Eliminate 280-byte memset from ReadDir iterator
This guy:
1536ab1b38/library/std/src/sys/unix/fs.rs (L589)
It turns out `libc::dirent64` is quite big—https://docs.rs/libc/0.2.135/libc/struct.dirent64.html. In #103135 this memset accounted for 0.9% of the runtime of iterating a big directory.
Almost none of the big zeroed value is ever used. We memcpy a tiny prefix (19 bytes) into it, and then read just 9 bytes (`d_ino` and `d_type`) back out. We can read exactly those 9 bytes we need directly from the original entry_ptr instead.
## History
This code got added in #93459 and tweaked in #94272 and #94750.
Prior to #93459, there was no memset but a full 280 bytes were being copied from the entry_ptr.
<table><tr><td>copy 280 bytes</td></tr></table>
This was not legal because not all of those bytes might be initialized, or even allocated, depending on the length of the directory entry's name, leading to a segfault. That PR fixed the segfault by creating a new zeroed dirent64 and copying just the guaranteed initialized prefix into it.
<table><tr><td>memset 280 bytes</td><td>copy 19 bytes</td></tr></table>
However this was still buggy because it used `addr_of!((*entry_ptr).d_name)`, which is considered UB by Miri in the case that the full extent of entry_ptr is not in bounds of the same allocation. (Arguably this shouldn't be a requirement, but here we are.)
The UB got fixed by #94272 by replacing `addr_of` with some pointer manipulation based on `offset_from`, but still fundamentally the same operation.
<table><tr><td>memset 280 bytes</td><td>copy 19 bytes</td></tr></table>
Then #94750 noticed that only 9 of those 19 bytes were even being used, so we could pick out only those 9 to put in the ReadDir value.
<table><tr><td>memset 280 bytes</td><td>copy 19 bytes</td><td>copy 9 bytes</td></tr></table>
After my PR we just grab the 9 needed bytes directly from entry_ptr.
<table><tr><td>copy 9 bytes</td></tr></table>
The resulting code is more complex but I believe still worthwhile to land for the following reason. This is an extremely straightforward thing to accomplish in C and clearly libc assumes that; literally just `entry_ptr->d_name`. The extra work in comparison to accomplish it in Rust is not an example of any actual safety being provided by Rust. I believe it's useful to have uncovered that and think about what could be done in the standard library or language to support this obvious operation better.
## References
- https://man7.org/linux/man-pages/man3/readdir.3.html
fs::get_path solarish version.
similar to linux, albeit there is no /proc/self notion on solaris
based system thus flattening the difference for simplification sake.
Rollup of 8 pull requests
Successful merges:
- #101118 (fs::get_mode enable getting the data via fcntl/F_GETFL on major BSD)
- #102072 (Add `ptr::Alignment` type)
- #102799 (rustdoc: remove hover gap in file picker)
- #102820 (Show let-else suggestion on stable.)
- #102829 (rename `ImplItemKind::TyAlias` to `ImplItemKind::Type`)
- #102831 (Don't use unnormalized type in `Ty::fn_sig` call in rustdoc `clean_middle_ty`)
- #102834 (Remove unnecessary `lift`/`lift_to_tcx` calls from rustdoc)
- #102838 (remove cfg(bootstrap) from Miri)
Failed merges:
r? `@ghost`
`@rustbot` modify labels: rollup
std::io: migrate ReadBuf to BorrowBuf/BorrowCursor
This PR replaces `ReadBuf` (used by the `Read::read_buf` family of methods) with `BorrowBuf` and `BorrowCursor`.
The general idea is to split `ReadBuf` because its API is large and confusing. `BorrowBuf` represents a borrowed buffer which is mostly read-only and (other than for construction) deals only with filled vs unfilled segments. a `BorrowCursor` is a mostly write-only view of the unfilled part of a `BorrowBuf` which distinguishes between initialized and uninitialized segments. For `Read::read_buf`, the caller would create a `BorrowBuf`, then pass a `BorrowCursor` to `read_buf`.
In addition to the major API split, I've made the following smaller changes:
* Removed some methods entirely from the API (mostly the functionality can be replicated with two calls rather than a single one)
* Unified naming, e.g., by replacing initialized with init and assume_init with set_init
* Added an easy way to get the number of bytes written to a cursor (`written` method)
As well as simplifying the API (IMO), this approach has the following advantages:
* Since we pass the cursor by value, we remove the 'unsoundness footgun' where a malicious `read_buf` could swap out the `ReadBuf`.
* Since `read_buf` cannot write into the filled part of the buffer, we prevent the filled part shrinking or changing which could cause underflow for the caller or unexpected behaviour.
## Outline
```rust
pub struct BorrowBuf<'a>
impl Debug for BorrowBuf<'_>
impl<'a> From<&'a mut [u8]> for BorrowBuf<'a>
impl<'a> From<&'a mut [MaybeUninit<u8>]> for BorrowBuf<'a>
impl<'a> BorrowBuf<'a> {
pub fn capacity(&self) -> usize
pub fn len(&self) -> usize
pub fn init_len(&self) -> usize
pub fn filled(&self) -> &[u8]
pub fn unfilled<'this>(&'this mut self) -> BorrowCursor<'this, 'a>
pub fn clear(&mut self) -> &mut Self
pub unsafe fn set_init(&mut self, n: usize) -> &mut Self
}
pub struct BorrowCursor<'buf, 'data>
impl<'buf, 'data> BorrowCursor<'buf, 'data> {
pub fn clone<'this>(&'this mut self) -> BorrowCursor<'this, 'data>
pub fn capacity(&self) -> usize
pub fn written(&self) -> usize
pub fn init_ref(&self) -> &[u8]
pub fn init_mut(&mut self) -> &mut [u8]
pub fn uninit_mut(&mut self) -> &mut [MaybeUninit<u8>]
pub unsafe fn as_mut(&mut self) -> &mut [MaybeUninit<u8>]
pub unsafe fn advance(&mut self, n: usize) -> &mut Self
pub fn ensure_init(&mut self) -> &mut Self
pub unsafe fn set_init(&mut self, n: usize) -> &mut Self
pub fn append(&mut self, buf: &[u8])
}
```
## TODO
* ~~Migrate non-unix libs and tests~~
* ~~Naming~~
* ~~`BorrowBuf` or `BorrowedBuf` or `SliceBuf`? (We might want an owned equivalent for the async IO traits)~~
* ~~Should we rename the `readbuf` module? We might keep the name indicate it includes both the buf and cursor variations and someday the owned version too. Or we could change it. It is not publicly exposed, so it is not that important~~.
* ~~`read_buf` method: we read into the cursor now, so the `_buf` suffix is a bit weird.~~
* ~~Documentation~~
* Tests are incomplete (I adjusted existing tests, but did not add new ones).
cc https://github.com/rust-lang/rust/issues/78485, https://github.com/rust-lang/rust/issues/94741
supersedes: https://github.com/rust-lang/rust/pull/95770, https://github.com/rust-lang/rust/pull/93359fixes#93305
Fix HorizonOS regression in FileTimes
The changes in #98246 caused a regression for multiple Newlib-based systems. This is just a fix including HorizonOS to the list of targets which require a workaround.
``@AzureMarker`` ``@ian-h-chamberlain``
r? ``@nagisa``
Support setting file accessed/modified timestamps
Add `struct FileTimes` to contain the relevant file timestamps, since
most platforms require setting all of them at once. (This also allows
for future platform-specific extensions such as setting creation time.)
Add `File::set_file_time` to set the timestamps for a `File`.
Implement the `sys` backends for UNIX, macOS (which needs to fall back
to `futimes` before macOS 10.13 because it lacks `futimens`), Windows,
and WASI.
Add `struct FileTimes` to contain the relevant file timestamps, since
most platforms require setting all of them at once. (This also allows
for future platform-specific extensions such as setting creation time.)
Add `File::set_file_time` to set the timestamps for a `File`.
Implement the `sys` backends for UNIX, macOS (which needs to fall back
to `futimes` before macOS 10.13 because it lacks `futimens`), Windows,
and WASI.
remove_dir_all: use fallback implementation on Miri
Fixes https://github.com/rust-lang/miri/issues/1966
The new implementation requires `openat`, `unlinkat`, and `fdopendir`. These cannot easily be shimmed in Miri since libstd does not expose APIs corresponding to them. So for now it is probably easiest to just use the fallback code in Miri. Nobody should run Miri as root anyway...
unix: reduce the size of DirEntry
On platforms where we call `readdir` instead of `readdir_r`, we store
the name as an allocated `CString` for variable length. There's no point
carrying around a full `dirent64` with its fixed-length `d_name` too.
On platforms where we call `readdir` instead of `readdir_r`, we store
the name as an allocated `CString` for variable length. There's no point
carrying around a full `dirent64` with its fixed-length `d_name` too.
Each recursive call was creating an `OsString` for a `&Path`, only for
it to be turned into a `CString` right away. Instead we can directly
pass `.name_cstr()`, saving two allocations each time.
This only affects the `slow` code path, if there is no `dirent.d_type` or if
the type is `DT_UNKNOWN`.
POSIX specifies that calling `unlink()` or `unlinkat(..., 0)` on a directory can
succeed:
> "The _path_ argument shall not name a directory unless the process has
> appropriate privileges and the implementation supports using _unlink()_ on
> directories."
This however can cause orphaned directories requiring an fsck e.g. on Illumos
UFS, so we have to avoid that in the common case. We now just try to recurse
into it first and unlink() if we can't open it as a directory.
unix: Use metadata for `DirEntry::file_type` fallback
When `DirEntry::file_type` fails to match a known `d_type`, we should
fall back to `DirEntry::metadata` instead of a bare `lstat`, because
this is faster and more reliable on targets with `fstatat`.
When `DirEntry::file_type` fails to match a known `d_type`, we should
fall back to `DirEntry::metadata` instead of a bare `lstat`, because
this is faster and more reliable on targets with `fstatat`.
