Re: [PATCH] Fix for gdb/PR 14808, vfork/exec inferior problem

[Don Breazeal <donb@codesourcery.com>]

On 5/15/2014 12:06 AM, Yao Qi wrote:
>  05/13/2014 07:34 AM, donb@codesourcery.com wrote:
>> @@ -649,6 +649,7 @@ handle_vfork_child_exec_or_exit (int exec)
>>  	  struct cleanup *old_chain;
>>  	  struct program_space *pspace;
>>  	  struct address_space *aspace;
>> +	  struct inferior *parent_inf;
>>  
> 
> Local parent_inf is only used in the "if (exec)" block below, so better
> to declare it there.
OK, done.

--snip--

>> +		 where the parent called vfork.  Now that the child has
>> +		 called exec and we are detaching from the parent, the
>> +		 parent inferior needs to have its own pspace and aspace
> 
> parent inferior has its own pspace, but may not have its own aspace,
> depending on gdbarch_has_shared_address_space.
Fixed.

> 
>> +		 so that changes in the child don't affect it.  We have
>> +		 to give the new spaces to the parent since we saved the
>> +		 child's spaces as the current spaces above.  Even though
>> +		 we are detaching the parent, we want to keep the
>> +		 corresponding entry in the inferiors list intact.  */
>> +	      parent_inf = current_inferior ();
>> +	      parent_inf->aspace = new_address_space ();
> 
> Rather than creating a new address space, use maybe_new_address_space, like
>> +	      parent_inf->pspace = add_program_space (parent_inf->aspace);
> 
> 	      parent_inf->pspace
> 		= add_program_space (maybe_new_address_space ());
> 	      parent_inf->aspace = parent_inf->pspace->aspace;
Done.

> 
>> +	      parent_inf->removable = inf->removable;
> 
> Field removable of parent inferior should be unchanged, IMO.
It turns out this assignment was a NOP.  Removed.

> 
>> +	      set_current_program_space (parent_inf->pspace);
>> +	      clone_program_space (parent_inf->pspace, pspace);
> 
> Do we need to unlink parent and child?  I am not very sure.
> 
> 	      /* Break the bonds.  */
> 	      inf->vfork_parent->vfork_child = NULL;
> 
We don't need to do it here, since this happens in target_detach.  Since parent_inf is the parent, it's fork_child pointer is set.  The unlinking happens in inferior.c:exit_inferior_1, which is ultimately called via target_detach.  The child's vfork_parent pointer is cleared by existing code near the end of handle_vfork_child_exec_or_exit.

Here is the updated patch.
--Don

gdb/
2014-05-15  Don Breazeal  <donb@codesourcery.com>

	* infrun.c (handle_vfork_child_exec_or_exit): For the case
	of a vfork where we follow the child and detach the parent,
	and the child execs, create a new pspace for the parent
	inferior.

gdb/testsuite
2014-05-15  Don Breazeal  <donb@codesourcery.com>

	* gdb.base/foll-vfork.exp (vfork_relations_in_info_inferiors):
	Test that after a vfork and child exec, the parent's exec file
	name has not been changed.

---
 gdb/infrun.c                          |   44 ++++++++++++++++++++++----------
 gdb/testsuite/gdb.base/foll-vfork.exp |   11 ++++++++
 2 files changed, 41 insertions(+), 14 deletions(-)

diff --git a/gdb/infrun.c b/gdb/infrun.c
index ab39b6e..20ab003 100644
--- a/gdb/infrun.c
+++ b/gdb/infrun.c
@@ -665,27 +665,42 @@ handle_vfork_child_exec_or_exit (int exec)
 	  else
 	    old_chain = save_current_space_and_thread ();
 
-	  /* We're letting loose of the parent.  */
+	  /* Make the parent the current inferior for target_detach.  */
 	  tp = any_live_thread_of_process (inf->vfork_parent->pid);
 	  switch_to_thread (tp->ptid);
 
-	  /* We're about to detach from the parent, which implicitly
-	     removes breakpoints from its address space.  There's a
-	     catch here: we want to reuse the spaces for the child,
-	     but, parent/child are still sharing the pspace at this
-	     point, although the exec in reality makes the kernel give
-	     the child a fresh set of new pages.  The problem here is
-	     that the breakpoints module being unaware of this, would
-	     likely chose the child process to write to the parent
-	     address space.  Swapping the child temporarily away from
-	     the spaces has the desired effect.  Yes, this is "sort
-	     of" a hack.  */
-
+	  /* The child inferior may be dead, so avoid giving the
+	     breakpoints module the option to write through to it
+	     by swapping the child temporarily away from the spaces
+	     (cloning a program space resets breakpoints).  */
 	  pspace = inf->pspace;
 	  aspace = inf->aspace;
 	  inf->aspace = NULL;
 	  inf->pspace = NULL;
 
+	  if (exec)
+	    {
+	      struct inferior *parent_inf;
+
+	      /* The parent and child inferiors have been sharing
+		 program and address space structures from the point
+		 where the parent called vfork.  Now that the child has
+		 called exec and we are detaching from the parent, the
+		 parent inferior needs to have its own pspace so that
+		 changes in the child don't affect it.  We have to give
+		 the new pspace to the parent (instead of the child)
+		 since we saved the child's spaces as the current spaces
+		 above.  Even though we are detaching the parent, we
+		 want to keep the corresponding entry in the inferiors
+		 list intact.  */
+	      parent_inf = current_inferior ();
+	      parent_inf->pspace
+	       = add_program_space (maybe_new_address_space ());
+	      parent_inf->aspace = parent_inf->pspace->aspace;
+	      set_current_program_space (parent_inf->pspace);
+	      clone_program_space (parent_inf->pspace, pspace);
+	    }
+
 	  if (debug_infrun || info_verbose)
 	    {
 	      target_terminal_ours ();
@@ -702,9 +717,10 @@ handle_vfork_child_exec_or_exit (int exec)
 				  inf->vfork_parent->pid);
 	    }
 
+	  /* Detach the parent.  */
 	  target_detach (NULL, 0);
 
-	  /* Put it back.  */
+	  /* Put the child spaces back.  */
 	  inf->pspace = pspace;
 	  inf->aspace = aspace;
 
diff --git a/gdb/testsuite/gdb.base/foll-vfork.exp b/gdb/testsuite/gdb.base/foll-vfork.exp
index fe3663c..e9b0110 100644
--- a/gdb/testsuite/gdb.base/foll-vfork.exp
+++ b/gdb/testsuite/gdb.base/foll-vfork.exp
@@ -442,6 +442,17 @@ proc vfork_relations_in_info_inferiors { variant } {
 	   pass $test
        }
    }
+
+   # Make sure the exec file name of the vfork parent is not
+   # changed when the child's is changed.
+   if { $variant == "exec" } {
+       set test "exec file name change"
+       gdb_test_multiple "info inferiors" $test {
+	   -re " 2 .*vforked-prog.*  1 .*foll-vfork.*$gdb_prompt " {
+	       pass $test
+	   }
+       }
+   }
 }}
 
 proc do_vfork_and_follow_parent_tests {} {
-- 
1.7.0.4