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 current tsc_init() clears the TSC feature bit if the TSC khz
cannot be calculated, causing us to panic in
arch/x86/kernel/cpu/bugs.c check_config(). We should simply mark it
unstable.
Frankly, someone should take an axe to this code. mark_tsc_unstable()
not only marks it unstable, but sets tsc_enabled to 0, which seems
redundant but is actually important here because means it won't be
used by sched_clock() either. Perhaps a tristate enum "UNUSABLE,
UNSTABLE, OK" would be clearer, and separate mark_tsc_unstable() and
mark_tsc_broken() functions?
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
In time_cpufreq_notifier() the cpu id to act upon is held in freq->cpu. Use it
instead of smp_processor_id() in the call to set_cyc2ns_scale().
This makes the preempt_*able() unnecessary and lets set_cyc2ns_scale() update
the intended cpu's cyc2ns.
Related mail/thread: http://lkml.org/lkml/2007/12/7/130
Signed-off-by: Karsten Wiese <fzu@wemgehoertderstaat.de>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
revert:
| commit 47001d6033
| Author: Thomas Gleixner <tglx@linutronix.de>
| Date: Tue Apr 1 19:45:18 2008 +0200
|
| x86: tsc prevent time going backwards
it has been identified to cause suspend regression - and the
commit fixes a longstanding bug that existed before 2.6.25 was
opened - so it can wait some more until the effects are better
understood.
Signed-off-by: Ingo Molnar <mingo@elte.hu>
We already catch most of the TSC problems by sanity checks, but there
is a subtle bug which has been in the code for ever. 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>
notsc is ignored in 32-bit kernels if CONFIG_X86_TSC is on.. which is
bad, fix it.
Signed-off-by: Pavel Machek <pavel@suse.cz>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
After a lot of discussions with AMD it turns out that TSC
on Fam10h CPUs is synchronized when the CONSTANT_TSC cpuid bit is set.
Or rather that if there are ever systems where that is not
true it would be their BIOS' task to disable the bit.
So finally use TSC gettimeofday on Fam10h by default.
Or rather it is always used now on CPUs where the AMD
specific CONSTANT_TSC bit is set.
This gives a nice speed bost for gettimeofday() on these systems
which tends to be by far the most common v/syscall.
On a Fam10h system here TSC gtod uses about 20% of the CPU time of
acpi_pm based gtod(). This was measured on 32bit, on 64bit
it is even better because TSC gtod() can use a vsyscall
and stay in ring 3, which acpi_pm doesn't.
The Intel check simply checks for CONSTANT_TSC too without hardcoding
Intel vendor. This is equivalent on 64bit because all 64bit capable Intel
CPUs will have CONSTANT_TSC set.
On Intel there is no CPU supplied CONSTANT_TSC bit currently,
but we synthesize one based on hardcoded knowledge which steppings
have p-state invariant TSC.
So the new logic is now: On CPUs which have the AMD specific
CONSTANT_TSC bit set or on Intel CPUs which are new enough
to be known to have p-state invariant TSC always use
TSC based gettimeofday()
Cc: lenb@kernel.org
Signed-off-by: Andi Kleen <ak@suse.de>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
scale the sched_clock() cyc_2_nsec scaling factor according to
CPU frequency changes.
[ mingo@elte.hu: simplified it and fixed it for SMP. ]
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
The previous patch wasn't correctly handling the 'count' variable. If
a CPU gave bad results on the 1st or 2nd run but good results on the
3rd, it wouldn't do the correct thing. No idea if any such CPU
exists, but the patch below handles that case by discarding the bad
runs.
If a bad result (too quick, or too slow) occurs on any of the 3 runs
it will be discarded.
Also updated some comments to explain what's going on.
Signed-off-by: Dave Johnson <djohnson@sw.starentnetworks.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
I ran into this problem on a system that was unable to obtain NTP sync
because the clock was running very slow (over 10000ppm slow). ntpd had
declared all of its peers 'reject' with 'peer_dist' reason.
On investigation, the tsc_khz variable was significantly incorrect
causing xtime to run slow. After a reboot tsc_khz was correct so I
did a reboot test to see how often the problem occurred:
Test was done on a 2000 Mhz Xeon system. Of 689 reboots, 8 of them
had unacceptable tsc_khz values (>500ppm):
range of tsc_khz # of boots % of boots
---------------- ---------- ----------
< 1999750 0 0.000%
1999750 - 1999800 21 3.048%
1999800 - 1999850 166 24.128%
1999850 - 1999900 241 35.029%
1999900 - 1999950 211 30.669%
1999950 - 2000000 42 6.105%
2000000 - 2000000 0 0.000%
2000050 - 2000100 0 0.000%
[...]
2000100 - 2015000 1 0.145% << BAD
2015000 - 2030000 6 0.872% << BAD
2030000 - 2045000 1 0.145% << BAD
2045000 < 0 0.000%
The worst boot was 2032.577 Mhz, over 1.5% off!
It appears that on rare occasions, mach_countup() is taking longer to
complete than necessary.
I suspect that this is caused by the CPU taking a periodic SMI
interrupt right at the end of the 30ms calibration loop. This would
cause the loop to delay while the SMI BIOS hander runs. The resulting
TSC value is beyond what it actually should be resulting in a higher
tsc_khz.
The below patch makes native_calculate_cpu_khz() take the best
(shortest duration, lowest khz) run of it's 3 calibration loops. If a
SMI goes off causing a bad result (long duration, higher khz) it will
be discarded.
With the patch applied, 300 boots of the same system produce good
results:
range of tsc_khz # of boots % of boots
---------------- ---------- ----------
< 1999750 0 0.000%
1999750 - 1999800 30 10.000%
1999800 - 1999850 166 55.333%
1999850 - 1999900 89 29.667%
1999900 - 1999950 15 5.000%
1999950 < 0 0.000%
Problem was found and tested against 2.6.18. Patch is against 2.6.22.
Signed-off-by: Dave Johnson <djohnson@sw.starentnetworks.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
* git://git.kernel.org/pub/scm/linux/kernel/git/bunk/trivial: (74 commits)
fix do_sys_open() prototype
sysfs: trivial: fix sysfs_create_file kerneldoc spelling mistake
Documentation: Fix typo in SubmitChecklist.
Typo: depricated -> deprecated
Add missing profile=kvm option to Documentation/kernel-parameters.txt
fix typo about TBI in e1000 comment
proc.txt: Add /proc/stat field
small documentation fixes
Fix compiler warning in smount example program from sharedsubtree.txt
docs/sysfs: add missing word to sysfs attribute explanation
documentation/ext3: grammar fixes
Documentation/java.txt: typo and grammar fixes
Documentation/filesystems/vfs.txt: typo fix
include/asm-*/system.h: remove unused set_rmb(), set_wmb() macros
trivial copy_data_pages() tidy up
Fix typo in arch/x86/kernel/tsc_32.c
file link fix for Pegasus USB net driver help
remove unused return within void return function
Typo fixes retrun -> return
x86 hpet.h: remove broken links
...
cpu_data is currently an array defined using NR_CPUS. This means that
we overallocate since we will rarely really use maximum configured cpus.
When NR_CPU count is raised to 4096 the size of cpu_data becomes
3,145,728 bytes.
These changes were adopted from the sparc64 (and ia64) code. An
additional field was added to cpuinfo_x86 to be a non-ambiguous cpu
index. This corresponds to the index into a cpumask_t as well as the
per_cpu index. It's used in various places like show_cpuinfo().
cpu_data is defined to be the boot_cpu_data structure for the NON-SMP
case.
Signed-off-by: Mike Travis <travis@sgi.com>
Acked-by: Christoph Lameter <clameter@sgi.com>
Cc: Andi Kleen <ak@suse.de>
Cc: James Bottomley <James.Bottomley@steeleye.com>
Cc: Dmitry Torokhov <dtor@mail.ru>
Cc: "Antonino A. Daplas" <adaplas@pol.net>
Cc: Mark M. Hoffman <mhoffman@lightlink.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>
Since the x86 merge, lots of files that referenced their own filenames
are no longer correct. Rather than keep them up to date, just delete
them, as they add no real value.
Additionally:
- fix up comment formatting in scx200_32.c
- Remove a credit from myself in setup_64.c from a time when we had no SCM
- remove longwinded history from tsc_32.c which can be figured out from
git.
Signed-off-by: Dave Jones <davej@redhat.com>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
