On 6/18/2018 10:49 AM, Ido Goshen wrote: > I'm really not sure that just setting the pcaps/dumpers.num_of_queue w/o > actually creating multiple queues is good enough. > If one uses N queues there are good changes he is using N cores. > To be consistent with DPDK behavior it should be safe to concurrently tx to > different queues. > If pcap_sendpacket to the same pcap_t handle is not thread safe then it will > require to pcap_open_live() for each queue > If using multiple pcap_open_live() then it will cause the rx out direction > problem again
Please do not reply top post, this thread become very hard to follow. I am not suggesting just increase num_of_queue, simply what I was suggesting it add capability to create multiple queues by using "iface" devarg, that is all. This should solve your problem because with "iface" devarg I don't get Tx packets back with my Rx handler. This has a downside that you create queues as pairs, you can't get 4Rx 1Tx queues etc.. So I agree with your suggestion, but there was some comments to your initial patch [1], can you please address them and send a new version? [1] https://mails.dpdk.org/archives/dev/2018-June/103476.html Thanks. > > -----Original Message----- > From: Ferruh Yigit <ferruh.yi...@intel.com> > Sent: Monday, June 18, 2018 11:13 AM > To: Ido Goshen <i...@cgstowernetworks.com> > Cc: dev@dpdk.org > Subject: Re: [PATCH v2] net/pcap: rx_iface_in stream type support > > On 6/16/2018 10:27 AM, Ido Goshen wrote: >> Is pcap_sendpacket() to the same pcap_t handle thread-safe? I couldn't find >> clear answer so I'd rather assume not. >> If it's not thread-safe then supporting multiple "iface"'s will require >> multiple pcap_open_live()'s and we are back in the same place. > > I am not suggesting extra multi thread safety. > > Currently in "iface" path, following is hardcoded: > pcaps.num_of_queue = 1; > dumpers.num_of_queue = 1; > > It can be possible to update that path to support multiple queue while using > "iface" devargs. > >> >>>> I am not sure exiting behavior is intentional, which is capturing sent >>>> packages in Rx pcap handler to same interface. >>>> Are you aware of any use case of existing behavior? Perhaps it can be >>>> option to set PCAP_D_IN by default for rx_iface argument. >> Even if unintentional I find it very useful for testing, as this way it's >> very easy to send traffic to the app by tcpreplay on the same host the app >> is running on. >> Using tcpreplay is in the out direction that will not be captured if >> PCAP_D_IN is set. >> If PCAP_D_IN is the only option then it will require external device (or >> some networking trick) to send packets to the app. >> So, I'd say it is good for testing but less good for real >> functionality > > OK to keep it if there is a usecase. > >> >> -----Original Message----- >> From: Ferruh Yigit <ferruh.yi...@intel.com> >> Sent: Friday, June 15, 2018 3:53 PM >> To: Ido Goshen <i...@cgstowernetworks.com> >> Cc: dev@dpdk.org >> Subject: Re: [PATCH v2] net/pcap: rx_iface_in stream type support >> >> On 6/14/2018 9:44 PM, Ido Goshen wrote: >>> I think we are starting to mix two things One is how to configure >>> pcap eth dev with multiple queues and I totally agree it would have been >>> nicer to just say something like "max_tx_queues =N" instead of needing to >>> write "tx_iface" N times, but as it was already supported in that way (for >>> any reason?) I wasn't trying to enhance or change it. >>> The other issue is pcap direction API, which I was trying to expose to >>> users of dpdk pcap device. >> >> Hi Ido, >> >> Assuming "iface" argument solves the direction issue, I am suggestion adding >> multiqueue support to "iface" argument as a solution to your problem. >> >> I am not suggesting using new arguments like "max_tx_queues =N", "iface" can >> be used same as how rx/tx_ifcase used now, provide it multiple times. >> >>> Refer to https://www.tcpdump.org/manpages/pcap_setdirection.3pcap.txt >>> or man tcpdump for -P/--direction in|out|inout option, Actually I >>> think a more realistic emulation of a physical device (non-virtual) >>> would be to capture only the incoming direction (set PCAP_D_IN), >>> again the existing behavior is very useful too and I didn't try to >>> change or eliminate it but just add additional stream type option >> >> I am not sure exiting behavior is intentional, which is capturing sent >> packages in Rx pcap handler to same interface. >> Are you aware of any use case of existing behavior? Perhaps it can be option >> to set PCAP_D_IN by default for rx_iface argument. >> >>> >>> -----Original Message----- >>> From: Ferruh Yigit <ferruh.yi...@intel.com> >>> Sent: Thursday, June 14, 2018 9:09 PM >>> To: Ido Goshen <i...@cgstowernetworks.com> >>> Cc: dev@dpdk.org >>> Subject: Re: [PATCH v2] net/pcap: rx_iface_in stream type support >>> >>> On 6/14/2018 6:14 PM, Ido Goshen wrote: >>>> I use "rx_iface","tx_iface" (and not just "iface") in order to have >>>> multiple TX queues I just gave a simplified setting with 1 queue My >>>> app does a full mesh between the ports (not fixed pairs like l2fwd) >>>> so all the forwarding lcores can tx to the same port simultaneously and as >>>> DPDK docs say: >>>> "Multiple logical cores should never share receive or transmit queues for >>>> interfaces since this would require global locks and hinder performance." >>>> For example if I have 3 ports handled by 3 cores it'll be >>>> myapp -c 7 -n1 --no-huge \ >>>> >>>> --vdev=eth_pcap0,rx_iface=eth0,tx_iface=eth0,tx_iface=eth0,tx_iface=eth0 \ >>>> >>>> --vdev=eth_pcap0,rx_iface=eth1,tx_iface=eth1,tx_iface=eth1,tx_iface=eth1 \ >>>> >>>> --vdev=eth_pcap0,rx_iface=eth2,tx_iface=eth2,tx_iface=eth2,tx_iface=eth2 \ >>>> -- -p 7 >>>> Is there another way to achieve multiple queues in pcap vdev? >>> >>> If you want to use multiple core you need multiple queues, as you said, and >>> above is the way to create multiple queues for pcap. >>> >>> Currently "iface" argument only supports single interface in a hardcoded >>> way, but technically it should be possible to update it to support multiple >>> queue. >>> >>> So if "iface" arguments works for you, it can be better to add multi queue >>> support to "iface" instead of introducing a new device argument. >>> >>>> >>>> I do see that using "iface" behaves differently - I'll try to >>>> investigate why >>> >>> pcap_open_live() is called for both arguments, for "rx_iface/tx_iface" pair >>> it has been called twice one for each. Not sure if pcap library returns >>> same handler or two different handlers for this case since iface name is >>> same. >>> For "iface" argument pcap_open_live() called once, so we have single >>> handler for both Rx & Tx. This may be difference. >>> >>>> And still even when using "iface" I also see packets that are >>>> transmitted out of eth1 (e.g. tcpreplay -i eth1 packets.pcap) and >>>> not only packets that are received (e.g. ping from far end to eth0 >>>> ip) >>> >>> This is interesting, I have tried with external packet generator, "iface" >>> was working as expected for me. >>> >>>> >>>> >>>> -----Original Message----- >>>> From: Ferruh Yigit <ferruh.yi...@intel.com> >>>> Sent: Wednesday, June 13, 2018 1:57 PM >>>> To: Ido Goshen <i...@cgstowernetworks.com> >>>> Cc: dev@dpdk.org >>>> Subject: Re: [PATCH v2] net/pcap: rx_iface_in stream type support >>>> >>>> On 6/5/2018 6:10 PM, Ido Goshen wrote: >>>>> The problem is if a dpdk app uses the same iface(s) both as rx_iface and >>>>> tx_iface then it will receive back the packets it sends. >>>>> If my app sends a packet to portid=X with rte_eth_tx_burst() then I >>>>> wouldn't expect to receive it back by rte_eth_rx_burst() for that same >>>>> portid=X (assuming of course there's no external loopback) This is >>>>> coming from the default nature of pcap that like a sniffer captures both >>>>> incoming and outgoing direction. >>>>> The patch provides an option to limit pcap rx_iface to get only incoming >>>>> traffic which is more like a real (non-pcap) dpdk device. >>>>> >>>>> for example: >>>>> when using existing *rx_iface* >>>>> l2fwd -c 3 -n1 --no-huge >>>>> --vdev=eth_pcap0,rx_iface=eth1,tx_iface=eth1 >>>>> --vdev=eth_pcap1,rx_iface=dummy0,tx_iface=dummy0 -- -p 3 -T 1 >>>>> sending only 1 single packet into eth1 will end in an infinite loop >>>>> - >>>> >>>> If you are using same interface for both Rx & Tx, why not using "iface=xxx" >>>> argument, can you please test with following: >>>> >>>> l2fwd -c 3 -n1 --no-huge --vdev=eth_pcap0,iface=eth1 >>>> --vdev=eth_pcap1,iface=dummy0 -- -p 3 -T 1 >>>> >>>> >>>> I can't reproduce the issue with above command. >>>> >>>> Thanks, >>>> ferruh >>>> >>> >> >