This changes the behavior of rename-file etc. slightly.
The old behavior mostly disagreed with the documentation, and had
a race condition bug that could allow attackers to modify victims'
write-protected directories (Bug#27986).
* doc/lispref/files.texi (Changing Files): Document that in
rename-file etc., NEWFILE is special if it is a directory name.
* etc/NEWS: Document the change in behavior.
* src/fileio.c (directory_like): Remove. All uses removed.
(expand_cp_target): Test only whether NEWNAME is a directory name,
not whether it is currently a directory. This avoids a race.
(Fcopy_file, Frename_file, Fadd_name_to_file, Fmake_symbolic_link):
Document behavior if NEWNAME is a directory name.
(Frename_file): Simplify now that the destdir behavior occurs
only when NEWNAME is a directory name.
* test/lisp/net/tramp-tests.el (tramp-test11-copy-file)
(tramp-test12-rename-file, tramp--test-check-files):
Adjust tests to match new behavior.
This partially reverts my 2016-05-30 patch. Apparently MinGW64
still requires pacifications that GCC 7.1.1 x86-64 (Fedora 26)
does not. Also, pacify tparam.c, which isn’t used on Fedora.
* lib-src/etags.c (process_file_name, TeX_commands):
* src/buffer.c (fix_overlays_before):
* src/data.c (Fmake_variable_buffer_local, cons_to_unsigned)
(cons_to_signed):
* src/editfns.c (Ftranslate_region_internal):
Prefer UNINIT to some stray value, as this simplifies
code-reading later.
* src/eval.c (CACHEABLE): New macro.
(internal_lisp_condition_case): Use it.
* src/tparam.c (tparam1): Use FALLTHROUGH to pacify GCC.
* src/window.c (window_scroll_pixel_based): If screen position is
to be preserved, make sure its recorded Y coordinate is outside
the scroll margin. (Bug#28342)
* src/w32font.c (SUBRANGE): Use unsigned arithmetic for
bit-shifting, to avoid compiler warnings.
(w32font_text_extents): Tell GCC NGLYPHS is non-negative, to avoid
a warning. For details of the warning, see
http://lists.gnu.org/archive/html/emacs-devel/2017-09/msg00093.html.
* src/term.c (keys) [WINDOWSNT]: Don't define, as it is not used
in that build.
* src/sound.c (sound_perror): Ifdef away on WINDOWSNT, as this
function is not used in that build.
* configure.ac: Disable -Wsuggest-attribute=format on MS-Windows.
* src/fileio.c (directory_file_name) [DOS_NT]: Fix the DOS_NT case
to be consistent with last change.
* test/src/fileio-tests.el (fileio-tests--odd-symlink-chars):
Disable on MS-Windows.
(fileio-tests--directory-file-name-dos-nt)
(fileio-tests--file-name-as-directory-dos-nt): New tests.
* src/fileio.c (directory_file_name): For "///" and longer,
return "/", not "//", as per POSIX.
* test/src/fileio-tests.el (fileio-tests--directory-file-name)
(fileio-tests--file-name-as-directory): New tests.
Problem reported by Martin Rudalics in:
http://lists.gnu.org/archive/html/emacs-devel/2017-09/msg00014.html
* src/xterm.c (xm_scroll_callback, xaw_jump_callback)
(x_set_toolkit_scroll_bar_thumb)
(x_set_toolkit_horizontal_scroll_bar_thumb):
Go back to using ‘float’ temporaries rather than ‘double’.
Although quite possibly this masks an underlying bug,
we lack time to look into that now.
* src/macfont.h (CGContextSetFontSmoothingStyle): Remove version
check.
* src/macfont.m (macfont_draw): Remove version check, and test for
existence of CGContextSetFontSmoothingStyle.
* src/xterm.c (handle_one_xevent): Restrict earlier fix of
Bug#24963 and Bug#25887 to avoid that a non-GTK Emacs won't
react to state changes received via ConfigureNotify.
Check for a pointer anywhere within the object, as opposed to just
the start of the object. This is needed for gcc -Os -flto on
x86-64 (Bug#28213). This change means that the garbage collector
is more conservative, and will incorrectly keep objects that it
does not need to, but that is better than incorrectly discarding
objects that should be kept.
* src/alloc.c (ADVANCE, VINDEX): Now functions, not macros;
this is easier to debug.
(setup_on_free_list): Rename from SETUP_ON_FREE_LIST.
Now a function with two args, not a macro with three.
All callers changed.
(live_string_holding, live_cons_holding, live_symbol_holding)
(live_misc_holding, live_vector_holding, live_buffer_holding):
New functions, which check for any object containing the addressed
byte, not just for an object at the given address.
(live_string_p, live_cons_p, live_symbol_p, live_misc_p)
(live_vector_p, live_buffer_p):
Redefine in terms of the new functions.
(live_float_p): Refactor slightly to match the new functions.
(mark_maybe_object, mark_maybe_pointer): Use the new functions.
Don’t bother checking mark bits, as mark_object already does that,
and omitting the checks here simplifies the code. Although
mark_maybe_object can continue to insist that tagged pointers
still address the start of the object, mark_maybe_pointer now is
more conservative and checks for pointers anywhere into an object.
This is needed for gcc -Os -flto on x86-64; otherwise, GC misses part
of the stack when scanning for heap roots, causing Emacs to crash
later (Bug#28213). The problem is that Emacs's hack for getting an
address near the stack top does not work when link-time optimization
moves stack variables around.
* configure.ac (HAVE___BUILTIN_FRAME_ADDRESS): New macro.
* lib-src/make-docfile.c (DEFUN_noinline): New constant.
(write_globals, scan_c_stream): Support noinline.
* src/alloc.c (NEAR_STACK_TOP): New macro.
(SET_STACK_TOP_ADDRESS): Use it.
(flush_stack_call_func, Fgarbage_collect): Now noinline.
This is needed for gcc -Os -flto on x86-64 (Bug#28213).
* src/emacs.c (main): Align stack-bottom variable as a pointer,
since mark_memory requires this.
* src/gtkutil.c (xg_initialize): Don’t set deprecated and ignored
gtk-menu-bar-accel setting in new versions of GTK+. Use g_object_set
instead of deprecated gtk_settngs_set_string_property otherwise.
* src/gtkutil.c (my_log_handler): Don’t define in new versions of
GTK+.
(xg_set_geometry): Always use gtk_window_move in new versions of GTK+.
* src/xterm.c (syms_of_xterm): Document that x-gtk-use-window-move
is ignored.
* lisp/subr.el (x-gtk-use-window-move): Make obsolete.
This lets Emacs deal with arbitrary local symlinks without
mishandling their contents (Bug#28156). For example,
(progn (shell-command "ln -fs '~' 'x'") (rename-file "x" "/tmp/x"))
now consistently creates a symbolic link from '/tmp/x' to '~'.
Formerly, it did that only if the working directory was on the
same filesystem as /tmp; otherwise, it expanded the '~' to
the user's home directory.
* lisp/dired.el (dired-get-filename): Use files--name-absolute-system-p
instead of rolling our own code.
* lisp/files.el (files--name-absolute-system-p): New function.
(file-truename, file-chase-links): Use it to avoid mishandling
symlink contents that begin with ~.
(copy-directory, move-file-to-trash):
Use concat rather than expand-file-name, to avoid mishandling
symlink contents that begin with ~.
* src/fileio.c (Fmake_symbolic_link): Do not expand leading "~" in the
target unless interactive. Strip leading "/:" if interactive.
(emacs_readlinkat): Do not prepend "/:" to the link target if
it starts with "/" and contains ":" before NUL.
* test/src/fileio-tests.el (try-link): Rename from try-char,
and accept a string instead of a char. All uses changed.
(fileio-tests--symlink-failure): Also test leading ~, and "/:",
to test the new behavior.
* doc/lispref/files.texi (File Attributes):
Document file-attributes atomicity.
* etc/NEWS: Document the fix.
* src/dired.c (file_attributes): New args DIRNAME and FILENAME,
for diagnostics. All callers changed. On platforms like
GNU/Linux that support O_PATH, fix a race condition in
file-attributes and similar functions, so that these functions do
not return nonsense if a directory entry is replaced while getting
its attributes. On non-GNU platforms, do a better (though not
perfect) job of detecting the race, and return nil if detected.
* src/fileio.c (expand_and_dir_to_file): Simplify by omitting 2nd
argument, since in practice it always has the default value. All
callers changed. Prefer C99 style decls in nearby code.
* src/xterm.c (xm_scroll_callback, xaw_jump_callback)
(x_set_toolkit_scroll_bar_thumb)
(x_set_toolkit_horizontal_scroll_bar_thumb): Prefer ‘double’ to
‘float’ for individual local floating-point temporaries.
