[patch] fixing leaking struct value* in py-value.c

[Aman Gupta]

Bug: http://sourceware.org/bugzilla/show_bug.cgi?id=10896

Given:

$ cat simple.c
#include <stdlib.h>
#include <stdio.h>
#include <unistd.h>

struct blah {
char test[2];
int i;
float none;
} *list_of_blahs;

int main(){
list_of_blahs = malloc(sizeof(struct blah) * 2000000);
printf("pid: %d\n", getpid());
while(1){
sleep(1);
}
}


$ cat simple.py
gdb.execute('p list_of_blahs')
list = gdb.history(0)
end = list + 1000000

while list < end:
list['i']
list += 1


$ ./simple &
[1] 16572
$ gdb -ex 'py execfile("./simple.py")' ./simple 16572

The gdb process will leak a large amount of memory (>512mb RSS) while
executing the python loop. Some (but not all) of this memory is
returned after the loop completes and free_all_values() is called for
the next command. The following patch against archer-tromey-python
keeps the memory usage constant during the loop.

diff --git a/gdb/python/py-value.c b/gdb/python/py-value.c
index 14efd79..f305b01 100644
--- a/gdb/python/py-value.c
+++ b/gdb/python/py-value.c
@@ -143,7 +143,7 @@ valpy_new (PyTypeObject *subtype, PyObject *args,
PyObject *keywords)
}

value_obj->value = value;
- value_incref (value);
+ release_value_or_incref (value);
value_obj->address = NULL;
value_obj->type = NULL;
note_value (value_obj);
@@ -267,6 +267,7 @@ valpy_cast (PyObject *self, PyObject *args)
{
PyObject *type_obj;
struct type *type;
+ PyObject *res_obj = NULL;
struct value *res_val = NULL; /* Initialize to appease gcc warning. */
volatile struct gdb_exception except;

@@ -305,8 +306,10 @@ valpy_getitem (PyObject *self, PyObject *key)
{
value_object *self_value = (value_object *) self;
char *field = NULL;
+ PyObject *res_obj = NULL;
struct value *res_val = NULL;
volatile struct gdb_exception except;
+ struct value *mark = value_mark ();

if (gdbpy_is_string (key))
{
@@ -344,9 +347,13 @@ valpy_getitem (PyObject *self, PyObject *key)
}

xfree (field);
- GDB_PY_HANDLE_EXCEPTION (except);
+ GDB_PY_HANDLE_EXCEPTION_AND_CLEANUP (except, mark);
+
+ if (res_val)
+ res_obj = value_to_value_object (res_val);

- return res_val ? value_to_value_object (res_val) : NULL;
+ value_free_to_mark (mark);
+ return res_obj;
}

static int
@@ -427,8 +434,10 @@ enum valpy_opcode
static PyObject *
valpy_binop (enum valpy_opcode opcode, PyObject *self, PyObject *other)
{
+ PyObject *res_obj = NULL;
struct value *res_val = NULL; /* Initialize to appease gcc warning. */
volatile struct gdb_exception except;
+ struct value *mark = value_mark ();

TRY_CATCH (except, RETURN_MASK_ALL)
{
@@ -519,9 +528,13 @@ valpy_binop (enum valpy_opcode opcode, PyObject
*self, PyObject *other)
break;
}
}
- GDB_PY_HANDLE_EXCEPTION (except);
+ GDB_PY_HANDLE_EXCEPTION_AND_CLEANUP (except, mark);
+
+ if (res_val)
+ res_obj = value_to_value_object (res_val);

- return res_val ? value_to_value_object (res_val) : NULL;
+ value_free_to_mark (mark);
+ return res_obj;
}

static PyObject *
@@ -684,6 +697,7 @@ valpy_richcompare (PyObject *self, PyObject *other, int op)
int result = 0;
struct value *value_other;
volatile struct gdb_exception except;
+ struct value *mark = value_mark ();

if (other == Py_None)
/* Comparing with None is special. From what I can tell, in Python
@@ -742,6 +756,7 @@ valpy_richcompare (PyObject *self, PyObject *other, int op)
break;
}
}
+ value_free_to_mark (mark);
GDB_PY_HANDLE_EXCEPTION (except);

/* In this case, the Python exception has already been set. */
@@ -851,7 +866,7 @@ value_to_value_object (struct value *val)
if (val_obj != NULL)
{
val_obj->value = val;
- value_incref (val);
+ release_value_or_incref (val);
val_obj->address = NULL;
val_obj->type = NULL;
note_value (val_obj);
diff --git a/gdb/python/python-internal.h b/gdb/python/python-internal.h
index 47662d9..309ee5b 100644
--- a/gdb/python/python-internal.h
+++ b/gdb/python/python-internal.h
@@ -149,11 +149,17 @@ PyObject *gdbpy_parameter_value (enum var_types, void *);
/* Use this after a TRY_EXCEPT to throw the appropriate Python
exception. */
#define GDB_PY_HANDLE_EXCEPTION(Exception) \
+ GDB_PY_HANDLE_EXCEPTION_AND_CLEANUP(Exception, NULL)
+
+#define GDB_PY_HANDLE_EXCEPTION_AND_CLEANUP(Exception, mark) \
do { \
- if (Exception.reason < 0) \
+ if (Exception.reason < 0) { \
+ if (mark) \
+ value_free_to_mark (mark); \
return PyErr_Format (Exception.reason == RETURN_QUIT \
? PyExc_KeyboardInterrupt : PyExc_RuntimeError, \
"%s", Exception.message); \
+ } \
} while (0)

/* Use this after a TRY_EXCEPT to throw the appropriate Python
diff --git a/gdb/value.c b/gdb/value.c
index 2f31185..9187a7f 100644
--- a/gdb/value.c
+++ b/gdb/value.c
@@ -691,7 +691,7 @@ free_all_values (void)
/* Remove VAL from the chain all_values
so it will not be freed automatically. */

-void
+int
release_value (struct value *val)
{
struct value *v;
@@ -699,7 +699,7 @@ release_value (struct value *val)
if (all_values == val)
{
all_values = val->next;
- return;
+ return 1;
}

for (v = all_values; v; v = v->next)
@@ -707,9 +707,21 @@ release_value (struct value *val)
if (v->next == val)
{
v->next = val->next;
- break;
+ return 1;
}
}
+
+ return 0;
+}
+
+/* Release VAL or increment its reference count if
+ it was released already */
+
+void
+release_value_or_incref (struct value *val)
+{
+ if (release_value (val) == 0)
+ value_incref (val);
}

/* Release all values up to mark */
diff --git a/gdb/value.h b/gdb/value.h
index 0da7031..5705210 100644
--- a/gdb/value.h
+++ b/gdb/value.h
@@ -595,7 +595,9 @@ extern void value_free (struct value *val);

extern void free_all_values (void);

-extern void release_value (struct value *val);
+extern int release_value (struct value *val);
+
+extern void release_value_or_incref (struct value *val);

extern int record_latest_value (struct value *val);

[Phil Muldoon]

Post by Aman Gupta
The gdb process will leak a large amount of memory (>512mb RSS) while
executing the python loop. Some (but not all) of this memory is
returned after the loop completes and free_all_values() is called for
the next command. The following patch against archer-tromey-python
keeps the memory usage constant during the loop.

On my tests, the consumption noted is far in excess of that, so good
find.

Post by Aman Gupta
diff --git a/gdb/python/py-value.c b/gdb/python/py-value.c
index 14efd79..f305b01 100644
--- a/gdb/python/py-value.c
+++ b/gdb/python/py-value.c
@@ -143,7 +143,7 @@ valpy_new (PyTypeObject *subtype, PyObject *args,
PyObject *keywords)

Your mailer appears to be wrapping patches. So this means the patch
command will complain of a malformed patch. In this case it was
easily fixed in an editor, but a heads up in any case.

Post by Aman Gupta
#define GDB_PY_HANDLE_EXCEPTION(Exception) \
+ GDB_PY_HANDLE_EXCEPTION_AND_CLEANUP(Exception, NULL)
+
+#define GDB_PY_HANDLE_EXCEPTION_AND_CLEANUP(Exception, mark) \
do { \
- if (Exception.reason < 0) \
+ if (Exception.reason < 0) { \
+ if (mark) \
+ value_free_to_mark (mark); \
return PyErr_Format (Exception.reason == RETURN_QUIT \
? PyExc_KeyboardInterrupt : PyExc_RuntimeError, \
"%s", Exception.message); \
+ } \
} while (0)

When I try to compile against archer-tromey-python with:

../archer/configure && make CFLAGS="-O0 -g3"

I get a compile error:

cc1: warnings being treated as errors
../../archer/gdb/python/py-symbol.c: In function ‘gdbpy_lookup_symbol’:
../../archer/gdb/python/py-symbol.c:200: error: implicit declaration of function ‘value_free_to_mark’
make[2]: *** [py-symbol.o] Error 1

This is because GDB_PY_HANDLE_EXCEPTION is used in many sources files,
and value_free_to_mark is defined in value.{h|c}. And as you've
taken over the definition of GDB_PY_HANDLE_EXCEPTION the substitution
has some side-effects. So this needs to be fixed first.

Did you spot or note any regressions from a full test run by comparing
the gdb.sum files before and after the patch? A note indicating the
platform/architecture of testing, and any regressions spotted would be
great, thanks.

Also, this patch needs a ChangeLog.

Cheers,

Phil

[Aman Gupta]

Thanks for the feedback. Here is an updated patch against the latest
archer-tromey-python with the fixes you mentioned.

I ran `make check` before and after the patch on a "Linux 2.6.32.3 #3
SMP Thu Feb 4 23:11:14 UTC 2010 x86_64 GNU/Linux" and there were no
changes to the gdb.sum file (other than 3 lines with newer
timestamps).

diff --git a/gdb/ChangeLog b/gdb/ChangeLog
index 7968e93..62030a5 100644
--- a/gdb/ChangeLog
+++ b/gdb/ChangeLog
@@ -1,3 +1,13 @@
[phone] Aman Gupta <aman-vFcy+***@public.gmane.org>
+ * gdb/value.c (release_value): Return 0 on success.
+ (release_value_or_incref): New function, incref value if already
+ released.
+ * gdb/python/py-value.c (valpy_new, value_to_value_object): Use
+ release_value_or_incref() instead of value_incref() to prevent leaks.
+ (valpy_richcompare, valpy_binop, valpy_getitem): Call value_mark()
+ before calling into gdb, and make sure to value_free_to_mark() before
+ returning or raising any exceptions.
+
2010-10-04 Doug Evans <dje-hpIqsD4AKlfQT0dZR+***@public.gmane.org>

* cc-with-index.sh: New file.
diff --git a/gdb/python/py-value.c b/gdb/python/py-value.c
index 0aeea7c..4cf3c5b 100644
--- a/gdb/python/py-value.c
+++ b/gdb/python/py-value.c
@@ -149,7 +149,7 @@ valpy_new (PyTypeObject *subtype, PyObject *args,
PyObject *keywords)
}

value_obj->value = value;
- value_incref (value);
+ release_value_or_incref (value);
value_obj->address = NULL;
value_obj->type = NULL;
value_obj->dynamic_type = NULL;
@@ -444,8 +444,10 @@ valpy_getitem (PyObject *self, PyObject *key)
{
value_object *self_value = (value_object *) self;
char *field = NULL;
+ PyObject *res_obj = NULL;
struct value *res_val = NULL;
volatile struct gdb_exception except;
+ struct value *mark = value_mark ();

if (gdbpy_is_string (key))
{
@@ -485,9 +487,15 @@ valpy_getitem (PyObject *self, PyObject *key)
}

xfree (field);
+ if (except.reason < 0)
+ value_free_to_mark (mark);
GDB_PY_HANDLE_EXCEPTION (except);

- return res_val ? value_to_value_object (res_val) : NULL;
+ if (res_val)
+ res_obj = value_to_value_object (res_val);
+
+ value_free_to_mark (mark);
+ return res_obj;
}

static int
@@ -623,8 +631,10 @@ enum valpy_opcode
static PyObject *
valpy_binop (enum valpy_opcode opcode, PyObject *self, PyObject *other)
{
+ PyObject *res_obj = NULL;
struct value *res_val = NULL; /* Initialize to appease gcc warning. */
volatile struct gdb_exception except;
+ struct value *mark = value_mark ();

TRY_CATCH (except, RETURN_MASK_ALL)
{
@@ -715,9 +725,16 @@ valpy_binop (enum valpy_opcode opcode, PyObject
*self, PyObject *other)
break;
}
}
+
+ if (except.reason < 0)
+ value_free_to_mark (mark);
GDB_PY_HANDLE_EXCEPTION (except);

- return res_val ? value_to_value_object (res_val) : NULL;
+ if (res_val)
+ res_obj = value_to_value_object (res_val);
+
+ value_free_to_mark (mark);
+ return res_obj;
}

static PyObject *
@@ -878,6 +895,7 @@ valpy_richcompare (PyObject *self, PyObject *other, int op)
int result = 0;
struct value *value_other;
volatile struct gdb_exception except;
+ struct value *mark = value_mark ();

if (other == Py_None)
/* Comparing with None is special. From what I can tell, in Python
@@ -936,6 +954,7 @@ valpy_richcompare (PyObject *self, PyObject *other, int op)
break;
}
}
+ value_free_to_mark (mark);
GDB_PY_HANDLE_EXCEPTION (except);

/* In this case, the Python exception has already been set. */
@@ -1059,7 +1078,7 @@ value_to_value_object (struct value *val)
if (val_obj != NULL)
{
val_obj->value = val;
- value_incref (val);
+ release_value_or_incref (val);
val_obj->address = NULL;
val_obj->type = NULL;
val_obj->dynamic_type = NULL;
diff --git a/gdb/value.c b/gdb/value.c
index 55fcd38..0d98ba8 100644
--- a/gdb/value.c
+++ b/gdb/value.c
@@ -737,9 +737,10 @@ free_value_chain (struct value *v)
}

/* Remove VAL from the chain all_values
- so it will not be freed automatically. */
+ so it will not be freed automatically.
+ Result of 0 indicates the value was found and released. */

-void
+int
release_value (struct value *val)
{
struct value *v;
@@ -748,7 +749,7 @@ release_value (struct value *val)
{
all_values = val->next;
val->next = NULL;
- return;
+ return 0;
}

for (v = all_values; v; v = v->next)
@@ -757,12 +758,25 @@ release_value (struct value *val)
{
v->next = val->next;
val->next = NULL;
- break;
+ return 0;
}
}
+
+ return -1;
+}
+
+/* Release VAL or increment its reference count if
+ it was released already */
+
+void
+release_value_or_incref (struct value *val)
+{
+ if (release_value (val) != 0)
+ value_incref (val);
}

/* Release all values up to mark */
+
struct value *
value_release_to_mark (struct value *mark)
{
diff --git a/gdb/value.h b/gdb/value.h
index 12cbc0a..fc5a8b3 100644
--- a/gdb/value.h
+++ b/gdb/value.h
@@ -641,7 +641,9 @@ extern void free_all_values (void);

extern void free_value_chain (struct value *v);

-extern void release_value (struct value *val);
+extern int release_value (struct value *val);
+
+extern void release_value_or_incref (struct value *val);

extern int record_latest_value (struct value *val);

Post by Phil Muldoon

Post by Aman Gupta
The gdb process will leak a large amount of memory (>512mb RSS) while
executing the python loop. Some (but not all) of this memory is
returned after the loop completes and free_all_values() is called for
the next command. The following patch against archer-tromey-python
keeps the memory usage constant during the loop.

On my tests, the consumption noted is far in excess of that, so good
find.

Post by Aman Gupta
diff --git a/gdb/python/py-value.c b/gdb/python/py-value.c
index 14efd79..f305b01 100644
--- a/gdb/python/py-value.c
+++ b/gdb/python/py-value.c
@@ -143,7 +143,7 @@ valpy_new (PyTypeObject *subtype, PyObject *args,
PyObject *keywords)

Your mailer appears to be wrapping patches.  So this means the patch
command will complain of a malformed patch.  In this case it was
easily fixed in an editor, but a heads up in any case.

Post by Aman Gupta
 #define GDB_PY_HANDLE_EXCEPTION(Exception)                           \
+    GDB_PY_HANDLE_EXCEPTION_AND_CLEANUP(Exception, NULL)
+
+#define GDB_PY_HANDLE_EXCEPTION_AND_CLEANUP(Exception, mark)                         \
     do {                                                             \
-      if (Exception.reason < 0)                                              \
+      if (Exception.reason < 0) {                                    \
+        if (mark)                                            \
+          value_free_to_mark (mark);                                         \
      return PyErr_Format (Exception.reason == RETURN_QUIT            \
                           ? PyExc_KeyboardInterrupt : PyExc_RuntimeError, \
                           "%s", Exception.message);                  \
+      }                                              \
     } while (0)

../archer/configure && make CFLAGS="-O0 -g3"
cc1: warnings being treated as errors
../../archer/gdb/python/py-symbol.c:200: error: implicit declaration of function ‘value_free_to_mark’
make[2]: *** [py-symbol.o] Error 1
This is because GDB_PY_HANDLE_EXCEPTION is used in many sources files,
and value_free_to_mark is defined in value.{h|c}. And as you've
taken over the definition of GDB_PY_HANDLE_EXCEPTION the substitution
has some side-effects. So this needs to be fixed first.
Did you spot or note any regressions from a full test run by comparing
the gdb.sum files before and after the patch?  A note indicating the
platform/architecture of testing, and any regressions spotted would be
great, thanks.
Also, this patch needs a ChangeLog.
Cheers,
Phil