gdb/python: fix memory leak in python inferior code

Message ID 20210905085309.2153900-1-andrew.burgess@embecosm.com
State New
Headers show
Series
  • gdb/python: fix memory leak in python inferior code
Related show

Commit Message

Andrew Burgess Sept. 5, 2021, 8:53 a.m.
When a user creates a gdb.Inferior object for the first time a new
Python object is created.  This object is then cached within GDB's
inferior object using the registry mechanism (see
inferior_to_inferior_object in py-inferior.c, specifically the calls
to inferior_data and set_inferior_data).

This reference to the Python object, held within the real inferior
object, counts as a reference to the Python object, and so holds the
reference count of the Python object above zero.

At the same time, the Python object maintains a pointer back to GDB's
real inferior object.

When GDB's inferior object is deleted (say the inferior exits) then
py_free_inferior is called (thanks to the registry system), this
function looks up the Python object and sets the inferior pointer to
nullptr and finally deletes reduces the reference count on the Python
object.

If at this point the user still holds a reference to the Python object
then nothing happens.  The gdb.Inferior Python object is now in the
non-valid state (see infpy_is_valid in py-inferior.c), but otherwise,
everything is fine.

However, if there are no further references to the gdb.Inferior
object, or, once the user has given up all their references to the
gdb.Inferior object, then infpy_dealloc is called.

This function currently checks to see if the inferior pointer within
the gdb.Inferior object is nullptr or not.  If the pointer is nullptr
then infpy_dealloc immediately returns.

Only when the inferior point in the gdb.Inferior is not nullptr do
we (a) set the gdb.Inferior reference inside GDB's inferior to
nullptr, and (b) call the underlying Python tp_free function.

There are a number things wrong here:

  1. The gdb.Inferior reference within GDB's inferior object hold a
  reference count, thus, setting this reference to nullptr without
  first decrementing the reference count would leak a reference.  This
  could be solved by having the inferior hold a gdbpy_ref object,
  however...

  2. As GDB's inferior holds a reference then infpy_dealloc will never
  be called until GDB's inferior object is deleted, calls
  py_free_inferior, and so gives up the reference.  At this point
  there is no longer a need to call set_inferior_data to set the field
  back to NULL, that field must have been cleared in order to get the
  reference count to zero, which means...

  3. If we know that py_free_inferior must be called before
  infpy_dealloc, then we know that the inferior pointer in
  gdb.Inferior will always be nullptr when infpy_dealloc is called,
  this means that the call to the underlying tp_free function will
  always be skipped.  Skipping this call at all always seems like a
  bad idea, but it turns out we will always skip it.

Given all of the above, I assert that the gdb.Inferior field will
always be nullptr when infpy_dealloc is called.  That's what this
patch does.

I assume that failing to call tp_free will mean that Python leaks some
memory associated with the gdb.Inferior object, but I don't know any
way that I could show that in a test.  As such, I've got no new tests
for this commit, but if anyone has any ideas I'd be happy to add
something.
---
 gdb/python/py-inferior.c | 16 ++++++++++++----
 1 file changed, 12 insertions(+), 4 deletions(-)

-- 
2.25.4

Comments

Tom de Vries via Gdb-patches Sept. 7, 2021, 3:10 p.m. | #1
On 2021-09-05 4:53 a.m., Andrew Burgess wrote:
> When a user creates a gdb.Inferior object for the first time a new

> Python object is created.  This object is then cached within GDB's

> inferior object using the registry mechanism (see

> inferior_to_inferior_object in py-inferior.c, specifically the calls

> to inferior_data and set_inferior_data).

> 

> This reference to the Python object, held within the real inferior

> object, counts as a reference to the Python object, and so holds the

> reference count of the Python object above zero.

> 

> At the same time, the Python object maintains a pointer back to GDB's

> real inferior object.

> 

> When GDB's inferior object is deleted (say the inferior exits) then

> py_free_inferior is called (thanks to the registry system), this

> function looks up the Python object and sets the inferior pointer to

> nullptr and finally deletes reduces the reference count on the Python

> object.

> 

> If at this point the user still holds a reference to the Python object

> then nothing happens.  The gdb.Inferior Python object is now in the

> non-valid state (see infpy_is_valid in py-inferior.c), but otherwise,

> everything is fine.

> 

> However, if there are no further references to the gdb.Inferior

> object, or, once the user has given up all their references to the

> gdb.Inferior object, then infpy_dealloc is called.

> 

> This function currently checks to see if the inferior pointer within

> the gdb.Inferior object is nullptr or not.  If the pointer is nullptr

> then infpy_dealloc immediately returns.

> 

> Only when the inferior point in the gdb.Inferior is not nullptr do

> we (a) set the gdb.Inferior reference inside GDB's inferior to

> nullptr, and (b) call the underlying Python tp_free function.

> 

> There are a number things wrong here:

> 

>   1. The gdb.Inferior reference within GDB's inferior object hold a

>   reference count, thus, setting this reference to nullptr without

>   first decrementing the reference count would leak a reference.  This

>   could be solved by having the inferior hold a gdbpy_ref object,

>   however...

> 

>   2. As GDB's inferior holds a reference then infpy_dealloc will never

>   be called until GDB's inferior object is deleted, calls

>   py_free_inferior, and so gives up the reference.  At this point

>   there is no longer a need to call set_inferior_data to set the field

>   back to NULL, that field must have been cleared in order to get the

>   reference count to zero, which means...

> 

>   3. If we know that py_free_inferior must be called before

>   infpy_dealloc, then we know that the inferior pointer in

>   gdb.Inferior will always be nullptr when infpy_dealloc is called,

>   this means that the call to the underlying tp_free function will

>   always be skipped.  Skipping this call at all always seems like a

>   bad idea, but it turns out we will always skip it.

> 

> Given all of the above, I assert that the gdb.Inferior field will

> always be nullptr when infpy_dealloc is called.  That's what this

> patch does.

> 

> I assume that failing to call tp_free will mean that Python leaks some

> memory associated with the gdb.Inferior object, but I don't know any

> way that I could show that in a test.  As such, I've got no new tests

> for this commit, but if anyone has any ideas I'd be happy to add

> something.


I tried to read this a few times, and I'm stuck on not understanding how
there isn't a reference cycle here.  If the Python object holds a
reference on the inferior, and the inferior holds a reference on the
Python object (through the registry), how can that work?  Or is it a
weak reference in one of the directions?

Simon
Andrew Burgess Sept. 8, 2021, 9:26 a.m. | #2
* Simon Marchi <simon.marchi@polymtl.ca> [2021-09-07 11:10:20 -0400]:

> 

> 

> On 2021-09-05 4:53 a.m., Andrew Burgess wrote:

> > When a user creates a gdb.Inferior object for the first time a new

> > Python object is created.  This object is then cached within GDB's

> > inferior object using the registry mechanism (see

> > inferior_to_inferior_object in py-inferior.c, specifically the calls

> > to inferior_data and set_inferior_data).

> > 

> > This reference to the Python object, held within the real inferior

> > object, counts as a reference to the Python object, and so holds the

> > reference count of the Python object above zero.

> > 

> > At the same time, the Python object maintains a pointer back to GDB's

> > real inferior object.

> > 

> > When GDB's inferior object is deleted (say the inferior exits) then

> > py_free_inferior is called (thanks to the registry system), this

> > function looks up the Python object and sets the inferior pointer to

> > nullptr and finally deletes reduces the reference count on the Python

> > object.

> > 

> > If at this point the user still holds a reference to the Python object

> > then nothing happens.  The gdb.Inferior Python object is now in the

> > non-valid state (see infpy_is_valid in py-inferior.c), but otherwise,

> > everything is fine.

> > 

> > However, if there are no further references to the gdb.Inferior

> > object, or, once the user has given up all their references to the

> > gdb.Inferior object, then infpy_dealloc is called.

> > 

> > This function currently checks to see if the inferior pointer within

> > the gdb.Inferior object is nullptr or not.  If the pointer is nullptr

> > then infpy_dealloc immediately returns.

> > 

> > Only when the inferior point in the gdb.Inferior is not nullptr do

> > we (a) set the gdb.Inferior reference inside GDB's inferior to

> > nullptr, and (b) call the underlying Python tp_free function.

> > 

> > There are a number things wrong here:

> > 

> >   1. The gdb.Inferior reference within GDB's inferior object hold a

> >   reference count, thus, setting this reference to nullptr without

> >   first decrementing the reference count would leak a reference.  This

> >   could be solved by having the inferior hold a gdbpy_ref object,

> >   however...

> > 

> >   2. As GDB's inferior holds a reference then infpy_dealloc will never

> >   be called until GDB's inferior object is deleted, calls

> >   py_free_inferior, and so gives up the reference.  At this point

> >   there is no longer a need to call set_inferior_data to set the field

> >   back to NULL, that field must have been cleared in order to get the

> >   reference count to zero, which means...

> > 

> >   3. If we know that py_free_inferior must be called before

> >   infpy_dealloc, then we know that the inferior pointer in

> >   gdb.Inferior will always be nullptr when infpy_dealloc is called,

> >   this means that the call to the underlying tp_free function will

> >   always be skipped.  Skipping this call at all always seems like a

> >   bad idea, but it turns out we will always skip it.

> > 

> > Given all of the above, I assert that the gdb.Inferior field will

> > always be nullptr when infpy_dealloc is called.  That's what this

> > patch does.

> > 

> > I assume that failing to call tp_free will mean that Python leaks some

> > memory associated with the gdb.Inferior object, but I don't know any

> > way that I could show that in a test.  As such, I've got no new tests

> > for this commit, but if anyone has any ideas I'd be happy to add

> > something.

> 

> I tried to read this a few times, and I'm stuck on not understanding how

> there isn't a reference cycle here.  If the Python object holds a

> reference on the inferior, and the inferior holds a reference on the

> Python object (through the registry), how can that work?  Or is it a

> weak reference in one of the directions?


There's no "reference" from the Python object to the inferior in the
same way there's a "reference" from the inferior to the Python object.

That is the inferior (via the registry) holds a reference to the
Python object.  This means that the Python object can't be deleted
until the reference in the inferior is released (that reference holds
the refcount above zero).

However, the Python object only holds a pointer back to the inferior,
even though inferiors are a child of refcounted_object, the pointer
within the Python object isn't handled as a refcounted_object, just a
raw pointer.

So, the inferior's refcount can hit zero with no consideration for the
pointer held within the Python object, but that's OK, when the
inferior is deleted we clean up the registry.  This calls into the
Python code, which sets the inferior pointer to nullptr, and
decrements the refcount on the Python object.

When the user gives up their last gdb.Inferior reference the Python
object can then be free'd.

Hope that makes things a little clearer,

Thanks,
Andrew

Patch

diff --git a/gdb/python/py-inferior.c b/gdb/python/py-inferior.c
index cfbc2f6574f..c36f275b23e 100644
--- a/gdb/python/py-inferior.c
+++ b/gdb/python/py-inferior.c
@@ -864,12 +864,20 @@  static void
 infpy_dealloc (PyObject *obj)
 {
   inferior_object *inf_obj = (inferior_object *) obj;
-  struct inferior *inf = inf_obj->inferior;
 
-  if (! inf)
-    return;
+  /* The inferior itself holds a reference to this Python object, which
+     will keep the reference count of this object above zero until GDB
+     deletes the inferior and py_free_inferior is called.
+
+     Once py_free_inferior has been called then the link between this
+     Python object and the inferior is set to nullptr, and then the
+     reference count on this Python object is decremented.
+
+     The result of all this is that the link between this Python object and
+     the inferior should always have been set to nullptr before this
+     function is called.  */
+  gdb_assert (inf_obj->inferior == nullptr);
 
-  set_inferior_data (inf, infpy_inf_data_key, NULL);
   Py_TYPE (obj)->tp_free (obj);
 }