Thanks Manish! I am building a test kernel now, and I will let you know once it is ready to test.
If we get good test results, I will submit the patch for SRU to the Ubuntu kernels once the patch has hit mainline. ** Description changed: - With QL41xxx and Ubuntu DNS server DNS failures are seen when updated to - the latest Ubuntu kernel 20.04.1 LTS version 5.4.0-52-generic. Issue was - not observed with 4.5 ubuntu-linux. + BugLink: https://bugs.launchpad.net/bugs/1909062 - Problem Definition: - OS Version: /etc/os-release shows Ubuntu 18.04.4 LTS, but Booted kernel is the latest Ubuntu 20.04.1 LTS version 5.4.0-52-generic - NIC: 2 dual-port (4) QLogic Corp. FastLinQ QL41000 Series 10/25/40/50GbE Controller [1077:8070] (rev 02) - Inbox driver qede v8.37.0.20 + [Impact] - Complete Detailed Problem Description: - Anything that uses the internal Kubernetes DNS server fails. If an external DNS server is used resolution works for non-Kubernetes IPs. + For users with QLogic QL41xxx series NICs, such as the FastLinQ QL41000 + Series 10/25/40/50GbE Controller, when they upgrade from the 4.15 kernel + to the 5.4 kernel, Kubernetes Internal DNS requests will fail, due to + these packets getting corrupted. - Similar issue is described in this article. - https://github.com/kubernetes/kubernetes/issues/95365 + Kubernetes uses IPIP tunnelled packets for internal DNS resolution, and + this particular packet type is not supported for hardware tx checksum + offload, and the packets end up corrupted when the qede driver attempts + to checksum them. - Below patch recently on upstream fixes this - - [Note that issue was introduced by driver's tunnel offload support which was added in after 4.5 kernel] + This only affects internal Kubernetes DNS, as regular DNS lookups to + regular external domains will succeed, due to them not using IPIP packet + types. + + [Fix] + + Marvell has developed a fix for the qede driver, which checks the packet + type, and if it is IPPROTO_IPIP, then csum offloads are disabled for + socket buffers of type IPIP. commit 5d5647dad259bb416fd5d3d87012760386d97530 Author: Manish Chopra <mani...@marvell.com> - Date: Mon Dec 21 06:55:30 2020 -0800 + Date: Mon Dec 21 06:55:30 2020 -0800 + Subject: qede: fix offload for IPIP tunnel packets + Link: https://git.kernel.org/pub/scm/linux/kernel/git/netdev/net.git/commit/?id=5d5647dad259bb416fd5d3d87012760386d97530 - qede: fix offload for IPIP tunnel packets + This commit is currently in the netdev tree, awaiting merge to mainline. + The commit is queued for upstream stable. - IPIP tunnels packets are unknown to device, - hence these packets are incorrectly parsed and - caused the packet corruption, so disable offlods - for such packets at run time. + [Testcase] - Signed-off-by: Manish Chopra <mani...@marvell.com> - Signed-off-by: Sudarsana Kalluru <skall...@marvell.com> - Signed-off-by: Igor Russkikh <irussk...@marvell.com> - Link: https://lore.kernel.org/r/20201221145530.7771-1-mani...@marvell.com - Signed-off-by: Jakub Kicinski <k...@kernel.org> + The system must have a QLogic QL41xxx series NIC fitted, and needs to be + a part of a Kubernetes cluster. - Thanks, - Manish + Firstly, get a list of all devices in the system: + + $ sudo ifconfig + + Next, set all devices down with: + + $ sudo ifconfig <device> down + + Next, bring up the QLogic QL41xxx device: + + $ sudo ifconfig <qlogic nic device> up + + Then, attempt to lookup an internal Kubernetes domain: + + $ nslookup <internal kubernetes domain address> + + Without the patch, the connection will time out: + + ;; connection timed out; no servers could be reached + + If we look at packet traces with tcpdump, we see it leaves the source, + but never arrives at the destination. + + There is a test kernel available in the following ppa: + + https://launchpad.net/~mruffell/+archive/ubuntu/sf297772-test + + If you install it, then Kubernetes internal DNS lookups will succeed. + + [Where problems could occur] + + If a regression were to occur, then users of the qede driver would be + affected. This is limited to those with QLogic QL41xxx series NICs. The + patch explicitly checks for IPIP type packets, so only those particular + packets would be affected. + + Since IPIP type packets are uncommon, it would not cause a total outage + on regression, since most packets are not IPIP tunnelled. It could + potentially cause problems for users who frequently handle VPN or + Kubernetes internal DNS traffic. + + A workaround would be to use ethtool to disable tx csum offload for all + packet types, or to revert to an older kernel. ** Tags added: sts -- You received this bug notification because you are a member of Kernel Packages, which is subscribed to linux in Ubuntu. https://bugs.launchpad.net/bugs/1909062 Title: qede: Kubernetes Internal DNS Failure due to QL41xxx NIC not supporting IPIP tx csum offload Status in linux package in Ubuntu: Confirmed Status in linux source package in Focal: In Progress Status in linux source package in Groovy: In Progress Bug description: BugLink: https://bugs.launchpad.net/bugs/1909062 [Impact] For users with QLogic QL41xxx series NICs, such as the FastLinQ QL41000 Series 10/25/40/50GbE Controller, when they upgrade from the 4.15 kernel to the 5.4 kernel, Kubernetes Internal DNS requests will fail, due to these packets getting corrupted. Kubernetes uses IPIP tunnelled packets for internal DNS resolution, and this particular packet type is not supported for hardware tx checksum offload, and the packets end up corrupted when the qede driver attempts to checksum them. This only affects internal Kubernetes DNS, as regular DNS lookups to regular external domains will succeed, due to them not using IPIP packet types. [Fix] Marvell has developed a fix for the qede driver, which checks the packet type, and if it is IPPROTO_IPIP, then csum offloads are disabled for socket buffers of type IPIP. commit 5d5647dad259bb416fd5d3d87012760386d97530 Author: Manish Chopra <mani...@marvell.com> Date: Mon Dec 21 06:55:30 2020 -0800 Subject: qede: fix offload for IPIP tunnel packets Link: https://git.kernel.org/pub/scm/linux/kernel/git/netdev/net.git/commit/?id=5d5647dad259bb416fd5d3d87012760386d97530 This commit is currently in the netdev tree, awaiting merge to mainline. The commit is queued for upstream stable. [Testcase] The system must have a QLogic QL41xxx series NIC fitted, and needs to be a part of a Kubernetes cluster. Firstly, get a list of all devices in the system: $ sudo ifconfig Next, set all devices down with: $ sudo ifconfig <device> down Next, bring up the QLogic QL41xxx device: $ sudo ifconfig <qlogic nic device> up Then, attempt to lookup an internal Kubernetes domain: $ nslookup <internal kubernetes domain address> Without the patch, the connection will time out: ;; connection timed out; no servers could be reached If we look at packet traces with tcpdump, we see it leaves the source, but never arrives at the destination. There is a test kernel available in the following ppa: https://launchpad.net/~mruffell/+archive/ubuntu/sf297772-test If you install it, then Kubernetes internal DNS lookups will succeed. [Where problems could occur] If a regression were to occur, then users of the qede driver would be affected. This is limited to those with QLogic QL41xxx series NICs. The patch explicitly checks for IPIP type packets, so only those particular packets would be affected. Since IPIP type packets are uncommon, it would not cause a total outage on regression, since most packets are not IPIP tunnelled. It could potentially cause problems for users who frequently handle VPN or Kubernetes internal DNS traffic. A workaround would be to use ethtool to disable tx csum offload for all packet types, or to revert to an older kernel. To manage notifications about this bug go to: https://bugs.launchpad.net/ubuntu/+source/linux/+bug/1909062/+subscriptions -- Mailing list: https://launchpad.net/~kernel-packages Post to : kernel-packages@lists.launchpad.net Unsubscribe : https://launchpad.net/~kernel-packages More help : https://help.launchpad.net/ListHelp
