From mboxrd@z Thu Jan 1 00:00:00 1970 From: yong.zhang0@gmail.com (Yong Zhang) Date: Fri, 27 May 2011 15:01:38 +0800 Subject: [BUG] "sched: Remove rq->lock from the first half of ttwu()" locks up on ARM In-Reply-To: <1306430633.2497.91.camel@laptop> References: <1306343335.21578.65.camel@twins> <1306358128.21578.107.camel@twins> <1306405979.1200.63.camel@twins> <1306407759.27474.207.camel@e102391-lin.cambridge.arm.com> <1306409575.1200.71.camel@twins> <1306412511.1200.90.camel@twins> <20110526154508.GA13788@redhat.com> <1306425584.2497.81.camel@laptop> <1306426148.2497.83.camel@laptop> <20110526170422.GA18413@redhat.com> <1306430264.2497.88.camel@laptop> <1306430633.2497.91.camel@laptop> Message-ID: To: linux-arm-kernel@lists.infradead.org List-Id: linux-arm-kernel.lists.infradead.org On Fri, May 27, 2011 at 1:23 AM, Peter Zijlstra wrote: > > We'd end up with something like the below, which isn't too different > from what I've now got queued. > > It has the extra cpu == smp_processor_id() check, but I'm not sure this > whole case is worth the trouble. I could go stick some counters in to > verify how often all this happens I guess. > > --- > ?arch/x86/include/asm/system.h | ? ?2 ++ > ?kernel/sched.c ? ? ? ? ? ? ? ?| ? 14 +++++++++++--- > ?kernel/sched_debug.c ? ? ? ? ?| ? ?7 +++++++ > ?3 files changed, 20 insertions(+), 3 deletions(-) > > diff --git a/arch/x86/include/asm/system.h b/arch/x86/include/asm/system.h > index c2ff2a1..2c597e8 100644 > --- a/arch/x86/include/asm/system.h > +++ b/arch/x86/include/asm/system.h > @@ -10,6 +10,8 @@ > ?#include > ?#include > > +#define __ARCH_WANT_INTERRUPTS_ON_CTXSW > + > ?/* entries in ARCH_DLINFO: */ > ?#if defined(CONFIG_IA32_EMULATION) || !defined(CONFIG_X86_64) > ?# define AT_VECTOR_SIZE_ARCH 2 > diff --git a/kernel/sched.c b/kernel/sched.c > index 2d12893..e4f7a9f 100644 > --- a/kernel/sched.c > +++ b/kernel/sched.c > @@ -2636,9 +2636,17 @@ try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags) > ? ? ? ? ? ? ? ? * to spin on ->on_cpu if p is current, since that would > ? ? ? ? ? ? ? ? * deadlock. > ? ? ? ? ? ? ? ? */ > - ? ? ? ? ? ? ? if (p == current) { > - ? ? ? ? ? ? ? ? ? ? ? ttwu_queue(p, cpu); > - ? ? ? ? ? ? ? ? ? ? ? goto stat; > + ? ? ? ? ? ? ? if (cpu == smp_processor_id()) { > + ? ? ? ? ? ? ? ? ? ? ? struct rq *rq; > + > + ? ? ? ? ? ? ? ? ? ? ? rq = __task_rq_lock(p); > + ? ? ? ? ? ? ? ? ? ? ? if (p->on_cpu) { As Oleg has said, I also think we don't need this check. > + ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ttwu_activate(rq, p, ENQUEUE_WAKEUP); > + ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ttwu_do_wakeup(rq, p, wake_flags); And the difference with ttwu_queue() is ttwu_queue() calls ttwu_activate() with another flag ENQUEUE_WAKING, so if we call ->task_waking() before ttwu_queue(), I guess it will work too. But I like this version, because we call ->task_waking() and ttwu_activate() on the local cpu, that means the calculations on vruntime in that two functions are accumulated into noop. Thanks, Yong > + ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? __task_rq_unlock(rq); > + ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? goto stat; > + ? ? ? ? ? ? ? ? ? ? ? } > + ? ? ? ? ? ? ? ? ? ? ? __task_rq_unlock(rq); > ? ? ? ? ? ? ? ?} > ?#endif > ? ? ? ? ? ? ? ?cpu_relax(); > diff --git a/kernel/sched_debug.c b/kernel/sched_debug.c > index a6710a1..f0ff1de 100644 > --- a/kernel/sched_debug.c > +++ b/kernel/sched_debug.c > @@ -332,6 +332,13 @@ static int sched_debug_show(struct seq_file *m, void *v) > ? ? ? ? ? ? ? ?(int)strcspn(init_utsname()->version, " "), > ? ? ? ? ? ? ? ?init_utsname()->version); > > +#ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW > + ? ? ? SEQ_printf(m, "__ARCH_WANT_INTERRUPTS_ON_CTXSW\n"); > +#endif > +#ifdef __ARCH_WANT_UNLOCKED_CTXSW > + ? ? ? SEQ_printf(m, "__ARCH_WANT_UNLOCKED_CTXSW\n"); > +#endif > + > ?#define P(x) \ > ? ? ? ?SEQ_printf(m, "%-40s: %Ld\n", #x, (long long)(x)) > ?#define PN(x) \ > > > -- Only stand for myself