linux-libnuma test failing in SLES 12 build root

[dgloe]

In a SLES 12 build root the linux-libnuma make check test is failing with this output:

linux-libnuma: linux-libnuma.c:70: main: Assertion `numa_bitmask_equal(bitmask, numa_all_nodes_ptr)' failed.
./wrapper.sh: line 14: 25529 Aborted (core dumped) "$@"

This is using hwloc 1.11.4.

[dgloe]

When I try again with debugging information turned on I get a different error before linux-libnuma (moved to #214).

linux-libnuma: linux-libnuma.c:64: main: Assertion `hwloc_bitmap_isequal(set, set2)' failed.
./wrapper.sh: line 14: 21061 Aborted                 (core dumped) "$@"

#4  0x0000000000401d3f in main () at linux-libnuma.c:64
        topology = 0x604960
        set = 0x607ca0
        set2 = 0x606fc0
        nocpunomemnodeset = 0x6073e0
        nocpubutmemnodeset = 0x607370
        nomembutcpunodeset = 0x606d70
        nomembutcpucpuset = 0x608040
        node = <optimized out>
        bitmask = <optimized out>
        bitmask2 = <optimized out>
        mask = 1
        maxnode = 4204333
        i = <optimized out>
        __PRETTY_FUNCTION__ = "main"

[bgoglin]

get_last_cpu_location is a completely different function, I am ignoring it here.
Please add this patch:

diff --git a/tests/linux-libnuma.c b/tests/linux-libnuma.c
index 9750c6e..7a49bfb 100644
--- a/tests/linux-libnuma.c
+++ b/tests/linux-libnuma.c
@@ -67,6 +67,7 @@ int main(void)
   bitmask = hwloc_cpuset_to_linux_libnuma_bitmask(topology, set);
   /* numa_all_nodes_ptr contains NODES with no CPU but with memory */
   hwloc_bitmap_foreach_begin(i, nocpubutmemnodeset) { numa_bitmask_setbit(bitmask, i); } hwloc_bitmap_foreach_end();
+printf("%ld != %ld\n", *bitmask->maskp, *numa_all_nodes_ptr->maskp);
   assert(numa_bitmask_equal(bitmask, numa_all_nodes_ptr));
   numa_bitmask_free(bitmask);

then rebuild the test with make -C tests linux-libnuma, and run it manually with tests/linux-libnuma, and report the output.

Also please send the file foo.xml generated by lstopo foo.xml so that we know what your machine looks like.

[dgloe]

When I run it again it fails earlier, at hwloc_bitmap_isequal(set, set2). So the additional printf doesn't help.

foo.xml.txt

[bgoglin]

Do you have NUMA enabled in your kernel? (grep CONFIG_NUMA /boot/config-*)
If not, we may want to just disable this test in that case.

Otherwise:

diff --git a/tests/linux-libnuma.c b/tests/linux-libnuma.c
index 9750c6e..8992e47 100644
--- a/tests/linux-libnuma.c
+++ b/tests/linux-libnuma.c
@@ -61,6 +61,14 @@ int main(void)
   hwloc_cpuset_from_linux_libnuma_bitmask(topology, set2, numa_all_nodes_ptr);
   /* numa_all_nodes_ptr doesn't contain NODES with CPU but no memory */
   hwloc_bitmap_or(set2, set2, nomembutcpucpuset);
+{
+  char *a, *b, *c;
+  hwloc_bitmap_asprintf(&a, set);
+  hwloc_bitmap_asprintf(&b, set2);
+  hwloc_bitmap_asprintf(&c, nomembutcpucpuset);
+  printf("nomembutcpucpuset %s\n", c);
+  printf("numa_all_nodes_ptr gave %s != %s\n", b, a);
+}
   assert(hwloc_bitmap_isequal(set, set2));
   hwloc_bitmap_free(set2);

[dgloe]

The host system has

CONFIG_NUMA=y
CONFIG_NUMA_EMU=y

The chroot doesn't have a /boot/config-* file.

With /proc mounted it's getting past the first assert. The first printf you suggested is giving:

abuild@cfosbld03:tests $ ./linux-libnuma
1 != 3
linux-libnuma: linux-libnuma.c:72: main: Assertion `numa_bitmask_equal(bitmask, numa_all_nodes_ptr)' failed.
Aborted (core dumped)

[bgoglin]

Your lstopo output and the printf says there is a single core ad hyperthread in the chroot, while libnuma says there are two of them. Is the chroot somehow restricted to a single core and hyperthread while the machine actually has two of them?

[bgoglin]

I am going to ignore/close this issue since it looks like libnuma doesn't properly report things in that chroot, and I don't think we care about libnuma<->hwloc interoperability regression checks in such an build environment anyway.