Discussion: Guidelines for printing to stdout #3093
Description
TLDR: What should we recommend for printing to stdout?
print!/println!?write!/writeln!to locked stdout?write!/writeln!on aBufWriter?- Some other solution?
Printing to stdout sounds simple, but Rust (rightfully) exposes a lot of complexity that other languages attempt to hide. We have several reasons to care about this complexity:
- We want to gracefully handle a variety of edge cases (such as Panic in multiple programs when redirecting stdout to /dev/full #2925).
- We want to make the utils as fast or faster than GNU (or at the very least in the same ballpark as GNU) and IO can be a significant bottleneck. This can really make a difference, for instance in this new PR: join: improve performance #3092
The print!/println! macros are relatively unsuitable for both of these cases. It will lock stdout for every call and only does line buffering by default (I believe). Additionally, it will panic on a write error, so handling edge cases becomes difficult (if not impossible).
Click to see the function that `println!` calls
Source: https://doc.rust-lang.org/src/std/io/stdio.rs.html
/// Write `args` to the capture buffer if enabled and possible, or `global_s`
/// otherwise. `label` identifies the stream in a panic message.
///
/// This function is used to print error messages, so it takes extra
/// care to avoid causing a panic when `local_s` is unusable.
/// For instance, if the TLS key for the local stream is
/// already destroyed, or if the local stream is locked by another
/// thread, it will just fall back to the global stream.
///
/// However, if the actual I/O causes an error, this function does panic.
fn print_to<T>(args: fmt::Arguments<'_>, global_s: fn() -> T, label: &str)
where
T: Write,
{
if OUTPUT_CAPTURE_USED.load(Ordering::Relaxed)
&& OUTPUT_CAPTURE.try_with(|s| {
// Note that we completely remove a local sink to write to in case
// our printing recursively panics/prints, so the recursive
// panic/print goes to the global sink instead of our local sink.
s.take().map(|w| {
let _ = w.lock().unwrap_or_else(|e| e.into_inner()).write_fmt(args);
s.set(Some(w));
})
}) == Ok(Some(()))
{
// Succesfully wrote to capture buffer.
return;
}
if let Err(e) = global_s().write_fmt(args) {
panic!("failed printing to {}: {}", label, e);
}
}There is a an alternative using Write on StdoutLock (possibly wrapped in a BufWriter). This gives us much more control with locking, buffering and error handling as write! returns a result. However, this imposes a surprisingly heavy maintenance burden.
The current version of ls uses the BufWriter approach and illustrates this burden nicely. First, the BufWriter is passed to many functions, which is not horrible, but becomes quite annoying. Second, the errors were reported in the wrong order, because they were written to an unbuffered stderr, but interleaved with a buffered stdout. The output that was generated before an error was therefore buffered, then the error was written and only afterwards was the buffered output written to the terminal. Currently, this is fixed by explicitly flushing the output before writing an error. This works, but is not ideal and not something I'd want to see implemented in every util.
More info on this issue for
lscan be found in #2809 and #2785
So I'd like to discuss this and then set our conclusions as guidelines for all utils in the project. Any ideas are welcome! Also, if I got anything wrong with this post please correct me! I would also love to hear how other Rust programs deal with this issue.