W dniu 2022-04-07 o 22:51, Lech Perczak pisze:
Hi Bjørn,
Many thanks you for your review! Answers inline.
W dniu 2022-04-07 o 08:25, Bjørn Mork pisze:
Lech Perczak <lech.perczak@xxxxxxxxx> writes:
+static int zte_rndis_rx_fixup(struct usbnet *dev, struct sk_buff *skb)
+{
+ return rndis_rx_fixup(dev, skb) && usbnet_cdc_zte_rx_fixup(dev,
skb);
+}
Does this work as expected? Only the last ethernet packet in the rndis
frame will end up being handled by usbnet_cdc_zte_rx_fixup(). The
others are cloned and submitted directly to usbnet_skb_return().
I've got some positive reports from at least two owners of the device
- I don't have one myself. In the meantime asked them to run tests
with high traffic, because this should most probably manifest itself
in that scenario easily - my wild guess is that the modem doesn't use
batching, but you are most certainly right in the general case. And
for testing on older modems, we can probably only count on Kristian.
I don't know how to best solve that, but maybe add another
RNDIS_DRIVER_DATA_x flag and test that in rndis_rx_fixup? I.e something
like
bool fixup_dst = dev->driver_info->data &
RNDIS_DRIVER_DATA_FIXUP_DST:
..
/* try to return all the packets in the batch */
skb2 = skb_clone(skb, GFP_ATOMIC);
if (unlikely(!skb2))
break;
skb_pull(skb, msg_len - sizeof *hdr);
skb_trim(skb2, data_len);
if (fixup_dst)
usbnet_cdc_zte_rx_fixup(dev, skb2);
usbnet_skb_return(dev, skb2);
}
if (fixup_dst)
usbnet_cdc_zte_rx_fixup(dev, skb);
/* caller will usbnet_skb_return the remaining packet */
return 1;
}
I'll consider that. My concern with that approach is degradation of
performance by testing for that flag, both for ZTE and non-ZTE
devices, for each and every packet. But this might be the only
solution, as I cannot catch the n-1 sk_buffs from the batch
mid-flight, only the last one. The only other way that currently comes
to my mind, is to duplicate rndis_rx_fixup, with added calls to
usbnet_cdc_zte_rx_fixup in the right places. But the amount of
duplicated code by doing so would be huge, so I'd like to avoid that
as well.
I will definitely send a V2 after I decide on a solution and do some
testing, including high downlink traffic.
Bjørn
Hi Bjørn,
I implemented the fix according to your suggestion and did some testing.
Although I don't have a full MF286R myself, I used my Raspberry Pi
Zero's USB gadget
to simulate the modem's RNDIS interface, and compared three scenarios:
- generic VID/PID with "locally administered" bit off,
- generic VID/PID with "locally administered" bit on,
- ZTE VID/PID (of MF286R's modem) with "locally administered" bit on.
Of course, only the last one activated the MAC fixup path.
For testing I used one of my modem-less MF286A cross-flashed to MF286R
using current
OpenWrt master - which are exactly the same hardware, modulo the
internal modem.
In all three scenarios, when running iperf3 server on the Pi Zero,
I got constant 150Mbps of traffic in both directions, with iperf3 client
running on the
router itself. When router was uploading data to my "modem", CPU usage
was around 66%.
When downloading, the total usage would hit 100%, with about 15%
attributed to syscalls,
and about 85% attributed to softirq.
When using iperf3 client on a PC connected to the router, and enabling
flow offload,
the softirq load would drop to around 75%, and CPU would idle for the
rest of time,
in both directions, but the downlink speed would drop to around 125Mbps,
with upload the
same as if running iperf3 on router itself.
I compared all of this against a build without this patchset, with
scenario one - getting
exactly the same performance.
So, suumming up - it seems my concerns about performance were
exaggerated, so I decided to just
introduce the check inside zte_rndis_fixup(), just as suggested in this
thread. V2 coming shortly.
One strange quirk I noticed while testing, is that when "locally
administered" bit in MAC address
was set, the interface would get "usb" prefix on the host side, and
"eth" otherwise.
--
Pozdrawiam/Kind regards,
Lech Perczak
