loops_per_jiffy*Hz is not always 1 second exactly
it depends on the realization of _delay() .
delay_tsc is used as _delay() in arch/x86/lib/delay.c
It makes loop loops_per_jiffy larger than exception
and causes one thread can not obtain the spin lock for
a long time which may trigger HARD LOCKUP in this case.
So we use cpu_clock() which is more accurate.
Signed-off-by: Chuansheng Liu <[email protected]>
Signed-off-by: xiaoming wang <[email protected]>
---
kernel/locking/spinlock_debug.c | 9 ++++++---
1 files changed, 6 insertions(+), 3 deletions(-)
diff --git a/kernel/locking/spinlock_debug.c b/kernel/locking/spinlock_debug.c
index 0374a59..471d26c 100644
--- a/kernel/locking/spinlock_debug.c
+++ b/kernel/locking/spinlock_debug.c
@@ -105,10 +105,13 @@ static inline void debug_spin_unlock(raw_spinlock_t *lock)
static void __spin_lock_debug(raw_spinlock_t *lock)
{
- u64 i;
- u64 loops = loops_per_jiffy * HZ;
+ u64 t;
+ u64 one_second = 1000000000;
+ u32 this_cpu = raw_smp_processor_id();
+
+ t = cpu_clock(this_cpu);
- for (i = 0; i < loops; i++) {
+ while (cpu_clock(this_cpu) - t < one_second) {
if (arch_spin_trylock(&lock->raw_lock))
return;
__delay(1);
--
1.7.1
On Wed, Apr 30, 2014 at 06:40:38PM -0400, Wang, Xiaoming wrote:
> loops_per_jiffy*Hz is not always 1 second exactly
> it depends on the realization of _delay() .
> delay_tsc is used as _delay() in arch/x86/lib/delay.c
> It makes loop loops_per_jiffy larger than exception
> and causes one thread can not obtain the spin lock for
> a long time which may trigger HARD LOCKUP in this case.
> So we use cpu_clock() which is more accurate.
>
> Signed-off-by: Chuansheng Liu <[email protected]>
> Signed-off-by: xiaoming wang <[email protected]>
> ---
> kernel/locking/spinlock_debug.c | 9 ++++++---
> 1 files changed, 6 insertions(+), 3 deletions(-)
>
> diff --git a/kernel/locking/spinlock_debug.c b/kernel/locking/spinlock_debug.c
> index 0374a59..471d26c 100644
> --- a/kernel/locking/spinlock_debug.c
> +++ b/kernel/locking/spinlock_debug.c
> @@ -105,10 +105,13 @@ static inline void debug_spin_unlock(raw_spinlock_t *lock)
>
> static void __spin_lock_debug(raw_spinlock_t *lock)
> {
> - u64 i;
> - u64 loops = loops_per_jiffy * HZ;
> + u64 t;
> + u64 one_second = 1000000000;
> + u32 this_cpu = raw_smp_processor_id();
> +
> + t = cpu_clock(this_cpu);
>
> - for (i = 0; i < loops; i++) {
> + while (cpu_clock(this_cpu) - t < one_second) {
> if (arch_spin_trylock(&lock->raw_lock))
> return;
> __delay(1);
Yep, and now you've broken support for archs that fall back to jiffies
for cpu_clock :-), jiffies need not progress if you've got IRQs
disabled.
On Wed, Apr 30, 2014 at 06:40:38PM -0400, Wang, Xiaoming wrote:
> loops_per_jiffy*Hz is not always 1 second exactly
> it depends on the realization of _delay() .
> delay_tsc is used as _delay() in arch/x86/lib/delay.c
This just states delay() is broken. The primary response should be to
try and fix that, no?
Hello Peter,
> -----Original Message-----
> From: Peter Zijlstra [mailto:[email protected]]
> Sent: Wednesday, April 30, 2014 4:55 PM
> To: Wang, Xiaoming
> Cc: [email protected]; [email protected]; Liu, Chuansheng
> Subject: Re: [PATCH] lib/spinlock_debug: Tweak the loop time to fit different
> _delay()
>
> On Wed, Apr 30, 2014 at 06:40:38PM -0400, Wang, Xiaoming wrote:
> > loops_per_jiffy*Hz is not always 1 second exactly
> > it depends on the realization of _delay() .
> > delay_tsc is used as _delay() in arch/x86/lib/delay.c
>
> This just states delay() is broken. The primary response should be to
> try and fix that, no?
delay(1s_count) is accurate, but delay(1) is not accurate indeed, since executing
some instruction, then the 1 cycle delay maybe be used already.
On Wed, Apr 30, 2014 at 09:30:55AM +0000, Liu, Chuansheng wrote:
> > From: Peter Zijlstra [mailto:[email protected]]
> > Sent: Wednesday, April 30, 2014 4:55 PM
> > To: Wang, Xiaoming
> > Cc: [email protected]; [email protected]; Liu, Chuansheng
> > Subject: Re: [PATCH] lib/spinlock_debug: Tweak the loop time to fit different
> > _delay()
> >
> > On Wed, Apr 30, 2014 at 06:40:38PM -0400, Wang, Xiaoming wrote:
> > > loops_per_jiffy*Hz is not always 1 second exactly
> > > it depends on the realization of _delay() .
> > > delay_tsc is used as _delay() in arch/x86/lib/delay.c
> >
> > This just states delay() is broken. The primary response should be to
> > try and fix that, no?
>
>
> delay(1s_count) is accurate, but delay(1) is not accurate indeed, since executing
> some instruction, then the 1 cycle delay maybe be used already.
OK, so there's (finally) a problem statement, so is there anything sane
we can do about that?
But yes, a trylock is a cmpxchg, and a cmpxchg on a contended cacheline
can be _much_ longer than one loop.
Now the real problem is coming up with something that'll work for all
architectures.