On 02/29/2016 08:33 PM, Zhang Chen wrote: > > > On 02/29/2016 03:11 PM, Jason Wang wrote: >> >> On 02/24/2016 05:03 PM, Zhang Chen wrote: >>> If queue=rx, filter-redirector will get the packet that guest send, >>> then redirect >>> to outdev(if none, do nothing). but queue=rx/tx/all not related to >>> indev. please >>> look the flow chart below. queue=xxx just work for one >>> way(filter->outdev). >>> >>> filter >>> + >>> | >>> | >>> redirector | >>> +-------------------------+ >>> | | | >>> | | | >>> | | | >>> indev +----------------+ +----------------> outdev >>> | | | >>> | | | >>> | | | >>> +-------------------------+ >>> | >>> | >>> v >>> filter >>> >>> | >>> >>> | >>> >>> v >>> filter ........ filter ...... guest >>> >> This looks a violation on the assumption of current filter behavior. >> Each filter should only talk to the 'next' or 'prev' filter on the chain >> (depends on the direction) or netdev when queue=rx or netdev's peer when >> queue=tx. >> >> And in fact there's subtle differences with your patch: >> >> When queue='all' since you force nf->netdev as sender, direction is >> NET_FILTER_DIRECTION_TX, the packet will be passed to 'next' filter on >> the chain. >> When queue='rx', direction is NET_FILTER_DIRECTION_RX, the packet will >> be pass to 'prev' filter on the chain. >> >> So as you can see, 'all' is ambiguous here. I think we should keep >> current behavior by redirecting traffic to netdev when queue='rx'. For >> queue='all', maybe we need redirect the traffic to both netdev and >> netdev's peer. >> >> > > OK, I will change usage to : > > -filter-redirector,id=r0,netdev=hn0,queue=tx/rx/all,outdev=s1,indev=s0,in_direction=tx/rx > > How about this?
Looks like in_direction complicates the issue and is unnecessary. In fact, it could be achieved by having another redirector. > > I will fix it in V3. > > Thanks > zhangchen
