`FileDescriptor.duplicate(as:retryOnInterrupt:)` writes its first
parameter entry as ``- `target`:`` rather than `- target:`. DocC matches
parameter documentation by the bare identifier, so the backticked entry
is not associated with the parameter: the description is dropped from
the rendered parameter list and DocC reports the parameter as
undocumented.
The `duplicate(as:options:retryOnInterrupt:)` overload twenty lines
below documents the same parameter without backticks, which is what
this changes the first overload to match.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MDv2d3fEYCVWGpu7AgD4qq
The docs for FilePermissions.setUserID and setGroupID referred to the setuid(2) and setgid(2) man pages. Those pages describe system calls that change a process's IDs. The permission bits are documented in chmod(2) as S_ISUID and S_ISGID. This matches how saveText already refers to S_ISVTX.
Use String(unsafeUninitializedCapacity:initializingUTF8With:), through
the existing backwards-compatible helper, so the message is written
straight into the string's storage instead of a stack buffer that is
then copied. Also update the copyright year range of Errno.swift.
musl's message for an unknown error number doesn't include the number,
so skip that check when building against Musl, as on WASI.
strerror_s on Windows and bionic's strerror_r truncate without
returning an error, so detect a full buffer in the shim instead.
Also avoid the deprecated String(cString:) array overload in the test.
strerror(3) isn't guaranteed to be thread-safe: for an unknown error
number, it may format the message into a static buffer shared by all
threads, so concurrent calls to Errno.description could return each
other's text.
Add a csystem_strerror_r shim to CSystem that writes into a
caller-provided buffer. It hides the difference between the XSI
strerror_r (Darwin, BSDs, musl, WASI) and the GNU variant that glibc
and bionic provide under _GNU_SOURCE, and uses strerror_s on Windows.
Errno.description now formats into a stack buffer through this shim.
Also add tests for Errno.description, which were a TODO.
Fixes#156
When an operation failed, the blocking consume methods dropped its completion and only reported the `Errno`. The caller could not tell which request failed, since the `context` was lost. This breaks any user that matches completions to requests.
This PR makes `blockingConsumeCompletions` pass the completion of a failed operation together with its error. Errors of the ring itself, such as `.timeout`, are still passed without a completion. Additionally, `blockingConsumeCompletion(timeout:)` now returns the completion of a failed operation instead of throwing. It only throws errors of the ring itself.
I also added `Completion.error`, which is `nil` on success and the decoded `Errno` otherwise to make it easy to get to the error of a completion.
Fixes#364Fixes#395
Darwin.Mach's mach_task_self_ is a shared global; Swift 6.3
reports it as not concurrency-safe when imported. Use a
preconcurrency import for older Darwin SDKs.
Fixes#373
Co-authored-by: Guillaume Lessard <guillaume.lessard@apple.com>
The gate os(WASI) && (swift(<6.1) || !_runtime) incorrectly falls
through to pthread APIs on Swift 6.1+ non-threaded WASI, where
pthread_* is not in scope. Guard solely on !_runtime.
Fixes#370