Updating a BPF_MAP_TYPE_HASH_OF_MAPS or BPF_MAP_TYPE_ARRAY_OF_MAPS via bpf_map_update_elem() is very expensive. In one of our workloads, we're inserting ~1400 maps of type BPF_MAP_TYPE_ARRAY into a BPF_MAP_TYPE_ARRAY_OF_MAPS. This takes ~21 seconds on a single thread, with an average of ~15ms per call: Function Name: map_update_elem Number of calls: 1369 Total time: 21s 182ms 966µs Maximum: 47ms 937µs Average: 15ms 473µs Minimum: 7µs Profiling shows that nearly all of this time is going to synchronize_rcu(), via maybe_wait_bpf_programs() in map_update_elem(). The call to synchronize_rcu() is done to ensure that after bpf_map_update_elem() returns, no BPF programs are still looking at the old value of the map, per commit 1ae80cf31938 ("bpf: wait for running BPF programs when updating map-in-map"). As discussed on the bpf mailing list, replace synchronize_rcu() with synchronize_rcu_expedited(). This is 175x faster: it now takes an average of 88 microseconds per call, for a total of 127 milliseconds in the same benchmark: Function Name: map_update_elem Number of calls: 1439 Total time: 127ms 626µs Maximum: 445µs Average: 88µs Minimum: 10µs Link: https://lore.kernel.org/bpf/CAH6OuBR=w2kybK6u7aH_35B=Bo1PCukeMZefR=7V4Z2tJNK--Q@mail.gmail.com/ Signed-off-by: Ritesh Oedayrajsingh Varma --- kernel/bpf/syscall.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/kernel/bpf/syscall.c b/kernel/bpf/syscall.c index d5851800b3de..ea4c19ae3edc 100644 --- a/kernel/bpf/syscall.c +++ b/kernel/bpf/syscall.c @@ -158,7 +158,7 @@ static void maybe_wait_bpf_programs(struct bpf_map *map) */ if (map->map_type == BPF_MAP_TYPE_HASH_OF_MAPS || map->map_type == BPF_MAP_TYPE_ARRAY_OF_MAPS) - synchronize_rcu(); + synchronize_rcu_expedited(); } static void unpin_uptr_kaddr(void *kaddr) -- 2.52.0