We already catch most of the TSC problems by sanity checks, but there
is a subtle bug which has been in the code forever. This can cause
time jumps in the range of hours.
This was reported in:
http://lkml.org/lkml/2007/8/23/96
and
http://lkml.org/lkml/2008/3/31/23
I was able to reproduce the problem with a gettimeofday loop test on a
dual core and a quad core machine which both have sychronized
TSCs. The TSCs seems not to be perfectly in sync though, but the
kernel is not able to detect the slight delta in the sync check. Still
there exists an extremly small window where this delta can be observed
with a real big time jump. So far I was only able to reproduce this
with the vsyscall gettimeofday implementation, but in theory this
might be observable with the syscall based version as well.
CPU 0 updates the clock source variables under xtime/vyscall lock and
CPU1, where the TSC is slighty behind CPU0, is reading the time right
after the seqlock was unlocked.
The clocksource reference data was updated with the TSC from CPU0 and
the value which is read from TSC on CPU1 is less than the reference
data. This results in a huge delta value due to the unsigned
subtraction of the TSC value and the reference value. This algorithm
can not be changed due to the support of wrapping clock sources like
pm timer.
The huge delta is converted to nanoseconds and added to xtime, which
is then observable by the caller. The next gettimeofday call on CPU1
will show the correct time again as now the TSC has advanced above the
reference value.
To prevent this TSC specific wreckage we need to compare the TSC value
against the reference value and return the latter when it is larger
than the actual TSC value.
I pondered to mark the TSC unstable when the readout is smaller than
the reference value, but this would render an otherwise good and fast
clocksource unusable without a real good reason.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
The printk() can deadlock because it can wake up klogd(), and
task enqueueing will try to read the time in order to set a hrtimer.
Reported-by: Marcin Slusarz <marcin.slusarz@gmail.com>
Debugged-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
The first version of the ntp_interval/tick_length inconsistent usage patch was
recently merged as bbe4d18ac2http://git.kernel.org/gitweb.cgi?p=linux/kernel/git/torvalds/linux-2.6.git;a=commit;h=bbe4d18ac2e058c56adb0cd71f49d9ed3216a405
While the fix did greatly improve the situation, it was correctly pointed out
by Roman that it does have a small bug: If the users change clocksources after
the system has been running and NTP has made corrections, the correctoins made
against the old clocksource will be applied against the new clocksource,
causing error.
The second attempt, which corrects the issue in the NTP_INTERVAL_LENGTH
definition has also made it up-stream as commit
e13a2e61ddhttp://git.kernel.org/gitweb.cgi?p=linux/kernel/git/torvalds/linux-2.6.git;a=commit;h=e13a2e61dd5152f5499d2003470acf9c838eab84
Roman has correctly pointed out that CLOCK_TICK_ADJUST is calculated
based on the PIT's frequency, and isn't really relevant to non-PIT
driven clocksources (that is, clocksources other then jiffies and pit).
This patch reverts both of those changes, and simply removes
CLOCK_TICK_ADJUST.
This does remove the granularity error correction for users of PIT and Jiffies
clocksource users, but the granularity error but for the majority of users, it
should be within the 500ppm range NTP can accommodate for.
For systems that have granularity errors greater then 500ppm, the
"ntp_tick_adj=" boot option can be used to compensate.
[johnstul@us.ibm.com: provided changelog]
[mattilinnanvuori@yahoo.com: maek ntp_tick_adj static]
Signed-off-by: Roman Zippel <zippel@linux-m68k.org>
Acked-by: john stultz <johnstul@us.ibm.com>
Signed-off-by: Matti Linnanvuori <mattilinnanvuori@yahoo.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Cc: mingo@elte.hu
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Fix typo in comments.
BTW: I have to fix coding style in arch/ia64/kernel/time.c also, otherwise
checkpatch.pl will be complaining.
Signed-off-by: Li Zefan <lizf@cn.fujitsu.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: john stultz <johnstul@us.ibm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Function timekeeping_is_continuous() no longer checks flag
CLOCK_IS_CONTINUOUS, and it checks CLOCK_SOURCE_VALID_FOR_HRES now. So rename
the function accordingly.
Signed-off-by: Li Zefan <lizf@cn.fujitsu.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: john stultz <johnstul@us.ibm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
xtime_cache needs to be updated whenever xtime and or wall_to_monotic
are changed. Otherwise users of xtime_cache might see a stale (and in
the case of timezone changes utterly wrong) value until the next
update happens.
Fixup the obvious places, which miss this update.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Acked-by: John Stultz <johnstul@us.ibm.com>
Tested-by: Dhaval Giani <dhaval@linux.vnet.ibm.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
I recently noticed on one of my boxes that when synched with an NTP
server, the drift value reported for the system was ~283ppm. While in
some cases, clock hardware can be that bad, it struck me as unusual as
the system was using the acpi_pm clocksource, which is one of the more
trustworthy and accurate clocksources on x86 hardware.
I brought up another system and let it sync to the same NTP server, and
I noticed a similar 280some ppm drift.
In looking at the code, I found that the acpi_pm's constant frequency
was being computed correctly at boot-up, however once the system was up,
even without the ntp daemon running, the clocksource's frequency was
being modified by the clocksource_adjust() function.
Digging deeper, I realized that in the code that keeps track of how much
the clocksource is skewing from the ntp desired time, we were using
different lengths to establish how long an time interval was.
The clocksource was being setup with the following interval:
NTP_INTERVAL_LENGTH = NSEC_PER_SEC/NTP_INTERVAL_FREQ
While the ntp code was using the tick_length_base value:
tick_length_base ~= (tick_usec * NSEC_PER_USEC * USER_HZ)
/NTP_INTERVAL_FREQ
The subtle difference is:
(tick_usec * NSEC_PER_USEC * USER_HZ) != NSEC_PER_SEC
This difference in calculation was causing the clocksource correction
code to apply a correction factor to the clocksource so the two
intervals were the same, however this results in the actual frequency of
the clocksource to be made incorrect. I believe this difference would
affect all clocksources, although to differing degrees depending on the
clocksource resolution.
The issue was introduced when my HZ free ntp patch landed in 2.6.21-rc1,
so my apologies for the mistake, and for not noticing it until now.
The following patch, corrects the clocksource's initialization code so
it uses the same interval length as the code in ntp.c. After applying
this patch, the drift value for the same system went from ~283ppm to
only 2.635ppm.
I believe this patch to be good, however it does affect all arches and
I've only tested on x86, so some caution is advised. I do think it would
be a likely candidate for a stable 2.6.24.x release.
Any thoughts or feedback would be appreciated.
Signed-off-by: John Stultz <johnstul@us.ibm.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
- getnstimeofday() was just a wrapper around __get_realtime_clock_ts()
- Replace calls to __get_realtime_clock_ts() by calls to getnstimeofday()
- Fix bogus reference to get_realtime_clock_ts(), which never existed
Signed-off-by: Geert Uytterhoeven <Geert.Uytterhoeven@sonycom.com>
Cc: john stultz <johnstul@us.ibm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
All kobjects require a dynamically allocated name now. We no longer
need to keep track if the name is statically assigned, we can just
unconditionally free() all kobject names on cleanup.
Signed-off-by: Kay Sievers <kay.sievers@vrfy.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@suse.de>
- remove the no longer required __attribute__((weak)) of xtime_lock
- remove the following no longer used EXPORT_SYMBOL's:
- xtime
- xtime_lock
Signed-off-by: Adrian Bunk <bunk@kernel.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: john stultz <johnstul@us.ibm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
improve performance of sys_time(). sys_time() returns time in seconds,
but it does so by calling do_gettimeofday() and then returning the
tv_sec portion of the GTOD time. But the data structure "xtime", which
is updated by every timer/scheduler tick, already offers HZ granularity
time.
the patch improves the sysbench oltp macrobenchmark by 4-5% on an AMD
dual-core system:
v2.6.23:
#threads
1: transactions: 4073 (407.23 per sec.)
2: transactions: 8530 (852.81 per sec.)
3: transactions: 8321 (831.88 per sec.)
4: transactions: 8407 (840.58 per sec.)
5: transactions: 8070 (806.74 per sec.)
v2.6.23 + sys_time-speedup.patch:
1: transactions: 4281 (428.09 per sec.)
2: transactions: 8910 (890.85 per sec.)
3: transactions: 8659 (865.79 per sec.)
4: transactions: 8676 (867.34 per sec.)
5: transactions: 8532 (852.91 per sec.)
and by 4-5% on an Intel dual-core system too:
2.6.23:
1: transactions: 4560 (455.94 per sec.)
2: transactions: 10094 (1009.30 per sec.)
3: transactions: 9755 (975.36 per sec.)
4: transactions: 9859 (985.78 per sec.)
5: transactions: 9701 (969.72 per sec.)
2.6.23 + sys_time-speedup.patch:
1: transactions: 4779 (477.84 per sec.)
2: transactions: 10103 (1010.14 per sec.)
3: transactions: 10141 (1013.93 per sec.)
4: transactions: 10371 (1036.89 per sec.)
5: transactions: 10178 (1017.50 per sec.)
(the more CPUs the system has, the more speedup this patch gives for
this particular workload.)
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Timekeeping resume adjusts xtime by adding the slept time in seconds and
resets the reference value of the clock source (clock->cycle_last).
clock->cycle last is used to calculate the delta between the last xtime
update and the readout of the clock source in __get_nsec_offset(). xtime
plus the offset is the current time. The resume code ignores the delta
which had already elapsed between the last xtime update and the actual
time of suspend. If the suspend time is short, then we can see time
going backwards on resume.
Suspend:
offs_s = clock->read() - clock->cycle_last;
now = xtime + offs_s;
timekeeping_suspend_time = read_rtc();
Resume:
sleep_time = read_rtc() - timekeeping_suspend_time;
xtime.tv_sec += sleep_time;
clock->cycle_last = clock->read();
offs_r = clock->read() - clock->cycle_last;
now = xtime + offs_r;
if sleep_time_seconds == 0 and offs_r < offs_s, then time goes
backwards.
Fix this by storing the offset from the last xtime update and add it to
xtime during resume, when we reset clock->cycle_last:
sleep_time = read_rtc() - timekeeping_suspend_time;
xtime.tv_sec += sleep_time;
xtime += offs_s; /* Fixup xtime offset at suspend time */
clock->cycle_last = clock->read();
offs_r = clock->read() - clock->cycle_last;
now = xtime + offs_r;
Thanks to Marcelo for tracking this down on the OLPC and providing the
necessary details to analyze the root cause.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Cc: John Stultz <johnstul@us.ibm.com>
Cc: Tosatti <marcelo@kvack.org>
Lockdep complains about the access of rtc in timekeeping_suspend
inside the interrupt disabled region of the write locked xtime lock.
Move the access outside.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Cc: John Stultz <johnstul@us.ibm.com>
This avoids xtime lag seen with dynticks, because while 'xtime' itself
is still not updated often, we keep a 'xtime_cache' variable around that
contains the approximate real-time that _is_ updated each time we do a
'update_wall_time()', and is thus never off by more than one tick.
IOW, this restores the original semantics for 'xtime' users, as long as
you use the proper abstraction functions (ie 'current_kernel_time()' or
'get_seconds()' depending on whether you want a timespec or just the
seconds field).
[ Updated Patch. As penance for my sins I've also yanked another #ifdef
that was added to avoid the xtime lag w/ hrtimers. ]
Signed-off-by: John Stultz <johnstul@us.ibm.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
This avoids use of the kernel-internal "xtime" variable directly outside
of the actual time-related functions. Instead, use the helper functions
that we already have available to us.
This doesn't actually change any behaviour, but this will allow us to
fix the fact that "xtime" isn't updated very often with CONFIG_NO_HZ
(because much of the realtime information is maintained as separate
offsets to 'xtime'), which has caused interfaces that use xtime directly
to get a time that is out of sync with the real-time clock by up to a
third of a second or so.
Signed-off-by: John Stultz <johnstul@us.ibm.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Remove time_interpolator code (This is generic code, but
only user was ia64. It has been superseded by the
CONFIG_GENERIC_TIME code).
Signed-off-by: Bob Picco <bob.picco@hp.com>
Signed-off-by: John Stultz <johnstul@us.ibm.com>
Signed-off-by: Peter Keilty <peter.keilty@hp.com>
Signed-off-by: Tony Luck <tony.luck@intel.com>
clocksource_adjust() has a clock argument, which shadows the file global clock
variable. Fix this up.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Cc: john stultz <johnstul@us.ibm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
The commits
411187fb05 (GTOD: persistent clock support)
c1d370e167 (i386: use GTOD persistent clock
support)
changed the monotonic time so that it no longer jumps after resume, but it's
not possible to use it for boot time and process start time calculations then.
Also, the uptime no longer increases during suspend.
I add a variable to track the wall_to_monotonic changes, a function to get the
real boot time and a function to get the boot based time from the monotonic
one.
[akpm@linux-foundation.org: remove exports, add comment]
Signed-off-by: Tomas Janousek <tjanouse@redhat.com>
Cc: Tomas Smetana <tsmetana@redhat.com>
Cc: John Stultz <johnstul@us.ibm.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
The time keeping code move to kernel/time/timekeeping.c broke the
clocksource resume logic patch, which got applied to the old file by a
fuzzy application. Fix it up and move the clocksource_resume() call to
the appropriate place.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
[ tssk, tssk, everybody should use --fuzz=0 ]
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Move the timekeeping code out of kernel/timer.c and into
kernel/time/timekeeping.c. I made no cleanups or other changes in transit.
[akpm@linux-foundation.org: build fix]
Signed-off-by: John Stultz <johnstul@us.ibm.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
