The hrtimer_interrupt hang logic adjusts min_delta_ns based on the
execution time of the hrtimer callbacks.
This is error-prone for virtual machines, where a guest vcpu can be
scheduled out during the execution of the callbacks (and the callbacks
themselves can do operations that translate to blocking operations in
the hypervisor), which in can lead to large min_delta_ns rendering the
system unusable.
Replace the current heuristics with something more reliable. Allow the
interrupt code to try 3 times to catch up with the lost time. If that
fails use the total time spent in the interrupt handler to defer the
next timer interrupt so the system can catch up with other things
which got delayed. Limit that deferment to 100ms.
The retry events and the maximum time spent in the interrupt handler
are recorded and exposed via /proc/timer_list
Inspired by a patch from Marcelo.
Reported-by: Michael Tokarev <mjt@tls.msk.ru>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Tested-by: Marcelo Tosatti <mtosatti@redhat.com>
Cc: kvm@vger.kernel.org
* 'timers-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/linux-2.6-tip:
timers, init: Limit the number of per cpu calibration bootup messages
posix-cpu-timers: optimize and document timer_create callback
clockevents: Add missing include to pacify sparse
x86: vmiclock: Fix printk format
x86: Fix printk format due to variable type change
sparc: fix printk for change of variable type
clocksource/events: Fix fallout of generic code changes
nohz: Allow 32-bit machines to sleep for more than 2.15 seconds
nohz: Track last do_timer() cpu
nohz: Prevent clocksource wrapping during idle
nohz: Type cast printk argument
mips: Use generic mult/shift factor calculation for clocks
clocksource: Provide a generic mult/shift factor calculation
clockevents: Use u32 for mult and shift factors
nohz: Introduce arch_needs_cpu
nohz: Reuse ktime in sub-functions of tick_check_idle.
time: Remove xtime_cache
time: Implement logarithmic time accumulation
Since commit 0a544198 "timekeeping: Move NTP adjusted clock multiplier
to struct timekeeper" the clock multiplier of vsyscall is updated with
the unmodified clock multiplier of the clock source and not with the
NTP adjusted multiplier of the timekeeper.
This causes user space observerable time warps:
new CLOCK-warp maximum: 120 nsecs, 00000025c337c537 -> 00000025c337c4bf
Add a new argument "mult" to update_vsyscall() and hand in the
timekeeping internal NTP adjusted multiplier.
Signed-off-by: Lin Ming <ming.m.lin@intel.com>
Cc: "Zhang Yanmin" <yanmin_zhang@linux.intel.com>
Cc: Martin Schwidefsky <schwidefsky@de.ibm.com>
Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Cc: Tony Luck <tony.luck@intel.com>
LKML-Reference: <1258436990.17765.83.camel@minggr.sh.intel.com>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
powerpc grew a new warning due to the type change of clockevent->mult.
The architectures which use parts of the generic time keeping
infrastructure tripped over my wrong assumption that
clocksource_register is only used when GENERIC_TIME=y.
I should have looked and also I should have known better. These
renitent Gaul villages are racking my nerves. Some serious deprecating
is due.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
In the dynamic tick code, "max_delta_ns" (member of the
"clock_event_device" structure) represents the maximum sleep time
that can occur between timer events in nanoseconds.
The variable, "max_delta_ns", is defined as an unsigned long
which is a 32-bit integer for 32-bit machines and a 64-bit
integer for 64-bit machines (if -m64 option is used for gcc).
The value of max_delta_ns is set by calling the function
"clockevent_delta2ns()" which returns a maximum value of LONG_MAX.
For a 32-bit machine LONG_MAX is equal to 0x7fffffff and in
nanoseconds this equates to ~2.15 seconds. Hence, the maximum
sleep time for a 32-bit machine is ~2.15 seconds, where as for
a 64-bit machine it will be many years.
This patch changes the type of max_delta_ns to be "u64" instead of
"unsigned long" so that this variable is a 64-bit type for both 32-bit
and 64-bit machines. It also changes the maximum value returned by
clockevent_delta2ns() to KTIME_MAX. Hence this allows a 32-bit
machine to sleep for longer than ~2.15 seconds. Please note that this
patch also changes "min_delta_ns" to be "u64" too and although this is
unnecessary, it makes the patch simpler as it avoids to fixup all
callers of clockevent_delta2ns().
[ tglx: changed "unsigned long long" to u64 as we use this data type
through out the time code ]
Signed-off-by: Jon Hunter <jon-hunter@ti.com>
Cc: John Stultz <johnstul@us.ibm.com>
LKML-Reference: <1250617512-23567-3-git-send-email-jon-hunter@ti.com>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
The previous patch which limits the sleep time to the maximum
deferment time of the time keeping clocksource has some limitations on
SMP machines: if all CPUs are idle then for all CPUs the maximum sleep
time is limited.
Solve this by keeping track of which cpu had the do_timer() duty
assigned last and limit the sleep time only for this cpu.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
LKML-Reference: <new-submission>
Cc: Jon Hunter <jon-hunter@ti.com>
Cc: John Stultz <johnstul@us.ibm.com>
The dynamic tick allows the kernel to sleep for periods longer than a
single tick, but it does not limit the sleep time currently. In the
worst case the kernel could sleep longer than the wrap around time of
the time keeping clock source which would result in losing track of
time.
Prevent this by limiting it to the safe maximum sleep time of the
current time keeping clock source. The value is calculated when the
clock source is registered.
[ tglx: simplified the code a bit and massaged the commit msg ]
Signed-off-by: Jon Hunter <jon-hunter@ti.com>
Cc: John Stultz <johnstul@us.ibm.com>
LKML-Reference: <1250617512-23567-2-git-send-email-jon-hunter@ti.com>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
On some archs local_softirq_pending() has a data type of unsigned long
on others its unsigned int. Type cast it to (unsigned int) in the
printk to avoid the compiler warning.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
LKML-Reference: <new-submission>
MIPS has two functions to calculcate the mult/shift factors for clock
sources and clock events at run time. ARM needs such functions as
well.
Implement a function which calculates the mult/shift factors based on
the frequencies to which and from which is converted. The function
also has a parameter to specify the minimum conversion range in
seconds. This range is guaranteed not to produce a 64bit overflow when
a value is multiplied with the calculated mult factor. The larger the
conversion range the less becomes the conversion accuracy.
Provide two inline wrappers which handle clock events and clock
sources. For clock events the "from" frequency is nano seconds per
second which corresponds to 1GHz and "to" is the device frequency. For
clock sources "from" is the device frequency and "to" is nano seconds
per second.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Tested-by: Mikael Pettersson <mikpe@it.uu.se>
Acked-by: Ralf Baechle <ralf@linux-mips.org>
Acked-by: Linus Walleij <linus.walleij@stericsson.com>
Cc: John Stultz <johnstul@us.ibm.com>
LKML-Reference: <20091111134229.766673305@linutronix.de>
The mult and shift factors of clock events differ in their data type
from those of clock sources for no reason. u32 is sufficient for
both. shift is always <= 32 and mult is limited to 2^32-1 to avoid
64bit multiplication overflows in the conversion.
Preparatory patch for a generic mult/shift factor calculation
function.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Tested-by: Mikael Pettersson <mikpe@it.uu.se>
Acked-by: Ralf Baechle <ralf@linux-mips.org>
Acked-by: Linus Walleij <linus.walleij@stericsson.com>
Cc: John Stultz <johnstul@us.ibm.com>
LKML-Reference: <20091111134229.725664788@linutronix.de>
This patch was generated by
git grep -E -i -l 's(le|el)ct' | xargs -r perl -p -i -e 's/([Ss])(le|el)ct/$1elect/
with only skipping net/netfilter/xt_SECMARK.c and
include/linux/netfilter/xt_SECMARK.h which have a struct member called
selctx.
Signed-off-by: Uwe Kleine-König <u.kleine-koenig@pengutronix.de>
Signed-off-by: Jiri Kosina <jkosina@suse.cz>
Allow the architecture to request a normal jiffy tick when the system
goes idle and tick_nohz_stop_sched_tick is called . On s390 the hook is
used to prevent the system going fully idle if there has been an
interrupt other than a clock comparator interrupt since the last wakeup.
On s390 the HiperSockets response time for 1 connection ping-pong goes
down from 42 to 34 microseconds. The CPU cost decreases by 27%.
Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com>
LKML-Reference: <20090929122533.402715150@de.ibm.com>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
On a system with NOHZ=y tick_check_idle calls tick_nohz_stop_idle and
tick_nohz_update_jiffies. Given the right conditions (ts->idle_active
and/or ts->tick_stopped) both function get a time stamp with ktime_get.
The same time stamp can be reused if both function require one.
On s390 this change has the additional benefit that gcc inlines the
tick_nohz_stop_idle function into tick_check_idle. The number of
instructions to execute tick_check_idle drops from 225 to 144
(without the ktime_get optimization it is 367 vs 215 instructions).
before:
0) | tick_check_idle() {
0) | tick_nohz_stop_idle() {
0) | ktime_get() {
0) | read_tod_clock() {
0) 0.601 us | }
0) 1.765 us | }
0) 3.047 us | }
0) | ktime_get() {
0) | read_tod_clock() {
0) 0.570 us | }
0) 1.727 us | }
0) | tick_do_update_jiffies64() {
0) 0.609 us | }
0) 8.055 us | }
after:
0) | tick_check_idle() {
0) | ktime_get() {
0) | read_tod_clock() {
0) 0.617 us | }
0) 1.773 us | }
0) | tick_do_update_jiffies64() {
0) 0.593 us | }
0) 4.477 us | }
Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com>
Cc: john stultz <johnstul@us.ibm.com>
LKML-Reference: <20090929122533.206589318@de.ibm.com>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
After m68k's task_thread_info() doesn't refer to current,
it's possible to remove sched.h from interrupt.h and not break m68k!
Many thanks to Heiko Carstens for allowing this.
Signed-off-by: Alexey Dobriyan <adobriyan@gmail.com>
Commit f2e21c9610 had unfortunate side
effects with cpufreq governors on some systems.
If the system did not switch into NOHZ mode ts->inidle is not set when
tick_nohz_stop_sched_tick() is called from the idle routine. Therefor
all subsequent calls from irq_exit() to tick_nohz_stop_sched_tick()
fail to call tick_nohz_start_idle(). This results in bogus idle
accounting information which is passed to cpufreq governors.
Set the inidle flag unconditionally of the NOHZ active state to keep
the idle time accounting correct in any case.
[ tglx: Added comment and tweaked the changelog ]
Reported-by: Steven Noonan <steven@uplinklabs.net>
Signed-off-by: Eero Nurkkala <ext-eero.nurkkala@nokia.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>
Cc: Greg KH <greg@kroah.com>
Cc: Steven Noonan <steven@uplinklabs.net>
Cc: stable@kernel.org
LKML-Reference: <1254907901.30157.93.camel@eenurkka-desktop>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
With the prior logarithmic time accumulation patch, xtime will now
always be within one "tick" of the current time, instead of
possibly half a second off.
This removes the need for the xtime_cache value, which always
stored the time at the last interrupt, so this patch cleans that up
removing the xtime_cache related code.
This is a bit simpler, but still could use some wider testing.
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: <1254525855.7741.95.camel@localhost.localdomain>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
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>
* 'timers-fixes-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/linux-2.6-tip:
clocksource: Resume clocksource without taking the clocksource mutex
git commit 75c5158f70 converted the clocksource spinlock to a
mutex. This causes the following BUG:
BUG: sleeping function called from invalid context at
kernel/mutex.c:280 in_atomic(): 0, irqs_disabled(): 1, pid: 2473,
name: pm-suspend 2 locks held by pm-suspend/2473:
#0: (&buffer->mutex){......}, at: [<ffffffff8115ab13>]
sysfs_write_file+0x3c/0x137
#1: (pm_mutex){......}, at: [<ffffffff810865b5>]
enter_state+0x39/0x130 Pid: 2473, comm: pm-suspend Not tainted 2.6.31
#1 Call Trace:
[<ffffffff810792f0>] ? __debug_show_held_locks+0x22/0x24
[<ffffffff8104a2ef>] __might_sleep+0x107/0x10b
[<ffffffff8141fca9>] mutex_lock_nested+0x25/0x43
[<ffffffff81073537>] clocksource_resume+0x1c/0x60
[<ffffffff81072902>] timekeeping_resume+0x1e/0x1c8
[<ffffffff812aee62>] __sysdev_resume+0x25/0xcf
[<ffffffff812aef79>] sysdev_resume+0x6d/0xae
[<ffffffff810864f8>] suspend_devices_and_enter+0x12b/0x1af
[<ffffffff8108665b>] enter_state+0xdf/0x130
[<ffffffff81085dc3>] state_store+0xb6/0xd3
[<ffffffff81204c73>] kobj_attr_store+0x17/0x19
[<ffffffff8115abd2>] sysfs_write_file+0xfb/0x137
[<ffffffff811057d2>] vfs_write+0xae/0x10b
[<ffffffff81208392>] ? __up_read+0x1a/0x7f
[<ffffffff811058ef>] sys_write+0x4a/0x6e
[<ffffffff81011b82>] system_call_fastpath+0x16/0x1b
clocksource_resume is called early in the resume process, there is
only one cpu, no processes are running and the interrupts are
disabled. It is therefore possible to resume the clocksources
without taking the clocksource mutex.
Reported-by: Xiaotian Feng <xtfeng@gmail.com>
Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com>
Tested-by: Michal Schmidt <mschmidt@redhat.com>
Cc: Xiaotian Feng <xtfeng@gmail.com>
Cc: John Stultz <johnstul@us.ibm.com>
LKML-Reference: <20090924172952.49697825@mschwide.boeblingen.de.ibm.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
