time: logrithmic time accumulation
PATCH 1/2 time: logrithmic time accumulation
Accumulating one tick at a time works well unless we're using NOHZ. Then it can be an issue, since we may have to run through the loop a few thousand times, which can increase timer interrupt caused latency.
The current solution was to accumulate in half-second intervals with NOHZ. This kept the number of loops down, however it did slightly change how we make NTP adjustments. While not an issue with NTPd users, as NTPd makes adjustments over a longer period of time, other adjtimex() users have noticed the half-second granularity with which we can apply frequency changes to the clock.
For instance, if a application tries to apply a 100ppm frequency correction for 20ms to correct a 2us offset, with NOHZ they either get no correction, or a 50us correction.
Now, there will always be some granularity error for applying frequency corrections. However with users sensitive to this error have seen a 50-500x increase with NOHZ compared to running without NOHZ.
So I figured I'd try another approach then just simply increasing the interval. My approach is to consume the time interval logarithmically. This reduces the number of times through the loop needed keeping latency down, while still preserving the original granularity error for adjtimex() changes.
Further, this change allows us to remove the xtime_cache code (patch to follow), as xtime is always within one tick of the current time, instead of the half-second updates it saw before.
An earlier version of this patch has been shipping to x86 users in the RedHat MRG releases for awhile without issue, but I've reworked this version to be even more careful about avoiding possible overflows if the shift value gets too large.
Since this is not the most trivial code, and its slightly different then whats been tested for awhile, it would be good to get this into some trees for testing. Be it -tip or -mm, either would work. If there's no problems it could be a 2.6.33 or 2.6.34 item.
Any comments or feedback would be appreciated!
include/linux/timex.h +0 -4
@@ -261,11 +261,7 @@ static inline int ntp_synced(void)#define NTP_SCALE_SHIFT 32 -#ifdef CONFIG_NO_HZ -#define NTP_INTERVAL_FREQ (2) -#else #define NTP_INTERVAL_FREQ (HZ) -#endif #define NTP_INTERVAL_LENGTH (NSEC_PER_SEC/NTP_INTERVAL_FREQ)/* Returns how long ticks are at present, in ns / 2^NTP_SCALE_SHIFT. */
kernel/time/timekeeping.c +60 -25
105 lines not on this page · raw message
PATCH 2/2 time: remove xtime_cache
With the prior logarithmic time accumulation patch, xtime will now always be within one "tick" of the current time, instead of possibly half a second off.
This removes the need for the xtime_cache value, which always stored the time at the last interrupt, so this patch cleans that up removing the xtime_cache related code.
This is a bit simpler, but still could use some wider testing.
Any comments or feedback would be appreciated!
thanks
-john
kernel/time.c +0 -1
@@ -136,7 +136,6 @@ static inline void warp_clock(void)write_seqlock_irq (&xtime_lock); wall_to_monotonic.tv_sec -= sys_tz.tz_minuteswest *60 ; xtime.tv_sec += sys_tz.tz_minuteswest *60 ; -update_xtime_cache (0 );write_sequnlock_irq (&xtime_lock);clock_was_set (); }
kernel/time/timekeeping.c +4 -23
@@ -164,13 +164,6 @@ struct timespec raw_time;/* flag for if timekeeping is suspended */ int __read_mostly timekeeping_suspended; -static struct timespec xtime_cache __attribute__ ((aligned (16))); -void update_xtime_cache (u64 nsec) -{ - xtime_cache = xtime; -timespec_add_ns (&xtime_cache, nsec); -} -/* must hold xtime_lock */ void timekeeping_leap_insert (int leapsecond) {
@@ -331,8 +324,6 @@ int do_settimeofday(struct timespec *tv) xtime = *tv; -update_xtime_cache (0 ); - timekeeper.ntp_error =0 ;ntp_clear ();
@@ -547,7 +538,6 @@ void __init timekeeping_init(void) }set_normalized_timespec (&wall_to_monotonic, -boot.tv_sec, -boot.tv_nsec); -update_xtime_cache (0 ); total_sleep_time.tv_sec =0 ; total_sleep_time.tv_nsec =0 ;write_sequnlock_irqrestore (&xtime_lock, flags);
@@ -581,7 +571,6 @@ static int timekeeping_resume(struct sys_device *dev) wall_to_monotonic =timespec_sub (wall_to_monotonic, ts); total_sleep_time =timespec_add_safe (total_sleep_time, ts); } -update_xtime_cache (0 );/* re-base the last cycle value */ timekeeper.clock->cycle_last = timekeeper.clock->read (timekeeper.clock); timekeeper.ntp_error =0 ;
@@ -721,7 +710,6 @@ static void timekeeping_adjust(s64 offset) timekeeper.ntp_error_shift; } -/** * logarithmic_accumulation - shifted accumulation of cycles *
@@ -765,7 +753,6 @@ static cycle_t logarithmic_accumulation(cycle_t offset, int shift)return offset; } -/** * update_wall_time - Uses the current clocksource to increment the wall time *
@@ -775,7 +762,6 @@ void update_wall_time(void) {struct clocksource *clock;cycle_t offset; -u64 nsecs;int shift =0 , maxshift;/* Make sure we're fully resumed: */
@@ -841,9 +827,6 @@ void update_wall_time(void) timekeeper.ntp_error += timekeeper.xtime_nsec << timekeeper.ntp_error_shift; - nsecs =clocksource_cyc2ns (offset, timekeeper.mult, timekeeper.shift); -update_xtime_cache (nsecs); -/* check to see if there is a new clocksource to use */ update_vsyscall (&xtime, timekeeper.clock); }
@@ -880,13 +863,13 @@ void monotonic_to_bootbased(struct timespec *ts)unsigned long get_seconds (void ) { -return xtime_cache.tv_sec; +return xtime.tv_sec; }EXPORT_SYMBOL (get_seconds);struct timespec __current_kernel_time (void ) { -return xtime_cache; +return xtime; }struct timespec current_kernel_time (void )
@@ -896,8 +879,7 @@ struct timespec current_kernel_time(void)do { seq =read_seqbegin (&xtime_lock); - - now = xtime_cache; + now = xtime; }while (read_seqretry (&xtime_lock, seq));return now;
@@ -911,8 +893,7 @@ struct timespec get_monotonic_coarse(void)do { seq =read_seqbegin (&xtime_lock); - - now = xtime_cache; + now = xtime; mono = wall_to_monotonic; }while (read_seqretry (&xtime_lock, seq));
[tip:timers/core] time: Remove xtime_cache
Commit-ID: 5335c1c371aa32fcef97c4979f04b065772194ab Gitweb: http://git.kernel.org/tip/5335c1c371aa32fcef97c4979f04b065772194ab Author: john stultz <[email protected]> AuthorDate: Fri, 2 Oct 2009 16:24:15 -0700 Committer: Ingo Molnar <[email protected]> CommitDate: Sun, 4 Oct 2009 19:31:39 +0200
time: Remove xtime_cache
With the prior logarithmic time accumulation patch, xtime will now always be within one "tick" of the current time, instead of possibly half a second off.
This removes the need for the xtime_cache value, which always stored the time at the last interrupt, so this patch cleans that up removing the xtime_cache related code.
This is a bit simpler, but still could use some wider testing.
LKML-Reference: <[email protected]>
--- kernel/time.c | 1 - kernel/time/timekeeping.c | 27 ++++----------------------- 2 files changed, 4 insertions(+), 24 deletions(-)
kernel/time.c +0 -1
@@ -136,7 +136,6 @@ static inline void warp_clock(void)write_seqlock_irq (&xtime_lock); wall_to_monotonic.tv_sec -= sys_tz.tz_minuteswest *60 ; xtime.tv_sec += sys_tz.tz_minuteswest *60 ; -update_xtime_cache (0 );write_sequnlock_irq (&xtime_lock);clock_was_set (); }
kernel/time/timekeeping.c +4 -23
@@ -164,13 +164,6 @@ struct timespec raw_time;/* flag for if timekeeping is suspended */ int __read_mostly timekeeping_suspended; -static struct timespec xtime_cache __attribute__ ((aligned (16))); -void update_xtime_cache (u64 nsec) -{ - xtime_cache = xtime; -timespec_add_ns (&xtime_cache, nsec); -} -/* must hold xtime_lock */ void timekeeping_leap_insert (int leapsecond) {
@@ -331,8 +324,6 @@ int do_settimeofday(struct timespec *tv) xtime = *tv; -update_xtime_cache (0 ); - timekeeper.ntp_error =0 ;ntp_clear ();
@@ -547,7 +538,6 @@ void __init timekeeping_init(void) }set_normalized_timespec (&wall_to_monotonic, -boot.tv_sec, -boot.tv_nsec); -update_xtime_cache (0 ); total_sleep_time.tv_sec =0 ; total_sleep_time.tv_nsec =0 ;write_sequnlock_irqrestore (&xtime_lock, flags);
@@ -581,7 +571,6 @@ static int timekeeping_resume(struct sys_device *dev) wall_to_monotonic =timespec_sub (wall_to_monotonic, ts); total_sleep_time =timespec_add_safe (total_sleep_time, ts); } -update_xtime_cache (0 );/* re-base the last cycle value */ timekeeper.clock->cycle_last = timekeeper.clock->read (timekeeper.clock); timekeeper.ntp_error =0 ;
@@ -721,7 +710,6 @@ static void timekeeping_adjust(s64 offset) timekeeper.ntp_error_shift; } -/** * logarithmic_accumulation - shifted accumulation of cycles *
@@ -765,7 +753,6 @@ static cycle_t logarithmic_accumulation(cycle_t offset, int shift)return offset; } -/** * update_wall_time - Uses the current clocksource to increment the wall time *
@@ -775,7 +762,6 @@ void update_wall_time(void) {struct clocksource *clock;cycle_t offset; -u64 nsecs;int shift =0 , maxshift;/* Make sure we're fully resumed: */
@@ -841,9 +827,6 @@ void update_wall_time(void) timekeeper.ntp_error += timekeeper.xtime_nsec << timekeeper.ntp_error_shift; - nsecs =clocksource_cyc2ns (offset, timekeeper.mult, timekeeper.shift); -update_xtime_cache (nsecs); -/* check to see if there is a new clocksource to use */ update_vsyscall (&xtime, timekeeper.clock); }
@@ -880,13 +863,13 @@ void monotonic_to_bootbased(struct timespec *ts)unsigned long get_seconds (void ) { -return xtime_cache.tv_sec; +return xtime.tv_sec; }EXPORT_SYMBOL (get_seconds);struct timespec __current_kernel_time (void ) { -return xtime_cache; +return xtime; }struct timespec current_kernel_time (void )
@@ -896,8 +879,7 @@ struct timespec current_kernel_time(void)do { seq =read_seqbegin (&xtime_lock); - - now = xtime_cache; + now = xtime; }while (read_seqretry (&xtime_lock, seq));return now;
@@ -911,8 +893,7 @@ struct timespec get_monotonic_coarse(void)do { seq =read_seqbegin (&xtime_lock); - - now = xtime_cache; + now = xtime; mono = wall_to_monotonic; }while (read_seqretry (&xtime_lock, seq));
[tip:timers/core] time: Remove xtime_cache
In reply to john stultzCommit-ID: 7bc7d637452383d56ba4368d4336b0dde1bb476d Gitweb: http://git.kernel.org/tip/7bc7d637452383d56ba4368d4336b0dde1bb476d Author: john stultz <[email protected]> AuthorDate: Fri, 2 Oct 2009 16:24:15 -0700 Committer: Ingo Molnar <[email protected]> CommitDate: Mon, 5 Oct 2009 13:52:02 +0200
time: Remove xtime_cache
With the prior logarithmic time accumulation patch, xtime will now always be within one "tick" of the current time, instead of possibly half a second off.
This removes the need for the xtime_cache value, which always stored the time at the last interrupt, so this patch cleans that up removing the xtime_cache related code.
This is a bit simpler, but still could use some wider testing.
LKML-Reference: <[email protected]>
--- kernel/time.c | 1 - kernel/time/timekeeping.c | 27 ++++----------------------- 2 files changed, 4 insertions(+), 24 deletions(-)
kernel/time.c +0 -1
@@ -136,7 +136,6 @@ static inline void warp_clock(void)write_seqlock_irq (&xtime_lock); wall_to_monotonic.tv_sec -= sys_tz.tz_minuteswest *60 ; xtime.tv_sec += sys_tz.tz_minuteswest *60 ; -update_xtime_cache (0 );write_sequnlock_irq (&xtime_lock);clock_was_set (); }
kernel/time/timekeeping.c +4 -23
@@ -164,13 +164,6 @@ struct timespec raw_time;/* flag for if timekeeping is suspended */ int __read_mostly timekeeping_suspended; -static struct timespec xtime_cache __attribute__ ((aligned (16))); -void update_xtime_cache (u64 nsec) -{ - xtime_cache = xtime; -timespec_add_ns (&xtime_cache, nsec); -} -/* must hold xtime_lock */ void timekeeping_leap_insert (int leapsecond) {
@@ -331,8 +324,6 @@ int do_settimeofday(struct timespec *tv) xtime = *tv; -update_xtime_cache (0 ); - timekeeper.ntp_error =0 ;ntp_clear ();
@@ -547,7 +538,6 @@ void __init timekeeping_init(void) }set_normalized_timespec (&wall_to_monotonic, -boot.tv_sec, -boot.tv_nsec); -update_xtime_cache (0 ); total_sleep_time.tv_sec =0 ; total_sleep_time.tv_nsec =0 ;write_sequnlock_irqrestore (&xtime_lock, flags);
@@ -581,7 +571,6 @@ static int timekeeping_resume(struct sys_device *dev) wall_to_monotonic =timespec_sub (wall_to_monotonic, ts); total_sleep_time =timespec_add_safe (total_sleep_time, ts); } -update_xtime_cache (0 );/* re-base the last cycle value */ timekeeper.clock->cycle_last = timekeeper.clock->read (timekeeper.clock); timekeeper.ntp_error =0 ;
@@ -721,7 +710,6 @@ static void timekeeping_adjust(s64 offset) timekeeper.ntp_error_shift; } -/** * logarithmic_accumulation - shifted accumulation of cycles *
@@ -765,7 +753,6 @@ static cycle_t logarithmic_accumulation(cycle_t offset, int shift)return offset; } -/** * update_wall_time - Uses the current clocksource to increment the wall time *
@@ -775,7 +762,6 @@ void update_wall_time(void) {struct clocksource *clock;cycle_t offset; -u64 nsecs;int shift =0 , maxshift;/* Make sure we're fully resumed: */
@@ -841,9 +827,6 @@ void update_wall_time(void) timekeeper.ntp_error += timekeeper.xtime_nsec << timekeeper.ntp_error_shift; - nsecs =clocksource_cyc2ns (offset, timekeeper.mult, timekeeper.shift); -update_xtime_cache (nsecs); -/* check to see if there is a new clocksource to use */ update_vsyscall (&xtime, timekeeper.clock); }
@@ -880,13 +863,13 @@ void monotonic_to_bootbased(struct timespec *ts)unsigned long get_seconds (void ) { -return xtime_cache.tv_sec; +return xtime.tv_sec; }EXPORT_SYMBOL (get_seconds);struct timespec __current_kernel_time (void ) { -return xtime_cache; +return xtime; }struct timespec current_kernel_time (void )
@@ -896,8 +879,7 @@ struct timespec current_kernel_time(void)do { seq =read_seqbegin (&xtime_lock); - - now = xtime_cache; + now = xtime; }while (read_seqretry (&xtime_lock, seq));return now;
@@ -911,8 +893,7 @@ struct timespec get_monotonic_coarse(void)do { seq =read_seqbegin (&xtime_lock); - - now = xtime_cache; + now = xtime; mono = wall_to_monotonic; }while (read_seqretry (&xtime_lock, seq));
Re: [PATCH 2/2] time: remove xtime_cache
In reply to john stultz@John Stultz
I also backported this patch to 2.6.31.2-rt13, could you please look it over and see if it looks okay to you?
@Thomas
Same as the previous patch, please consider queuing this for -rt14 and we can drop it in the future because John Stultz submitted it upstream.
Thanks
Quoted 1 line
From
868ee3f7346a90ae3b529ac24996f059cc322a82Mon Sep 17 00:00:00 2001
From: tip-bot for john stultz <[email protected]>
Date: Mon, 5 Oct 2009 11:54:53 +0000
Subject: [PATCH] time: Remove xtime_cache
Commit-ID: 7bc7d637452383d56ba4368d4336b0dde1bb476d Gitweb: http://git.kernel.org/tip/7bc7d637452383d56ba4368d4336b0dde1bb476d Author: john stultz <[email protected]> AuthorDate: Fri, 2 Oct 2009 16:24:15 -0700 Committer: Ingo Molnar <[email protected]> CommitDate: Mon, 5 Oct 2009 13:52:02 +0200
time: Remove xtime_cache
With the prior logarithmic time accumulation patch, xtime will now always be within one "tick" of the current time, instead of possibly half a second off.
This removes the need for the xtime_cache value, which always stored the time at the last interrupt, so this patch cleans that up removing the xtime_cache related code.
This is a bit simpler, but still could use some wider testing.
LKML-Reference: <[email protected]>
--- kernel/time.c | 1 - kernel/time/timekeeping.c | 20 ++------------------ 2 files changed, 2 insertions(+), 19 deletions(-)
kernel/time.c +0 -1
@@ -136,7 +136,6 @@ static inline void warp_clock(void)write_atomic_seqlock_irq (&xtime_lock); wall_to_monotonic.tv_sec -= sys_tz.tz_minuteswest *60 ; xtime.tv_sec += sys_tz.tz_minuteswest *60 ; -update_xtime_cache (0 );write_atomic_sequnlock_irq (&xtime_lock);clock_was_set (); }
kernel/time/timekeeping.c +0 -15
@@ -48,16 +48,8 @@ static unsigned long total_sleep_time; /* seconds *//* flag for if timekeeping is suspended */ int __read_mostly timekeeping_suspended; -static struct timespec xtime_cache __attribute__ ((aligned (16))); -void update_xtime_cache (u64 nsec) -{ - xtime_cache = xtime; -timespec_add_ns (&xtime_cache, nsec); -} -struct clocksource *clock; -#ifdef CONFIG_GENERIC_TIME /** * clocksource_forward_now - update clock to the current time
@@ -233,7 +225,6 @@ int do_settimeofday(struct timespec *tv) xtime = *tv; -update_xtime_cache (0 ); clock->error =0 ;ntp_clear ();
@@ -435,7 +426,6 @@ void __init timekeeping_init(void) xtime.tv_nsec =0 ;set_normalized_timespec (&wall_to_monotonic, -xtime.tv_sec, -xtime.tv_nsec); -update_xtime_cache (0 ); total_sleep_time =0 ;write_atomic_sequnlock_irqrestore (&xtime_lock, flags); }
@@ -467,7 +457,6 @@ static int timekeeping_resume(struct sys_device *dev) wall_to_monotonic.tv_sec -= sleep_length; total_sleep_time += sleep_length; } -update_xtime_cache (0 );/* re-base the last cycle value */ clock->cycle_last =0 ; clock->cycle_last =clocksource_read (clock);
@@ -608,7 +597,6 @@ static void clocksource_adjust(s64 offset) (NTP_SCALE_SHIFT - clock->shift); } -/** * logarithmic_accumulation - shifted accumulation of cycles *
@@ -652,7 +640,6 @@ static cycle_t logarithmic_accumulation(cycle_t offset, int shift)return offset; } -/** * update_wall_time - Uses the current clocksource to increment the wall time *
@@ -661,7 +648,6 @@ static cycle_t logarithmic_accumulation(cycle_t offset, int shift)void update_wall_time (void ) {cycle_t offset; -u64 nsecs;int shift =0 , maxshift;/* Make sure we're fully resumed: */
@@ -725,8 +711,6 @@ void update_wall_time(void) clock->xtime_nsec -= (s64 )xtime.tv_nsec << clock->shift; clock->error += clock->xtime_nsec << (NTP_SCALE_SHIFT - clock->shift); -update_xtime_cache (cyc2ns (clock, offset));
Quoted 164 lines · john stultz
Accumulating one tick at a time works well unless we're using NOHZ. Then it can be an issue, since we may have to run through the loop a few thousand times, which can increase timer interrupt caused latency.
The current solution was to accumulate in half-second intervals with NOHZ. This kept the number of loops down, however it did slightly change how we make NTP adjustments. While not an issue with NTPd users, as NTPd makes adjustments over a longer period of time, other
adjtimex()users have noticed the half-second granularity with which we can apply frequency changes to the clock.For instance, if a application tries to apply a 100ppm frequency correction for 20ms to correct a 2us offset, with NOHZ they either get no correction, or a 50us correction.
Now, there will always be some granularity error for applying frequency corrections. However with users sensitive to this error have seen a 50-500x increase with NOHZ compared to running without NOHZ.
So I figured I'd try another approach then just simply increasing the interval. My approach is to consume the time interval logarithmically. This reduces the number of times through the loop needed keeping latency down, while still preserving the original granularity error for
adjtimex()changes.Further, this change allows us to remove the
xtime_cachecode (patch to follow), as xtime is always within one tick of the current time, instead of the half-second updates it saw before.An earlier version of this patch has been shipping to x86 users in the RedHat MRG releases for awhile without issue, but I've reworked this version to be even more careful about avoiding possible overflows if the shift value gets too large.
Since this is not the most trivial code, and its slightly different then whats been tested for awhile, it would be good to get this into some trees for testing. Be it -tip or -mm, either would work. If there's no problems it could be a 2.6.33 or 2.6.34 item.
Any comments or feedback would be appreciated!
Signed-off-by John Stultz <[email protected]>include/linux/timex.h +0
-4@@ -261,11 +261,7 @@ static inline int ntp_synced(void)#define NTP_SCALE_SHIFT 32 -#ifdef CONFIG_NO_HZ -#define NTP_INTERVAL_FREQ (2) -#else #define NTP_INTERVAL_FREQ (HZ) -#endif #define NTP_INTERVAL_LENGTH (NSEC_PER_SEC/NTP_INTERVAL_FREQ)/* Returns how long ticks are at present, in ns / 2^NTP_SCALE_SHIFT. */ kernel/time/timekeeping.c +58
-20@@ -721,6 +721,51 @@ static void timekeeping_adjust(s64 offset) timekeeper.ntp_error_shift; } + +/** +* logarithmic_accumulation - shifted accumulation of cycles +* +* This functions accumulates a shifted interval of cycles into +* into a shifted interval nanoseconds. Allows for O(log) accumulation +* loop. +* +* Returns the unconsumed cycles. +*/ +static cycle_t logarithmic_accumulation (cycle_t offset,int shift) +{ +u64 nsecps = (u64 )NSEC_PER_SEC << timekeeper.shift; + +/* If the offset is smaller then a shifted interval, do nothing */ +if (offset < timekeeper.cycle_interval<<shift) +return offset; + +/* accumulate one shifted interval */ + offset -= timekeeper.cycle_interval << shift; + timekeeper.clock->cycle_last += timekeeper.cycle_interval << shift; + + timekeeper.xtime_nsec += timekeeper.xtime_interval << shift; +while (timekeeper.xtime_nsec >= nsecps) { + timekeeper.xtime_nsec -= nsecps; + xtime.tv_sec++; +second_overflow (); + } + +/* accumulate into raw time */ + raw_time.tv_nsec += timekeeper.raw_interval << shift;; +while (raw_time.tv_nsec >=NSEC_PER_SEC ) { + raw_time.tv_nsec -=NSEC_PER_SEC ; + raw_time.tv_sec++; + } + +/* accumulate error between NTP and clock interval */ + timekeeper.ntp_error += tick_length << shift; + timekeeper.ntp_error -= timekeeper.xtime_interval << + (timekeeper.ntp_error_shift + shift); + +return offset; +} + +/** * update_wall_time - Uses the current clocksource to increment the wall time * @@ -731,6 +776,7 @@ void update_wall_time(void)struct clocksource *clock;cycle_t offset;u64 nsecs; +int shift =0 , maxshift;/* Make sure we're fully resumed: */ if (unlikely (timekeeping_suspended))@@ -744,33 +790,22 @@ void update_wall_time(void)#endif timekeeper.xtime_nsec = (s64 )xtime.tv_nsec << timekeeper.shift; -/* normally this loop will run just once, however in the -* case of lost or late ticks, it will accumulate correctly. +/* +* With NO_HZ we may have to accumulate many cycle_intervals +* (think "ticks") worth of time at once. To do this efficiently, +* we calculate the largest doubling multiple of cycle_intervals +* that is smaller then the offset. We then accumulate that +* chunk in one go, and then try to consume the next smaller +* doubled multiple. */ + shift =ilog2 (offset) -ilog2 (timekeeper.cycle_interval); + shift =max (0 , shift); +/* Bound shift to one less then what overflows tick_length */ + maxshift = (8 *sizeof (tick_length) - (ilog2 (tick_length)+1 )) -1 ; + shift =min (shift, maxshift);while (offset >= timekeeper.cycle_interval) { -u64 nsecps = (u64 )NSEC_PER_SEC << timekeeper.shift; - -/* accumulate one interval */ - offset -= timekeeper.cycle_interval; - clock->cycle_last += timekeeper.cycle_interval; - - timekeeper.xtime_nsec += timekeeper.xtime_interval; -if (timekeeper.xtime_nsec >= nsecps) { - timekeeper.xtime_nsec -= nsecps; - xtime.tv_sec++; -second_overflow (); - } - - raw_time.tv_nsec += timekeeper.raw_interval; -if (raw_time.tv_nsec >=NSEC_PER_SEC ) { - raw_time.tv_nsec -=NSEC_PER_SEC ; - raw_time.tv_sec++; - }
kernel/time/timekeeping.c void update_wall_time(void) · line 811
Quoted 10 lines
kernel/time/timekeeping.c +2
-5- -/* accumulate error between NTP and clock interval */ - timekeeper.ntp_error += tick_length; - timekeeper.ntp_error -= timekeeper.xtime_interval << - timekeeper.ntp_error_shift; + offset =logarithmic_accumulation (offset, shift); + shift--; }/* correct the clock when NTP error is too big */
There are several (6) trailing whitespace errors that checkpatch exposes, but other than that:
Quoted 174 lines · John Kacur
Quoted 170 lines · john stultz
Accumulating one tick at a time works well unless we're using NOHZ. Then it can be an issue, since we may have to run through the loop a few thousand times, which can increase timer interrupt caused latency.
The current solution was to accumulate in half-second intervals with NOHZ. This kept the number of loops down, however it did slightly change how we make NTP adjustments. While not an issue with NTPd users, as NTPd makes adjustments over a longer period of time, other
adjtimex()users have noticed the half-second granularity with which we can apply frequency changes to the clock.For instance, if a application tries to apply a 100ppm frequency correction for 20ms to correct a 2us offset, with NOHZ they either get no correction, or a 50us correction.
Now, there will always be some granularity error for applying frequency corrections. However with users sensitive to this error have seen a 50-500x increase with NOHZ compared to running without NOHZ.
So I figured I'd try another approach then just simply increasing the interval. My approach is to consume the time interval logarithmically. This reduces the number of times through the loop needed keeping latency down, while still preserving the original granularity error for
adjtimex()changes.Further, this change allows us to remove the
xtime_cachecode (patch to follow), as xtime is always within one tick of the current time, instead of the half-second updates it saw before.An earlier version of this patch has been shipping to x86 users in the RedHat MRG releases for awhile without issue, but I've reworked this version to be even more careful about avoiding possible overflows if the shift value gets too large.
Since this is not the most trivial code, and its slightly different then whats been tested for awhile, it would be good to get this into some trees for testing. Be it -tip or -mm, either would work. If there's no problems it could be a 2.6.33 or 2.6.34 item.
Any comments or feedback would be appreciated!
Signed-off-by John Stultz <[email protected]>include/linux/timex.h +0
-4@@ -261,11 +261,7 @@ static inline int ntp_synced(void)#define NTP_SCALE_SHIFT 32 -#ifdef CONFIG_NO_HZ -#define NTP_INTERVAL_FREQ (2) -#else #define NTP_INTERVAL_FREQ (HZ) -#endif #define NTP_INTERVAL_LENGTH (NSEC_PER_SEC/NTP_INTERVAL_FREQ)/* Returns how long ticks are at present, in ns / 2^NTP_SCALE_SHIFT. */ kernel/time/timekeeping.c +60
-25@@ -721,6 +721,51 @@ static void timekeeping_adjust(s64 offset) timekeeper.ntp_error_shift; } + +/** +* logarithmic_accumulation - shifted accumulation of cycles +* +* This functions accumulates a shifted interval of cycles into +* into a shifted interval nanoseconds. Allows for O(log) accumulation +* loop. +* +* Returns the unconsumed cycles. +*/ +static cycle_t logarithmic_accumulation (cycle_t offset,int shift) +{ +u64 nsecps = (u64 )NSEC_PER_SEC << timekeeper.shift; + +/* If the offset is smaller then a shifted interval, do nothing */ +if (offset < timekeeper.cycle_interval<<shift) +return offset; + +/* accumulate one shifted interval */ + offset -= timekeeper.cycle_interval << shift; + timekeeper.clock->cycle_last += timekeeper.cycle_interval << shift; + + timekeeper.xtime_nsec += timekeeper.xtime_interval << shift; +while (timekeeper.xtime_nsec >= nsecps) { + timekeeper.xtime_nsec -= nsecps; + xtime.tv_sec++; +second_overflow (); + } + +/* accumulate into raw time */ + raw_time.tv_nsec += timekeeper.raw_interval << shift;; +while (raw_time.tv_nsec >=NSEC_PER_SEC ) { + raw_time.tv_nsec -=NSEC_PER_SEC ; + raw_time.tv_sec++; + } + +/* accumulate error between NTP and clock interval */ + timekeeper.ntp_error += tick_length << shift; + timekeeper.ntp_error -= timekeeper.xtime_interval << + (timekeeper.ntp_error_shift + shift); + +return offset; +} + +/** * update_wall_time - Uses the current clocksource to increment the wall time * @@ -731,6 +776,7 @@ void update_wall_time(void)struct clocksource *clock;cycle_t offset;u64 nsecs; +int shift =0 , maxshift;/* Make sure we're fully resumed: */ if (unlikely (timekeeping_suspended))@@ -744,33 +790,22 @@ void update_wall_time(void)#endif timekeeper.xtime_nsec = (s64 )xtime.tv_nsec << timekeeper.shift; -/* normally this loop will run just once, however in the -* case of lost or late ticks, it will accumulate correctly. +/* +* With NO_HZ we may have to accumulate many cycle_intervals +* (think "ticks") worth of time at once. To do this efficiently, +* we calculate the largest doubling multiple of cycle_intervals +* that is smaller then the offset. We then accumulate that +* chunk in one go, and then try to consume the next smaller +* doubled multiple. */ + shift =ilog2 (offset) -ilog2 (timekeeper.cycle_interval); + shift =max (0 , shift); +/* Bound shift to one less then what overflows tick_length */ + maxshift = (8 *sizeof (tick_length) - (ilog2 (tick_length)+1 )) -1 ; + shift =min (shift, maxshift);while (offset >= timekeeper.cycle_interval) { -u64 nsecps = (u64 )NSEC_PER_SEC << timekeeper.shift; - -/* accumulate one interval */ - offset -= timekeeper.cycle_interval; - clock->cycle_last += timekeeper.cycle_interval; - - timekeeper.xtime_nsec += timekeeper.xtime_interval; -if (timekeeper.xtime_nsec >= nsecps) { - timekeeper.xtime_nsec -= nsecps; - xtime.tv_sec++; -second_overflow (); - } - - raw_time.tv_nsec += timekeeper.raw_interval; -if (raw_time.tv_nsec >=NSEC_PER_SEC ) { - raw_time.tv_nsec -=NSEC_PER_SEC ; - raw_time.tv_sec++; - } - -/* accumulate error between NTP and clock interval */ - timekeeper.ntp_error += tick_length; - timekeeper.ntp_error -= timekeeper.xtime_interval << - timekeeper.ntp_error_shift; + offset =logarithmic_accumulation (offset, shift); + shift--; }
Quoted 7 lines
kernel/time/timekeeping.c void update_wall_time(void) · line 811
Quoted 4 lines
kernel/time/timekeeping.c +0
-0/* correct the clock when NTP error is too big */ There are several (6) trailing whitespace errors that checkpatch exposes, but other than that:
Yep, i fixed those already. (also the few inconsistent comment capitalizations)
The two patches held up fine in -tip testing so far.
Quoted 1 line
Signed-off-by John Kacur <[email protected]>
Thanks - i changed that to Reviewed-by (SoB is for being part of the patch-flow) and added it to the two commits.
Ingo
@John Stultz
I backported your patch to 2.6.31.2-rt13, could you please look it over and see if it looks okay to you?
@Thomas
Could you please consider queuing this up for -rt14.
Since John submitted it upstream, we will be able to drop it again in the future.
Thanks
Quoted 1 line
From
8090f669e58901c1b0c5e8bac4160eaaf7990f4dMon Sep 17 00:00:00 2001
From: tip-bot for john stultz <[email protected]>
Date: Mon, 5 Oct 2009 11:54:38 +0000
Subject: [PATCH] time: Implement logarithmic time accumulation
Commit-ID: a092ff0f90cae22b2ac8028ecd2c6f6c1a9e4601 Gitweb: http://git.kernel.org/tip/a092ff0f90cae22b2ac8028ecd2c6f6c1a9e4601 Author: john stultz <[email protected]> AuthorDate: Fri, 2 Oct 2009 16:17:53 -0700 Committer: Ingo Molnar <[email protected]> CommitDate: Mon, 5 Oct 2009 13:51:48 +0200
time: Implement logarithmic time accumulation
Accumulating one tick at a time works well unless we're using NOHZ. Then it can be an issue, since we may have to run through the loop a few thousand times, which can increase timer interrupt caused latency.
The current solution was to accumulate in half-second intervals with NOHZ. This kept the number of loops down, however it did slightly change how we make NTP adjustments. While not an issue with NTPd users, as NTPd makes adjustments over a longer period of time, other adjtimex() users have noticed the half-second granularity with which we can apply frequency changes to the clock.
For instance, if a application tries to apply a 100ppm frequency correction for 20ms to correct a 2us offset, with NOHZ they either get no correction, or a 50us correction.
Now, there will always be some granularity error for applying frequency corrections. However with users sensitive to this error have seen a 50-500x increase with NOHZ compared to running without NOHZ.
So I figured I'd try another approach then just simply increasing the interval. My approach is to consume the time interval logarithmically. This reduces the number of times through the loop needed keeping latency down, while still preserving the original granularity error for adjtimex() changes.
Further, this change allows us to remove the xtime_cache code (patch to follow), as xtime is always within one tick of the current time, instead of the half-second updates it saw before.
An earlier version of this patch has been shipping to x86 users in the RedHat MRG releases for awhile without issue, but I've reworked this version to be even more careful about avoiding possible overflows if the shift value gets too large.
LKML-Reference: <[email protected]>
--- include/linux/timex.h | 4 -- kernel/time/timekeeping.c | 83 +++++++++++++++++++++++++++++++++------------ 2 files changed, 61 insertions(+), 26 deletions(-)
include/linux/timex.h +0 -4
@@ -261,11 +261,7 @@ static inline int ntp_synced(void)#define NTP_SCALE_SHIFT 32 -#ifdef CONFIG_NO_HZ -#define NTP_INTERVAL_FREQ (2) -#else #define NTP_INTERVAL_FREQ (HZ) -#endif #define NTP_INTERVAL_LENGTH (NSEC_PER_SEC/NTP_INTERVAL_FREQ)/* Returns how long ticks are at present, in ns / 2^NTP_SCALE_SHIFT. */
kernel/time/timekeeping.c +59 -18
94 lines not on this page · raw message
Quoted 71 lines · On Mon, 5 Oct 2009 12:43:37 +0200 (CEST) John Kacur wrote:
Quoted 67 lines · john stultz
Accumulating one tick at a time works well unless we're using NOHZ. Then it can be an issue, since we may have to run through the loop a few thousand times, which can increase timer interrupt caused latency.
The current solution was to accumulate in half-second intervals with NOHZ. This kept the number of loops down, however it did slightly change how we make NTP adjustments. While not an issue with NTPd users, as NTPd makes adjustments over a longer period of time, other
adjtimex()users have noticed the half-second granularity with which we can apply frequency changes to the clock.For instance, if a application tries to apply a 100ppm frequency correction for 20ms to correct a 2us offset, with NOHZ they either get no correction, or a 50us correction.
Now, there will always be some granularity error for applying frequency corrections. However with users sensitive to this error have seen a 50-500x increase with NOHZ compared to running without NOHZ.
So I figured I'd try another approach then just simply increasing the interval. My approach is to consume the time interval logarithmically. This reduces the number of times through the loop needed keeping latency down, while still preserving the original granularity error for
adjtimex()changes.Further, this change allows us to remove the
xtime_cachecode (patch to follow), as xtime is always within one tick of the current time, instead of the half-second updates it saw before.An earlier version of this patch has been shipping to x86 users in the RedHat MRG releases for awhile without issue, but I've reworked this version to be even more careful about avoiding possible overflows if the shift value gets too large.
Since this is not the most trivial code, and its slightly different then whats been tested for awhile, it would be good to get this into some trees for testing. Be it -tip or -mm, either would work. If there's no problems it could be a 2.6.33 or 2.6.34 item.
Any comments or feedback would be appreciated!
Signed-off-by John Stultz <[email protected]>include/linux/timex.h +0
-4@@ -261,11 +261,7 @@ static inline int ntp_synced(void)#define NTP_SCALE_SHIFT 32 -#ifdef CONFIG_NO_HZ -#define NTP_INTERVAL_FREQ (2) -#else #define NTP_INTERVAL_FREQ (HZ) -#endif #define NTP_INTERVAL_LENGTH (NSEC_PER_SEC/NTP_INTERVAL_FREQ)/* Returns how long ticks are at present, in ns / 2^NTP_SCALE_SHIFT. */ kernel/time/timekeeping.c +0
-0@@ -721,6 +721,51 @@ static void timekeeping_adjust(s64 offset) timekeeper.ntp_error_shift; }
Quoted 2 lines
kernel/time/timekeeping.c static void timekeeping_adjust(s64 offset) · line 725
Quoted 2 lines
kernel/time/timekeeping.c +2
-0+ +/**
"/**" means "beginning of kernel-doc", so those function parameters need to be described here also...
Quoted 101 lines
Quoted 101 lines
kernel/time/timekeeping.c +58
-25+* logarithmic_accumulation - shifted accumulation of cycles +* +* This functions accumulates a shifted interval of cycles into +* into a shifted interval nanoseconds. Allows for O(log) accumulation +* loop. +* +* Returns the unconsumed cycles. +*/ +static cycle_t logarithmic_accumulation (cycle_t offset,int shift) +{ +u64 nsecps = (u64 )NSEC_PER_SEC << timekeeper.shift; + +/* If the offset is smaller then a shifted interval, do nothing */ +if (offset < timekeeper.cycle_interval<<shift) +return offset; + +/* accumulate one shifted interval */ + offset -= timekeeper.cycle_interval << shift; + timekeeper.clock->cycle_last += timekeeper.cycle_interval << shift; + + timekeeper.xtime_nsec += timekeeper.xtime_interval << shift; +while (timekeeper.xtime_nsec >= nsecps) { + timekeeper.xtime_nsec -= nsecps; + xtime.tv_sec++; +second_overflow (); + } + +/* accumulate into raw time */ + raw_time.tv_nsec += timekeeper.raw_interval << shift;; +while (raw_time.tv_nsec >=NSEC_PER_SEC ) { + raw_time.tv_nsec -=NSEC_PER_SEC ; + raw_time.tv_sec++; + } + +/* accumulate error between NTP and clock interval */ + timekeeper.ntp_error += tick_length << shift; + timekeeper.ntp_error -= timekeeper.xtime_interval << + (timekeeper.ntp_error_shift + shift); + +return offset; +} + +/** * update_wall_time - Uses the current clocksource to increment the wall time * @@ -731,6 +776,7 @@ void update_wall_time(void)struct clocksource *clock;cycle_t offset;u64 nsecs; +int shift =0 , maxshift;/* Make sure we're fully resumed: */ if (unlikely (timekeeping_suspended))@@ -744,33 +790,22 @@ void update_wall_time(void)#endif timekeeper.xtime_nsec = (s64 )xtime.tv_nsec << timekeeper.shift; -/* normally this loop will run just once, however in the -* case of lost or late ticks, it will accumulate correctly. +/* +* With NO_HZ we may have to accumulate many cycle_intervals +* (think "ticks") worth of time at once. To do this efficiently, +* we calculate the largest doubling multiple of cycle_intervals +* that is smaller then the offset. We then accumulate that +* chunk in one go, and then try to consume the next smaller +* doubled multiple. */ + shift =ilog2 (offset) -ilog2 (timekeeper.cycle_interval); + shift =max (0 , shift); +/* Bound shift to one less then what overflows tick_length */ + maxshift = (8 *sizeof (tick_length) - (ilog2 (tick_length)+1 )) -1 ; + shift =min (shift, maxshift);while (offset >= timekeeper.cycle_interval) { -u64 nsecps = (u64 )NSEC_PER_SEC << timekeeper.shift; - -/* accumulate one interval */ - offset -= timekeeper.cycle_interval; - clock->cycle_last += timekeeper.cycle_interval; - - timekeeper.xtime_nsec += timekeeper.xtime_interval; -if (timekeeper.xtime_nsec >= nsecps) { - timekeeper.xtime_nsec -= nsecps; - xtime.tv_sec++; -second_overflow (); - } - - raw_time.tv_nsec += timekeeper.raw_interval; -if (raw_time.tv_nsec >=NSEC_PER_SEC ) { - raw_time.tv_nsec -=NSEC_PER_SEC ; - raw_time.tv_sec++; - } - -/* accumulate error between NTP and clock interval */ - timekeeper.ntp_error += tick_length; - timekeeper.ntp_error -= timekeeper.xtime_interval << - timekeeper.ntp_error_shift; + offset =logarithmic_accumulation (offset, shift); + shift--; }
Quoted 8 lines
kernel/time/timekeeping.c void update_wall_time(void) · line 811
Quoted 4 lines
kernel/time/timekeeping.c +0
-0/* correct the clock when NTP error is too big */ There are several (6) trailing whitespace errors that checkpatch exposes, but other than that:
Signed-off-by John Kacur <[email protected]>
---
~Randy
[tip:timers/core] time: Implement logarithmic time accumulation
In reply to john stultzCommit-ID: 8a45af8eb809a0d3512877296bd128e73fcee379 Gitweb: http://git.kernel.org/tip/8a45af8eb809a0d3512877296bd128e73fcee379 Author: john stultz <[email protected]> AuthorDate: Fri, 2 Oct 2009 16:17:53 -0700 Committer: Ingo Molnar <[email protected]> CommitDate: Sun, 4 Oct 2009 19:31:39 +0200
time: Implement logarithmic time accumulation
Accumulating one tick at a time works well unless we're using NOHZ. Then it can be an issue, since we may have to run through the loop a few thousand times, which can increase timer interrupt caused latency.
The current solution was to accumulate in half-second intervals with NOHZ. This kept the number of loops down, however it did slightly change how we make NTP adjustments. While not an issue with NTPd users, as NTPd makes adjustments over a longer period of time, other adjtimex() users have noticed the half-second granularity with which we can apply frequency changes to the clock.
For instance, if a application tries to apply a 100ppm frequency correction for 20ms to correct a 2us offset, with NOHZ they either get no correction, or a 50us correction.
Now, there will always be some granularity error for applying frequency corrections. However with users sensitive to this error have seen a 50-500x increase with NOHZ compared to running without NOHZ.
So I figured I'd try another approach then just simply increasing the interval. My approach is to consume the time interval logarithmically. This reduces the number of times through the loop needed keeping latency down, while still preserving the original granularity error for adjtimex() changes.
Further, this change allows us to remove the xtime_cache code (patch to follow), as xtime is always within one tick of the current time, instead of the half-second updates it saw before.
An earlier version of this patch has been shipping to x86 users in the RedHat MRG releases for awhile without issue, but I've reworked this version to be even more careful about avoiding possible overflows if the shift value gets too large.
LKML-Reference: <[email protected]>
--- include/linux/timex.h | 4 -- kernel/time/timekeeping.c | 85 +++++++++++++++++++++++++++++++------------- 2 files changed, 60 insertions(+), 29 deletions(-)
include/linux/timex.h +0 -4
@@ -261,11 +261,7 @@ static inline int ntp_synced(void)#define NTP_SCALE_SHIFT 32 -#ifdef CONFIG_NO_HZ -#define NTP_INTERVAL_FREQ (2) -#else #define NTP_INTERVAL_FREQ (HZ) -#endif #define NTP_INTERVAL_LENGTH (NSEC_PER_SEC/NTP_INTERVAL_FREQ)/* Returns how long ticks are at present, in ns / 2^NTP_SCALE_SHIFT. */
kernel/time/timekeeping.c +60 -25
105 lines not on this page · raw message
[tip:timers/core] time: Implement logarithmic time accumulation
In reply to john stultzCommit-ID: a092ff0f90cae22b2ac8028ecd2c6f6c1a9e4601 Gitweb: http://git.kernel.org/tip/a092ff0f90cae22b2ac8028ecd2c6f6c1a9e4601 Author: john stultz <[email protected]> AuthorDate: Fri, 2 Oct 2009 16:17:53 -0700 Committer: Ingo Molnar <[email protected]> CommitDate: Mon, 5 Oct 2009 13:51:48 +0200
time: Implement logarithmic time accumulation
Accumulating one tick at a time works well unless we're using NOHZ. Then it can be an issue, since we may have to run through the loop a few thousand times, which can increase timer interrupt caused latency.
The current solution was to accumulate in half-second intervals with NOHZ. This kept the number of loops down, however it did slightly change how we make NTP adjustments. While not an issue with NTPd users, as NTPd makes adjustments over a longer period of time, other adjtimex() users have noticed the half-second granularity with which we can apply frequency changes to the clock.
For instance, if a application tries to apply a 100ppm frequency correction for 20ms to correct a 2us offset, with NOHZ they either get no correction, or a 50us correction.
Now, there will always be some granularity error for applying frequency corrections. However with users sensitive to this error have seen a 50-500x increase with NOHZ compared to running without NOHZ.
So I figured I'd try another approach then just simply increasing the interval. My approach is to consume the time interval logarithmically. This reduces the number of times through the loop needed keeping latency down, while still preserving the original granularity error for adjtimex() changes.
Further, this change allows us to remove the xtime_cache code (patch to follow), as xtime is always within one tick of the current time, instead of the half-second updates it saw before.
An earlier version of this patch has been shipping to x86 users in the RedHat MRG releases for awhile without issue, but I've reworked this version to be even more careful about avoiding possible overflows if the shift value gets too large.
LKML-Reference: <[email protected]>
--- include/linux/timex.h | 4 -- kernel/time/timekeeping.c | 85 +++++++++++++++++++++++++++++++------------- 2 files changed, 60 insertions(+), 29 deletions(-)
include/linux/timex.h +0 -4
@@ -261,11 +261,7 @@ static inline int ntp_synced(void)#define NTP_SCALE_SHIFT 32 -#ifdef CONFIG_NO_HZ -#define NTP_INTERVAL_FREQ (2) -#else #define NTP_INTERVAL_FREQ (HZ) -#endif #define NTP_INTERVAL_LENGTH (NSEC_PER_SEC/NTP_INTERVAL_FREQ)/* Returns how long ticks are at present, in ns / 2^NTP_SCALE_SHIFT. */
kernel/time/timekeeping.c +60 -25
105 lines not on this page · raw message