itimerval *ovalue);
 extern int do_getitimer(int which, struct itimerval *value);
 extern void getnstimeofday (struct timespec *tv);
+extern u64 do_monotonic_clock(void);
 
 extern struct timespec timespec_trunc(struct timespec t, unsigned gran);
 
diff -urpN 2.6.13-rc3-base/include/linux/timer.h 2.6.13-rc3-dev/include/linux/timer.h
--- 2.6.13-rc3-base/include/linux/timer.h	2005-07-13 15:52:14.000000000 -0700
+++ 2.6.13-rc3-dev/include/linux/timer.h	2005-07-14 12:44:40.000000000 -0700
@@ -11,6 +11,7 @@ struct timer_base_s;
 struct timer_list {
 	struct list_head entry;
 	unsigned long expires;
+	u64 expires_nsecs;
 
 	unsigned long magic;
 
@@ -27,6 +28,7 @@ extern struct timer_base_s __init_timer_
 #define TIMER_INITIALIZER(_function, _expires, _data) {		\
 		.function = (_function),			\
 		.expires = (_expires),				\
+		.expires_nsecs = 0,				\
 		.data = (_data),				\
 		.base = &__init_timer_base,			\
 		.magic = TIMER_MAGIC,				\
@@ -51,30 +53,15 @@ static inline int timer_pending(const st
 
 extern void add_timer_on(struct timer_list *timer, int cpu);
 extern int del_timer(struct timer_list * timer);
-extern int __mod_timer(struct timer_list *timer, unsigned long expires);
+extern int __mod_timer(struct timer_list *timer);
 extern int mod_timer(struct timer_list *timer, unsigned long expires);
+extern void add_timer(struct timer_list *timer);
+extern int set_timer_nsecs(struct timer_list *timer, u64 expires_nsecs);
+extern void set_timer_on_nsecs(struct timer_list *timer, u64 expires_nsecs,
+								int cpu);
 
 extern unsigned long next_timer_interrupt(void);
 
-/***
- * add_timer - start a timer
- * @timer: the timer to be added
- *
- * The kernel will do a ->function(->data) callback from the
- * timer interrupt at the ->expired point in the future. The
- * current time is 'jiffies'.
- *
- * The timer's ->expired, ->function (and if the handler uses it, ->data)
- * fields must be set prior calling this function.
- *
- * Timers with an ->expired field in the past will be executed in the next
- * timer tick.
- */
-static inline void add_timer(struct timer_list * timer)
-{
-	__mod_timer(timer, timer->expires);
-}
-
 #ifdef CONFIG_SMP
   extern int try_to_del_timer_sync(struct timer_list *timer);
   extern int del_timer_sync(struct timer_list *timer);
diff -urpN 2.6.13-rc3-base/kernel/time.c 2.6.13-rc3-dev/kernel/time.c
--- 2.6.13-rc3-base/kernel/time.c	2005-07-13 15:51:57.000000000 -0700
+++ 2.6.13-rc3-dev/kernel/time.c	2005-07-14 12:44:40.000000000 -0700
@@ -589,3 +589,21 @@ EXPORT_SYMBOL(get_jiffies_64);
 #endif
 
 EXPORT_SYMBOL(jiffies);
+
+u64 do_monotonic_clock(void)
+{
+	struct timespec now, now_w2m;
+	unsigned long seq;
+
+	getnstimeofday(&now);
+
+	do {
+		seq = read_seqbegin(&xtime_lock);
+		now_w2m = wall_to_monotonic;
+	} while (read_seqretry(&xtime_lock, seq));
+
+	return (u64)(now.tv_sec + now_w2m.tv_sec) * NSEC_PER_SEC +
+				(now.tv_nsec + now_w2m.tv_nsec);
+}
+
+EXPORT_SYMBOL_GPL(do_monotonic_clock);
diff -urpN 2.6.13-rc3-base/kernel/timer.c 2.6.13-rc3-dev/kernel/timer.c
--- 2.6.13-rc3-base/kernel/timer.c	2005-07-13 15:52:14.000000000 -0700
+++ 2.6.13-rc3-dev/kernel/timer.c	2005-07-14 12:44:40.000000000 -0700
@@ -56,6 +56,15 @@ static void time_interpolator_update(lon
 #define TVR_SIZE (1 << TVR_BITS)
 #define TVN_MASK (TVN_SIZE - 1)
 #define TVR_MASK (TVR_SIZE - 1)
+/*
+ * Modifying TIMERINTERVAL_BITS changes the software resolution of
+ * soft-timers. While 20 bits would be closer to a millisecond, there
+ * are performance gains from allowing a software resolution finer than
+ * the hardware (HZ=1000)
+ */
+#define TIMERINTERVAL_BITS 19
+#define TIMERINTERVAL_SIZE (1 << TIMERINTERVAL_BITS)
+#define TIMERINTERVAL_MASK (TIMERINTERVAL_SIZE - 1)
 
 struct timer_base_s {
 	spinlock_t lock;
@@ -72,7 +81,7 @@ typedef struct tvec_root_s {
 
 struct tvec_t_base_s {
 	struct timer_base_s t_base;
-	unsigned long timer_jiffies;
+	unsigned long last_timer_time;
 	tvec_root_t tv1;
 	tvec_t tv2;
 	tvec_t tv3;
@@ -114,11 +123,88 @@ static inline void check_timer(struct ti
 		check_timer_failed(timer);
 }
 
+/*
+ * nsecs_to_timerintervals_ceiling - convert nanoseconds to timerintervals
+ * @n: number of nanoseconds to convert
+ *
+ * This is where changes to TIMERINTERVAL_BITS affect the soft-timer
+ * subsystem.
+ *
+ * Some explanation of the math is necessary:
+ * Rather than do decimal arithmetic, we shift for the sake of speed.
+ * This does mean that the actual requestable sleeps are
+ * 	2^(sizeof(unsigned long)*8 - TIMERINTERVAL_BITS)
+ * timerintervals.
+ *
+ * The conditional takes care of the corner case where we request a 0
+ * nanosecond sleep; if the quantity were unsigned, we would not
+ * propogate the carry and force a wrap when adding the 1.
+ *
+ * To prevent timers from being expired early, we:
+ *	Take the ceiling when we add; and
+ *	Take the floor when we expire.
+ */
+static inline unsigned long nsecs_to_timerintervals_ceiling(u64 nsecs)
+{
+	if (nsecs)
+		return (unsigned long)(((nsecs - 1) >> TIMERINTERVAL_BITS) + 1);
+	else
+		return 0UL;
+}
+
+/*
+ * nsecs_to_timerintervals_floor - convert nanoseconds to timerintervals
+ * @n: number of nanoseconds to convert
+ *
+ * This is where changes to TIMERINTERVAL_BITS affect the soft-timer
+ * subsystem.
+ *
+ * Some explanation of the math is necessary:
+ * Rather than do decimal arithmetic, we shift for the sake of speed.
+ * This does mean that the actual requestable sleeps are
+ * 	2^(sizeof(unsigned long)*8 - TIMERINTERVAL_BITS)
+ *
+ * There is no special case for 0 in the floor function, since we do not
+ * do any subtraction or addition of 1
+ *
+ * To prevent timers from being expired early, we:
+ *	Take the ceiling when we add; and
+ *	Take the floor when we expire.
+ */
+static inline unsigned long nsecs_to_timerintervals_floor(u64 nsecs)
+{
+	return (unsigned long)(nsecs >> TIMERINTERVAL_BITS);
+}
+
+/*
+ * jiffies_to_timerintervals - convert absolute jiffies to timerintervals
+ * @abs_jiffies: number of jiffies to convert
+ *
+ * First, we convert the absolute jiffies parameter to a relative
+ * jiffies value. To maintain precision, we convert the relative
+ * jiffies value to a relative nanosecond value and then convert that
+ * to a relative soft-timer interval unit value. We then add this
+ * relative value to the current time according to the timeofday-
+ * subsystem, converted to soft-timer interval units.
+ *
+ * We only use this function when adding timers, so we are free to
+ * always use the ceiling version of nsecs_to_timerintervals.
+ *
+ * This function only exists to support deprecated interfaces. Once
+ * those interfaces have been converted to the alternatives, it should
+ * be removed.
+ */
+static inline unsigned long jiffies_to_timerintervals(unsigned long abs_jiffies)
+{
+	unsigned long relative_jiffies = abs_jiffies - jiffies;
+	return nsecs_to_timerintervals_ceiling(do_monotonic_clock() +
+		jiffies_to_nsecs(relative_jiffies));
+}
 
 static void internal_add_timer(tvec_base_t *base, struct timer_list *timer)
 {
-	unsigned long expires = timer->expires;
-	unsigned long idx = expires - base->timer_jiffies;
+	unsigned long expires = nsecs_to_timerintervals_ceiling(timer->expires_nsecs);
+	unsigned long idx = expires - base->last_timer_time;
 	struct list_head *vec;
 
 	if (idx < TVR_SIZE) {
@@ -138,7 +224,7 @@ static void internal_add_timer(tvec_base
 		 * Can happen if you add a timer with expires == jiffies,
 		 * or you set a timer to go off in the past
 		 */
-		vec = base->tv1.vec + (base->timer_jiffies & TVR_MASK);
+		vec = base->tv1.vec + (base->last_timer_time & TVR_MASK);
 	} else {
 		int i;
 		/* If the timeout is larger than 0xffffffff on 64-bit
@@ -146,7 +232,7 @@ static void internal_add_timer(tvec_base
 		 */
 		if (idx > 0xffffffffUL) {
 			idx = 0xffffffffUL;
-			expires = idx + base->timer_jiffies;
+			expires = idx + base->last_timer_time;
 		}
 		i = (expires >> (TVR_BITS + 3 * TVN_BITS)) & TVN_MASK;
 		vec = base->tv5.vec + i;
@@ -222,7 +308,7 @@ static timer_base_t *lock_timer_base(str
 	}
 }
 
-int __mod_timer(struct timer_list *timer, unsigned long expires)
+int __mod_timer(struct timer_list *timer)
 {
 	timer_base_t *base;
 	tvec_base_t *new_base;
@@ -261,7 +347,7 @@ int __mod_timer(struct timer_list *timer
 		}
 	}
 
-	timer->expires = expires;
+	/* expires should be in timerintervals, and is currently ignored? */
 	internal_add_timer(new_base, timer);
 	spin_unlock_irqrestore(&new_base->t_base.lock, flags);
 
@@ -281,21 +367,50 @@ void add_timer_on(struct timer_list *tim
 {
 	tvec_base_t *base = &per_cpu(tvec_bases, cpu);
   	unsigned long flags;
-
+  
   	BUG_ON(timer_pending(timer) || !timer->function);
 
 	check_timer(timer);
 
 	spin_lock_irqsave(&base->t_base.lock, flags);
+	timer->expires_nsecs = do_monotonic_clock() +
+		jiffies_to_nsecs(timer->expires - jiffies);
 	timer->base = &base->t_base;
 	internal_add_timer(base, timer);
 	spin_unlock_irqrestore(&base->t_base.lock, flags);
 }
 
+/***
+ * add_timer - start a timer
+ * @timer: the timer to be added
+ *
+ * The kernel will do a ->function(->data) callback from the
+ * timer interrupt at the ->expired point in the future. The
+ * current time is 'jiffies'.
+ *
+ * The timer's ->expired, ->function (and if the handler uses it, ->data)
+ * fields must be set prior calling this function.
+ *
+ * Timers with an ->expired field in the past will be executed in the next
+ * timer tick.
+ *
+ * The callers of add_timer() should be aware that the interface is now
+ * deprecated. set_timer_nsecs() is the single interface for adding and
+ * modifying timers.
+ */
+void add_timer(struct timer_list * timer)
+{
+	timer->expires_nsecs = do_monotonic_clock() +
+		jiffies_to_nsecs(timer->expires - jiffies);
+	__mod_timer(timer);
+}
+
+EXPORT_SYMBOL(add_timer);
 
 /***
  * mod_timer - modify a timer's timeout
  * @timer: the timer to be modified
+ * @expires: absolute time, in jiffies, when timer should expire
  *
  * mod_timer is a more efficient way to update the expire field of an
  * active timer (if the timer is inactive it will be activated)
@@ -311,6 +426,10 @@ void add_timer_on(struct timer_list *tim
  * The function returns whether it has modified a pending timer or not.
  * (ie. mod_timer() of an inactive timer returns 0, mod_timer() of an
  * active timer returns 1.)
+ *
+ * The callers of mod_timer() should be aware that the interface is now
+ * deprecated. set_timer_nsecs() is the single interface for adding and
+ * modifying timers.
  */
 int mod_timer(struct timer_list *timer, unsigned long expires)
 {
@@ -318,6 +437,9 @@ int mod_timer(struct timer_list *timer, 
 
 	check_timer(timer);
 
+	timer->expires_nsecs = do_monotonic_clock() +
+		jiffies_to_nsecs(expires - jiffies);
+
 	/*
 	 * This is a common optimization triggered by the
 	 * networking code - if the timer is re-modified
@@ -326,10 +448,56 @@ int mod_timer(struct timer_list *timer, 
 	if (timer->expires == expires && timer_pending(timer))
 		return 1;
 
-	return __mod_timer(timer, expires);
+	return __mod_timer(timer);
 }
 
-EXPORT_SYMBOL(mod_timer);
+/*
+ * set_timer_nsecs - modify a timer's timeout in nsecs
+ * @timer: the timer to be modified
+ *
+ * set_timer_nsecs replaces both add_timer and mod_timer. The caller
+ * should call do_monotonic_clock() to determine the absolute timeout
+ * necessary.
+ */
+int set_timer_nsecs(struct timer_list *timer, u64 expires_nsecs)
+{
+	BUG_ON(!timer->function);
+
+	check_timer(timer);
+
+	if (timer_pending(timer) && timer->expires_nsecs == expires_nsecs)
+		return 1;
+
+	timer->expires_nsecs = expires_nsecs;
+
+	return __mod_timer(timer);
+}
+
+EXPORT_SYMBOL_GPL(set_timer_nsecs);
+
+/***
+ * set_timer_on_nsecs - start a timer on a particular CPU
+ * @timer: the timer to be added
+ * @expires_nsecs: absolute time in nsecs when timer should expire
+ * @cpu: the CPU to start it on
+ *
+ * This is not very scalable on SMP. Double adds are not possible.
+ */
+void set_timer_on_nsecs(struct timer_list *timer, u64 expires_nsecs, int cpu)
+{
+	tvec_base_t *base = &per_cpu(tvec_bases, cpu);
+  	unsigned long flags;
+
+  	BUG_ON(timer_pending(timer) || !timer->function);
+
+	check_timer(timer);
+
+	spin_lock_irqsave(&base->t_base.lock, flags);
+	timer->expires_nsecs = expires_nsecs;
+	timer->base = &base->t_base;
+	internal_add_timer(base, timer);
+	spin_unlock_irqrestore(&base->t_base.lock, flags);
+}
 
 /***
  * del_timer - deactive a timer.
@@ -455,17 +623,17 @@ static int cascade(tvec_base_t *base, tv
  * This function cascades all vectors and executes all expired timer
  * vectors.
  */
-#define INDEX(N) (base->timer_jiffies >> (TVR_BITS + N * TVN_BITS)) & TVN_MASK
+#define INDEX(N) (base->last_timer_time >> (TVR_BITS + N * TVN_BITS)) & TVN_MASK
 
-static inline void __run_timers(tvec_base_t *base)
+static inline void __run_timers(tvec_base_t *base, unsigned long current_timer_time)
 {
 	struct timer_list *timer;
 
 	spin_lock_irq(&base->t_base.lock);
-	while (time_after_eq(jiffies, base->timer_jiffies)) {
+	while (time_after_eq(current_timer_time, base->last_timer_time)) {
 		struct list_head work_list = LIST_HEAD_INIT(work_list);
 		struct list_head *head = &work_list;
- 		int index = base->timer_jiffies & TVR_MASK;
+ 		int index = base->last_timer_time & TVR_MASK;
  
 		/*
 		 * Cascade timers:
@@ -475,7 +643,7 @@ static inline void __run_timers(tvec_bas
 				(!cascade(base, &base->tv3, INDEX(1))) &&
 					!cascade(base, &base->tv4, INDEX(2)))
 			cascade(base, &base->tv5, INDEX(3));
-		++base->timer_jiffies; 
+		++base->last_timer_time;
 		list_splice_init(base->tv1.vec + index, &work_list);
 		while (!list_empty(head)) {
 			void (*fn)(unsigned long);
@@ -524,20 +692,20 @@ unsigned long next_timer_interrupt(void)
 
 	base = &__get_cpu_var(tvec_bases);
 	spin_lock(&base->t_base.lock);
-	expires = base->timer_jiffies + (LONG_MAX >> 1);
+	expires = base->last_timer_time + (LONG_MAX >> 1);
 	list = 0;
 
 	/* Look for timer events in tv1. */
-	j = base->timer_jiffies & TVR_MASK;
+	j = base->last_timer_time & TVR_MASK;
 	do {
 		list_for_each_entry(nte, base->tv1.vec + j, entry) {
 			expires = nte->expires;
-			if (j < (base->timer_jiffies & TVR_MASK))
+			if (j < (base->last_timer_time & TVR_MASK))
 				list = base->tv2.vec + (INDEX(0));
 			goto found;
 		}
 		j = (j + 1) & TVR_MASK;
-	} while (j != (base->timer_jiffies & TVR_MASK));
+	} while (j != (base->last_timer_time & TVR_MASK));
 
 	/* Check tv2-tv5. */
 	varray[0] = &base->tv2;
@@ -910,10 +1078,15 @@ EXPORT_SYMBOL(xtime_lock);
  */
 static void run_timer_softirq(struct softirq_action *h)
 {
+	unsigned long current_timer_time;
 	tvec_base_t *base = &__get_cpu_var(tvec_bases);
 
-	if (time_after_eq(jiffies, base->timer_jiffies))
-		__run_timers(base);
+	/* cache the converted current time, rounding down */
+	current_timer_time =
+		nsecs_to_timerintervals_floor(do_monotonic_clock());
+
+	if (time_after_eq(current_timer_time, base->last_timer_time))
+		__run_timers(base, current_timer_time);
 }
 
 /*

[RFC][PATCH 3/4] new human-time schedule_timeout() functions

[Nishanth Aravamudan]

From: Nishanth Aravamudan <nacc@us.ibm.com>

Description: Add new human-time schedule_timeout() style functions,
along with the appropriate constants/prototypes.

Signed-off-by: Nishanth Aravamudan <nacc@us.ibm.com>

---

 include/linux/sched.h |    7 ++
 include/linux/time.h  |    4 +
 kernel/timer.c        |  147 ++++++++++++++++++++++++++++++++++++++++++++++++++
 3 files changed, 158 insertions(+)

diff -urpN 2.6.13-rc3-base/include/linux/sched.h 2.6.13-rc3-dev/include/linux/sched.h
--- 2.6.13-rc3-base/include/linux/sched.h	2005-07-13 15:52:14.000000000 -0700
+++ 2.6.13-rc3-dev/include/linux/sched.h	2005-07-14 12:45:15.000000000 -0700
@@ -182,7 +182,14 @@ extern void scheduler_tick(void);
 extern int in_sched_functions(unsigned long addr);
 
 #define	MAX_SCHEDULE_TIMEOUT	LONG_MAX
+#define	MAX_SCHEDULE_TIMEOUT_NSECS	((u64)(-1))
+#define	MAX_SCHEDULE_TIMEOUT_USECS	ULONG_MAX
+#define	MAX_SCHEDULE_TIMEOUT_MSECS	UINT_MAX
+
 extern signed long FASTCALL(schedule_timeout(signed long timeout));
+extern u64 FASTCALL(schedule_timeout_nsecs(u64 timeout_nsecs));
+extern unsigned long FASTCALL(schedule_timeout_usecs(unsigned long timeout_usecs));
+extern unsigned int FASTCALL(schedule_timeout_msecs(unsigned int timeout_msecs));
 asmlinkage void schedule(void);
 
 struct namespace;
diff -urpN 2.6.13-rc3-base/include/linux/time.h 2.6.13-rc3-dev/include/linux/time.h
--- 2.6.13-rc3-base/include/linux/time.h	2005-07-14 12:45:07.000000000 -0700
+++ 2.6.13-rc3-dev/include/linux/time.h	2005-07-14 12:45:15.000000000 -0700
@@ -36,6 +36,10 @@ struct timezone {
 #define NSEC_PER_SEC (1000000000L)
 #endif
 
+#ifndef NSEC_PER_MSEC
+#define NSEC_PER_MSEC (1000000L)
+#endif
+
 #ifndef NSEC_PER_USEC
 #define NSEC_PER_USEC (1000L)
 #endif
diff -urpN 2.6.13-rc3-base/kernel/timer.c 2.6.13-rc3-dev/kernel/timer.c
--- 2.6.13-rc3-base/kernel/timer.c	2005-07-14 12:45:07.000000000 -0700
+++ 2.6.13-rc3-dev/kernel/timer.c	2005-07-14 12:45:15.000000000 -0700
@@ -1271,6 +1271,10 @@ static void process_timeout(unsigned lon
  * value will be %MAX_SCHEDULE_TIMEOUT.
  *
  * In all cases the return value is guaranteed to be non-negative.
+ *
+ * The callers of schedule_timeout() should be aware that the interface
+ * is now deprecated. schedule_timeout_{msecs,usecs,nsecs}() are now the
+ * interfaces for relative timeout requests.
  */
 fastcall signed long __sched schedule_timeout(signed long timeout)
 {
@@ -1326,6 +1330,149 @@ fastcall signed long __sched schedule_ti
 
 EXPORT_SYMBOL(schedule_timeout);
 
+/**
+ * schedule_timeout_nsecs - sleep until timeout
+ * @timeout_nsecs: timeout value in nanoseconds
+ *
+ * Make the current task sleep until @timeout_nsecs nsecs have
+ * elapsed. The routine will return immediately unless
+ * the current task state has been set (see set_current_state()).
+ *
+ * You can set the task state as follows -
+ *
+ * %TASK_UNINTERRUPTIBLE - at least @timeout_nsecs nsecs are guaranteed
+ * to pass before the routine returns. The routine will return 0
+ *
+ * %TASK_INTERRUPTIBLE - the routine may return early if a signal is
+ * delivered to the current task. In this case the remaining time
+ * in nsecs will be returned, or 0 if the timer expired in time
+ *
+ * The current task state is guaranteed to be TASK_RUNNING when this
+ * routine returns.
+ *
+ * Specifying a @timeout value of %MAX_SCHEDULE_TIMEOUT_NSECS will
+ * schedule the CPU away without a bound on the timeout. In this case
+ * the return value will be %MAX_SCHEDULE_TIMEOUT_NSECS.
+ */
+fastcall u64 __sched schedule_timeout_nsecs(u64 timeout_nsecs)
+{
+	struct timer_list timer;
+	u64 expires;
+
+	if (timeout_nsecs == MAX_SCHEDULE_TIMEOUT_NSECS) {
+		schedule();
+		goto out;
+	}
+
+	expires = do_monotonic_clock() + timeout_nsecs;
+
+	init_timer(&timer);
+	timer.data = (unsigned long) current;
+	timer.function = process_timeout;
+
+	set_timer_nsecs(&timer, expires);
+	schedule();
+	del_singleshot_timer_sync(&timer);
+
+	timeout_nsecs = do_monotonic_clock();
+	if (expires < timeout_nsecs)
+		timeout_nsecs = (u64)0UL;
+	else
+		timeout_nsecs = expires - timeout_nsecs;
+out:
+	return timeout_nsecs;
+}
+
+EXPORT_SYMBOL_GPL(schedule_timeout_nsecs);
+
+/**
+ * schedule_timeout_usecs - sleep until timeout
+ * @timeout_usecs: timeout value in nanoseconds
+ *
+ * Make the current task sleep until @timeout_usecs usecs have
+ * elapsed. The routine will return immediately unless
+ * the current task state has been set (see set_current_state()).
+ *
+ * You can set the task state as follows -
+ *
+ * %TASK_UNINTERRUPTIBLE - at least @timeout_usecs usecs are guaranteed
+ * to pass before the routine returns. The routine will return 0
+ *
+ * %TASK_INTERRUPTIBLE - the routine may return early if a signal is
+ * delivered to the current task. In this case the remaining time
+ * in usecs will be returned, or 0 if the timer expired in time
+ *
+ * The current task state is guaranteed to be TASK_RUNNING when this
+ * routine returns.
+ *
+ * Specifying a @timeout value of %MAX_SCHEDULE_TIMEOUT_USECS will
+ * schedule the CPU away without a bound on the timeout. In this case
+ * the return value will be %MAX_SCHEDULE_TIMEOUT_USECS.
+ */
+fastcall inline unsigned long __sched schedule_timeout_usecs(unsigned long timeout_usecs)
+{
+	u64 timeout_nsecs;
+
+	if (timeout_usecs == MAX_SCHEDULE_TIMEOUT_USECS)
+		timeout_nsecs = MAX_SCHEDULE_TIMEOUT_NSECS;
+	else
+		timeout_nsecs = timeout_usecs * (u64)NSEC_PER_USEC;
+	/*
+	 * Make sure to round up by subtracting one before division and
+	 * adding one after
+	 */
+	timeout_nsecs = schedule_timeout_nsecs(timeout_nsecs) - 1;
+	do_div(timeout_nsecs, NSEC_PER_USEC);
+	timeout_usecs = (unsigned long)timeout_nsecs + 1UL;
+	return timeout_usecs;
+}
+
+EXPORT_SYMBOL_GPL(schedule_timeout_usecs);
+
+/**
+ * schedule_timeout_msecs - sleep until timeout
+ * @timeout_msecs: timeout value in nanoseconds
+ *
+ * Make the current task sleep until @timeout_msecs msecs have
+ * elapsed. The routine will return immediately unless
+ * the current task state has been set (see set_current_state()).
+ *
+ * You can set the task state as follows -
+ *
+ * %TASK_UNINTERRUPTIBLE - at least @timeout_msecs msecs are guaranteed
+ * to pass before the routine returns. The routine will return 0
+ *
+ * %TASK_INTERRUPTIBLE - the routine may return early if a signal is
+ * delivered to the current task. In this case the remaining time
+ * in msecs will be returned, or 0 if the timer expired in time
+ *
+ * The current task state is guaranteed to be TASK_RUNNING when this
+ * routine returns.
+ *
+ * Specifying a @timeout value of %MAX_SCHEDULE_TIMEOUT_MSECS will
+ * schedule the CPU away without a bound on the timeout. In this case
+ * the return value will be %MAX_SCHEDULE_TIMEOUT_MSECS.
+ */
+fastcall inline unsigned int __sched schedule_timeout_msecs(unsigned int timeout_msecs)
+{
+	u64 timeout_nsecs;
+
+	if (timeout_msecs == MAX_SCHEDULE_TIMEOUT_MSECS)
+		timeout_nsecs = MAX_SCHEDULE_TIMEOUT_NSECS;
+	else
+		timeout_nsecs = timeout_msecs * (u64)NSEC_PER_MSEC;
+	/*
+	 * Make sure to round up by subtracting one before division and
+	 * adding one after
+	 */
+	timeout_nsecs = schedule_timeout_nsecs(timeout_nsecs) - 1;
+	do_div(timeout_nsecs, NSEC_PER_MSEC);
+	timeout_msecs = (unsigned int)timeout_nsecs + 1;
+	return timeout_msecs;
+}
+
+EXPORT_SYMBOL_GPL(schedule_timeout_msecs);
+
 /* Thread ID - the internal kernel "pid" */
 asmlinkage long sys_gettid(void)
 {

[RFC][PATCH 4/4] convert sys_nanosleep() to use set_timer_nsecs()

[Nishanth Aravamudan]

From: Nishanth Aravamudan <nacc@us.ibm.com>

Description: Add timespec and timeval conversion functions for
nanoseconds. Convert sys_nanosleep() to use schedule_timeout_nsecs().

Signed-off-by: Nishanth Aravamudan <nacc@us.ibm.com>

---

 include/linux/time.h |   33 +++++++++++++++++++++++++++++++++
 kernel/timer.c       |   24 ++++++++++++------------
 2 files changed, 45 insertions(+), 12 deletions(-)

diff -urpN 2.6.13-rc3-base/include/linux/time.h 2.6.13-rc3-dev/include/linux/time.h
--- 2.6.13-rc3-base/include/linux/time.h	2005-07-14 12:46:46.000000000 -0700
+++ 2.6.13-rc3-dev/include/linux/time.h	2005-07-14 12:48:25.000000000 -0700
@@ -2,6 +2,7 @@
 #define _LINUX_TIME_H
 
 #include <linux/types.h>
+#include <asm/div64.h>
 
 #ifdef __KERNEL__
 #include <linux/seqlock.h>
@@ -126,6 +127,38 @@ set_normalized_timespec (struct timespec
 	ts->tv_nsec = nsec;
 }
 
+/* Inline helper functions */
+static inline struct timeval nsecs_to_timeval(u64 ns)
+{
+	struct timeval tv;
+	tv.tv_sec = div_long_long_rem(ns, NSEC_PER_SEC, &tv.tv_usec);
+	tv.tv_usec = (tv.tv_usec + NSEC_PER_USEC/2) / NSEC_PER_USEC;
+	return tv;
+}
+
+static inline struct timespec nsecs_to_timespec(u64 ns)
+{
+	struct timespec ts;
+	ts.tv_sec = div_long_long_rem(ns, NSEC_PER_SEC, &ts.tv_nsec);
+	return ts;
+}
+
+static inline u64 timespec_to_nsecs(struct timespec* ts)
+{
+	u64 ret;
+	ret = ((u64)ts->tv_sec) * NSEC_PER_SEC;
+	ret += (u64)ts->tv_nsec;
+	return ret;
+}
+
+static inline u64 timeval_to_nsecs(struct timeval* tv)
+{
+	u64 ret;
+	ret = ((u64)tv->tv_sec) * NSEC_PER_SEC;
+	ret += ((u64)tv->tv_usec) * NSEC_PER_USEC;
+	return ret;
+}
+
 #endif /* __KERNEL__ */
 
 #define NFDBITS			__NFDBITS
diff -urpN 2.6.13-rc3-base/kernel/timer.c 2.6.13-rc3-dev/kernel/timer.c
--- 2.6.13-rc3-base/kernel/timer.c	2005-07-14 12:46:46.000000000 -0700
+++ 2.6.13-rc3-dev/kernel/timer.c	2005-07-14 12:48:25.000000000 -0700
@@ -1481,21 +1481,21 @@ asmlinkage long sys_gettid(void)
 
 static long __sched nanosleep_restart(struct restart_block *restart)
 {
-	unsigned long expire = restart->arg0, now = jiffies;
+	u64 expire = restart->arg0, now = do_monotonic_clock();
 	struct timespec __user *rmtp = (struct timespec __user *) restart->arg1;
 	long ret;
 
 	/* Did it expire while we handled signals? */
-	if (!time_after(expire, now))
+	if (now > expire)
 		return 0;
 
-	current->state = TASK_INTERRUPTIBLE;
-	expire = schedule_timeout(expire - now);
+	set_current_state(TASK_INTERRUPTIBLE);
+	expire = schedule_timeout_nsecs(expire - now);
 
 	ret = 0;
 	if (expire) {
 		struct timespec t;
-		jiffies_to_timespec(expire, &t);
+		t = nsecs_to_timespec(expire);
 
 		ret = -ERESTART_RESTARTBLOCK;
 		if (rmtp && copy_to_user(rmtp, &t, sizeof(t)))
@@ -1508,7 +1508,7 @@ static long __sched nanosleep_restart(st
 asmlinkage long sys_nanosleep(struct timespec __user *rqtp, struct timespec __user *rmtp)
 {
 	struct timespec t;
-	unsigned long expire;
+	u64 expire;
 	long ret;
 
 	if (copy_from_user(&t, rqtp, sizeof(t)))
@@ -1517,20 +1517,20 @@ asmlinkage long sys_nanosleep(struct tim
 	if ((t.tv_nsec >= 1000000000L) || (t.tv_nsec < 0) || (t.tv_sec < 0))
 		return -EINVAL;
 
-	expire = timespec_to_jiffies(&t) + (t.tv_sec || t.tv_nsec);
-	current->state = TASK_INTERRUPTIBLE;
-	expire = schedule_timeout(expire);
+	expire = timespec_to_nsecs(&t);
+	set_current_state(TASK_INTERRUPTIBLE);
+	expire = schedule_timeout_nsecs(expire);
 
 	ret = 0;
 	if (expire) {
 		struct restart_block *restart;
-		jiffies_to_timespec(expire, &t);
+		t = nsecs_to_timespec(expire);
 		if (rmtp && copy_to_user(rmtp, &t, sizeof(t)))
 			return -EFAULT;
 
 		restart = &current_thread_info()->restart_block;
 		restart->fn = nanosleep_restart;
-		restart->arg0 = jiffies + expire;
+		restart->arg0 = do_monotonic_clock() + expire;
 		restart->arg1 = (unsigned long) rmtp;
 		ret = -ERESTART_RESTARTBLOCK;
 	}
@@ -1642,7 +1642,7 @@ static void __devinit init_timers_cpu(in
 	for (j = 0; j < TVR_SIZE; j++)
 		INIT_LIST_HEAD(base->tv1.vec + j);
 
-	base->timer_jiffies = jiffies;
+	base->last_timer_time = 0UL;
 }
 
 #ifdef CONFIG_HOTPLUG_CPU

Re: [RFC][PATCH 1/4] add jiffies_to_nsecs() helper and fix up size of usecs

[Dave Hansen]

On Thu, 2005-07-14 at 13:28 -0700, Nishanth Aravamudan wrote:
> +static inline u64 jiffies_to_nsecs(const unsigned long j)
> +{
> +#if HZ <= NSEC_PER_SEC && !(NSEC_PER_SEC % HZ)
> +	return (NSEC_PER_SEC / HZ) * (u64)j;
> +#elif HZ > NSEC_PER_SEC && !(HZ % NSEC_PER_SEC)
> +	return ((u64)j + (HZ / NSEC_PER_SEC) - 1)/(HZ / NSEC_PER_SEC);
> +#else
> +	return ((u64)j * NSEC_PER_SEC) / HZ;
> +#endif
> +}

That might look a little better something like:

static inline u64 jiffies_to_nsecs(const unsigned long __j)
{
	u64 j = __j;

	if (HZ <= NSEC_PER_SEC && !(NSEC_PER_SEC % HZ))
		return (NSEC_PER_SEC / HZ) * j;
	else if (HZ > NSEC_PER_SEC && !(HZ % NSEC_PER_SEC))
		return (j + (HZ / NSEC_PER_SEC) - 1)/(HZ / NSEC_PER_SEC);
	else
		return (j * NSEC_PER_SEC) / HZ;
}

Compilers are smart :)

-- Dave

Re: [RFC][PATCH 1/4] add jiffies_to_nsecs() helper and fix up size of usecs

[Nishanth Aravamudan]

On 14.07.2005 [13:54:47 -0700], Dave Hansen wrote:
> On Thu, 2005-07-14 at 13:28 -0700, Nishanth Aravamudan wrote:
> > +static inline u64 jiffies_to_nsecs(const unsigned long j)
> > +{
> > +#if HZ <= NSEC_PER_SEC && !(NSEC_PER_SEC % HZ)
> > +	return (NSEC_PER_SEC / HZ) * (u64)j;
> > +#elif HZ > NSEC_PER_SEC && !(HZ % NSEC_PER_SEC)
> > +	return ((u64)j + (HZ / NSEC_PER_SEC) - 1)/(HZ / NSEC_PER_SEC);
> > +#else
> > +	return ((u64)j * NSEC_PER_SEC) / HZ;
> > +#endif
> > +}
> 
> That might look a little better something like:
> 
> static inline u64 jiffies_to_nsecs(const unsigned long __j)
> {
> 	u64 j = __j;
> 
> 	if (HZ <= NSEC_PER_SEC && !(NSEC_PER_SEC % HZ))
> 		return (NSEC_PER_SEC / HZ) * j;
> 	else if (HZ > NSEC_PER_SEC && !(HZ % NSEC_PER_SEC))
> 		return (j + (HZ / NSEC_PER_SEC) - 1)/(HZ / NSEC_PER_SEC);
> 	else
> 		return (j * NSEC_PER_SEC) / HZ;
> }
> 
> Compilers are smart :)

Well, I was trying to keep it similar to the other conversion functions.
I guess the compiler can evaluate the conditional full of constants at
compile-time regardless of whether it is #if or if ().

I can make these changes if others would like them as well.

Thanks,
Nish

Re: [RFC][PATCH 0/4] new human-time soft-timer subsystem

[Roman Zippel]

Hi,

On Thu, 14 Jul 2005, Nishanth Aravamudan wrote:

> We no longer use jiffies (the variable) as the basis for determining
> what "time" a timer should expire or when it should be added. Instead,
> we use a new function, do_monotonic_clock(), which is simply a wrapper
> for getnstimeofday().

And suddenly a simple 32bit integer becomes a complex 64bit integer, which 
requires hardware access to read a timer and additional conversion into ns.
Why is suddenly everyone so obsessed with molesting something simple and 
cute as jiffies?

bye, Roman

Re: [RFC][PATCH 1/4] add jiffies_to_nsecs() helper and fix up size of usecs

[Pavel Machek]

Hi!

> > > +static inline u64 jiffies_to_nsecs(const unsigned long j)
> > > +{
> > > +#if HZ <= NSEC_PER_SEC && !(NSEC_PER_SEC % HZ)
> > > +	return (NSEC_PER_SEC / HZ) * (u64)j;
> > > +#elif HZ > NSEC_PER_SEC && !(HZ % NSEC_PER_SEC)
> > > +	return ((u64)j + (HZ / NSEC_PER_SEC) - 1)/(HZ / NSEC_PER_SEC);
> > > +#else
> > > +	return ((u64)j * NSEC_PER_SEC) / HZ;
> > > +#endif
> > > +}
> > 
> > That might look a little better something like:
> > 
> > static inline u64 jiffies_to_nsecs(const unsigned long __j)
> > {
> > 	u64 j = __j;
> > 
> > 	if (HZ <= NSEC_PER_SEC && !(NSEC_PER_SEC % HZ))
> > 		return (NSEC_PER_SEC / HZ) * j;
> > 	else if (HZ > NSEC_PER_SEC && !(HZ % NSEC_PER_SEC))
> > 		return (j + (HZ / NSEC_PER_SEC) - 1)/(HZ / NSEC_PER_SEC);
> > 	else
> > 		return (j * NSEC_PER_SEC) / HZ;
> > }
> > 
> > Compilers are smart :)
> 
> Well, I was trying to keep it similar to the other conversion functions.
> I guess the compiler can evaluate the conditional full of constants at
> compile-time regardless of whether it is #if or if ().
> 
> I can make these changes if others would like them as well.

Yes, please. And feel free to convert nearby functions, too ;-).
								Pavel
-- 
teflon -- maybe it is a trademark, but it should not be.

Re: [RFC][PATCH 1/4] add jiffies_to_nsecs() helper and fix up size of usecs

[Nishanth Aravamudan]

On 15.07.2005 [14:14:25 +0200], Pavel Machek wrote:
> Hi!
> 
> > > > +static inline u64 jiffies_to_nsecs(const unsigned long j)
> > > > +{
> > > > +#if HZ <= NSEC_PER_SEC && !(NSEC_PER_SEC % HZ)
> > > > +	return (NSEC_PER_SEC / HZ) * (u64)j;
> > > > +#elif HZ > NSEC_PER_SEC && !(HZ % NSEC_PER_SEC)
> > > > +	return ((u64)j + (HZ / NSEC_PER_SEC) - 1)/(HZ / NSEC_PER_SEC);
> > > > +#else
> > > > +	return ((u64)j * NSEC_PER_SEC) / HZ;
> > > > +#endif
> > > > +}
> > > 
> > > That might look a little better something like:
> > > 
> > > static inline u64 jiffies_to_nsecs(const unsigned long __j)
> > > {
> > > 	u64 j = __j;
> > > 
> > > 	if (HZ <= NSEC_PER_SEC && !(NSEC_PER_SEC % HZ))
> > > 		return (NSEC_PER_SEC / HZ) * j;
> > > 	else if (HZ > NSEC_PER_SEC && !(HZ % NSEC_PER_SEC))
> > > 		return (j + (HZ / NSEC_PER_SEC) - 1)/(HZ / NSEC_PER_SEC);
> > > 	else
> > > 		return (j * NSEC_PER_SEC) / HZ;
> > > }
> > > 
> > > Compilers are smart :)
> > 
> > Well, I was trying to keep it similar to the other conversion functions.
> > I guess the compiler can evaluate the conditional full of constants at
> > compile-time regardless of whether it is #if or if ().
> > 
> > I can make these changes if others would like them as well.
> 
> Yes, please. And feel free to convert nearby functions, too ;-).

I have a patch to make this change for all the jiffies <--> human-time
functions, but have a problem. I noticed that these functions, in the
if/else form (as opposed to #if/#else) will warn about division-by-zero
problems, as (HZ / MSEC_PER_SEC), (HZ / USEC_PER_SEC) & (HZ /
NSEC_PER_SEC) are all 0 if HZ < 1000 (which, of course, is the default
now :) ). Any suggestions? Just leave the functions as is? Even then,
I'm going to update this patch to use USEC_PER_SEC and MSEC_PER_SEC in
the other conversion functions like I use NSEC_PER_SEC in the first
version.

Thanks,
Nish

Re: [RFC][PATCH 0/4] new human-time soft-timer subsystem

[Nishanth Aravamudan]

On 15.07.2005 [00:28:44 +0200], Roman Zippel wrote:
> Hi,
> 
> On Thu, 14 Jul 2005, Nishanth Aravamudan wrote:
> 
> > We no longer use jiffies (the variable) as the basis for determining
> > what "time" a timer should expire or when it should be added. Instead,
> > we use a new function, do_monotonic_clock(), which is simply a wrapper
> > for getnstimeofday().
> 
> And suddenly a simple 32bit integer becomes a complex 64bit integer, which 
> requires hardware access to read a timer and additional conversion into ns.
> Why is suddenly everyone so obsessed with molesting something simple and 
> cute as jiffies?

Thanks for the feedback, Roman. I know the 64-bit operations are
critical from a performance perspective and may be excessive from a
pragmatic perspective. Maybe an alternative would be to only provide
*microsecond* resolution in the software, which I currently assume is
storable in an unsigned long (a little over an hour?). We could then
provide a supplemental interface for those sleeps which would exceed
this time, either via looping or a 64-bit parameter for this special
interface.

Would that perhaps be a better alternative from the 64-bit perspective?

We could do this one better, perhaps, by basically doing exactly what
jiffies does now, but storing a time value (in microseconds) instead of
a count of the number of ticks (jiffies' current interpretation). This
would perhaps be a 64-bit op, but that is the case current with
jiffies_64++ (or jiffies_64 += jiffies_increment). I will work on some
patches to do something to this effect and will bring it up during the
time/timer talk (Saturday at 13h30).

Thanks again,
Nish

[RFC][UPDATE PATCH 2/4] human-time soft-timer core changes

[Nishanth Aravamudan]

On 14.07.2005 [13:40:11 -0700], Nishanth Aravamudan wrote:
> From: Nishanth Aravamudan <nacc@us.ibm.com>
> 
> Description: The core revision to the soft-timer subsystem to divorce it
> from the timer interrupt in software, i.e. jiffies. Instead, use
> getnstimeofday() (via do_monotonic_clock()) as the basis for addition
> and expiration of timers. Add a new unit, the timerinterval, which is
> a 2^TIMERINTERVAL_BITS nanoseconds in length. The converted value in
> timerintervals is used where we would have used the timer's expires
> member before. Add set_timer_nsecs() and set_timer_nsecs_on() functions
> to directly request nanosecond delays. These functions replace
> add_timer(), mod_timer() and add_timer_on().
> 
> Signed-off-by: Nishanth Aravamudan <nacc@us.ibm.com>

Sigh, one version of my development patches removed the export of
mod_timer(). Of course, I forgot to revert that hunk before sending it
out. If anyone (maybe not likely) is testing out these patches, please
use this version.

Thanks,
Nish

---

 include/linux/time.h  |    1 
 include/linux/timer.h |   27 +-----
 kernel/time.c         |   18 ++++
 kernel/timer.c        |  215 +++++++++++++++++++++++++++++++++++++++++++++-----
 4 files changed, 221 insertions(+), 40 deletions(-)

diff -urpN 2.6.13-rc3-base/include/linux/time.h 2.6.13-rc3-dev/include/linux/time.h
--- 2.6.13-rc3-base/include/linux/time.h	2005-03-01 23:38:12.000000000 -0800
+++ 2.6.13-rc3-dev/include/linux/time.h	2005-07-14 12:44:40.000000000 -0700
@@ -103,6 +103,7 @@ struct itimerval;
 extern int do_setitimer(int which, struct itimerval *value, struct itimerval *ovalue);
 extern int do_getitimer(int which, struct itimerval *value);
 extern void getnstimeofday (struct timespec *tv);
+extern u64 do_monotonic_clock(void);
 
 extern struct timespec timespec_trunc(struct timespec t, unsigned gran);
 
diff -urpN 2.6.13-rc3-base/include/linux/timer.h 2.6.13-rc3-dev/include/linux/timer.h
--- 2.6.13-rc3-base/include/linux/timer.h	2005-07-13 15:52:14.000000000 -0700
+++ 2.6.13-rc3-dev/include/linux/timer.h	2005-07-14 12:44:40.000000000 -0700
@@ -11,6 +11,7 @@ struct timer_base_s;
 struct timer_list {
 	struct list_head entry;
 	unsigned long expires;
+	u64 expires_nsecs;
 
 	unsigned long magic;
 
@@ -27,6 +28,7 @@ extern struct timer_base_s __init_timer_
 #define TIMER_INITIALIZER(_function, _expires, _data) {		\
 		.function = (_function),			\
 		.expires = (_expires),				\
+		.expires_nsecs = 0,				\
 		.data = (_data),				\
 		.base = &__init_timer_base,			\
 		.magic = TIMER_MAGIC,				\
@@ -51,30 +53,15 @@ static inline int timer_pending(const st
 
 extern void add_timer_on(struct timer_list *timer, int cpu);
 extern int del_timer(struct timer_list * timer);
-extern int __mod_timer(struct timer_list *timer, unsigned long expires);
+extern int __mod_timer(struct timer_list *timer);
 extern int mod_timer(struct timer_list *timer, unsigned long expires);
+extern void add_timer(struct timer_list *timer);
+extern int set_timer_nsecs(struct timer_list *timer, u64 expires_nsecs);
+extern void set_timer_on_nsecs(struct timer_list *timer, u64 expires_nsecs,
+								int cpu);
 
 extern unsigned long next_timer_interrupt(void);
 
-/***
- * add_timer - start a timer
- * @timer: the timer to be added
- *
- * The kernel will do a ->function(->data) callback from the
- * timer interrupt at the ->expired point in the future. The
- * current time is 'jiffies'.
- *
- * The timer's ->expired, ->function (and if the handler uses it, ->data)
- * fields must be set prior calling this function.
- *
- * Timers with an ->expired field in the past will be executed in the next
- * timer tick.
- */
-static inline void add_timer(struct timer_list * timer)
-{
-	__mod_timer(timer, timer->expires);
-}
-
 #ifdef CONFIG_SMP
   extern int try_to_del_timer_sync(struct timer_list *timer);
   extern int del_timer_sync(struct timer_list *timer);
diff -urpN 2.6.13-rc3-base/kernel/time.c 2.6.13-rc3-dev/kernel/time.c
--- 2.6.13-rc3-base/kernel/time.c	2005-07-13 15:51:57.000000000 -0700
+++ 2.6.13-rc3-dev/kernel/time.c	2005-07-14 12:44:40.000000000 -0700
@@ -589,3 +589,21 @@ EXPORT_SYMBOL(get_jiffies_64);
 #endif
 
 EXPORT_SYMBOL(jiffies);
+
+u64 do_monotonic_clock(void)
+{
+	struct timespec now, now_w2m;
+	unsigned long seq;
+
+	getnstimeofday(&now);
+
+	do {
+		seq = read_seqbegin(&xtime_lock);
+		now_w2m = wall_to_monotonic;
+	} while (read_seqretry(&xtime_lock, seq));
+
+	return (u64)(now.tv_sec + now_w2m.tv_sec) * NSEC_PER_SEC +
+				(now.tv_nsec + now_w2m.tv_nsec);
+}
+
+EXPORT_SYMBOL_GPL(do_monotonic_clock);
diff -urpN 2.6.13-rc3-base/kernel/timer.c 2.6.13-rc3-dev/kernel/timer.c
--- 2.6.13-rc3-base/kernel/timer.c	2005-07-13 15:52:14.000000000 -0700
+++ 2.6.13-rc3-dev/kernel/timer.c	2005-07-14 12:44:40.000000000 -0700
@@ -56,6 +56,15 @@ static void time_interpolator_update(lon
 #define TVR_SIZE (1 << TVR_BITS)
 #define TVN_MASK (TVN_SIZE - 1)
 #define TVR_MASK (TVR_SIZE - 1)
+/*
+ * Modifying TIMERINTERVAL_BITS changes the software resolution of
+ * soft-timers. While 20 bits would be closer to a millisecond, there
+ * are performance gains from allowing a software resolution finer than
+ * the hardware (HZ=1000)
+ */
+#define TIMERINTERVAL_BITS 19
+#define TIMERINTERVAL_SIZE (1 << TIMERINTERVAL_BITS)
+#define TIMERINTERVAL_MASK (TIMERINTERVAL_SIZE - 1)
 
 struct timer_base_s {
 	spinlock_t lock;
@@ -72,7 +81,7 @@ typedef struct tvec_root_s {
 
 struct tvec_t_base_s {
 	struct timer_base_s t_base;
-	unsigned long timer_jiffies;
+	unsigned long last_timer_time;
 	tvec_root_t tv1;
 	tvec_t tv2;
 	tvec_t tv3;
@@ -114,11 +123,88 @@ static inline void check_timer(struct ti
 		check_timer_failed(timer);
 }
 
+/*
+ * nsecs_to_timerintervals_ceiling - convert nanoseconds to timerintervals
+ * @n: number of nanoseconds to convert
+ *
+ * This is where changes to TIMERINTERVAL_BITS affect the soft-timer
+ * subsystem.
+ *
+ * Some explanation of the math is necessary:
+ * Rather than do decimal arithmetic, we shift for the sake of speed.
+ * This does mean that the actual requestable sleeps are
+ * 	2^(sizeof(unsigned long)*8 - TIMERINTERVAL_BITS)
+ * timerintervals.
+ *
+ * The conditional takes care of the corner case where we request a 0
+ * nanosecond sleep; if the quantity were unsigned, we would not
+ * propogate the carry and force a wrap when adding the 1.
+ *
+ * To prevent timers from being expired early, we:
+ *	Take the ceiling when we add; and
+ *	Take the floor when we expire.
+ */
+static inline unsigned long nsecs_to_timerintervals_ceiling(u64 nsecs)
+{
+	if (nsecs)
+		return (unsigned long)(((nsecs - 1) >> TIMERINTERVAL_BITS) + 1);
+	else
+		return 0UL;
+}
+
+/*
+ * nsecs_to_timerintervals_floor - convert nanoseconds to timerintervals
+ * @n: number of nanoseconds to convert
+ *
+ * This is where changes to TIMERINTERVAL_BITS affect the soft-timer
+ * subsystem.
+ *
+ * Some explanation of the math is necessary:
+ * Rather than do decimal arithmetic, we shift for the sake of speed.
+ * This does mean that the actual requestable sleeps are
+ * 	2^(sizeof(unsigned long)*8 - TIMERINTERVAL_BITS)
+ *
+ * There is no special case for 0 in the floor function, since we do not
+ * do any subtraction or addition of 1
+ *
+ * To prevent timers from being expired early, we:
+ *	Take the ceiling when we add; and
+ *	Take the floor when we expire.
+ */
+static inline unsigned long nsecs_to_timerintervals_floor(u64 nsecs)
+{
+	return (unsigned long)(nsecs >> TIMERINTERVAL_BITS);
+}
+
+/*
+ * jiffies_to_timerintervals - convert absolute jiffies to timerintervals
+ * @abs_jiffies: number of jiffies to convert
+ *
+ * First, we convert the absolute jiffies parameter to a relative
+ * jiffies value. To maintain precision, we convert the relative
+ * jiffies value to a relative nanosecond value and then convert that
+ * to a relative soft-timer interval unit value. We then add this
+ * relative value to the current time according to the timeofday-
+ * subsystem, converted to soft-timer interval units.
+ *
+ * We only use this function when adding timers, so we are free to
+ * always use the ceiling version of nsecs_to_timerintervals.
+ *
+ * This function only exists to support deprecated interfaces. Once
+ * those interfaces have been converted to the alternatives, it should
+ * be removed.
+ */
+static inline unsigned long jiffies_to_timerintervals(unsigned long abs_jiffies)
+{
+	unsigned long relative_jiffies = abs_jiffies - jiffies;
+	return nsecs_to_timerintervals_ceiling(do_monotonic_clock() +
+		jiffies_to_nsecs(relative_jiffies));
+}
 
 static void internal_add_timer(tvec_base_t *base, struct timer_list *timer)
 {
-	unsigned long expires = timer->expires;
-	unsigned long idx = expires - base->timer_jiffies;
+	unsigned long expires = nsecs_to_timerintervals_ceiling(timer->expires_nsecs);
+	unsigned long idx = expires - base->last_timer_time;
 	struct list_head *vec;
 
 	if (idx < TVR_SIZE) {
@@ -138,7 +224,7 @@ static void internal_add_timer(tvec_base
 		 * Can happen if you add a timer with expires == jiffies,
 		 * or you set a timer to go off in the past
 		 */
-		vec = base->tv1.vec + (base->timer_jiffies & TVR_MASK);
+		vec = base->tv1.vec + (base->last_timer_time & TVR_MASK);
 	} else {
 		int i;
 		/* If the timeout is larger than 0xffffffff on 64-bit
@@ -146,7 +232,7 @@ static void internal_add_timer(tvec_base
 		 */
 		if (idx > 0xffffffffUL) {
 			idx = 0xffffffffUL;
-			expires = idx + base->timer_jiffies;
+			expires = idx + base->last_timer_time;
 		}
 		i = (expires >> (TVR_BITS + 3 * TVN_BITS)) & TVN_MASK;
 		vec = base->tv5.vec + i;
@@ -222,7 +308,7 @@ static timer_base_t *lock_timer_base(str
 	}
 }
 
-int __mod_timer(struct timer_list *timer, unsigned long expires)
+int __mod_timer(struct timer_list *timer)
 {
 	timer_base_t *base;
 	tvec_base_t *new_base;
@@ -261,7 +347,7 @@ int __mod_timer(struct timer_list *timer
 		}
 	}
 
-	timer->expires = expires;
+	/* expires should be in timerintervals, and is currently ignored? */
 	internal_add_timer(new_base, timer);
 	spin_unlock_irqrestore(&new_base->t_base.lock, flags);
 
@@ -281,21 +367,50 @@ void add_timer_on(struct timer_list *tim
 {
 	tvec_base_t *base = &per_cpu(tvec_bases, cpu);
   	unsigned long flags;
-
+  
   	BUG_ON(timer_pending(timer) || !timer->function);
 
 	check_timer(timer);
 
 	spin_lock_irqsave(&base->t_base.lock, flags);
+	timer->expires_nsecs = do_monotonic_clock() +
+		jiffies_to_nsecs(timer->expires - jiffies);
 	timer->base = &base->t_base;
 	internal_add_timer(base, timer);
 	spin_unlock_irqrestore(&base->t_base.lock, flags);
 }
 
+/***
+ * add_timer - start a timer
+ * @timer: the timer to be added
+ *
+ * The kernel will do a ->function(->data) callback from the
+ * timer interrupt at the ->expired point in the future. The
+ * current time is 'jiffies'.
+ *
+ * The timer's ->expired, ->function (and if the handler uses it, ->data)
+ * fields must be set prior calling this function.
+ *
+ * Timers with an ->expired field in the past will be executed in the next
+ * timer tick.
+ *
+ * The callers of add_timer() should be aware that the interface is now
+ * deprecated. set_timer_nsecs() is the single interface for adding and
+ * modifying timers.
+ */
+void add_timer(struct timer_list * timer)
+{
+	timer->expires_nsecs = do_monotonic_clock() +
+		jiffies_to_nsecs(timer->expires - jiffies);
+	__mod_timer(timer);
+}
+
+EXPORT_SYMBOL(add_timer);
 
 /***
  * mod_timer - modify a timer's timeout
  * @timer: the timer to be modified
+ * @expires: absolute time, in jiffies, when timer should expire
  *
  * mod_timer is a more efficient way to update the expire field of an
  * active timer (if the timer is inactive it will be activated)
@@ -311,6 +426,10 @@ void add_timer_on(struct timer_list *tim
  * The function returns whether it has modified a pending timer or not.
  * (ie. mod_timer() of an inactive timer returns 0, mod_timer() of an
  * active timer returns 1.)
+ *
+ * The callers of mod_timer() should be aware that the interface is now
+ * deprecated. set_timer_nsecs() is the single interface for adding and
+ * modifying timers.
  */
 int mod_timer(struct timer_list *timer, unsigned long expires)
 {
@@ -318,6 +437,9 @@ int mod_timer(struct timer_list *timer, 
 
 	check_timer(timer);
 
+	timer->expires_nsecs = do_monotonic_clock() +
+		jiffies_to_nsecs(expires - jiffies);
+
 	/*
 	 * This is a common optimization triggered by the
 	 * networking code - if the timer is re-modified
@@ -326,10 +448,56 @@ int mod_timer(struct timer_list *timer, 
 	if (timer->expires == expires && timer_pending(timer))
 		return 1;
 
-	return __mod_timer(timer, expires);
+	return __mod_timer(timer);
 }
 
EXPORT_SYMBOL(mod_timer);
+
+/*
+ * set_timer_nsecs - modify a timer's timeout in nsecs
+ * @timer: the timer to be modified
+ *
+ * set_timer_nsecs replaces both add_timer and mod_timer. The caller
+ * should call do_monotonic_clock() to determine the absolute timeout
+ * necessary.
+ */
+int set_timer_nsecs(struct timer_list *timer, u64 expires_nsecs)
+{
+	BUG_ON(!timer->function);
+
+	check_timer(timer);
+
+	if (timer_pending(timer) && timer->expires_nsecs == expires_nsecs)
+		return 1;
+
+	timer->expires_nsecs = expires_nsecs;
+
+	return __mod_timer(timer);
+}
+
+EXPORT_SYMBOL_GPL(set_timer_nsecs);
+
+/***
+ * set_timer_on_nsecs - start a timer on a particular CPU
+ * @timer: the timer to be added
+ * @expires_nsecs: absolute time in nsecs when timer should expire
+ * @cpu: the CPU to start it on
+ *
+ * This is not very scalable on SMP. Double adds are not possible.
+ */
+void set_timer_on_nsecs(struct timer_list *timer, u64 expires_nsecs, int cpu)
+{
+	tvec_base_t *base = &per_cpu(tvec_bases, cpu);
+  	unsigned long flags;
+
+  	BUG_ON(timer_pending(timer) || !timer->function);
+
+	check_timer(timer);
+
+	spin_lock_irqsave(&base->t_base.lock, flags);
+	timer->expires_nsecs = expires_nsecs;
+	timer->base = &base->t_base;
+	internal_add_timer(base, timer);
+	spin_unlock_irqrestore(&base->t_base.lock, flags);
+}
 
 /***
  * del_timer - deactive a timer.
@@ -455,17 +623,17 @@ static int cascade(tvec_base_t *base, tv
  * This function cascades all vectors and executes all expired timer
  * vectors.
  */
-#define INDEX(N) (base->timer_jiffies >> (TVR_BITS + N * TVN_BITS)) & TVN_MASK
+#define INDEX(N) (base->last_timer_time >> (TVR_BITS + N * TVN_BITS)) & TVN_MASK
 
-static inline void __run_timers(tvec_base_t *base)
+static inline void __run_timers(tvec_base_t *base, unsigned long current_timer_time)
 {
 	struct timer_list *timer;
 
 	spin_lock_irq(&base->t_base.lock);
-	while (time_after_eq(jiffies, base->timer_jiffies)) {
+	while (time_after_eq(current_timer_time, base->last_timer_time)) {
 		struct list_head work_list = LIST_HEAD_INIT(work_list);
 		struct list_head *head = &work_list;
- 		int index = base->timer_jiffies & TVR_MASK;
+ 		int index = base->last_timer_time & TVR_MASK;
  
 		/*
 		 * Cascade timers:
@@ -475,7 +643,7 @@ static inline void __run_timers(tvec_bas
 				(!cascade(base, &base->tv3, INDEX(1))) &&
 					!cascade(base, &base->tv4, INDEX(2)))
 			cascade(base, &base->tv5, INDEX(3));
-		++base->timer_jiffies; 
+		++base->last_timer_time;
 		list_splice_init(base->tv1.vec + index, &work_list);
 		while (!list_empty(head)) {
 			void (*fn)(unsigned long);
@@ -524,20 +692,20 @@ unsigned long next_timer_interrupt(void)
 
 	base = &__get_cpu_var(tvec_bases);
 	spin_lock(&base->t_base.lock);
-	expires = base->timer_jiffies + (LONG_MAX >> 1);
+	expires = base->last_timer_time + (LONG_MAX >> 1);
 	list = 0;
 
 	/* Look for timer events in tv1. */
-	j = base->timer_jiffies & TVR_MASK;
+	j = base->last_timer_time & TVR_MASK;
 	do {
 		list_for_each_entry(nte, base->tv1.vec + j, entry) {
 			expires = nte->expires;
-			if (j < (base->timer_jiffies & TVR_MASK))
+			if (j < (base->last_timer_time & TVR_MASK))
 				list = base->tv2.vec + (INDEX(0));
 			goto found;
 		}
 		j = (j + 1) & TVR_MASK;
-	} while (j != (base->timer_jiffies & TVR_MASK));
+	} while (j != (base->last_timer_time & TVR_MASK));
 
 	/* Check tv2-tv5. */
 	varray[0] = &base->tv2;
@@ -910,10 +1078,15 @@ EXPORT_SYMBOL(xtime_lock);
  */
 static void run_timer_softirq(struct softirq_action *h)
 {
+	unsigned long current_timer_time;
 	tvec_base_t *base = &__get_cpu_var(tvec_bases);
 
-	if (time_after_eq(jiffies, base->timer_jiffies))
-		__run_timers(base);
+	/* cache the converted current time, rounding down */
+	current_timer_time =
+		nsecs_to_timerintervals_floor(do_monotonic_clock());
+
+	if (time_after_eq(current_timer_time, base->last_timer_time))
+		__run_timers(base, current_timer_time);
 }
 
 /*