mingw and same-inode

[Eric Blake]

I'm currently testing these two patches, as mingw prerequisites before I can 
get linkat() working.  In particular, mingw is lousy at SAME_INODE, since all 
three of [fl]stat produce st_ino == 0 for all files (then again, mingw never 
claimed POSIX compliance!).  Code was always taking the identical-file path, 
even for distinct files.

I'm also preparing a followup patch for coreutils usage of SAME_INODE.  
Thoughts before I apply this?

>From 0719361730021a335bb932bfe4ce2f0524f43649 Mon Sep 17 00:00:00 2001
From: Eric Blake <address@hidden>
Date: Wed, 23 Sep 2009 14:13:00 -0600
Subject: [PATCH 1/2] lstat: avoid mingw compilation error

The openat module required lstat, but then repeated the checks
for whether lstat worked.  When cross-compiling to mingw, this
led to a wrong answer and tried compiling lstat.c, even though it
was not necessary, with a result of a compiler warning about
undeclared lstat.

* m4/lstat.m4 (gl_FUNC_LSTAT): Avoid duplicate calls to
AC_FUNC_LSTAT_FOLLOWS_SLASHED_SYMLINK, and deal with missing
lstat ourselves.
* lib/lstat.c [!HAVE_LSTAT]: Do nothing if <sys/stat.h> override
was adequate.
* m4/sys_stat_h.m4 (gl_HEADER_SYS_STAT_H): Let lstat module handle
the checks for lstat.

Signed-off-by: Eric Blake <address@hidden>
---
 ChangeLog        |    9 +++++++++
 lib/lstat.c      |   23 ++++++++++++++++-------
 m4/lstat.m4      |   10 ++++++----
 m4/sys_stat_h.m4 |   12 +-----------
 4 files changed, 32 insertions(+), 22 deletions(-)

diff --git a/ChangeLog b/ChangeLog
index 8a4aebf..7ed6953 100644
--- a/ChangeLog
+++ b/ChangeLog
@@ -1,5 +1,14 @@
 2009-09-23  Eric Blake  <address@hidden>

+       lstat: avoid mingw compilation error
+       * m4/lstat.m4 (gl_FUNC_LSTAT): Avoid duplicate calls to
+       AC_FUNC_LSTAT_FOLLOWS_SLASHED_SYMLINK, and deal with missing
+       lstat ourselves.
+       * lib/lstat.c [!HAVE_LSTAT]: Do nothing if <sys/stat.h> override
+       was adequate.
+       * m4/sys_stat_h.m4 (gl_HEADER_SYS_STAT_H): Let lstat module handle
+       the checks for lstat.
+
        link: fix test failure on Solaris 9
        * lib/link.c (rpl_link): Don't assume link will catch bogus
        trailing slash on source.
diff --git a/lib/lstat.c b/lib/lstat.c
index a05f674..27a0cb5 100644
--- a/lib/lstat.c
+++ b/lib/lstat.c
@@ -20,11 +20,18 @@

 #include <config.h>

+#if !HAVE_LSTAT
+/* On systems that lack symlinks, our replacement <sys/stat.h> already
+   defined lstat as stat, so there is nothing further to do other than
+   avoid an empty file.  */
+typedef int dummy;
+#else /* HAVE_LSTAT */
+
 /* Get the original definition of lstat.  It might be defined as a macro.  */
-#define __need_system_sys_stat_h
-#include <sys/types.h>
-#include <sys/stat.h>
-#undef __need_system_sys_stat_h
+# define __need_system_sys_stat_h
+# include <sys/types.h>
+# include <sys/stat.h>
+# undef __need_system_sys_stat_h

 static inline int
 orig_lstat (const char *filename, struct stat *buf)
@@ -33,10 +40,10 @@ orig_lstat (const char *filename, struct stat *buf)
 }

 /* Specification.  */
-#include <sys/stat.h>
+# include <sys/stat.h>

-#include <string.h>
-#include <errno.h>
+# include <string.h>
+# include <errno.h>

 /* lstat works differently on Linux and Solaris systems.  POSIX (see
    `pathname resolution' in the glossary) requires that programs like
@@ -81,3 +88,5 @@ rpl_lstat (const char *file, struct stat *sbuf)
     }
   return stat (file, sbuf);
 }
+
+#endif /* HAVE_LSTAT */
diff --git a/m4/lstat.m4 b/m4/lstat.m4
index 74c78fe..089d0ff 100644
--- a/m4/lstat.m4
+++ b/m4/lstat.m4
@@ -1,4 +1,4 @@
-# serial 19
+# serial 20

 # Copyright (C) 1997-2001, 2003-2009 Free Software Foundation, Inc.
 #
@@ -12,15 +12,17 @@ AC_DEFUN([gl_FUNC_LSTAT],
 [
   AC_REQUIRE([gl_SYS_STAT_H_DEFAULTS])
   dnl If lstat does not exist, the replacement <sys/stat.h> does
-  dnl "#define lstat stat", and lstat.c does not need to be compiled.
+  dnl "#define lstat stat", and lstat.c is a no-op.
   AC_CHECK_FUNCS_ONCE([lstat])
   if test $ac_cv_func_lstat = yes; then
-    AC_FUNC_LSTAT_FOLLOWS_SLASHED_SYMLINK
-    dnl Note: AC_FUNC_LSTAT_FOLLOWS_SLASHED_SYMLINK does AC_LIBOBJ([lstat]).
+    AC_REQUIRE([AC_FUNC_LSTAT_FOLLOWS_SLASHED_SYMLINK])
     if test $ac_cv_func_lstat_dereferences_slashed_symlink = no; then
+      dnl Note: AC_FUNC_LSTAT_FOLLOWS_SLASHED_SYMLINK does AC_LIBOBJ([lstat]).
       REPLACE_LSTAT=1
     fi
     # Prerequisites of lib/lstat.c.
     AC_REQUIRE([AC_C_INLINE])
+  else
+    HAVE_LSTAT=0
   fi
 ])
diff --git a/m4/sys_stat_h.m4 b/m4/sys_stat_h.m4
index df7b238..147c517 100644
--- a/m4/sys_stat_h.m4
+++ b/m4/sys_stat_h.m4
@@ -1,4 +1,4 @@
-# sys_stat_h.m4 serial 16   -*- Autoconf -*-
+# sys_stat_h.m4 serial 17   -*- Autoconf -*-
 dnl Copyright (C) 2006-2009 Free Software Foundation, Inc.
 dnl This file is free software; the Free Software Foundation
 dnl gives unlimited permission to copy and/or distribute it,
@@ -11,16 +11,6 @@ AC_DEFUN([gl_HEADER_SYS_STAT_H],
 [
   AC_REQUIRE([gl_SYS_STAT_H_DEFAULTS])

-  dnl Check for lstat.  Systems that lack it (mingw) also lack symlinks, so
-  dnl stat is a good replacement.
-  AC_CHECK_FUNCS_ONCE([lstat])
-  if test $ac_cv_func_lstat = yes; then
-    HAVE_LSTAT=1
-  else
-    HAVE_LSTAT=0
-  fi
-  AC_SUBST([HAVE_LSTAT])
-
   dnl For the mkdir substitute.
   AC_REQUIRE([AC_C_INLINE])

--- 
1.6.1.2


>From 5538c910e1acd307a3d3e69f6d80044b1f3d935a Mon Sep 17 00:00:00 2001
From: Eric Blake <address@hidden>
Date: Wed, 23 Sep 2009 14:51:29 -0600
Subject: [PATCH 2/2] same-inode: make SAME_INODE tri-state, to port to mingw

Mingw has the annoying habit (already documented in
doc/posix-functions/*stat) that st_ino is always 0.
This means that naive uses of SAME_INODE(a,b) would
succeed, even on distinct files.

Here's an analysis of all gnulib modules that used
the macro before this commit:

chdir-safer is safe - SAME_INODE protected by HAVE_READLINK
cycle-check - mingw has no dir hard links and no symlinks,
 so no directory cycles can occur, and we should ignore -1
fts - SAME_INODE protected by FTS_DEBUG
hash-triple - using -1 gives more hash collisions, but
 the results are still correct
openat-proc - SAME_INODE protected by stat("/proc/self")
same - no dir cycles, so files are only same with identical name
link-follow.m4 - configure test already correct on mingw
test-canonicalize* - test already passes on mingw
test-[l]stat - test already passes on mingw

* NEWS: Mention this change.
* lib/same-inode.h (same-inode.h): Recognize mingw limitation of
st_ino always being 0.
* lib/cycle-check.h (CYCLE_CHECK_REFLECT_CHDIR_UP): Update caller.
* lib/cycle-check.c (cycle_check): Likewise.
* lib/same.c (same_name): Likewise.

Signed-off-by: Eric Blake <address@hidden>
---
 ChangeLog         |    8 ++++++++
 NEWS              |    3 +++
 lib/cycle-check.c |    5 +++--
 lib/cycle-check.h |    4 ++--
 lib/same-inode.h  |   16 +++++++++++++---
 lib/same.c        |    7 +++++--
 6 files changed, 34 insertions(+), 9 deletions(-)

diff --git a/ChangeLog b/ChangeLog
index 7ed6953..a729fdf 100644
--- a/ChangeLog
+++ b/ChangeLog
@@ -1,5 +1,13 @@
 2009-09-23  Eric Blake  <address@hidden>

+       same-inode: make SAME_INODE tri-state, to port to mingw
+       * NEWS: Mention this change.
+       * lib/same-inode.h (same-inode.h): Recognize mingw limitation of
+       st_ino always being 0.
+       * lib/cycle-check.h (CYCLE_CHECK_REFLECT_CHDIR_UP): Update caller.
+       * lib/cycle-check.c (cycle_check): Likewise.
+       * lib/same.c (same_name): Likewise.
+
        lstat: avoid mingw compilation error
        * m4/lstat.m4 (gl_FUNC_LSTAT): Avoid duplicate calls to
        AC_FUNC_LSTAT_FOLLOWS_SLASHED_SYMLINK, and deal with missing
diff --git a/NEWS b/NEWS
index 62c631f..87fc884 100644
--- a/NEWS
+++ b/NEWS
@@ -6,6 +6,9 @@ User visible incompatible changes

 Date        Modules         Changes

+2009-09-23  same-inode      The macro SAME_INODE is now tri-state, adding -1
+                            for unknown.
+
 2009-09-16  canonicalize-lgpl
                             The include file is changed from "canonicalize.h"
                             to <stdlib.h>.
diff --git a/lib/cycle-check.c b/lib/cycle-check.c
index e120460..a5b6df1 100644
--- a/lib/cycle-check.c
+++ b/lib/cycle-check.c
@@ -1,6 +1,7 @@
 /* help detect directory cycles efficiently

-   Copyright (C) 2003, 2004, 2005, 2006 Free Software Foundation, Inc.
+   Copyright (C) 2003, 2004, 2005, 2006, 2009 Free Software
+   Foundation, Inc.

    This program is free software: you can redistribute it and/or modify
    it under the terms of the GNU General Public License as published by
@@ -62,7 +63,7 @@ cycle_check (struct cycle_check_state *state, struct stat 
const *sb)
   /* If the current directory ever happens to be the same
      as the one we last recorded for the cycle detection,
      then it's obviously part of a cycle.  */
-  if (state->chdir_counter && SAME_INODE (*sb, state->dev_ino))
+  if (state->chdir_counter && SAME_INODE (*sb, state->dev_ino) == 1)
     return true;

   /* If the number of `descending' chdir calls is a power of two,
diff --git a/lib/cycle-check.h b/lib/cycle-check.h
index ee3bf21..ea9b9c4 100644
--- a/lib/cycle-check.h
+++ b/lib/cycle-check.h
@@ -1,6 +1,6 @@
 /* help detect directory cycles efficiently

-   Copyright (C) 2003, 2004, 2006 Free Software Foundation, Inc.
+   Copyright (C) 2003, 2004, 2006, 2009 Free Software Foundation, Inc.

    This program is free software: you can redistribute it and/or modify
    it under the terms of the GNU General Public License as published by
@@ -41,7 +41,7 @@ bool cycle_check (struct cycle_check_state *state, struct 
stat const *sb);
       /* You must call cycle_check at least once before using this macro.  */ \
       if ((State)->chdir_counter == 0)                         \
         abort ();                                              \
-      if (SAME_INODE ((State)->dev_ino, SB_subdir))            \
+      if (SAME_INODE ((State)->dev_ino, SB_subdir) == 1)       \
        {                                                       \
          (State)->dev_ino.st_dev = (SB_dir).st_dev;            \
          (State)->dev_ino.st_ino = (SB_dir).st_ino;            \
diff --git a/lib/same-inode.h b/lib/same-inode.h
index 0632711..4b89dd1 100644
--- a/lib/same-inode.h
+++ b/lib/same-inode.h
@@ -1,6 +1,6 @@
 /* Determine whether two stat buffers refer to the same file.

-   Copyright (C) 2006 Free Software Foundation, Inc.
+   Copyright (C) 2006, 2009 Free Software Foundation, Inc.

    This program is free software: you can redistribute it and/or modify
    it under the terms of the GNU General Public License as published by
@@ -18,8 +18,18 @@
 #ifndef SAME_INODE_H
 # define SAME_INODE_H 1

+/* Perform a tri-state query on whether STAT_BUF_1 and STAT_BUF_2
+   represent the same file.  Return 1 for equal, 0 for distinct, and
+   -1 for indeterminate (the latter is generally possible only on
+   mingw).  Algorithms that use this macro must be prepared to handle
+   the indeterminate case without wrong results.  For example, if an
+   optimization is possible if two files are the same but unsafe if
+   distinct, use SAME_INODE()==1; whereas for an optimization that is
+   possible only for distinct files, use !SAME_INODE().  */
+
 # define SAME_INODE(Stat_buf_1, Stat_buf_2) \
-   ((Stat_buf_1).st_ino == (Stat_buf_2).st_ino \
-    && (Stat_buf_1).st_dev == (Stat_buf_2).st_dev)
+   (((Stat_buf_1).st_ino == (Stat_buf_2).st_ino         \
+     && (Stat_buf_1).st_dev == (Stat_buf_2).st_dev)     \
+    ? 1 - 2 * !(Stat_buf_1).st_ino : 0)

 #endif
diff --git a/lib/same.c b/lib/same.c
index af3a95e..5251fb8 100644
--- a/lib/same.c
+++ b/lib/same.c
@@ -1,7 +1,7 @@
 /* Determine whether two file names refer to the same file.

-   Copyright (C) 1997, 1998, 1999, 2000, 2002, 2003, 2004, 2005, 2006 Free
-   Software Foundation, Inc.
+   Copyright (C) 1997, 1998, 1999, 2000, 2002, 2003, 2004, 2005, 2006,
+   2009 Free Software Foundation, Inc.

    This program is free software: you can redistribute it and/or modify
    it under the terms of the GNU General Public License as published by
@@ -97,6 +97,9 @@ same_name (const char *source, const char *dest)
        }

       same = SAME_INODE (source_dir_stats, dest_dir_stats);
+      if (same < 0)
+       same = (identical_basenames
+               && strcmp (source_basename, dest_basename) == 0);

 #if ! _POSIX_NO_TRUNC && HAVE_PATHCONF && defined _PC_NAME_MAX
       if (same && ! identical_basenames)
-- 
1.6.1.2