FILE's lock is now always .#FILE and may be a regular file.

* etc/NEWS: Document this.
* nt/inc/unistd.h (O_NOFOLLOW): New macro.
* src/filelock.c: Include <c-ctype.h>.
(MAX_LFINFO): New top-level constant.
(lock_info_type): Remove members pid, boot_time.  Add members at,
dot, colon.  Change user member to be the entire buffer, not a
pointer.  This allows us to handle the case where a foreign
pid or boot time exceeds the local range.  All uses changed.
(LINKS_MIGHT_NOT_WORK): New constant.
(FREE_LOCK_INFO): Remove, as the pieces no longer need freeing.
(defined_WINDOWSNT): Remove.
(MAKE_LOCK_NAME, file_in_lock_file_name):
Always use .#FILE (not .#-FILE) for the file lock,
even if it is a regular file.
(rename_lock_file): New function.
(create_lock_file): Use it.
(create_lock_file, read_lock_data):
Prefer a symbolic link for the lock file, falling back on a
regular file if symlinks don't work.  Do not try to create
symlinks on MS-Windows, due to security hassles.  Stick with
POSIXish functions (open, read, write, close, fchmod, readlink, symlink,
link, rename, unlink, mkstemp) when creating locks, as a GNUish
host may be using a Windowsish file system, and cannot use
MS-Windows-only system calls.  Fall back on mktemp if mkstemp
doesn't work.  Don't fail merely because of a symlink-contents
length limit in the current file system; fall back on regular
files.  Increase the symlink contents length limit to 8 KiB, this
should be big enough for any real use and doesn't crunch the
stack.
(create_lock_file, lock_file_1, read_lock_data):
Simplify allocation of lock file buffers now that they fit in 8 KiB.
(lock_file_1): Return error number, not bool.  All callers changed.
(ELOOP): New macro, if not already defined.
(read_lock_data): Return size of lock file contents, not Lisp object.
All callers changed.  Handle a race condition if some other process
replaces a regular-file lock with a symlink lock or vice versa,
while we're trying to read the lock.
(current_lock_owner): Parse contents more carefully, to help avoid
confusing a regular-file lock with some other application's use
of the file.  Check for lock file contents being too long, or
not parsing correctly.
(current_lock_owner, lock_file):
Allow foreign pid and boot times that exceed the local range.
(current_lock_owner, lock_if_free, lock_file):
Simplify allocation of lock file contents.
* src/w32.c (sys_rename_replace): New function, containing most of
the contents of the old sys_rename.
(sys_rename): Use it.
(fchmod): New dummy function.
* src/w32.h (sys_rename_replace, fchmod): New decls.

Fixes: debbugs:13807

1 files changed, 52 insertions, 0 deletions

diff --git a/src/ChangeLog b/src/ChangeLog
index c4b4703e59d..af03fa54291 100644
--- a/src/ChangeLog
+++ b/src/ChangeLog
@@ -1,3 +1,55 @@
+2013-03-05  Paul Eggert  <eggert@cs.ucla.edu>
+
+        FILE's lock is now always .#FILE and may be a regular file (Bug#13807).
+        * filelock.c: Include <c-ctype.h>.
+        (MAX_LFINFO): New top-level constant.
+        (lock_info_type): Remove members pid, boot_time.  Add members at,
+        dot, colon.  Change user member to be the entire buffer, not a
+        pointer.  This allows us to handle the case where a foreign
+        pid or boot time exceeds the local range.  All uses changed.
+        (LINKS_MIGHT_NOT_WORK): New constant.
+        (FREE_LOCK_INFO): Remove, as the pieces no longer need freeing.
+        (defined_WINDOWSNT): Remove.
+        (MAKE_LOCK_NAME, file_in_lock_file_name):
+        Always use .#FILE (not .#-FILE) for the file lock,
+        even if it is a regular file.
+        (rename_lock_file): New function.
+        (create_lock_file): Use it.
+        (create_lock_file, read_lock_data):
+        Prefer a symbolic link for the lock file, falling back on a
+        regular file if symlinks don't work.  Do not try to create
+        symlinks on MS-Windows, due to security hassles.  Stick with
+        POSIXish functions (open, read, write, close, fchmod, readlink, symlink,
+        link, rename, unlink, mkstemp) when creating locks, as a GNUish
+        host may be using a Windowsish file system, and cannot use
+        MS-Windows-only system calls.  Fall back on mktemp if mkstemp
+        doesn't work.  Don't fail merely because of a symlink-contents
+        length limit in the current file system; fall back on regular
+        files.  Increase the symlink contents length limit to 8 KiB, this
+        should be big enough for any real use and doesn't crunch the
+        stack.
+        (create_lock_file, lock_file_1, read_lock_data):
+        Simplify allocation of lock file buffers now that they fit in 8 KiB.
+        (lock_file_1): Return error number, not bool.  All callers changed.
+        (ELOOP): New macro, if not already defined.
+        (read_lock_data): Return size of lock file contents, not Lisp object.
+        All callers changed.  Handle a race condition if some other process
+        replaces a regular-file lock with a symlink lock or vice versa,
+        while we're trying to read the lock.
+        (current_lock_owner): Parse contents more carefully, to help avoid
+        confusing a regular-file lock with some other application's use
+        of the file.  Check for lock file contents being too long, or
+        not parsing correctly.
+        (current_lock_owner, lock_file):
+        Allow foreign pid and boot times that exceed the local range.
+        (current_lock_owner, lock_if_free, lock_file):
+        Simplify allocation of lock file contents.
+        * w32.c (sys_rename_replace): New function, containing most of
+        the contents of the old sys_rename.
+        (sys_rename): Use it.
+        (fchmod): New dummy function.
+        * w32.h (sys_rename_replace, fchmod): New decls.
+
 2013-03-05  Eli Zaretskii  <eliz@gnu.org>
 
         * bidi.c (bidi_resolve_explicit_1): Don't call CHAR_TO_BYTE or