annotate etc/DEBUG @ 83294:a8c8fab4072c

I give up; attempt to fix annoying crashes by reverting xdisp.c changes in patch-328. * src/xdisp.c (with_echo_area_buffer, set_message, set_message_1, echo_area_display): Revert change applied in patch-328. git-archimport-id: [e-mail]--2004/emacs--multi-tty--0--patch-334

Debugging GNU Emacs
Copyright (c) 1985, 2000, 2001, 2005 Free Software Foundation, Inc.

   Permission is granted to anyone to make or distribute verbatim copies
   of this document as received, in any medium, provided that the
   copyright notice and permission notice are preserved,
   and that the distributor grants the recipient permission
   for further redistribution as permitted by this notice.

   Permission is granted to distribute modified versions
   of this document, or of portions of it,
   under the above conditions, provided also that they
   carry prominent notices stating who last changed them.

[People who debug Emacs on Windows using native Windows debuggers
should read the Windows-specific section near the end of this
document.]

** When you debug Emacs with GDB, you should start it in the directory
where the executable was made.  That directory has a .gdbinit file
that defines various "user-defined" commands for debugging Emacs.

** When you are trying to analyze failed assertions, it will be
essential to compile Emacs either completely without optimizations or
at least (when using GCC) with the -fno-crossjumping option.  Failure
to do so may make the compiler recycle the same abort call for all
assertions in a given function, rendering the stack backtrace useless
for identifying the specific failed assertion.

** It is a good idea to run Emacs under GDB (or some other suitable
debugger) *all the time*.  Then, when Emacs crashes, you will be able
to debug the live process, not just a core dump.  (This is especially
important on systems which don't support core files, and instead print
just the registers and some stack addresses.)

** If Emacs hangs, or seems to be stuck in some infinite loop, typing
"kill -TSTP PID", where PID is the Emacs process ID, will cause GDB to
kick in, provided that you run under GDB.

** Getting control to the debugger

`Fsignal' is a very useful place to put a breakpoint in.
All Lisp errors go through there.

It is useful, when debugging, to have a guaranteed way to return to
the debugger at any time.  When using X, this is easy: type C-z at the
window where Emacs is running under GDB, and it will stop Emacs just
as it would stop any ordinary program.  When Emacs is running in a
terminal, things are not so easy.

The src/.gdbinit file in the Emacs distribution arranges for SIGINT
(C-g in Emacs) to be passed to Emacs and not give control back to GDB.
On modern POSIX systems, you can override that with this command:

   handle SIGINT stop nopass

After this `handle' command, SIGINT will return control to GDB.  If
you want the C-g to cause a QUIT within Emacs as well, omit the
`nopass'.

A technique that can work when `handle SIGINT' does not is to store
the code for some character into the variable stop_character.  Thus,

    set stop_character = 29

makes Control-] (decimal code 29) the stop character.
Typing Control-] will cause immediate stop.  You cannot
use the set command until the inferior process has been started.
Put a breakpoint early in `main', or suspend the Emacs,
to get an opportunity to do the set command.

When Emacs is running in a terminal, it is useful to use a separate terminal
for the debug session.  This can be done by starting Emacs as usual, then
attaching to it from gdb with the `attach' command which is explained in the
node "Attach" of the GDB manual.

** Examining Lisp object values.

When you have a live process to debug, and it has not encountered a
fatal error, you can use the GDB command `pr'.  First print the value
in the ordinary way, with the `p' command.  Then type `pr' with no
arguments.  This calls a subroutine which uses the Lisp printer.

You can also use `pp value' to print the emacs value directly.

Note: It is not a good idea to try `pr' or `pp' if you know that Emacs
is in deep trouble: its stack smashed (e.g., if it encountered SIGSEGV
due to stack overflow), or crucial data structures, such as `obarray',
corrupted, etc.  In such cases, the Emacs subroutine called by `pr'
might make more damage, like overwrite some data that is important for
debugging the original problem.

Also, on some systems it is impossible to use `pr' if you stopped
Emacs while it was inside `select'.  This is in fact what happens if
you stop Emacs while it is waiting.  In such a situation, don't try to
use `pr'.  Instead, use `s' to step out of the system call.  Then
Emacs will be between instructions and capable of handling `pr'.

If you can't use `pr' command, for whatever reason, you can fall back
on lower-level commands.  Use the `xtype' command to print out the
data type of the last data value.  Once you know the data type, use
the command that corresponds to that type.  Here are these commands:

    xint xptr xwindow xmarker xoverlay xmiscfree xintfwd xboolfwd xobjfwd
    xbufobjfwd xkbobjfwd xbuflocal xbuffer xsymbol xstring xvector xframe
    xwinconfig xcompiled xcons xcar xcdr xsubr xprocess xfloat xscrollbar

Each one of them applies to a certain type or class of types.
(Some of these types are not visible in Lisp, because they exist only
internally.)

Each x... command prints some information about the value, and
produces a GDB value (subsequently available in $) through which you
can get at the rest of the contents.

In general, most of the rest of the contents will be additional Lisp
objects which you can examine in turn with the x... commands.

Even with a live process, these x...  commands are useful for
examining the fields in a buffer, window, process, frame or marker.
Here's an example using concepts explained in the node "Value History"
of the GDB manual to print values associated with the variable
called frame.  First, use these commands:

    cd src
    gdb emacs
    b set_frame_buffer_list
    r -q

Then Emacs hits the breakpoint:

    (gdb) p frame
    $1 = 139854428
    (gdb) xtype
    Lisp_Vectorlike
    PVEC_FRAME
    (gdb) xframe
    $2 = (struct frame *) 0x8560258
    (gdb) p *$
    $3 = {
      size = 1073742931,
      next = 0x85dfe58,
      name = 140615219,
      [...]
    }
    (gdb) p $3->name
    $4 = 140615219

Now we can use `pr' to print the name of the frame:

    (gdb) pr
    "[e-mail]"

The Emacs C code heavily uses macros defined in lisp.h.  So suppose
we want the address of the l-value expression near the bottom of
`add_command_key' from keyboard.c:

  XVECTOR (this_command_keys)->contents[this_command_key_count++] = key;

XVECTOR is a macro, and therefore GDB does not know about it.
GDB cannot evaluate "p XVECTOR (this_command_keys)".

However, you can use the xvector command in GDB to get the same
result.  Here is how:

    (gdb) p this_command_keys
    $1 = 1078005760
    (gdb) xvector
    $2 = (struct Lisp_Vector *) 0x411000
    0
    (gdb) p $->contents[this_command_key_count]
    $3 = 1077872640
    (gdb) p &$
    $4 = (int *) 0x411008

Here's a related example of macros and the GDB `define' command.
There are many Lisp vectors such as `recent_keys', which contains the
last 100 keystrokes.  We can print this Lisp vector

p recent_keys
pr

But this may be inconvenient, since `recent_keys' is much more verbose
than `C-h l'.  We might want to print only the last 10 elements of
this vector.  `recent_keys' is updated in keyboard.c by the command

  XVECTOR (recent_keys)->contents[recent_keys_index] = c;

So we define a GDB command `xvector-elts', so the last 10 keystrokes
are printed by

    xvector-elts recent_keys recent_keys_index 10

where you can define xvector-elts as follows:

    define xvector-elts
    set $i = 0
    p $arg0
    xvector
    set $foo = $
    while $i < $arg2
    p $foo->contents[$arg1-($i++)]
    pr
    end
    document xvector-elts
    Prints a range of elements of a Lisp vector.
    xvector-elts  v n i
    prints `i' elements of the vector `v' ending at the index `n'.
    end

** Getting Lisp-level backtrace information within GDB

The most convenient way is to use the `xbacktrace' command.  This
shows the names of the Lisp functions that are currently active.

If that doesn't work (e.g., because the `backtrace_list' structure is
corrupted), type "bt" at the GDB prompt, to produce the C-level
backtrace, and look for stack frames that call Ffuncall.  Select them
one by one in GDB, by typing "up N", where N is the appropriate number
of frames to go up, and in each frame that calls Ffuncall type this:

   p *args
   pr

This will print the name of the Lisp function called by that level
of function calling.

By printing the remaining elements of args, you can see the argument
values.  Here's how to print the first argument:

   p args[1]
   pr

If you do not have a live process, you can use xtype and the other
x...  commands such as xsymbol to get such information, albeit less
conveniently.  For example:

   p *args
   xtype

and, assuming that "xtype" says that args[0] is a symbol:

   xsymbol

** Debugging what happens while preloading and dumping Emacs

Type `gdb temacs' and start it with `r -batch -l loadup dump'.

If temacs actually succeeds when running under GDB in this way, do not
try to run the dumped Emacs, because it was dumped with the GDB
breakpoints in it.

** Debugging `temacs'

Debugging `temacs' is useful when you want to establish whether a
problem happens in an undumped Emacs.  To run `temacs' under a
debugger, type "gdb temacs", then start it with `r -batch -l loadup'.

** If you encounter X protocol errors

Try evaluating (x-synchronize t).  That puts Emacs into synchronous
mode, where each Xlib call checks for errors before it returns.  This
mode is much slower, but when you get an error, you will see exactly
which call really caused the error.

You can start Emacs in a synchronous mode by invoking it with the -xrm
option, like this:

    emacs -xrm "emacs.synchronous: true"

Setting a breakpoint in the function `x_error_quitter' and looking at
the backtrace when Emacs stops inside that function will show what
code causes the X protocol errors.

Some bugs related to the X protocol disappear when Emacs runs in a
synchronous mode.  To track down those bugs, we suggest the following
procedure:

  - Run Emacs under a debugger and put a breakpoint inside the
    primitive function which, when called from Lisp, triggers the X
    protocol errors.  For example, if the errors happen when you
    delete a frame, put a breakpoint inside `Fdelete_frame'.

  - When the breakpoint breaks, step through the code, looking for
    calls to X functions (the ones whose names begin with "X" or
    "Xt" or "Xm").

  - Insert calls to `XSync' before and after each call to the X
    functions, like this:

       XSync (f->output_data.x->display_info->display, 0);

    where `f' is the pointer to the `struct frame' of the selected
    frame, normally available via XFRAME (selected_frame).  (Most
    functions which call X already have some variable that holds the
    pointer to the frame, perhaps called `f' or `sf', so you shouldn't
    need to compute it.)

    If your debugger can call functions in the program being debugged,
    you should be able to issue the calls to `XSync' without recompiling
    Emacs.  For example, with GDB, just type:

       call XSync (f->output_data.x->display_info->display, 0)

    before and immediately after the suspect X calls.  If your
    debugger does not support this, you will need to add these pairs
    of calls in the source and rebuild Emacs.

    Either way, systematically step through the code and issue these
    calls until you find the first X function called by Emacs after
    which a call to `XSync' winds up in the function
    `x_error_quitter'.  The first X function call for which this
    happens is the one that generated the X protocol error.

  - You should now look around this offending X call and try to figure
    out what is wrong with it.

** If Emacs causes errors or memory leaks in your X server

You can trace the traffic between Emacs and your X server with a tool
like xmon, available at ftp://ftp.x.org/contrib/devel_tools/.

Xmon can be used to see exactly what Emacs sends when X protocol errors
happen.  If Emacs causes the X server memory usage to increase you can
use xmon to see what items Emacs creates in the server (windows,
graphical contexts, pixmaps) and what items Emacs delete.  If there
are consistently more creations than deletions, the type of item
and the activity you do when the items get created can give a hint where
to start debugging.

** If the symptom of the bug is that Emacs fails to respond

Don't assume Emacs is `hung'--it may instead be in an infinite loop.
To find out which, make the problem happen under GDB and stop Emacs
once it is not responding.  (If Emacs is using X Windows directly, you
can stop Emacs by typing C-z at the GDB job.)  Then try stepping with
`step'.  If Emacs is hung, the `step' command won't return.  If it is
looping, `step' will return.

If this shows Emacs is hung in a system call, stop it again and
examine the arguments of the call.  If you report the bug, it is very
important to state exactly where in the source the system call is, and
what the arguments are.

If Emacs is in an infinite loop, try to determine where the loop
starts and ends.  The easiest way to do this is to use the GDB command
`finish'.  Each time you use it, Emacs resumes execution until it
exits one stack frame.  Keep typing `finish' until it doesn't
return--that means the infinite loop is in the stack frame which you
just tried to finish.

Stop Emacs again, and use `finish' repeatedly again until you get back
to that frame.  Then use `next' to step through that frame.  By
stepping, you will see where the loop starts and ends.  Also, examine
the data being used in the loop and try to determine why the loop does
not exit when it should.

** If certain operations in Emacs are slower than they used to be, here
is some advice for how to find out why.

Stop Emacs repeatedly during the slow operation, and make a backtrace
each time.  Compare the backtraces looking for a pattern--a specific
function that shows up more often than you'd expect.

If you don't see a pattern in the C backtraces, get some Lisp
backtrace information by typing "xbacktrace" or by looking at Ffuncall
frames (see above), and again look for a pattern.

When using X, you can stop Emacs at any time by typing C-z at GDB.
When not using X, you can do this with C-g.  On non-Unix platforms,
such as MS-DOS, you might need to press C-BREAK instead.

** If GDB does not run and your debuggers can't load Emacs.

On some systems, no debugger can load Emacs with a symbol table,
perhaps because they all have fixed limits on the number of symbols
and Emacs exceeds the limits.  Here is a method that can be used
in such an extremity.  Do

    nm -n temacs > nmout
    strip temacs
    adb temacs
    0xd:i
    0xe:i
    14:i
    17:i
    :r -l loadup   (or whatever)

It is necessary to refer to the file `nmout' to convert
numeric addresses into symbols and vice versa.

It is useful to be running under a window system.
Then, if Emacs becomes hopelessly wedged, you can create
another window to do kill -9 in.  kill -ILL is often
useful too, since that may make Emacs dump core or return
to adb.


** Debugging incorrect screen updating.

To debug Emacs problems that update the screen wrong, it is useful
to have a record of what input you typed and what Emacs sent to the
screen.  To make these records, do

(open-dribble-file "~/.dribble")
(open-termscript "~/.termscript")

The dribble file contains all characters read by Emacs from the
terminal, and the termscript file contains all characters it sent to
the terminal.  The use of the directory `~/' prevents interference
with any other user.

If you have irreproducible display problems, put those two expressions
in your ~/.emacs file.  When the problem happens, exit the Emacs that
you were running, kill it, and rename the two files.  Then you can start
another Emacs without clobbering those files, and use it to examine them.

An easy way to see if too much text is being redrawn on a terminal is to
evaluate `(setq inverse-video t)' before you try the operation you think
will cause too much redrawing.  This doesn't refresh the screen, so only
newly drawn text is in inverse video.

The Emacs display code includes special debugging code, but it is
normally disabled.  You can enable it by building Emacs with the
pre-processing symbol GLYPH_DEBUG defined.  Here's one easy way,
suitable for Unix and GNU systems, to build such a debugging version:

	 MYCPPFLAGS='-DGLYPH_DEBUG=1' make

Building Emacs like that activates many assertions which scrutinize
display code operation more than Emacs does normally.  (To see the
code which tests these assertions, look for calls to the `xassert'
macros.)  Any assertion that is reported to fail should be
investigated.

Building with GLYPH_DEBUG defined also defines several helper
functions which can help debugging display code.  One such function is
`dump_glyph_matrix'.  If you run Emacs under GDB, you can print the
contents of any glyph matrix by just calling that function with the
matrix as its argument.  For example, the following command will print
the contents of the current matrix of the window whose pointer is in
`w':

  (gdb) p dump_glyph_matrix (w->current_matrix, 2)

(The second argument 2 tells dump_glyph_matrix to print the glyphs in
a long form.)  You can dump the selected window's current glyph matrix
interactively with "M-x dump-glyph-matrix RET"; see the documentation
of this function for more details.

Several more functions for debugging display code are available in
Emacs compiled with GLYPH_DEBUG defined; type "C-h f dump- TAB" and
"C-h f trace- TAB" to see the full list.

When you debug display problems running emacs under X, you can use
the `ff' command to flush all pending display updates to the screen.


** Debugging LessTif

If you encounter bugs whereby Emacs built with LessTif grabs all mouse
and keyboard events, or LessTif menus behave weirdly, it might be
helpful to set the `DEBUGSOURCES' and `DEBUG_FILE' environment
variables, so that one can see what LessTif was doing at this point.
For instance

  export DEBUGSOURCES="RowColumn.c:MenuShell.c:MenuUtil.c"
  export DEBUG_FILE=/usr/tmp/LESSTIF_TRACE
  emacs &

causes LessTif to print traces from the three named source files to a
file in `/usr/tmp' (that file can get pretty large).  The above should
be typed at the shell prompt before invoking Emacs, as shown by the
last line above.

Running GDB from another terminal could also help with such problems.
You can arrange for GDB to run on one machine, with the Emacs display
appearing on another.  Then, when the bug happens, you can go back to
the machine where you started GDB and use the debugger from there.


** Debugging problems which happen in GC

The array `last_marked' (defined on alloc.c) can be used to display up
to 500 last objects marked by the garbage collection process.
Whenever the garbage collector marks a Lisp object, it records the
pointer to that object in the `last_marked' array.  The variable
`last_marked_index' holds the index into the `last_marked' array one
place beyond where the pointer to the very last marked object is
stored.

The single most important goal in debugging GC problems is to find the
Lisp data structure that got corrupted.  This is not easy since GC
changes the tag bits and relocates strings which make it hard to look
at Lisp objects with commands such as `pr'.  It is sometimes necessary
to convert Lisp_Object variables into pointers to C struct's manually.
Use the `last_marked' array and the source to reconstruct the sequence
that objects were marked.

Once you discover the corrupted Lisp object or data structure, it is
useful to look at it in a fresh Emacs session and compare its contents
with a session that you are debugging.

** Debugging problems with non-ASCII characters

If you experience problems which seem to be related to non-ASCII
characters, such as \201 characters appearing in the buffer or in your
files, set the variable byte-debug-flag to t.  This causes Emacs to do
some extra checks, such as look for broken relations between byte and
character positions in buffers and strings; the resulting diagnostics
might pinpoint the cause of the problem.

** Debugging the TTY (non-windowed) version

The most convenient method of debugging the character-terminal display
is to do that on a window system such as X.  Begin by starting an
xterm window, then type these commands inside that window:

  $ tty
  $ echo $TERM

Let's say these commands print "/dev/ttyp4" and "xterm", respectively.

Now start Emacs (the normal, windowed-display session, i.e. without
the `-nw' option), and invoke "M-x gdb RET emacs RET" from there.  Now
type these commands at GDB's prompt:

  (gdb) set args -nw -t /dev/ttyp4
  (gdb) set environment TERM xterm
  (gdb) run

The debugged Emacs should now start in no-window mode with its display
directed to the xterm window you opened above.

Similar arrangement is possible on a character terminal by using the
`screen' package.

** Running Emacs built with malloc debugging packages

If Emacs exhibits bugs that seem to be related to use of memory
allocated off the heap, it might be useful to link Emacs with a
special debugging library, such as Electric Fence (a.k.a. efence) or
GNU Checker, which helps find such problems.

Emacs compiled with such packages might not run without some hacking,
because Emacs replaces the system's memory allocation functions with
its own versions, and because the dumping process might be
incompatible with the way these packages use to track allocated
memory.  Here are some of the changes you might find necessary
(SYSTEM-NAME and MACHINE-NAME are the names of your OS- and
CPU-specific headers in the subdirectories of `src'):

  - In src/s/SYSTEM-NAME.h add "#define SYSTEM_MALLOC".

  - In src/m/MACHINE-NAME.h add "#define CANNOT_DUMP" and
    "#define CANNOT_UNEXEC".

  - Configure with a different --prefix= option.  If you use GCC,
    version 2.7.2 is preferred, as some malloc debugging packages
    work a lot better with it than with 2.95 or later versions.

  - Type "make" then "make -k install".

  - If required, invoke the package-specific command to prepare
    src/temacs for execution.

  - cd ..; src/temacs

(Note that this runs `temacs' instead of the usual `emacs' executable.
This avoids problems with dumping Emacs mentioned above.)

Some malloc debugging libraries might print lots of false alarms for
bitfields used by Emacs in some data structures.  If you want to get
rid of the false alarms, you will have to hack the definitions of
these data structures on the respective headers to remove the `:N'
bitfield definitions (which will cause each such field to use a full
int).

** Some suggestions for debugging on MS Windows:

   (written by Marc Fleischeuers, Geoff Voelker and Andrew Innes)

To debug Emacs with Microsoft Visual C++, you either start emacs from
the debugger or attach the debugger to a running emacs process.

To start emacs from the debugger, you can use the file bin/debug.bat.
The Microsoft Developer studio will start and under Project, Settings,
Debug, General you can set the command-line arguments and Emacs's
startup directory.  Set breakpoints (Edit, Breakpoints) at Fsignal and
other functions that you want to examine.  Run the program (Build,
Start debug).  Emacs will start and the debugger will take control as
soon as a breakpoint is hit.

You can also attach the debugger to an already running Emacs process.
To do this, start up the Microsoft Developer studio and select Build,
Start debug, Attach to process.  Choose the Emacs process from the
list.  Send a break to the running process (Debug, Break) and you will
find that execution is halted somewhere in user32.dll.  Open the stack
trace window and go up the stack to w32_msg_pump.  Now you can set
breakpoints in Emacs (Edit, Breakpoints).  Continue the running Emacs
process (Debug, Step out) and control will return to Emacs, until a
breakpoint is hit.

To examine the contents of a Lisp variable, you can use the function
'debug_print'.  Right-click on a variable, select QuickWatch (it has
an eyeglass symbol on its button in the toolbar), and in the text
field at the top of the window, place 'debug_print(' and ')' around
the expression.  Press 'Recalculate' and the output is sent to stderr,
and to the debugger via the OutputDebugString routine.  The output
sent to stderr should be displayed in the console window that was
opened when the emacs.exe executable was started.  The output sent to
the debugger should be displayed in the 'Debug' pane in the Output
window.  If Emacs was started from the debugger, a console window was
opened at Emacs' startup; this console window also shows the output of
'debug_print'.

For example, start and run Emacs in the debugger until it is waiting
for user input.  Then click on the `Break' button in the debugger to
halt execution.  Emacs should halt in `ZwUserGetMessage' waiting for
an input event.  Use the `Call Stack' window to select the procedure
`w32_msp_pump' up the call stack (see below for why you have to do
this).  Open the QuickWatch window and enter
"debug_print(Vexec_path)".  Evaluating this expression will then print
out the contents of the Lisp variable `exec-path'.

If QuickWatch reports that the symbol is unknown, then check the call
stack in the `Call Stack' window.  If the selected frame in the call
stack is not an Emacs procedure, then the debugger won't recognize
Emacs symbols.  Instead, select a frame that is inside an Emacs
procedure and try using `debug_print' again.

If QuickWatch invokes debug_print but nothing happens, then check the
thread that is selected in the debugger.  If the selected thread is
not the last thread to run (the "current" thread), then it cannot be
used to execute debug_print.  Use the Debug menu to select the current
thread and try using debug_print again.  Note that the debugger halts
execution (e.g., due to a breakpoint) in the context of the current
thread, so this should only be a problem if you've explicitly switched
threads.

It is also possible to keep appropriately masked and typecast Lisp
symbols in the Watch window, this is more convenient when steeping
though the code.  For instance, on entering apply_lambda, you can
watch (struct Lisp_Symbol *) (0xfffffff & args[0]).

Optimizations often confuse the MS debugger.  For example, the
debugger will sometimes report wrong line numbers, e.g., when it
prints the backtrace for a crash.  It is usually best to look at the
disassembly to determine exactly what code is being run--the
disassembly will probably show several source lines followed by a
block of assembler for those lines.  The actual point where Emacs
crashes will be one of those source lines, but not neccesarily the one
that the debugger reports.

Another problematic area with the MS debugger is with variables that
are stored in registers: it will sometimes display wrong values for
those variables.  Usually you will not be able to see any value for a
register variable, but if it is only being stored in a register
temporarily, you will see an old value for it.  Again, you need to
look at the disassembly to determine which registers are being used,
and look at those registers directly, to see the actual current values
of these variables.

;;; arch-tag: fbf32980-e35d-481f-8e4c-a2eca2586e6b