Fix artifact directories not having traversal permissions

[staticfloat]

It turns out that Windows requires the executable bit set if the BYPASS_TRAVERSE_CHECKING privilege is not attached to a user's account. It's simply more correct to ensure that our directories have this bit set, and unfortunately our filemode() function call is not complete on Windows and does not include this bit, so we manually add it in on Windows.

[giordano]

Is this related to JuliaLang/julia#52272?

[staticfloat]

It may not fix all cases, but it is one way such issues arise.

[StefanKarpinski]

Should we also be setting this bit on directories in Tar.extract?

[staticfloat]

In my tests, Tar.extract didn’t set any permissions on directories on windows.

[ViralBShah]

Shall we merge this and have it backported to the upcoming 1.10.7 (if there is still time?) and 1.11 as well?

[davidanthoff]

Why are we still modifying file permissions on Windows at all? I'm still completely convinced that we should do absolutely nothing to modify file permissions on Windows systems, see JuliaLang/julia#52272 (comment) for my explanation of that.

[topolarity]

Libuv seems to agree with you: https://github.com/libuv/libuv/blob/31ea3411ccab6db63a3f573523be61f80f984d50/src/win/fs.c#L1840-L1849

I'm a bit surprised this even does anything.. I think libuv calls _wchmod, which is documented as ignoring anything that's not _S_IWRITE or _S_IREAD (in particular _S_IEXEC is being set here)

[staticfloat]

Looking at this again, shouldn't this actually be 0o111?

[staticfloat]

I continue to believe that it's better to have chmod() do something than nothing, and that bugs in implementation should just be fixed so the behavior is as close to reasonable as possible. However, I also understand the argument that this is an unusual position for opensource-on-windows-software to take, and I could see the argument for reverting this behavior in future releases of Julia.

However, at least for the present releases, I think this change is purely an improvement. (filemode() not returning +x on Windows is pretty clearly a bug, working around that seems like a fine thing to do. Alternatively, fixing filemode() could also be a good thing to do, but that's more invasive.)

[topolarity]

I continue to believe that it's better to have chmod() do something than nothing...
However, at least for the present releases, I think this change is purely an improvement.

I'm confused.. maybe I chased through the source code wrong, but based on the libuv implementation it really looks like the chmod call (the one implemented by Microsoft) that this is passed to ought to do nothing with the new flag.

[staticfloat]

I'm confused.. maybe I chased through the source code wrong, but based on the libuv implementation it really looks like the chmod call (the one implemented by Microsoft) that this is passed to ought to do nothing with the new flag.

You want to take a look at this function, our fork has diverged from upstream libuv on this, because of work that Jameson and I did.