diff options
| author | Jason A. Donenfeld <Jason@zx2c4.com> | 2024-09-29 04:55:49 +0200 |
|---|---|---|
| committer | Shuah Khan <skhan@linuxfoundation.org> | 2024-10-01 08:30:58 -0600 |
| commit | a18c835779e1a2ecf8e83c18f5af6a3b05699aaa (patch) | |
| tree | a8be7c94c81a8c548b49fdd7467bec8e9b7d126f /tools | |
| parent | 45a8897db67d43aad3a17a792fe06a6306f483d6 (diff) | |
selftests: vDSO: align getrandom states to cache line
This prevents false sharing, which makes a large difference on machines
with several NUMA nodes, such as on a dual socket Intel(R) Xeon(R) Gold
6338 CPU @ 2.00GHz, where the "bench-multi" test goes from 2.7s down to
1.9s. While this is just test code, it also forms the basis of how folks
will wind up implementing this in libraries, so we should implement this
simple cache alignment improvement here.
Suggested-by: Florian Weimer <fweimer@redhat.com>
Cc: Adhemerval Zanella <adhemerval.zanella@linaro.org>
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
Signed-off-by: Shuah Khan <skhan@linuxfoundation.org>
Diffstat (limited to 'tools')
| -rw-r--r-- | tools/testing/selftests/vDSO/vdso_test_getrandom.c | 8 |
1 files changed, 5 insertions, 3 deletions
diff --git a/tools/testing/selftests/vDSO/vdso_test_getrandom.c b/tools/testing/selftests/vDSO/vdso_test_getrandom.c index 72a1d9b43a84..e5e83dbec589 100644 --- a/tools/testing/selftests/vDSO/vdso_test_getrandom.c +++ b/tools/testing/selftests/vDSO/vdso_test_getrandom.c @@ -59,10 +59,12 @@ static void *vgetrandom_get_state(void) size_t page_size = getpagesize(); size_t new_cap; size_t alloc_size, num = sysconf(_SC_NPROCESSORS_ONLN); /* Just a decent heuristic. */ + size_t state_size_aligned, cache_line_size = sysconf(_SC_LEVEL1_DCACHE_LINESIZE) ?: 1; void *new_block, *new_states; - alloc_size = (num * vgrnd.params.size_of_opaque_state + page_size - 1) & (~(page_size - 1)); - num = (page_size / vgrnd.params.size_of_opaque_state) * (alloc_size / page_size); + state_size_aligned = (vgrnd.params.size_of_opaque_state + cache_line_size - 1) & (~(cache_line_size - 1)); + alloc_size = (num * state_size_aligned + page_size - 1) & (~(page_size - 1)); + num = (page_size / state_size_aligned) * (alloc_size / page_size); new_block = mmap(0, alloc_size, vgrnd.params.mmap_prot, vgrnd.params.mmap_flags, -1, 0); if (new_block == MAP_FAILED) goto out; @@ -78,7 +80,7 @@ static void *vgetrandom_get_state(void) if (((uintptr_t)new_block & (page_size - 1)) + vgrnd.params.size_of_opaque_state > page_size) new_block = (void *)(((uintptr_t)new_block + page_size - 1) & (~(page_size - 1))); vgrnd.states[i] = new_block; - new_block += vgrnd.params.size_of_opaque_state; + new_block += state_size_aligned; } vgrnd.len = num; goto success; |