Use fstatat to check long path sizes in fcntl::readlinkat

[ghost]

Without this, the lstat call ignores the dirfd and checks the wrong file. I wanted to write a unit test for this but I'm not sure how to do it reliably, AFAIK the maximum path size can vary between filesystems.

[asomers]

Can you explain in a little more detail what the problem is? Or better yet, show that unit test you came up with, even if it doesn't work everywhere.

[ghost]

The readlink and readlinkat syscalls read out a symlink to a fixed size buffer. If the buffer size is too small, then the syscall has to be performed again with a larger buffer. As a first step to mitigate this, inner_readlink will try to call lstat to read the size of the symlink from the filesystem. If the lstat call fails then it will fall back to enlarging the buffer in a loop and retrying until it succeeds.

This will work fine if we called it as fcntl::readlink, but the problem comes in if this is a call to fcntl::readlinkat. The readlinkat syscall expects a dirfd and a relative path. If we only call lstat then the dirfd is ignored and lstat will be trying to access the wrong file. On second thought, it's not really feasible to write a unit test for this since this can never cause the whole function to fail, it just causes it to be slower in some circumstances, since it will always enter the loop if the lstat call fails. The solution to this is: When we have a dirfd, use the corresponding stat-family syscall that takes a dirfd, which is fstatat.

There is another special case on Linux, where readlinkat will read the symlink straight from the fd if it's passed an empty string as a path. fstatat has similar behavior, but only if it's passed an AT_EMPTY_PATH flag, so we have to reconcile that here too.

Reference:

• https://man7.org/linux/man-pages/man2/readlinkat.2.html
• https://man7.org/linux/man-pages/man2/fstatat.2.html

[asomers]

This looks good. I just need to figure out why CI is hanging.

[asomers]

CI is hanging due to a regression in rustup. rust-lang/rustup#2774

[asomers]

If you rebase on main, CI should be fixed.

[asomers]

bors r+

[bors]

Build succeeded:

• Android aarch64
• Android arm
• Android armv7
• Android i686
• Android x86_64
• FreeBSD amd64 & i686
• iOS
• Linux aarch64
• Linux arm gnueabi
• Linux arm-musleabi
• Linux armv7 gnueabihf
• Linux i686
• Linux i686 musl
• Linux MIPS
• Linux MIPS64
• Linux MIPS64 el
• Linux mipsel
• Linux powerpc
• Linux powerpc64le
• Linux s390x
• Linux x86_64
• Linux x86_64 musl
• Minver
• NetBSD x86_64
• OSX i686
• OSX x86_64
• Redox x86_64
• Rust Stable