Vetting/agent/probes/inventory.go

// Package probes collects hardware facts from a booted Linux system.
//
// Inventory stage consumers call a sequence of probe functions; each
// probe accepts a Logger so it can emit narrative lines (what tool it's
// running, what it found, what it skipped) into the run log. Probes are
// tolerant of missing tools or sysfs gaps — they log a warn line and
// return zero values rather than failing the whole stage.
package probes

import (
	"bufio"
	"fmt"
	"io"
	"os"
	"os/exec"
	"path/filepath"
	"regexp"
	"runtime"
	"strconv"
	"strings"

	"vetting/internal/spec"
)

// Logger is the subset of the agent's logForwarder that probes need.
// Zero-valued fields are safe to call via log* helpers below — missing
// functions are skipped silently so unit tests and older callers (e.g.
// Collect) don't have to wire anything up.
type Logger struct {
	Info func(string)
	Warn func(string)
}

func (l Logger) info(s string) {
	if l.Info != nil {
		l.Info(s)
	}
}

func (l Logger) warn(s string) {
	if l.Warn != nil {
		l.Warn(s)
	}
}

// Collect runs every probe with a no-op logger and returns the merged
// inventory. Kept for callers that only want the data without the
// narrative; the agent's runner calls the per-probe functions directly.
func Collect() (*spec.Inventory, error) {
	return CollectWithLogger(Logger{}), nil
}

func CollectWithLogger(log Logger) *spec.Inventory {
	inv := &spec.Inventory{}
	inv.CPU = CPU(log)
	inv.Memory = Memory(log)
	inv.Disks = Disks(log)
	inv.NICs = NICs(log)
	inv.GPUs = GPUs(log)
	inv.System = System(log)
	inv.Baseboard = Baseboard(log)
	inv.PSU = PSU(log)
	inv.OS = OS(log)
	return inv
}

// ----- CPU --------------------------------------------------------------

// CPU reads /proc/cpuinfo. We pull model, vendor, stepping, physical
// core count (via "cpu cores" — the per-package count is duplicated on
// every logical CPU line, so any occurrence wins), and a filtered flag
// subset relevant to virtualization / encryption gates.
func CPU(log Logger) spec.CPUSpec {
	log.info("  CPU: reading /proc/cpuinfo")
	c := spec.CPUSpec{LogicalCores: runtime.NumCPU()}
	f, err := os.Open("/proc/cpuinfo")
	if err != nil {
		log.warn("  CPU: /proc/cpuinfo unreadable: " + err.Error())
		return c
	}
	defer func() { _ = f.Close() }()
	parseCPUInfoInto(f, &c)
	flagSummary := "none"
	if len(c.Flags) > 0 {
		flagSummary = strings.Join(c.Flags, ",")
	}
	log.info(fmt.Sprintf("  CPU: model=%q vendor=%s cores=%d threads=%d flags=%s",
		c.Model, c.Vendor, c.PhysicalCores, c.LogicalCores, flagSummary))
	return c
}

// cpuFlagAllow is the short-list of /proc/cpuinfo flags we care about —
// everything else is either ubiquitous or too noisy to keep in a
// reportable inventory artifact.
var cpuFlagAllow = map[string]bool{
	"vmx":     true,
	"svm":     true,
	"aes":     true,
	"sse4_2":  true,
	"avx":     true,
	"avx2":    true,
	"avx512f": true,
}

func parseCPUInfoInto(r io.Reader, c *spec.CPUSpec) {
	sc := bufio.NewScanner(r)
	sc.Buffer(make([]byte, 0, 64*1024), 1024*1024)
	for sc.Scan() {
		line := sc.Text()
		k, v, ok := strings.Cut(line, ":")
		if !ok {
			continue
		}
		key := strings.TrimSpace(k)
		val := strings.TrimSpace(v)
		switch key {
		case "model name":
			if c.Model == "" {
				c.Model = val
			}
		case "vendor_id":
			if c.Vendor == "" {
				c.Vendor = val
			}
		case "stepping":
			if c.Stepping == "" {
				c.Stepping = val
			}
		case "cpu cores":
			if c.PhysicalCores == 0 {
				if n, err := strconv.Atoi(val); err == nil {
					c.PhysicalCores = n
				}
			}
		case "flags", "Features":
			if len(c.Flags) == 0 {
				for _, f := range strings.Fields(val) {
					if cpuFlagAllow[f] {
						c.Flags = append(c.Flags, f)
					}
				}
			}
		}
	}
}

// ----- Memory -----------------------------------------------------------

// Memory captures total GiB from /proc/meminfo and per-DIMM rows from
// dmidecode -t 17. When dmidecode is unavailable or returns nothing the
// total still lands via /proc/meminfo so the spec diff for total_gib
// stays meaningful.
func Memory(log Logger) spec.MemorySpec {
	m := spec.MemorySpec{}
	if total, ok := readMemTotalGiB(); ok {
		m.TotalGiB = total
	}
	if _, err := exec.LookPath("dmidecode"); err != nil {
		log.warn("  Memory: dmidecode not installed; per-DIMM skipped")
		log.info(fmt.Sprintf("  Memory: %dGiB total (per-slot unavailable)", m.TotalGiB))
		return m
	}
	log.info("  Memory: dmidecode -t 17")
	out, err := exec.Command("dmidecode", "-t", "17").Output()
	if err != nil {
		log.warn("  Memory: dmidecode -t 17 failed: " + err.Error())
		log.info(fmt.Sprintf("  Memory: %dGiB total (per-slot unavailable)", m.TotalGiB))
		return m
	}
	m.Modules = parseDIMMs(strings.NewReader(string(out)), log)
	populated := 0
	for _, d := range m.Modules {
		if d.Populated {
			populated++
		}
	}
	log.info(fmt.Sprintf("  Memory: %dGiB total, %d of %d slots populated",
		m.TotalGiB, populated, len(m.Modules)))
	return m
}

func readMemTotalGiB() (int, bool) {
	f, err := os.Open("/proc/meminfo")
	if err != nil {
		return 0, false
	}
	defer func() { _ = f.Close() }()
	sc := bufio.NewScanner(f)
	for sc.Scan() {
		fields := strings.Fields(sc.Text())
		if len(fields) >= 2 && fields[0] == "MemTotal:" {
			if kb, err := strconv.ParseInt(fields[1], 10, 64); err == nil {
				return int(kb / 1024 / 1024), true
			}
		}
	}
	return 0, false
}

// parseDIMMs reads dmidecode -t 17 output. Each handle is one "Memory
// Device"; an empty slot reports Size: "No Module Installed". Slot
// locator is Locator (or Bank Locator fallback).
func parseDIMMs(r io.Reader, log Logger) []spec.DIMMSpec {
	blocks := parseDmidecodeAllSections(r, "Memory Device")
	out := make([]spec.DIMMSpec, 0, len(blocks))
	for _, kv := range blocks {
		slot := strings.TrimSpace(kv["Locator"])
		if slot == "" {
			slot = strings.TrimSpace(kv["Bank Locator"])
		}
		d := spec.DIMMSpec{Slot: slot}
		size := strings.TrimSpace(kv["Size"])
		if size == "" || strings.EqualFold(size, "No Module Installed") || strings.EqualFold(size, "Not Installed") {
			log.info("  Memory: slot=" + nonEmpty(slot, "?") + " empty")
			out = append(out, d)
			continue
		}
		d.Populated = true
		d.SizeGB = parseDmidecodeSize(size)
		if speed := strings.TrimSpace(kv["Configured Memory Speed"]); speed != "" {
			d.SpeedMTS = parseMTS(speed)
		}
		if d.SpeedMTS == 0 {
			if speed := strings.TrimSpace(kv["Speed"]); speed != "" {
				d.SpeedMTS = parseMTS(speed)
			}
		}
		d.Manufacturer = cleanDmidecodeValue(kv["Manufacturer"])
		d.PartNumber = cleanDmidecodeValue(kv["Part Number"])
		log.info(fmt.Sprintf("  Memory: slot=%s %dGiB %dMT/s %s PN=%s",
			nonEmpty(slot, "?"), d.SizeGB, d.SpeedMTS, nonEmpty(d.Manufacturer, "?"), nonEmpty(d.PartNumber, "?")))
		out = append(out, d)
	}
	return out
}

// parseDmidecodeSize turns "16 GB" / "8192 MB" / "32 GiB" into GB.
// Treats MB as MiB-ish by dividing by 1024 (dmidecode reports binary
// megabytes even when labeled "MB"). Returns 0 on parse failure.
func parseDmidecodeSize(s string) int {
	parts := strings.Fields(s)
	if len(parts) < 2 {
		return 0
	}
	n, err := strconv.Atoi(parts[0])
	if err != nil {
		return 0
	}
	unit := strings.ToLower(parts[1])
	// dmidecode reports binary megabytes even when the suffix is "MB";
	// accept both "MB" and "MiB" variants. Match on first letter for
	// unit, which covers "GB"/"GiB"/"G" equally well.
	switch unit[0] {
	case 't':
		return n * 1024
	case 'g':
		return n
	case 'm':
		return n / 1024
	case 'k':
		return n / (1024 * 1024)
	}
	return 0
}

// parseMTS pulls the MT/s integer from "3200 MT/s" or "2400 MHz" —
// dmidecode's "Speed" field uses either unit across vendors.
func parseMTS(s string) int {
	parts := strings.Fields(s)
	if len(parts) == 0 {
		return 0
	}
	n, err := strconv.Atoi(parts[0])
	if err != nil {
		return 0
	}
	return n
}

func cleanDmidecodeValue(s string) string {
	s = strings.TrimSpace(s)
	if s == "" {
		return ""
	}
	// dmidecode emits literal strings like "Not Specified" / "Unknown"
	// when the SMBIOS field is blank — suppress them so the inventory
	// doesn't parrot SMBIOS nulls as real values.
	low := strings.ToLower(s)
	if low == "not specified" || low == "unknown" || low == "none" || low == "to be filled by o.e.m." {
		return ""
	}
	return s
}

func nonEmpty(s, fallback string) string {
	if s == "" {
		return fallback
	}
	return s
}

// ----- Disks ------------------------------------------------------------

// Disks walks /sys/class/block and picks out real block devices (no
// partitions, no loop/ram). For each it reads size (512B sectors),
// serial, model, transport (by inspecting the sysfs symlink target),
// and rotational flag.
func Disks(log Logger) []spec.DiskSpec {
	log.info("  Disks: walking /sys/class/block")
	entries, err := os.ReadDir("/sys/class/block")
	if err != nil {
		log.warn("  Disks: /sys/class/block unreadable: " + err.Error())
		return nil
	}
	var out []spec.DiskSpec
	for _, e := range entries {
		name := e.Name()
		if !isRealDisk(name) {
			continue
		}
		base := filepath.Join("/sys/class/block", name)
		size := diskSizeGB(base)
		serial := diskSerial(name)
		model := diskModel(name)
		transport := diskTransport(name)
		rotational := diskRotational(base)
		if size == 0 && serial == "" && model == "" {
			continue
		}
		out = append(out, spec.DiskSpec{
			Serial:     serial,
			SizeGB:     size,
			Model:      model,
			Transport:  transport,
			Rotational: rotational,
		})
		log.info(fmt.Sprintf("  Disks: %s model=%q %dGB %s rotational=%t serial=%s",
			name, model, size, nonEmpty(transport, "?"), rotational, nonEmpty(serial, "?")))
	}
	return out
}

func isRealDisk(name string) bool {
	if strings.HasPrefix(name, "loop") || strings.HasPrefix(name, "ram") ||
		strings.HasPrefix(name, "zram") || strings.HasPrefix(name, "dm-") {
		return false
	}
	partPath := filepath.Join("/sys/class/block", name, "partition")
	if _, err := os.Stat(partPath); err == nil {
		return false
	}
	return true
}

func diskSizeGB(base string) int {
	b, err := os.ReadFile(filepath.Join(base, "size"))
	if err != nil {
		return 0
	}
	sectors, err := strconv.ParseInt(strings.TrimSpace(string(b)), 10, 64)
	if err != nil {
		return 0
	}
	return int(sectors * 512 / 1_000_000_000)
}

func diskSerial(name string) string {
	for _, rel := range []string{
		filepath.Join("/sys/block", name, "device", "serial"),
		filepath.Join("/sys/block", name, "device", "vpd_pg80"),
		filepath.Join("/sys/block", name, "serial"),
	} {
		if b, err := os.ReadFile(rel); err == nil {
			s := sanitizeASCII(string(b))
			if s != "" {
				return s
			}
		}
	}
	cmd := exec.Command("udevadm", "info", "--query=property", "--name="+name)
	out, err := cmd.Output()
	if err != nil {
		return ""
	}
	for _, line := range strings.Split(string(out), "\n") {
		if v, ok := strings.CutPrefix(line, "ID_SERIAL_SHORT="); ok {
			return sanitizeASCII(v)
		}
	}
	return ""
}

func diskModel(name string) string {
	for _, rel := range []string{
		filepath.Join("/sys/block", name, "device", "model"),
		filepath.Join("/sys/block", name, "device", "inquiry"),
	} {
		if b, err := os.ReadFile(rel); err == nil {
			s := sanitizeASCII(string(b))
			if s != "" {
				return s
			}
		}
	}
	return ""
}

// diskTransport inspects /sys/class/block/<name>'s symlink target. The
// path segment after "/devices/" gives the bus: "pci.../nvme/..." →
// nvme, "pci.../ata.../host..." → sata, "pci.../virtio..." → virtio,
// "pci.../host.../target..." (no "ata") → scsi/sas.
func diskTransport(name string) string {
	target, err := os.Readlink(filepath.Join("/sys/class/block", name))
	if err != nil {
		return ""
	}
	t := strings.ToLower(target)
	switch {
	case strings.Contains(t, "/nvme/"):
		return "nvme"
	case strings.Contains(t, "/usb"):
		return "usb"
	case strings.Contains(t, "/virtio"):
		return "virtio"
	case strings.Contains(t, "/ata"):
		return "sata"
	case strings.Contains(t, "/mmc"):
		return "mmc"
	case strings.Contains(t, "/host"):
		return "scsi"
	}
	return ""
}

func diskRotational(base string) bool {
	b, err := os.ReadFile(filepath.Join(base, "queue/rotational"))
	if err != nil {
		return false
	}
	return strings.TrimSpace(string(b)) == "1"
}

// ----- NICs -------------------------------------------------------------

func NICs(log Logger) []spec.NICSpec {
	log.info("  NICs: walking /sys/class/net")
	root := "/sys/class/net"
	entries, err := os.ReadDir(root)
	if err != nil {
		log.warn("  NICs: /sys/class/net unreadable: " + err.Error())
		return nil
	}
	var out []spec.NICSpec
	for _, e := range entries {
		name := e.Name()
		if name == "lo" {
			continue
		}
		base := filepath.Join(root, name)
		mac := readLine(filepath.Join(base, "address"))
		if mac == "" || mac == "00:00:00:00:00:00" {
			continue
		}
		speed := 0
		if b, err := os.ReadFile(filepath.Join(base, "speed")); err == nil {
			if mbps, err := strconv.Atoi(strings.TrimSpace(string(b))); err == nil && mbps > 0 {
				speed = mbps / 1000
			}
		}
		driver, pci := nicDriverAndPCI(base)
		n := spec.NICSpec{
			MAC:       strings.ToLower(mac),
			SpeedGbps: speed,
			Driver:    driver,
			PCIAddr:   pci,
		}
		out = append(out, n)
		suffix := ""
		if speed == 0 {
			suffix = " (no link)"
		}
		log.info(fmt.Sprintf("  NICs: %s mac=%s %dGbps driver=%s pci=%s%s",
			name, n.MAC, speed, nonEmpty(driver, "?"), nonEmpty(pci, "?"), suffix))
	}
	return out
}

// nicDriverAndPCI reads DRIVER= from the device/uevent file and resolves
// the device symlink for the PCI address (last path segment).
func nicDriverAndPCI(base string) (driver, pci string) {
	if b, err := os.ReadFile(filepath.Join(base, "device/uevent")); err == nil {
		for _, line := range strings.Split(string(b), "\n") {
			if v, ok := strings.CutPrefix(line, "DRIVER="); ok {
				driver = strings.TrimSpace(v)
				break
			}
		}
	}
	if target, err := os.Readlink(filepath.Join(base, "device")); err == nil {
		// target looks like "../../../0000:02:00.0" — last element is pci addr.
		pci = filepath.Base(strings.TrimSpace(target))
	}
	return driver, pci
}

// ----- GPUs -------------------------------------------------------------

// GPUs leans on lspci for the base list; for each match it parses the
// PCI address and runs `lspci -vv -s <addr>` to pull VRAM from the
// prefetchable memory region size. Driver comes from the kernel-driver
// line lspci emits when verbose.
func GPUs(log Logger) []spec.GPUSpec {
	if _, err := exec.LookPath("lspci"); err != nil {
		log.warn("  GPUs: lspci not installed; skipped")
		return nil
	}
	log.info("  GPUs: lspci -mm -nn")
	out, err := exec.Command("lspci", "-D", "-mm", "-nn").Output()
	if err != nil {
		log.warn("  GPUs: lspci failed: " + err.Error())
		return nil
	}
	var gpus []spec.GPUSpec
	for _, line := range strings.Split(string(out), "\n") {
		low := strings.ToLower(line)
		if !strings.Contains(low, "vga compatible controller") &&
			!strings.Contains(low, "3d controller") {
			continue
		}
		// With -D the first quoted field is the PCI address.
		fields := splitQuoted(line)
		if len(fields) < 3 {
			continue
		}
		// -D -mm emits the address as the first non-quoted token.
		// splitQuoted strips unquoted whitespace, so the address is
		// indexed at 0 and the quoted class/vendor/device follow.
		// On older lspci the first field is the slot in quotes instead;
		// cover both: if fields[0] doesn't look like an address, strip
		// it from the space-separated prefix.
		addr := ""
		if pciAddrRE.MatchString(fields[0]) {
			addr = fields[0]
		} else if i := strings.IndexByte(line, '"'); i > 0 {
			prefix := strings.TrimSpace(line[:i])
			if pciAddrRE.MatchString(prefix) {
				addr = prefix
			}
		}
		model := stripPCIID(fields[2])
		if !pciAddrRE.MatchString(fields[0]) {
			// With no -D support the device is at [1], model at [2]: same
			// as old code path. Keep existing behavior.
			model = stripPCIID(fields[2])
		}
		model = sanitizeASCII(model)
		if model == "" {
			continue
		}
		g := spec.GPUSpec{Model: model, PCIAddr: addr}
		if addr != "" {
			if vv, err := exec.Command("lspci", "-vv", "-s", addr).Output(); err == nil {
				g.VRAMGiB, g.Driver = parseLspciVerbose(string(vv))
			}
		}
		gpus = append(gpus, g)
		log.info(fmt.Sprintf("  GPUs: %s pci=%s driver=%s vram=%dGiB",
			model, nonEmpty(addr, "?"), nonEmpty(g.Driver, "?"), g.VRAMGiB))
	}
	return gpus
}

var pciAddrRE = regexp.MustCompile(`^[0-9a-fA-F]{4}:[0-9a-fA-F]{2}:[0-9a-fA-F]{2}\.[0-9a-fA-F]$`)

// parseLspciVerbose pulls the largest prefetchable memory region (VRAM)
// in GiB and the "Kernel driver in use" name from an lspci -vv block.
// The memory line looks like: "Memory at ... (64-bit, prefetchable) [size=8G]"
func parseLspciVerbose(s string) (vramGiB int, driver string) {
	sizeRE := regexp.MustCompile(`prefetchable\) \[size=(\d+)([KMGT])`)
	for _, line := range strings.Split(s, "\n") {
		trim := strings.TrimSpace(line)
		if v, ok := strings.CutPrefix(trim, "Kernel driver in use:"); ok {
			driver = strings.TrimSpace(v)
			continue
		}
		if m := sizeRE.FindStringSubmatch(line); m != nil {
			n, _ := strconv.Atoi(m[1])
			switch m[2] {
			case "T":
				n *= 1024
			case "G":
				// already GiB
			case "M":
				n /= 1024
			case "K":
				n /= (1024 * 1024)
			}
			if n > vramGiB {
				vramGiB = n
			}
		}
	}
	return vramGiB, driver
}

// ----- System / Baseboard -----------------------------------------------

func System(log Logger) spec.SystemSpec {
	if _, err := exec.LookPath("dmidecode"); err != nil {
		log.warn("  System: dmidecode not installed; skipped")
		return spec.SystemSpec{}
	}
	log.info("  System: dmidecode -t system")
	out, err := exec.Command("dmidecode", "-t", "system").Output()
	if err != nil {
		log.warn("  System: dmidecode -t system failed: " + err.Error())
		return spec.SystemSpec{}
	}
	kv := parseDmidecodeSection(strings.NewReader(string(out)), "System Information")
	if kv == nil {
		log.warn("  System: no System Information section in dmidecode output")
		return spec.SystemSpec{}
	}
	s := spec.SystemSpec{
		Manufacturer: cleanDmidecodeValue(kv["Manufacturer"]),
		ProductName:  cleanDmidecodeValue(kv["Product Name"]),
		SerialNumber: cleanDmidecodeValue(kv["Serial Number"]),
		UUID:         cleanDmidecodeValue(kv["UUID"]),
	}
	log.info(fmt.Sprintf("  System: manufacturer=%s product=%q serial=%s uuid=%s",
		nonEmpty(s.Manufacturer, "?"), s.ProductName,
		nonEmpty(s.SerialNumber, "?"), nonEmpty(s.UUID, "?")))
	return s
}

func Baseboard(log Logger) spec.BaseboardSpec {
	if _, err := exec.LookPath("dmidecode"); err != nil {
		log.warn("  Baseboard: dmidecode not installed; skipped")
		return spec.BaseboardSpec{}
	}
	log.info("  Baseboard: dmidecode -t baseboard")
	out, err := exec.Command("dmidecode", "-t", "baseboard").Output()
	if err != nil {
		log.warn("  Baseboard: dmidecode -t baseboard failed: " + err.Error())
		return spec.BaseboardSpec{}
	}
	kv := parseDmidecodeSection(strings.NewReader(string(out)), "Base Board Information")
	if kv == nil {
		log.warn("  Baseboard: no Base Board Information section in dmidecode output")
		return spec.BaseboardSpec{}
	}
	b := spec.BaseboardSpec{
		Manufacturer: cleanDmidecodeValue(kv["Manufacturer"]),
		ProductName:  cleanDmidecodeValue(kv["Product Name"]),
		SerialNumber: cleanDmidecodeValue(kv["Serial Number"]),
	}
	log.info(fmt.Sprintf("  Baseboard: manufacturer=%s product=%q serial=%s",
		nonEmpty(b.Manufacturer, "?"), b.ProductName, nonEmpty(b.SerialNumber, "?")))
	return b
}

// ----- PSU --------------------------------------------------------------

func PSU(log Logger) []spec.PowerSupplySpec {
	if _, err := exec.LookPath("dmidecode"); err != nil {
		log.warn("  PSU: dmidecode not installed; skipped")
		return nil
	}
	log.info("  PSU: dmidecode -t 39")
	out, err := exec.Command("dmidecode", "-t", "39").Output()
	if err != nil {
		log.warn("  PSU: dmidecode -t 39 failed: " + err.Error())
		return nil
	}
	blocks := parseDmidecodeAllSections(strings.NewReader(string(out)), "System Power Supply")
	if len(blocks) == 0 {
		log.warn("  PSU: dmidecode -t 39 returned no data (mini-PC or SMBIOS lacks type 39)")
		return nil
	}
	var out2 []spec.PowerSupplySpec
	for _, kv := range blocks {
		p := spec.PowerSupplySpec{
			Slot:         cleanDmidecodeValue(kv["Location"]),
			Manufacturer: cleanDmidecodeValue(kv["Manufacturer"]),
			Model:        cleanDmidecodeValue(kv["Model Part Number"]),
			Status:       cleanDmidecodeValue(kv["Status"]),
		}
		if p.Model == "" {
			p.Model = cleanDmidecodeValue(kv["Name"])
		}
		if watts := cleanDmidecodeValue(kv["Max Power Capacity"]); watts != "" {
			// "760 W" or "760W"
			clean := strings.TrimSuffix(strings.TrimSpace(watts), "W")
			clean = strings.TrimSpace(clean)
			if n, err := strconv.Atoi(strings.Fields(clean)[0]); err == nil {
				p.MaxWatts = n
			}
		}
		out2 = append(out2, p)
		log.info(fmt.Sprintf("  PSU: slot=%s manufacturer=%s model=%q %dW status=%q",
			nonEmpty(p.Slot, "?"), nonEmpty(p.Manufacturer, "?"), p.Model, p.MaxWatts, p.Status))
	}
	return out2
}

// ----- OS ---------------------------------------------------------------

func OS(log Logger) spec.OSSpec {
	o := spec.OSSpec{}
	// /proc/sys/kernel/osrelease is equivalent to `uname -r` and a plain
	// file read — keeps the probe buildable on non-linux dev hosts
	// without syscall.Uname gymnastics. Missing on non-linux; falls back
	// to running `uname -r`.
	if b, err := os.ReadFile("/proc/sys/kernel/osrelease"); err == nil {
		o.Kernel = strings.TrimSpace(string(b))
	} else if out, err := exec.Command("uname", "-r").Output(); err == nil {
		o.Kernel = strings.TrimSpace(string(out))
	}
	if b, err := os.ReadFile("/etc/os-release"); err == nil {
		o.Distribution, o.Version = parseOSRelease(string(b))
	}
	log.info(fmt.Sprintf("  OS: kernel=%s distro=%q version=%s",
		nonEmpty(o.Kernel, "?"), o.Distribution, nonEmpty(o.Version, "?")))
	return o
}

// parseOSRelease returns PRETTY_NAME (distribution) and VERSION_ID
// (version). Quotes are stripped per the spec.
func parseOSRelease(s string) (dist, version string) {
	sc := bufio.NewScanner(strings.NewReader(s))
	for sc.Scan() {
		line := sc.Text()
		k, v, ok := strings.Cut(line, "=")
		if !ok {
			continue
		}
		v = strings.Trim(strings.TrimSpace(v), `"`)
		switch strings.TrimSpace(k) {
		case "PRETTY_NAME":
			dist = v
		case "VERSION_ID":
			version = v
		}
	}
	return dist, version
}

// ----- shared -----------------------------------------------------------

func splitQuoted(line string) []string {
	// Extended variant: splits space-separated tokens outside quotes AND
	// preserves the content inside quotes as single tokens. Needed so
	// `lspci -D -mm -nn` output (leading PCI address token outside any
	// quotes) parses cleanly alongside the quoted class/vendor fields.
	var out []string
	var cur strings.Builder
	inQ := false
	flush := func() {
		if cur.Len() > 0 {
			out = append(out, cur.String())
			cur.Reset()
		}
	}
	for _, r := range line {
		switch {
		case r == '"':
			if inQ {
				out = append(out, cur.String())
				cur.Reset()
			}
			inQ = !inQ
		case r == ' ' && !inQ:
			flush()
		default:
			cur.WriteRune(r)
		}
	}
	flush()
	return out
}

func readLine(path string) string {
	b, err := os.ReadFile(path)
	if err != nil {
		return ""
	}
	return strings.TrimSpace(string(b))
}

func sanitizeASCII(s string) string {
	var b strings.Builder
	b.Grow(len(s))
	for i := 0; i < len(s); i++ {
		c := s[i]
		if c >= 0x20 && c <= 0x7E {
			b.WriteByte(c)
		}
	}
	return strings.TrimSpace(b.String())
}

var pciIDTail = regexp.MustCompile(` *\[[0-9a-fA-F]{4}\]$`)

func stripPCIID(s string) string {
	return pciIDTail.ReplaceAllString(s, "")
}