This seems to have been designed to minimize changes to struct _dirdesc which used to be public in <dirent.h>. That seems a reasonable thing to do.

Fixing the nanoseconds (also PR 218018) will be a separate diff.

Since you have not committed this yet, another request here: since you made clock_nanosleep() a cancellation point (correctly), please add it to the list in share/man/man3/pthread_testcancel.3.

Hmm, I thought of something else. It would help future extensions and making applications correct if an invalid flags argument (that is, not 0 and not TIMER_ABSTIME) returns error EINVAL instead of behaving like one of the two.

Looks good to me as a separate process.

I like the tests, even though they take a fair bit of time to execute.

This is correct for architectures where int and long are the same size, but otherwise there may still be surprising behaviour with strings like 4294967295.

The example is possible with existing functionality:

The idea looks good to me.

Language fixes from wblock.

In D9096#188667, @emaste wrote:

nftw is POSIX so perhaps slightly preferable. I'm indifferent.

I'm somewhat surprised to see this change, since getty's output is generally only seen by administrators nowadays, so complicated code to improve its cosmetics is less useful.

Apart from the one minor issue I do not see problems (which does not mean there aren't any).

In D9169#189490, @cem wrote:

In D9169#189478, @jilles wrote:

A quick query at https://codesearch.debian.net/ suggests there is quite a bit of code that tries to detect the two qsort_r() variants. It would not surprise me if some high-profile port broke.

Hm, it would surprise me. It should continue to detect at least one variant as working, and using either variant is fine.

I like the idea of handling this the same way as basename()/dirname().

This is not a review of the idea to support both qsort_r() variants like this. At this time I can only say it seems somewhat of a dangerous idea.

I like the idea of adding -R support, but I think the way symlinks are handled is dangerous: with the proposed code, setfacl -R will follow symlinks. This should be changed to match symlink(7), adding -L, -H and -P options.

Defining _POSIX_PRIORITY_SCHEDULING as 0 in sys/sys/unistd.h will tell applications that compile-time support is available but run-time support needs to be verified using sysconf(), and will tell lib/libc/gen/sysconf.c to query for support using sysctl. Therefore, that seems the right solution to me.

Looks good to me otherwise.