Avoid renameat2 on Android

[glbrntt]

Motivation:

Until very recently Android didn't support the renameat2 syscall, this means that on most systems NIO will fail to compile. renameat2 is used to delay the materialization of files on Linux. On Darwin a different code path is used which we can make use of on Android too.

Modifications:

• Update a few if-defs to force Android onto the alternative path

Result:

NIO builds on Android

[glbrntt]

@finagolfin could you verify this fixes your build?

[finagolfin]

I'll apply it locally and let you know in a little bit, thanks.

[finagolfin]

This isn't building either, some issues with optionals, will try modifying this patch and let you know within an hour.

[glbrntt]

This isn't building either, some issues with optionals, will try modifying this patch and let you know within an hour.

Suspect you'll need this commit too: 600a210

[finagolfin]

Another problem is that Bionic doesn't have confstr() at all:

/home/swift-nio/Sources/NIOFileSystem/Internal/System Calls/Syscalls.swift:389:12: error: cannot find 'confstr' in scope
    return confstr(name, buffer, size)                                                                                                                                        ^~~~~~~

I think this pull works to fix the renameat2 issue, but then other compilation errors from your large pull pop up on Android, looking into them now.

[glbrntt]

Another problem is that Bionic doesn't have confstr() at all:

/home/swift-nio/Sources/NIOFileSystem/Internal/System Calls/Syscalls.swift:389:12: error: cannot find 'confstr' in scope
    return confstr(name, buffer, size)                                                                                                                                        ^~~~~~~

We can workaround confstr too as it looks like we're only using it on Darwin.

I think this pull works to fix the renameat2 issue, but then other compilation errors from your large pull pop up on Android, looking into them now.

Great, thanks for looking!

[finagolfin]

I don't see other issues related to renameat2, so I think this pull works. If you can ifdef confstr() out on Android too, please go ahead and merge these changes.

I will work through the remaining compilation issues related to optionals, because the latest Android NDK added nullability annotations all over Bionic that your pull will have to be slightly modified for, and submit another pull for that later.

[glbrntt]

I don't see other issues related to renameat2, so I think this pull works. If you can ifdef confstr() out on Android too, please go ahead and merge these changes.

Great, will do!

I will work through the remaining compilation issues related to optionals, because the latest Android NDK added nullability annotations all over Bionic that your pull will have to be slightly modified for, and submit another pull for that later.

Do you have any timeline on this? We'd like to tag a release and would rather not break you!

[finagolfin]

Turns out this doesn't compile on Android, because rename(from:to:options:) is only defined on Darwin.

[glbrntt]

Oops, yes, this variant is Darwin only as it has the .exclusive flag. I think we can avoid this by not supporting "atomic" file creation on Android when exclusive create is set (when it isn't set we can do a normal rename without the exclusive flag).

[finagolfin]

Do you have any timeline on this?

Most likely later today.

We'd like to tag a release and would rather not break you!

I mostly build trunk right now, so shouldn't be an issue.

[finagolfin]

Applied this locally and all the remaining errors appear to be type and optional mismatches related to slightly different Android declarations, so you can go ahead and merge this when you like.

[Lukasa]

I just want to sync up here a bit: musl also lacks renameat2 as a libc function, but it does have it as a system call. I'm thinking that we'll want to omit the call through libc and instead make the system call directly. @finagolfin do we think that's gonna cause any problems on android?

[finagolfin]

No, I think that should work, as I don't think that syscall is filtered out on Android.

[Lukasa]

See #2628

[finagolfin]

Tried that Musl syscall pull out, seems to work well after adding my patch for the type/optional mismatches and the confstr() commit from this pull, except I see several tests failing because they call linkat(), not sure where that's coming from.

[finagolfin]

I will look into the linkat() testing issues on Android in the next couple days. Meanwhile, if @glbrntt would reduce this pull to the confstr() commit and merge, that would be helpful.

[finagolfin]

Thanks