[v2] Fix failure to detach if threads exit while detaching on linux

In this v2:

Fixed some style issues,
Remove stub target requirement
Fixed include of nat/linux-procfs in inf-ptrace

--
This patches fixes detaching on linux when some threads exit while we're
detaching with GDB and GDBServer.

What happened before is that as GDB/GDBserver would be detaching threads
one thread at a time and allowing them to continue, if one of these
detached threads called exit for example and the other threads were
destroyed GDB/GDBserver would still try and detach these exited threads
and fail with a message like: "Can't detach process." as ptrace could not
execute the operation.

This patch uses check_ptrace_stopped_lwp_gone or
linux_proc_pid_is_trace_stopped_nowarn like is used in the resume case to
avoid an error if this function detects that the ptrace failure is normal
since the thread has exited.

This patch adds the gdb.threads/detach-gone-thread.exp test for this case.

Tested on x86-linux with {unix, native-gdbserver, native-extended-gdbserver}

gdb/gdbserver/ChangeLog:

	* linux-low.c (check_ptrace_stopped_lwp_gone) Move up to be used
	by linux_detach_one_lwp.
	(linux_detach_one_lwp): Report the error only if
	check_ptrace_stopped_lwp_gone is false.

gdb/ChangeLog:

	* linux-nat.c (check_ptrace_stopped_lwp_gone): Move up to be used
	by detach_callback.
	(detach_callback): Report the error only if
	check_ptrace_stopped_lwp_gone is false.
	(linux_nat_detach): Likewise.

gdb/testsuite/ChangeLog:

	* gdb.threads/detach-gone-thread.c: New file.
	* gdb.threads/detach-gone-thread.exp: New file.
---
 gdb/gdbserver/linux-low.c                        | 73 ++++++++++---------
 gdb/linux-nat.c                                  | 93 +++++++++++++++---------
 gdb/testsuite/gdb.threads/detach-gone-thread.c   | 47 ++++++++++++
 gdb/testsuite/gdb.threads/detach-gone-thread.exp | 52 +++++++++++++
 4 files changed, 193 insertions(+), 72 deletions(-)
 create mode 100644 gdb/testsuite/gdb.threads/detach-gone-thread.c
 create mode 100644 gdb/testsuite/gdb.threads/detach-gone-thread.exp

On 06/03/2016 03:49 PM, Antoine Tremblay wrote:

> @@ -1491,7 +1523,28 @@ linux_nat_detach (struct target_ops *ops, const char *args, int from_tty)
>        linux_fork_detach (args, from_tty);
>      }
>    else
> -    linux_ops->to_detach (ops, args, from_tty);
> +    {
> +      TRY
> +	{
> +	  linux_ops->to_detach (ops, args, from_tty);
> +	}
> +      CATCH (ex, RETURN_MASK_ERROR)
> +	{
> +	  if (!check_ptrace_stopped_lwp_gone (main_lwp))
> +	    {
> +	      delete_lwp (main_lwp->ptid);

Is deleting the right thing to do?  I think we'll likely
crash, the next time we try to do something (e.g., try
to detach again).  Seems like on the gdbserver side,
the lwp isn't deleted in this case.

> +	      throw_exception (ex);
> +	    }
> +	  /* Ignore the error since the thread is gone already.  */
> +	  else
> +	    {
> +	      inferior_ptid = null_ptid;
> +	      detach_inferior (pid);
> +	      inf_child_maybe_unpush_target (ops);

Please factor these three lines out of inf_ptrace_attach into
a separate function, exported from inf-ptrace.h and called
from both places.  E.g.:

  extern void inf_ptrace_attach_success (struct target_ops *ops);

> +	    }
> +	}
> +    }
> +  delete_lwp (main_lwp->ptid);
>  }
>  

> +clean_restart ${testfile}
> +
> +if ![runto_main] {
> +    fail "Can't run to main to check for trace support"

"to check for trace support" is stale.

> +    return -1
> +}
> +
> +gdb_breakpoint "_exit"
> +gdb_continue_to_breakpoint "_exit" ".*_exit.*"
> +
> +# Since the gdbserver-native board ends remote debugging on detach,
> +# we can't use gdb_test as this function intercepts the Ending remote
> +# debugging message and tries to restart gdb.

Eh, that looks like a bogus think for gdb_test_multiple to be doing.

Seems like it's been there since "forever":

commit 19fa4a0af34ff07f260fc75e903ab92d8ca03d33
Author:     Mike Werner <mtw@cygnus>
AuthorDate: Sun Feb 21 20:03:55 1993 +0000

        * gdb/testsuite: Initial creation of gdb/testsuite.
        Migrated dejagnu testcases and support files for testing nm to
        gdb/testsuite from deja-gnu.  These files were moved "as is"
        with no modifications.  This migration is part of a major overhaul
        of dejagnu.  The modifications to these testcases, etc., which
        will allow them to work with the new version of dejagnu will be
        made in a future update.

I'd support just removing that bit.

In any case, there's no need to avoid gdb_test/gdb_test_multiple.
Any user-specified pattern overrides gdb_test_multiple's internal
patterns.  So you just need to intercept the  "Ending remote debugging"
message yourself.

> However, it can't do that
> +# as gdb_exit on this board will call monitor exit which will fail as
> +# we're not remote debugging anymore after a detach in this case.
> +send_gdb "detach\n"
> +gdb_expect {

set test "detach"
gdb_test_multiple $test $test {
    -re "Detaching from .*, process $decimal\r\nEnding remote debugging\.\r\n$gdb_prompt $" {
        # This is what you get with "target remote".
        pass $test
    }
    -re "Detaching from .*, process $decimal\r\n$gdb_prompt $" {
        pass $test
    }
}

> +    -re "Detaching from .*, process $decimal\r\nEnding remote debugging\." {

Missing $gdb_prompt match.

> +	#Don't try to exit gdbserver see above comment.
> +	set gdbserver_reconnect_p 1

I don't understand this one.  It shouldn't be needed.

> +	pass "detach non-extended"

Please use the same test message for all branches:

  pass $test

> +    }
> +    -re "Detaching from .*, process $decimal\r\n$gdb_prompt $" {
> +	pass "detach"
> +    }
> +    -re ".*$gdb_prompt $" {
> +        fail "detach"
> +    }
> +    timeout {
> +	fail "detach timeout"

FYI, this would have been the standard:

> +	fail "$test (timeout)"

> +    }
> +}
> 

Thanks,
Pedro Alves

Pedro Alves writes:

> On 06/03/2016 03:49 PM, Antoine Tremblay wrote:
>
>> @@ -1491,7 +1523,28 @@ linux_nat_detach (struct target_ops *ops, const char *args, int from_tty)
>>        linux_fork_detach (args, from_tty);
>>      }
>>    else
>> -    linux_ops->to_detach (ops, args, from_tty);
>> +    {
>> +      TRY
>> +	{
>> +	  linux_ops->to_detach (ops, args, from_tty);
>> +	}
>> +      CATCH (ex, RETURN_MASK_ERROR)
>> +	{
>> +	  if (!check_ptrace_stopped_lwp_gone (main_lwp))
>> +	    {
>> +	      delete_lwp (main_lwp->ptid);
>
> Is deleting the right thing to do?  I think we'll likely
> crash, the next time we try to do something (e.g., try
> to detach again).  Seems like on the gdbserver side,
> the lwp isn't deleted in this case.
>

Indeed, OK.

>> +	      throw_exception (ex);
>> +	    }
>> +	  /* Ignore the error since the thread is gone already.  */
>> +	  else
>> +	    {
>> +	      inferior_ptid = null_ptid;
>> +	      detach_inferior (pid);
>> +	      inf_child_maybe_unpush_target (ops);
>
> Please factor these three lines out of inf_ptrace_attach into
> a separate function, exported from inf-ptrace.h and called
> from both places.  E.g.:
>

OK.

>   extern void inf_ptrace_attach_success (struct target_ops *ops);
>

OK. (detach_success)

>> +	    }
>> +	}
>> +    }
>> +  delete_lwp (main_lwp->ptid);
>>  }
>>  
>
>> +clean_restart ${testfile}
>> +
>> +if ![runto_main] {
>> +    fail "Can't run to main to check for trace support"
>
> "to check for trace support" is stale.
>

Oops thanks.

>> +    return -1
>> +}
>> +
>> +gdb_breakpoint "_exit"
>> +gdb_continue_to_breakpoint "_exit" ".*_exit.*"
>> +
>> +# Since the gdbserver-native board ends remote debugging on detach,
>> +# we can't use gdb_test as this function intercepts the Ending remote
>> +# debugging message and tries to restart gdb.
>
> Eh, that looks like a bogus think for gdb_test_multiple to be doing.
>
> Seems like it's been there since "forever":
>
> commit 19fa4a0af34ff07f260fc75e903ab92d8ca03d33
> Author:     Mike Werner <mtw@cygnus>
> AuthorDate: Sun Feb 21 20:03:55 1993 +0000
>
>         * gdb/testsuite: Initial creation of gdb/testsuite.
>         Migrated dejagnu testcases and support files for testing nm to
>         gdb/testsuite from deja-gnu.  These files were moved "as is"
>         with no modifications.  This migration is part of a major overhaul
>         of dejagnu.  The modifications to these testcases, etc., which
>         will allow them to work with the new version of dejagnu will be
>         made in a future update.
>
> I'd support just removing that bit.
>
> In any case, there's no need to avoid gdb_test/gdb_test_multiple.
> Any user-specified pattern overrides gdb_test_multiple's internal
> patterns.  So you just need to intercept the  "Ending remote debugging"
> message yourself.
>

Right, I tested that before without sucess but had not the perfect regex
yet...

I'll use this now. thanks.

>> However, it can't do that
>> +# as gdb_exit on this board will call monitor exit which will fail as
>> +# we're not remote debugging anymore after a detach in this case.
>> +send_gdb "detach\n"
>> +gdb_expect {
>
> set test "detach"
> gdb_test_multiple $test $test {
>     -re "Detaching from .*, process $decimal\r\nEnding remote debugging\.\r\n$gdb_prompt $" {
>         # This is what you get with "target remote".
>         pass $test
>     }
>     -re "Detaching from .*, process $decimal\r\n$gdb_prompt $" {
>         pass $test
>     }
> }
>
>> +    -re "Detaching from .*, process $decimal\r\nEnding remote debugging\." {
>
> Missing $gdb_prompt match.
>

OK.

>> +	#Don't try to exit gdbserver see above comment.
>> +	set gdbserver_reconnect_p 1
>
> I don't understand this one.  It shouldn't be needed.
>

This one is needed since on the teardown of the .exp test gdb_exit is
called and calls monitor_exit which fails with a timeout this avoids
this and the original gdb_exit is called.

>> +	pass "detach non-extended"
>
> Please use the same test message for all branches:
>
>   pass $test
>
>> +    }
>> +    -re "Detaching from .*, process $decimal\r\n$gdb_prompt $" {
>> +	pass "detach"
>> +    }
>> +    -re ".*$gdb_prompt $" {
>> +        fail "detach"
>> +    }
>> +    timeout {
>> +	fail "detach timeout"
>
> FYI, this would have been the standard:
>
>> +	fail "$test (timeout)"
>
>> +    }
>> +}
>>
Noted, thx.

On 06/03/2016 04:32 PM, Antoine Tremblay wrote:
>>> >> +	#Don't try to exit gdbserver see above comment.
>>> >> +	set gdbserver_reconnect_p 1
>> >
>> > I don't understand this one.  It shouldn't be needed.
>> >
> This one is needed since on the teardown of the .exp test gdb_exit is
> called and calls monitor_exit which fails with a timeout this avoids
> this and the original gdb_exit is called.

Then that strikes me as the wrong thing to do.  We should instead
expect that gdbserver exits cleanly, and then clear server_spawn_id.

E.g.:

proc test_server_exit {} {
  global server_spawn_id
  if ![info exists server_spawn_id] {
    return
  }
  
  set test "server exits"
  gdb_expect {
     -i $server_spawn_id
     eof {
        pass $test
	wait -i $server_spawn_id
        unset server_spawn_id
     }
     timeout {
       fail "$test (timeout)"
     }
   }
  }
}

and then:

    -re "Detaching from .*, process $decimal\r\nEnding remote debugging\.\r\n$gdb_prompt $" {
        # This is what you get with "target remote".
        pass $test
        # If testing with gdbserver, make sure it manages to exit.
        test_server_exit
    }

Thanks,
Pedro Alves

Pedro Alves writes:

> On 06/03/2016 04:32 PM, Antoine Tremblay wrote:
>>>> >> +	#Don't try to exit gdbserver see above comment.
>>>> >> +	set gdbserver_reconnect_p 1
>>> >
>>> > I don't understand this one.  It shouldn't be needed.
>>> >
>> This one is needed since on the teardown of the .exp test gdb_exit is
>> called and calls monitor_exit which fails with a timeout this avoids
>> this and the original gdb_exit is called.
>
> Then that strikes me as the wrong thing to do.  We should instead
> expect that gdbserver exits cleanly, and then clear server_spawn_id.
>
> E.g.:
>
> proc test_server_exit {} {
>   global server_spawn_id
>   if ![info exists server_spawn_id] {
>     return
>   }
>   
>   set test "server exits"
>   gdb_expect {
>      -i $server_spawn_id
>      eof {
>         pass $test
> 	wait -i $server_spawn_id
>         unset server_spawn_id
>      }
>      timeout {
>        fail "$test (timeout)"
>      }
>    }
>   }
> }
>
> and then:
>
>     -re "Detaching from .*, process $decimal\r\nEnding remote debugging\.\r\n$gdb_prompt $" {
>         # This is what you get with "target remote".
>         pass $test
>         # If testing with gdbserver, make sure it manages to exit.
>         test_server_exit
>     }
>

Thanks does seems much better !

However, testing it it triggers the test_server_exit timeout case, I'm
guessing because eof already happened and we're waiting for nothing ?

I'm tempted just to unset the var... but ideas are welcome..

Regards,
Antoine

On 06/03/2016 04:58 PM, Antoine Tremblay wrote:

> However, testing it it triggers the test_server_exit timeout case, I'm
> guessing because eof already happened and we're waiting for nothing ?
> 
> I'm tempted just to unset the var... but ideas are welcome..

Would be easier to try if I had access to an updated patch.. :-)

Thanks,
Pedro Alves