Auto merge of #51050 - symphorien:fstatat, r=kennytm
std::fs::DirEntry.metadata(): use fstatat instead of lstat when possible
When reading a directory with `read_dir`, querying metadata for a resulting `DirEntry` is done by building the whole path and then `lstat`ing it, which requires the kernel to resolve the whole path. Instead, one
can use the file descriptor to the enumerated directory and use `fstatat`. This make the resolving step
unnecessary.
This PR implements using `fstatat` on linux, android and emscripten.
## Compatibility across targets
`fstatat` is POSIX.
* Linux >= 2.6.19 according to https://linux.die.net/man/2/fstatat
* android according to https://android.googlesource.com/platform/bionic/+/master/libc/libc.map.txt#392
* emscripten according to https://github.com/kripken/emscripten/blob/
7f89560101843198787530731f40a65288f6f15f/system/include/libc/sys/stat.h#L76
The man page says "A similar system call exists on Solaris." but I haven't found it.
## Compatibility with old platforms
This was introduced with glibc 2.4 according to the man page. The only information I could find about the minimal version of glibc rust must support is this discussion https://internals.rust-lang.org/t/bumping-glibc-requirements-for-the-rust-toolchain/5111/10
The conclusion, if I understand correctly, is that currently rust supports glibc >= 2.3.4 but the "real" requirement is Centos 5 with glibc 2.5. This PR would make the minimal version 2.4, so this should be fine.
## Benefit
I did the following silly benchmark:
```rust
use std::io;
use std::fs;
use std::os::linux::fs::MetadataExt;
use std::time::Instant;
fn main() -> Result<(), io::Error> {
let mut n = 0;
let mut size = 0;
let start = Instant::now();
for entry in fs::read_dir("/nix/store/.links")? {
let entry = entry?;
let stat = entry.metadata()?;
size += stat.st_size();
n+=1;
}
println!("{} files, size {}, time {:?}", n, size, Instant::now().duration_since(start));
Ok(())
}
```
On warm cache, with current rust nightly:
```
1014099 files, size
76895290022, time Duration { secs: 2, nanos:
65832118 }
```
(between 2.1 and 2.9 seconds usually)
With this PR:
```
1014099 files, size
76895290022, time Duration { secs: 1, nanos:
581662953 }
```
(1.5 to 1.6 seconds usually).
approximately 40% faster :)
On cold cache there is not much to gain because path lookup (which we spare) would have been a cache hit:
Before
```
1014099 files, size
76895290022, time Duration { secs: 391, nanos:
739874992 }
```
After
```
1014099 files, size
76895290022, time Duration { secs: 388, nanos:
431567396 }
```
## Testing
The tests were run on linux `x86_64`
```
python x.py test src/tools/tidy
./x.py test src/libstd
```
and the above benchmark.
I did not test any other target.