[lpneyon]

Follow-up to #885, as mentioned there.

Linux 7.2 added diag/ attribute groups with error and recovery counters for NVMe controllers, namespace block devices and native-multipath namespace heads. This adds methods to read them on demand:

Method	sysfs source
NVMeDevice.Diag()	/sys/class/nvme/<ctrl>/diag/{command_error_count,reset_count,reconnect_count} (reconnect_count is fabrics-only)
NVMeDevice.NamespaceDiag(ns)	/sys/class/nvme/<ctrl>/<ns>/diag/{command_retries_count,command_error_count,multipath_failover_count}
NVMeSubsystem.NamespaceDiag(ns)	/sys/class/nvme-subsystem/<subsys>/<ns>/diag/{io_requeue_no_usable_path_count,io_fail_no_available_path_count}

The head counters show I/O that was requeued or failed because every path was down, which the per-path counters cannot show.

NVMeClass() and NVMeSubsystemClass() read exactly the same files as before; the diag files are only read when a method is called. NVMeDevice and NVMeSubsystem gain a Path field with their sysfs directory for the methods to read from. Every counter is a pointer and stays nil when the kernel does not provide the attribute, so older kernels and PCIe controllers (no reconnect_count) work unchanged. Namespace names that do not match the NVMe naming pattern are rejected.

Tests: fixtures extended with diag/ attributes for controllers, namespaces and a multipath head; new tests cover each method, a kernel without diag/, and invalid names. Test pointers use new(uint64(...)) following #865.

The node_exporter nvme and nvmesubsystem collector changes that expose these are ready and will follow once this is merged.

[SuperQ]

I'm wondering if we should split this out into a separate function from NVMeClass().

Adding a whole new dataset is going to increase the amount of load caused by all the file reading/parsing when it's less used.

Maybe we could implement this as a functions on NVMeDevice

Something like this:

func (d *NVMeDevice) NamespaceDiag() NVMeNamespaceDiag {
}

[lpneyon]

Thanks for the suggestion. I've reworked it along those lines: NVMeClass() and NVMeSubsystemClass() no longer read any diag/ files, and the counters are read on demand through methods:

func (d NVMeDevice) Diag() (NVMeControllerDiag, error)
func (d NVMeDevice) NamespaceDiag(namespace string) (NVMeNamespaceDiag, error)
func (s NVMeSubsystem) NamespaceDiag(namespace string) (NVMeSubsystemNamespaceDiag, error)

NVMeDevice and NVMeSubsystem now carry a Path field so the methods know where to read. I used value receivers so they can be called directly on NVMeClass map entries, and added an error return since the reads can fail. PR description updated to match.

[SuperQ]

Very nice, I like it.