comparison README.multi-tty @ 83013:e77d1a63471b

Don't select on stdin unconditionally. src/process.c (init_process): Don't add stdin to input_wait_mask. git-archimport-id: [e-mail]--2004/emacs--multi-tty--0--patch-53

83012:4aa172a45af1

   secondary tty support, but it does not seem to do anything useful.
   (Update: Look again. X unconditionally enables this, maybe that's
   why raw terminal support is broken again.  I really do need to
   understand input.)

** I have seen a case when Emacs with multiple ttys fell into a loop
   eating 100% of CPU time.  Strace showed this loop:

	getpid()                                = 30284
	kill(30284, SIGIO)                      = 0
	--- SIGIO (I/O possible) @ 0 (0) ---

   Update: This may have been caused by checking for nread != 0
   instead of nread > 0 after calling read_socket_hook in
   read_avail_input.

** emacsclient -t from an Emacs term buffer does not work, complains
   about face problems.  This can even lock up Emacs (if the recursive
   frame sets single_kboard).  Update: the face problems are caused by
   bugs in term.el, not in multi-tty.  The lockup is caused by
   single_kboard mode, and is not easily solvable.  The best thing to
   do is to simply refuse to create a tty frame of type `eterm'.

DIARY OF CHANGES
----------------

(ex-TODO items with explanations.)

-- Introduce a new struct for terminal devices.

   (Done, see struct tty_output.  The list of members is not yet
   complete.)

-- Change the bootstrap procedure to initialize tty_list.

   (Done, but needs review.)

-- Change make-terminal-frame to support specifying another tty.

   (Done, new frame parameters: `tty' and `tty-type'.)

-- Implement support for reading from multiple terminals.

   (Done, read_avail_input tries to read from each terminal, until one
   succeeds.  MULTI_KBOARD is not used.  Secondary terminals don't send
   SIGIO!)

   (Update: They do, now.)

   (Update2: After enabling X, they don't.)

-- other-frame should cycle through the frames on the `current'
   terminal only.

   (Done, by trivially modifiying next_frame and prev_frame.)

-- Support different terminal sizes.

   (Done, no problem.)

-- Make sure terminal resizes are handled gracefully.  (Could be
   problematic.)

   (Done.  We don't get automatic SIGWINCH for additional ttys,
   though.)

-- Extend emacsclient to automatically open a new tty when it connects
   to Emacs.

   (Done.  It's an ugly hack, needs more work.)

-- Redisplay must refresh the topmost frame on *all* terminals, not
   just the initial terminal.

   (Done, but introduced an ugly redisplay problems.  Ugh.)

-- Fix redisplay problems.

   (Done; it turned out that the entire Wcm structure must be moved
   inside tty_output.  Why didn't I catch this earlier?)

-- Provide a way for emacsclient to tell Emacs that the tty has been
   resized.

   (Done, simply forward the SIGWINCH signal.)

-- Each keypress should automatically select the frame corresponding
   to the terminal that it was coming from.  This means that Emacs
   must know from which terminal the last keyboard event came from.

   (Done, it was quite simple, the input event system already
   supported multiple frames.)

-- Fix SIGIO issue with secondary terminals.

   (Done, emacsclient signals Emacs after writing to the proxy pseudo
   terminal.  Note that this means that multi-tty does not work with
   raw ttys!)

   (Update: This is bullshit.  There is a read_input_waiting function,
   extend that somehow.)

   (Update of update: The first update was not right either, extending
   read_input_waiting was not necessary.  Secondary ttys do seem to
   send signals on input.)

   (Update^3: Not any more.)

-- Make make-terminal-frame look up the `tty' and `tty-type' frame
   parameters from the currently selected terminal before the global
   default.

   (Done.)

-- Put all cached terminal escape sequences into struct tty_output.
   Currently, they are still stored in global variables, so we don't
   really support multiple terminal types.

   (Done.  It was not fun.)

-- Implement sane error handling after initialization.  (Currently
   emacs exits if you specify a bad terminal type.)  The helpful error
   messages must still be provided when Emacs starts.

   (Done.)

-- Implement terminal deletion, i.e., deleting local frames, closing
   the tty device and restoring its previous state without exiting
   Emacs.

   (Done, but at the moment only called when an error happens during
   initialization.  There is a memory corruption error around this
   somewhere.)  (Update: now it is fully enabled.)

-- Implement automatic deletion of terminals when the last frame on
   that terminal is closed.

   (Done.)

-- Restore tty screen after closing the terminal.

   (Done, we do the same as Emacs 21.2 for all terminals.)

-- 'TERM=dumb src/emacs' does not restore the terminal state.

   (Done.)

-- C-g should work on secondary terminals.

   (Done, but the binding is not configurable.)

-- Deal with SIGHUP in Emacs and in emacsclient.  (After this, the
   server-frames may be removed from server.el.)

   (Done, nothing to do.  It seems that Emacs does not receive SIGHUP
   from secondary ttys, which is actually a good thing.)  (Update: I
   think it would be a bad idea to remove server-frames.)

-- Change emacsclient/server.el to support the -t argument better,
   i.e. automatically close the socket when the frame is closed.

   (Seems to be working OK.)

-- Fix mysterious memory corruption error with tty deletion.  To
   trigger it, try the following shell command:

	while true; do TERM=no-such-terminal-definition emacsclient -h; done

   Emacs usually dumps core after a few dozen iterations.  (The bug
   seems to be related to the xfreeing or bzeroing of
   tty_output.Wcm.  Maybe there are outside references to struct Wcm?
   Why were these vars collected into a struct before multi-tty
   support?)

   (Done.  Whew.  It turned out that the problem had nothing to do
   with hypothetical external references to Wcm, or any other
   tty_output component; it was simply that delete_tty closed the
   filehandles of secondary ttys twice, resulting in fclose doubly
   freeing memory.  Utterly trivial matter.  I love the C's memory
   management, it puts hair on your chest.)

-- Support raw secondary terminals.  (Note that SIGIO works only on
   the controlling terminal.) Hint: extend read_input_waiting for
   multiple ttys and hopefully this will be fixed.

   (Done, it seems to have been working already for some time.  It
   seems F_SETOWN does work, after all.  Not sure what made it fail
   earlier, but it seems to be fixed (there were several changes
   around request_sigio, maybe one of them did it).
   read_input_waiting is only used in sys_select, don't change
   it.)  (Update: After adding X support, it's broken again.)

-- Find out why does Emacs abort when it wants to close its
   controlling tty.  Hint: chan_process[] array.  Hey, maybe
   noninterrupt-IO would work, too?  Update: no, there is no process
   for stdin/out.

   (Done.  Added add/delete_keyboard_wait_descriptor to
   term_init/delete_tty.  The hint was right, in a way.)

-- Issue with SIGIO: it needs to be disabled during redisplay.  See if
   fcntl kernel behaviour could be emulated by emacsclient.

   (Done.  Simply disabled the SIGIO emulation hack in emacsclient.)
   (Update: it was added back.)

-- server.el: There are issues with saving files in buffers of closed
   clients.  Try editing a file with emacsclient -f, and (without
   saving it) do a delete-frame.  The frame is closed without
   question, and a surprising confirmation prompt appears in another
   frame.

   (Done.  delete-frame now asks for confirmation if it still has
   pending buffers, and modified buffers don't seem to be deleted.)

-- emacsclient.el, server.el: Handle eval or file open errors when
   doing -t.

   (Done.)

-- Make parts of struct tty_output accessible from Lisp.  The device
   name and the type is sufficient.

   (Done, see frame-tty-name and frame-tty-type.)

-- Export delete_tty to the Lisp environment, for emacsclient.

   (Done, see delete-tty.)

-- Get rid of the accessor macros in termchar.h, or define macros for
   all members.

   (Done.)

-- Move device-specific parameters (like costs) commonly used by
   device backends to a common, device-dependent structure.

   (Done.  See struct display_method in termhooks.h.)

-- Fix X support.

   (Done.  Well, it seems to be working.)

-- Allow simultaneous X and tty frames.  (Handling input could be
   tricky.  Or maybe not.)

   (Done.  Allowed, that is.  It is currently extremely unstable, to
   the point of being unusable.  The rif variable causes constant
   core dumps.  Handling input is indeed tricky.)

-- Rewrite multi-tty input in terms of MULTI_KBOARD.

   (Done.  In fact, there was no need to rewrite anything, I just
   added a kboard member to tty_display_info, and initialized the
   frame's kboard from there.)

-- Fix rif issue with X-tty combo sessions.  IMHO the best thing to do
   is to get rid of that global variable (and use the value value in
   display_method, which is guaranteed to be correct).

   (Done, did exactly that.  Core dumps during combo sessions became
   much rarer.  In fact, I have not yet met a single one.)

-- Add multi-tty support to talk.el.

   (Done.)

-- Clean up the source of emacsclient.  It is a mess.

   (Done, eliminated stupid proxy-pty kludge.)

-- Fix faces on tty frames during X-tty combo sessions.  There is an
   init_frame_faces call in init_sys_modes, see if there is a problem
   with it.

   (Done, there was a stupid mistake in
   Ftty_supports_face_attributes_p. Colors are broken, though.)

-- C-x 5 2, C-x 5 o, C-x 5 0 on an emacsclient frame unexpectedly
   exits emacsclient.  This is a result of trying to be clever with
   delete-frame-functions.

   (Fixed, added delete-tty-after-functions, and changed server.el to
   use it.)

-- Something with (maybe) multi-keyboard support broke function keys
   and arrows on ttys during X+tty combo sessions.  Debug this.

   (I can't reproduce it, maybe the terminal type was wrong.)

-- Fix input from raw ttys (again).

   (Now it seems to work all right.)

-- During an X-tty combo session, a (message "Hello") from a tty frame
   goes to the X frame.  Fix this.

   (Done.  There was a safeguard against writing to the initial
   terminal frame during bootstrap which prevented echo_area_display
   from working correctly on a tty frame during a combo session.)

-- If there are no frames on its controlling terminal, Emacs should
   exit if the uses presses C-c there.

   (Done, as far as possible.  See the SIGTERM comment in
   interrupt_signal on why this seems to be impossible to solve this
   in general.)

-- During an X session, Emacs seems to read from stdin.  Also, Emacs
   fails to start without a controlling tty.

   (Fixed by replacing the troublesome termcap display with a dummy
   bootstrap display during bootstrap.

-- Do tty output through struct display, like graphical display
   backends.

   (Done.)

-- Define an output_initial value for output_method for the initial
   frame that is dumped with Emacs.  Checking for this frame (e.g. in
   cmd_error_internal) is ugly.

   (Done, broking interactive temacs.)

-- The command `emacsclient -t -e '(delete-frame)'' fails to exit.

   (Fixed.)

-- frame-creation-function should always create a frame that is on the
   same display as the selected frame.  Maybe frame-creation-function
   should simply be removed and make-frame changed to do the right
   thing.

   (Done, with a nice hack.  frame-creation-function is now frame-local.)

-- Fix C-g on raw ttys.

   (Done.  I disabled the interrupt/quit keys on all secondary
   terminals, so Emacs sees C-g as normal input.  This looks like an
   overkill, because emacsclient has extra code to pass SIGINT to
   Emacs, so C-g should remain the interrupt/quit key on emacsclient
   frames.  See the next entry why implementing this distinction would
   be a bad idea.)

-- Make sure C-g goes to the right frame with ttys.  This is hard, as
   SIGINT doesn't have a tty parameter. :-(

   (Done, the previous change fixes this as a pleasant side effect.)


;;; arch-tag: 8da1619e-2e79-41a8-9ac9-a0485daad17d

83013:e77d1a63471b

   secondary tty support, but it does not seem to do anything useful.
   (Update: Look again. X unconditionally enables this, maybe that's
   why raw terminal support is broken again.  I really do need to
   understand input.)

** emacsclient -t from an Emacs term buffer does not work, complains
   about face problems.  This can even lock up Emacs (if the recursive
   frame sets single_kboard).  Update: the face problems are caused by
   bugs in term.el, not in multi-tty.  The lockup is caused by
   single_kboard mode, and is not easily solvable.  The best thing to
   do is to simply refuse to create a tty frame of type `eterm'.

DIARY OF CHANGES
----------------

(ex-TODO items with explanations.)

-- Introduce a new struct for terminal devices.

   (Done, see struct tty_output.  The list of members is not yet
   complete.)

-- Change the bootstrap procedure to initialize tty_list.

   (Done, but needs review.)

-- Change make-terminal-frame to support specifying another tty.

   (Done, new frame parameters: `tty' and `tty-type'.)

-- Implement support for reading from multiple terminals.

   (Done, read_avail_input tries to read from each terminal, until one
   succeeds.  MULTI_KBOARD is not used.  Secondary terminals don't send
   SIGIO!)

   (Update: They do, now.)

   (Update2: After enabling X, they don't.)

-- other-frame should cycle through the frames on the `current'
   terminal only.

   (Done, by trivially modifiying next_frame and prev_frame.)

-- Support different terminal sizes.

   (Done, no problem.)

-- Make sure terminal resizes are handled gracefully.  (Could be
   problematic.)

   (Done.  We don't get automatic SIGWINCH for additional ttys,
   though.)

-- Extend emacsclient to automatically open a new tty when it connects
   to Emacs.

   (Done.  It's an ugly hack, needs more work.)

-- Redisplay must refresh the topmost frame on *all* terminals, not
   just the initial terminal.

   (Done, but introduced an ugly redisplay problems.  Ugh.)

-- Fix redisplay problems.

   (Done; it turned out that the entire Wcm structure must be moved
   inside tty_output.  Why didn't I catch this earlier?)

-- Provide a way for emacsclient to tell Emacs that the tty has been
   resized.

   (Done, simply forward the SIGWINCH signal.)

-- Each keypress should automatically select the frame corresponding
   to the terminal that it was coming from.  This means that Emacs
   must know from which terminal the last keyboard event came from.

   (Done, it was quite simple, the input event system already
   supported multiple frames.)

-- Fix SIGIO issue with secondary terminals.

   (Done, emacsclient signals Emacs after writing to the proxy pseudo
   terminal.  Note that this means that multi-tty does not work with
   raw ttys!)

   (Update: This is bullshit.  There is a read_input_waiting function,
   extend that somehow.)

   (Update of update: The first update was not right either, extending
   read_input_waiting was not necessary.  Secondary ttys do seem to
   send signals on input.)

   (Update^3: Not any more.)

-- Make make-terminal-frame look up the `tty' and `tty-type' frame
   parameters from the currently selected terminal before the global
   default.

   (Done.)

-- Put all cached terminal escape sequences into struct tty_output.
   Currently, they are still stored in global variables, so we don't
   really support multiple terminal types.

   (Done.  It was not fun.)

-- Implement sane error handling after initialization.  (Currently
   emacs exits if you specify a bad terminal type.)  The helpful error
   messages must still be provided when Emacs starts.

   (Done.)

-- Implement terminal deletion, i.e., deleting local frames, closing
   the tty device and restoring its previous state without exiting
   Emacs.

   (Done, but at the moment only called when an error happens during
   initialization.  There is a memory corruption error around this
   somewhere.)  (Update: now it is fully enabled.)

-- Implement automatic deletion of terminals when the last frame on
   that terminal is closed.

   (Done.)

-- Restore tty screen after closing the terminal.

   (Done, we do the same as Emacs 21.2 for all terminals.)

-- 'TERM=dumb src/emacs' does not restore the terminal state.

   (Done.)

-- C-g should work on secondary terminals.

   (Done, but the binding is not configurable.)

-- Deal with SIGHUP in Emacs and in emacsclient.  (After this, the
   server-frames may be removed from server.el.)

   (Done, nothing to do.  It seems that Emacs does not receive SIGHUP
   from secondary ttys, which is actually a good thing.)  (Update: I
   think it would be a bad idea to remove server-frames.)

-- Change emacsclient/server.el to support the -t argument better,
   i.e. automatically close the socket when the frame is closed.

   (Seems to be working OK.)

-- Fix mysterious memory corruption error with tty deletion.  To
   trigger it, try the following shell command:

	while true; do TERM=no-such-terminal-definition emacsclient -h; done

   Emacs usually dumps core after a few dozen iterations.  (The bug
   seems to be related to the xfreeing or bzeroing of
   tty_output.Wcm.  Maybe there are outside references to struct Wcm?
   Why were these vars collected into a struct before multi-tty
   support?)

   (Done.  Whew.  It turned out that the problem had nothing to do
   with hypothetical external references to Wcm, or any other
   tty_output component; it was simply that delete_tty closed the
   filehandles of secondary ttys twice, resulting in fclose doubly
   freeing memory.  Utterly trivial matter.  I love the C's memory
   management, it puts hair on your chest.)

-- Support raw secondary terminals.  (Note that SIGIO works only on
   the controlling terminal.) Hint: extend read_input_waiting for
   multiple ttys and hopefully this will be fixed.

   (Done, it seems to have been working already for some time.  It
   seems F_SETOWN does work, after all.  Not sure what made it fail
   earlier, but it seems to be fixed (there were several changes
   around request_sigio, maybe one of them did it).
   read_input_waiting is only used in sys_select, don't change
   it.)  (Update: After adding X support, it's broken again.)

-- Find out why does Emacs abort when it wants to close its
   controlling tty.  Hint: chan_process[] array.  Hey, maybe
   noninterrupt-IO would work, too?  Update: no, there is no process
   for stdin/out.

   (Done.  Added add/delete_keyboard_wait_descriptor to
   term_init/delete_tty.  The hint was right, in a way.)

-- Issue with SIGIO: it needs to be disabled during redisplay.  See if
   fcntl kernel behaviour could be emulated by emacsclient.

   (Done.  Simply disabled the SIGIO emulation hack in emacsclient.)
   (Update: it was added back.)

-- server.el: There are issues with saving files in buffers of closed
   clients.  Try editing a file with emacsclient -f, and (without
   saving it) do a delete-frame.  The frame is closed without
   question, and a surprising confirmation prompt appears in another
   frame.

   (Done.  delete-frame now asks for confirmation if it still has
   pending buffers, and modified buffers don't seem to be deleted.)

-- emacsclient.el, server.el: Handle eval or file open errors when
   doing -t.

   (Done.)

-- Make parts of struct tty_output accessible from Lisp.  The device
   name and the type is sufficient.

   (Done, see frame-tty-name and frame-tty-type.)

-- Export delete_tty to the Lisp environment, for emacsclient.

   (Done, see delete-tty.)

-- Get rid of the accessor macros in termchar.h, or define macros for
   all members.

   (Done.)

-- Move device-specific parameters (like costs) commonly used by
   device backends to a common, device-dependent structure.

   (Done.  See struct display_method in termhooks.h.)

-- Fix X support.

   (Done.  Well, it seems to be working.)

-- Allow simultaneous X and tty frames.  (Handling input could be
   tricky.  Or maybe not.)

   (Done.  Allowed, that is.  It is currently extremely unstable, to
   the point of being unusable.  The rif variable causes constant
   core dumps.  Handling input is indeed tricky.)

-- Rewrite multi-tty input in terms of MULTI_KBOARD.

   (Done.  In fact, there was no need to rewrite anything, I just
   added a kboard member to tty_display_info, and initialized the
   frame's kboard from there.)

-- Fix rif issue with X-tty combo sessions.  IMHO the best thing to do
   is to get rid of that global variable (and use the value value in
   display_method, which is guaranteed to be correct).

   (Done, did exactly that.  Core dumps during combo sessions became
   much rarer.  In fact, I have not yet met a single one.)

-- Add multi-tty support to talk.el.

   (Done.)

-- Clean up the source of emacsclient.  It is a mess.

   (Done, eliminated stupid proxy-pty kludge.)

-- Fix faces on tty frames during X-tty combo sessions.  There is an
   init_frame_faces call in init_sys_modes, see if there is a problem
   with it.

   (Done, there was a stupid mistake in
   Ftty_supports_face_attributes_p. Colors are broken, though.)

-- C-x 5 2, C-x 5 o, C-x 5 0 on an emacsclient frame unexpectedly
   exits emacsclient.  This is a result of trying to be clever with
   delete-frame-functions.

   (Fixed, added delete-tty-after-functions, and changed server.el to
   use it.)

-- Something with (maybe) multi-keyboard support broke function keys
   and arrows on ttys during X+tty combo sessions.  Debug this.

   (I can't reproduce it, maybe the terminal type was wrong.)

-- Fix input from raw ttys (again).

   (Now it seems to work all right.)

-- During an X-tty combo session, a (message "Hello") from a tty frame
   goes to the X frame.  Fix this.

   (Done.  There was a safeguard against writing to the initial
   terminal frame during bootstrap which prevented echo_area_display
   from working correctly on a tty frame during a combo session.)

-- If there are no frames on its controlling terminal, Emacs should
   exit if the uses presses C-c there.

   (Done, as far as possible.  See the SIGTERM comment in
   interrupt_signal on why this seems to be impossible to solve this
   in general.)

-- During an X session, Emacs seems to read from stdin.  Also, Emacs
   fails to start without a controlling tty.

   (Fixed by replacing the troublesome termcap display with a dummy
   bootstrap display during bootstrap.

-- Do tty output through struct display, like graphical display
   backends.

   (Done.)

-- Define an output_initial value for output_method for the initial
   frame that is dumped with Emacs.  Checking for this frame (e.g. in
   cmd_error_internal) is ugly.

   (Done, broking interactive temacs.)

-- The command `emacsclient -t -e '(delete-frame)'' fails to exit.

   (Fixed.)

-- frame-creation-function should always create a frame that is on the
   same display as the selected frame.  Maybe frame-creation-function
   should simply be removed and make-frame changed to do the right
   thing.

   (Done, with a nice hack.  frame-creation-function is now frame-local.)

-- Fix C-g on raw ttys.

   (Done.  I disabled the interrupt/quit keys on all secondary
   terminals, so Emacs sees C-g as normal input.  This looks like an
   overkill, because emacsclient has extra code to pass SIGINT to
   Emacs, so C-g should remain the interrupt/quit key on emacsclient
   frames.  See the next entry why implementing this distinction would
   be a bad idea.)

-- Make sure C-g goes to the right frame with ttys.  This is hard, as
   SIGINT doesn't have a tty parameter. :-(

   (Done, the previous change fixes this as a pleasant side effect.)

-- I have seen a case when Emacs with multiple ttys fell into a loop
   eating 100% of CPU time.  Strace showed this loop:

	getpid()                                = 30284
	kill(30284, SIGIO)                      = 0
	--- SIGIO (I/O possible) @ 0 (0) ---

   Update: This may have been caused by checking for nread != 0
   instead of nread > 0 after calling read_socket_hook in
   read_avail_input.

   (Fixed.  This was caused by unconditionally including stdin in
   input_wait_mask in init_process.  The select call in
   wait_reading_process_input always returned immediately, indicating
   that there is pending input from stdin, which nobody read.

   Note that the above strace output seems to be an unrelated but
   similar bug.  I think that is now fixed.)

;;; arch-tag: 8da1619e-2e79-41a8-9ac9-a0485daad17d