mirror of
https://git.kernel.org/pub/scm/linux/kernel/git/stable/linux.git
synced 2024-09-30 06:10:56 +00:00
57030a0b62
Even if there is more rx data waiting on the chip, the rx napi poll fn
will never run more than once - it will always read a few buffers, then
bail out and re-arm interrupts. Which results in ping-pong between napi
and interrupt.
This defeats the purpose of napi, and is bad for performance.
Fix by making the rx napi poll behave identically to other ethernet
drivers:
1. initialize rx napi polling with an arbitrary budget (64).
2. in the polling fn, return full weight if rx queue is not depleted,
this tells the napi core to "keep polling".
3. update the rx tail ("ring the doorbell") once for every 8 processed
rx ring buffers.
Thanks to Jakub Kicinski, Eric Dumazet and Andrew Lunn for their expert
opinions and suggestions.
Tested with 20 seconds of full bandwidth receive (iperf3):
rx irqs softirqs(NET_RX)
-----------------------------
before 23827 33620
after 129 4081
Tested-by: Sven Van Asbroeck <thesven73@gmail.com> # lan7430
Fixes:
|
||
|---|---|---|
| .. | ||
| enc28j60.c | ||
| enc28j60_hw.h | ||
| encx24j600-regmap.c | ||
| encx24j600.c | ||
| encx24j600_hw.h | ||
| Kconfig | ||
| lan743x_ethtool.c | ||
| lan743x_ethtool.h | ||
| lan743x_main.c | ||
| lan743x_main.h | ||
| lan743x_ptp.c | ||
| lan743x_ptp.h | ||
| Makefile | ||