Johannes Berg <johannes@xxxxxxxxxxxxxxxx> writes: > On Wed, 2019-12-11 at 15:04 +0100, Toke Høiland-Jørgensen wrote: >> Johannes Berg <johannes@xxxxxxxxxxxxxxxx> writes: >> >> > Btw, there's *another* issue. You said in the commit log: >> > >> > This patch does *not* include any mechanism to wake a throttled TXQ again, >> > on the assumption that this will happen anyway as a side effect of whatever >> > freed the skb (most commonly a TX completion). >> > >> > Thinking about this some more, I'm not convinced that this assumption >> > holds. You could have been stopped due to the global limit, and now you >> > wake some queue but the TXQ is empty - now you should reschedule some >> > *other* TXQ since the global limit had kicked in, not the per-TXQ limit, >> > and prevented dequeuing, no? >> >> Well if you hit the global limit that means you have 24ms worth of data >> queued in the hardware; those should be completed in turn, and enable >> more to be dequeued, no? > > Yes, but on which queues? > > Say you have some queues - some (Q1-Qn) got a LOT of traffic, and > another (Q0) just has some interactive traffic. > > You could then end up in a situation where you have 24ms queued up on > Q1-Qn (with n high enough to not have hit the per-queue AQL limit), > right? > > Say also the last frame on Q0 was dequeued by the hardware, but the > tx_dequeue() got NULL because of the AQL limit having been eaten up by > all the packets on Q1-Qn. > > Now you'll no longer get a new dequeue attempt on Q0 (it was already > empty last time, so no hardware reclaim to trigger new dequeues), and a > new dequeue on the *other* queues will not do anything for this queue. Oh, right, I see; yeah, that could probably happen. I guess we could either kick all available queues whenever the global limit goes from "above" to "below"; or we could remove the "return NULL" logic from tx_dequeue() and rely on next_txq() to throttle. I think the latter is probably simpler, but I'm a little worried that the throttling will become too lax (because the driver can keep dequeueing in the same scheduling round)... -Toke
