Re: [PATCH v2 2/3] selftests/mm: fix ksm NUMA merge test for systems with memoryless NUMA nodes
From: Sayali Patil <hidden>
Date: 2026-06-30 10:50:57
Also in:
linux-kselftest, linux-mm, lkml
On 30/06/26 15:02, Sayali Patil wrote:
quoted hunk ↗ jump to hunk
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) <david@kernel.org> Signed-off-by: David Hildenbrand (Arm) <david@kernel.org> Signed-off-by: Sayali Patil <redacted> --- 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