a1cbcaa9ea
The sched_clock_remote() implementation has the following inatomicity problem on 32bit systems when accessing the remote scd->clock, which is a 64bit value. CPU0 CPU1 sched_clock_local() sched_clock_remote(CPU0) ... remote_clock = scd[CPU0]->clock read_low32bit(scd[CPU0]->clock) cmpxchg64(scd->clock,...) read_high32bit(scd[CPU0]->clock) While the update of scd->clock is using an atomic64 mechanism, the readout on the remote cpu is not, which can cause completely bogus readouts. It is a quite rare problem, because it requires the update to hit the narrow race window between the low/high readout and the update must go across the 32bit boundary. The resulting misbehaviour is, that CPU1 will see the sched_clock on CPU1 ~4 seconds ahead of it's own and update CPU1s sched_clock value to this bogus timestamp. This stays that way due to the clamping implementation for about 4 seconds until the synchronization with CLOCK_MONOTONIC undoes the problem. The issue is hard to observe, because it might only result in a less accurate SCHED_OTHER timeslicing behaviour. To create observable damage on realtime scheduling classes, it is necessary that the bogus update of CPU1 sched_clock happens in the context of an realtime thread, which then gets charged 4 seconds of RT runtime, which results in the RT throttler mechanism to trigger and prevent scheduling of RT tasks for a little less than 4 seconds. So this is quite unlikely as well. The issue was quite hard to decode as the reproduction time is between 2 days and 3 weeks and intrusive tracing makes it less likely, but the following trace recorded with trace_clock=global, which uses sched_clock_local(), gave the final hint: <idle>-0 0d..30 400269.477150: hrtimer_cancel: hrtimer=0xf7061e80 <idle>-0 0d..30 400269.477151: hrtimer_start: hrtimer=0xf7061e80 ... irq/20-S-587 1d..32 400273.772118: sched_wakeup: comm= ... target_cpu=0 <idle>-0 0dN.30 400273.772118: hrtimer_cancel: hrtimer=0xf7061e80 What happens is that CPU0 goes idle and invokes sched_clock_idle_sleep_event() which invokes sched_clock_local() and CPU1 runs a remote wakeup for CPU0 at the same time, which invokes sched_remote_clock(). The time jump gets propagated to CPU0 via sched_remote_clock() and stays stale on both cores for ~4 seconds. There are only two other possibilities, which could cause a stale sched clock: 1) ktime_get() which reads out CLOCK_MONOTONIC returns a sporadic wrong value. 2) sched_clock() which reads the TSC returns a sporadic wrong value. #1 can be excluded because sched_clock would continue to increase for one jiffy and then go stale. #2 can be excluded because it would not make the clock jump forward. It would just result in a stale sched_clock for one jiffy. After quite some brain twisting and finding the same pattern on other traces, sched_clock_remote() remained the only place which could cause such a problem and as explained above it's indeed racy on 32bit systems. So while on 64bit systems the readout is atomic, we need to verify the remote readout on 32bit machines. We need to protect the local->clock readout in sched_clock_remote() on 32bit as well because an NMI could hit between the low and the high readout, call sched_clock_local() and modify local->clock. Thanks to Siegfried Wulsch for bearing with my debug requests and going through the tedious tasks of running a bunch of reproducer systems to generate the debug information which let me decode the issue. Reported-by: Siegfried Wulsch <Siegfried.Wulsch@rovema.de> Acked-by: Peter Zijlstra <peterz@infradead.org> Cc: Steven Rostedt <rostedt@goodmis.org> Link: http://lkml.kernel.org/r/alpine.LFD.2.02.1304051544160.21884@ionos Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Cc: stable@vger.kernel.org
377 lines
8.9 KiB
C
377 lines
8.9 KiB
C
/*
|
|
* sched_clock for unstable cpu clocks
|
|
*
|
|
* Copyright (C) 2008 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
|
|
*
|
|
* Updates and enhancements:
|
|
* Copyright (C) 2008 Red Hat, Inc. Steven Rostedt <srostedt@redhat.com>
|
|
*
|
|
* Based on code by:
|
|
* Ingo Molnar <mingo@redhat.com>
|
|
* Guillaume Chazarain <guichaz@gmail.com>
|
|
*
|
|
*
|
|
* What:
|
|
*
|
|
* cpu_clock(i) provides a fast (execution time) high resolution
|
|
* clock with bounded drift between CPUs. The value of cpu_clock(i)
|
|
* is monotonic for constant i. The timestamp returned is in nanoseconds.
|
|
*
|
|
* ######################### BIG FAT WARNING ##########################
|
|
* # when comparing cpu_clock(i) to cpu_clock(j) for i != j, time can #
|
|
* # go backwards !! #
|
|
* ####################################################################
|
|
*
|
|
* There is no strict promise about the base, although it tends to start
|
|
* at 0 on boot (but people really shouldn't rely on that).
|
|
*
|
|
* cpu_clock(i) -- can be used from any context, including NMI.
|
|
* sched_clock_cpu(i) -- must be used with local IRQs disabled (implied by NMI)
|
|
* local_clock() -- is cpu_clock() on the current cpu.
|
|
*
|
|
* How:
|
|
*
|
|
* The implementation either uses sched_clock() when
|
|
* !CONFIG_HAVE_UNSTABLE_SCHED_CLOCK, which means in that case the
|
|
* sched_clock() is assumed to provide these properties (mostly it means
|
|
* the architecture provides a globally synchronized highres time source).
|
|
*
|
|
* Otherwise it tries to create a semi stable clock from a mixture of other
|
|
* clocks, including:
|
|
*
|
|
* - GTOD (clock monotomic)
|
|
* - sched_clock()
|
|
* - explicit idle events
|
|
*
|
|
* We use GTOD as base and use sched_clock() deltas to improve resolution. The
|
|
* deltas are filtered to provide monotonicity and keeping it within an
|
|
* expected window.
|
|
*
|
|
* Furthermore, explicit sleep and wakeup hooks allow us to account for time
|
|
* that is otherwise invisible (TSC gets stopped).
|
|
*
|
|
*
|
|
* Notes:
|
|
*
|
|
* The !IRQ-safetly of sched_clock() and sched_clock_cpu() comes from things
|
|
* like cpufreq interrupts that can change the base clock (TSC) multiplier
|
|
* and cause funny jumps in time -- although the filtering provided by
|
|
* sched_clock_cpu() should mitigate serious artifacts we cannot rely on it
|
|
* in general since for !CONFIG_HAVE_UNSTABLE_SCHED_CLOCK we fully rely on
|
|
* sched_clock().
|
|
*/
|
|
#include <linux/spinlock.h>
|
|
#include <linux/hardirq.h>
|
|
#include <linux/export.h>
|
|
#include <linux/percpu.h>
|
|
#include <linux/ktime.h>
|
|
#include <linux/sched.h>
|
|
|
|
/*
|
|
* Scheduler clock - returns current time in nanosec units.
|
|
* This is default implementation.
|
|
* Architectures and sub-architectures can override this.
|
|
*/
|
|
unsigned long long __attribute__((weak)) sched_clock(void)
|
|
{
|
|
return (unsigned long long)(jiffies - INITIAL_JIFFIES)
|
|
* (NSEC_PER_SEC / HZ);
|
|
}
|
|
EXPORT_SYMBOL_GPL(sched_clock);
|
|
|
|
__read_mostly int sched_clock_running;
|
|
|
|
#ifdef CONFIG_HAVE_UNSTABLE_SCHED_CLOCK
|
|
__read_mostly int sched_clock_stable;
|
|
|
|
struct sched_clock_data {
|
|
u64 tick_raw;
|
|
u64 tick_gtod;
|
|
u64 clock;
|
|
};
|
|
|
|
static DEFINE_PER_CPU_SHARED_ALIGNED(struct sched_clock_data, sched_clock_data);
|
|
|
|
static inline struct sched_clock_data *this_scd(void)
|
|
{
|
|
return &__get_cpu_var(sched_clock_data);
|
|
}
|
|
|
|
static inline struct sched_clock_data *cpu_sdc(int cpu)
|
|
{
|
|
return &per_cpu(sched_clock_data, cpu);
|
|
}
|
|
|
|
void sched_clock_init(void)
|
|
{
|
|
u64 ktime_now = ktime_to_ns(ktime_get());
|
|
int cpu;
|
|
|
|
for_each_possible_cpu(cpu) {
|
|
struct sched_clock_data *scd = cpu_sdc(cpu);
|
|
|
|
scd->tick_raw = 0;
|
|
scd->tick_gtod = ktime_now;
|
|
scd->clock = ktime_now;
|
|
}
|
|
|
|
sched_clock_running = 1;
|
|
}
|
|
|
|
/*
|
|
* min, max except they take wrapping into account
|
|
*/
|
|
|
|
static inline u64 wrap_min(u64 x, u64 y)
|
|
{
|
|
return (s64)(x - y) < 0 ? x : y;
|
|
}
|
|
|
|
static inline u64 wrap_max(u64 x, u64 y)
|
|
{
|
|
return (s64)(x - y) > 0 ? x : y;
|
|
}
|
|
|
|
/*
|
|
* update the percpu scd from the raw @now value
|
|
*
|
|
* - filter out backward motion
|
|
* - use the GTOD tick value to create a window to filter crazy TSC values
|
|
*/
|
|
static u64 sched_clock_local(struct sched_clock_data *scd)
|
|
{
|
|
u64 now, clock, old_clock, min_clock, max_clock;
|
|
s64 delta;
|
|
|
|
again:
|
|
now = sched_clock();
|
|
delta = now - scd->tick_raw;
|
|
if (unlikely(delta < 0))
|
|
delta = 0;
|
|
|
|
old_clock = scd->clock;
|
|
|
|
/*
|
|
* scd->clock = clamp(scd->tick_gtod + delta,
|
|
* max(scd->tick_gtod, scd->clock),
|
|
* scd->tick_gtod + TICK_NSEC);
|
|
*/
|
|
|
|
clock = scd->tick_gtod + delta;
|
|
min_clock = wrap_max(scd->tick_gtod, old_clock);
|
|
max_clock = wrap_max(old_clock, scd->tick_gtod + TICK_NSEC);
|
|
|
|
clock = wrap_max(clock, min_clock);
|
|
clock = wrap_min(clock, max_clock);
|
|
|
|
if (cmpxchg64(&scd->clock, old_clock, clock) != old_clock)
|
|
goto again;
|
|
|
|
return clock;
|
|
}
|
|
|
|
static u64 sched_clock_remote(struct sched_clock_data *scd)
|
|
{
|
|
struct sched_clock_data *my_scd = this_scd();
|
|
u64 this_clock, remote_clock;
|
|
u64 *ptr, old_val, val;
|
|
|
|
#if BITS_PER_LONG != 64
|
|
again:
|
|
/*
|
|
* Careful here: The local and the remote clock values need to
|
|
* be read out atomic as we need to compare the values and
|
|
* then update either the local or the remote side. So the
|
|
* cmpxchg64 below only protects one readout.
|
|
*
|
|
* We must reread via sched_clock_local() in the retry case on
|
|
* 32bit as an NMI could use sched_clock_local() via the
|
|
* tracer and hit between the readout of
|
|
* the low32bit and the high 32bit portion.
|
|
*/
|
|
this_clock = sched_clock_local(my_scd);
|
|
/*
|
|
* We must enforce atomic readout on 32bit, otherwise the
|
|
* update on the remote cpu can hit inbetween the readout of
|
|
* the low32bit and the high 32bit portion.
|
|
*/
|
|
remote_clock = cmpxchg64(&scd->clock, 0, 0);
|
|
#else
|
|
/*
|
|
* On 64bit the read of [my]scd->clock is atomic versus the
|
|
* update, so we can avoid the above 32bit dance.
|
|
*/
|
|
sched_clock_local(my_scd);
|
|
again:
|
|
this_clock = my_scd->clock;
|
|
remote_clock = scd->clock;
|
|
#endif
|
|
|
|
/*
|
|
* Use the opportunity that we have both locks
|
|
* taken to couple the two clocks: we take the
|
|
* larger time as the latest time for both
|
|
* runqueues. (this creates monotonic movement)
|
|
*/
|
|
if (likely((s64)(remote_clock - this_clock) < 0)) {
|
|
ptr = &scd->clock;
|
|
old_val = remote_clock;
|
|
val = this_clock;
|
|
} else {
|
|
/*
|
|
* Should be rare, but possible:
|
|
*/
|
|
ptr = &my_scd->clock;
|
|
old_val = this_clock;
|
|
val = remote_clock;
|
|
}
|
|
|
|
if (cmpxchg64(ptr, old_val, val) != old_val)
|
|
goto again;
|
|
|
|
return val;
|
|
}
|
|
|
|
/*
|
|
* Similar to cpu_clock(), but requires local IRQs to be disabled.
|
|
*
|
|
* See cpu_clock().
|
|
*/
|
|
u64 sched_clock_cpu(int cpu)
|
|
{
|
|
struct sched_clock_data *scd;
|
|
u64 clock;
|
|
|
|
WARN_ON_ONCE(!irqs_disabled());
|
|
|
|
if (sched_clock_stable)
|
|
return sched_clock();
|
|
|
|
if (unlikely(!sched_clock_running))
|
|
return 0ull;
|
|
|
|
scd = cpu_sdc(cpu);
|
|
|
|
if (cpu != smp_processor_id())
|
|
clock = sched_clock_remote(scd);
|
|
else
|
|
clock = sched_clock_local(scd);
|
|
|
|
return clock;
|
|
}
|
|
|
|
void sched_clock_tick(void)
|
|
{
|
|
struct sched_clock_data *scd;
|
|
u64 now, now_gtod;
|
|
|
|
if (sched_clock_stable)
|
|
return;
|
|
|
|
if (unlikely(!sched_clock_running))
|
|
return;
|
|
|
|
WARN_ON_ONCE(!irqs_disabled());
|
|
|
|
scd = this_scd();
|
|
now_gtod = ktime_to_ns(ktime_get());
|
|
now = sched_clock();
|
|
|
|
scd->tick_raw = now;
|
|
scd->tick_gtod = now_gtod;
|
|
sched_clock_local(scd);
|
|
}
|
|
|
|
/*
|
|
* We are going deep-idle (irqs are disabled):
|
|
*/
|
|
void sched_clock_idle_sleep_event(void)
|
|
{
|
|
sched_clock_cpu(smp_processor_id());
|
|
}
|
|
EXPORT_SYMBOL_GPL(sched_clock_idle_sleep_event);
|
|
|
|
/*
|
|
* We just idled delta nanoseconds (called with irqs disabled):
|
|
*/
|
|
void sched_clock_idle_wakeup_event(u64 delta_ns)
|
|
{
|
|
if (timekeeping_suspended)
|
|
return;
|
|
|
|
sched_clock_tick();
|
|
touch_softlockup_watchdog();
|
|
}
|
|
EXPORT_SYMBOL_GPL(sched_clock_idle_wakeup_event);
|
|
|
|
/*
|
|
* As outlined at the top, provides a fast, high resolution, nanosecond
|
|
* time source that is monotonic per cpu argument and has bounded drift
|
|
* between cpus.
|
|
*
|
|
* ######################### BIG FAT WARNING ##########################
|
|
* # when comparing cpu_clock(i) to cpu_clock(j) for i != j, time can #
|
|
* # go backwards !! #
|
|
* ####################################################################
|
|
*/
|
|
u64 cpu_clock(int cpu)
|
|
{
|
|
u64 clock;
|
|
unsigned long flags;
|
|
|
|
local_irq_save(flags);
|
|
clock = sched_clock_cpu(cpu);
|
|
local_irq_restore(flags);
|
|
|
|
return clock;
|
|
}
|
|
|
|
/*
|
|
* Similar to cpu_clock() for the current cpu. Time will only be observed
|
|
* to be monotonic if care is taken to only compare timestampt taken on the
|
|
* same CPU.
|
|
*
|
|
* See cpu_clock().
|
|
*/
|
|
u64 local_clock(void)
|
|
{
|
|
u64 clock;
|
|
unsigned long flags;
|
|
|
|
local_irq_save(flags);
|
|
clock = sched_clock_cpu(smp_processor_id());
|
|
local_irq_restore(flags);
|
|
|
|
return clock;
|
|
}
|
|
|
|
#else /* CONFIG_HAVE_UNSTABLE_SCHED_CLOCK */
|
|
|
|
void sched_clock_init(void)
|
|
{
|
|
sched_clock_running = 1;
|
|
}
|
|
|
|
u64 sched_clock_cpu(int cpu)
|
|
{
|
|
if (unlikely(!sched_clock_running))
|
|
return 0;
|
|
|
|
return sched_clock();
|
|
}
|
|
|
|
u64 cpu_clock(int cpu)
|
|
{
|
|
return sched_clock_cpu(cpu);
|
|
}
|
|
|
|
u64 local_clock(void)
|
|
{
|
|
return sched_clock_cpu(0);
|
|
}
|
|
|
|
#endif /* CONFIG_HAVE_UNSTABLE_SCHED_CLOCK */
|
|
|
|
EXPORT_SYMBOL_GPL(cpu_clock);
|
|
EXPORT_SYMBOL_GPL(local_clock);
|