PPPoE WANs do not take full advantage of NIC driver queues for receiving traffic
On PPPoE WANs packets are only received on one NIC driver queue (queue0) while packets are transmitted from all queues (queue0 and queue1). This has been observed on multiple systems with PPPoE-based WANs and igb(4) NICs, though it may also affect L2TP and PPTP type WANs since they all use mpd.
On my WAN (igb1 with PPPoE on top):
dev.igb.1.queue0.tx_packets: 2535085 dev.igb.1.queue0.rx_packets: 5365670 dev.igb.1.queue1.tx_packets: 2711996 dev.igb.1.queue1.rx_packets: 0
Other non-PPPoE interfaces on the same unit, including additional WANs, show activity in all driver queues.
On another, different, firewall with igb(4) NICs and PPPoE, the same condition is present:
dev.igb.0.queue0.tx_packets: 8504662 dev.igb.0.queue0.rx_packets: 29342831 dev.igb.0.queue1.tx_packets: 3543617 dev.igb.0.queue1.rx_packets: 0
On lower speed WANs there are no negative effects, but on higher-speed PPPoE WANs such as gigabit links it can cause some disparity where traffic is transmitted at the expected rate but not received at the expected rate.
How to check (on systems with igb NICs):
sysctl -a | grep '\.igb\..*x_pack'
Look for the PPPoE WAN physical interface and check if has activity in the tx queue1 and rx queue1, or only the tx queue1 (or higher queues)
Currently only tested on pfSense 2.2.3-RELEASE amd64 images, needs more feedback/testing to confirm if it happens on additional versions/architectures.
#1 Updated by Steve Wheeler almost 2 years ago
Seems likely to be this:
"Unfortunately, RSS is usually capable of hashing IPv4 and IPv4 traffic (L3+L4). All other traffic like PPPoE or MPLS or .. is usually received by queue 0."
There is a patch suggested there.
#3 Updated by Jim Thompson over 1 year ago
- Assignee set to Jim Thompson
- Priority changed from Normal to Low
- Target version changed from 2.3 to Future
Fixing this likely requires an in-kernel RSS (Toeplitz) implementation. Such a thing is coming for FreeBSD (Adrian is working on it for the upper layers of the stack), but it's going to be a while before it's ready to interface to netisr.
Priority dropped to "low". Will review when we're based on 11.
#7 Updated by Jim Pingle 4 months ago
- Target version changed from Future to 2.4.0
It still happens on 2.4, actually it's a little worse since it doesn't appear to transmit on the additional queues like it did previously:
dev.igb.1.queue1.rx_packets: 0 dev.igb.1.queue1.tx_packets: 0 dev.igb.1.queue0.rx_packets: 1978785 dev.igb.1.queue0.tx_packets: 1959503
Earlier note on the ticket said to review once we're on FreeBSD 11, so I'll set the target to 2.4 but it may need to be pushed again depending on what we find.
#9 Updated by Vladimir Putin 4 months ago
According to this
Script that can solve CPU 1-core overload problem, but not default queue problem.
#!/bin/sh # PROVIDE: cpuset-igb # REQUIRE: FILESYSTEMS # BEFORE: netif # KEYWORD: nojail case "$1" in *start) echo "Binding igb(4) IRQs to CPUs" cpus=`sysctl -n kern.smp.cpus` vmstat -ai | sed -E '/^irq.*que/!d; s/^irq([0-9]+): igb([0-9]+):que ([0-9]+).*/\1 \2 \3/' |\ while read irq igb que do cpuset -l $(( ($igb+$que) % $cpus )) -x $irq done ;; esac
#12 Updated by J P about 2 months ago
This should definitely be marked as a bug and not a feature.
Has anybody tried the igb driver patch from https://wiki.freebsd.org/NetworkPerformanceTuning ?
The link is dead and I can't find any archived copies. That, in combination with the script above, should at least fix this temporarily for users of NICs that use the igb driver.
I found a forum post (https://forum.pfsense.org/index.php?topic=114123.0) that shows someone using the igb driver that is reporting all queues are used with PPPoE. Can someone test if this actually still happens?
Here is the FreeBSD bug tracker pointing out this issue https://bugs.freebsd.org/bugzilla/show_bug.cgi?id=203856
The issue is over a year old (October 2015) and hasn't seen much progress.
#13 Updated by David Wood about 2 months ago
The problem with the patch mentioned in comment 12 was that it was a kludge for igb(4) only, not a fix for the underlying issue which, so far as I remember, affects PPPoE on all multi-queue NICs. I'm not even sure the kludge will apply cleanly on FreeBSD 11.x, let alone work, as there are new in-kernel RSS features and various fixes/updates to the igb(4) driver. I suspect the patch has disappeared from a combination of failing to be a universal solution, lack of support and code rot.
I'm fairly certain that the pfSense team will take the approach of fixing this issue properly for all scenarios or doing nothing. An unsupported patch for a single NIC driver is unlikely to qualify for inclusion in a production firewall distribution.
Annoying as this issue undoubtedly is, I expect it will miss the cut for pfSense 2.4-RELEASE. There comes a point where developers have to decide an upcoming release is feature complete in order to close out the remaining bugs and ship.