New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Replace pthread RwLock
with custom implementation
#110211
Conversation
Hey! It looks like you've submitted a new PR for the library teams! If this PR contains changes to any Examples of
|
☔ The latest upstream changes (presumably #110324) made this pull request unmergeable. Please resolve the merge conflicts. |
I've found the bottleneck! Because the The implementation now uses a separate |
Few functions here with boolean args looks like have compile time known arguments values (i.e. true/false), so probably optimizer can see it and inline it's values. Maybe try to make this arguments const generics instead and see if this will work better? |
At least locally, the performance doesn't change at all (that makes sense, the condition in |
☔ The latest upstream changes (presumably #118996) made this pull request unmergeable. Please resolve the merge conflicts. |
☔ The latest upstream changes (presumably #117285) made this pull request unmergeable. Please resolve the merge conflicts. |
☔ The latest upstream changes (presumably #120496) made this pull request unmergeable. Please resolve the merge conflicts. |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
I finally had the time to review this. It looks great! Simpler and a lot more readable than I was expecting. ^^
Just a few small comments:
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Looks great to me!
Since this is a bunch of new unsafe+atomics code, I'd feel slightly more comfortable if another reviewer took a look at it too: @Amanieu, would you have time to also take a look at the soundness of this lock implementation?
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
r=me with the small nit addressed
@bors r+ |
I recall for mutex we stuck to pthreads on some platforms because that gives the scheduler more information and it can apply priority boosting when a high-priority thread wants to take a lock from a low-priority thread (specifically, on macOS). Do similar concerns not apply here? |
I was just thinking about the same thing, but I think PI is not an issue here. For higher-priority readers, it can't practically be supported, as you'd have to memorize all current readers. For higher-priority writers, while it's possible to support PI in principle, it appears that no platform actually does so (macOS doesn't (I think), NetBSD doesn't, QNX doesn't). I can't find any at least, and the POSIX standard does not mandate or facilitate PI for rwlocks, on the contrary, it warns against priority inversions. |
☀️ Test successful - checks-actions |
Replace pthread `RwLock` with custom implementation This is one of the last items in #93740. I'm doing `RwLock` first because it is more self-contained and has less tradeoffs to make. The motivation is explained in the documentation, but in short: the pthread rwlock is slow and buggy and `std` can do much better. I considered implementing a parking lot, as was discussed in the tracking issue, but settled for the queue-based version because writing self-balancing binary trees is not fun in Rust... This is a rather complex change, so I have added quite a bit of documentation to help explain it. Please point out any part that could be explained better. ~~The read performance is really good, I'm getting 4x the throughput of the pthread version and about the same performance as usync/parking_lot on an Apple M1 Max in the usync benchmark suite, but the write performance still falls way behind what usync and parking_lot achieve. I tried using a separate queue lock like what usync uses, but that didn't help. I'll try to investigate further in the future, but I wanted to get some eyes on this first.~~ [Resolved](rust-lang/rust#110211 (comment)) r? `@m-ou-se` CC `@kprotty`
Finished benchmarking commit (b17491c): comparison URL. Overall result: ✅ improvements - no action needed@rustbot label: -perf-regression Instruction countThis is a highly reliable metric that was used to determine the overall result at the top of this comment.
Max RSS (memory usage)This benchmark run did not return any relevant results for this metric. CyclesThis benchmark run did not return any relevant results for this metric. Binary sizeThis benchmark run did not return any relevant results for this metric. Bootstrap: 664.937s -> 663.718s (-0.18%) |
So now that we have this, is it even worth still also having the Futex-based rwlock for Linux? |
Probably yes? I don't have a Linux machine to run benchmarks on, but my guess would be that there is little difference in performance between the implementations. The Linux version is a lot simpler, so I don't think there is any harm in keeping it. The true benefit of this version is that we can finally get rid of the unsound lock code on SGX and provide a fast and well-tested (heh!) lock for all the weird platforms out there (e.g. xous just uses spinning right now). |
Replace pthread `RwLock` with custom implementation This is one of the last items in #93740. I'm doing `RwLock` first because it is more self-contained and has less tradeoffs to make. The motivation is explained in the documentation, but in short: the pthread rwlock is slow and buggy and `std` can do much better. I considered implementing a parking lot, as was discussed in the tracking issue, but settled for the queue-based version because writing self-balancing binary trees is not fun in Rust... This is a rather complex change, so I have added quite a bit of documentation to help explain it. Please point out any part that could be explained better. ~~The read performance is really good, I'm getting 4x the throughput of the pthread version and about the same performance as usync/parking_lot on an Apple M1 Max in the usync benchmark suite, but the write performance still falls way behind what usync and parking_lot achieve. I tried using a separate queue lock like what usync uses, but that didn't help. I'll try to investigate further in the future, but I wanted to get some eyes on this first.~~ [Resolved](rust-lang/rust#110211 (comment)) r? `@m-ou-se` CC `@kprotty`
Use queue-based `RwLock` on more platforms This switches over Windows 7, SGX and Xous to the queue-based `RwLock` implementation added in rust-lang#110211, thereby fixing rust-lang#121949 for Windows 7 and partially resolving rust-lang#114581 on SGX. TEEOS can't currently be switched because it doesn't have a good thread parking implementation. CC `@roblabla` `@raoulstrackx` `@xobs` Could you help me test this, please? r? `@ChrisDenton` the Windows stuff should be familiar to you
Rollup merge of rust-lang#123811 - joboet:queue_em_up, r=ChrisDenton Use queue-based `RwLock` on more platforms This switches over Windows 7, SGX and Xous to the queue-based `RwLock` implementation added in rust-lang#110211, thereby fixing rust-lang#121949 for Windows 7 and partially resolving rust-lang#114581 on SGX. TEEOS can't currently be switched because it doesn't have a good thread parking implementation. CC `@roblabla` `@raoulstrackx` `@xobs` Could you help me test this, please? r? `@ChrisDenton` the Windows stuff should be familiar to you
Replace pthread `RwLock` with custom implementation This is one of the last items in #93740. I'm doing `RwLock` first because it is more self-contained and has less tradeoffs to make. The motivation is explained in the documentation, but in short: the pthread rwlock is slow and buggy and `std` can do much better. I considered implementing a parking lot, as was discussed in the tracking issue, but settled for the queue-based version because writing self-balancing binary trees is not fun in Rust... This is a rather complex change, so I have added quite a bit of documentation to help explain it. Please point out any part that could be explained better. ~~The read performance is really good, I'm getting 4x the throughput of the pthread version and about the same performance as usync/parking_lot on an Apple M1 Max in the usync benchmark suite, but the write performance still falls way behind what usync and parking_lot achieve. I tried using a separate queue lock like what usync uses, but that didn't help. I'll try to investigate further in the future, but I wanted to get some eyes on this first.~~ [Resolved](rust-lang/rust#110211 (comment)) r? `@m-ou-se` CC `@kprotty`
Pkgsrc changes: * Adapt checksums and patches, some have beene intregrated upstream. Upstream chnages: Version 1.78.0 (2024-05-02) =========================== Language -------- - [Stabilize `#[cfg(target_abi = ...)]`] (rust-lang/rust#119590) - [Stabilize the `#[diagnostic]` namespace and `#[diagnostic::on_unimplemented]` attribute] (rust-lang/rust#119888) - [Make async-fn-in-trait implementable with concrete signatures] (rust-lang/rust#120103) - [Make matching on NaN a hard error, and remove the rest of `illegal_floating_point_literal_pattern`] (rust-lang/rust#116284) - [static mut: allow mutable reference to arbitrary types, not just slices and arrays] (rust-lang/rust#117614) - [Extend `invalid_reference_casting` to include references casting to bigger memory layout] (rust-lang/rust#118983) - [Add `non_contiguous_range_endpoints` lint for singleton gaps after exclusive ranges] (rust-lang/rust#118879) - [Add `wasm_c_abi` lint for use of older wasm-bindgen versions] (rust-lang/rust#117918) This lint currently only works when using Cargo. - [Update `indirect_structural_match` and `pointer_structural_match` lints to match RFC] (rust-lang/rust#120423) - [Make non-`PartialEq`-typed consts as patterns a hard error] (rust-lang/rust#120805) - [Split `refining_impl_trait` lint into `_reachable`, `_internal` variants] (rust-lang/rust#121720) - [Remove unnecessary type inference when using associated types inside of higher ranked `where`-bounds] (rust-lang/rust#119849) - [Weaken eager detection of cyclic types during type inference] (rust-lang/rust#119989) - [`trait Trait: Auto {}`: allow upcasting from `dyn Trait` to `dyn Auto`] (rust-lang/rust#119338) Compiler -------- - [Made `INVALID_DOC_ATTRIBUTES` lint deny by default] (rust-lang/rust#111505) - [Increase accuracy of redundant `use` checking] (rust-lang/rust#117772) - [Suggest moving definition if non-found macro_rules! is defined later] (rust-lang/rust#121130) - [Lower transmutes from int to pointer type as gep on null] (rust-lang/rust#121282) Target changes: - [Windows tier 1 targets now require at least Windows 10] (rust-lang/rust#115141) - [Enable CMPXCHG16B, SSE3, SAHF/LAHF and 128-bit Atomics in tier 1 Windows] (rust-lang/rust#120820) - [Add `wasm32-wasip1` tier 2 (without host tools) target] (rust-lang/rust#120468) - [Add `wasm32-wasip2` tier 3 target] (rust-lang/rust#119616) - [Rename `wasm32-wasi-preview1-threads` to `wasm32-wasip1-threads`] (rust-lang/rust#122170) - [Add `arm64ec-pc-windows-msvc` tier 3 target] (rust-lang/rust#119199) - [Add `armv8r-none-eabihf` tier 3 target for the Cortex-R52] (rust-lang/rust#110482) - [Add `loongarch64-unknown-linux-musl` tier 3 target] (rust-lang/rust#121832) Refer to Rust's [platform support page][platform-support-doc] for more information on Rust's tiered platform support. Libraries --------- - [Bump Unicode to version 15.1.0, regenerate tables] (rust-lang/rust#120777) - [Make align_offset, align_to well-behaved in all cases] (rust-lang/rust#121201) - [PartialEq, PartialOrd: document expectations for transitive chains] (rust-lang/rust#115386) - [Optimize away poison guards when std is built with panic=abort] (rust-lang/rust#100603) - [Replace pthread `RwLock` with custom implementation] (rust-lang/rust#110211) - [Implement unwind safety for Condvar on all platforms] (rust-lang/rust#121768) - [Add ASCII fast-path for `char::is_grapheme_extended`] (rust-lang/rust#121138) Stabilized APIs --------------- - [`impl Read for &Stdin`] (https://doc.rust-lang.org/stable/std/io/struct.Stdin.html#impl-Read-for-%26Stdin) - [Accept non `'static` lifetimes for several `std::error::Error` related implementations] (rust-lang/rust#113833) - [Make `impl<Fd: AsFd>` impl take `?Sized`] (rust-lang/rust#114655) - [`impl From<TryReserveError> for io::Error`] (https://doc.rust-lang.org/stable/std/io/struct.Error.html#impl-From%3CTryReserveError%3E-for-Error) These APIs are now stable in const contexts: - [`Barrier::new()`] (https://doc.rust-lang.org/stable/std/sync/struct.Barrier.html#method.new) Cargo ----- - [Stabilize lockfile v4](rust-lang/cargo#12852) - [Respect `rust-version` when generating lockfile] (rust-lang/cargo#12861) - [Control `--charset` via auto-detecting config value] (rust-lang/cargo#13337) - [Support `target.<triple>.rustdocflags` officially] (rust-lang/cargo#13197) - [Stabilize global cache data tracking] (rust-lang/cargo#13492) Misc ---- - [rustdoc: add `--test-builder-wrapper` arg to support wrappers such as RUSTC_WRAPPER when building doctests] (rust-lang/rust#114651) Compatibility Notes ------------------- - [Many unsafe precondition checks now run for user code with debug assertions enabled] (rust-lang/rust#120594) This change helps users catch undefined behavior in their code, though the details of how much is checked are generally not stable. - [riscv only supports split_debuginfo=off for now] (rust-lang/rust#120518) - [Consistently check bounds on hidden types of `impl Trait`] (rust-lang/rust#121679) - [Change equality of higher ranked types to not rely on subtyping] (rust-lang/rust#118247) - [When called, additionally check bounds on normalized function return type] (rust-lang/rust#118882) - [Expand coverage for `arithmetic_overflow` lint] (rust-lang/rust#119432) Internal Changes ---------------- These changes do not affect any public interfaces of Rust, but they represent significant improvements to the performance or internals of rustc and related tools. - [Update to LLVM 18](rust-lang/rust#120055) - [Build `rustc` with 1CGU on `x86_64-pc-windows-msvc`] (rust-lang/rust#112267) - [Build `rustc` with 1CGU on `x86_64-apple-darwin`] (rust-lang/rust#112268) - [Introduce `run-make` V2 infrastructure, a `run_make_support` library and port over 2 tests as example] (rust-lang/rust#113026) - [Windows: Implement condvar, mutex and rwlock using futex] (rust-lang/rust#121956)
This is one of the last items in #93740. I'm doing
RwLock
first because it is more self-contained and has less tradeoffs to make. The motivation is explained in the documentation, but in short: the pthread rwlock is slow and buggy andstd
can do much better. I considered implementing a parking lot, as was discussed in the tracking issue, but settled for the queue-based version because writing self-balancing binary trees is not fun in Rust...This is a rather complex change, so I have added quite a bit of documentation to help explain it. Please point out any part that could be explained better.
The read performance is really good, I'm getting 4x the throughput of the pthread version and about the same performance as usync/parking_lot on an Apple M1 Max in the usync benchmark suite, but the write performance still falls way behind what usync and parking_lot achieve. I tried using a separate queue lock like what usync uses, but that didn't help. I'll try to investigate further in the future, but I wanted to get some eyes on this first.Resolvedr? @m-ou-se
CC @kprotty