std: Avoid usage of `Once` in `Instant`
This commit removes usage of `Once` from the internal implementation of
time utilities on OSX and Windows. It turns out that we accidentally hit
a deadlock today (#59020) via events that look like:
* A thread invokes `park_timeout`
* Internally, only on OSX, `park_timeout` calls `Instant::elapsed`
* Inside of `Instant::elapsed` on OSX we enter a `Once` to initialize
global timer data
* Inside of `Once`, it attempts to `park`
This means on the same stack frame, when there's contention, we're
calling `park` from inside `park_timeout`, causing a deadlock!
The solution implemented in this commit was to remove usage of `Once`
and instead just do a small dance with atomics. There's no real need we
need to guarantee that the global information is only learned once, only
that it's only *stored* once. This implementation may have multiple
threads invoke `mach_timebase_info`, but only one will store the global
information which will amortize the cost for all other threads.
A similar fix has been applied to windows to be uniform across our
implementations, but looking at the code on Windows no deadlock was
possible. This is purely just a consistency update for Windows and in
theory a slightly leaner implementation.
Closes#59020
This commit removes usage of `Once` from the internal implementation of
time utilities on OSX and Windows. It turns out that we accidentally hit
a deadlock today (#59020) via events that look like:
* A thread invokes `park_timeout`
* Internally, only on OSX, `park_timeout` calls `Instant::elapsed`
* Inside of `Instant::elapsed` on OSX we enter a `Once` to initialize
global timer data
* Inside of `Once`, it attempts to `park`
This means on the same stack frame, when there's contention, we're
calling `park` from inside `park_timeout`, causing a deadlock!
The solution implemented in this commit was to remove usage of `Once`
and instead just do a small dance with atomics. There's no real need we
need to guarantee that the global information is only learned once, only
that it's only *stored* once. This implementation may have multiple
threads invoke `mach_timebase_info`, but only one will store the global
information which will amortize the cost for all other threads.
A similar fix has been applied to windows to be uniform across our
implementations, but looking at the code on Windows no deadlock was
possible. This is purely just a consistency update for Windows and in
theory a slightly leaner implementation.
Closes#59020
This commit fills out the `std::fs` module and implementation for WASI.
Not all APIs are implemented, such as permissions-related ones and
`canonicalize`, but all others APIs have been implemented and very
lightly tested so far. We'll eventually want to run a more exhaustive
test suite!
For now the highlights of this commit are:
* The `std::fs::File` type is now backed by `WasiFd`, a raw WASI file
descriptor.
* All APIs in `std::fs` (except permissions/canonicalize) have
implementations for the WASI target.
* A suite of unstable extension traits were added to
`std::os::wasi::fs`. These traits expose the raw filesystem
functionality of WASI, namely `*at` syscalls (opening a file relative
to an already opened one, for example). Additionally metadata only
available on wasi is exposed through these traits.
Perhaps one of the most notable parts is the implementation of
path-taking APIs. WASI actually has no fundamental API that just takes a
path, but rather everything is relative to a previously opened file
descriptor. To allow existing APIs to work (that only take a path) WASI
has a few syscalls to learn about "pre opened" file descriptors by the
runtime. We use these to build a map of existing directory names to file
descriptors, and then when using a path we try to anchor it at an
already-opened file.
This support is very rudimentary though and is intended to be shared
with C since it's likely to be so tricky. For now though the C library
doesn't expose quite an API for us to use, so we implement it for now
and will swap it out as soon as one is available.
fs::copy() unix: set file mode early
A convenience method like fs::copy() should try to prevent pitfalls a
normal user doesn't think about.
In case of an empty umask, setting the file mode early prevents
temporarily world readable or even writeable files,
because the default mode is 0o666.
In case the target is a named pipe or special device node, setting the
file mode can lead to unwanted side effects, like setting permissons on
`/dev/stdout` or for root setting permissions on `/dev/null`.
copy_file_range() returns EINVAL, if the destination is a FIFO/pipe or
a device like "/dev/null", so fallback to io::copy, too.
Fixes: https://github.com/rust-lang/rust/issues/26933
Fixed: https://github.com/rust-lang/rust/issues/37885
Simplify checked_duration_since
This follows the same design as we updated to in #56490. Internally, all the system specific time implementations are checked, no panics. Then the panicking publicly exported API can just call the checked version of itself and make do with a single panic (`expect`) at the top.
Since the internal sys implementations are now checked, this gets rid of the extra `if self >= &earlier` check in `checked_duration_since`. Except likely making the generated machine code simpler, it also reduces the algorithm from "Check panic condition -> call possibly panicking method" to just "call non panicking method".
Added two test cases:
* Edge case: Make sure `checked_duration_since` on two equal `Instant`s produce a zero duration, not a `None`.
* Most common/intended usage: Make sure `later.checked_duration_since(earlier)`, returns an expected value.
A convenience method like fs::copy() should try to prevent pitfalls a
normal user doesn't think about.
In case of an empty umask, setting the file mode early prevents
temporarily world readable or even writeable files,
because the default mode is 0o666.
In case the target is a named pipe or special device node, setting the
file mode can lead to unwanted side effects, like setting permissons on
`/dev/stdout` or for root setting permissions on `/dev/null`.
copy_file_range() returns EINVAL, if the destination is a FIFO/pipe or
a device like "/dev/null", so fallback to io::copy, too.
Use `fcopyfile` on MacOS instead of `copyfile`.
Fixes: https://github.com/rust-lang/rust/issues/26933
Fixed: https://github.com/rust-lang/rust/issues/37885
Rollup of 18 pull requests
Successful merges:
- #59106 (Add peer_addr function to UdpSocket)
- #59170 (Add const generics to rustdoc)
- #59172 (Update and clean up several parts of CONTRIBUTING.md)
- #59190 (consistent naming for Rhs type parameter in libcore/ops)
- #59236 (Rename miri component to miri-preview)
- #59266 (Do not complain about non-existing fields after parse recovery)
- #59273 (some small HIR doc improvements)
- #59291 (Make Option<ThreadId> no larger than ThreadId, with NonZeroU64)
- #59297 (convert field/method confusion help to suggestions)
- #59304 (Move some bench tests back from libtest)
- #59309 (Add messages for different verbosity levels. Output copy actions.)
- #59321 (Unify E0109, E0110 and E0111)
- #59322 (Tweak incorrect escaped char diagnostic)
- #59323 (use suggestions for "enum instead of variant" error)
- #59327 (Add NAN test to docs)
- #59329 (cleanup: Remove compile-fail-fulldeps directory again)
- #59347 (Move one test from run-make-fulldeps to ui)
- #59360 (Add tracking issue number for `seek_convenience`)
Failed merges:
r? @ghost
Add peer_addr function to UdpSocket
Fixes#59104
This is my first pull request to Rust, so opening early for some feedback.
My biggest question is: where do I add tests?
Any comments very much appreciated!
Change `std::fs::copy` to use `copyfile` on MacOS and iOS
`copyfile` on MacOS is similar to `CopyFileEx` on Windows. It supports copying resource forks, extended attributes, and file ACLs, none of which are copied by the current generic unix implementation.
The API is available from MacOS 10.7 and iOS 4.3 (and possibly earlier but I haven't checked).
Closes#58895.
Add vectored read and write support
This functionality has lived for a while in the tokio ecosystem, where
it can improve performance by minimizing copies.
r? @alexcrichton
Refactor Windows stdio and remove stdin double buffering
I was looking for something nice and small to work on, tried to tackle a few FIXME's in Windows stdio, and things grew from there.
This part of the standard library contains some tricky code, and has changed over the years to handle more corner cases. It could use some refactoring and extra comments.
Changes/fixes:
- Made `StderrRaw` `pub(crate)`, to remove the `Write` implementations on `sys::Stderr` (used unsynchronised for panic output).
- Remove the unused `Read` implementation on `sys::windows::stdin`
- The `windows::stdio::Output` enum made sense when we cached the handles, but we can use simple functions like `is_console` now that we get the handle on every read/write
- `write` can now calculate the number of written bytes as UTF-8 when we can't write all `u16`s.
- If `write` could only write one half of a surrogate pair, attempt another write for the other because user code can't reslice in any way that would allow us to write it otherwise.
- Removed the double buffering on stdin. Documentation on the unexposed `StdinRaw` says: 'This handle is not synchronized or buffered in any fashion'; which is now true.
- `sys::windows::Stdin` now always only partially fills its buffer, so we can guarantee any arbitrary UTF-16 can be re-encoded without losing any data.
- `sys::windows::STDIN_BUF_SIZE` is slightly larger to compensate. There should be no real change in the number of syscalls the buffered `Stdin` does. This buffer is a little larger, while the extra buffer on Stdin is gone.
- `sys::windows::Stdin` now attempts to handle unpaired surrogates at its buffer boundary.
- `sys::windows::Stdin` no langer allocates for its buffer, but the UTF-16 decoding still does.
### Testing
I did some manual testing of reading and writing to console. The console does support UTF-16 in some sense, but doesn't supporting displaying characters outside the BMP.
- compile stage 1 stdlib with a tiny value for `MAX_BUFFER_SIZE` to make it easier to catch corner cases
- run a simple test program that reads on stdin, and echo's to stdout
- write some lines with plenty of ASCII and emoji in a text editor
- copy and paste in console to stdin
- return with `\r\n\` or CTRL-Z
- copy and paste in text editor
- check it round-trips
-----
Fixes https://github.com/rust-lang/rust/issues/23344. All but one of the suggestions in that issue are now implemented. the missing one is:
> * When reading data, we require the entire set of input to be valid UTF-16. We should instead attempt to read as much of the input as possible as valid UTF-16, only returning an error for the actual invalid elements. For example if we read 10 elements, 5 of which are valid UTF-16, the 6th is bad, and then the remaining are all valid UTF-16, we should probably return the first 5 on a call to `read`, then return an error, then return the remaining on the next call to `read`.
Stdin in Console mode is dealing with text directly input by a user. In my opinion getting an unpaired surrogate is quite unlikely in that case, and a valid reason to error on the entire line of input (which is probably short). Dealing with it is incompatible with an unbuffered stdin, which seems the more interesting guarantee to me.
Simplify the unix `Weak` functionality
- We can avoid allocation by adding a NUL to the function name.
- We can get `Option<F>` directly, rather than aliasing the inner `AtomicUsize`.
This is a non-POSIX extension implemented in Solaris and in glibc 2.29.
With this we can still use `posix_spawn()` when `Command::current_dir()`
has been set, otherwise we fallback to `fork(); chdir(); exec()`.
Cosmetic improvements to doc comments
This has been factored out from https://github.com/rust-lang/rust/pull/58036 to only include changes to documentation comments (throughout the rustc codebase).
r? @steveklabnik
Once you're happy with this, maybe we could get it through with r=1, so it doesn't constantly get invalidated? (I'm not sure this will be an issue, but just in case...) Anyway, thanks for your advice so far!
OSX: fix#57534 registering thread dtors while running thread dtors
r? @alexcrichton
- "fast" `thread_local` destructors get run even on the main thread
- "fast" `thread_local` dtors, can initialize other `thread_local`'s
One corner case where this fix doesn't work, is when a C++ `thread_local` triggers the initialization of a rust `thread_local`.
I did not add any std::thread specific flag to indicate that the thread is currently exiting, which would be checked before registering a new dtor (I didn't really know where to stick that). I think this does the trick tho!
Let me know if anything needs tweaking/fixing/etc.
resolves this for macos: https://github.com/rust-lang/rust/issues/28129
fixes: https://github.com/rust-lang/rust/issues/57534
