Previously, system uptime is truncated to seconds, and then the
subtraction from `time.Now()` is performed. Because uptime does not roll
over to the next second at the same instant as `time.Now()`, then
`BootTime()` ends up not being precise, and often varies by 1 second.
This commit does the subtraction before truncating to seconds, which
results in a significantly lower chance of variance in `BootTime()`.
The order of init function execution is dependant on the order that the
source files are passed to the compiler. This causes issues when
building under other build systems, such as bazel or buck, as they are
not guarenteed to maintain the same file order as the default go tool.
When fetching stats on all processes at once there's a non-trivial amount of
time spent in the `BootTime` call. But since this value should never change
during a live process, we can use a cached version for all subsequent calls.
This is mostly intended for Linux, where we are returning the OS version
in the PlatformVersion field, which seems reasonable. Often it is still
useful to know which Linux kernel is running.
For FreeBSD and Darwin the kernel version matches the platform version,
since they previously used the kernel version for the platform version.
For Windows the kernel version is empty, since there is no clear way
to determine it.
On supported hosts the value returned is a UUID (case preserving
from the value of the underlying OS).
For Linux this is generated once, randomly per boot. For FreeBSD and
Darwin this is a more durable value that should persist across reboots.
