linux/kernel/time/timekeeping.c

906 lines
23 KiB
C
Raw Normal View History

/*
* linux/kernel/time/timekeeping.c
*
* Kernel timekeeping code and accessor functions
*
* This code was moved from linux/kernel/timer.c.
* Please see that file for copyright and history logs.
*
*/
#include <linux/module.h>
#include <linux/interrupt.h>
#include <linux/percpu.h>
#include <linux/init.h>
#include <linux/mm.h>
#include <linux/sched.h>
#include <linux/sysdev.h>
#include <linux/clocksource.h>
#include <linux/jiffies.h>
#include <linux/time.h>
#include <linux/tick.h>
#include <linux/stop_machine.h>
/* Structure holding internal timekeeping values. */
struct timekeeper {
/* Current clocksource used for timekeeping. */
struct clocksource *clock;
/* The shift value of the current clocksource. */
int shift;
/* Number of clock cycles in one NTP interval. */
cycle_t cycle_interval;
/* Number of clock shifted nano seconds in one NTP interval. */
u64 xtime_interval;
/* shifted nano seconds left over when rounding cycle_interval */
s64 xtime_remainder;
/* Raw nano seconds accumulated per NTP interval. */
u32 raw_interval;
/* Clock shifted nano seconds remainder not stored in xtime.tv_nsec. */
u64 xtime_nsec;
/* Difference between accumulated time and NTP time in ntp
* shifted nano seconds. */
s64 ntp_error;
/* Shift conversion between clock shifted nano seconds and
* ntp shifted nano seconds. */
int ntp_error_shift;
/* NTP adjusted clock multiplier */
u32 mult;
};
struct timekeeper timekeeper;
/**
* timekeeper_setup_internals - Set up internals to use clocksource clock.
*
* @clock: Pointer to clocksource.
*
* Calculates a fixed cycle/nsec interval for a given clocksource/adjustment
* pair and interval request.
*
* Unless you're the timekeeping code, you should not be using this!
*/
static void timekeeper_setup_internals(struct clocksource *clock)
{
cycle_t interval;
u64 tmp, ntpinterval;
timekeeper.clock = clock;
clock->cycle_last = clock->read(clock);
/* Do the ns -> cycle conversion first, using original mult */
tmp = NTP_INTERVAL_LENGTH;
tmp <<= clock->shift;
ntpinterval = tmp;
tmp += clock->mult/2;
do_div(tmp, clock->mult);
if (tmp == 0)
tmp = 1;
interval = (cycle_t) tmp;
timekeeper.cycle_interval = interval;
/* Go back from cycles -> shifted ns */
timekeeper.xtime_interval = (u64) interval * clock->mult;
timekeeper.xtime_remainder = ntpinterval - timekeeper.xtime_interval;
timekeeper.raw_interval =
((u64) interval * clock->mult) >> clock->shift;
timekeeper.xtime_nsec = 0;
timekeeper.shift = clock->shift;
timekeeper.ntp_error = 0;
timekeeper.ntp_error_shift = NTP_SCALE_SHIFT - clock->shift;
/*
* The timekeeper keeps its own mult values for the currently
* active clocksource. These value will be adjusted via NTP
* to counteract clock drifting.
*/
timekeeper.mult = clock->mult;
}
/* Timekeeper helper functions. */
static inline s64 timekeeping_get_ns(void)
{
cycle_t cycle_now, cycle_delta;
struct clocksource *clock;
/* read clocksource: */
clock = timekeeper.clock;
cycle_now = clock->read(clock);
/* calculate the delta since the last update_wall_time: */
cycle_delta = (cycle_now - clock->cycle_last) & clock->mask;
/* return delta convert to nanoseconds using ntp adjusted mult. */
return clocksource_cyc2ns(cycle_delta, timekeeper.mult,
timekeeper.shift);
}
static inline s64 timekeeping_get_ns_raw(void)
{
cycle_t cycle_now, cycle_delta;
struct clocksource *clock;
/* read clocksource: */
clock = timekeeper.clock;
cycle_now = clock->read(clock);
/* calculate the delta since the last update_wall_time: */
cycle_delta = (cycle_now - clock->cycle_last) & clock->mask;
/* return delta convert to nanoseconds using ntp adjusted mult. */
return clocksource_cyc2ns(cycle_delta, clock->mult, clock->shift);
}
/*
* This read-write spinlock protects us from races in SMP while
* playing with xtime.
*/
__cacheline_aligned_in_smp DEFINE_SEQLOCK(xtime_lock);
/*
* The current time
* wall_to_monotonic is what we need to add to xtime (or xtime corrected
* for sub jiffie times) to get to monotonic time. Monotonic is pegged
* at zero at system boot time, so wall_to_monotonic will be negative,
* however, we will ALWAYS keep the tv_nsec part positive so we can use
* the usual normalization.
*
* wall_to_monotonic is moved after resume from suspend for the monotonic
* time not to jump. We need to add total_sleep_time to wall_to_monotonic
* to get the real boot based time offset.
*
* - wall_to_monotonic is no longer the boot time, getboottime must be
* used instead.
*/
static struct timespec xtime __attribute__ ((aligned (16)));
static struct timespec wall_to_monotonic __attribute__ ((aligned (16)));
static struct timespec total_sleep_time;
/*
* The raw monotonic time for the CLOCK_MONOTONIC_RAW posix clock.
*/
struct timespec raw_time;
/* flag for if timekeeping is suspended */
int __read_mostly timekeeping_suspended;
/* must hold xtime_lock */
void timekeeping_leap_insert(int leapsecond)
{
xtime.tv_sec += leapsecond;
wall_to_monotonic.tv_sec -= leapsecond;
update_vsyscall(&xtime, &wall_to_monotonic, timekeeper.clock,
timekeeper.mult);
}
/**
* timekeeping_forward_now - update clock to the current time
*
* Forward the current clock to update its state since the last call to
* update_wall_time(). This is useful before significant clock changes,
* as it avoids having to deal with this time offset explicitly.
*/
static void timekeeping_forward_now(void)
{
cycle_t cycle_now, cycle_delta;
struct clocksource *clock;
s64 nsec;
clock = timekeeper.clock;
cycle_now = clock->read(clock);
cycle_delta = (cycle_now - clock->cycle_last) & clock->mask;
clock->cycle_last = cycle_now;
nsec = clocksource_cyc2ns(cycle_delta, timekeeper.mult,
timekeeper.shift);
/* If arch requires, add in gettimeoffset() */
nsec += arch_gettimeoffset();
timespec_add_ns(&xtime, nsec);
nsec = clocksource_cyc2ns(cycle_delta, clock->mult, clock->shift);
timespec_add_ns(&raw_time, nsec);
}
/**
* getnstimeofday - Returns the time of day in a timespec
* @ts: pointer to the timespec to be set
*
* Returns the time of day in a timespec.
*/
void getnstimeofday(struct timespec *ts)
{
unsigned long seq;
s64 nsecs;
WARN_ON(timekeeping_suspended);
do {
seq = read_seqbegin(&xtime_lock);
*ts = xtime;
nsecs = timekeeping_get_ns();
/* If arch requires, add in gettimeoffset() */
nsecs += arch_gettimeoffset();
} while (read_seqretry(&xtime_lock, seq));
timespec_add_ns(ts, nsecs);
}
EXPORT_SYMBOL(getnstimeofday);
ktime_t ktime_get(void)
{
unsigned int seq;
s64 secs, nsecs;
WARN_ON(timekeeping_suspended);
do {
seq = read_seqbegin(&xtime_lock);
secs = xtime.tv_sec + wall_to_monotonic.tv_sec;
nsecs = xtime.tv_nsec + wall_to_monotonic.tv_nsec;
nsecs += timekeeping_get_ns();
} while (read_seqretry(&xtime_lock, seq));
/*
* Use ktime_set/ktime_add_ns to create a proper ktime on
* 32-bit architectures without CONFIG_KTIME_SCALAR.
*/
return ktime_add_ns(ktime_set(secs, 0), nsecs);
}
EXPORT_SYMBOL_GPL(ktime_get);
/**
* ktime_get_ts - get the monotonic clock in timespec format
* @ts: pointer to timespec variable
*
* The function calculates the monotonic clock from the realtime
* clock and the wall_to_monotonic offset and stores the result
* in normalized timespec format in the variable pointed to by @ts.
*/
void ktime_get_ts(struct timespec *ts)
{
struct timespec tomono;
unsigned int seq;
s64 nsecs;
WARN_ON(timekeeping_suspended);
do {
seq = read_seqbegin(&xtime_lock);
*ts = xtime;
tomono = wall_to_monotonic;
nsecs = timekeeping_get_ns();
} while (read_seqretry(&xtime_lock, seq));
set_normalized_timespec(ts, ts->tv_sec + tomono.tv_sec,
ts->tv_nsec + tomono.tv_nsec + nsecs);
}
EXPORT_SYMBOL_GPL(ktime_get_ts);
/**
* do_gettimeofday - Returns the time of day in a timeval
* @tv: pointer to the timeval to be set
*
* NOTE: Users should be converted to using getnstimeofday()
*/
void do_gettimeofday(struct timeval *tv)
{
struct timespec now;
getnstimeofday(&now);
tv->tv_sec = now.tv_sec;
tv->tv_usec = now.tv_nsec/1000;
}
EXPORT_SYMBOL(do_gettimeofday);
/**
* do_settimeofday - Sets the time of day
* @tv: pointer to the timespec variable containing the new time
*
* Sets the time of day to the new time and update NTP and notify hrtimers
*/
int do_settimeofday(struct timespec *tv)
{
struct timespec ts_delta;
unsigned long flags;
if ((unsigned long)tv->tv_nsec >= NSEC_PER_SEC)
return -EINVAL;
write_seqlock_irqsave(&xtime_lock, flags);
timekeeping_forward_now();
ts_delta.tv_sec = tv->tv_sec - xtime.tv_sec;
ts_delta.tv_nsec = tv->tv_nsec - xtime.tv_nsec;
wall_to_monotonic = timespec_sub(wall_to_monotonic, ts_delta);
xtime = *tv;
timekeeper.ntp_error = 0;
ntp_clear();
update_vsyscall(&xtime, &wall_to_monotonic, timekeeper.clock,
timekeeper.mult);
write_sequnlock_irqrestore(&xtime_lock, flags);
/* signal hrtimers about time change */
clock_was_set();
return 0;
}
EXPORT_SYMBOL(do_settimeofday);
/**
* change_clocksource - Swaps clocksources if a new one is available
*
* Accumulates current time interval and initializes new clocksource
*/
static int change_clocksource(void *data)
{
struct clocksource *new, *old;
new = (struct clocksource *) data;
timekeeping_forward_now();
if (!new->enable || new->enable(new) == 0) {
old = timekeeper.clock;
timekeeper_setup_internals(new);
if (old->disable)
old->disable(old);
}
return 0;
}
/**
* timekeeping_notify - Install a new clock source
* @clock: pointer to the clock source
*
* This function is called from clocksource.c after a new, better clock
* source has been registered. The caller holds the clocksource_mutex.
*/
void timekeeping_notify(struct clocksource *clock)
{
if (timekeeper.clock == clock)
return;
stop_machine(change_clocksource, clock, NULL);
tick_clock_notify();
}
/**
* ktime_get_real - get the real (wall-) time in ktime_t format
*
* returns the time in ktime_t format
*/
ktime_t ktime_get_real(void)
{
struct timespec now;
getnstimeofday(&now);
return timespec_to_ktime(now);
}
EXPORT_SYMBOL_GPL(ktime_get_real);
/**
* getrawmonotonic - Returns the raw monotonic time in a timespec
* @ts: pointer to the timespec to be set
*
* Returns the raw monotonic time (completely un-modified by ntp)
*/
void getrawmonotonic(struct timespec *ts)
{
unsigned long seq;
s64 nsecs;
do {
seq = read_seqbegin(&xtime_lock);
nsecs = timekeeping_get_ns_raw();
*ts = raw_time;
} while (read_seqretry(&xtime_lock, seq));
timespec_add_ns(ts, nsecs);
}
EXPORT_SYMBOL(getrawmonotonic);
/**
* timekeeping_valid_for_hres - Check if timekeeping is suitable for hres
*/
int timekeeping_valid_for_hres(void)
{
unsigned long seq;
int ret;
do {
seq = read_seqbegin(&xtime_lock);
ret = timekeeper.clock->flags & CLOCK_SOURCE_VALID_FOR_HRES;
} while (read_seqretry(&xtime_lock, seq));
return ret;
}
/**
* timekeeping_max_deferment - Returns max time the clocksource can be deferred
*
* Caller must observe xtime_lock via read_seqbegin/read_seqretry to
* ensure that the clocksource does not change!
*/
u64 timekeeping_max_deferment(void)
{
return timekeeper.clock->max_idle_ns;
}
/**
* read_persistent_clock - Return time from the persistent clock.
*
* Weak dummy function for arches that do not yet support it.
* Reads the time from the battery backed persistent clock.
* Returns a timespec with tv_sec=0 and tv_nsec=0 if unsupported.
*
* XXX - Do be sure to remove it once all arches implement it.
*/
void __attribute__((weak)) read_persistent_clock(struct timespec *ts)
{
ts->tv_sec = 0;
ts->tv_nsec = 0;
}
/**
* read_boot_clock - Return time of the system start.
*
* Weak dummy function for arches that do not yet support it.
* Function to read the exact time the system has been started.
* Returns a timespec with tv_sec=0 and tv_nsec=0 if unsupported.
*
* XXX - Do be sure to remove it once all arches implement it.
*/
void __attribute__((weak)) read_boot_clock(struct timespec *ts)
{
ts->tv_sec = 0;
ts->tv_nsec = 0;
}
/*
* timekeeping_init - Initializes the clocksource and common timekeeping values
*/
void __init timekeeping_init(void)
{
struct clocksource *clock;
unsigned long flags;
struct timespec now, boot;
read_persistent_clock(&now);
read_boot_clock(&boot);
write_seqlock_irqsave(&xtime_lock, flags);
ntp_init();
clock = clocksource_default_clock();
if (clock->enable)
clock->enable(clock);
timekeeper_setup_internals(clock);
xtime.tv_sec = now.tv_sec;
xtime.tv_nsec = now.tv_nsec;
raw_time.tv_sec = 0;
raw_time.tv_nsec = 0;
if (boot.tv_sec == 0 && boot.tv_nsec == 0) {
boot.tv_sec = xtime.tv_sec;
boot.tv_nsec = xtime.tv_nsec;
}
set_normalized_timespec(&wall_to_monotonic,
-boot.tv_sec, -boot.tv_nsec);
total_sleep_time.tv_sec = 0;
total_sleep_time.tv_nsec = 0;
write_sequnlock_irqrestore(&xtime_lock, flags);
}
/* time in seconds when suspend began */
static struct timespec timekeeping_suspend_time;
/**
* timekeeping_resume - Resumes the generic timekeeping subsystem.
* @dev: unused
*
* This is for the generic clocksource timekeeping.
* xtime/wall_to_monotonic/jiffies/etc are
* still managed by arch specific suspend/resume code.
*/
static int timekeeping_resume(struct sys_device *dev)
{
unsigned long flags;
struct timespec ts;
read_persistent_clock(&ts);
clocksource_resume();
write_seqlock_irqsave(&xtime_lock, flags);
if (timespec_compare(&ts, &timekeeping_suspend_time) > 0) {
ts = timespec_sub(ts, timekeeping_suspend_time);
xtime = timespec_add(xtime, ts);
wall_to_monotonic = timespec_sub(wall_to_monotonic, ts);
total_sleep_time = timespec_add(total_sleep_time, ts);
}
/* re-base the last cycle value */
timekeeper.clock->cycle_last = timekeeper.clock->read(timekeeper.clock);
timekeeper.ntp_error = 0;
timekeeping_suspended = 0;
write_sequnlock_irqrestore(&xtime_lock, flags);
touch_softlockup_watchdog();
clockevents_notify(CLOCK_EVT_NOTIFY_RESUME, NULL);
/* Resume hrtimers */
hres_timers_resume();
return 0;
}
static int timekeeping_suspend(struct sys_device *dev, pm_message_t state)
{
unsigned long flags;
read_persistent_clock(&timekeeping_suspend_time);
write_seqlock_irqsave(&xtime_lock, flags);
timekeeping_forward_now();
timekeeping_suspended = 1;
write_sequnlock_irqrestore(&xtime_lock, flags);
clockevents_notify(CLOCK_EVT_NOTIFY_SUSPEND, NULL);
clocksource_suspend();
return 0;
}
/* sysfs resume/suspend bits for timekeeping */
static struct sysdev_class timekeeping_sysclass = {
.name = "timekeeping",
.resume = timekeeping_resume,
.suspend = timekeeping_suspend,
};
static struct sys_device device_timer = {
.id = 0,
.cls = &timekeeping_sysclass,
};
static int __init timekeeping_init_device(void)
{
int error = sysdev_class_register(&timekeeping_sysclass);
if (!error)
error = sysdev_register(&device_timer);
return error;
}
device_initcall(timekeeping_init_device);
/*
* If the error is already larger, we look ahead even further
* to compensate for late or lost adjustments.
*/
static __always_inline int timekeeping_bigadjust(s64 error, s64 *interval,
s64 *offset)
{
s64 tick_error, i;
u32 look_ahead, adj;
s32 error2, mult;
/*
* Use the current error value to determine how much to look ahead.
* The larger the error the slower we adjust for it to avoid problems
* with losing too many ticks, otherwise we would overadjust and
* produce an even larger error. The smaller the adjustment the
* faster we try to adjust for it, as lost ticks can do less harm
* here. This is tuned so that an error of about 1 msec is adjusted
* within about 1 sec (or 2^20 nsec in 2^SHIFT_HZ ticks).
*/
error2 = timekeeper.ntp_error >> (NTP_SCALE_SHIFT + 22 - 2 * SHIFT_HZ);
error2 = abs(error2);
for (look_ahead = 0; error2 > 0; look_ahead++)
error2 >>= 2;
/*
* Now calculate the error in (1 << look_ahead) ticks, but first
* remove the single look ahead already included in the error.
*/
tick_error = tick_length >> (timekeeper.ntp_error_shift + 1);
tick_error -= timekeeper.xtime_interval >> 1;
error = ((error - tick_error) >> look_ahead) + tick_error;
/* Finally calculate the adjustment shift value. */
i = *interval;
mult = 1;
if (error < 0) {
error = -error;
*interval = -*interval;
*offset = -*offset;
mult = -1;
}
for (adj = 0; error > i; adj++)
error >>= 1;
*interval <<= adj;
*offset <<= adj;
return mult << adj;
}
/*
* Adjust the multiplier to reduce the error value,
* this is optimized for the most common adjustments of -1,0,1,
* for other values we can do a bit more work.
*/
static void timekeeping_adjust(s64 offset)
{
s64 error, interval = timekeeper.cycle_interval;
int adj;
error = timekeeper.ntp_error >> (timekeeper.ntp_error_shift - 1);
if (error > interval) {
error >>= 2;
if (likely(error <= interval))
adj = 1;
else
adj = timekeeping_bigadjust(error, &interval, &offset);
} else if (error < -interval) {
error >>= 2;
if (likely(error >= -interval)) {
adj = -1;
interval = -interval;
offset = -offset;
} else
adj = timekeeping_bigadjust(error, &interval, &offset);
} else
return;
timekeeper.mult += adj;
timekeeper.xtime_interval += interval;
timekeeper.xtime_nsec -= offset;
timekeeper.ntp_error -= (interval - offset) <<
timekeeper.ntp_error_shift;
}
Revert "time: Remove xtime_cache" This reverts commit 7bc7d637452383d56ba4368d4336b0dde1bb476d, as requested by John Stultz. Quoting John: "Petr Titěra reported an issue where he saw odd atime regressions with 2.6.33 where there were a full second worth of nanoseconds in the nanoseconds field. He also reviewed the time code and narrowed down the problem: unhandled overflow of the nanosecond field caused by rounding up the sub-nanosecond accumulated time. Details: * At the end of update_wall_time(), we currently round up the sub-nanosecond portion of accumulated time when storing it into xtime. This was added to avoid time inconsistencies caused when the sub-nanosecond portion was truncated when storing into xtime. Unfortunately we don't handle the possible second overflow caused by that rounding. * Previously the xtime_cache code hid this overflow by normalizing the xtime value when storing into the xtime_cache. * We could try to handle the second overflow after the rounding up, but since this affects the timekeeping's internal state, this would further complicate the next accumulation cycle, causing small errors in ntp steering. As much as I'd like to get rid of it, the xtime_cache code is known to work. * The correct fix is really to include the sub-nanosecond portion in the timekeeping accessor function, so we don't need to round up at during accumulation. This would greatly simplify the accumulation code. Unfortunately, we can't do this safely until the last three non-GENERIC_TIME arches (sparc32, arm, cris) are converted (those patches are in -mm) and we kill off the spots where arches set xtime directly. This is all 2.6.34 material, so I think reverting the xtime_cache change is the best approach for now. Many thanks to Petr for both reporting and finding the issue!" Reported-by: Petr Titěra <P.Titera@century.cz> Requested-by: john stultz <johnstul@us.ibm.com> Cc: Ingo Molnar <mingo@elte.hu> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-12-22 22:10:37 +00:00
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. Signed-off-by: John Stultz <johnstul@us.ibm.com> Acked-by: Thomas Gleixner <tglx@linutronix.de> Reviewed-by: John Kacur <jkacur@redhat.com> Cc: Clark Williams <williams@redhat.com> Cc: Martin Schwidefsky <schwidefsky@de.ibm.com> Cc: Andrew Morton <akpm@linux-foundation.org> LKML-Reference: <1254525473.7741.88.camel@localhost.localdomain> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2009-10-02 23:17:53 +00:00
/**
* 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;
u64 raw_nsecs;
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. Signed-off-by: John Stultz <johnstul@us.ibm.com> Acked-by: Thomas Gleixner <tglx@linutronix.de> Reviewed-by: John Kacur <jkacur@redhat.com> Cc: Clark Williams <williams@redhat.com> Cc: Martin Schwidefsky <schwidefsky@de.ibm.com> Cc: Andrew Morton <akpm@linux-foundation.org> LKML-Reference: <1254525473.7741.88.camel@localhost.localdomain> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2009-10-02 23:17:53 +00:00
/* 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 raw time */
raw_nsecs = timekeeper.raw_interval << shift;
raw_nsecs += raw_time.tv_nsec;
if (raw_nsecs >= NSEC_PER_SEC) {
u64 raw_secs = raw_nsecs;
raw_nsecs = do_div(raw_secs, NSEC_PER_SEC);
raw_time.tv_sec += raw_secs;
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. Signed-off-by: John Stultz <johnstul@us.ibm.com> Acked-by: Thomas Gleixner <tglx@linutronix.de> Reviewed-by: John Kacur <jkacur@redhat.com> Cc: Clark Williams <williams@redhat.com> Cc: Martin Schwidefsky <schwidefsky@de.ibm.com> Cc: Andrew Morton <akpm@linux-foundation.org> LKML-Reference: <1254525473.7741.88.camel@localhost.localdomain> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2009-10-02 23:17:53 +00:00
}
raw_time.tv_nsec = raw_nsecs;
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. Signed-off-by: John Stultz <johnstul@us.ibm.com> Acked-by: Thomas Gleixner <tglx@linutronix.de> Reviewed-by: John Kacur <jkacur@redhat.com> Cc: Clark Williams <williams@redhat.com> Cc: Martin Schwidefsky <schwidefsky@de.ibm.com> Cc: Andrew Morton <akpm@linux-foundation.org> LKML-Reference: <1254525473.7741.88.camel@localhost.localdomain> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2009-10-02 23:17:53 +00:00
/* Accumulate error between NTP and clock interval */
timekeeper.ntp_error += tick_length << shift;
timekeeper.ntp_error -=
(timekeeper.xtime_interval + timekeeper.xtime_remainder) <<
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. Signed-off-by: John Stultz <johnstul@us.ibm.com> Acked-by: Thomas Gleixner <tglx@linutronix.de> Reviewed-by: John Kacur <jkacur@redhat.com> Cc: Clark Williams <williams@redhat.com> Cc: Martin Schwidefsky <schwidefsky@de.ibm.com> Cc: Andrew Morton <akpm@linux-foundation.org> LKML-Reference: <1254525473.7741.88.camel@localhost.localdomain> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2009-10-02 23:17:53 +00:00
(timekeeper.ntp_error_shift + shift);
return offset;
}
Revert "time: Remove xtime_cache" This reverts commit 7bc7d637452383d56ba4368d4336b0dde1bb476d, as requested by John Stultz. Quoting John: "Petr Titěra reported an issue where he saw odd atime regressions with 2.6.33 where there were a full second worth of nanoseconds in the nanoseconds field. He also reviewed the time code and narrowed down the problem: unhandled overflow of the nanosecond field caused by rounding up the sub-nanosecond accumulated time. Details: * At the end of update_wall_time(), we currently round up the sub-nanosecond portion of accumulated time when storing it into xtime. This was added to avoid time inconsistencies caused when the sub-nanosecond portion was truncated when storing into xtime. Unfortunately we don't handle the possible second overflow caused by that rounding. * Previously the xtime_cache code hid this overflow by normalizing the xtime value when storing into the xtime_cache. * We could try to handle the second overflow after the rounding up, but since this affects the timekeeping's internal state, this would further complicate the next accumulation cycle, causing small errors in ntp steering. As much as I'd like to get rid of it, the xtime_cache code is known to work. * The correct fix is really to include the sub-nanosecond portion in the timekeeping accessor function, so we don't need to round up at during accumulation. This would greatly simplify the accumulation code. Unfortunately, we can't do this safely until the last three non-GENERIC_TIME arches (sparc32, arm, cris) are converted (those patches are in -mm) and we kill off the spots where arches set xtime directly. This is all 2.6.34 material, so I think reverting the xtime_cache change is the best approach for now. Many thanks to Petr for both reporting and finding the issue!" Reported-by: Petr Titěra <P.Titera@century.cz> Requested-by: john stultz <johnstul@us.ibm.com> Cc: Ingo Molnar <mingo@elte.hu> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-12-22 22:10:37 +00:00
/**
* update_wall_time - Uses the current clocksource to increment the wall time
*
* Called from the timer interrupt, must hold a write on xtime_lock.
*/
void update_wall_time(void)
{
struct clocksource *clock;
cycle_t offset;
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. Signed-off-by: John Stultz <johnstul@us.ibm.com> Acked-by: Thomas Gleixner <tglx@linutronix.de> Reviewed-by: John Kacur <jkacur@redhat.com> Cc: Clark Williams <williams@redhat.com> Cc: Martin Schwidefsky <schwidefsky@de.ibm.com> Cc: Andrew Morton <akpm@linux-foundation.org> LKML-Reference: <1254525473.7741.88.camel@localhost.localdomain> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2009-10-02 23:17:53 +00:00
int shift = 0, maxshift;
/* Make sure we're fully resumed: */
if (unlikely(timekeeping_suspended))
return;
clock = timekeeper.clock;
#ifdef CONFIG_ARCH_USES_GETTIMEOFFSET
offset = timekeeper.cycle_interval;
#else
offset = (clock->read(clock) - clock->cycle_last) & clock->mask;
#endif
timekeeper.xtime_nsec = (s64)xtime.tv_nsec << timekeeper.shift;
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. Signed-off-by: John Stultz <johnstul@us.ibm.com> Acked-by: Thomas Gleixner <tglx@linutronix.de> Reviewed-by: John Kacur <jkacur@redhat.com> Cc: Clark Williams <williams@redhat.com> Cc: Martin Schwidefsky <schwidefsky@de.ibm.com> Cc: Andrew Morton <akpm@linux-foundation.org> LKML-Reference: <1254525473.7741.88.camel@localhost.localdomain> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2009-10-02 23:17:53 +00:00
/*
* 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.
*/
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. Signed-off-by: John Stultz <johnstul@us.ibm.com> Acked-by: Thomas Gleixner <tglx@linutronix.de> Reviewed-by: John Kacur <jkacur@redhat.com> Cc: Clark Williams <williams@redhat.com> Cc: Martin Schwidefsky <schwidefsky@de.ibm.com> Cc: Andrew Morton <akpm@linux-foundation.org> LKML-Reference: <1254525473.7741.88.camel@localhost.localdomain> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2009-10-02 23:17:53 +00:00
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) {
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. Signed-off-by: John Stultz <johnstul@us.ibm.com> Acked-by: Thomas Gleixner <tglx@linutronix.de> Reviewed-by: John Kacur <jkacur@redhat.com> Cc: Clark Williams <williams@redhat.com> Cc: Martin Schwidefsky <schwidefsky@de.ibm.com> Cc: Andrew Morton <akpm@linux-foundation.org> LKML-Reference: <1254525473.7741.88.camel@localhost.localdomain> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2009-10-02 23:17:53 +00:00
offset = logarithmic_accumulation(offset, shift);
if(offset < timekeeper.cycle_interval<<shift)
shift--;
}
/* correct the clock when NTP error is too big */
timekeeping_adjust(offset);
time: catch xtime_nsec underflows and fix them Impact: fix time warp bug Alex Shi, along with Yanmin Zhang have been noticing occasional time inconsistencies recently. Through their great diagnosis, they found that the xtime_nsec value used in update_wall_time was occasionally going negative. After looking through the code for awhile, I realized we have the possibility for an underflow when three conditions are met in update_wall_time(): 1) We have accumulated a second's worth of nanoseconds, so we incremented xtime.tv_sec and appropriately decrement xtime_nsec. (This doesn't cause xtime_nsec to go negative, but it can cause it to be small). 2) The remaining offset value is large, but just slightly less then cycle_interval. 3) clocksource_adjust() is speeding up the clock, causing a corrective amount (compensating for the increase in the multiplier being multiplied against the unaccumulated offset value) to be subtracted from xtime_nsec. This can cause xtime_nsec to underflow. Unfortunately, since we notify the NTP subsystem via second_overflow() whenever we accumulate a full second, and this effects the error accumulation that has already occured, we cannot simply revert the accumulated second from xtime nor move the second accumulation to after the clocksource_adjust call without a change in behavior. This leaves us with (at least) two options: 1) Simply return from clocksource_adjust() without making a change if we notice the adjustment would cause xtime_nsec to go negative. This would work, but I'm concerned that if a large adjustment was needed (due to the error being large), it may be possible to get stuck with an ever increasing error that becomes too large to correct (since it may always force xtime_nsec negative). This may just be paranoia on my part. 2) Catch xtime_nsec if it is negative, then add back the amount its negative to both xtime_nsec and the error. This second method is consistent with how we've handled earlier rounding issues, and also has the benefit that the error being added is always in the oposite direction also always equal or smaller then the correction being applied. So the risk of a corner case where things get out of control is lessened. This patch fixes bug 11970, as tested by Yanmin Zhang http://bugzilla.kernel.org/show_bug.cgi?id=11970 Reported-by: alex.shi@intel.com Signed-off-by: John Stultz <johnstul@us.ibm.com> Acked-by: "Zhang, Yanmin" <yanmin_zhang@linux.intel.com> Tested-by: "Zhang, Yanmin" <yanmin_zhang@linux.intel.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-02 02:34:41 +00:00
/*
* Since in the loop above, we accumulate any amount of time
* in xtime_nsec over a second into xtime.tv_sec, its possible for
* xtime_nsec to be fairly small after the loop. Further, if we're
* slightly speeding the clocksource up in timekeeping_adjust(),
time: catch xtime_nsec underflows and fix them Impact: fix time warp bug Alex Shi, along with Yanmin Zhang have been noticing occasional time inconsistencies recently. Through their great diagnosis, they found that the xtime_nsec value used in update_wall_time was occasionally going negative. After looking through the code for awhile, I realized we have the possibility for an underflow when three conditions are met in update_wall_time(): 1) We have accumulated a second's worth of nanoseconds, so we incremented xtime.tv_sec and appropriately decrement xtime_nsec. (This doesn't cause xtime_nsec to go negative, but it can cause it to be small). 2) The remaining offset value is large, but just slightly less then cycle_interval. 3) clocksource_adjust() is speeding up the clock, causing a corrective amount (compensating for the increase in the multiplier being multiplied against the unaccumulated offset value) to be subtracted from xtime_nsec. This can cause xtime_nsec to underflow. Unfortunately, since we notify the NTP subsystem via second_overflow() whenever we accumulate a full second, and this effects the error accumulation that has already occured, we cannot simply revert the accumulated second from xtime nor move the second accumulation to after the clocksource_adjust call without a change in behavior. This leaves us with (at least) two options: 1) Simply return from clocksource_adjust() without making a change if we notice the adjustment would cause xtime_nsec to go negative. This would work, but I'm concerned that if a large adjustment was needed (due to the error being large), it may be possible to get stuck with an ever increasing error that becomes too large to correct (since it may always force xtime_nsec negative). This may just be paranoia on my part. 2) Catch xtime_nsec if it is negative, then add back the amount its negative to both xtime_nsec and the error. This second method is consistent with how we've handled earlier rounding issues, and also has the benefit that the error being added is always in the oposite direction also always equal or smaller then the correction being applied. So the risk of a corner case where things get out of control is lessened. This patch fixes bug 11970, as tested by Yanmin Zhang http://bugzilla.kernel.org/show_bug.cgi?id=11970 Reported-by: alex.shi@intel.com Signed-off-by: John Stultz <johnstul@us.ibm.com> Acked-by: "Zhang, Yanmin" <yanmin_zhang@linux.intel.com> Tested-by: "Zhang, Yanmin" <yanmin_zhang@linux.intel.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-02 02:34:41 +00:00
* its possible the required corrective factor to xtime_nsec could
* cause it to underflow.
*
* Now, we cannot simply roll the accumulated second back, since
* the NTP subsystem has been notified via second_overflow. So
* instead we push xtime_nsec forward by the amount we underflowed,
* and add that amount into the error.
*
* We'll correct this error next time through this function, when
* xtime_nsec is not as small.
*/
if (unlikely((s64)timekeeper.xtime_nsec < 0)) {
s64 neg = -(s64)timekeeper.xtime_nsec;
timekeeper.xtime_nsec = 0;
timekeeper.ntp_error += neg << timekeeper.ntp_error_shift;
time: catch xtime_nsec underflows and fix them Impact: fix time warp bug Alex Shi, along with Yanmin Zhang have been noticing occasional time inconsistencies recently. Through their great diagnosis, they found that the xtime_nsec value used in update_wall_time was occasionally going negative. After looking through the code for awhile, I realized we have the possibility for an underflow when three conditions are met in update_wall_time(): 1) We have accumulated a second's worth of nanoseconds, so we incremented xtime.tv_sec and appropriately decrement xtime_nsec. (This doesn't cause xtime_nsec to go negative, but it can cause it to be small). 2) The remaining offset value is large, but just slightly less then cycle_interval. 3) clocksource_adjust() is speeding up the clock, causing a corrective amount (compensating for the increase in the multiplier being multiplied against the unaccumulated offset value) to be subtracted from xtime_nsec. This can cause xtime_nsec to underflow. Unfortunately, since we notify the NTP subsystem via second_overflow() whenever we accumulate a full second, and this effects the error accumulation that has already occured, we cannot simply revert the accumulated second from xtime nor move the second accumulation to after the clocksource_adjust call without a change in behavior. This leaves us with (at least) two options: 1) Simply return from clocksource_adjust() without making a change if we notice the adjustment would cause xtime_nsec to go negative. This would work, but I'm concerned that if a large adjustment was needed (due to the error being large), it may be possible to get stuck with an ever increasing error that becomes too large to correct (since it may always force xtime_nsec negative). This may just be paranoia on my part. 2) Catch xtime_nsec if it is negative, then add back the amount its negative to both xtime_nsec and the error. This second method is consistent with how we've handled earlier rounding issues, and also has the benefit that the error being added is always in the oposite direction also always equal or smaller then the correction being applied. So the risk of a corner case where things get out of control is lessened. This patch fixes bug 11970, as tested by Yanmin Zhang http://bugzilla.kernel.org/show_bug.cgi?id=11970 Reported-by: alex.shi@intel.com Signed-off-by: John Stultz <johnstul@us.ibm.com> Acked-by: "Zhang, Yanmin" <yanmin_zhang@linux.intel.com> Tested-by: "Zhang, Yanmin" <yanmin_zhang@linux.intel.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-02 02:34:41 +00:00
}
/*
* Store full nanoseconds into xtime after rounding it up and
* add the remainder to the error difference.
*/
xtime.tv_nsec = ((s64) timekeeper.xtime_nsec >> timekeeper.shift) + 1;
timekeeper.xtime_nsec -= (s64) xtime.tv_nsec << timekeeper.shift;
timekeeper.ntp_error += timekeeper.xtime_nsec <<
timekeeper.ntp_error_shift;
/*
* Finally, make sure that after the rounding
* xtime.tv_nsec isn't larger then NSEC_PER_SEC
*/
if (unlikely(xtime.tv_nsec >= NSEC_PER_SEC)) {
xtime.tv_nsec -= NSEC_PER_SEC;
xtime.tv_sec++;
second_overflow();
}
Revert "time: Remove xtime_cache" This reverts commit 7bc7d637452383d56ba4368d4336b0dde1bb476d, as requested by John Stultz. Quoting John: "Petr Titěra reported an issue where he saw odd atime regressions with 2.6.33 where there were a full second worth of nanoseconds in the nanoseconds field. He also reviewed the time code and narrowed down the problem: unhandled overflow of the nanosecond field caused by rounding up the sub-nanosecond accumulated time. Details: * At the end of update_wall_time(), we currently round up the sub-nanosecond portion of accumulated time when storing it into xtime. This was added to avoid time inconsistencies caused when the sub-nanosecond portion was truncated when storing into xtime. Unfortunately we don't handle the possible second overflow caused by that rounding. * Previously the xtime_cache code hid this overflow by normalizing the xtime value when storing into the xtime_cache. * We could try to handle the second overflow after the rounding up, but since this affects the timekeeping's internal state, this would further complicate the next accumulation cycle, causing small errors in ntp steering. As much as I'd like to get rid of it, the xtime_cache code is known to work. * The correct fix is really to include the sub-nanosecond portion in the timekeeping accessor function, so we don't need to round up at during accumulation. This would greatly simplify the accumulation code. Unfortunately, we can't do this safely until the last three non-GENERIC_TIME arches (sparc32, arm, cris) are converted (those patches are in -mm) and we kill off the spots where arches set xtime directly. This is all 2.6.34 material, so I think reverting the xtime_cache change is the best approach for now. Many thanks to Petr for both reporting and finding the issue!" Reported-by: Petr Titěra <P.Titera@century.cz> Requested-by: john stultz <johnstul@us.ibm.com> Cc: Ingo Molnar <mingo@elte.hu> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-12-22 22:10:37 +00:00
/* check to see if there is a new clocksource to use */
update_vsyscall(&xtime, &wall_to_monotonic, timekeeper.clock,
timekeeper.mult);
}
/**
* getboottime - Return the real time of system boot.
* @ts: pointer to the timespec to be set
*
* Returns the time of day in a timespec.
*
* This is based on the wall_to_monotonic offset and the total suspend
* time. Calls to settimeofday will affect the value returned (which
* basically means that however wrong your real time clock is at boot time,
* you get the right time here).
*/
void getboottime(struct timespec *ts)
{
struct timespec boottime = {
.tv_sec = wall_to_monotonic.tv_sec + total_sleep_time.tv_sec,
.tv_nsec = wall_to_monotonic.tv_nsec + total_sleep_time.tv_nsec
};
set_normalized_timespec(ts, -boottime.tv_sec, -boottime.tv_nsec);
}
EXPORT_SYMBOL_GPL(getboottime);
/**
* monotonic_to_bootbased - Convert the monotonic time to boot based.
* @ts: pointer to the timespec to be converted
*/
void monotonic_to_bootbased(struct timespec *ts)
{
*ts = timespec_add(*ts, total_sleep_time);
}
EXPORT_SYMBOL_GPL(monotonic_to_bootbased);
unsigned long get_seconds(void)
{
return xtime.tv_sec;
}
EXPORT_SYMBOL(get_seconds);
struct timespec __current_kernel_time(void)
{
return xtime;
}
struct timespec __get_wall_to_monotonic(void)
{
return wall_to_monotonic;
}
struct timespec current_kernel_time(void)
{
struct timespec now;
unsigned long seq;
do {
seq = read_seqbegin(&xtime_lock);
Revert "time: Remove xtime_cache" This reverts commit 7bc7d637452383d56ba4368d4336b0dde1bb476d, as requested by John Stultz. Quoting John: "Petr Titěra reported an issue where he saw odd atime regressions with 2.6.33 where there were a full second worth of nanoseconds in the nanoseconds field. He also reviewed the time code and narrowed down the problem: unhandled overflow of the nanosecond field caused by rounding up the sub-nanosecond accumulated time. Details: * At the end of update_wall_time(), we currently round up the sub-nanosecond portion of accumulated time when storing it into xtime. This was added to avoid time inconsistencies caused when the sub-nanosecond portion was truncated when storing into xtime. Unfortunately we don't handle the possible second overflow caused by that rounding. * Previously the xtime_cache code hid this overflow by normalizing the xtime value when storing into the xtime_cache. * We could try to handle the second overflow after the rounding up, but since this affects the timekeeping's internal state, this would further complicate the next accumulation cycle, causing small errors in ntp steering. As much as I'd like to get rid of it, the xtime_cache code is known to work. * The correct fix is really to include the sub-nanosecond portion in the timekeeping accessor function, so we don't need to round up at during accumulation. This would greatly simplify the accumulation code. Unfortunately, we can't do this safely until the last three non-GENERIC_TIME arches (sparc32, arm, cris) are converted (those patches are in -mm) and we kill off the spots where arches set xtime directly. This is all 2.6.34 material, so I think reverting the xtime_cache change is the best approach for now. Many thanks to Petr for both reporting and finding the issue!" Reported-by: Petr Titěra <P.Titera@century.cz> Requested-by: john stultz <johnstul@us.ibm.com> Cc: Ingo Molnar <mingo@elte.hu> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-12-22 22:10:37 +00:00
now = xtime;
} while (read_seqretry(&xtime_lock, seq));
return now;
}
EXPORT_SYMBOL(current_kernel_time);
struct timespec get_monotonic_coarse(void)
{
struct timespec now, mono;
unsigned long seq;
do {
seq = read_seqbegin(&xtime_lock);
Revert "time: Remove xtime_cache" This reverts commit 7bc7d637452383d56ba4368d4336b0dde1bb476d, as requested by John Stultz. Quoting John: "Petr Titěra reported an issue where he saw odd atime regressions with 2.6.33 where there were a full second worth of nanoseconds in the nanoseconds field. He also reviewed the time code and narrowed down the problem: unhandled overflow of the nanosecond field caused by rounding up the sub-nanosecond accumulated time. Details: * At the end of update_wall_time(), we currently round up the sub-nanosecond portion of accumulated time when storing it into xtime. This was added to avoid time inconsistencies caused when the sub-nanosecond portion was truncated when storing into xtime. Unfortunately we don't handle the possible second overflow caused by that rounding. * Previously the xtime_cache code hid this overflow by normalizing the xtime value when storing into the xtime_cache. * We could try to handle the second overflow after the rounding up, but since this affects the timekeeping's internal state, this would further complicate the next accumulation cycle, causing small errors in ntp steering. As much as I'd like to get rid of it, the xtime_cache code is known to work. * The correct fix is really to include the sub-nanosecond portion in the timekeeping accessor function, so we don't need to round up at during accumulation. This would greatly simplify the accumulation code. Unfortunately, we can't do this safely until the last three non-GENERIC_TIME arches (sparc32, arm, cris) are converted (those patches are in -mm) and we kill off the spots where arches set xtime directly. This is all 2.6.34 material, so I think reverting the xtime_cache change is the best approach for now. Many thanks to Petr for both reporting and finding the issue!" Reported-by: Petr Titěra <P.Titera@century.cz> Requested-by: john stultz <johnstul@us.ibm.com> Cc: Ingo Molnar <mingo@elte.hu> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-12-22 22:10:37 +00:00
now = xtime;
mono = wall_to_monotonic;
} while (read_seqretry(&xtime_lock, seq));
set_normalized_timespec(&now, now.tv_sec + mono.tv_sec,
now.tv_nsec + mono.tv_nsec);
return now;
}