2002-11-09 09:58:06

by Ravikiran G Thirumalai

Subject: [patch] get/put_cpu in up need not disable preemption

AFAICS, get_cpu, put_cpu and put_cpu_no_resched need not disable
preemption on a uniprocessor. Foll patch removes the disable/enable
premeption stuff for the UP case. Tested on a PIII 4 way for both
UP and SMP configs. Pls apply.

Thanks,
Kiran


diff -ruN -X dontdiff linux-2.5.46/include/linux/smp.h get_cpu-2.5.46/include/linux/smp.h
--- linux-2.5.46/include/linux/smp.h Tue Nov 5 04:00:31 2002
+++ get_cpu-2.5.46/include/linux/smp.h Sat Nov 9 12:27:46 2002
@@ -78,6 +78,11 @@
extern void unregister_cpu_notifier(struct notifier_block *nb);

int cpu_up(unsigned int cpu);
+
+#define get_cpu() ({ preempt_disable(); smp_processor_id(); })
+#define put_cpu() preempt_enable()
+#define put_cpu_no_resched() preempt_enable_no_resched()
+
#else /* !SMP */

/*
@@ -106,10 +111,11 @@
static inline void unregister_cpu_notifier(struct notifier_block *nb)
{
}
-#endif /* !SMP */

-#define get_cpu() ({ preempt_disable(); smp_processor_id(); })
-#define put_cpu() preempt_enable()
-#define put_cpu_no_resched() preempt_enable_no_resched()
+#define get_cpu() smp_processor_id()
+#define put_cpu() do { } while (0)
+#define put_cpu_no_resched() do { } while (0)
+
+#endif /* !SMP */

#endif /* __LINUX_SMP_H */

2002-11-09 18:47:08

by Robert Love

Subject: Re: [patch] get/put_cpu in up need not disable preemption

On Sat, 2002-11-09 at 05:06, Ravikiran G Thirumalai wrote:

> AFAICS, get_cpu, put_cpu and put_cpu_no_resched need not disable
> preemption on a uniprocessor. Foll patch removes the disable/enable
> premeption stuff for the UP case. Tested on a PIII 4 way for both
> UP and SMP configs. Pls apply.

No, it needs to.

Per-CPU data can alleviate the need for a lock. On SMP, a per-CPU
variable does not need a lock since it is impossible for another task to
enter the same critical section and access the same variable, from
another CPU. On preempt it is fully possible.

For example, this is a critical section, and you do not want two threads
on the same CPU concurrently inside:

extern struct my_struct[NR_CPUS];
int cpu = get_cpu();

do_stuff(my_struct[cpu]);
do_more_stuff(my_struct[cpu]);
foo = my_struct[cpu].bar;

put_cpu();

So get_cpu() must disable preemption even on UP. The problem you are
thinking of (being preempted and returning on a different CPU) is only a
subset of the issues. All processor accesses must be atomic.

Robert Love