[patchv2] Fix SIGTERM signal safety (PR gdb/15358) [refresh]

  Hi,

the post several minutes ago (not the year old one) had accidentally corrupted
testcase.

------------------------------------------------------------------------------

reposting with no changes, the last post is a year old one, it had been already
approved by Tom that time so I will check it in in some time.
	[patch] Fix SIGTERM signal safety (PR gdb/15358)
	https://sourceware.org/ml/gdb-patches/2013-07/msg00094.html
	Message-ID: <20130702200010.GA23478@host2.jankratochvil.net>

The statement about PR remote/15297 dependency is IMO false, this patch works
on its own.  The PR remote/15297 patch remains unfinished/pending, but that is
off-topic here.

------------------------------------------------------------------------------

gdb deadlock due to gdb calling calloc() in signal handler
http://sourceware.org/bugzilla/show_bug.cgi?id=15358

Patch depends in its functionality on:
	[patchv2 2/2] Fix CTRL-C for remote.c (PR remote/15297)
	http://sourceware.org/ml/gdb-patches/2013-06/msg00943.html
	Message-ID: <20130630181110.GB29548@host2.jankratochvil.net>

async mode seems easy.

The sync mode is a bit difficult - assuming it is safe to call quit_force from
any place of QUIT;.  OTOH the patch above assumes it can do:
	  if (check_quit_flag ())
	    send_interrupt_sequence ();
which clears quit_flag but we used set_quit_flag () to call quit_force, not
just to throw the quit exception.  This is why QUIT now checks also for
SYNC_QUIT_FORCE_RUN.

The change in linux-nat.c comes from testing i386-linux-nat.c (therefore
32-bit host GDB).  i386_linux_resume there calls QUIT; via target_read ().
This is a bug on its own, filed as:
	http://sourceware.org/bugzilla/show_bug.cgi?id=15713

But I have seen another bug in linux-nat.c, it was depending on PTRACE_KILL
but at least Linux kernel ptrace expert Oleg Nesterov considers PTRACE_KILL
superseded by kill(SIGKILL).  Therefore I used there (also) more safe SIGKILL 
so that the possibly inconsistent state of inferior from i386_linux_resume
does not matter.


No regressions on {x86_64,x86_64-m32,i686}-fedora19pre-linux-gnu and in
gdbserver mode.


Thanks,
Jan


gdb/
2013-07-02  Jan Kratochvil  <jan.kratochvil@redhat.com>

	PR gdb/15358
	* defs.h (sync_quit_force_run): New declaration.
	(QUIT): Check also SYNC_QUIT_FORCE_RUN.
	* event-top.c (async_sigterm_handler): New declaration.
	(async_sigterm_token): New variable.
	(async_init_signals): Create also async_sigterm_token.
	(async_sigterm_handler): New function.
	(sync_quit_force_run): New variable.
	(handle_sigterm): Replace quit_force call by other calls.
	* linux-nat.c (linux_nat_kill): Use kill_callback first.
	Extend the comment for stop_callback.
	* utils.c (quit): Call quit_force if SYNC_QUIT_FORCE_RUN.

gdb/testsuite/
2013-07-02  Jan Kratochvil  <jan.kratochvil@redhat.com>

	PR gdb/15358
	* gdb.base/gdb-sigterm.c: New file.
	* gdb.base/gdb-sigterm.exp: New file.
  

Jan Kratochvil writes:
 > [...]

Hi.
I have some comments and requests for this patch.

A high level comment is that I guess this patch means it's
now more important to fix the spots of gdb where QUITs aren't
called for long times (I was working on a common one, debug info
reading, but had to defer that to work on something else).

 > The sync mode is a bit difficult - assuming it is safe to call quit_force from
 > any place of QUIT;.  OTOH the patch above assumes it can do:
 > 	  if (check_quit_flag ())
 > 	    send_interrupt_sequence ();
 > which clears quit_flag but we used set_quit_flag () to call quit_force, not
 > just to throw the quit exception.  This is why QUIT now checks also for
 > SYNC_QUIT_FORCE_RUN.
 > 
 > The change in linux-nat.c comes from testing i386-linux-nat.c (therefore
 > 32-bit host GDB).  i386_linux_resume there calls QUIT; via target_read ().
 > This is a bug on its own, filed as:
 > 	http://sourceware.org/bugzilla/show_bug.cgi?id=15713
 > 
 > But I have seen another bug in linux-nat.c, it was depending on PTRACE_KILL
 > but at least Linux kernel ptrace expert Oleg Nesterov considers PTRACE_KILL
 > superseded by kill(SIGKILL).  Therefore I used there (also) more safe SIGKILL 
 > so that the possibly inconsistent state of inferior from i386_linux_resume
 > does not matter.
 > 
 > 
 > No regressions on {x86_64,x86_64-m32,i686}-fedora19pre-linux-gnu and in
 > gdbserver mode.
 > 
 > 
 > Thanks,
 > Jan
 > 
 > 
 > gdb/
 > 2013-07-02  Jan Kratochvil  <jan.kratochvil@redhat.com>
 > 
 > 	PR gdb/15358
 > 	* defs.h (sync_quit_force_run): New declaration.
 > 	(QUIT): Check also SYNC_QUIT_FORCE_RUN.
 > 	* event-top.c (async_sigterm_handler): New declaration.
 > 	(async_sigterm_token): New variable.
 > 	(async_init_signals): Create also async_sigterm_token.
 > 	(async_sigterm_handler): New function.
 > 	(sync_quit_force_run): New variable.
 > 	(handle_sigterm): Replace quit_force call by other calls.
 > 	* linux-nat.c (linux_nat_kill): Use kill_callback first.
 > 	Extend the comment for stop_callback.
 > 	* utils.c (quit): Call quit_force if SYNC_QUIT_FORCE_RUN.
 > 
 > gdb/testsuite/
 > 2013-07-02  Jan Kratochvil  <jan.kratochvil@redhat.com>
 > 
 > 	PR gdb/15358
 > 	* gdb.base/gdb-sigterm.c: New file.
 > 	* gdb.base/gdb-sigterm.exp: New file.
 > 
 > diff --git a/gdb/defs.h b/gdb/defs.h
 > index 480133d..47da43a 100644
 > --- a/gdb/defs.h
 > +++ b/gdb/defs.h
 > @@ -171,6 +171,9 @@ extern int check_quit_flag (void);
 >  /* * Set the quit flag.  */
 >  extern void set_quit_flag (void);
 >  
 > +/* Flag that function quit should call quit_force.  */
 > +extern volatile int sync_quit_force_run;
 > +
 >  extern int immediate_quit;
 >  
 >  extern void quit (void);
 > @@ -183,7 +186,7 @@ extern void quit (void);
 >     needed.  */
 >  
 >  #define QUIT { \
 > -  if (check_quit_flag ()) quit (); \
 > +  if (check_quit_flag () || sync_quit_force_run) quit (); \
 >    if (deprecated_interactive_hook) deprecated_interactive_hook (); \
 >  }
 >  
 > diff --git a/gdb/event-top.c b/gdb/event-top.c
 > index fbe89bd..af1562c 100644
 > --- a/gdb/event-top.c
 > +++ b/gdb/event-top.c
 > @@ -72,6 +72,7 @@ static void async_float_handler (gdb_client_data);
 >  #ifdef STOP_SIGNAL
 >  static void async_stop_sig (gdb_client_data);
 >  #endif
 > +static void async_sigterm_handler (gdb_client_data arg);
 >  
 >  /* Readline offers an alternate interface, via callback
 >     functions.  These are all included in the file callback.c in the
 > @@ -135,6 +136,7 @@ static struct async_signal_handler *sigfpe_token;
 >  #ifdef STOP_SIGNAL
 >  static struct async_signal_handler *sigtstp_token;
 >  #endif
 > +static struct async_signal_handler *async_sigterm_token;
 >  
 >  /* Structure to save a partially entered command.  This is used when
 >     the user types '\' at the end of a command line.  This is necessary
 > @@ -770,6 +772,8 @@ async_init_signals (void)
 >      create_async_signal_handler (async_stop_sig, NULL);
 >  #endif
 >  
 > +  async_sigterm_token =
 > +    create_async_signal_handler (async_sigterm_handler, NULL);
 >  }

I realize SIGTERM isn't handled exactly like SIGINT, but IWBN
to keep all the SIGTERM bits together in this function.
How about moving the above snippet to after this:

  signal (SIGINT, handle_sigint);
  sigint_token =
    create_async_signal_handler (async_request_quit, NULL);
  signal (SIGTERM, handle_sigterm);

btw, the rule has been effectively changed to require = being
put on the next line.
While I certainly wouldn't ask for whitespace change in this patch,
[IOW I'm not asking for the existing code to be fixed]
I'm curious what the rule is for handling this here.
For the patch at hand, should it be written as this:

  async_sigterm_token =
    create_async_signal_handler (async_sigterm_handler, NULL);

or this:

  async_sigterm_token
    = create_async_signal_handler (async_sigterm_handler, NULL);

?
I honestly don't know.
It's not important of course, just wondering.

 >  
 >  /* Tell the event loop what to do if SIGINT is received.
 > @@ -797,13 +801,31 @@ handle_sigint (int sig)
 >    gdb_call_async_signal_handler (sigint_token, immediate_quit);
 >  }
 >  
 > +/* Handle GDB exit upon receiving SIGTERM if target_can_async_p ().  */
 > +
 > +static void
 > +async_sigterm_handler (gdb_client_data arg)
 > +{
 > +  quit_force (NULL, stdin == instream);
 > +}
 > +
 > +/* See defs.h.  */
 > +volatile int sync_quit_force_run;
 > +
 >  /* Quit GDB if SIGTERM is received.
 >     GDB would quit anyway, but this way it will clean up properly.  */
 >  void
 >  handle_sigterm (int sig)
 >  {
 >    signal (sig, handle_sigterm);
 > -  quit_force ((char *) 0, stdin == instream);

I think we need a comment here explaining *why* we can't
just call quit_force here.

 > +
 > +  if (target_can_async_p ())
 > +    mark_async_signal_handler (async_sigterm_token);
 > +  else
 > +    {
 > +      sync_quit_force_run = 1;
 > +      set_quit_flag ();
 > +    }
 >  }
 >  
 >  /* Do the quit.  All the checks have been done by the caller.  */
 > diff --git a/gdb/linux-nat.c b/gdb/linux-nat.c
 > index b615423..ec84188 100644
 > --- a/gdb/linux-nat.c
 > +++ b/gdb/linux-nat.c
 > @@ -3777,8 +3777,15 @@ linux_nat_kill (struct target_ops *ops)
 >      {
 >        ptid_t ptid = pid_to_ptid (ptid_get_pid (inferior_ptid));
 >  
 > +      /* Kill all LWP's before trying to stop them.  In rare cases the
 > +	 lwp_info state may not match the inferior and
 > +	 stop_wait_callback could lock up.  */
 > +      iterate_over_lwps (ptid, kill_callback, NULL);
 > +
 >        /* Stop all threads before killing them, since ptrace requires
 > -	 that the thread is stopped to sucessfully PTRACE_KILL.  */
 > +	 that the thread is stopped to sucessfully PTRACE_KILL.
 > +	 kill_callback normally already turned the inferior into a zombie
 > +	 except for old Linux kernels 2.4.x.  */
 >        iterate_over_lwps (ptid, stop_callback, NULL);
 >        /* ... and wait until all of them have reported back that
 >  	 they're no longer running.  */

This part feels sufficiently outside the scope of this patch
that IWBN if this were a separate patch.

 > diff --git a/gdb/utils.c b/gdb/utils.c
 > index 9d068c5..364470c 100644
 > --- a/gdb/utils.c
 > +++ b/gdb/utils.c
 > @@ -999,6 +999,12 @@ print_sys_errmsg (const char *string, int errcode)
 >  void
 >  quit (void)
 >  {
 > +  if (sync_quit_force_run)
 > +    {
 > +      sync_quit_force_run = 0;
 > +      quit_force (NULL, stdin == instream);
 > +    }
 > +

Bleah.  Not your problem, and obviously I'm not suggesting
fixing the names in this patch,  but "quit" is overloaded.
quit is really quit_current_command (or some such),
and quit_force is really quit_and_exit (or some such).

I raise the issue because the patch is conflating them,
and I'm wondering if that's a good idea.

 > [...]
 >
 > diff --git a/gdb/testsuite/gdb.base/gdb-sigterm.exp b/gdb/testsuite/gdb.base/gdb-sigterm.exp
 > new file mode 100644
 > index 0000000..8baeb96
 > --- /dev/null
 > +++ b/gdb/testsuite/gdb.base/gdb-sigterm.exp
 > @@ -0,0 +1,94 @@
 > +# This testcase is part of GDB, the GNU debugger.
 > +#
 > +# Copyright 2013 Free Software Foundation, Inc.
 > +#
 > +# This program is free software; you can redistribute it and/or modify
 > +# it under the terms of the GNU General Public License as published by
 > +# the Free Software Foundation; either version 3 of the License, or
 > +# (at your option) any later version.
 > +#
 > +# This program is distributed in the hope that it will be useful,
 > +# but WITHOUT ANY WARRANTY; without even the implied warranty of
 > +# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
 > +# GNU General Public License for more details.
 > +#
 > +# You should have received a copy of the GNU General Public License
 > +# along with this program.  If not, see <http://www.gnu.org/licenses/>.
 > +
 > +standard_testfile
 > +
 > +if { [build_executable ${testfile}.exp ${testfile}] == -1 } {
 > +    return -1
 > +}
 > +
 > +proc do_test { pass } {
 > +    global testfile gdb_prompt binfile pf_prefix
 > +
 > +    if ![runto_main] {
 > +	return -1
 > +    }
 > +
 > +    gdb_breakpoint "${testfile}.c:[gdb_get_line_number "loop-line" ${testfile}.c]" \
 > +		   temporary
 > +
 > +    # gdb_continue_to_breakpoint would print a pass message.
 > +    gdb_test "continue" "Temporary breakpoint .* loop-line .*" ""
 > +
 > +    gdb_test_no_output "set range-stepping off" ""
 > +    gdb_test_no_output "set debug infrun 1" ""
 > +    gdb_test_no_output "set debug lin-lwp 1" ""
 > +
 > +    set test "run a bit #$pass"
 > +    set abort 1
 > +    gdb_test_multiple "step" $test {
 > +	-re "infrun: stepping inside range" {
 > +	    # Suppress pass $test
 > +	    verbose -log "$pf_prefix $test"
 > +	    set abort 0
 > +	}
 > +    }
 > +    if $abort {
 > +	return
 > +    }
 > +
 > +    set gdb_pid [exp_pid -i [board_info host fileid]]
 > +    remote_exec host "kill -TERM ${gdb_pid}"
 > +
 > +    set test "expect eof #$pass"
 > +    set abort 1
 > +    set stepping 0
 > +    gdb_test_multiple "" $test {
 > +	eof {
 > +	    verbose -log "$pf_prefix $test"
 > +	    set abort 0
 > +	}
 > +	-re "infrun: stepping inside range" {
 > +	    incr stepping
 > +	    if { $stepping > 200 } {
 > +		fail "$test (stepping inside range $stepping times)"
 > +	    } else {
 > +		exp_continue
 > +	    }
 > +	}
 > +    }
 > +    if $abort {
 > +	return
 > +    }
 > +}
 > +
 > +
 > +for {set pass 0} {$pass < 50} {incr pass} {

How come 50 passes?
At the least a comment is required explaining why.

 > +
 > +    clean_restart ${testfile}
 > +    gdb_test_no_output "set target-async off" ""
 > +    with_test_prefix "sync" {
 > +        do_test $pass
 > +    }
 > +
 > +    clean_restart ${testfile}
 > +    gdb_test_no_output "set target-async on" ""
 > +    with_test_prefix "async" {
 > +        do_test $pass
 > +    }
 > +}
 > +pass "$pass SIGTERM passes"