get_nprocs question

get_nprocs question

[Andreas Jaeger]

[-- Attachment #1: Type: text/plain, Size: 663 bytes --]


What exactly is the semantics of get_nprocs, the manual states:
The @code{get_nprocs} function returns the number of available processors.

Does available mean the number of all processors the kernel runs on?
Or does it mean the number of CPUs upon which the process and its
children will be permitted to execute?

In the latter case we can use the sched_getaffinity system call for
get_nprocs if it's available.  If this is ok, I'll prepare a patch...

Andreas
-- 
 Andreas Jaeger, aj@suse.de, http://www.suse.de/~aj
  SuSE Linux AG, Maxfeldstr. 5, 90409 Nürnberg, Germany
   GPG fingerprint = 93A3 365E CE47 B889 DF7F  FED1 389A 563C C272 A126

[-- Attachment #2: Type: application/pgp-signature, Size: 188 bytes --]

Re: get_nprocs question

[Andreas Jaeger]

[-- Attachment #1: Type: text/plain, Size: 2616 bytes --]

Andreas Jaeger <aj@suse.de> writes:

> What exactly is the semantics of get_nprocs, the manual states:
> The @code{get_nprocs} function returns the number of available processors.
>
> Does available mean the number of all processors the kernel runs on?
> Or does it mean the number of CPUs upon which the process and its
> children will be permitted to execute?
>
> In the latter case we can use the sched_getaffinity system call for
> get_nprocs if it's available.  If this is ok, I'll prepare a patch...

Here's the proposed patch.  Is this ok?

Andreas

2004-03-16  Andreas Jaeger  <aj@suse.de>
            Andi Kleen  <ak@suse.de>

        * sysdeps/unix/sysv/linux/getsysstats.c (popcnt): New.
        (__get_nprocs): Use sched_getaffinity to get number of processors.

============================================================
Index: sysdeps/unix/sysv/linux/getsysstats.c
--- sysdeps/unix/sysv/linux/getsysstats.c	4 Sep 2003 08:25:11 -0000	1.27
+++ sysdeps/unix/sysv/linux/getsysstats.c	16 Mar 2004 20:04:23 -0000
@@ -1,5 +1,5 @@
 /* Determine various system internal values, Linux version.
-   Copyright (C) 1996-2001, 2002, 2003 Free Software Foundation, Inc.
+   Copyright (C) 1996-2001, 2002, 2003, 2004 Free Software Foundation, Inc.
    This file is part of the GNU C Library.
    Contributed by Ulrich Drepper <drepper@cygnus.com>, 1996.
 
@@ -133,6 +133,26 @@ get_proc_path (char *buffer, size_t bufs
   while (0)
 #endif
 
+static unsigned char pop4 [16] =
+{
+  0, 1, 1, 2, 1, 2, 2, 3, 1, 2, 2, 3, 2, 3, 3, 4
+};
+
+/* Population count: count number of bits set.  */
+static int
+popcnt (unsigned char *buffer, int num)
+{
+  int count = 0;
+  int i;
+  for (i = 0; i < num; i++)
+    {
+      unsigned char w = buffer[i];
+      count += pop4[w >> 4] + pop4[w & 0xF];
+    }
+  return count;
+}
+
+
 int
 __get_nprocs ()
 {
@@ -141,7 +161,17 @@ __get_nprocs ()
   const char *proc_path;
   int result = 1;
 
-  /* XXX Here will come a test for the new system call.  */
+#ifdef __NR_sched_getaffinity
+  cpu_set_t cpuset;
+
+  result = INLINE_SYSCALL (sched_getaffinity, 3, __getpid(), sizeof (cpu_set_t),
+			   &cpuset);
+  if (result != -1)
+    {
+      return popcnt ((unsigned char *) &cpuset, result);
+    }
+
+#endif
 
   /* Get mount point of proc filesystem.  */
   proc_path = get_proc_path (buffer, sizeof buffer);

-- 
 Andreas Jaeger, aj@suse.de, http://www.suse.de/~aj
  SuSE Linux AG, Maxfeldstr. 5, 90409 Nürnberg, Germany
   GPG fingerprint = 93A3 365E CE47 B889 DF7F  FED1 389A 563C C272 A126

[-- Attachment #2: Type: application/pgp-signature, Size: 188 bytes --]

Re: get_nprocs question

[Ulrich Drepper]

Andreas Jaeger wrote:

> Here's the proposed patch.  Is this ok?

No.  The value returned never had a thread-specific value but with
getaffinity it would have.  The definition should stay as it it, the
affinity issue reducing the possible number of processors the
process/thread can run on is a different issue which needs a different
set of interfaces.

-- 
➧ Ulrich Drepper ➧ Red Hat, Inc. ➧ 444 Castro St ➧ Mountain View, CA ❖

Re: get_nprocs question

[Andreas Jaeger]

[-- Attachment #1: Type: text/plain, Size: 1008 bytes --]

Ulrich Drepper <drepper@redhat.com> writes:

> Andreas Jaeger wrote:
>
>> Here's the proposed patch.  Is this ok?
>
> No.  The value returned never had a thread-specific value but with
> getaffinity it would have.  The definition should stay as it it, the
> affinity issue reducing the possible number of processors the
> process/thread can run on is a different issue which needs a different
> set of interfaces.

OK, then just one more question: What is the difference between these
two:

/* Return number of configured processors.  */
extern int get_nprocs_conf (void) __THROW;

/* Return number of available processors.  */
extern int get_nprocs (void) __THROW;

How can configured be different from available?  My change above would
make it different but perhaps not as desired...

Andreas
-- 
 Andreas Jaeger, aj@suse.de, http://www.suse.de/~aj
  SuSE Linux AG, Maxfeldstr. 5, 90409 Nürnberg, Germany
   GPG fingerprint = 93A3 365E CE47 B889 DF7F  FED1 389A 563C C272 A126

[-- Attachment #2: Type: application/pgp-signature, Size: 188 bytes --]

Re: get_nprocs question

[Ulrich Drepper]

Andreas Jaeger wrote:

> OK, then just one more question: What is the difference between these
> two:
> 
> /* Return number of configured processors.  */
> extern int get_nprocs_conf (void) __THROW;
> 
> /* Return number of available processors.  */
> extern int get_nprocs (void) __THROW;

The number of processors the OS knows or knew about need not be the same
 as the number which is currently in use by any process.  This is all
from long before there was affinity.  Processors can be disabled (think
hotplug).  The SPARC kernel always had the distinction, x86 and may
others don't.

-- 
➧ Ulrich Drepper ➧ Red Hat, Inc. ➧ 444 Castro St ➧ Mountain View, CA ❖