collector: add nvmesubsystem collector for NVMe-oF path health

[sradco]

Summary

Add a new disabled-by-default nvmesubsystem collector that reads /sys/class/nvme-subsystem/ to expose NVMe over Fabrics subsystem path health metrics.

Exposed metrics

Metric	Type	Labels	Description
node_nvmesubsystem_info	Gauge (info)	subsystem, nqn, model, serial, iopolicy	Subsystem identity
node_nvmesubsystem_namespace_info	Gauge (info)	subsystem, device	Maps namespace block device to subsystem
node_nvmesubsystem_paths	Gauge	subsystem	Total controller paths
node_nvmesubsystem_paths_live	Gauge	subsystem	Controller paths in live state
node_nvmesubsystem_path_state	Gauge	subsystem, controller, transport, state	Per-controller state

The namespace_info metric enables precise correlation between NVMe block devices (e.g. nvme0n1) and their parent subsystems, which is needed for workload-aware storage path health alerting in Kubernetes environments.

Enable

--collector.nvmesubsystem

No special permissions required — reads only world-readable sysfs attributes.

Dependencies

• sysfs: add NVMeSubsystemClass for NVMe-oF subsystem parsing procfs#797 (NVMe-oF subsystem sysfs parsing with namespace discovery)
• Stacked on collector: add dmmultipath collector for DM-multipath sysfs metrics #3581 (dmmultipath collector, shared test helpers)

Test plan

• Unit tests for metrics emission
• Tests for no-devices case
• Shared test helpers extracted to avoid collision with dmmultipath tests

[sradco]

Hi @SuperQ , I created this PR for a new multipath collector.
I would appreciate your review.

[jsafrane]

The code looks good to me. I tested it with 2 NVMe over TCP devices, I got:

node_nvmesubsystem_info{iopolicy="numa",model="Linux",nqn="tempdisk",serial="bab529a0f32e397e1319",subsystem="nvme-subsys0"} 1
node_nvmesubsystem_path_state{controller="nvme0",state="connecting",subsystem="nvme-subsys0",transport="tcp"} 0
node_nvmesubsystem_path_state{controller="nvme0",state="dead",subsystem="nvme-subsys0",transport="tcp"} 0
node_nvmesubsystem_path_state{controller="nvme0",state="live",subsystem="nvme-subsys0",transport="tcp"} 1
node_nvmesubsystem_path_state{controller="nvme0",state="resetting",subsystem="nvme-subsys0",transport="tcp"} 0
node_nvmesubsystem_path_state{controller="nvme0",state="unknown",subsystem="nvme-subsys0",transport="tcp"} 0
node_nvmesubsystem_path_state{controller="nvme1",state="connecting",subsystem="nvme-subsys0",transport="tcp"} 0
node_nvmesubsystem_path_state{controller="nvme1",state="dead",subsystem="nvme-subsys0",transport="tcp"} 0
node_nvmesubsystem_path_state{controller="nvme1",state="live",subsystem="nvme-subsys0",transport="tcp"} 1
node_nvmesubsystem_path_state{controller="nvme1",state="resetting",subsystem="nvme-subsys0",transport="tcp"} 0
node_nvmesubsystem_path_state{controller="nvme1",state="unknown",subsystem="nvme-subsys0",transport="tcp"} 0
node_nvmesubsystem_paths_live{subsystem="nvme-subsys0"} 2
node_nvmesubsystem_paths_total{subsystem="nvme-subsys0"} 2

Which looks reasonable.

[jsafrane]

for the record, I checked that this is a complete list of all states reported by the kernel today.

[sradco]

PR is failing since its based on prometheus/procfs#797, which is still not merged