Crash (abort) doing completion on filenames >= 256 bytes

[g2p]

This is possible on bcachefs (stat --file-system /mnt/bcachefs says 512 bytes in a dirent, koverstreet/bcachefs#733 has a test script; not sure if that limit includes a final NUL).

Create the file like this:

i=256; touch "$(printf "%05d" ${i})$(head -c $((${i}-5)) /dev/zero | tr '\0' 'a')"

The backtrace looks like (as of 49023ad):

   5: fish::common::cstr2wcstring
             at ./src/common.rs:1185:18
   6: fish::wutil::dir_iter::DirIter::next
             at ./src/wutil/dir_iter.rs:282:27
   7: fish::wildcard::expander::WildCardExpander::expand_trailing_slash
             at ./src/wildcard.rs:667:41
   8: fish::wildcard::expander::WildCardExpander::expand
             at ./src/wildcard.rs:528:21

Apparently there are 256 byte slices in dirent structs returned from the libc crate: https://docs.rs/libc/latest/libc/struct.dirent.html
gdb can printf the rest of the string so the slice length is incorrect.
std's own read_dir will probably work better.

[g2p]

The issues using dirent from the libc crate are discussed here: rust-lang/libc#2669

The stdlib now does this (before &raw there were complex macros):

            let name = CStr::from_ptr((&raw const (*entry_ptr).d_name).cast());
            let name_bytes = name.to_bytes();

And the alternative is to wrap stdlib and re-insert dotdot when needed.

[g2p]

&raw needs Rust 1.82 and fish seems to limit itself to 1.70 though.
https://blog.rust-lang.org/2024/10/17/Rust-1.82.0.html#native-syntax-for-creating-a-raw-pointer

Rust version discussion: #10230

[g2p]

Nix, which fish already uses, also has a dir module that would simplify the wutil::dir_iter code. std is more correct wrt unsafe code but doesn't expose the dirfd interface or rewinding. And rustix's Dir::read looks even better, it doesn't inherit problems from the libc crate's struct because the default backend does the syscalls directly.

[faho]

Before you go off into the weeds here: We try to go light on dependencies, there's no need to exhaustively survey everything.

This seems like it should be fixed for 4.0.1, which means we want to keep it simple.

The nix module would be a choice, as would be just Writing The Code for this specific case.

[g2p]

The nix module would be a choice, as would be just Writing The Code for this specific case.

Yeah. I'm just annoyed that it's a choice between &raw (clear and correct, but you'll need to bump to rust 1.82) and doing something that goes against unsafe guidelines (directly or through nix).

[faho]

We're not gonna bump the MSRV for 4.0.x, so the latter it is, apparently.

[g2p]

I've sent a patch for the minimal thing.

[faho]

And in 4.0.1 as f336cf5.

Thanks!