feat(scanner): add option not to follow symlinks

[lutzky]

Never consider a symlink to be a subdirectory; this prevents hanging when the symlink points to a slow/inaccessible filesystem.

Description

Instead of using is_dir, which follows symlinks, use symlink_metadata, which doesn't.

Motivation and Context

If the current directory has a symlink to a slow filesystem (e.g. an inaccessible network filesystem), is_dir hangs indefinitely, disregarding scan_timeout.

This should help with some instances of #312.

Screenshots (if appropriate):

n/a

How Has This Been Tested?

• I have tested using MacOS
• I have tested using Linux
• I have tested using Windows

I've tested by running starship timings with network disconnected in a directory which includes a symlink to a network filesystem. Without my change it hangs, with my change it doesn't.

Checklist:

• I have updated the documentation accordingly
• I have updated the tests accordingly

[davidkna]

Following symlinks does seem somewhat useful, I assume there is no option to ignore cross-device symlinks only?

[lutzky]

The trick is determining whether or not a symlink is cross-device without calling stat on the destination (because that's the bit that might fail). https://github.com/stemjail/mnt-rs seems to be able to do it, but doesn't appear very portable. WDYT?

[davidkna]

The trick is determining whether or not a symlink is cross-device without calling stat on the destination (because that's the bit that might fail). https://github.com/stemjail/mnt-rs seems to be able to do it, but doesn't appear very portable. WDYT?

Try using systemstat which we already depend on, relevant functions will likely return an error on unsupported operating systems which should be handled.

[lutzky]

I've tried out some stuff in a99335a; WDYT? Reasonable or too brittle?

[lutzky]

There's an issue with systemstat: mounts proceeds to immediately stat each filesystem, causing it to wait on network. 😞

I guess one path forward is submitting a patch to systemstat to make a no-stat-just-the-files option... WDYT?

[davidkna]

I guess one path forward is submitting a patch to systemstat to make a no-stat-just-the-files option... WDYT?

The crate is somewhat minimalistic, so I'm uncertain if the maintainers will be receptive to such a change.

[lutzky]

So, given these limitations, how about we go back to the "don't follow symlinks in this case" option? Perhaps as a configurable?

[davidkna]

I think making it configurable is fine, unless you want to try making the timeout itself more reliable.

[lutzky]

Making the timeout itself more reliable is more challenging (D is called "uninterruptable" for a reason, you'd end up accumulating stale threads, it might be worth doing but would be messy). For now, I've made it configurable. PTAL

[lutzky]

Finally added tests. The failures on nightly are unrelated to my changes, occur on master as well, and should be fixed by time-rs/time#622. Do my changes seem OK now?

[davidkna]

LGTM

[lutzky]

Still blocked on #5532 for passing the checks.

[davidkna]

Thanks for the contribution @lutzky!