The PS5 GbE NIC has no INTx pin and its MSI sometimes never fires after a
cold boot. The driver already works around that with a 1 Hz link-poll
watchdog that kicks NAPI by hand.
But gbe_poll() re-checks GBE_REG_ISR after completing NAPI and reschedules
itself whenever a GBE_IER_MASK cause is still latched, without ever
acknowledging (W1C) the ISR -- only the hardware handler gbe_isr() does.
When the MSI is dead gbe_isr() never runs, so the latched cause (e.g.
RX_DONE) is never cleared, the re-check stays true forever, and NAPI
reschedules itself ~30k times/s, pinning ksoftirqd on that CPU at 100%.
Acknowledge the pending causes before re-kicking so a stale, undelivered
cause can no longer spin NAPI.
While the MSI is dead also poll the rings every GBE_POLL_FALLBACK_MS (10 ms)
instead of once per second, so interactive RX latency drops from ~1 s to a
few ms; fall back to HZ as soon as real IRQs start arriving.
Credits:
- slidybat: shared his base, from which a few fixes were merged in
- cow: endless testing, corrections when I went wrong, catching the
upload-under-load issue, and being an absolute legend
MassZero0andGitHub
Andy Nguyen
rmuxandGitHub
eeplyandGitHub
sleirsgoevyandGitHub