squiid/rust
1
0
Fork 0
Code Issues Pull requests Projects Activity

Auto merge of #88652 - AGSaidi:linux-aarch64-should-be-actually-monotonic, r=yaahc

Browse source 
linux/aarch64 Now() should be actually_monotonic()

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)
for v8.4 systems that don't support FEAT_LSE2 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.

A small sample program using glommio improves by 70x on a 32 core Graviton2
system with this change.
This commit is contained in:
bors 2021-10-17 09:30:30 +00:00
commit 1d6f24210c
2 changed files with 15 additions and 0 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

1
library/std/src/sys/unix/time.rs
View file

@ -303,6 +303,7 @@ mod inner {
pub fn actually_monotonic() -> bool {
(cfg!(target_os = "linux") && cfg!(target_arch = "x86_64"))
|| (cfg!(target_os = "linux") && cfg!(target_arch = "x86"))
|| (cfg!(target_os = "linux") && cfg!(target_arch = "aarch64"))
|| cfg!(target_os = "fuchsia")
}

14
library/std/src/time.rs
View file

@ -268,6 +268,20 @@ impl Instant {
//
// To hopefully mitigate the impact of this, a few platforms are
// excluded as "these at least haven't gone backwards yet".
//
// 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 recieve 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 which cause time to
// not monoticially 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
if time::Instant::actually_monotonic() {
return Instant(os_now);
}