Enforce no-move rule of ReentrantMutex using Pin and fix UB in stdio
A `sys_common::ReentrantMutex` may not be moved after initializing it with `.init()`. This was not enforced, but only stated as a requirement in the comments on the unsafe functions. This change enforces this no-moving rule using `Pin`, by changing `&self` to a `Pin` in the `init()` and `lock()` functions.
This uncovered a bug I introduced in #77154: stdio.rs (the only user of ReentrantMutex) called `init()` on its ReentrantMutexes while constructing them in the intializer of `SyncOnceCell::get_or_init`, which would move them afterwards. Interestingly, the ReentrantMutex unit tests already had the same bug, so this invalid usage has been tested on all (CI-tested) platforms for a long time. Apparently this doesn't break badly on any of the major platforms, but it does break the rules.\*
To be able to keep using SyncOnceCell, this adds a `SyncOnceCell::get_or_init_pin` function, which makes it possible to work with pinned values inside a (pinned) SyncOnceCell. Whether this function should be public or not and what its exact behaviour and interface should be if it would be public is something I'd like to leave for a separate issue or PR. In this PR, this function is internal-only and marked with `pub(crate)`.
\* Note: That bug is now included in 1.48, while this patch can only make it to ~~1.49~~ 1.50. We should consider the implications of 1.48 shipping with a wrong usage of `pthread_mutex_t` / `CRITICAL_SECTION` / .. which technically invokes UB according to their specification. The risk is very low, considering the objects are not 'used' (locked) before the move, and the ReentrantMutex unit tests have verified this works fine in practice.
Edit: This has been backported and included in 1.48. And soon 1.49 too.
---
In future changes, I want to push this usage of Pin further inside `sys` instead of only `sys_common`, and apply it to all 'unmovable' objects there (`Mutex`, `Condvar`, `RwLock`). Also, while `sys_common`'s mutexes and condvars are already taken care of by #77147 and #77648, its `RwLock` should still be made movable or get pinned.
Use is_write_vectored to optimize the write_vectored implementation for BufWriter
In case when the underlying writer does not have an efficient implementation `write_vectored`, the present implementation of
`write_vectored` for `BufWriter` may still forward vectored writes directly to the writer depending on the total length of the data. This misses the advantage of buffering, as the actually written slice may be small.
Provide an alternative code path for the non-vectored case, where the slices passed to `BufWriter` are coalesced in the buffer before being flushed to the underlying writer with plain `write` calls. The buffer is only bypassed if an individual slice's length is at least as large as the buffer.
Remove a FIXME comment referring to #72919 as the issue has been closed with an explanation provided.
The code in io::stdio before this change misused the ReentrantMutexes,
by calling init() on them and moving them afterwards. Now that
ReentrantMutex requires Pin for init(), this mistake is no longer easy
to make.
In particular, IntoIneerError only currently provides .error() which
returns a reference, not an owned value. This is not helpful and
means that a caller of BufWriter::into_inner cannot acquire an owned
io::Error which seems quite wrong.
Signed-off-by: Ian Jackson <ijackson@chiark.greenend.org.uk>
Do what write does and optimize for the most likely case:
slices are much smaller than the buffer. If a slice does not fit
completely in the remaining capacity of the buffer, it is left out
rather than buffered partially. Special treatment is only left for
oversized slices that are written directly to the underlying writer.
Now that BufWriter always claims to support vectored writes,
look through it at the wrapped writer to decide whether to
use vectored writes for LineWriter.
If the underlying writer does not support efficient vectored output,
do it differently: always try to coalesce the slices in the buffer
until one comes that does not fit entirely. Flush the buffer before
the first slice if needed.
These referred to a “`Write`er”—extra *e*. Presumably a copy-paste
holdover from “`Read`er”.
Test Plan:
Running ``git grep '`\?[Ww]rite`\?er'`` no longer finds any results.
wchargin-branch: io-write-docs
Simplify output capturing
This is a sequence of incremental improvements to the unstable/internal `set_panic` and `set_print` mechanism used by the `test` crate:
1. Remove the `LocalOutput` trait and use `Arc<Mutex<dyn Write>>` instead of `Box<dyn LocalOutput>`. In practice, all implementations of `LocalOutput` were just `Arc<Mutex<..>>`. This simplifies some logic and removes all custom `Sink` implementations such as `library/test/src/helpers/sink.rs`. Also removes a layer of indirection, as the outermost `Box` is now gone. It also means that locking now happens per `write_fmt`, not per individual `write` within. (So `"{} {}\n"` now results in one `lock()`, not four or more.)
2. Since in all cases the `dyn Write`s were just `Vec<u8>`s, replace the type with `Arc<Mutex<Vec<u8>>>`. This simplifies things more, as error handling and flushing can be removed now. This also removes the hack needed in the default panic handler to make this work with `::realstd`, as (unlike `Write`) `Vec<u8>` is from `alloc`, not `std`.
3. Replace the `RefCell`s by regular `Cell`s. The `RefCell`s were mostly used as `mem::replace(&mut *cell.borrow_mut(), something)`, which is just `Cell::replace`. This removes an unecessary bookkeeping and makes the code a bit easier to read.
4. Merge `set_panic` and `set_print` into a single `set_output_capture`. Neither the test crate nor rustc (the only users of this feature) have a use for using these separately. Merging them simplifies things even more. This uses a new function name and feature name, to make it clearer this is internal and not supposed to be used by other crates.
Might be easier to review per commit.
specialize io::copy to use copy_file_range, splice or sendfile
Fixes#74426.
Also covers #60689 but only as an optimization instead of an official API.
The specialization only covers std-owned structs so it should avoid the problems with #71091
Currently linux-only but it should be generalizable to other unix systems that have sendfile/sosplice and similar.
There is a bit of optimization potential around the syscall count. Right now it may end up doing more syscalls than the naive copy loop when doing short (<8KiB) copies between file descriptors.
The test case executes the following:
```
[pid 103776] statx(3, "", AT_STATX_SYNC_AS_STAT|AT_EMPTY_PATH, STATX_ALL, {stx_mask=STATX_ALL|STATX_MNT_ID, stx_attributes=0, stx_mode=S_IFREG|0644, stx_size=17, ...}) = 0
[pid 103776] write(4, "wxyz", 4) = 4
[pid 103776] write(4, "iklmn", 5) = 5
[pid 103776] copy_file_range(3, NULL, 4, NULL, 5, 0) = 5
```
0-1 `stat` calls to identify the source file type. 0 if the type can be inferred from the struct from which the FD was extracted
𝖬 `write` to drain the `BufReader`/`BufWriter` wrappers. only happen when buffers are present. 𝖬 ≾ number of wrappers present. If there is a write buffer it may absorb the read buffer contents first so only result in a single write. Vectored writes would also be an option but that would require more invasive changes to `BufWriter`.
𝖭 `copy_file_range`/`splice`/`sendfile` until file size, EOF or the byte limit from `Take` is reached. This should generally be *much* more efficient than the read-write loop and also have other benefits such as DMA offload or extent sharing.
## Benchmarks
```
OLD
test io::tests::bench_file_to_file_copy ... bench: 21,002 ns/iter (+/- 750) = 6240 MB/s [ext4]
test io::tests::bench_file_to_file_copy ... bench: 35,704 ns/iter (+/- 1,108) = 3671 MB/s [btrfs]
test io::tests::bench_file_to_socket_copy ... bench: 57,002 ns/iter (+/- 4,205) = 2299 MB/s
test io::tests::bench_socket_pipe_socket_copy ... bench: 142,640 ns/iter (+/- 77,851) = 918 MB/s
NEW
test io::tests::bench_file_to_file_copy ... bench: 14,745 ns/iter (+/- 519) = 8889 MB/s [ext4]
test io::tests::bench_file_to_file_copy ... bench: 6,128 ns/iter (+/- 227) = 21389 MB/s [btrfs]
test io::tests::bench_file_to_socket_copy ... bench: 13,767 ns/iter (+/- 3,767) = 9520 MB/s
test io::tests::bench_socket_pipe_socket_copy ... bench: 26,471 ns/iter (+/- 6,412) = 4951 MB/s
```
It was only ever used with Vec<u8> anyway. This simplifies some things.
- It no longer needs to be flushed, because that's a no-op anyway for
a Vec<u8>.
- Writing to a Vec<u8> never fails.
- No #[cfg(test)] code is needed anymore to use `realstd` instead of
`std`, because Vec comes from alloc, not std (like Write).
Capture output from threads spawned in tests
This is revival of #75172.
Original text:
> Fixes#42474.
>
> r? `@dtolnay` since you expressed interest in this, but feel free to redirect if you aren't the right person anymore.
---
Closes#75172.
Refactor io/buffered.rs into submodules
This pull request splits `BufWriter`, `BufReader`, `LineWriter`, and `LineWriterShim` (along with their associated tests) into separate submodules. It contains no functional changes. This change is being made in anticipation of adding another type of buffered writer which can be switched between line- and block-buffering mode.
Part of a series of pull requests resolving #60673.
The thread local LOCAL_STDOUT and LOCAL_STDERR are only used by the test
crate to capture output from tests when running them in the same process
in differen threads. However, every program will check these variables
on every print, even outside of testing.
This involves allocating a thread local key, and registering a thread
local destructor. This can be somewhat expensive.
This change keeps a global flag (LOCAL_STREAMS) which will be set to
true when either of these local streams is used. (So, effectively only
in test and benchmark runs.) When this flag is off, these thread locals
are not even looked at and therefore will not be initialized on the
first output on every thread, which also means no thread local
destructors will be registered.
Remove std::io::lazy::Lazy in favour of SyncOnceCell
The (internal) std::io::lazy::Lazy was used to lazily initialize the stdout and stdin buffers (and mutexes). It uses atexit() to register a destructor to flush the streams on exit, and mark the streams as 'closed'. Using the stream afterwards would result in a panic.
Stdout uses a LineWriter which contains a BufWriter that will flush the buffer on drop. This one is important to be executed during shutdown, to make sure no buffered output is lost. It also forbids access to stdout afterwards, since the buffer is already flushed and gone.
Stdin uses a BufReader, which does not implement Drop. It simply forgets any previously read data that was not read from the buffer yet. This means that in the case of stdin, the atexit() function's only effect is making stdin inaccessible to the program, such that later accesses result in a panic. This is uncessary, as it'd have been safe to access stdin during shutdown of the program.
---
This change removes the entire io::lazy module in favour of SyncOnceCell. SyncOnceCell's fast path is much faster (a single atomic operation) than locking a sys_common::Mutex on every access like Lazy did.
However, SyncOnceCell does not use atexit() to drop the contained object during shutdown.
As noted above, this is not a problem for stdin. It simply means stdin is now usable during shutdown.
The atexit() call for stdout is moved to the stdio module. Unlike the now-removed Lazy struct, SyncOnceCell does not have a 'gone and unusable' state that panics. Instead of adding this again, this simply replaces the buffer with one with zero capacity. This effectively flushes the old buffer *and* makes any writes afterwards pass through directly without touching a buffer, making print!() available during shutdown without panicking.
---
In addition, because the contents of the SyncOnceCell are no longer dropped, we can now use `&'static` instead of `Arc` in `Stdout` and `Stdin`. This also saves two levels of indirection in `stdin()` and `stdout()`, since Lazy effectively stored a `Box<Arc<T>>`, and SyncOnceCell stores the `T` directly.
The (internal) std::io::lazy::Lazy was used to lazily initialize the
stdout and stdin buffers (and mutexes). It uses atexit() to register a
destructor to flush the streams on exit, and mark the streams as
'closed'. Using the stream afterwards would result in a panic.
Stdout uses a LineWriter which contains a BufWriter that will flush the
buffer on drop. This one is important to be executed during shutdown,
to make sure no buffered output is lost. It also forbids access to
stdout afterwards, since the buffer is already flushed and gone.
Stdin uses a BufReader, which does not implement Drop. It simply forgets
any previously read data that was not read from the buffer yet. This
means that in the case of stdin, the atexit() function's only effect is
making stdin inaccessible to the program, such that later accesses
result in a panic. This is uncessary, as it'd have been safe to access
stdin during shutdown of the program.
---
This change removes the entire io::lazy module in favour of
SyncOnceCell. SyncOnceCell's fast path is much faster (a single atomic
operation) than locking a sys_common::Mutex on every access like Lazy
did.
However, SyncOnceCell does not use atexit() to drop the contained object
during shutdown.
As noted above, this is not a problem for stdin. It simply means stdin
is now usable during shutdown.
The atexit() call for stdout is moved to the stdio module. Unlike the
now-removed Lazy struct, SyncOnceCell does not have a 'gone and
unusable' state that panics. Instead of adding this again, this simply
replaces the buffer with one with zero capacity. This effectively
flushes the old buffer *and* makes any writes afterwards pass through
directly without touching a buffer, making print!() available during
shutdown without panicking.