@@ -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;
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> --- tools/testing/selftests/vDSO/vdso_test_getrandom.c | 8 +++++--- 1 file changed, 5 insertions(+), 3 deletions(-)