Re: [PATCH net-next v1 3/5] net: stmmac: support fp parameter of tc-taprio

On Wed, Jul 31, 2024 at 06:43:14PM +0800, Furong Xu wrote:

tc-taprio can select whether traffic classes are express or preemptible.

After some traffic tests, MAC merge layer statistics are all good.

Local device:
ethtool --include-statistics --json --show-mm eth1
[ {
        "ifname": "eth1",
        "pmac-enabled": true,
        "tx-enabled": true,
        "tx-active": true,
        "tx-min-frag-size": 60,
        "rx-min-frag-size": 60,
        "verify-enabled": true,
        "verify-time": 100,
        "max-verify-time": 128,
        "verify-status": "SUCCEEDED",
        "statistics": {
            "MACMergeFrameAssErrorCount": 0,
            "MACMergeFrameSmdErrorCount": 0,
            "MACMergeFrameAssOkCount": 0,
            "MACMergeFragCountRx": 0,
            "MACMergeFragCountTx": 1398,
            "MACMergeHoldCount": 15783

In order for readers to really understand this output (including me),
could you also post the associated tc-taprio command, please?

You deleted the code that treated the Set-And-Hold-MAC GCL command -
and according to 802.1Q, that is the only source of Hold requests.
I _think_ that as a side effect of your reimplementation, every time the
gate for TC 0 opens, the HoldCount bumps by one. Would that be a correct
description?

The more unfortunate part is that I haven't yet come across a NIC
hardware design that would behave completely as you'd expect w.r.t. Hold
requests. In the case of DWMAC, I would expect that with a taprio
schedule that lacks any Set-And-Hold-MAC command, the HoldCount would
stay at zero. I'm not sure, given the way they piggy back onto gate 0
for Hold/Release, that this is possible :(

At least HoldCount stays constant with a tc-mqprio offload, right?

        }
    } ]