ng_eiface(4): Set IFF_KNOWSEPOCH flag to avoid enter net epoch recursively.

[afedorov]

Summary

ng_eiface(4) RX path:

ng_eiface_rcvdata() -> ifp->input -> ether_input().

ng_eiface_rcvdata() is always called with NET_EPOCH as the netgraph(4) data path. But ether_input() checks the interface's IFF_KNOWSEPOCH flag to decide whether or not to enter NET_EPOCH. This results in a double entry to NET_EPOCH.

The patch slightly increases RX PPS: ~ 100 Kpps.

Test Plan

Two machines:
First M1:

CPU: Intel(R) Xeon(R) Gold 6226R CPU @ 2.90GHz (2900.00-MHz K8-class CPU)

ixl2: flags=8863<UP,BROADCAST,RUNNING,SIMPLEX,MULTICAST> metric 0 mtu 1500
        options=4e507bb<RXCSUM,TXCSUM,VLAN_MTU,VLAN_HWTAGGING,JUMBO_MTU,VLAN_HWCSUM,TSO4,TSO6,LRO,VLAN_HWFILTER,VLAN_HWTSO,RXCSUM_IPV6,TXCSUM_IPV6,NOMAP>
        ether 3c:ec:ef:3d:77:6a
        media: Ethernet autoselect (10GBase-KR <full-duplex>)
        status: active
        nd6 options=29<PERFORMNUD,IFDISABLED,AUTO_LINKLOCAL>

Second M2:

CPU: Intel(R) Xeon(R) Gold 6226R CPU @ 2.90GHz (2900.00-MHz K8-class CPU)

ixl2: flags=8863<UP,BROADCAST,RUNNING,SIMPLEX,MULTICAST> metric 0 mtu 1500
        options=4e507bb<RXCSUM,TXCSUM,VLAN_MTU,VLAN_HWTAGGING,JUMBO_MTU,VLAN_HWCSUM,TSO4,TSO6,LRO,VLAN_HWFILTER,VLAN_HWTSO,RXCSUM_IPV6,TXCSUM_IPV6,NOMAP>
        ether 3c:ec:ef:3d:77:76
        media: Ethernet autoselect (10GBase-KR <full-duplex>)
        status: active
        nd6 options=29<PERFORMNUD,IFDISABLED,AUTO_LINKLOCAL>

M1:

ifconfig ixl2 up
./pkt-gen -i ixl2 -f tx

M2:

ifconfig ixl2 up
kldload ng_ether
ngctl mkpeer ixl2: eiface lower ether
ifconfig ngeth0 up

Check PPS using systat(1) - M2:

Before patch:

               /0   /1   /2   /3   /4   /5   /6   /7   /8   /9   /10
Load Average   ||||||||||||

 Interface           Traffic               Peak                Total
    ngeth0  in      3.289 Mp/s          3.308 Mp/s          129.722 Mp
            out     0.000 Kp/s          0.000 Kp/s            0.000 Kp

      ixl2  in      15.065 Kp/s         15.263 Mp/s          472.285 Mp
            out     0.000 Kp/s          0.000 Kp/s          256.000 Tp

With patch:

               /0   /1   /2   /3   /4   /5   /6   /7   /8   /9   /10
Load Average   ||||

 Interface           Traffic               Peak                Total
    ngeth0  in      3.489 Mp/s          3.518 Mp/s          250.177 Mp
            out     0.000 Kp/s          0.000 Kp/s            0.000 Kp

      ixl2  in     15.105 Mp/s         15.129 Mp/s         1016.303 Mp
            out     0.000 Kp/s          0.000 Kp/s          256.000 Tp

[zlei]

Generally looks good to me.

The patch slightly increases RX PPS: ~ 100 Kpps.

From to ?

sys/netgraph/ng_eiface.c

415

Nits, no need for the parentheses.

[afedorov]

In D39369#896103, @zlei wrote:

Generally looks good to me.

The patch slightly increases RX PPS: ~ 100 Kpps.

From to ?

I added a test plan ^^^

[zlei]

In D39369#896126, @afedorov wrote:

In D39369#896103, @zlei wrote:

Generally looks good to me.

The patch slightly increases RX PPS: ~ 100 Kpps.

From to ?

I added a test plan ^^^

I meant an absolute increasing (increases RX PPS: ~ 100 Kpps) is not valuable in commit log, or even worse without details of test environment.

Anyway the patch is good and expected to win performance.

[afedorov]

Abandon due to D39510

IFF_KNOWSEPOCH is enabled by default.