fix(agent): read /proc/uptime on linux instead of sysinfo(2)

[alecrubin]

Fixes #1948.

host.Uptime() is unix.Sysinfo() in gopsutil, and lxcfs can't intercept a syscall, so every LXC container reports the host's uptime. /proc/uptime is virtualized correctly, it just wasn't being read.

This reads /proc/uptime on linux and falls back to host.Uptime() if the file is missing or unparseable. Other platforms go through a stub that calls host.Uptime() exactly as before, so nothing changes off linux.

Built a small binary that calls both ways and ran it on my Proxmox 9 box (lxcfs 7.0.0-pve1), containers had been up about an hour:

=== PVE host
host.Uptime()  [sysinfo(2), current] :   268189 s
/proc/uptime   [this PR]             :   268188 s

=== inside CT 404
host.Uptime()  [sysinfo(2), current] :   268191 s
/proc/uptime   [this PR]             :     3880 s

=== inside CT 403
host.Uptime()  [sysinfo(2), current] :   268192 s
/proc/uptime   [this PR]             :     3886 s

Same on bare metal, correct in containers.

Tests cover the parse (typical, zero, no trailing newline, single field, large value) and the three fallback paths (missing file, empty file, unparseable). uptimeFilePath is a package var so the tests can point it at a fixture, same as mdraidSysfsRoot does.

Ran go build and go vet for both linux and darwin, and the tests in a golang:1.26 container.

[henrygd]

Thanks a lot, this looks like a good approach.

One change before merge: validate the parsed float before converting it. strconv.ParseFloat accepts NaN, infinities, and values beyond the uint64 range without returning an error, which can yield a huge bogus uptime instead of using the fallback.

seconds, err := strconv.ParseFloat(fields[0], 64)
if err != nil ||
	math.IsNaN(seconds) ||
	math.IsInf(seconds, 0) ||
	seconds < 0 ||
	seconds >= 1<<64 {
	return host.Uptime()
}
return uint64(seconds), nil

And if possible, please add tests covering NaN, +Inf, negative values, and an out-of-range value such as 1e20.

[alecrubin]

Thanks a lot, this looks like a good approach.

One change before merge: validate the parsed float before converting it. strconv.ParseFloat accepts NaN, infinities, and values beyond the uint64 range without returning an error, which can yield a huge bogus uptime instead of using the fallback.

seconds, err := strconv.ParseFloat(fields[0], 64)
if err != nil ||
	math.IsNaN(seconds) ||
	math.IsInf(seconds, 0) ||
	seconds < 0 ||
	seconds >= 1<<64 {
	return host.Uptime()
}
return uint64(seconds), nil

And if possible, please add tests covering NaN, +Inf, negative values, and an out-of-range value such as 1e20.

Thanks for the review. Done in 666b6f6, took your validation as-is. Added fallback tests for NaN, +Inf, -Inf, a negative value, and 1e20. ParseFloat parses all of those without an error so each one hits its own guard instead of the err != nil path. Ran the tests in a golang:1.26 container, all passing.

[henrygd]

Perfect, thank you! Appreciate your work.