Time in UNIX system calls counts from the epoch, 1970-01-01. The timespec
struct used in various system calls represents this as a number of seconds and
a number of nanoseconds. Nanoseconds are required to be between 0 and
999_999_999, because the portion outside that range should be represented in
the seconds field; if nanoseconds were larger than 999_999_999, the seconds
field should go up instead.
Suppose you ask for the time 1969-12-31, what time is that? On UNIX systems
that support times before the epoch, that's seconds=-86400, one day before the
epoch. But now, suppose you ask for the time 1969-12-31 23:59:00.1. In other
words, a tenth of a second after one minute before the epoch. On most UNIX
systems, that's represented as seconds=-60, nanoseconds=100_000_000. The macOS
bug is that it returns seconds=-59, nanoseconds=-900_000_000.
While that's in some sense an accurate description of the time (59.9 seconds
before the epoch), that violates the invariant of the timespec data structure:
nanoseconds must be between 0 and 999999999. This causes this assertion in the
Rust standard library.
So, on macOS, if we get a Timespec value with seconds less than or equal to
zero, and nanoseconds between -999_999_999 and -1 (inclusive), we can add
1_000_000_000 to the nanoseconds and subtract 1 from the seconds, and then
convert. The resulting timespec value is still accepted by macOS, and when fed
back into the OS, produces the same results. (If you set a file's mtime with
that timestamp, then read it back, you get back the one with negative
nanoseconds again.)
Co-authored-by: Josh Triplett <josh@joshtriplett.org>
Replace `gettimeofday` with `clock_gettime(CLOCK_REALTIME)` on:
```
all(target_os = "macos", not(target_arch = "aarch64")),
target_os = "ios",
target_os = "watchos",
target_os = "tvos"
))]
```
`gettimeofday` was first used in
cc367edd95
which predated the introduction of `clock_gettime` support in macOS
10.12 Sierra which became the minimum supported version in
58bbca958d.
Fix `checked_{add,sub}_duration` incorrectly returning `None` when `other` has more than `i64::MAX` seconds
Use `checked_{add,sub}_unsigned` in `checked_{add,sub}_duration` so that the correct result is returned when adding/subtracting durations with more than `i64::MAX` seconds.
This removes all mutex/atomics based workarounds for non-monotonic clocks and makes the previously panicking methods saturating instead.
Effectively this moves the monotonization from `Instant` construction to the comparisons.
This has some observable effects, especially on platforms without monotonic clocks:
* Incorrectly ordered Instant comparisons no longer panic. This may hide some programming errors until someone actually looks at the resulting `Duration`
* `checked_duration_since` will now return `None` in more cases. Previously it only happened when one compared instants obtained in the wrong order or
manually created ones. Now it also does on backslides.
The upside is reduced complexity and lower overhead of `Instant::now`.
While issues have been seen on arm64 platforms the Arm architecture requires
that the counter monotonically increases and that it must provide a uniform
view of system time (e.g. it must not be possible for a core to receive a
message from another core with a time stamp and observe time going backwards
(ARM DDI 0487G.b D11.1.2). While there have been a few 64bit SoCs that have
bugs (#49281, #56940) which cause time to not monotonically increase, these have
been fixed in the Linux kernel and we shouldn't penalize all Arm SoCs for those
who refuse to update their kernels:
SUN50I_ERRATUM_UNKNOWN1 - Allwinner A64 / Pine A64 - fixed in 5.1
FSL_ERRATUM_A008585 - Freescale LS2080A/LS1043A - fixed in 4.10
HISILICON_ERRATUM_161010101 - Hisilicon 1610 - fixed in 4.11
ARM64_ERRATUM_858921 - Cortex A73 - fixed in 4.12
255a3f3e18 std: Force `Instant::now()` to be monotonic added a mutex to work around
this problem and a small test program using glommio shows the majority of time spent
acquiring and releasing this Mutex. 3914a7b0da tries to improve this, but actually
makes it worse on big systems as for 128b atomics a ldxp/stxp pair (and
successful loop) is required which is expensive as a lock and because of how
the load/store-exclusives scale on large Arm systems is both unfair to threads
and tends to go backwards in performance.
