[dpdk-dev] [PATCH 0/5] vmxnet3 pmd fixes/improvement

[Yong Wang]

Thomas/Waterman,

I couldn't reproduce the reported issue on v1.8.0-rc1 and both l2fwd and l3fwd works fine using the same command posted.

# dpdk_nic_bind.py --status

Network devices using DPDK-compatible driver
============================================
0000:0b:00.0 'VMXNET3 Ethernet Controller' drv=igb_uio unused=
0000:13:00.0 'VMXNET3 Ethernet Controller' drv=igb_uio unused=

Network devices using kernel driver
===================================
0000:02:00.0 '82545EM Gigabit Ethernet Controller (Copper)' if=eth2 drv=e1000 unused=igb_uio *Active*

Other network devices
=====================
<none>

#  ./build/l3fwd-vf -c 0x6 -n 4 -- -p 0x3 -config "(0,0,1),(1,0,2)"
...
EAL: TSC frequency is ~2800101 KHz
EAL: Master core 1 is ready (tid=ee3c6840)
EAL: Core 2 is ready (tid=de1ff700)
EAL: PCI device 0000:02:00.0 on NUMA socket -1
EAL:   probe driver: 8086:100f rte_em_pmd
EAL:   0000:02:00.0 not managed by UIO driver, skipping
EAL: PCI device 0000:0b:00.0 on NUMA socket -1
EAL:   probe driver: 15ad:7b0 rte_vmxnet3_pmd
EAL:   PCI memory mapped at 0x7f8bee3dd000
EAL:   PCI memory mapped at 0x7f8bee3dc000
EAL:   PCI memory mapped at 0x7f8bee3da000
EAL: PCI device 0000:13:00.0 on NUMA socket -1
EAL:   probe driver: 15ad:7b0 rte_vmxnet3_pmd
EAL:   PCI memory mapped at 0x7f8bee3d9000
EAL:   PCI memory mapped at 0x7f8bee3d8000
EAL:   PCI memory mapped at 0x7f8bee3d6000
Initializing port 0 ... Creating queues: nb_rxq=1 nb_txq=1...  Address:00:0C:29:72:C6:7E, Allocated mbuf pool on socket 0
LPM: Adding route 0x01010100 / 24 (0)
LPM: Adding route 0x02010100 / 24 (1)
LPM: Adding route 0x03010100 / 24 (2)
LPM: Adding route 0x04010100 / 24 (3)
LPM: Adding route 0x05010100 / 24 (4)
LPM: Adding route 0x06010100 / 24 (5)
LPM: Adding route 0x07010100 / 24 (6)
LPM: Adding route 0x08010100 / 24 (7)
txq=0,0,0 
Initializing port 1 ... Creating queues: nb_rxq=1 nb_txq=1...  Address:00:0C:29:72:C6:88, txq=1,0,0 

Initializing rx queues on lcore 1 ... rxq=0,0,0 
Initializing rx queues on lcore 2 ... rxq=1,0,0 
done: Port 0
done: Port 1
L3FWD: entering main loop on lcore 2
L3FWD:  -- lcoreid=2 portid=1 rxqueueid=0
L3FWD: entering main loop on lcore 1
L3FWD:  -- lcoreid=1 portid=0 rxqueueid=0

I don't have the exact setup but I suspect this is related as the errors looks like a tx queue param used is not supported by vmxnet3 backend.  The patchset does not touch the txq config path so it's not clear how it breaks rte_eth_tx_queue_setup().  So my question to Waterman:
(1) Is this a regression on the same branch, i.e. running the unpatched build works but failed with the patch applied?
(2) By any chance did you change the following struct in main.c for those sample programs to a different value, in particular txq_flags?

static const struct rte_eth_txconf tx_conf = {
        .tx_thresh = {
                .pthresh = TX_PTHRESH,
                .hthresh = TX_HTHRESH,
                .wthresh = TX_WTHRESH,
        },
        .tx_free_thresh = 0, /* Use PMD default values */
        .tx_rs_thresh = 0, /* Use PMD default values */
        .txq_flags = (ETH_TXQ_FLAGS_NOMULTSEGS |   <== any changes here?
                      ETH_TXQ_FLAGS_NOVLANOFFL |
                      ETH_TXQ_FLAGS_NOXSUMSCTP |
                      ETH_TXQ_FLAGS_NOXSUMUDP |
                      ETH_TXQ_FLAGS_NOXSUMTCP)
};

Thanks,
Yong
________________________________________
From: Thomas Monjalon <thomas.monjalon at 6wind.com>
Sent: Tuesday, October 28, 2014 7:40 AM
To: Yong Wang
Cc: dev at dpdk.org; Cao, Waterman
Subject: Re: [dpdk-dev] [PATCH 0/5] vmxnet3 pmd fixes/improvement

Hi Yong,

Is there any progress with this patchset?

Thanks
--
Thomas

2014-10-22 07:07, Cao, Waterman:
> Hi Yong,
>
>       We verified your patch with VMWare ESXi 5.5 and found VMware L2fwd and L3fwd cmd can't run.
>     But We use DPDK1.7_rc1 package to validate VMware regression, It works fine.
> .
> 1.[Test Environment]:
>  - VMware ESXi 5.5;
>  - 2 VM
>  - FC20 on Host / FC20-64 on VM
>  - Crown Pass server (E2680 v2 ivy bridge )
>  - Niantic 82599
>
> 2. [Test Topology]:
>       Create 2VMs (Fedora 18, 64bit) .
>     We pass through one physical port(Niantic 82599) to each VM, and also create one virtual device: vmxnet3 in each VM.
>       To connect with two VMs, we use one vswitch to connect two vmxnet3 interface.
>     Then, PF1 and vmxnet3A are in VM1; PF2 and vmxnet3B are in VM2.
>       The traffic flow for l2fwd/l3fwd is as below::
>       Ixia -> PF1 -> vmxnet3A -> vswitch -> vmxnet3B -> PF2 -> Ixia. (traffic generator)
>
> 3.[ Test Step]:
>
> tar dpdk1.8.rc1 ,compile and run;
>
> L2fwd:  ./build/l2fwd -c f -n 4 -- -p 0x3
> L3fwd:  ./build/l3fwd-vf -c 0x6 -n 4 -- -p 0x3 -config "(0,0,1),(1,0,2)"
>
> 4.[Error log]:
>
> ---VMware L2fwd:---
>
> EAL:   0000:0b:00.0 not managed by UIO driver, skipping
> EAL: PCI device 0000:13:00.0 on NUMA socket -1
> EAL:   probe driver: 8086:10fb rte_ixgbe_pmd
> EAL:   PCI memory mapped at 0x7f678ae6e000
> EAL:   PCI memory mapped at 0x7f678af34000
> PMD: eth_ixgbe_dev_init(): MAC: 2, PHY: 17, SFP+: 5
> PMD: eth_ixgbe_dev_init(): port 0 vendorID=0x8086 deviceID=0x10fb
> EAL: PCI device 0000:1b:00.0 on NUMA socket -1
> EAL:   probe driver: 15ad:7b0 rte_vmxnet3_pmd
> EAL:   PCI memory mapped at 0x7f678af33000
> EAL:   PCI memory mapped at 0x7f678af32000
> EAL:   PCI memory mapped at 0x7f678af30000
> Lcore 0: RX port 0
> Lcore 1: RX port 1
> Initializing port 0... PMD: ixgbe_dev_rx_queue_setup(): sw_ring=0x7f670b0f5580 hw_ring=0x7f6789fe5280 dma_addr=0x373e5280
> PMD: ixgbe_dev_rx_queue_setup(): Rx Burst Bulk Alloc Preconditions are satisfied. Rx Burst Bulk Alloc function will be used on port=0, queue=0.
> PMD: ixgbe_dev_rx_queue_setup(): Vector rx enabled, please make sure RX burst size no less than 32.
> PMD: ixgbe_dev_tx_queue_setup(): sw_ring=0x7f670b0f3480 hw_ring=0x7f671b820080 dma_addr=0x100020080
> PMD: ixgbe_dev_tx_queue_setup(): Using simple tx code path
> PMD: ixgbe_dev_tx_queue_setup(): Vector tx enabled.
> done:
> Port 0, MAC address: 90:E2:BA:4A:33:78
>
> Initializing port 1... EAL: Error - exiting with code: 1
>   Cause: rte_eth_tx_queue_setup:err=-22, port=1
>
> ---VMware L3fwd:---
>
> EAL: TSC frequency is ~2793265 KHz
> EAL: Master core 1 is ready (tid=9f49a880)
> EAL: Core 2 is ready (tid=1d7f2700)
> EAL: PCI device 0000:0b:00.0 on NUMA socket -1
> EAL:   probe driver: 15ad:7b0 rte_vmxnet3_pmd
> EAL:   0000:0b:00.0 not managed by UIO driver, skipping
> EAL: PCI device 0000:13:00.0 on NUMA socket -1
> EAL:   probe driver: 8086:10fb rte_ixgbe_pmd
> EAL:   PCI memory mapped at 0x7f079f3e4000
> EAL:   PCI memory mapped at 0x7f079f4aa000
> PMD: eth_ixgbe_dev_init(): MAC: 2, PHY: 17, SFP+: 5
> PMD: eth_ixgbe_dev_init(): port 0 vendorID=0x8086 deviceID=0x10fb
> EAL: PCI device 0000:1b:00.0 on NUMA socket -1
> EAL:   probe driver: 15ad:7b0 rte_vmxnet3_pmd
> EAL:   PCI memory mapped at 0x7f079f4a9000
> EAL:   PCI memory mapped at 0x7f079f4a8000
> EAL:   PCI memory mapped at 0x7f079f4a6000
> Initializing port 0 ... Creating queues: nb_rxq=1 nb_txq=1...  Address:90:E2:BA:4A:33:78, Allocated mbuf pool on socket 0
> LPM: Adding route 0x01010100 / 24 (0)
> LPM: Adding route 0x02010100 / 24 (1)
> LPM: Adding route 0x03010100 / 24 (2)
> LPM: Adding route 0x04010100 / 24 (3)
> LPM: Adding route 0x05010100 / 24 (4)
> LPM: Adding route 0x06010100 / 24 (5)
> LPM: Adding route 0x07010100 / 24 (6)
> LPM: Adding route 0x08010100 / 24 (7)
> txq=0,0,0 PMD: ixgbe_dev_tx_queue_setup(): sw_ring=0x7f071f6f3c80 hw_ring=0x7f079e5e5280 dma_addr=0x373e5280
> PMD: ixgbe_dev_tx_queue_setup(): Using simple tx code path
> PMD: ixgbe_dev_tx_queue_setup(): Vector tx enabled.
>
> Initializing port 1 ... Creating queues: nb_rxq=1 nb_txq=1...  Address:00:0C:29:F0:90:41, txq=1,0,0 EAL: Error - exiting with code: 1
>   Cause: rte_eth_tx_queue_setup: err=-22, port=1
>
>
> Can you help to recheck this patch with latest DPDK code?
>
> Regards
> Waterman
>
> -----Original Message-----
> >From: dev [mailto:dev-bounces at dpdk.org] On Behalf Of Yong Wang
> >Sent: Wednesday, October 22, 2014 6:10 AM
> >To: Patel, Rashmin N; Stephen Hemminger
> >Cc: dev at dpdk.org
> >Subject: Re: [dpdk-dev] [PATCH 0/5] vmxnet3 pmd fixes/improvement
> >
> >Rashmin/Stephen,
> >
> >Since you have worked on vmxnet3 pmd drivers, I wonder if you can help review this set of patches.  Any other reviews/test verifications are welcome of course.  We have reviewed/tested all patches internally.
> >
> >Yong
> >________________________________________
> >From: dev <dev-bounces at dpdk.org> on behalf of Yong Wang <yongwang at vmware.com>
> >Sent: Monday, October 13, 2014 2:00 PM
> >To: Thomas Monjalon
> >Cc: dev at dpdk.org
> >Subject: Re: [dpdk-dev] [PATCH 0/5] vmxnet3 pmd fixes/improvement
> >
> >Only the last one is performance related and it merely tries to give hints to the compiler to hopefully make branch prediction more efficient.  It also moves a constant assignment out of the pkt polling loop.
> >
> >We did performance evaluation on a Nehalem box with 4cores at 2.8GHz x 2 socket:
> >On the DPDK-side, it's running some l3 forwarding apps in a VM on ESXi with one core assigned for polling.  The client side is pktgen/dpdk, pumping 64B tcp packets at line rate.  Before the patch, we are seeing ~900K PPS with 65% cpu of a core used for DPDK.  After the patch, we are seeing the same pkt rate with only 45% of a core used.  CPU usage is collected factoring our the idle loop cost.  The packet rate is a result of the mode we used for vmxnet3 (pure emulation mode running default number of hypervisor contexts).  I can add these info in the review request.
> >
> >Yong
> >________________________________________
> >From: Thomas Monjalon <thomas.monjalon at 6wind.com>
> >Sent: Monday, October 13, 2014 1:29 PM
> >To: Yong Wang
> >Cc: dev at dpdk.org
> >Subject: Re: [dpdk-dev] [PATCH 0/5] vmxnet3 pmd fixes/improvement
> >
> >Hi,
> >
> >2014-10-12 23:23, Yong Wang:
> >> This patch series include various fixes and improvement to the
> >> vmxnet3 pmd driver.
> >>
> >> Yong Wang (5):
> >>   vmxnet3: Fix VLAN Rx stripping
> >>   vmxnet3: Add VLAN Tx offload
> >>   vmxnet3: Fix dev stop/restart bug
> >>   vmxnet3: Add rx pkt check offloads
> >>   vmxnet3: Some perf improvement on the rx path
> >
> >Please, could describe what is the performance gain for these patches?
> >Benchmark numbers would be appreciated.
> >
> >Thanks
> >--
> >Thomas