Hello,
We have again encountered a bug with the Intel E810-C 100Gx2 cards. The latest issue detected is that they are not truly compatible with Debian 11, as reported on Intel's official site. As I mentioned in previous posts, when trying to compile the Ice drivers (using the `make install` command) on Debian 11, multiple symbolic link errors appear. The same issue occurs on Debian 12.
We have now migrated to Ubuntu 22.04, which Intel also recommends for compatibility, and we have noticed that when installing the Ice drivers, there are virtually no symbolic link errors. Additionally, we have observed improved performance in handling traffic, as we are forwarding traffic of over 30 Gbps with more than 4 million packets per second through the Linux kernel stack.
# lsb_release -a
No LSB modules are available.
Distributor ID: Ubuntu
Description: Ubuntu 22.04.4 LTS
Release: 22.04
Codename: jammy
However, we are now facing a new issue with the Intel E810-C cards and Ice drivers. We use XDP and BPF, and when disabling XDP in some cases (though not always), the drivers freeze. We inject the code in XDPDRV mode. This issue occurs both in Ubuntu and Debian.
ip link set dev "$interface" xdpdrv off
In the Linux kernel log, you can observe the crash (when disabling XDP).
NETDEV WATCHDOG: enp3s0f0 (ice): transmit queue 0 timed out
We investigated a bit and it seems that other users of the same E810-C card and Ice drivers have noticed or experienced the same behavior when using XDP. In fact, there is a recent report from Red Hat where it seems they have also reported this to Intel.
The drivers we are using are the second-to-last ones, published on Intel's official site. The ones on SourceForge probably don't make any changes, as only version 1.14.13 has added support for SLES15sp6. According to the changelogs, no corrections related to XDP have been applied.
driver: ice
version: 1.14.11
The solution (which causes downtime) that we found when detecting this behavior is to disable the Ice drivers and then re-enable them to get the drivers working again:
modprobe -r ice; modprobe ice
This is causing issues with our DoS/DDoS mitigation project, as it is necessary to be able to reload XDP when applying new rules.
Could you assist with this and provide an estimated resolution time? I understand that other members, such as RedHat, have reported something very similar to what we have detected.
We are considering switching network card brands. We have Intel throughout our edges, and with the i40e cards, we have not encountered any similar incidents with XDP.
Photo of the Intel E810-C 100Gx2 network card:
Best regards, and I look forward to hearing from your engineers.
Link Copied
