On 2019/5/28 9:17, Stephen Hemminger wrote: > On Tue, 28 May 2019 09:04:18 +0800 > Yunsheng Lin <linyunsh...@huawei.com> wrote: > >> On 2019/5/27 22:58, Stephen Hemminger wrote: >>> On Mon, 27 May 2019 09:47:54 +0800 >>> Yunsheng Lin <linyunsh...@huawei.com> wrote: >>> >>>> When user has configured a large number of virtual netdev, such >>>> as 4K vlans, the carrier on/off operation of the real netdev >>>> will also cause it's virtual netdev's link state to be processed >>>> in linkwatch. Currently, the processing is done in a work queue, >>>> which may cause worker starvation problem for other work queue. >>>> >>>> This patch releases the cpu when link watch worker has processed >>>> a fixed number of netdev' link watch event, and schedule the >>>> work queue again when there is still link watch event remaining. >>>> >>>> Signed-off-by: Yunsheng Lin <linyunsh...@huawei.com> >>> >>> Why not put link watch in its own workqueue so it is scheduled >>> separately from the system workqueue? >> >> From testing and debuging, the workqueue runs on the cpu where the >> workqueue is schedule when using normal workqueue, even using its >> own workqueue instead of system workqueue. So if the cpu is busy >> processing the linkwatch event, it is not able to process other >> workqueue' work when the workqueue is scheduled on the same cpu. >> >> Using unbound workqueue may solve the cpu starvation problem. >> But the __linkwatch_run_queue is called with rtnl_lock, so if it >> takes a lot time to process, other need to take the rtnl_lock may >> not be able to move forward. > > Agree with the starvation issue. My cocern is that large number of > events that end up being delayed would impact things that are actually > watching for link events (like routing daemons).
Agreed. I am not familiar with above use cases, it would be very helpful if someone can help testing the impact of above use case. > > It probably would be not accepted to do rtnl_unlock/sched_yield/rtnl_lock > in the loop, but that is another alternative. Yes. But seems not very efficient to do rtnl_unlock/sched_yield/rtnl_lock for very linkwatch_do_dev. > > > > . >
