Qualcomm Fast Path For LEDE


#487

I'm "glad" to hear that. Thanks for reviewing!


#488

With gwlim's version, if I rmmod fast-classifier (only shortcut_fe_cm remains), htop reports double load. If I insmod it back then all goes back to normal. Maybe it helps.
If I'll have time today I test all combination with dissent1's version.


#489

Yes, as I’ve pointed out sfe-cm is a bit faster because it lacks some checks and has more simple logic of offloading rules creation, that may be the reason. Other than that f-c supports bridge offloading.

Well you can’t rmmod the sfe-cm completely, it has leftover rules that it has created for SFE main module. You should not load it at all instead of rmmoding it. The same goes for f-c. You can check modules startup scripts in /etc/modules.d


#490

Oh, well, anyway. :slight_smile:
I compiled fe-cm as well with f-c (with your SFE build), here are the results:
during the 20 mins test: down 8900 KB/s , up 1300 KB/s , connections ~70, htop values:

rmmod fast-classifier (only shortcut-fe-cm)
~92% !!!!!!!
even rmmod shortcut-fe-cm
~92%
insmod shortcut-fe-cm back
~92% !!!!!!!
insmod fast-classifier back
~55%
rmmod shortcut-fe-cm (only fast-classifier)
~92% !!!! :smiley:
rmmod fast-classifier , and insmod fast-classifier back
~55%

Hope it helps!


#491

And I'd like to see your results as well, along with hardware, connection, etc. Even better if your router is supported by gwlim's aug builds, because you can easily compare the 2 (mips optimization doesn't really matter here).

Good thing about SQM is that it puts such a load on our router that is clearly visible.
You need to find out what your initial SQM settings are to get around 90% of sirq usage.
You have to fire up a torrent client with similar settings/usage between tests and looking at CPU % value in htop.
Note that the number of connections also matter not just the up/down speed (higher number of connection puts more load with same speed).


#492

As I’ve said, you can’t rmmod, as it leads to a broken behavior.

When you are running my patchset, sfe doesn’t accelerate ingress flows flagged by QoS. When you are running gwlim’s patchset SQM ingress is not working, because packets are bypassed the QoS stack - SFE grabs it thus not letting qdisc rules to be applied. That could also be the reason.
Another reason - if you are torrenting - there are a lot of small flows appearing and disappearing, but f-c decides to offload only when 128 packets received (you can adjust it by echoing needed value into offload_at_packets).
When using gwlims patchset, basically you have only sfe-cm really working and f-c is hanging around (in the bar probably :slight_smile: )


#493

Can someone quickly help me get the concept of this ?

If i have a Wireless Access Point that uses a bridge interface between its wifi interfaces and the lan.
Will this be accelerated by SFE ?
For now it just displays:
size=0 offload=0 offload_no_match=0 offloaded=0 done=0 offl_dbg_msg_fail=0 done_dbg_msg_fail=0
and
NO_IIF = 2891 CT_DESTROY_MISS = 559 so that does mean it doesnt find the proper device ?

Summary

1: lo: <LOOPBACK,UP,LOWER_UP> mtu 65536 qdisc noqueue state UNKNOWN qlen 1
link/loopback 00:00:00:00:00:00 brd 00:00:00:00:00:00
inet 127.0.0.1/8 scope host lo
valid_lft forever preferred_lft forever
inet6 ::1/128 scope host
valid_lft forever preferred_lft forever
2: eth0: <BROADCAST,MULTICAST,UP,LOWER_UP> mtu 1500 qdisc fq_codel state UP qlen 1000
link/ether 80:2a:a8:c0:24:78 brd ff:ff:ff:ff:ff:ff
inet6 fe80::822a:a8ff:fec0:2478/64 scope link
valid_lft forever preferred_lft forever
3: ifb0: <BROADCAST,NOARP> mtu 1500 qdisc noop state DOWN qlen 32
link/ether f6:c1:51:b9:9f:e1 brd ff:ff:ff:ff:ff:ff
4: ifb1: <BROADCAST,NOARP> mtu 1500 qdisc noop state DOWN qlen 32
link/ether 0e:f8:51:22:50:e1 brd ff:ff:ff:ff:ff:ff
7: br-lan: <BROADCAST,MULTICAST,UP,LOWER_UP> mtu 1500 qdisc noqueue state UP qlen 1000
link/ether 80:2a:a8:c0:24:78 brd ff:ff:ff:ff:ff:ff
inet 10.0.0.5/24 brd 10.0.0.255 scope global br-lan
valid_lft forever preferred_lft forever
8: eth0.1@eth0: <BROADCAST,MULTICAST,UP,LOWER_UP> mtu 1500 qdisc noqueue master br-lan state UP qlen 1000
link/ether 80:2a:a8:c0:24:78 brd ff:ff:ff:ff:ff:ff
9: eth0.2@eth0: <BROADCAST,MULTICAST,UP,LOWER_UP> mtu 1500 qdisc noqueue state UP qlen 1000
link/ether 80:2a:a8:c0:24:78 brd ff:ff:ff:ff:ff:ff
inet6 fe80::822a:a8ff:fec0:2478/64 scope link
valid_lft forever preferred_lft forever
10: wlan0: <BROADCAST,MULTICAST,UP,LOWER_UP> mtu 1500 qdisc mq master br-lan state UP qlen 1000
link/ether 80:2a:a8:c2:24:78 brd ff:ff:ff:ff:ff:ff
inet6 fe80::822a:a8ff:fec2:2478/64 scope link
valid_lft forever preferred_lft forever
11: wlan1: <BROADCAST,MULTICAST,UP,LOWER_UP> mtu 1500 qdisc noqueue master br-lan state UP qlen 1000
link/ether 80:2a:a8:c1:24:78 brd ff:ff:ff:ff:ff:ff
inet6 fe80::822a:a8ff:fec1:2478/64 scope link
valid_lft forever preferred_lft forever
12: wlan1-1: <BROADCAST,MULTICAST,UP,LOWER_UP> mtu 1500 qdisc noqueue master br-lan state UP qlen 1000
link/ether 82:2a:a8:c1:24:78 brd ff:ff:ff:ff:ff:ff
inet6 fe80::802a:a8ff:fec1:2478/64 scope link
valid_lft forever preferred_lft forever

Also im getting the following output on my USG (unifi security gateway)
size=70 offload=0 offload_no_match=0 offloaded=2636 done=2488 offl_dbg_msg_fail=2636 done_dbg_msg_fail=2488

The two identical counts (2636) worry me a little bit.

CT_NO_CONFIRM = 54749
TCP_NOT_ASSURED = 49271
TCP_NOT_ESTABLISHED = 926
UNKNOW_PROTOCOL = 5
NO_SRC_DEV = 1800
NO_DEST_DEV = 1173803
WAIT_FOR_ACCELERATION = 996227
UPDATE_PROTOCOL_FAIL = 926
CT_DESTROY_MISS = 118532

#494

You need to echo 1 > /sys/fast_classifier/skip_to_bridge_ingress

That’s ok, the 2nd value is that the debug message that the connection has been offloaded had failed to be sent to a debug daemon


#495

Hey thanks for that so it should start working if i set that or wont it work at all ?


#496

Yes that symbol enables incoming traffic on bridge interfaces to be offloaded


#497

I keep getting NO_IIF errors and its still at 0 i must be doing sth wrong.

Could that be related to the interface being a switch with vlans for both ports ?


#498

No_iif error also include locally generated packets and bridge egress, so it’s not a problem.
Bridge connections are not tracked, so it should not appear in those debug statistics. You can try measuring sirq. As far as I get it, bridge is already fast, so there won’t be such a noticeable gain.


#499

A son of that WRT54G v2.0, WRT54GL v1.1 powered with Tomato RAF is able to NAT ~60 Mbit/s! :smiley:
I don't know RAF firmware is available to your WRT.


#500

I am using gwlim's 32MB ram build for TP-Link 1043NDv1 and I have noticed a interesting WiFi behaviour when setting up several (in my scenario - 4) virtual Access Point SSIDs.
WiFi speed is different when connected the main AP SSID name (the one that is first on the list and has more configuration options available) and the other ones.The first is up to 2x faster then the other virtual APs.
Is it possible that the patch is not working on virtual Access Points but only on the main one?
Has anyone else tested similar scenarios with virtual AP on the same radio and encountered wifi speed differences?

Thanks in advance for your help.


#501

Using this on mvebu wrt3200acm would speed up things?


#502

Hello i have install your firmware in my tp link wr1043ndv3....when i go to 192.168.1.1 it says that no password is set...when i click in "bold" to set password it doesn't redirects me to set password...just refresh the page....i reboot my pc...router...nothing change.
What is user and password????
To login and change them after....i install this.


#503

This is the trickiest part!

Just simply login with empty password field ... and set password after login!

User is root, the password field should be left empty.


#504

Yes....i forgot it that user is root :sleeping:
I dont think that with SQM enable in download part doesn't work...in upload part it works.
I am missing something for download?
Also my ping decreased from 20ms to 12ms!!!


#505

The CPU cores are fast enough to route at 1Gbps without software help.


#506

Sorry for the late reply, I needed time to test things.

Thanks for the confirmation.

That's the reason behind such a performance diff.
It turned out during stress test (6 download, ~250 connection) that SQM ingress doesn't work with gwlim build (I'll update my latest post(s)), it completely ate up all the bandwidth:

  • not with vpn
  • nor without vpn

Interesting. How can I determine (without just trying) what is the required value.

I just put back the build that includes your patch (everything is the same apart from SFE patches) and I'll use it in regular basis. I'll report back after a week if all goes well.
Thanks