On Wed, 2021-08-11 at 12:59 +0000, Xueming(Steven) Li wrote: > > > -----Original Message----- > > From: Ferruh Yigit <ferruh.yi...@intel.com> > > Sent: Wednesday, August 11, 2021 8:04 PM > > To: Xueming(Steven) Li <xuemi...@nvidia.com>; Jerin Jacob > > <jerinjac...@gmail.com> > > Cc: dpdk-dev <dev@dpdk.org>; NBU-Contact-Thomas Monjalon > > <tho...@monjalon.net>; Andrew Rybchenko > > <andrew.rybche...@oktetlabs.ru> > > Subject: Re: [dpdk-dev] [PATCH v1] ethdev: introduce shared Rx queue > > > > On 8/11/2021 9:28 AM, Xueming(Steven) Li wrote: > > > > > > > > > > -----Original Message----- > > > > From: Jerin Jacob <jerinjac...@gmail.com> > > > > Sent: Wednesday, August 11, 2021 4:03 PM > > > > To: Xueming(Steven) Li <xuemi...@nvidia.com> > > > > Cc: dpdk-dev <dev@dpdk.org>; Ferruh Yigit <ferruh.yi...@intel.com>; > > > > NBU-Contact-Thomas Monjalon <tho...@monjalon.net>; Andrew Rybchenko > > > > <andrew.rybche...@oktetlabs.ru> > > > > Subject: Re: [dpdk-dev] [PATCH v1] ethdev: introduce shared Rx queue > > > > > > > > On Mon, Aug 9, 2021 at 7:46 PM Xueming(Steven) Li <xuemi...@nvidia.com> > > > > wrote: > > > > > > > > > > Hi, > > > > > > > > > > > -----Original Message----- > > > > > > From: Jerin Jacob <jerinjac...@gmail.com> > > > > > > Sent: Monday, August 9, 2021 9:51 PM > > > > > > To: Xueming(Steven) Li <xuemi...@nvidia.com> > > > > > > Cc: dpdk-dev <dev@dpdk.org>; Ferruh Yigit <ferruh.yi...@intel.com>; > > > > > > NBU-Contact-Thomas Monjalon <tho...@monjalon.net>; Andrew Rybchenko > > > > > > <andrew.rybche...@oktetlabs.ru> > > > > > > Subject: Re: [dpdk-dev] [PATCH v1] ethdev: introduce shared Rx > > > > > > queue > > > > > > > > > > > > On Mon, Aug 9, 2021 at 5:18 PM Xueming Li <xuemi...@nvidia.com> > > > > > > wrote: > > > > > > > > > > > > > > In current DPDK framework, each RX queue is pre-loaded with mbufs > > > > > > > for incoming packets. When number of representors scale out in a > > > > > > > switch domain, the memory consumption became significant. Most > > > > > > > important, polling all ports leads to high cache miss, high > > > > > > > latency and low throughput. > > > > > > > > > > > > > > This patch introduces shared RX queue. Ports with same > > > > > > > configuration in a switch domain could share RX queue set by > > > > > > > specifying sharing group. > > > > > > > Polling any queue using same shared RX queue receives packets from > > > > > > > all member ports. Source port is identified by mbuf->port. > > > > > > > > > > > > > > Port queue number in a shared group should be identical. Queue > > > > > > > index is > > > > > > > 1:1 mapped in shared group. > > > > > > > > > > > > > > Share RX queue is supposed to be polled on same thread. > > > > > > > > > > > > > > Multiple groups is supported by group ID. > > > > > > > > > > > > Is this offload specific to the representor? If so can this name be > > > > > > changed specifically to representor? > > > > > > > > > > Yes, PF and representor in switch domain could take advantage. > > > > > > > > > > > If it is for a generic case, how the flow ordering will be > > > > > > maintained? > > > > > > > > > > Not quite sure that I understood your question. The control path of > > > > > is almost same as before, PF and representor port still needed, rte > > > > > flows not impacted. > > > > > Queues still needed for each member port, descriptors(mbuf) will be > > > > > supplied from shared Rx queue in my PMD implementation. > > > > > > > > My question was if create a generic RTE_ETH_RX_OFFLOAD_SHARED_RXQ > > > > offload, multiple ethdev receive queues land into the same receive > > > > queue, In that case, how the flow order is maintained for > > respective receive queues. > > > > > > I guess the question is testpmd forward stream? The forwarding logic has > > > to be changed slightly in case of shared rxq. > > > basically for each packet in rx_burst result, lookup source stream > > > according to mbuf->port, forwarding to target fs. > > > Packets from same source port could be grouped as a small burst to > > > process, this will accelerates the performance if traffic come from > > > limited ports. I'll introduce some common api to do shard rxq forwarding, > > > call it with packets handling callback, so it suites for all > > forwarding engine. Will sent patches soon. > > > > > > > All ports will put the packets in to the same queue (share queue), right? > > Does this means only single core will poll only, what will > > happen if there are multiple cores polling, won't it cause problem? > > This has been mentioned in commit log, the shared rxq is supposed to be > polling in single thread(core) - I think it should be "MUST". > Result is unexpected if there are multiple cores pooling, that's why I added > a polling schedule check in testpmd. > Similar for rx/tx burst function, a queue can't be polled on multiple > thread(core), and for performance concern, no such check in eal api. > > If users want to utilize multiple cores to distribute workloads, it's > possible to define more groups, queues in different group could be > could be polled on multiple cores. > > It's possible to poll every member port in group, but not necessary, any port > in group could be polled to get packets for all ports in group. > > If the member port subject to hot plug/remove, it's possible to create a > vdev with same queue number, copy rxq object and poll vdev > as a dedicate proxy for the group. > > > > > And if this requires specific changes in the application, I am not sure > > about the solution, can't this work in a transparent way to the > > application? > > Yes, we considered different options in design stage. One possible solution > is to cache received packets in rings, this can be done on > eth layer, but I'm afraid less benefits, user still has to be a ware of > multiple core polling. > This can be done as a wrapper PMD later, more efforts.
For people want to use shared rxq to save memory, they need to be conscious on the core polling rule: dedicate core for shared rxq, like the rule for rxq and txq. I'm afraid specific changes in application is a must, but not too much: polling one port per group is sufficient. Protections in data plane will definitely hurt performance :( > > > > > Overall, is this for optimizing memory for the port represontors? If so > > can't we have a port representor specific solution, reducing > > scope can reduce the complexity it brings? > > This feature supports both PF and representor, and yes, major issue is memory > of representors. Poll all representors also > introduces more core cache miss latency. This feature essentially aggregates > all ports in group as one port. > On the other hand, it's useful for rte flow to create offloading flows using > representor as a regular port ID. > As discussion with Jerin below, the major memory consumed by PF or representor is mbufs pre-filled to rxq. PMD can't assume all representors shares same memory pool, or share rxqs internally in PMD - user might schedule representors to different cores. Defining shared rxq flag and group looks a good direction. > It's great if any new solution/suggestion, my head buried in PMD code :) > > > > > > > If this offload is only useful for representor case, Can we make this > > > > offload specific to representor the case by changing its name and scope. > > > > > > It works for both PF and representors in same switch domain, for > > > application like OVS, few changes to apply. > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > Signed-off-by: Xueming Li <xuemi...@nvidia.com> > > > > > > > --- > > > > > > > doc/guides/nics/features.rst | 11 +++++++++++ > > > > > > > doc/guides/nics/features/default.ini | 1 + > > > > > > > doc/guides/prog_guide/switch_representation.rst | 10 ++++++++++ > > > > > > > lib/ethdev/rte_ethdev.c | 1 + > > > > > > > lib/ethdev/rte_ethdev.h | 7 +++++++ > > > > > > > 5 files changed, 30 insertions(+) > > > > > > > > > > > > > > diff --git a/doc/guides/nics/features.rst > > > > > > > b/doc/guides/nics/features.rst index a96e12d155..2e2a9b1554 100644 > > > > > > > --- a/doc/guides/nics/features.rst > > > > > > > +++ b/doc/guides/nics/features.rst > > > > > > > @@ -624,6 +624,17 @@ Supports inner packet L4 checksum. > > > > > > > > > > > > > > ``tx_offload_capa,tx_queue_offload_capa:DEV_TX_OFFLOAD_OUTER_UDP_CKSUM``. > > > > > > > > > > > > > > > > > > > > > +.. _nic_features_shared_rx_queue: > > > > > > > + > > > > > > > +Shared Rx queue > > > > > > > +--------------- > > > > > > > + > > > > > > > +Supports shared Rx queue for ports in same switch domain. > > > > > > > + > > > > > > > +* **[uses] rte_eth_rxconf,rte_eth_rxmode**: > > > > > > > ``offloads:RTE_ETH_RX_OFFLOAD_SHARED_RXQ``. > > > > > > > +* **[provides] mbuf**: ``mbuf.port``. > > > > > > > + > > > > > > > + > > > > > > > .. _nic_features_packet_type_parsing: > > > > > > > > > > > > > > Packet type parsing > > > > > > > diff --git a/doc/guides/nics/features/default.ini > > > > > > > b/doc/guides/nics/features/default.ini > > > > > > > index 754184ddd4..ebeb4c1851 100644 > > > > > > > --- a/doc/guides/nics/features/default.ini > > > > > > > +++ b/doc/guides/nics/features/default.ini > > > > > > > @@ -19,6 +19,7 @@ Free Tx mbuf on demand = > > > > > > > Queue start/stop = > > > > > > > Runtime Rx queue setup = > > > > > > > Runtime Tx queue setup = > > > > > > > +Shared Rx queue = > > > > > > > Burst mode info = > > > > > > > Power mgmt address monitor = > > > > > > > MTU update = > > > > > > > diff --git a/doc/guides/prog_guide/switch_representation.rst > > > > > > > b/doc/guides/prog_guide/switch_representation.rst > > > > > > > index ff6aa91c80..45bf5a3a10 100644 > > > > > > > --- a/doc/guides/prog_guide/switch_representation.rst > > > > > > > +++ b/doc/guides/prog_guide/switch_representation.rst > > > > > > > @@ -123,6 +123,16 @@ thought as a software "patch panel" > > > > > > > front-end for applications. > > > > > > > .. [1] `Ethernet switch device driver model (switchdev) > > > > > > > > > > > > > > <https://www.kernel.org/doc/Documentation/networking/switchdev.txt > > > > > > > > `_ > > > > > > > > > > > > > > +- Memory usage of representors is huge when number of representor > > > > > > > +grows, > > > > > > > + because PMD always allocate mbuf for each descriptor of Rx > > > > > > > queue. > > > > > > > + Polling the large number of ports brings more CPU load, cache > > > > > > > +miss and > > > > > > > + latency. Shared Rx queue can be used to share Rx queue between > > > > > > > +PF and > > > > > > > + representors in same switch domain. > > > > > > > +``RTE_ETH_RX_OFFLOAD_SHARED_RXQ`` > > > > > > > + is present in Rx offloading capability of device info. Setting > > > > > > > +the > > > > > > > + offloading flag in device Rx mode or Rx queue configuration to > > > > > > > +enable > > > > > > > + shared Rx queue. Polling any member port of shared Rx queue can > > > > > > > +return > > > > > > > + packets of all ports in group, port ID is saved in > > > > > > > ``mbuf.port``. > > > > > > > + > > > > > > > Basic SR-IOV > > > > > > > ------------ > > > > > > > > > > > > > > diff --git a/lib/ethdev/rte_ethdev.c b/lib/ethdev/rte_ethdev.c > > > > > > > index 9d95cd11e1..1361ff759a 100644 > > > > > > > --- a/lib/ethdev/rte_ethdev.c > > > > > > > +++ b/lib/ethdev/rte_ethdev.c > > > > > > > @@ -127,6 +127,7 @@ static const struct { > > > > > > > RTE_RX_OFFLOAD_BIT2STR(OUTER_UDP_CKSUM), > > > > > > > RTE_RX_OFFLOAD_BIT2STR(RSS_HASH), > > > > > > > RTE_ETH_RX_OFFLOAD_BIT2STR(BUFFER_SPLIT), > > > > > > > + RTE_ETH_RX_OFFLOAD_BIT2STR(SHARED_RXQ), > > > > > > > }; > > > > > > > > > > > > > > #undef RTE_RX_OFFLOAD_BIT2STR > > > > > > > diff --git a/lib/ethdev/rte_ethdev.h b/lib/ethdev/rte_ethdev.h > > > > > > > index d2b27c351f..a578c9db9d 100644 > > > > > > > --- a/lib/ethdev/rte_ethdev.h > > > > > > > +++ b/lib/ethdev/rte_ethdev.h > > > > > > > @@ -1047,6 +1047,7 @@ struct rte_eth_rxconf { > > > > > > > uint8_t rx_drop_en; /**< Drop packets if no descriptors > > > > > > > are available. */ > > > > > > > uint8_t rx_deferred_start; /**< Do not start queue with > > > > > > > rte_eth_dev_start(). */ > > > > > > > uint16_t rx_nseg; /**< Number of descriptions in rx_seg > > > > > > > array. > > > > > > > */ > > > > > > > + uint32_t shared_group; /**< Shared port group index in > > > > > > > + switch domain. */ > > > > > > > /** > > > > > > > * Per-queue Rx offloads to be set using DEV_RX_OFFLOAD_* > > > > > > > flags. > > > > > > > * Only offloads set on rx_queue_offload_capa or > > > > > > > rx_offload_capa @@ -1373,6 +1374,12 @@ struct rte_eth_conf { > > > > > > > #define DEV_RX_OFFLOAD_OUTER_UDP_CKSUM 0x00040000 > > > > > > > #define DEV_RX_OFFLOAD_RSS_HASH 0x00080000 > > > > > > > #define RTE_ETH_RX_OFFLOAD_BUFFER_SPLIT 0x00100000 > > > > > > > +/** > > > > > > > + * Rx queue is shared among ports in same switch domain to save > > > > > > > +memory, > > > > > > > + * avoid polling each port. Any port in group can be used to > > > > > > > receive packets. > > > > > > > + * Real source port number saved in mbuf->port field. > > > > > > > + */ > > > > > > > +#define RTE_ETH_RX_OFFLOAD_SHARED_RXQ 0x00200000 > > > > > > > > > > > > > > #define DEV_RX_OFFLOAD_CHECKSUM (DEV_RX_OFFLOAD_IPV4_CKSUM | \ > > > > > > > DEV_RX_OFFLOAD_UDP_CKSUM | \ > > > > > > > -- > > > > > > > 2.25.1 > > > > > > > >
