"Jason A. Donenfeld" <[email protected]> writes: > Hi Sebastian, > > Seems like you've identified two things, the use of need_resched, and > potentially surrounding napi_schedule in local_bh_{disable,enable}. > > Regarding need_resched, I pulled that out of other code that seemed to > have the "same requirements", as vaguely conceived. It indeed might > not be right. The intent is to have that worker running at maximum > throughput for extended periods of time, but not preventing other > threads from running elsewhere, so that, e.g., a user's machine > doesn't have a jenky mouse when downloading a file. > > What are the effects of unconditionally calling cond_resched() without > checking for if (need_resched())? Sounds like you're saying none at > all?
I believe so: AFAIU, you use need_resched() if you need to do some kind of teardown before the schedule point, like this example I was recently looking at: https://elixir.bootlin.com/linux/latest/source/net/bpf/test_run.c#L73 If you just need to maybe reschedule, you can just call cond_resched() and it'll do what it says on the tin: do a schedule if needed, and return immediately otherwise. > Regarding napi_schedule, I actually wasn't aware that it's requirement > to _only_ ever run from softirq was a strict one. When I switched to > using napi_schedule in this way, throughput really jumped up > significantly. Part of this indeed is from the batching, so that the > napi callback can then handle more packets in one go later. But I > assumed it was something inside of NAPI that was batching and > scheduling it, rather than a mistake on my part to call this from a wq > and not from a softirq. > > What, then, are the effects of surrounding that in > local_bh_{disable,enable} as you've done in the patch? You mentioned > one aspect is that it will "invoke wg_packet_rx_poll() where you see > only one skb." It sounds like that'd be bad for performance, though, > given that the design of napi is really geared toward batching. Heh, I wrote a whole long explanation he about variable batch sizes because you don't control when the NAPI is scheduled, etc... And then I noticed the while loop is calling ptr_ring_consume_bh(), which means that there's already a local_bh_disable/enable pair on every loop invocation. So you already have this :) Which of course raises the question of whether there's anything to gain from *adding* batching to the worker? Something like: #define BATCH_SIZE 8 void wg_packet_decrypt_worker(struct work_struct *work) { struct crypt_queue *queue = container_of(work, struct multicore_worker, work)->ptr; void *skbs[BATCH_SIZE]; bool again; int i; restart: local_bh_disable(); ptr_ring_consume_batched(&queue->ring, skbs, BATCH_SIZE); for (i = 0; i < BATCH_SIZE; i++) { struct sk_buff *skb = skbs[i]; enum packet_state state; if (!skb) break; state = likely(decrypt_packet(skb, PACKET_CB(skb)->keypair)) ? PACKET_STATE_CRYPTED : PACKET_STATE_DEAD; wg_queue_enqueue_per_peer_rx(skb, state); } again = !ptr_ring_empty(&queue->ring); local_bh_enable(); if (again) { cond_resched(); goto restart; } } Another thing that might be worth looking into is whether it makes sense to enable threaded NAPI for Wireguard. See: https://lore.kernel.org/r/[email protected] -Toke
