Re: make openat declarations consistent

[Jim Meyering]

Eric Blake wrote:
> Jim, what do you think of this patch, making it possible to avoid
> including "openat.h" if you only use the POSIX 2008 interfaces?

That's definitely the right thing to do.
Thanks for working on this.

I haven't looked at it carefully yet, but already
your mention of AT_EACCESS inspired me to fix (on coreutils' next)
a long-time infelicity: having to call euidaccess on the
full relative name of each file processed by rm without -f.
The expense of processing an N-component name at each iteration
would make "rm -r" an O(N^2) process.
I'm adjusting it to use faccessat with AT_EACCESS instead.

That brings up the lack of an faccessat replacement in gnulib.
For now, I may just punt and guard the use with #if,
(and added check for existence of that function)

diff --git a/src/remove.c b/src/remove.c
index 32f67a1..99e652a 100644
--- a/src/remove.c
+++ b/src/remove.c
@@ -172,6 +172,13 @@ write_protected_non_symlink (int fd_cwd,
         mess up with long file names). */

   {
+    /* Use faccessat if possible, so as to avoid the expense
+       of processing an N-component name.  */
+#if HAVE_FACCESSAT && AT_EACCESS
+    if (faccessat (fd_cwd, file, W_OK, AT_EACCESS) == 0)
+      return 0;
+#endif
+
     /* This implements #5: */
     size_t file_name_len = strlen (full_name);

But in the long run, we should provide a replacement function,
as we do for all other *at functions.