On 30/06/26 15:02, Sayali Patil wrote:
The KSM NUMA merge test allocates identical pages on different NUMA
nodes and verifies KSM behavior with merge_across_nodes enabled and
disabled.
On systems with memoryless NUMA nodes, for example:
#numactl -H
available: 2 nodes (0,4)
.....
node 0 cpus: 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
node 0 size: 14825 MB
node 0 free: 1382 MB
node 4 cpus:
node 4 size: 0 MB
node 4 free: 0 MB
the test may attempt to allocate memory on a node without memory,
causing numa_alloc_onnode() to fail and resulting in a spurious test
failure.
The test currently checks numa_num_configured_nodes() to determine
whether sufficient NUMA nodes are available. However, configured nodes
do not necessarily have memory.
Reuse the existing get_first_mem_node() and get_next_mem_node()
helpers to locate NUMA nodes that actually contain memory, and skip
the test when fewer than two such nodes are available.
Before patch:
---------------------------
running ./ksm_tests -N -m 1
---------------------------
mbind: Invalid argument
ok 1 KSM NUMA merging
Totals: pass:1 fail:0 xfail:0 xpass:0 skip:0 error:0
[PASS]
ok 1 ksm_tests -N -m 1
---------------------------
running ./ksm_tests -N -m 0
---------------------------
mbind: Invalid argument
not ok 1 KSM NUMA merging
Totals: pass:0 fail:1 xfail:0 xpass:0 skip:0 error:0
[FAIL]
not ok 2 ksm_tests -N -m 0 # exit=1
After patch:
---------------------------
running ./ksm_tests -N -m 1
---------------------------
At least 2 NUMA nodes with memory must be available
ok 1
SKIP KSM NUMA merging
Totals: pass:0 fail:0 xfail:0 xpass:0 skip:1 error:0
[PASS]
ok 1 ksm_tests -N -m 1
---------------------------
running ./ksm_tests -N -m 0
---------------------------
At least 2 NUMA nodes with memory must be available
ok 1
SKIP KSM NUMA merging
Totals: pass:0 fail:0 xfail:0 xpass:0 skip:1 error:0
[PASS]
ok 2 ksm_tests -N -m 0
Fixes: e3820ab252dd ("selftest/vm: fix ksm selftest to run with different NUMA
topologies")
Co-developed-by: David Hildenbrand (Arm) <[email protected]>
Signed-off-by: David Hildenbrand (Arm) <[email protected]>
Signed-off-by: Sayali Patil <[email protected]>
---
tools/testing/selftests/mm/ksm_tests.c | 16 +++++++++-------
1 file changed, 9 insertions(+), 7 deletions(-)
diff --git a/tools/testing/selftests/mm/ksm_tests.c
b/tools/testing/selftests/mm/ksm_tests.c
index a050f4840cfa..2ebbb544c671 100644
--- a/tools/testing/selftests/mm/ksm_tests.c
+++ b/tools/testing/selftests/mm/ksm_tests.c
@@ -440,9 +440,9 @@ static int get_next_mem_node(int node)
mem_node = i % (max_node + 1);
node_size = numa_node_size(mem_node, NULL);
if (node_size > 0)
- break;
+ return mem_node;
}
- return mem_node;
+ return -ENODEV;
}
static int get_first_mem_node(void)
@@ -455,8 +455,8 @@ static int check_ksm_numa_merge(int merge_type, int
mapping, int prot, int timeo
{
void *numa1_map_ptr, *numa2_map_ptr;
struct timespec start_time;
+ int first_node, second_node;
int page_count = 2;
- int first_node;
if (clock_gettime(CLOCK_MONOTONIC_RAW, &start_time)) {
ksft_perror("clock_gettime");
@@ -467,17 +467,19 @@ static int check_ksm_numa_merge(int merge_type, int
mapping, int prot, int timeo
ksft_print_msg("NUMA support not enabled\n");
return KSFT_SKIP;
}
- if (numa_num_configured_nodes() <= 1) {
- ksft_print_msg("At least 2 NUMA nodes must be available\n");
+ first_node = get_first_mem_node();
+ second_node = get_next_mem_node(first_node);
+
+ if (second_node < 0) {
+ ksft_print_msg("At least 2 NUMA nodes with memory must be
available\n");
return KSFT_SKIP;
}
if (ksm_write_sysfs(KSM_FP("merge_across_nodes"), merge_across_nodes))
return KSFT_FAIL;
/* allocate 2 pages in 2 different NUMA nodes and fill them with the same data */
- first_node = get_first_mem_node();
numa1_map_ptr = numa_alloc_onnode(page_size, first_node);
- numa2_map_ptr = numa_alloc_onnode(page_size,
get_next_mem_node(first_node));
+ numa2_map_ptr = numa_alloc_onnode(page_size, second_node);
if (!numa1_map_ptr || !numa2_map_ptr) {
ksft_perror("numa_alloc_onnode");
return KSFT_FAIL;
AI review comment:
If get_first_mem_node() doesn't find a node with memory, will it
return -ENODEV and pass it directly into get_next_mem_node()?
Looking at get_next_mem_node() with a negative node parameter:
tools/testing/selftests/mm/ksm_tests.c:get_next_mem_node() {
...
for (i = node + 1; i <= max_node + node; i++) {
mem_node = i % (max_node + 1);
node_size = numa_node_size(mem_node, NULL);
...
}
Because the modulo operator preserves the sign of a negative dividend,
starting the loop with a negative node value causes mem_node to become
negative. This means a negative node ID is then passed into libnuma's
numa_node_size() function.
Could we check if first_node is valid before attempting to find the
second node?
The additional check for get_first_mem_node() is not needed. After
numa_available() succeeds, a running system should always have at least
one NUMA node with memory, so get_first_mem_node() is expected to return
a valid node ID. The actual prerequisite for this test is the presence
of two NUMA nodes with memory, which is already validated by checking
the return value of get_next_mem_node().
Thanks,
Sayali
