why we need openat et al in the kernel

[Jim Meyering]

[ I have to preface this by saying I'm not interested in the
  attribute-related semantics of openat, but rather in the
  fd-relative-open--related semantics.  ]

Why do we need openat and related functions in the kernel?

  Without openat-like functions[1], it is impossible to process
  an arbitrarily deep hierarchy (efficiently) or even a single
  arbitrarily-distant file without changing the working directory.
  Even on systems with no PATH_MAX limitation, it is prohibitively
  expensive (O(depth**2)) to process a deep hierarchy without
  changing the working directory or using openat-like functions.

  What's wrong with changing the working directory?

    - it makes the code non-reentrant.  Imagine a function that processes
      a directory hierarchy.  In order to deal with an arbitrarily deep
      hierarchy, it uses open and fchdir as it traverses the tree.  If
      this function is called in a multi-threaded application, other
      threads may find the current working directory changed out from
      under them.

    - it may make it impossible to return to the original working directory.
      For example, consider the command `rm -r /tmp/deep dot-relative'.
      Removing /tmp/deep may require changing into /tmp/deep then
      /tmp/deep/sub, then /tmp/deep/sub/2, ...  If the initial working
      directory was not readable (so couldn't be opened) and getcwd fails,
      then there is no way to return and remove.  Yes, this is contrived :-)

Note that glibc (as of Nov 11, 2005) provides openat et. al., but that
its implementation relies on the /proc file system, which isn't always
usable, even on systems with the required kernel options: e.g., in a
restrictive chroot environment.

Why does openat have to be implemented in the kernel?
Because any emulation cannot be 100% effective.  Gnulib's
save_cwd fails if the working directory is not readable
and too deep for getcwd.  restore_cwd can fail if save_cwd
failed to get a file descriptor and the original directory
is gone or inaccessible.

So, is there any interest in adding these functions,
independent of the attribute-related debate?

Jim

[1] The list of functions Solaris provides: openat, fchownat, fstatat,
futimesat, renameat, unlinkat.  Plus, a library-level function: fdopendir.
In rewriting GNU rm (coreutils/src/remove.c) to use these functions,
I've identified one more that is required: euidaccessat.
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/