linux_dsm_epyc7002/include/linux/timekeeper_internal.h

156 lines
5.2 KiB
C
Raw Normal View History

/*
* You SHOULD NOT be including this unless you're vsyscall
* handling code or timekeeping internal code!
*/
#ifndef _LINUX_TIMEKEEPER_INTERNAL_H
#define _LINUX_TIMEKEEPER_INTERNAL_H
#include <linux/clocksource.h>
#include <linux/jiffies.h>
#include <linux/time.h>
/**
* struct tk_read_base - base structure for timekeeping readout
* @clock: Current clocksource used for timekeeping.
* @mask: Bitmask for two's complement subtraction of non 64bit clocks
* @cycle_last: @clock cycle value at last update
* @mult: (NTP adjusted) multiplier for scaled math conversion
* @shift: Shift value for scaled math conversion
* @xtime_nsec: Shifted (fractional) nano seconds offset for readout
* @base: ktime_t (nanoseconds) base time for readout
* @base_real: Nanoseconds base value for clock REALTIME readout
*
* This struct has size 56 byte on 64 bit. Together with a seqcount it
* occupies a single 64byte cache line.
*
* The struct is separate from struct timekeeper as it is also used
* for a fast NMI safe accessors.
*
* @base_real is for the fast NMI safe accessor to allow reading clock
* realtime from any context.
*/
struct tk_read_base {
struct clocksource *clock;
u64 mask;
u64 cycle_last;
u32 mult;
u32 shift;
u64 xtime_nsec;
ktime_t base;
u64 base_real;
};
/**
* struct timekeeper - Structure holding internal timekeeping values.
* @tkr_mono: The readout base structure for CLOCK_MONOTONIC
* @tkr_raw: The readout base structure for CLOCK_MONOTONIC_RAW
* @xtime_sec: Current CLOCK_REALTIME time in seconds
* @ktime_sec: Current CLOCK_MONOTONIC time in seconds
* @wall_to_monotonic: CLOCK_REALTIME to CLOCK_MONOTONIC offset
* @offs_real: Offset clock monotonic -> clock realtime
* @offs_boot: Offset clock monotonic -> clock boottime
* @offs_tai: Offset clock monotonic -> clock tai
* @tai_offset: The current UTC to TAI offset in seconds
* @clock_was_set_seq: The sequence number of clock was set events
time: Add history to cross timestamp interface supporting slower devices Another representative use case of time sync and the correlated clocksource (in addition to PTP noted above) is PTP synchronized audio. In a streaming application, as an example, samples will be sent and/or received by multiple devices with a presentation time that is in terms of the PTP master clock. Synchronizing the audio output on these devices requires correlating the audio clock with the PTP master clock. The more precise this correlation is, the better the audio quality (i.e. out of sync audio sounds bad). From an application standpoint, to correlate the PTP master clock with the audio device clock, the system clock is used as a intermediate timebase. The transforms such an application would perform are: System Clock <-> Audio clock System Clock <-> Network Device Clock [<-> PTP Master Clock] Modern Intel platforms can perform a more accurate cross timestamp in hardware (ART,audio device clock). The audio driver requires ART->system time transforms -- the same as required for the network driver. These platforms offload audio processing (including cross-timestamps) to a DSP which to ensure uninterrupted audio processing, communicates and response to the host only once every millsecond. As a result is takes up to a millisecond for the DSP to receive a request, the request is processed by the DSP, the audio output hardware is polled for completion, the result is copied into shared memory, and the host is notified. All of these operation occur on a millisecond cadence. This transaction requires about 2 ms, but under heavier workloads it may take up to 4 ms. Adding a history allows these slow devices the option of providing an ART value outside of the current interval. In this case, the callback provided is an accessor function for the previously obtained counter value. If get_system_device_crosststamp() receives a counter value previous to cycle_last, it consults the history provided as an argument in history_ref and interpolates the realtime and monotonic raw system time using the provided counter value. If there are any clock discontinuities, e.g. from calling settimeofday(), the monotonic raw time is interpolated in the usual way, but the realtime clock time is adjusted by scaling the monotonic raw adjustment. When an accessor function is used a history argument *must* be provided. The history is initialized using ktime_get_snapshot() and must be called before the counter values are read. Cc: Prarit Bhargava <prarit@redhat.com> Cc: Richard Cochran <richardcochran@gmail.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Ingo Molnar <mingo@kernel.org> Cc: Andy Lutomirski <luto@amacapital.net> Cc: kevin.b.stanton@intel.com Cc: kevin.j.clarke@intel.com Cc: hpa@zytor.com Cc: jeffrey.t.kirsher@intel.com Cc: netdev@vger.kernel.org Reviewed-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Christopher S. Hall <christopher.s.hall@intel.com> [jstultz: Fixed up cycles_t/cycle_t type confusion] Signed-off-by: John Stultz <john.stultz@linaro.org>
2016-02-22 18:15:23 +07:00
* @cs_was_changed_seq: The sequence number of clocksource change events
time: Prevent early expiry of hrtimers[CLOCK_REALTIME] at the leap second edge Currently, leapsecond adjustments are done at tick time. As a result, the leapsecond was applied at the first timer tick *after* the leapsecond (~1-10ms late depending on HZ), rather then exactly on the second edge. This was in part historical from back when we were always tick based, but correcting this since has been avoided since it adds extra conditional checks in the gettime fastpath, which has performance overhead. However, it was recently pointed out that ABS_TIME CLOCK_REALTIME timers set for right after the leapsecond could fire a second early, since some timers may be expired before we trigger the timekeeping timer, which then applies the leapsecond. This isn't quite as bad as it sounds, since behaviorally it is similar to what is possible w/ ntpd made leapsecond adjustments done w/o using the kernel discipline. Where due to latencies, timers may fire just prior to the settimeofday call. (Also, one should note that all applications using CLOCK_REALTIME timers should always be careful, since they are prone to quirks from settimeofday() disturbances.) However, the purpose of having the kernel do the leap adjustment is to avoid such latencies, so I think this is worth fixing. So in order to properly keep those timers from firing a second early, this patch modifies the ntp and timekeeping logic so that we keep enough state so that the update_base_offsets_now accessor, which provides the hrtimer core the current time, can check and apply the leapsecond adjustment on the second edge. This prevents the hrtimer core from expiring timers too early. This patch does not modify any other time read path, so no additional overhead is incurred. However, this also means that the leap-second continues to be applied at tick time for all other read-paths. Apologies to Richard Cochran, who pushed for similar changes years ago, which I resisted due to the concerns about the performance overhead. While I suspect this isn't extremely critical, folks who care about strict leap-second correctness will likely want to watch this. Potentially a -stable candidate eventually. Originally-suggested-by: Richard Cochran <richardcochran@gmail.com> Reported-by: Daniel Bristot de Oliveira <bristot@redhat.com> Reported-by: Prarit Bhargava <prarit@redhat.com> Signed-off-by: John Stultz <john.stultz@linaro.org> Cc: Richard Cochran <richardcochran@gmail.com> Cc: Jan Kara <jack@suse.cz> Cc: Jiri Bohac <jbohac@suse.cz> Cc: Shuah Khan <shuahkh@osg.samsung.com> Cc: Ingo Molnar <mingo@kernel.org> Link: http://lkml.kernel.org/r/1434063297-28657-4-git-send-email-john.stultz@linaro.org Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
2015-06-12 05:54:55 +07:00
* @next_leap_ktime: CLOCK_MONOTONIC time value of a pending leap-second
* @raw_sec: CLOCK_MONOTONIC_RAW time in seconds
* @cycle_interval: Number of clock cycles in one NTP interval
* @xtime_interval: Number of clock shifted nano seconds in one NTP
* interval.
* @xtime_remainder: Shifted nano seconds left over when rounding
* @cycle_interval
* @raw_interval: Shifted raw nano seconds accumulated per NTP interval.
* @ntp_error: Difference between accumulated time and NTP time in ntp
* shifted nano seconds.
* @ntp_error_shift: Shift conversion between clock shifted nano seconds and
* ntp shifted nano seconds.
* @last_warning: Warning ratelimiter (DEBUG_TIMEKEEPING)
* @underflow_seen: Underflow warning flag (DEBUG_TIMEKEEPING)
* @overflow_seen: Overflow warning flag (DEBUG_TIMEKEEPING)
*
* Note: For timespec(64) based interfaces 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.
*/
struct timekeeper {
struct tk_read_base tkr_mono;
struct tk_read_base tkr_raw;
u64 xtime_sec;
unsigned long ktime_sec;
struct timespec64 wall_to_monotonic;
ktime_t offs_real;
ktime_t offs_boot;
ktime_t offs_tai;
s32 tai_offset;
unsigned int clock_was_set_seq;
time: Add history to cross timestamp interface supporting slower devices Another representative use case of time sync and the correlated clocksource (in addition to PTP noted above) is PTP synchronized audio. In a streaming application, as an example, samples will be sent and/or received by multiple devices with a presentation time that is in terms of the PTP master clock. Synchronizing the audio output on these devices requires correlating the audio clock with the PTP master clock. The more precise this correlation is, the better the audio quality (i.e. out of sync audio sounds bad). From an application standpoint, to correlate the PTP master clock with the audio device clock, the system clock is used as a intermediate timebase. The transforms such an application would perform are: System Clock <-> Audio clock System Clock <-> Network Device Clock [<-> PTP Master Clock] Modern Intel platforms can perform a more accurate cross timestamp in hardware (ART,audio device clock). The audio driver requires ART->system time transforms -- the same as required for the network driver. These platforms offload audio processing (including cross-timestamps) to a DSP which to ensure uninterrupted audio processing, communicates and response to the host only once every millsecond. As a result is takes up to a millisecond for the DSP to receive a request, the request is processed by the DSP, the audio output hardware is polled for completion, the result is copied into shared memory, and the host is notified. All of these operation occur on a millisecond cadence. This transaction requires about 2 ms, but under heavier workloads it may take up to 4 ms. Adding a history allows these slow devices the option of providing an ART value outside of the current interval. In this case, the callback provided is an accessor function for the previously obtained counter value. If get_system_device_crosststamp() receives a counter value previous to cycle_last, it consults the history provided as an argument in history_ref and interpolates the realtime and monotonic raw system time using the provided counter value. If there are any clock discontinuities, e.g. from calling settimeofday(), the monotonic raw time is interpolated in the usual way, but the realtime clock time is adjusted by scaling the monotonic raw adjustment. When an accessor function is used a history argument *must* be provided. The history is initialized using ktime_get_snapshot() and must be called before the counter values are read. Cc: Prarit Bhargava <prarit@redhat.com> Cc: Richard Cochran <richardcochran@gmail.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Ingo Molnar <mingo@kernel.org> Cc: Andy Lutomirski <luto@amacapital.net> Cc: kevin.b.stanton@intel.com Cc: kevin.j.clarke@intel.com Cc: hpa@zytor.com Cc: jeffrey.t.kirsher@intel.com Cc: netdev@vger.kernel.org Reviewed-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Christopher S. Hall <christopher.s.hall@intel.com> [jstultz: Fixed up cycles_t/cycle_t type confusion] Signed-off-by: John Stultz <john.stultz@linaro.org>
2016-02-22 18:15:23 +07:00
u8 cs_was_changed_seq;
time: Prevent early expiry of hrtimers[CLOCK_REALTIME] at the leap second edge Currently, leapsecond adjustments are done at tick time. As a result, the leapsecond was applied at the first timer tick *after* the leapsecond (~1-10ms late depending on HZ), rather then exactly on the second edge. This was in part historical from back when we were always tick based, but correcting this since has been avoided since it adds extra conditional checks in the gettime fastpath, which has performance overhead. However, it was recently pointed out that ABS_TIME CLOCK_REALTIME timers set for right after the leapsecond could fire a second early, since some timers may be expired before we trigger the timekeeping timer, which then applies the leapsecond. This isn't quite as bad as it sounds, since behaviorally it is similar to what is possible w/ ntpd made leapsecond adjustments done w/o using the kernel discipline. Where due to latencies, timers may fire just prior to the settimeofday call. (Also, one should note that all applications using CLOCK_REALTIME timers should always be careful, since they are prone to quirks from settimeofday() disturbances.) However, the purpose of having the kernel do the leap adjustment is to avoid such latencies, so I think this is worth fixing. So in order to properly keep those timers from firing a second early, this patch modifies the ntp and timekeeping logic so that we keep enough state so that the update_base_offsets_now accessor, which provides the hrtimer core the current time, can check and apply the leapsecond adjustment on the second edge. This prevents the hrtimer core from expiring timers too early. This patch does not modify any other time read path, so no additional overhead is incurred. However, this also means that the leap-second continues to be applied at tick time for all other read-paths. Apologies to Richard Cochran, who pushed for similar changes years ago, which I resisted due to the concerns about the performance overhead. While I suspect this isn't extremely critical, folks who care about strict leap-second correctness will likely want to watch this. Potentially a -stable candidate eventually. Originally-suggested-by: Richard Cochran <richardcochran@gmail.com> Reported-by: Daniel Bristot de Oliveira <bristot@redhat.com> Reported-by: Prarit Bhargava <prarit@redhat.com> Signed-off-by: John Stultz <john.stultz@linaro.org> Cc: Richard Cochran <richardcochran@gmail.com> Cc: Jan Kara <jack@suse.cz> Cc: Jiri Bohac <jbohac@suse.cz> Cc: Shuah Khan <shuahkh@osg.samsung.com> Cc: Ingo Molnar <mingo@kernel.org> Link: http://lkml.kernel.org/r/1434063297-28657-4-git-send-email-john.stultz@linaro.org Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
2015-06-12 05:54:55 +07:00
ktime_t next_leap_ktime;
u64 raw_sec;
/* The following members are for timekeeping internal use */
u64 cycle_interval;
u64 xtime_interval;
s64 xtime_remainder;
u64 raw_interval;
/* The ntp_tick_length() value currently being used.
* This cached copy ensures we consistently apply the tick
* length for an entire tick, as ntp_tick_length may change
* mid-tick, and we don't want to apply that new value to
* the tick in progress.
*/
u64 ntp_tick;
/* Difference between accumulated time and NTP time in ntp
* shifted nano seconds. */
s64 ntp_error;
u32 ntp_error_shift;
timekeeping: Rework frequency adjustments to work better w/ nohz The existing timekeeping_adjust logic has always been complicated to understand. Further, since it was developed prior to NOHZ becoming common, its not surprising it performs poorly when NOHZ is enabled. Since Miroslav pointed out the problematic nature of the existing code in the NOHZ case, I've tried to refactor the code to perform better. The problem with the previous approach was that it tried to adjust for the total cumulative error using a scaled dampening factor. This resulted in large errors to be corrected slowly, while small errors were corrected quickly. With NOHZ the timekeeping code doesn't know how far out the next tick will be, so this results in bad over-correction to small errors, and insufficient correction to large errors. Inspired by Miroslav's patch, I've refactored the code to try to address the correction in two steps. 1) Check the future freq error for the next tick, and if the frequency error is large, try to make sure we correct it so it doesn't cause much accumulated error. 2) Then make a small single unit adjustment to correct any cumulative error that has collected over time. This method performs fairly well in the simulator Miroslav created. Major credit to Miroslav for pointing out the issue, providing the original patch to resolve this, a simulator for testing, as well as helping debug and resolve issues in my implementation so that it performed closer to his original implementation. Cc: Miroslav Lichvar <mlichvar@redhat.com> Cc: Richard Cochran <richardcochran@gmail.com> Cc: Prarit Bhargava <prarit@redhat.com> Reported-by: Miroslav Lichvar <mlichvar@redhat.com> Signed-off-by: John Stultz <john.stultz@linaro.org>
2013-12-07 08:25:21 +07:00
u32 ntp_err_mult;
#ifdef CONFIG_DEBUG_TIMEKEEPING
long last_warning;
/*
* These simple flag variables are managed
* without locks, which is racy, but they are
* ok since we don't really care about being
* super precise about how many events were
* seen, just that a problem was observed.
*/
int underflow_seen;
int overflow_seen;
#endif
};
#ifdef CONFIG_GENERIC_TIME_VSYSCALL
extern void update_vsyscall(struct timekeeper *tk);
extern void update_vsyscall_tz(void);
#elif defined(CONFIG_GENERIC_TIME_VSYSCALL_OLD)
extern void update_vsyscall_old(struct timespec *ts, struct timespec *wtm,
struct clocksource *c, u32 mult,
u64 cycle_last);
extern void update_vsyscall_tz(void);
#else
static inline void update_vsyscall(struct timekeeper *tk)
{
}
static inline void update_vsyscall_tz(void)
{
}
#endif
#endif /* _LINUX_TIMEKEEPER_INTERNAL_H */