Use CLOCK_MONOTONIC_RAW instead of CLOCK_MONOTONIC for Instant

[maxbla]

According to man pages, CLOCK_MONOTONIC_RAW doesn't get adjustments from NTP or adjtime, whereas CLOCK_MONOTONIC does. This seems ideal for Instant. CLOCK_MONOTONIC_RAW appears to prevent slewing or jumping, closer matching user expectations about Instant (despite the fact that Instant is explicitly allowed to jump forward and slew)

This wasn't done earlier due to Rust's minimum supported Linux kernel version (now 2.6.32) since CLOCK_MONOTONIC_RAW was added in 2.6.28. See also #37902

I'd also like to change thread::sleep's implementation to use clock_nanosleep(CLOCK_MONOTONIC_RAW) from it's current nanosleep(). Note that

POSIX.1 specifies that nanosleep() should measure time against the CLOCK_REALTIME clock. However, Linux measures the time using the CLOCK_MONOTONIC clock.

from linux.die. This has negative implications for the implementation of thread::sleep on BSD, which should probably use CLOCK_MONOTONIC.

Of course, CLOCK_MONATONIC_RAW is Linux-specific, so other Unix platforms will still use CLOCK_MONATONIC

[sfackler]

The kernel version is not the only reason to use CLOCK_MONOTONIC:

We probably don't want to use CLOCK_MONOTONIC_RAW for Instant since it is not guaranteed to follow real time (e.g. 1 second on CLOCK_MONOTONIC_RAW can be shorter/longer than 1 real second). CLOCK_MONOTONIC is sped up/slowed down by the kernel as necessary to match real time (through NTP).

#37902 (comment)

[maxbla]

The kernel version is not the only reason to use CLOCK_MONOTONIC: ...

Sorry, I managed to not read that part of that issue.

That said, CLOCK_MONOTONIC is not guaranteed to follow real time either. For example, if NTP is significantly different from your hardware clock, it jumps. Also, if you don't use NTP, or adjtime, I suspect the two clocks are equivalent. But I'm not hell-bent on changing it.

I am convinced that thread::sleep should use clock_nanosleep() instead of nanosleep, as this will make the BSDs behave nicer.

[maxbla]

thread::sleep() cannot use clock_nanosleep() because clock_nanosleep is not currently exposed by libc on bsd (see rust-lang/libc#1922). Further, FreeBSD 10 didn't even have clock_nanosleep.