Re: bzr emacs(24.4.50.1) crashes when using auto-revert-tail-mode

[Eli Zaretskii <eliz at gnu dot org>]

> Date: Thu, 19 Jun 2014 09:11:37 -0400
> From: Ken Brown
> 
> On 6/18/2014 1:46 PM, Ken Brown wrote:
> > On 6/18/2014 11:06 AM, Filipp Gunbin wrote:
> >> I'm not sure whether this is Cygwin-specific, but I'm not able to test
> >> it on other OS, so here are the steps to reproduce:
> >>
> >> emacs -Q
> >> C-x C-r <some_file>
> > 
> > You mean C-x C-f
> > 
> >> M-x auto-revert-tail-mode
> >> wait for few seconds -> emacs crashes
> > 
> > I can confirm this, but on 32-bit Cygwin only; there's no crash on 
> > 64-bit Cygwin.  (This is a refreshing change from the emacs crashes 
> > people have been reporting on 64-bit Cygwin.)
> > 
> > Here's the backtrace:
> > 
> > #0  0x00000000 in ?? ()
> > No symbol table info available.
> > #1  0x610f842f in pthread_mutex::lock (this=0x0)
> >      at /usr/src/debug/cygwin-1.7.30-1/winsup/cygwin/thread.cc:1745
> >          self = <optimized out>
> >          result = <optimized out>
> >          __PRETTY_FUNCTION__ = "int pthread_mutex::lock()"
> > #2  0x00000201 in ?? ()
> > No symbol table info available.
> > 
> > Lisp Backtrace:
> > "gfile-add-watch" (0x288dc8)
> > "file-notify-add-watch" (0x2890c8)
> > "byte-code" (0x289340)
> > "auto-revert-notify-add-watch" (0x289778)
> > "auto-revert-buffers" (0x289b4c)
> > "apply" (0x289b48)
> > "byte-code" (0x289dc0)
> > "timer-event-handler" (0x28a1fc)
> > 
> > I'll look into this.  In the meantime, you can work around it by 
> > configuring --with-file-notification=no.
> 
> I'm afraid I ran into a brick wall trying to debug this.  I wanted to see what gfile-add-watch was doing, so I ran emacs under gdb with a breakpoint at Fgfile_add_watch and then a breakpoint at g_file_monitor (a Glib function called by Fgfile_add_watch).  When I tried to step through this function, I hit an assertion violation in gdb.  This is repeatable.  A log of the gdb session is appended below.
> 
> The problem occurs only on 32-bit Cygwin.  On 64-bit Cygwin, I can step through Fgfile_add_watch without a problem.

Debugging Glib applications is not for the faint at heart.  Its
OO-like objects intentionally conceal their innards from the outside
world, and appear as opaque objects in the debugger, unless you
actually step deep enough into the methods.

I generally find myself unable to debug Glib, unless I build Glib
without optimizations and with -g3.

I'd suggest to begin with looking at this from the Emacs side.  Do I
understand correctly that this works with previous versions of w32
Emacs?  If someone can show which past version of Cygwin-w32 build of
Emacs worked with gfilenotify, then looking at the differences between
then and now might show the way.

If no one can tell what past version worked, then I suggest to build
Glib with -O0 -g3, and debug that.  From the crash traceback, it
sounds like pthreads might also be involved, so I recommend to build
that with -O0 -g3 as well.  Then see if the backtrace becomes more
useful.

Also, does the 64-bit build use the same versions of Glib and
pthreads?  If not, perhaps using the same versions will resolve the
problem.

> (gdb) b Fgfile_add_watch
> Breakpoint 1 at 0x5f7a3f: file /usr/src/debug/emacs-24.3.91-2/src/gfilenotify.c, line 170.
> (gdb) r -Q
> Starting program: /usr/bin/emacs-w32.exe -Q
> [...]
> Breakpoint 1, Fgfile_add_watch (file=-2145740495, flags=13798510,
>     callback=-2146880150)
>     at /usr/src/debug/emacs-24.3.91-2/src/gfilenotify.c:170
> 170       GFileMonitorFlags gflags = G_FILE_MONITOR_NONE;
> (gdb) b g_file_monitor
> Breakpoint 2 at 0x6a357e50: file /usr/src/debug/glib2.0-2.38.2-2/gio/gfile.c, line 5338.
> (gdb) c
> Continuing.
> [...]
> Breakpoint 2, g_file_monitor (file=0x80061530,
>     flags=(G_FILE_MONITOR_WATCH_MOUNTS | G_FILE_MONITOR_SEND_MOVED),
>     cancellable=0x0, error=0x0)
>     at /usr/src/debug/glib2.0-2.38.2-2/gio/gfile.c:5338
> 5338    {
> (gdb) n
> 5339      if (g_file_query_file_type (file, 0, cancellable) == G_FILE_TYPE_DIRECTORY)
> (gdb) n
> 5338    {
> (gdb) n
> 5339      if (g_file_query_file_type (file, 0, cancellable) == G_FILE_TYPE_DIRECTORY)
> (gdb) n
> 5340        return g_file_monitor_directory (file,
> (gdb) n
> 5339      if (g_file_query_file_type (file, 0, cancellable) == G_FILE_TYPE_DIRECTORY)
> (gdb) n
> 5341                                         flags & ~G_FILE_MONITOR_WATCH_HARD_LINKS,
> (gdb) n
> 5340        return g_file_monitor_directory (file,
> (gdb) n
> 5345    }
> (gdb) n
> 5340        return g_file_monitor_directory (file,
> (gdb) n
> g_file_monitor_directory (file=0x80061530,
>     flags=(G_FILE_MONITOR_WATCH_MOUNTS | G_FILE_MONITOR_SEND_MOVED),
>     cancellable=0x0, error=0x0)
>     at /usr/src/debug/glib2.0-2.38.2-2/gio/gfile.c:5235
> 5235    {
> (gdb) n
> 5238      g_return_val_if_fail (G_IS_FILE (file), NULL);
> (gdb) n
> /netrel/src/gdb-7.6.50-4/gdb/infrun.c:1942: internal-error: resume: Assertion `pc_in_thread_step_range (pc, tp)' failed.
> A problem internal to GDB has been detected,
> further debugging may prove unreliable.
> Quit this debugging session? (y or n) y

Why did you need to step through the Glib code?  AFAIU, the file
monitor did trigger, so what I would first look at is the data it
delivers back to Emacs, not how the monitor works internally.  Did you
try that?

--
Problem reports:       http://cygwin.com/problems.html
FAQ:                   http://cygwin.com/faq/
Documentation:         http://cygwin.com/docs.html
Unsubscribe info:      http://cygwin.com/ml/#unsubscribe-simple