Ingo,
Given that you are the maintainer of kernel/softlockup.c, I am seeking
advice as to how to properly fix this problem.
The short version of the problem is:
* CONFIG_HAVE_UNSTABLE_SCHED_CLOCK=y
* Attach to kgdb after boot
* Wait for 90 seconds
* Execute a continue in gdb
* You receive a warning about softlockup
The patch that follows is a lengthy analysis on the issue, but the
question here is what is the right way to fix this?
It seems that a provision is required in order to get the clock synced
up prior to touching the watch dog. It was not clear that it was a
good idea to unconditionally call the sched_clock_tick() from the
softlockup touch code.
Your input on this issue is greatly appreciated.
Thanks,
Jason.
----------
From: "Dongdong Deng" <[email protected]>
Subject: [PATCH] softlockup: add sched_clock_tick() to avoid kernel warning on kgdb resume
When CONFIG_HAVE_UNSTABLE_SCHED_CLOCK is set and the sched_clock()
was gets the time from hardware, such as from TSC, kgdb often
causes softlock warning messages on resuming or detaching from
a debug session.
Sequence of events in the problem case:
1) "cpu sched clock" and "hardware time" are at 100 seconds prior
to a call to kgdb_handle_exception()
2) Debugger waits in kgdb_handle_exception() for 80 seconds and on
exit the following is called
...
touch_softlockup_watchdog()
--> __raw_get_cpu_var(touch_timestamp) = 0;
3) "cpu sched clock" = 100s (it didn't be updated, because the
interrupt was disabled in kgdb) but the "hardware time" = 180s
4) The first timer interrupt after resuming from kgdb_handle_exception
updates the watchdog from the "cpu sched clock"
update_process_times()
{
...
run_local_timers()
--> softlockup_tick()
--> check (touch_timestamp == 0)
(it is "YES" here, we have set "touch_timestamp = 0"
at kgdb)
--> __touch_softlockup_watchdog()
***(A)--> reset "touch_timestamp" to "get_timestamp()"
(Here, the "touch_timestamp" will still be set to
100s.)
...
scheduler_tick()
***(B)--> sched_clock_tick()
(update "cpu sched clock" to "hardware time" = 180s)
...
}
5) The Second timer interrupt handler appears to have a large jump and
trips the softlockup warning.
update_process_times()
{
...
run_local_timers()
--> softlockup_tick()
--> "cpu sched clock" - "touch_timestamp" = 180s-100s > 60s
--> printk "soft lockup error messages"
...
}
note:
***(A) reset "touch_timestamp" to "get_timestamp(this_cpu)"
Why "touch_timestamp" will be 100s, not 180s ?
We enable "CONFIG_HAVE_UNSTABLE_SCHED_CLOCK",
so the call trace of get_timestamp() is:
get_timestamp(this_cpu)
-->cpu_clock(this_cpu)
-->sched_clock_cpu(this_cpu)
-->__update_sched_clock(sched_clock_data, now)
22 static u64 __update_sched_clock(struct sched_clock_data *scd, u64 =
now)
23 {
24 s64 delta = now - scd->tick_raw;
25 u64 clock, min_clock, max_clock;
26
27 WARN_ON_ONCE(!irqs_disabled());
28
29 if (unlikely(delta < 0))
30 delta = 0;
31
32 clock = scd->tick_gtod + delta;
33
34 min_clock = wrap_max(scd->tick_gtod, scd->clock);
35 max_clock = wrap_max(scd->clock, scd->tick_gtod + TICK_NSEC);
36
37 clock = wrap_max(clock, min_clock);
38 clock = wrap_min(clock, max_clock);
39
40 scd->clock = clock;
41
42 return scd->clock;
43 }
Please pay attention to __update_sched_clock() function, it uses the
GTOD tick value to create a window to filter crazy "now" values. So
if "now" values is too big for sched_clock_data, it will be ignored.
as the data of "step 3)",
"now" = sched_clock() = "hardware time" = 180s and
sched_clock_data = 100s.
180s is too big for 100s, it will be ignored.
That's why the touch_timestamp will be set to 100s, not 180s.
The fix is to simply invoke sched_clock_tick() to update "cpu sched
clock" on exit from kgdb_handle_exception.
Signed-off-by: Dongdong Deng <[email protected]>
Signed-off-by: Jason Wessel <[email protected]>
Cc: Ingo Molnar <[email protected]>
Cc: [email protected]
---
kernel/softlockup.c | 3 +++
1 file changed, 3 insertions(+)
--- a/kernel/softlockup.c
+++ b/kernel/softlockup.c
@@ -118,6 +118,9 @@ void softlockup_tick(void)
}
if (touch_timestamp == 0) {
+ /* If the time stamp was touched externally make sure the
+ * scheduler tick is up to date as well */
+ sched_clock_tick();
__touch_softlockup_watchdog();
return;
}
On Mon, 2009-07-27 at 15:03 -0500, Jason Wessel wrote:
>
> The fix is to simply invoke sched_clock_tick() to update "cpu sched
> clock" on exit from kgdb_handle_exception.
Is that a regular IRQ context, or is that NMI context?
> Signed-off-by: Dongdong Deng <[email protected]>
> Signed-off-by: Jason Wessel <[email protected]>
> Cc: Ingo Molnar <[email protected]>
> Cc: [email protected]
> ---
> kernel/softlockup.c | 3 +++
> 1 file changed, 3 insertions(+)
>
> --- a/kernel/softlockup.c
> +++ b/kernel/softlockup.c
> @@ -118,6 +118,9 @@ void softlockup_tick(void)
> }
>
> if (touch_timestamp == 0) {
> + /* If the time stamp was touched externally make sure the
> + * scheduler tick is up to date as well */
> + sched_clock_tick();
> __touch_softlockup_watchdog();
> return;
> }
>
Aside from the funny comment style (please fix) the fix does look
sensible.
Peter Zijlstra wrote:
> On Mon, 2009-07-27 at 15:03 -0500, Jason Wessel wrote:
>
>> The fix is to simply invoke sched_clock_tick() to update "cpu sched
>> clock" on exit from kgdb_handle_exception.
>>
>
> Is that a regular IRQ context, or is that NMI context?
>
>
The patch text there is incorrect (which I can fix), because there were
3 other versions of this particular patch. The version of this patch
that tried to use sched_clock_tick() from the exception context hung the
system because there was lock contention for CPUs that were not in the
running state.
In this version, the sched_clock_tick() added in the softlockup_tick(),
which did not result in a hard hang. I am pretty sure that
softlockup_tick() runs as a part of the timer interrupt processing from
run_local_timers(), so there was no further need to turn off the interrupts.
Jason.
>> Signed-off-by: Dongdong Deng <[email protected]>
>> Signed-off-by: Jason Wessel <[email protected]>
>> Cc: Ingo Molnar <[email protected]>
>> Cc: [email protected]
>> ---
>> kernel/softlockup.c | 3 +++
>> 1 file changed, 3 insertions(+)
>>
>> --- a/kernel/softlockup.c
>> +++ b/kernel/softlockup.c
>> @@ -118,6 +118,9 @@ void softlockup_tick(void)
>> }
>>
>> if (touch_timestamp == 0) {
>> + /* If the time stamp was touched externally make sure the
>> + * scheduler tick is up to date as well */
>> + sched_clock_tick();
>> __touch_softlockup_watchdog();
>> return;
>> }
>>
>>
>
> Aside from the funny comment style (please fix) the fix does look
> sensible.
>
No problem.
Thanks,
Jason.
Peter Zijlstra wrote:
> On Mon, 2009-07-27 at 15:03 -0500, Jason Wessel wrote:
>> The fix is to simply invoke sched_clock_tick() to update "cpu sched
>> clock" on exit from kgdb_handle_exception.
>
> Is that a regular IRQ context, or is that NMI context?
>
>> Signed-off-by: Dongdong Deng <[email protected]>
>> Signed-off-by: Jason Wessel <[email protected]>
>> Cc: Ingo Molnar <[email protected]>
>> Cc: [email protected]
>> ---
>> kernel/softlockup.c | 3 +++
>> 1 file changed, 3 insertions(+)
>>
>> --- a/kernel/softlockup.c
>> +++ b/kernel/softlockup.c
>> @@ -118,6 +118,9 @@ void softlockup_tick(void)
>> }
>>
>> if (touch_timestamp == 0) {
>> + /* If the time stamp was touched externally make sure the
>> + * scheduler tick is up to date as well */
>> + sched_clock_tick();
>> __touch_softlockup_watchdog();
>> return;
>> }
>>
>
> Aside from the funny comment style (please fix) the fix does look
> sensible.
It turns out that further testing of this patch shows a regression in
the ability to detect certain lockups. It is a direct result of the
way the scheduling code makes use of the touch_softlockup_watchdog().
With the above proposed patch the tick was getting updated after a
resume, but was also getting updated with the run_timers(), and if
that happened before the softlockup tick, no softlockup would get
reported (note that I was using some test code to induce softlockups).
The patch below is a bit more invasive but solves the problem by
allowing kgdb to request that the sched cpu clock is updated only when
returning from a state where we know we need to force the update.
Would this change be an acceptable solution to allow kgdb to
peacefully exist with the softlockup code?
Thanks,
Jason.
-----
From: Jason Wessel <[email protected]>
Subject: [PATCH] softlockup: add sched_clock_tick() to avoid kernel warning on kgdb resume
When CONFIG_HAVE_UNSTABLE_SCHED_CLOCK is set sched_clock() gets the
time from hardware, such as from TSC. In this configuration kgdb will
report a softlock warning messages on resuming or detaching from a
debug session.
Sequence of events in the problem case:
1) "cpu sched clock" and "hardware time" are at 100 sec prior
to a call to kgdb_handle_exception()
2) Debugger waits in kgdb_handle_exception() for 80 sec and on exit
the following is called ... touch_softlockup_watchdog() -->
__raw_get_cpu_var(touch_timestamp) = 0;
3) "cpu sched clock" = 100s (it was not updated, because the interrupt
was disabled in kgdb) but the "hardware time" = 180 sec
4) The first timer interrupt after resuming from kgdb_handle_exception
updates the watchdog from the "cpu sched clock"
update_process_times() { ... run_local_timers() --> softlockup_tick()
--> check (touch_timestamp == 0) (it is "YES" here, we have set
"touch_timestamp = 0" at kgdb) --> __touch_softlockup_watchdog()
***(A)--> reset "touch_timestamp" to "get_timestamp()" (Here, the
"touch_timestamp" will still be set to 100s.) ...
scheduler_tick() ***(B)--> sched_clock_tick() (update "cpu sched
clock" to "hardware time" = 180s) ... }
5) The Second timer interrupt handler appears to have a large jump and
trips the softlockup warning.
update_process_times() { ... run_local_timers() --> softlockup_tick()
--> "cpu sched clock" - "touch_timestamp" = 180s-100s > 60s --> printk
"soft lockup error messages" ... }
note: ***(A) reset "touch_timestamp" to "get_timestamp(this_cpu)"
Why "touch_timestamp" is 100 sec, instead of 180 sec?
With the CONFIG_HAVE_UNSTABLE_SCHED_CLOCK" set the call trace of
get_timestamp() is:
get_timestamp(this_cpu) -->cpu_clock(this_cpu)
-->sched_clock_cpu(this_cpu) -->__update_sched_clock(sched_clock_data,
now)
The __update_sched_clock() function uses the GTOD tick value to create
a window to normalize the "now" values. So if "now" values is too big
for sched_clock_data, it will be ignored.
The fix is to invoke sched_clock_tick() to update "cpu sched clock" in
order to recover from this state. This is done by introducing the
function touch_softlockup_watchdog_sync(), which allows kgdb to
request that the sched clock is updated when the watchdog thread runs
the first time after a resume from kgdb.
Signed-off-by: Jason Wessel <[email protected]>
Signed-off-by: Dongdong Deng <[email protected]>
Cc: Ingo Molnar <[email protected]>
Cc: [email protected]
---
include/linux/sched.h | 4 ++++
kernel/kgdb.c | 6 +++---
kernel/softlockup.c | 16 ++++++++++++++++
3 files changed, 23 insertions(+), 3 deletions(-)
--- a/kernel/softlockup.c
+++ b/kernel/softlockup.c
@@ -79,6 +79,14 @@ void touch_softlockup_watchdog(void)
}
EXPORT_SYMBOL(touch_softlockup_watchdog);
+static int softlock_touch_sync[NR_CPUS];
+
+void touch_softlockup_watchdog_sync(void)
+{
+ softlock_touch_sync[raw_smp_processor_id()] = 1;
+ __raw_get_cpu_var(touch_timestamp) = 0;
+}
+
void touch_all_softlockup_watchdogs(void)
{
int cpu;
@@ -118,6 +126,14 @@ void softlockup_tick(void)
}
if (touch_timestamp == 0) {
+ if (unlikely(softlock_touch_sync[this_cpu])) {
+ /*
+ * If the time stamp was touched atomically
+ * make sure the scheduler tick is up to date.
+ */
+ softlock_touch_sync[this_cpu] = 0;
+ sched_clock_tick();
+ }
__touch_softlockup_watchdog();
return;
}
--- a/include/linux/sched.h
+++ b/include/linux/sched.h
@@ -300,6 +300,7 @@ extern void sched_show_task(struct task_
#ifdef CONFIG_DETECT_SOFTLOCKUP
extern void softlockup_tick(void);
extern void touch_softlockup_watchdog(void);
+extern void touch_softlockup_watchdog_sync(void);
extern void touch_all_softlockup_watchdogs(void);
extern int proc_dosoftlockup_thresh(struct ctl_table *table, int write,
struct file *filp, void __user *buffer,
@@ -313,6 +314,9 @@ static inline void softlockup_tick(void)
static inline void touch_softlockup_watchdog(void)
{
}
+static inline void touch_softlockup_watchdog_sync(void)
+{
+}
static inline void touch_all_softlockup_watchdogs(void)
{
}
--- a/kernel/kgdb.c
+++ b/kernel/kgdb.c
@@ -590,7 +590,7 @@ static void kgdb_wait(struct pt_regs *re
/* Signal the primary CPU that we are done: */
atomic_set(&cpu_in_kgdb[cpu], 0);
- touch_softlockup_watchdog();
+ touch_softlockup_watchdog_sync();
clocksource_touch_watchdog();
local_irq_restore(flags);
}
@@ -1433,7 +1433,7 @@ acquirelock:
atomic_read(&kgdb_cpu_doing_single_step) != cpu) {
atomic_set(&kgdb_active, -1);
- touch_softlockup_watchdog();
+ touch_softlockup_watchdog_sync();
clocksource_touch_watchdog();
local_irq_restore(flags);
@@ -1526,7 +1526,7 @@ acquirelock:
kgdb_restore:
/* Free kgdb_active */
atomic_set(&kgdb_active, -1);
- touch_softlockup_watchdog();
+ touch_softlockup_watchdog_sync();
clocksource_touch_watchdog();
local_irq_restore(flags);
On Tue, Aug 4, 2009 at 10:16 PM, Ingo Molnar <[email protected]> wrote:
>
> * Jason Wessel <[email protected]> wrote:
>
>> Peter Zijlstra wrote:
>> > On Mon, 2009-07-27 at 15:03 -0500, Jason Wessel wrote:
>> >> The fix is to simply invoke sched_clock_tick() to update "cpu sched
>> >> clock" on exit from kgdb_handle_exception.
>> >
>> > Is that a regular IRQ context, or is that NMI context?
>> >
>> >> Signed-off-by: Dongdong Deng <[email protected]>
>> >> Signed-off-by: Jason Wessel <[email protected]>
>> >> Cc: Ingo Molnar <[email protected]>
>> >> Cc: [email protected]
>> >> ---
>> >> kernel/softlockup.c | 3 +++
>> >> 1 file changed, 3 insertions(+)
>> >>
>> >> --- a/kernel/softlockup.c
>> >> +++ b/kernel/softlockup.c
>> >> @@ -118,6 +118,9 @@ void softlockup_tick(void)
>> >> }
>> >>
>> >> if (touch_timestamp == 0) {
>> >> + /* If the time stamp was touched externally make sure the
>> >> + * scheduler tick is up to date as well */
>> >> + sched_clock_tick();
>> >> __touch_softlockup_watchdog();
>> >> return;
>> >> }
>> >>
>> >
>> > Aside from the funny comment style (please fix) the fix does look
>> > sensible.
>>
>> It turns out that further testing of this patch shows a regression in
>> the ability to detect certain lockups. It is a direct result of the
>> way the scheduling code makes use of the touch_softlockup_watchdog().
>> With the above proposed patch the tick was getting updated after a
>> resume, but was also getting updated with the run_timers(), and if
>> that happened before the softlockup tick, no softlockup would get
>> reported (note that I was using some test code to induce softlockups).
>>
>> The patch below is a bit more invasive but solves the problem by
>> allowing kgdb to request that the sched cpu clock is updated only when
>> returning from a state where we know we need to force the update.
>>
>> Would this change be an acceptable solution to allow kgdb to
>> peacefully exist with the softlockup code?
>>
>> Thanks,
>> Jason.
>>
>>
>> -----
>> From: Jason Wessel <[email protected]>
>> Subject: [PATCH] softlockup: add sched_clock_tick() to avoid kernel warning on kgdb resume
>>
>> When CONFIG_HAVE_UNSTABLE_SCHED_CLOCK is set sched_clock() gets the
>> time from hardware, such as from TSC. In this configuration kgdb will
>> report a softlock warning messages on resuming or detaching from a
>> debug session.
>>
>> Sequence of events in the problem case:
>>
>> 1) "cpu sched clock" and "hardware time" are at 100 sec prior
>> to a call to kgdb_handle_exception()
>>
>> 2) Debugger waits in kgdb_handle_exception() for 80 sec and on exit
>> the following is called ... touch_softlockup_watchdog() -->
>> __raw_get_cpu_var(touch_timestamp) = 0;
>>
>> 3) "cpu sched clock" = 100s (it was not updated, because the interrupt
>> was disabled in kgdb) but the "hardware time" = 180 sec
>>
>> 4) The first timer interrupt after resuming from kgdb_handle_exception
>> updates the watchdog from the "cpu sched clock"
>>
>> update_process_times() { ... run_local_timers() --> softlockup_tick()
>> --> check (touch_timestamp == 0) (it is "YES" here, we have set
>> "touch_timestamp = 0" at kgdb) --> __touch_softlockup_watchdog()
>> ***(A)--> reset "touch_timestamp" to "get_timestamp()" (Here, the
>> "touch_timestamp" will still be set to 100s.) ...
>>
>> scheduler_tick() ***(B)--> sched_clock_tick() (update "cpu sched
>> clock" to "hardware time" = 180s) ... }
>>
>> 5) The Second timer interrupt handler appears to have a large jump and
>> trips the softlockup warning.
>>
>> update_process_times() { ... run_local_timers() --> softlockup_tick()
>> --> "cpu sched clock" - "touch_timestamp" = 180s-100s > 60s --> printk
>> "soft lockup error messages" ... }
>>
>> note: ***(A) reset "touch_timestamp" to "get_timestamp(this_cpu)"
>>
>> Why "touch_timestamp" is 100 sec, instead of 180 sec?
>>
>> With the CONFIG_HAVE_UNSTABLE_SCHED_CLOCK" set the call trace of
>> get_timestamp() is:
>>
>> get_timestamp(this_cpu) -->cpu_clock(this_cpu)
>> -->sched_clock_cpu(this_cpu) -->__update_sched_clock(sched_clock_data,
>> now)
>>
>> The __update_sched_clock() function uses the GTOD tick value to create
>> a window to normalize the "now" values. So if "now" values is too big
>> for sched_clock_data, it will be ignored.
>>
>> The fix is to invoke sched_clock_tick() to update "cpu sched clock" in
>> order to recover from this state. This is done by introducing the
>> function touch_softlockup_watchdog_sync(), which allows kgdb to
>> request that the sched clock is updated when the watchdog thread runs
>> the first time after a resume from kgdb.
>>
>> Signed-off-by: Jason Wessel <[email protected]>
>> Signed-off-by: Dongdong Deng <[email protected]>
>> Cc: Ingo Molnar <[email protected]>
>> Cc: [email protected]
>>
>> ---
>> include/linux/sched.h | 4 ++++
>> kernel/kgdb.c | 6 +++---
>> kernel/softlockup.c | 16 ++++++++++++++++
>> 3 files changed, 23 insertions(+), 3 deletions(-)
>>
>> --- a/kernel/softlockup.c
>> +++ b/kernel/softlockup.c
>> @@ -79,6 +79,14 @@ void touch_softlockup_watchdog(void)
>> }
>> EXPORT_SYMBOL(touch_softlockup_watchdog);
>>
>> +static int softlock_touch_sync[NR_CPUS];
>> +
>> +void touch_softlockup_watchdog_sync(void)
>> +{
>> + softlock_touch_sync[raw_smp_processor_id()] = 1;
>> + __raw_get_cpu_var(touch_timestamp) = 0;
>> +}
>> +
>> void touch_all_softlockup_watchdogs(void)
>> {
>> int cpu;
>> @@ -118,6 +126,14 @@ void softlockup_tick(void)
>> }
>>
>> if (touch_timestamp == 0) {
>> + if (unlikely(softlock_touch_sync[this_cpu])) {
>> + /*
>> + * If the time stamp was touched atomically
>> + * make sure the scheduler tick is up to date.
>> + */
>> + softlock_touch_sync[this_cpu] = 0;
>> + sched_clock_tick();
>> + }
>
> Hm, this looks quite ugly. Peter, Thomas, can you think of a cleaner
> solution?
>
We need not to sync up sched_clock in softlockup_tick, it can just return
and sync up the timestamp in the next tick. This will not touch what should
be done by dynticks, scheduler time and so on and keep softlockup_tick
clean.
I have rewrote the patch based on above, please check the attachment.
Thanks,
Yong
> Ingo
> --
> To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
> the body of a message to [email protected]
> More majordomo info at http://vger.kernel.org/majordomo-info.html
> Please read the FAQ at http://www.tux.org/lkml/
>