Unify id-based thread parking implementations
Multiple platforms currently use thread-id-based parking implementations (NetBSD and SGX[^1]). Even though the strategy does not differ, these are duplicated for each platform, as the id is encoded into an atomic thread variable in different ways for each platform.
Since `park` is only called by one thread, it is possible to move the thread id into a separate field. By ensuring that the field is only written to once, before any other threads access it, these accesses can be unsynchronized, removing any restrictions on the size and niches of the thread id.
This PR also renames the internal `thread_parker` modules to `thread_parking`, as that name now better reflects their contents. I hope this does not add too much reviewing noise.
r? `@m-ou-se`
`@rustbot` label +T-libs
[^1]: SOLID supports this as well, I will switch it over in a follow-up PR.
Make sentinel value configurable in `library/std/src/sys_common/thread_local_key.rs`
This is an excerpt of a changeset for the QNX/Neutrino OS. To make the patch for QNX smaller and easier to review, I've extracted this change (which is OS independent). I would be surprised if no other OS is also affected.
All this patch does is to define a `const` for a sentinel value instead of using it directly at several places.
There are OSs that always return the lowest free value. The algorithm in `lazy_init` always avoids keys with the sentinel value.
In affected OSs, this means that each call to `lazy_init` will always request two keys from the OS and returns/frees the first one (with sentinel value) immediately afterwards.
By making the sentinel value configurable, affected OSs can use a different value than zero to prevent this performance issue.
On QNX/Neutrino, it is planned to use a different sentinel value:
```rust
// Define a sentinel value that is unlikely to be returned
// as a TLS key (but it may be returned).
#[cfg(not(target_os = "nto"))]
const KEY_SENTVAL: usize = 0;
// On QNX/Neutrino, 0 is always returned when currently not in use.
// Using 0 would mean to always create two keys and remote the first
// one (with value of 0) immediately afterwards.
#[cfg(target_os = "nto")]
const KEY_SENTVAL: usize = libc::PTHREAD_KEYS_MAX + 1;
```
It seems like no other OS defines `PTHREAD_KEYS_MAX` in Rusts libc, but `limits.h` on unix systems does.
Move `ReentrantMutex` to `std::sync`
If I understand #84187 correctly, `sys_common` should not contain platform-independent code, even if it is private.
Use a more efficient `Once` on platforms without threads
The current implementation uses an atomic queue and spins rather than panicking when calling `call_once` recursively. Since concurrency is not supported on platforms like WASM, `Once` can be implemented much more efficiently using just a single non-atomic state variable.
Allow blocking `Command::output`
### Problem
Currently, `Command::output` is internally implemented using `Command::spawn`. This is problematic because some targets (like UEFI) do not actually support multitasking and thus block while the program is executing. This coupling does not make much sense as `Command::output` is supposed to block until the execution is complete anyway and thus does not need to rely on a non-blocking `Child` or any other intermediate.
### Solution
This PR moves the implementation of `Command::output` to `std::sys`. This means targets can choose to implement only `Command::output` without having to implement `Command::spawn`.
### Additional Information
This was originally conceived when working on https://github.com/rust-lang/rust/pull/100316. Currently, the only target I know about that will benefit from this change is UEFI.
This PR can also be used to implement more efficient `Command::output` since the intermediate `Process` is not actually needed anymore, but that is outside the scope of this PR.
Since this is not a public API change, I'm not sure if an RFC is needed or not.
This allows decoupling `Command::spawn` and `Command::output`. This is
useful for targets which do support launching programs in blocking mode
but do not support multitasking (Eg: UEFI).
This was originally conceived when working on https://github.com/rust-lang/rust/pull/100316
Signed-off-by: Ayush Singh <ayushsingh1325@gmail.com>
Reimplement std's thread parker on top of events on SGX
Mutex and Condvar are being replaced by more efficient implementations, which need thread parking themselves (see #93740). Therefore, the generic `Parker` needs to be replaced on all platforms where the new lock implementation will be used.
SGX enclaves have a per-thread event state, which allows waiting for and setting specific bits. This is already used by the current mutex implementation. The thread parker can however be much more efficient, as it only needs to store the `TCS` address of one thread. This address is stored in a state variable, which can also be set to indicate the thread was already notified.
`park_timeout` does not guard against spurious wakeups like the current condition variable does. This is allowed by the API of `Parker`, and I think it is better to let users handle these wakeups themselves as the guarding is quite expensive and might not be necessary.
`@jethrogb` as you wrote the initial SGX support for `std`, I assume you are the target maintainer? Could you help me test this, please? Lacking a x86_64 chip, I can't run SGX.
There are OSs that always return the lowest free value.
The algorithm in `lazy_init` always avoids keys with the
sentinel value.
In affected OSs, this means that each call to `lazy_init`
will always request two keys from the OS and returns/frees
the first one (with sentinel value) immediately afterwards.
By making the sentinel value configurable, affected OSs can
use a different value than zero to prevent this performance
issue.
Extract WStrUnits to sys_common::wstr
This commit extracts WStrUnits from sys::windows::args to sys_common::wstr. This allows using the same structure for other targets which use wtf8 (example UEFI).
This was originally a part of https://github.com/rust-lang/rust/pull/100316
Signed-off-by: Ayush Singh <ayushsingh1325@gmail.com>
This commit extracts WStrUnits from sys::windows::args to sys_common::wstr. This
allows using the same structure for other targets which use wtf8 (example UEFI).
This was originally a part of https://github.com/rust-lang/rust/pull/100316
Signed-off-by: Ayush Singh <ayushsingh1325@gmail.com>
Optimize TLS on Windows
This implements the suggestion in the current TLS code to embed the linked list of destructors in the `StaticKey` structure to save allocations. Additionally, locking is avoided when no destructor needs to be run. By using one Windows-provided `Once` per key instead of a global lock, locking is more finely-grained (this unblocks #100579).
Reduce CString allocations in std as much as possible
Currently, every operation involving paths in `fs` allocates memory to hold the path before sending it through the syscall. This PR instead uses a stack allocation (chosen size is somewhat arbitrary) when the path is short before falling back to heap allocations for long paths.
Benchmarks show that the stack allocation is ~2x faster for short paths:
```
test sys::unix::fd::tests::bench_heap_path_alloc ... bench: 34 ns/iter (+/- 2)
test sys::unix::fd::tests::bench_stack_path_alloc ... bench: 15 ns/iter (+/- 1)
```
For long paths, I couldn't find any measurable difference.
---
I'd be surprised if I was the first to think of this, so I didn't fully flush out the PR. If this change is desirable, I'll make use of `run_with_cstr` across all platforms in every fs method (currently just unix open for testing). I also added an `impl From<FromBytesWithNulError>` which is presumably a no-no (or at least needs to be done in another PR).
---
Also see https://github.com/nix-rust/nix/pull/1655 with a bunch of discussion where I'm doing something similar.
std: use futex in `Once`
Now that we have efficient locks, let's optimize the rest of `sync` as well. This PR adds a futex-based implementation for `Once`, which drastically simplifies the implementation compared to the generic version, which is provided as fallback for platforms without futex (Windows only supports them on newer versions, so it uses the fallback for now).
Instead of storing a linked list of waiters, the new implementation adds another state (`QUEUED`), which is set when there are waiting threads. These now use `futex_wait` on that state and are woken by the running thread when it finishes and notices the `QUEUED` state, thereby avoiding unnecessary calls to `futex_wake_all`.
make Condvar, Mutex, RwLock const constructors work with the `unsupported` impl
applying this patch locally to the `rust-src` component fixes#98378
however, the solution seems wrong to me because PR #97791 didn't add any `rustc_const_stable` attribute to underlying implementations like `std::sys::unix::futex`, so I must be missing something about how const-stability is checked ... maybe the `restricted_std` feature (gate?) has an effect?
fixes#98378fixes#98293 (probably)
std: use `sync::RwLock` for internal statics
Since `sync::RwLock` is now `const`-constructible, it can be used for internal statics, removing the need for `sys_common::StaticRwLock`. This adds some extra allocations on platforms which need to box their locks (currently SGX and some UNIX), but these will become unnecessary with the lock improvements tracked in #93740.
Make `ReentrantMutex` movable and `const`
As `MovableMutex` is now `const`, it can be used to simplify the implementation and interface of the internal reentrant mutex type. Consequently, the standard error stream does not need to be wrapped in `OnceLock` and `OnceLock::get_or_init_pin()` can be removed.
std: use realstd fast key when building tests
Under `cfg(test)`, the `std` crate is not the actual standard library, just any old crate we are testing. It imports the real standard library as `realstd`, and then does some careful `cfg` magic so that the crate built for testing uses the `realstd` global state rather than having its own copy of that.
However, this was not done for all global state hidden in std: the 'fast' version of thread-local keys, at least on some platforms, also involves some global state. Specifically its macOS version has this [`static REGISTERED`](bc63d5a26a/library/std/src/sys/unix/thread_local_dtor.rs (L62)) that would get duplicated. So this PR imports the 'fast' key type from `realstd` rather than using the local copy, to ensure its internal state (and that of the functions it calls) does not get duplicated.
I also noticed that the `__OsLocalKeyInner` is unused under `cfg(target_thread_local)`, so I removed it for that configuration. There was a comment saying macOS picks between `__OsLocalKeyInner` and `__FastLocalKeyInner` at runtime, but I think that comment is outdated -- I found no trace of such a runtime switching mechanism, and the library still check-builds on apple targets with this PR. (I don't have a Mac so I cannot actually run it.)
Optimize `Wtf8Buf::into_string` for the case where it contains UTF-8.
Add a `is_known_utf8` flag to `Wtf8Buf`, which tracks whether the
string is known to contain UTF-8. This is efficiently computed in many
common situations, such as when a `Wtf8Buf` is constructed from a `String`
or `&str`, or with `Wtf8Buf::from_wide` which is already doing UTF-16
decoding and already checking for surrogates.
This makes `OsString::into_string` O(1) rather than O(N) on Windows in
common cases.
And, it eliminates the need to scan through the string for surrogates in
`Args::next` and `Vars::next`, because the strings are already being
translated with `Wtf8Buf::from_wide`.
Many things on Windows construct `OsString`s with `Wtf8Buf::from_wide`,
such as `DirEntry::file_name` and `fs::read_link`, so with this patch,
users of those functions can subsequently call `.into_string()` without
paying for an extra scan through the string for surrogates.
r? `@ghost`
Guarantee `try_reserve` preserves the contents on error
Update doc comments to make the guarantee explicit. However, some
implementations does not have the statement though.
* `HashMap`, `HashSet`: require guarantees on hashbrown side.
* `PathBuf`: simply redirecting to `OsString`.
Fixes#99606.
Update doc comments to make the guarantee explicit. However, some
implementations does not have the statement though.
* `HashMap`, `HashSet`: require guarantees on hashbrown side.
* `PathBuf`: simply redirecting to `OsString`.
Fixes#99606.
Implement network primitives with ideal Rust layout, not C system layout
This PR is the result of this internals forum thread: https://internals.rust-lang.org/t/why-are-socketaddrv4-socketaddrv6-based-on-low-level-sockaddr-in-6/13321.
Instead of basing `std:::net::{Ipv4Addr, Ipv6Addr, SocketAddrV4, SocketAddrV6}` on system (C) structs, they are encoded in a more optimal and idiomatic Rust way.
This changes the public API of std by introducing structural equality impls for all four types here, which means that `match ipv4addr { SOME_CONSTANT => ... }` will now compile, whereas previously this was an error. No other intentional changes are introduced to public API.
It's possible to observe the current layout of these types (e.g., by pointer casting); most but not all libraries which were found by Crater to do this have had updates issued and affected versions yanked. See report below.
### Benefits of this change
- It will become possible to move these fundamental network types from `std` into `core` ([RFC](https://github.com/rust-lang/rfcs/pull/2832)).
- Some methods that can't be made `const fn`s today can be made `const fn`s with this change.
- `SocketAddrV4` only occupies 6 bytes instead of 16 bytes.
- These simple primitives become easier to read and uses less `unsafe`.
- Makes these types support structural equality, which means you can now (for instance) match an `Ipv4Addr` against a constant
### ~Remaining~ Previous problems
This change obviously changes the memory layout of the types. And it turns out some libraries invalidly assumes the memory layout and does very dangerous pointer casts to convert them. These libraries will have undefined behaviour and perform invalid memory access until patched.
- [x] - `mio` - Issue: https://github.com/tokio-rs/mio/issues/1386.
- [x] `0.7` branch https://github.com/tokio-rs/mio/pull/1388
- [x] `0.7.6` published https://github.com/tokio-rs/mio/pull/1398
- [x] Yank all `0.7` versions older than `0.7.6`
- [x] Report `<0.7.6` to RustSec Advisory Database https://rustsec.org/advisories/RUSTSEC-2020-0081.html
- [x] - `socket2` - Issue: https://github.com/rust-lang/socket2-rs/issues/119.
- [x] `0.3.x` branch https://github.com/rust-lang/socket2-rs/pull/120
- [x] `0.3.16` published
- [x] `master` branch https://github.com/rust-lang/socket2-rs/pull/122
- [x] Yank all `0.3` versions older than `0.3.16`
- [x] Report `<0.3.16` to RustSec Advisory Database https://rustsec.org/advisories/RUSTSEC-2020-0079.html
- [x] - `net2` - Issue: https://github.com/deprecrated/net2-rs/issues/105
- [x] https://github.com/deprecrated/net2-rs/pull/106
- [x] `0.2.36` published
- [x] Yank all `0.2` versions older than `0.2.36`
- [x] Report `<0.2.36` to RustSec Advisory Database https://rustsec.org/advisories/RUSTSEC-2020-0078.html
- [x] - `miow` - Issue: https://github.com/yoshuawuyts/miow/issues/38
- [x] `0.3.x` - https://github.com/yoshuawuyts/miow/pull/39
- [x] `0.3.6` published
- [x] `0.2.x` - https://github.com/yoshuawuyts/miow/pull/40
- [x] `0.2.2` published
- [x] Yanked all `0.2` versions older than `0.2.2`
- [x] Yanked all `0.3` versions older than `0.3.6`
- [x] Report `<0.2.2` and `<0.3.6` to RustSec Advisory Database https://rustsec.org/advisories/RUSTSEC-2020-0080.html
- [x] - `quinn master` (aka what became 0.7) - https://github.com/quinn-rs/quinn/issues/968https://github.com/quinn-rs/quinn/pull/987
- [x] - `quinn 0.6` - https://github.com/quinn-rs/quinn/pull/1045
- [x] - `quinn 0.5` - https://github.com/quinn-rs/quinn/pull/1046
- [x] - Release `0.7.0`, `0.6.2` and `0.5.4`
- [x] - `nb-connect` - https://github.com/smol-rs/nb-connect/issues/1
- [x] - Release `1.0.3`
- [x] - Yank all versions older than `1.0.3`
- [x] - `shadowsocks-rust` - https://github.com/shadowsocks/shadowsocks-rust/issues/462
- [ ] - `rio` - https://github.com/spacejam/rio/issues/44
- [ ] - `seaslug` - https://github.com/spacejam/seaslug/issues/1
#### Fixed crate versions
All crates I have found that assumed the memory layout have been fixed and published. The crates and versions that will continue working even as/if this PR is merged is (please upgrade these to help unblock this PR):
* `net2 0.2.36`
* `socket2 0.3.16`
* `miow 0.2.2`
* `miow 0.3.6`
* `mio 0.7.6`
* `mio 0.6.23` - Never had the invalid assumption itself, but has now been bumped to only allow fixed dependencies (`net2` + `miow`)
* `nb-connect 1.0.3`
* `quinn 0.5.4`
* `quinn 0.6.2`
### Release notes draft
This release changes the memory layout of `Ipv4Addr`, `Ipv6Addr`, `SocketAddrV4` and `SocketAddrV6`. The standard library no longer implements these as the corresponding `libc` structs (`sockaddr_in`, `sockaddr_in6` etc.). This internal representation was never exposed, but some crates relied on it anyway by unsafely transmuting. This change will cause those crates to make invalid memory accesses. Notably `net2 <0.2.36`, `socket2 <0.3.16`, `mio <0.7.6`, `miow <0.3.6` and a few other crates are affected. All known affected crates have been patched and have had fixed versions published over a year ago. If any affected crate is still in your dependency tree, you need to upgrade them before using this version of Rust.
stdlib support for Apple WatchOS
This is a follow-up to https://github.com/rust-lang/rust/pull/95243 (Add Apple WatchOS compiler targets) that adds stdlib support for Apple WatchOS.
`@deg4uss3r`
`@nagisa`
