fix(input.intel_pmt): Handle telem devices without numa_node attribute #13977
Conversation
When the intel_pmt input plugin is enabled currently, it will cause telegraf to fail on startup like so: ``` telegraf[47878]: 2023-09-21T22:27:09Z E! [telegraf] Error running agent: could not initialize input inputs.intel_pmt: error while exploring pmt sysfs: error while evaluating symlink "/sys/class/intel_pmt/telem0/numa_node": lstat /sys/devices/pci0000:00/0000:00:0a.0/intel_vsec.telemetry.0/intel_pmt/telem0/numa_node: no such file or directory ``` While intel_pmt telem devices do not have a numa_node attribute, their parent intel_vsec devices do. For example the current behavior: ``` $ ls -l /sys/class/intel_pmt/telem0/device/numa_node ls: cannot access '/sys/class/intel_pmt/telem0/device/numa_node': No such file or directory ``` Versus traversing up to intel_vsec device: ``` $ ls -l /sys/class/intel_pmt/telem0/device/../numa_node -rw-r--r--. 1 root root 4096 Sep 18 14:13 /sys/class/intel_pmt/telem0/device/../numa_node ``` Thus update explorePmtInSysfs() to traverse up to the intel_vsec device to find numa_node. Note, filepath.Join() will interpret the `..` ahead of the filepath.EvalSymlinks, thus we evalSymlinks `/sys/class/intel_pmt/telem0/device` and then use filepath.Join() to traverse up. Tested on Fedora 38's 6.4.14-200.fc38.x86_64 kernel and OpenSUSE's 15.4's 5.14.21-150400.24.55-default.
|
Thanks so much for the pull request! |
telegraf-tiger
bot
added
fix
|
!signed-cla |
|
@jakubsikorski it would be good to have you review this if possible. |
|
Download PR build artifacts for linux_amd64.tar.gz, darwin_amd64.tar.gz, and windows_amd64.zip. 👍 This pull request doesn't change the Telegraf binary size 📦 Click here to get additional PR build artifactsArtifact URLs |
There was a problem hiding this comment.
Hi Ben,
It seems different versions of the kernel behave differently here. This depends on the pmt_telemetry/intel_vsec device driver.
If the driver is bus/auxiliary: there won't be a numa_node file and it will fail, just as you described.
If the driver is bus/platform: numa_node file will be present and it will work as expected.
Traversing up the driver will become bus/pci. It could be intel_vsec as you described in newer versions of the kernel or intel_pmt, but it will be a pci driver, which should always have a numa_node file, even if the node is unknown (value is -1 then).
So the changes you propose will indeed fix that.
Thanks for finding this.
powersj
added
the
ready for final review
srebhan
changed the title
When the intel_pmt input plugin is enabled currently, it will cause telegraf to fail on startup like so:
While intel_pmt telem devices do not have a numa_node attribute, their parent intel_vsec devices do. For example the current behavior:
Versus traversing up to intel_vsec device:
Thus update explorePmtInSysfs() to traverse up to the intel_vsec device to find numa_node.
Note, filepath.Join() will interpret the
..ahead of the filepath.EvalSymlinks, thus we evalSymlinks/sys/class/intel_pmt/telem0/deviceand then use filepath.Join() to traverse up.Tested on Fedora 38's 6.4.14-200.fc38.x86_64 kernel and OpenSUSE's 15.4's 5.14.21-150400.24.55-default.
resolves #13976