Tuning and Troubleshooting Network Cards

General Issues

mbuf / nmbclusters

To oversimplify, “mbufs” are network memory buffers; portions of RAM set aside for use by networking for moving data around.

The count of active mbufs is shown on the dashboard and is tracked by a graph under Status > Monitoring.

If the mbuf usage is at or near the maximum, or a crash happened with a report that referenced mbufs/uma/memory allocation, then increasing the amount of available mbufs is likely required.

There are two ways to increase mbufs, either by using System Tunables or by using /boot/loader.conf.local

The default value is typically 26584. To start, increase that to 131072. That number can be again be doubled or more as needed, but be careful not to exceed available kernel memory. On 64 bit systems with multiple GB of RAM, set it to 1 million (1000000).

Adding as a System Tunable

To add the value as a tunable:

  • Navigate to System > Advanced, System Tunables tab.

  • Click fa-pencil to edit the entry if kern.ipc.nmbclusters is already in the list

-OR-

  • Click fa-plus to create a new entry if it does not exist.

  • Set the Tunable field to kern.ipc.nmbclusters

  • Set the Value field to 131072 or the desired number.

  • Click Save

  • Click Apply Changes

The value will take effect immediately.

Adding to loader.conf.local

Create or edit /boot/loader.conf.local using a text editor in the shell, or Diagnostics > Edit File. In that file, add the following line:

kern.ipc.nmbclusters="131072"

The new value will not take effect until after a reboot.

PPPoE with Multi-Queue NICs

Network cards which support multiple queues rely on hashing to assign traffic to a particular queue. This works well with IPv4/IPv6 TCP and UDP traffic, for example, but fails with other protocols such as those used for PPPoE.

This can lead to a network card under performing with the default network settings on pfSense® software, as noted on #4821 and FreeBSD PR 203856.

Adding a System Tunable or loader.conf.local entry for net.isr.dispatch=deferred can lead to performance gains on such systems.

Additionally, tuning the values of net.isr.maxthreads and net.isr.numthreads may yield additional performance gains. Generally these are best left at default values matching the number of CPU cores, but depending on the workload may work better at lower values.

Warning

In the past, deferred mode has led to issues on 32-bit platforms, such as crashes/panics, especially with ALTQ. There have been no recent reports, however, so it should be safe on recent versions of pfSense.

TSO/LRO

The settings for Hardware TCP Segmentation Offload (TSO) and Hardware Large Receive Offload (LRO) under System > Advanced on the Networking tab default to checked (disabled) for good reason. Nearly all hardware/drivers have issues with these settings, and they can lead to throughput issues. Ensure the options are checked. Sometimes disabling via sysctl is also necessary.

In /boot/loader.conf.local - Add the following (or create the file if it does not exist):

net.inet.tcp.tso=0

MSI/MSIX

Message Signaled Interrupts are an alternative to classic style Interrupts for retrieving data from hardware. Some cards behave better with MSI, MSIX, or classic style Interrupts, but the card will try the “best” available choice (MSIX, then MSI, then Interrupts).

MSIX and MSI can be disabled via loader tunables. In /boot/loader.conf.local - Add the following (or create the file if it does not exist):

hw.pci.enable_msix=0
hw.pci.enable_msi=0

To nudge the card to use MSI, disable only MSIX. To nudge the card to use regular Interrupts, disable both MSI and MSIX.

IP Input Queue (intr_queue)

This will show the current setting:

sysctl net.inet.ip.intr_queue_maxlen

However, in largely loaded installations this may not be enough. Here is how to check:

sysctl net.inet.ip.intr_queue_drops

If the above shows values above 0, try doubling the current value of net.inet.ip.intr_queue_maxlen.

For example:

sysctl net.inet.ip.intr_queue_maxlen=3000

Keep performing the above until the point is found where drops are eliminated without any adverse effects.

Afterwards, add an entry under System > Advanced, System Tunables tab to set net.inet.ip.intr_queue_maxlen to 3000

Card-Specific Issues

Broadcom bce(4) Cards

Several users have noted issues with certain Broadcom network cards, especially those built into Dell hardware. If the bce cards in the firewall are behaving erratically, dropping packets, or causing system crashes, then the following tweaks may help, especially on amd64.

In /boot/loader.conf.local - Add the following (or create the file if it does not exist):

kern.ipc.nmbclusters="131072"
hw.bce.tso_enable=0
hw.pci.enable_msix=0

That will increase the amount of network memory buffers, disable TSO directly, and disable msix.

Packet loss with many (small) UDP packets

If a lot of packet loss is observed with UDP on bce cards, try changing the netisr settings. These can be set as system tunables under System > Advanced, on the System Tunables tab. On that page, add two new tunables:

net.isr.direct_force=1
net.isr.direct=1

Broadcom bge(4) Cards

See above, but change “bce” to “bge” in the setting names.

Chelsio cxgbe(4) Cards

It is possible to disable the allocation of resources that are not related to the router so that the network adapter can use its entire set of resources for the corresponding functions:

In /boot/loader.conf.local - Add the following (or create the file if it does not exist):

hw.cxgbe.toecaps_allowed=0
hw.cxgbe.rdmacaps_allowed=0
hw.cxgbe.iscsicaps_allowed=0
hw.cxgbe.fcoecaps_allowed=0

Intel igb(4) and em(4) Cards

Certain intel igb cards, especially multi-port cards, can very easily exhaust mbufs and cause kernel panics, especially on amd64. The following tweak will prevent this from being an issue:

In /boot/loader.conf.local - Add the following (or create the file if it does not exist):

kern.ipc.nmbclusters="1000000"

That will increase the amount of network memory buffers, allowing the driver enough headroom for its optimal operation.

Not all NICs and PHYs are the same, even if they share a common driver or chipset. pfSense tries to drive network cards as fast and efficiently as possible, and some hardware combinations are unable to handle the load properly when pushed past their limits, or in certain configurations or network environments. Even if the NICs and drivers claim to support certain features like multiple queues, they may fail in practice when they are used, either due to the hardware or a specific configuration that requires a single queue. In these cases, it may be necessary to reduce the queues to one per card. Accomplish this by placing the following in /boot/loader.conf.local:

hw.igb.num_queues=1

Intel ix(4) Cards

In /boot/loader.conf.local:

kern.ipc.nmbclusters="1000000"
kern.ipc.nmbjumbop="524288"

As a sysctl (system tunable):

hw.intr_storm_threshold=10000

On releases prior to pfSense 2.2, the following may be necessary. If using VLANs with Intel 10 Gb ix(4) cards, some features of the driver for VLANs may need to be disabled to work correctly. For instance, to apply these settings on NIC ix0, run the following:

ifconfig ix0 -vlanhwfilter

That alone should be enough, but in some cases it may be necessary to disable TSO in addition to VLAN hardware filtering, which can be done like so:

ifconfig ix0 -vlanhwfilter -vlanhwtso -tso

These commands may be placed into a shellcmd tag to execute at boot time to make the change persistent.

Flow Control

In some circumstances, flow control may need to be disabled. The exact method of this depends on the hardware involved, but here are a couple examples:

All of these go in /boot/loader.conf.local

cxgbe(4):

hw.cxgbe.pause_settings=0

em(4):

hw.em.fc_setting=0

igb(4):

hw.igb.fc_setting=0

ixgbe(4) (aka ix):

  • Either this (pfSense 2.3+):

    hw.ix.flow_control=0
    
  • Or this (pfSense 2.2.x and before):

    hw.ix.0.fc=0
    

For ix and others, the flow control value can be further tuned:

  • 0: No Flow Control

  • 1: Receive Pause

  • 2: Transmit Pause

  • 3: Full Flow Control, Default