Re: how do I determine the number of active CPUs?

From: David Schwartz (davids_at_webmaster.com)
Date: 08/21/03 

Date: Thu, 21 Aug 2003 01:35:58 -0700

"Kevin Easton" <kevin@-nospam-pcug.org.au> wrote in message
news:newscache$4npwjh$5ti$1@tomato.pcug.org.au...

> > Except 'sysconf' returns 1 when it can't tell how many CPUs you have.
> > Very bad if you have both SMP-safe and SMP-unsafe code and are deciding
> > which to use.

> Userland code which is SMP-unsafe is also UP-unsafe, since threads may
> be suspended and run at any point (userspace threads are preemptable). T
> his is in contrast to kernel space code, which isn't preempted.

    Threads cannot be suspended and run at any time. In particular, they
can't be suspended inside an instruction. It is trivial to write code that
is SMP-unsafe but UP-safe. Here's an example (x86 Linux):

#define LONG unsigned int

typedef struct { volatile int c; } atom;

LONG InterlockedIncrement(LONG *i)
{ // Add one to *i, return 0 if and only if *i is now 0
 unsigned char c;
 atom *a=(atom *) i;
 __asm__ __volatile__(
    "incl %1; setne %0"
    : "=qm" (c) // out
    : "m" (a->c) // in
    : "memory");
 return c;
}

    The 'incl' instruction involves a read, an increment, and then a write.
The thread cannot be pre-empted between the read and the subsequent write.
However, on an SMP machine, the cache line can ping-pong between the read
and the write. So this is user-space code that is UP-safe and SMP-unsafe.

    For a P4, the corresponding SMP-safe code takes about 50 times longer
than the code above. This is because all synchronizing instructions require
flushing and refill all piplines.

    I do write applications that decide at run-time whether they're on an
SMP or a UP system and use SMP-unsafe code when they're sure they're running
on UP.

    DS