xemacs-beta: src/signal.c comparison

comparison src/signal.c @ 853:2b6fa2618f76

[xemacs-hg @ 2002-05-28 08:44:22 by ben] merge my stderr-proc ws make-docfile.c: Fix places where we forget to check for EOF. code-init.el: Don't use CRLF conversion by default on process output. CMD.EXE and friends work both ways but Cygwin programs don't like the CRs. code-process.el, multicast.el, process.el: Removed. Improvements to call-process-internal: -- allows a buffer to be specified for input and stderr output -- use it on all systems -- implement C-g as documented -- clean up and comment call-process-region uses new call-process facilities; no temp file. remove duplicate funs in process.el. comment exactly how coding systems work and fix various problems. open-multicast-group now does similar coding-system frobbing to open-network-stream. dumped-lisp.el, faces.el, msw-faces.el: Fix some hidden errors due to code not being defined at the right time. xemacs.mak: Add -DSTRICT. ================================================================ ALLOW SEPARATION OF STDOUT AND STDERR IN PROCESSES ================================================================ Standard output and standard error can be processed separately in a process. Each can have its own buffer, its own mark in that buffer, and its filter function. You can specify a separate buffer for stderr in `start-process' to get things started, or use the new primitives: set-process-stderr-buffer process-stderr-buffer process-stderr-mark set-process-stderr-filter process-stderr-filter Also, process-send-region takes a 4th optional arg, a buffer. Currently always uses a pipe() under Unix to read the error output. (#### Would a PTY be better?) sysdep.h, sysproc.h, unexfreebsd.c, unexsunos4.c, nt.c, emacs.c, callproc.c, symsinit.h, sysdep.c, Makefile.in.in, process-unix.c: Delete callproc.c. Move child_setup() to process-unix.c. wait_for_termination() now only needed on a few really old systems. console-msw.h, event-Xt.c, event-msw.c, event-stream.c, event-tty.c, event-unixoid.c, events.h, process-nt.c, process-unix.c, process.c, process.h, procimpl.h: Rewrite the process methods to handle a separate channel for error input. Create Lstreams for reading in the error channel. Many process methods need change. In general the changes are fairly clear as they involve duplicating what's used for reading the normal stdout and changing for stderr -- although tedious, as such changes are required throughout the entire process code. Rewrote the code that reads process output to do two loops, one for stdout and one for stderr. gpmevent.c, tooltalk.c: set_process_filter takes an argument for stderr. ================================================================ NEW ERROR-TRAPPING MECHANISM ================================================================ Totally rewrite error trapping code to be unified and support more features. Basic function is call_trapping_problems(), which lets you specify, by means of flags, what sorts of problems you want trapped. these can include -- quit -- errors -- throws past the function -- creation of "display objects" (e.g. buffers) -- deletion of already-existing "display objects" (e.g. buffers) -- modification of already-existing buffers -- entering the debugger -- gc -- errors->warnings (ala suspended errors) etc. All other error funs rewritten in terms of this one. Various older mechanisms removed or rewritten. window.c, insdel.c, console.c, buffer.c, device.c, frame.c: When creating a display object, added call to note_object_created(), for use with trapping_problems mechanism. When deleting, call check_allowed_operation() and note_object deleted(). The trapping-problems code records the objects created since the call-trapping-problems began. Those objects can be deleted, but none others (i.e. previously existing ones). bytecode.c, cmdloop.c: internal_catch takes another arg. eval.c: Add long comments describing the "five lists" used to maintain state (backtrace, gcpro, specbind, etc.) in the Lisp engine. backtrace.h, eval.c: Implement trapping-problems mechanism, eliminate old mechanisms or redo in terms of new one. frame.c, gutter.c: Flush out the concept of "critical display section", defined by the in_display() var. Use an internal_bind() to get it reset, rather than just doing it at end, because there may be a non-local exit. event-msw.c, event-stream.c, console-msw.h, device.c, dialog-msw.c, frame.c, frame.h, intl.c, toolbar.c, menubar-msw.c, redisplay.c, alloc.c, menubar-x.c: Make use of new trapping-errors stuff and rewrite code based on old mechanisms. glyphs-widget.c, redisplay.h: Protect calling Lisp in redisplay. insdel.c: Protect hooks against deleting existing buffers. frame-msw.c: Use EQ, not EQUAL in hash tables whose keys are just numbers. Otherwise we run into stickiness in redisplay because internal_equal() can QUIT. ================================================================ SIGNAL, C-G CHANGES ================================================================ Here we change the way that C-g interacts with event reading. The idea is that a C-g occurring while we're reading a user event should be read as C-g, but elsewhere should be a QUIT. The former code did all sorts of bizarreness -- requiring that no QUIT occurs anywhere in event-reading code (impossible to enforce given the stuff called or Lisp code invoked), and having some weird system involving enqueue/dequeue of a C-g and interaction with Vquit_flag -- and it didn't work. Now, we simply enclose all code where we want C-g read as an event with {begin/end}_dont_check_for_quit(). This completely turns off the mechanism that checks (and may remove or alter) C-g in the read-ahead queues, so we just get the C-g normal. Signal.c documents this very carefully. cmdloop.c: Correct use of dont_check_for_quit to new scheme, remove old out-of-date comments. event-stream.c: Fix C-g handling to actually work. device-x.c: Disable quit checking when err out. signal.c: Cleanup. Add large descriptive comment. process-unix.c, process-nt.c, sysdep.c: Use QUIT instead of REALLY_QUIT. It's not necessary to use REALLY_QUIT and just confuses the issue. lisp.h: Comment quit handlers. ================================================================ CONS CHANGES ================================================================ free_cons() now takes a Lisp_Object not the result of XCONS(). car and cdr have been renamed so that they don't get used directly; go through XCAR(), XCDR() instead. alloc.c, dired.c, editfns.c, emodules.c, fns.c, glyphs-msw.c, glyphs-x.c, glyphs.c, keymap.c, minibuf.c, search.c, eval.c, lread.c, lisp.h: Correct free_cons calling convention: now takes Lisp_Object, not Lisp_Cons chartab.c: Eliminate direct use of ->car, ->cdr, should be black box. callint.c: Rewrote using EXTERNAL_LIST_LOOP to avoid use of Lisp_Cons. ================================================================ USE INTERNAL-BIND-* ================================================================ eval.c: Cleanups of these funs. alloc.c, fileio.c, undo.c, specifier.c, text.c, profile.c, lread.c, redisplay.c, menubar-x.c, macros.c: Rewrote to use internal_bind_int() and internal_bind_lisp_object() in place of whatever varied and cumbersome mechanisms were formerly there. ================================================================ SPECBIND SANITY ================================================================ backtrace.h: - Improved comments backtrace.h, bytecode.c, eval.c: Add new mechanism check_specbind_stack_sanity() for sanity checking code each time the catchlist or specbind stack change. Removed older prototype of same mechanism. ================================================================ MISC ================================================================ lisp.h, insdel.c, window.c, device.c, console.c, buffer.c: Fleshed out authorship. device-msw.c: Correct bad Unicode-ization. print.c: Be more careful when not initialized or in fatal error handling. search.c: Eliminate running_asynch_code, an FSF holdover. alloc.c: Added comments about gc-cons-threshold. dialog-x.c: Use begin_gc_forbidden() around code to build up a widget value tree, like in menubar-x.c. gui.c: Use Qunbound not Qnil as the default for gethash. lisp-disunion.h, lisp-union.h: Added warnings on use of VOID_TO_LISP(). lisp.h: Use ERROR_CHECK_STRUCTURES to turn on ERROR_CHECK_TRAPPING_PROBLEMS and ERROR_CHECK_TYPECHECK lisp.h: Add assert_with_message. lisp.h: Add macros for gcproing entire arrays. (You could do this before but it required manual twiddling the gcpro structure.) lisp.h: Add prototypes for new functions defined elsewhere.

author	ben
date	Tue, 28 May 2002 08:45:36 +0000
parents	a5954632b187
children	79c6ff3eef26

comparison

equal deleted inserted replaced

-:d83885ef293b
+:2b6fa2618f76
 /* Handling asynchronous signals.
 Copyright (C) 1992, 1993, 1994 Free Software Foundation, Inc.
-Copyright (C) 1995, 1996, 2001 Ben Wing.
+Copyright (C) 1995, 1996, 2001, 2002 Ben Wing.
 This file is part of XEmacs.
 XEmacs is free software; you can redistribute it and/or modify it
 under the terms of the GNU General Public License as published by the
 {
 kick_status_notify ();
 }
 else
 /* call1 GC-protects its arguments */
-call1_trapping_errors ("Error in asynchronous timeout callback",
+call1_trapping_problems ("Error in asynchronous timeout callback",
-			   fun, arg);
+			     fun, arg, INHIBIT_GC);
 waiting_for_user_input_p = 0;
 unbind_to (specco);
 }
 /**********************************************************************/
 /* called from QUIT when something_happened gets set (as a result of
 a signal) */
-int
+void
 check_what_happened (void)
 {
+#ifdef ERROR_CHECK_TRAPPING_PROBLEMS
+if (in_display
+&& !((get_inhibit_flags () & INTERNAL_INHIBIT_ERRORS)
+	   && (get_inhibit_flags () & INTERNAL_INHIBIT_THROWS)))
+assert_with_message
+(0, "QUIT called from within redisplay without being properly wrapped");
+#endif
 /* No GC can happen anywhere here.  handle_async_timeout_signal()
 prevents GC (from asynch timeout handler), so does check_quit()
 (from processing a message such as WM_INITMENU as a result of
 draining the message queue).  establish_slow_interrupt_timer() is
 too low-level to do anything that might invoke QUIT or call Lisp
 {
 slowed_interrupt_timeout_happened = 0;
 establish_slow_interrupt_timer ();
 }
-return check_quit ();
+check_quit ();
 }
 #ifdef SIGIO
 /* Signal handler for SIGIO. */
 /**********************************************************************/
 /*                        Control-G checking                          */
 /**********************************************************************/
+/* Note: The code to handle QUIT is divided between lisp.h and signal.c.
+There is also some special-case code in the async timer code in
+event-stream.c to notice when the poll-for-quit (and poll-for-sigchld)
+timers have gone off. */
+/* OK, here's an overview of how this convoluted stuff works:
+[1] Scattered throughout the XEmacs core code are calls to the macro QUIT;
+This macro checks to see whether a C-g has recently been pressed and
+not yet handled, and if so, it handles the C-g by calling signal_quit(),
+which invokes the standard Fsignal() code, with the error being Qquit.
+Lisp code can establish handlers for this (using condition-case), but
+normally there is no handler, and so execution is thrown back to the
+innermost enclosing event loop. (One of the things that happens when
+entering an event loop is that a condition-case is established that
+catches *all* calls to `signal', including this one.)
+[2] How does the QUIT macro check to see whether C-g has been pressed;
+obviously this needs to be extremely fast.  Now for some history.
+In early Lemacs as inherited from the FSF going back 15 years or
+more, there was a great fondness for using SIGIO (which is sent
+whenever there is I/O available on a given socket, tty, etc.).
+In fact, in GNU Emacs, perhaps even today, all reading of events
+from the X server occurs inside the SIGIO handler!  This is crazy,
+but not completely relevant.  What is relevant is that similar
+stuff happened inside the SIGIO handler for C-g: it searched
+through all the pending (i.e. not yet delivered to XEmacs yet)
+X events for one that matched C-g.  When it saw a match, it set
+Vquit_flag to Qt.  On TTY's, C-g is actually mapped to be the
+interrupt character (i.e. it generates SIGINT), and XEmacs's
+handler for this signal sets Vquit_flag to Qt.  Then, sometime
+later after the signal handlers finished and a QUIT macro was
+called, the macro noticed the setting of Vquit_flag and used
+this as an indication to call signal_quit().  What signal_quit()
+actually does is set Vquit_flag to Qnil (so that we won't get
+repeated interruptions from a single C-g press) and then calls
+the equivalent of (signal 'quit nil).
+[3] Another complication is introduced in that Vquit_flag is actually
+exported to Lisp as `quit-flag'.  This allows users some level of
+control over whether and when C-g is processed as quit, esp. in
+combination with `inhibit-quit'.  This is another Lisp variable,
+and if set to non-nil, it inhibits signal_quit() from getting
+called, meaning that the C-g gets essentially ignored.  But not
+completely: Because the resetting of `quit-flag' happens only
+in signal_quit(), which isn't getting called, the C-g press is
+still noticed, and as soon as `inhibit-quit' is set back to nil,
+a quit will be signalled at the next QUIT macro.  Thus, what
+`inhibit-quit' really does is defer quits until after the quit-
+inhibitted period.
+[4] Another consideration, introduced by XEmacs, is critical quitting.
+If you press Control-Shift-G instead of just C-g, `quit-flag' is
+set to `critical' instead of to t.  When QUIT processes this value,
+it *ignores* the value of `inhibit-quit'.  This allows you to quit
+even out of a quit-inhibitted section of code!  Furthermore, when
+signal_quit() notices that it was invoked as a result of a critical
+quit, it automatically invokes the debugger (which otherwise would
+only happen when `debug-on-quit' is set to t).
+[5] Well, I explained above about how `quit-flag' gets set correctly,
+but I began with a disclaimer stating that this was the old way
+of doing things.  What's done now?  Well, first of all, the SIGIO
+handler (which formerly checked all pending events to see if there's
+a C-g) now does nothing but set a flag -- or actually two flags,
+something_happened and quit_check_signal_happened.  There are two
+flags because the QUIT macro is now used for more than just handling
+QUIT; it's also used for running asynchronous timeout handlers that
+have recently expired, and perhaps other things.  The idea here is
+that the QUIT macros occur extremely often in the code, but only occur
+at places that are relatively safe -- in particular, if an error occurs,
+nothing will get completely trashed.
+[6] Now, let's look at QUIT again.
+UNFINISHED.  Note, however, that as of the point when this comment
+got committed to CVS (mid-2001), the interaction between reading
+C-g as an event and processing it as QUIT was overhauled to (for
+the first time) be understandable and actually work correctly.
+Now, the way things work is that if C-g is pressed while XEmacs is
+blocking at the top level, waiting for a user event, it will be
+read as an event; otherwise, it will cause QUIT. (This includes
+times when XEmacs is blocking, but not waiting for a user event,
+e.g. accept-process-output and wait_delaying_user_events().)
+Formerly, this was supposed to happen, but didn't always due to a
+bizarre and broken scheme, documented in next_event_internal
+like this:
+If we read a ^G, then set quit-flag but do not discard the ^G.
+The callers of next_event_internal() will do one of two things:
+-- set Vquit_flag to Qnil. (next-event does this.) This will
+cause the ^G to be treated as a normal keystroke.
+-- not change Vquit_flag but attempt to enqueue the ^G, at
+which point it will be discarded.  The next time QUIT is
+called, it will notice that Vquit_flag was set.
+This required weirdness in enqueue_command_event_1 like this:
+put the event on the typeahead queue, unless
+the event is the quit char, in which case the `QUIT'
+which will occur on the next trip through this loop is
+all the processing we should do - leaving it on the queue
+would cause the quit to be processed twice.
+And further weirdness elsewhere, none of which made any sense,
+and didn't work, because (e.g.) it required that QUIT never
+happen anywhere inside next_event_internal() or any callers when
+C-g should be read as a user event, which was impossible to
+implement in practice.
+Now what we do is fairly simple.  Callers of next_event_internal()
+that want C-g read as a user event call begin_dont_check_for_quit().
+next_event_internal(), when it gets a C-g, simply sets Vquit_flag
+(just as when a C-g is detected during the operation of QUIT or
+QUITP), and then tries to QUIT.  This will fail if blocked by the
+previous call, at which point next_event_internal() will return
+the C-g as an event.  To unblock things, first set Vquit_flag to
+nil (it was set to t when the C-g was read, and if we don't reset
+it, the next call to QUIT will quit), and then unbind_to() the
+depth returned by begin_dont_check_for_quit().  It makes no
+difference is QUIT is called a zillion times in next_event_internal()
+or anywhere else, because it's blocked and will never signal.
+--ben
+*/
 static Lisp_Object
 restore_dont_check_for_quit (Lisp_Object val)
 {
 dont_check_for_quit = XINT (val);
 return Qnil;
 /* As an optimization in QUIT_FLAG_SAYS_SHOULD_QUIT, we bind inhibit-quit
 to t -- it has to be checked anyway, and by doing this, we only need
 to check dont_check_for_quit when quit-flag == `critical', which is
 rare. */
 specbind (Qinhibit_quit, Qt);
-record_unwind_protect (restore_dont_check_for_quit,
+internal_bind_int (&dont_check_for_quit, 1);
-			 make_int (dont_check_for_quit));
-dont_check_for_quit = 1;
 return depth;
 }
-/* The effect of this function is to set Vquit_flag if the user pressed
+/* If we're inside of a begin_dont_check_for_quit() section, but want
-^G and discard the ^G, so as to not notice the same ^G again. */
+to temporarily enable quit-checking, call this.  This is used in
+particular when processing menu filters -- some menu filters do
+antisocial things like load large amounts of Lisp code (custom in
+particular), and we obviously want a way of breaking out of any
+problems.  If you do use this, you should really be trapping the
+throw() that comes from the quitting (as does the code that handles
+menus popping up). */
 int
+begin_do_check_for_quit (void)
+{
+int depth = specpdl_depth ();
+specbind (Qinhibit_quit, Qnil);
+internal_bind_int (&dont_check_for_quit, 0);
+/* #### should we set Vquit_flag to Qnil? */
+return depth;
+}
+/* The effect of this function is to set Vquit_flag appropriately if the
+user pressed C-g or Sh-C-g.  After this function finishes, Vquit_flag
+will be Qt for C-g, Qcritical for Sh-C-g, and unchanged otherwise.
+The C-g or Sh-C-g is discarded, so it won't be noticed again.
+*/
+void
 check_quit (void)
 {
-/* dont_check_for_quit is set in two circumstances:
+/* dont_check_for_quit is set in three circumstances:
 (1) when we are in the process of changing the window
 configuration.  The frame might be in an inconsistent state,
 which will cause assertion failures if we check for QUIT.
 (2) when we are reading events, and want to read the C-g
 as an event.  The normal check for quit will discard the C-g,
 which would be bad.
-[[#### C-g is still often read as quit, e.g. if you type C-x C-g (the
+(3) when we're going down with a fatal error.  we're most likely
-C-g happens during the sit-for in maybe_echo_keys(); even if we
+in an inconsistent state, and we definitely don't want to be
-attempt to inhibit quit here, there is still a check later on for
+interrupted. */
-QUIT.  To fix this properly requires a fairly substantial overhaul of
-the quit-checking code, which is probably not worth it.)]] not true,
+/* We should *not* conditionalize on Vinhibit_quit, or
-we just have to always do dont_check_for_quit around all code that
-reads events.  my stderr-proc ws already does this.
-We should *not* conditionalize on Vinhibit_quit, or
 critical-quit (Control-Shift-G) won't work right. */
 /* WARNING: Even calling check_quit(), without actually dispatching
 a quit signal, can result in arbitrary Lisp code getting executed
 -- at least under Windows. (Not to mention obvious Lisp
 XEMACS! mainCRTStartup + 180 bytes
 _start() line 171
 KERNEL32! BaseProcessStart@4 + 115547 bytes
 */
+int specdepth;
 if (dont_check_for_quit)
-return 0;
+return;
 if (quit_check_signal_happened)
 {
 /* Since arbitrary Lisp code may be executed, GC might happen,
 which would majorly fuck a lot of things, e.g. re_match()
 [string gets relocated] and lots of other code that's not
 prepared to handle GC in QUIT. */
-int specdepth = begin_gc_forbidden ();
+specdepth = begin_gc_forbidden ();
 quit_check_signal_happened = 0;
 event_stream_quit_p ();
 unbind_to (specdepth);
-return 1;
+}
-}
-else
-return 0;
 }
 void
 #endif /* not SIGIO and not DONT_POLL_FOR_QUIT */
 }
 #endif /* 0 */
-#if defined(HAVE_UNIX_PROCESSES) && !defined(SIGCHLD)
+#if defined (HAVE_UNIX_PROCESSES) && !defined (SIGCHLD)
 static void
 init_poll_for_sigchld (void)
 {
 /* Check for terminated processes every 1/4 of a second.
 #elif !defined (DONT_POLL_FOR_QUIT)
 init_poll_for_quit ();
 #endif
 }
-#if defined(HAVE_UNIX_PROCESSES) && !defined(SIGCHLD)
+#if defined (HAVE_UNIX_PROCESSES) && !defined (SIGCHLD)
 init_poll_for_sigchld ();
 #endif
 EMACS_UNBLOCK_ALL_SIGNALS ();

Mercurial > hg > xemacs-beta

comparison src/signal.c @ 853:2b6fa2618f76