On 2024-11-06 17:48:59 [+0800], Hou Tao wrote: > Hi, Hi, > Yes. The patch set still invokes check_and_free_fields() under the > bucket lock when updating an existing element in a pre-allocated htab. I > missed the hrtimer case. For the sleeping lock, you mean the > cpu_base->softirq_expiry_lock in hrtimer_cancel_waiting_running(), right Yes. > ? Instead of cancelling the timer in workqueue, maybe we could save the > old value temporarily in the bucket lock, and try to free it outside of > the bucket lock or disabling the extra_elems logic temporarily for the > case ? Well, it is up to you. Either: diff --git a/kernel/bpf/helpers.c b/kernel/bpf/helpers.c index 1a43d06eab286..b077af12fc9b4 100644 --- a/kernel/bpf/helpers.c +++ b/kernel/bpf/helpers.c @@ -1593,10 +1593,7 @@ void bpf_timer_cancel_and_free(void *val) * To avoid these issues, punt to workqueue context when we are in a * timer callback. */ - if (this_cpu_read(hrtimer_running)) - queue_work(system_unbound_wq, &t->cb.delete_work); - else - bpf_timer_delete_work(&t->cb.delete_work); + queue_work(system_unbound_wq, &t->cb.delete_work); } /* This function is called by map_delete/update_elem for individual element and Or something smarter where you cancel the timer outside of the bucket lock. Sebastian
