On Mon, 23 Oct 2023 19:24:52 +0800 "wuqiang.matt" <wuqiang.m...@bytedance.com> wrote:
> The objpool_push can only happen on local cpu node, so only the local > cpu can touch slot->tail and slot->last, which ensures the correctness > of using cmpxchg without lock prefix (using try_cmpxchg_local instead > of try_cmpxchg_acquire). > > Testing with IACA found the lock version of pop/push pair costs 16.46 > cycles and local-push version costs 15.63 cycles. Kretprobe throughput > is improved to 1.019 times of the lock version for x86_64 systems. > > OS: Debian 10 X86_64, Linux 6.6rc6 with freelist > HW: XEON 8336C x 2, 64 cores/128 threads, DDR4 3200MT/s > > 1T 2T 4T 8T 16T > lock: 29909085 59865637 119692073 239750369 478005250 > local: 30297523 60532376 121147338 242598499 484620355 > 32T 48T 64T 96T 128T > lock: 957553042 1435814086 1680872925 2043126796 2165424198 > local: 968526317 1454991286 1861053557 2059530343 2171732306 > Yeah, slot->tail is only used on the local CPU. This looks good to me. Acked-by: Masami Hiramatsu (Google) <mhira...@kernel.org> Thanks! > Signed-off-by: wuqiang.matt <wuqiang.m...@bytedance.com> > --- > lib/objpool.c | 2 +- > 1 file changed, 1 insertion(+), 1 deletion(-) > > diff --git a/lib/objpool.c b/lib/objpool.c > index ce0087f64400..a032701beccb 100644 > --- a/lib/objpool.c > +++ b/lib/objpool.c > @@ -166,7 +166,7 @@ objpool_try_add_slot(void *obj, struct objpool_head > *pool, int cpu) > head = READ_ONCE(slot->head); > /* fault caught: something must be wrong */ > WARN_ON_ONCE(tail - head > pool->nr_objs); > - } while (!try_cmpxchg_acquire(&slot->tail, &tail, tail + 1)); > + } while (!try_cmpxchg_local(&slot->tail, &tail, tail + 1)); > > /* now the tail position is reserved for the given obj */ > WRITE_ONCE(slot->entries[tail & slot->mask], obj); > -- > 2.40.1 > -- Masami Hiramatsu (Google) <mhira...@kernel.org>
