[lpneyon]

What

Two commits, both adding optional fields to PciDevice. Every field is nil when the kernel does not create the file:

1. Per-device AER counters: AerCorrectable, AerFatal and AerNonFatal, parsed from aer_dev_correctable, aer_dev_fatal and aer_dev_nonfatal under /sys/bus/pci/devices/<BDF>/.
2. Root Port AER totals: AerRootPortTotalErrCor, AerRootPortTotalErrFatal and AerRootPortTotalErrNonFatal, parsed from aer_rootport_total_err_{cor,fatal,nonfatal}. The kernel creates these only on AER-capable Root Ports. They count every error reported to the port by the devices below it.

Why

AER parsing already exists, but it is only reachable through /sys/class/net/<iface>/device. GPUs, NVMe drives and InfiniBand HCAs have no net-class entry, so procfs cannot read their AER counters today. A rising correctable count is often the first sign of a degrading PCIe link, before the device drops off the bus.

Implementation

• The two file parsers in net_class_aer.go now take a file path and report whether the file existed. The existing net-class functions are thin wrappers around them, so their behaviour is unchanged.
• The Root Port totals follow the pattern parsePciDevice already uses for optional files, such as SR-IOV and power management.

If you would rather keep PciDevices() to its current reads and expose AER through methods, as in #887, I am happy to rework it that way.

Testing

• TestPciDevices covers four cases: a device with non-zero AER counters, one with all-zero counters, one with no AER files, and Root Port totals on 0000:00:02.1.
• TestAerCounters is updated, because enp3s0f0 points at the 0000:a2:00.0 fixture, which now has non-zero values.
• go test ./... passes, and golangci-lint v2.14.0 reports 0 issues.
• Tested end to end on Linux 7.3 in QEMU (Ubuntu 26.10 guest, OS-native AER through _OSC). I injected 2 receiver errors, 1 bad TLP and 1 completion timeout into a pcie-root-port with pcie_aer_inject_error. The parsed values matched the kernel's counters exactly: RxErr 2, BadTLP 1 and CmpltTO 1, with Root Port totals of 3 correctable and 1 non-fatal.
• Not yet tested on physical hardware. QEMU's e1000e and igb cannot report AER upstream, because their Device Control error-reporting bits are read-only. So errors reported by an endpoint below a port were not covered.

Ref: Documentation/ABI/testing/sysfs-bus-pci-devices-aer_stats

[SuperQ]

Please do not tag mention maintainers directly. We have a number of maintainers and review as soon as we can. Direct pings are unnecessary and somewhat rude.

[lpneyon]

Sorry, didn't mean to be rude.