ls -i: print consistent inode numbers also for mount points

On most unix- and linux-based kernels, ls -i DIR_CONTAINING_MOUNT_POINT
would print the wrong inode number for any entry that is a mount point.
It would do that by relying on readdir's dirent.d_ino values, while
most readdir implementations return the inode number of the underlying,
inaccessible directory.  Thus, it is not consistent with what you'd
get when applying stat to the same entry.  This bug led to surprising
results like "ls -i" and "ls -i --color" printing different numbers (ls
must usually "stat" a file to colorize its name).  This change makes it
so that on offending systems, ls must stat non-command-line-arguments
for which otherwise it would be able to use "for free" dirent.d_ino
values.  Regardless of this change, ls is already required to stat every
command-line argument.  Note: versions of GNU ls prior to coreutils-6.0
did not perform the invalid optimization, and hence always printed
correct inode numbers.  Thus, for the sake of correctness, ls -i is
forgoing the readdir optimization, for any kernel (including linux!)
with POSIX-nonconforming readdir.  Note that currently, only Cygwin has
been agile enough to conform.

* src/ls.c (RELIABLE_D_INO): Define.
(print_dir): Use it.
For plenty of discussion, see this long thread:
http://thread.gmane.org/gmane.comp.gnu.coreutils.bugs/14020
This bug was introduced by the 2006-02-26 commit, 33eb3efe:
"In ls, avoid calling stat for --inode (-i), when possible."
* tests/ls/readdir-mountpoint-inode: New test.
* tests/Makefile.am (TESTS): Add it.
* tests/ls/stat-vs-dirent: Don't suppress failure of this test,
now that ls -i is fixed.  Though note that it doesn't test well,
since it compares only the always-stat'd command-line arguments.
* NEWS (Bug fixes): Mention it.

1 files changed, 22 insertions, 1 deletions

diff --git a/src/ls.c b/src/ls.c
index 6316dfa68..553090dc5 100644
--- a/src/ls.c
+++ b/src/ls.c
@@ -126,6 +126,26 @@
    Subtracting doesn't always work, due to overflow.  */
 #define longdiff(a, b) ((a) < (b) ? -1 : (a) > (b))
 
+/* Unix-based readdir implementations have historically returned a dirent.d_ino
+   value that is sometimes not equal to the stat-obtained st_ino value for
+   that same entry.  This error occurs for a readdir entry that refers
+   to a mount point.  readdir's error is to return the inode number of
+   the underlying directory -- one that typically cannot be stat'ed, as
+   long as a file system is mounted on that directory.  RELIABLE_D_INO
+   encapsulates whether we can use the more efficient approach of relying
+   on readdir-supplied d_ino values, or whether we must incur the cost of
+   calling stat or lstat to obtain each guaranteed-valid inode number.  */
+
+#ifndef READDIR_LIES_ABOUT_MOUNTPOINT_D_INO
+# define READDIR_LIES_ABOUT_MOUNTPOINT_D_INO 1
+#endif
+
+#if READDIR_LIES_ABOUT_MOUNTPOINT_D_INO
+# define RELIABLE_D_INO(dp) NOT_AN_INODE_NUMBER
+#else
+# define RELIABLE_D_INO(dp) D_INO (dp)
+#endif
+
 #if ! HAVE_STRUCT_STAT_ST_AUTHOR
 # define st_author st_uid
 #endif
@@ -2501,7 +2521,8 @@ print_dir (char const *name, char const *realname, bool command_line_arg)
 # endif
                 }
 #endif
-              total_blocks += gobble_file (next->d_name, type, D_INO (next),
+              total_blocks += gobble_file (next->d_name, type,
+                                           RELIABLE_D_INO (next),
                                            false, name);
 
               /* In this narrow case, print out each name right away, so