limitedReadCloser, introduced in #2271, releases its slot only on Close(). Because io.Copy reads to EOF and returns without calling Close(), any caller that fails to explicitly Close() holds a limiter slot per layer.

Some deadlocks within this repo were caught after v0.21.6 and fixed:

Each fix correctly adds defer r.Close().

External consumers of Compressed() are in the same position. For example, carvel-dev/imgpkg calls layer.Compressed() in a loop in imagetar/tar_writer.go, passes the result to io.Copy, and never closes on success. This is structurally similar to the fix in #2308. imgpkg deadlocks on any bundle with more than four layers when using go-containerregistry ≥ v0.21.6.

While these missing Close() calls were arguably bugs on their own, it would be better if the consequence were not a deadlock, just briefly leaked resources. Although Compressed returns an io.ReadCloser, the contract of the method should probably not be to expect a deadlock if the consumer fails to call Close on all paths.

Possible improvement to limitedReadCloser: Once Read() returns EOF (or any other non-nil error), the download is no longer in flight and the slot should be freed. Adding a Read() override to limitedReadCloser would fix this for all current and future consumers without breaking any existing caller:

func (l *limitedReadCloser) Read(p []byte) (int, error) {
    n, err := l.ReadCloser.Read(p)
    if err != nil {
        l.once.Do(l.release)
    }
    return n, err
}