numa_get_mems_allowed bug

numa_get_mems_allowed bug

[Michael Spiegel]

Happy Thanksgiving to all.  It appears that numa_get_mems_allowed()
does not use the MPOL_F_MEMS_ALLOWED flag.  The return value I receive
when calling this function is an empty nodemask.  I've put together a
hack that uses MPOL_F_MEMS_ALLOWED, and I get the correct nodemask.
The hack I put together does not do the proper error checking that is
performed by getpol().  I'm not sure what is the best patch to fix the
issue.  Perhaps add a third argument to getpol() that accepts an
unsigned long flag?

Cheers,
--Michael

Re: numa_get_mems_allowed bug

[Michael Spiegel]

[-- Attachment #1: Type: text/plain, Size: 866 bytes --]

To stimulate some discussion, here is the patch file for the issue.
It does the appropriate error checking.  The patch does not use
getpol(), my question was whether it should call the function or not.

--Michael

On Fri, Nov 26, 2010 at 10:24 PM, Michael Spiegel
<michael.m.spiegel@gmail.com> wrote:
> Happy Thanksgiving to all.  It appears that numa_get_mems_allowed()
> does not use the MPOL_F_MEMS_ALLOWED flag.  The return value I receive
> when calling this function is an empty nodemask.  I've put together a
> hack that uses MPOL_F_MEMS_ALLOWED, and I get the correct nodemask.
> The hack I put together does not do the proper error checking that is
> performed by getpol().  I'm not sure what is the best patch to fix the
> issue.  Perhaps add a third argument to getpol() that accepts an
> unsigned long flag?
>
> Cheers,
> --Michael
>

[-- Attachment #2: numactl-mems-allowed.patch --]
[-- Type: application/octet-stream, Size: 518 bytes --]

diff -ru numactl-2.0.6-rc3-clean/libnuma.c numactl-2.0.6-rc3/libnuma.c
--- numactl-2.0.6-rc3-clean/libnuma.c	2010-10-07 09:23:51.000000000 -0400
+++ numactl-2.0.6-rc3/libnuma.c	2010-11-29 14:13:56.823538600 -0500
@@ -1136,7 +1136,9 @@
 	 * can change, so query on each call.
 	 */
 	bmp = numa_allocate_nodemask();
-	getpol(NULL,  bmp);
+	if (get_mempolicy(NULL, bmp->maskp, bmp->size + 1, 0,
+				MPOL_F_MEMS_ALLOWED) < 0)
+		numa_error("get_mempolicy");
 	return bmp;
 }
 make_internal_alias(numa_get_mems_allowed);

Re: numa_get_mems_allowed bug

[Andi Kleen]

On Mon, Nov 29, 2010 at 02:19:29PM -0500, Michael Spiegel wrote:
> To stimulate some discussion, here is the patch file for the issue.
> It does the appropriate error checking.  The patch does not use
> getpol(), my question was whether it should call the function or not.

Patch looks good to me. getpol() is just a utility function.
-Andi

Re: numa_get_mems_allowed bug

[Cliff Wickman]

On Mon, Nov 29, 2010 at 08:35:11PM +0100, Andi Kleen wrote:
> On Mon, Nov 29, 2010 at 02:19:29PM -0500, Michael Spiegel wrote:
> > To stimulate some discussion, here is the patch file for the issue.
> > It does the appropriate error checking.  The patch does not use
> > getpol(), my question was whether it should call the function or not.
> 
> Patch looks good to me. getpol() is just a utility function.
> -Andi

Okay, thanks for the patch Michael.   And Andi's review.

It's rolled into ftp://oss.sgi.com/www/projects/libnuma/download/
    numactl-2.0.6-rc4.tar.gz

Version 2.0.6 is awaiting a new feature from Andi.  It includes Tim
Pepper's fixes.  See the CHANGES file for a list.

-Cliff
-- 
Cliff Wickman
SGI
cpw@sgi.com
(651) 683-3824