Re: [PATCH net v2] net: stmmac: Clear DMA errors for multi-channel interrupts

[Magnus Leatti <magnusleatti@kernelriot.com>]

On 11/10/26 09:30, Paolo Abeni wrote:
> AFAICS 'statistics' here means accounting of some corner cases errors. I
> guess this could be a preferable route.

Hi Paolo,

Went and measured it, since my numbers from this morning only had RBU
masked. YT6801@7.3.3, this time with RBU and RPS both out of
the dwmac4 abnormal mask.

Floods first (4x 64B UDP, 3 runs): 0 MAC irqs and ~3.2M of 9.6M frames
delivered, in line with v2 and with RBU-only. Still 0 with 8x UDP +
TCP TX on top.

Then RPS: each XDP attach/detach restarts RX and raises it once.
MAC irqs per attach/detach, 3 cycles, idle link:

  both masked:   0
  v2:            2
  RBU-only:      1.5k - 25k
  stock:         0.9k - 32k

The fun part: I expected RPS to take the board down the way RBU does,
and it didn't. On stock and RBU-only it's a burst that goes quiet by
itself, and the box stays reachable the whole time. My guess is the RX
handler acks it on the next frame. I did not try an XDP attach in the
middle of a flood...

So both masked is the only one that sits at exactly 0 everywhere I
looked. The price is rx_buf_unav_irq and rx_process_stopped_irq, which
stay at 0.

Scope: I masked them for everyone, like d3265c19b35d did for xgmac2.
Masking only under multi-MSI would give the same result here and would
keep the counters on single-irq platforms.

Usual caveats: one board, and it's a 5.2 core so only the 4.10 mask
ran.

Magnus