gdb: Don't leak memory with TYPE_ALLOC / TYPE_ZALLOC
Commit Message
* Tom Tromey <tom@tromey.com> [2018-09-11 14:38:44 -0600]:
> >>>>> "Andrew" == Andrew Burgess <andrew.burgess@embecosm.com> writes:
>
> Andrew> I ran into the following issue after runnig valgrind on GDB.
> Andrew> Calling TYPE_ALLOC or TYPE_ZALLOC currently allocates memory from
> Andrew> either the objfile obstack or by using malloc. The problem with this
> Andrew> is that types are allocated either on the objfile obstack, or on the
> Andrew> gdbarch obstack.
>
> Thank you for the patch. I found the analysis very clear, so thanks for
> that too.
>
> Andrew> This commit ensures that auxiliary type data is allocated from the
> Andrew> same obstack as the type itself, which should reduce leaked memory.
>
> I'm curious whether this is something you intend to do. gdb has many
> things that are allocated for the lifetime of the process...
No. This was more an observation that some allocated memory appeared
to be getting lost. From there I went to read the code, and the
comment on TYPE_ALLOC (about either using the objfile obstack OR
malloc) seemed wrong (as types are on objfile or gdbarch obstack).
Given I was looking at the code and the fix seemed obvious....
>
> Andrew> The one problem case that I found with this change was in eval.c,
> Andrew> where in one place we allocate a local type structure, and then used
> Andrew> TYPE_ZALLOC to allocate some space for the type. This local type is
> Andrew> neither object file owned, nor gdbarch owned, and so the updated
> Andrew> TYPE_ALLOC code is unable to find an objstack to allocate space on.
>
> Andrew> My proposed solution for this issue is that the space should be
> Andrew> allocated with a direct call to xzalloc. We could extend TYPE_ALLOC
> Andrew> to check for type->gdbarch being null, and then fall back to a direct
> Andrew> call to xzalloc, however, I think making the rare case of a local type
> Andrew> requiring special handling is not a bad thing, this serves to
> Andrew> highlight that clearing up the memory will require special handling
> Andrew> too.
>
> I went and refreshed my memory of this area.
>
> My first thought was that this would be dangerous because -- surely --
> there was code allocating a type on the heap and manually managing it.
> I was remembering Python, but I think we didn't do this because we
> bailed on the whole "type GC" idea. So, there no problems that I could
> find.
>
> I suppose if the type from fake_method ever ends up somewhere it
> shouldn't -- passed to something that uses TYPE_ALLOC -- gdb will crash.
>
> I wonder if get_type_arch should assert that the result is not NULL.
New patch does this.
>
> Andrew> * gdbtypes.h (TYPE_ALLOC): Allocate space on either the objfile
> Andrew> obstack, or the gdbarch obstack.
> Andrew> (TYPE_ZALLOC): Rewrite using TYPE_ALLOC.
> Andrew> * eval.c (fake_method::fake_method): Call xzalloc directly for a
> Andrew> type that is neither object file owned, nor gdbarch owned.
>
> Andrew> +/* See comment on TYPE_ALLOC. */
> Andrew> +
> Andrew> +#define TYPE_ZALLOC(t,size) (memset (TYPE_ALLOC (t, size), 0, size))
>
> Would you mind seeing what happens if you use obstack_zalloc here
> instead? It seems like that would give us some extra compile-time
> safety. If it fails, that would be interesting.
Unless I'm missing something (very possible) there's no suitable
obstack_zalloc to use here.
There's this function:
template <typename T> static inline T* obstack_zalloc (struct obstack *ob);
However, that requires an actual type, and all I have is a size at
this point. Without much wider changes I don't think I can use that
in TYPE_ZALLOC. Please let me know if I've missed something.
> I think the eval.c change should have a comment explaining why
> TYPE_ZALLOC is not used.
Done.
> This is ok with that nit changed. If obstack_zalloc works the I think
> that variant would be better.
So, once I started picking at this thread I found a few more cases
that could be fixed in gdbtypes.c, these are included in the revised
patch below.
I know that there's more that could be done in this area, but I don't
want this patch to grow too much.
Thanks,
Andrew
---
gdb: Don't leak memory with TYPE_ALLOC / TYPE_ZALLOC
This patch started as an observation from valgrind that GDB appeared
to be loosing track of some memory associated with types. An example
valgrind stack would be:
24 bytes in 1 blocks are possibly lost in loss record 419 of 5,361
at 0x4C2EA1E: calloc (vg_replace_malloc.c:711)
by 0x623D26: xcalloc (common-utils.c:85)
by 0x623D65: xzalloc(unsigned long) (common-utils.c:95)
by 0x72A066: make_function_type(type*, type**) (gdbtypes.c:510)
by 0x72A098: lookup_function_type(type*) (gdbtypes.c:521)
by 0x73635D: gdbtypes_post_init(gdbarch*) (gdbtypes.c:5439)
by 0x727590: gdbarch_data(gdbarch*, gdbarch_data*) (gdbarch.c:5230)
by 0x735B99: builtin_type(gdbarch*) (gdbtypes.c:5313)
by 0x514D95: elf_rel_plt_read(minimal_symbol_reader&, objfile*, bfd_symbol**) (elfread.c:542)
by 0x51662F: elf_read_minimal_symbols(objfile*, int, elfinfo const*) (elfread.c:1121)
by 0x5168A5: elf_symfile_read(objfile*, enum_flags<symfile_add_flag>) (elfread.c:1207)
by 0x8520F5: read_symbols(objfile*, enum_flags<symfile_add_flag>) (symfile.c:794)
When we look in make_function_type we find a call to TYPE_ZALLOC
(inside the INIT_FUNC_SPECIFIC macro). It is this call to TYPE_ZALLOC
that is allocating memory with xcalloc, that is then getting lost.
The problem is tht calling TYPE_ALLOC or TYPE_ZALLOC currently
allocates memory from either the objfile obstack or by using malloc.
The problem with this is that types are allocated either on the
objfile obstack, or on the gdbarch obstack.
As a result, if we discard a type associated with an objfile then
auxiliary data allocated with TYPE_(Z)ALLOC will be correctly
discarded. But, if we were ever to discard a gdbarch then any
auxiliary type data would be leaked. Right now there are very few
places in GDB where a gdbarch is ever discarded, but it shouldn't hurt
to close down these bugs as we spot them.
This commit ensures that auxiliary type data is allocated from the
same obstack as the type itself, which should reduce leaked memory.
The one problem case that I found with this change was in eval.c,
where in one place we allocate a local type structure, and then used
TYPE_ZALLOC to allocate some space for the type. This local type is
neither object file owned, nor gdbarch owned, and so the updated
TYPE_ALLOC code is unable to find an objstack to allocate space on.
My proposed solution for this issue is that the space should be
allocated with a direct call to xzalloc. We could extend TYPE_ALLOC
to check for type->gdbarch being null, and then fall back to a direct
call to xzalloc, however, I think that making this rare case of a
local type require special handling is not a bad thing, this serves to
highlight that clearing up the memory will require special handling
too.
This special case of a local type is interesting as the types owner
field (contained within the main_type) is completely null. While
reflecting on this I looked at how types use the get_type_arch
function. It seems clear that, based on how this is used, it is never
intended that null will be returned from this function. This only
goes to reinforce, how locally alloctaed types, with no owner, are
both special, and need to be handled carefully. To help spot errors
earlier, I added an assert into get_type_arch that the returned arch
is not null.
Inside gdbarch.c I found a few other places where auxiliary type data
was being allocated directly on the heap rather than on the types
obstack. I have fixed these to call TYPE_ALLOC now.
Finally, it is worth noting that as we don't clean up our gdbarch
objects yet, then this will not make much of an impact on the amount
of memory reported as lost at program termination time. Memory
allocated for auxiliary type information is still not freed, however,
it is now on the correct obstack. If we do ever start freeing our
gdbarch structures then the associated type data will be cleaned up
correctly.
Tested on X86-64 GNU/Linux with no regressions.
gdb/ChangeLog:
* eval.c (fake_method::fake_method): Call xzalloc directly for a
type that is neither object file owned, nor gdbarch owned.
* gdbtypes.c (get_type_gdbarch): Add an assert that returned
gdbarch is non-NULL.
(alloc_type_instance): Allocate non-objfile owned types on the
gdbarch obstack.
(copy_type_recursive): Allocate TYPE_FIELDS and TYPE_RANGE_DATA
using TYPE_ALLOC to ensure memory is allocated on the correct
obstack.
* gdbtypes.h (TYPE_ALLOC): Allocate space on either the objfile
obstack, or the gdbarch obstack.
(TYPE_ZALLOC): Rewrite using TYPE_ALLOC.
---
gdb/ChangeLog | 15 +++++++++++++++
gdb/eval.c | 6 +++++-
gdb/gdbtypes.c | 21 ++++++++++++++++-----
gdb/gdbtypes.h | 37 +++++++++++++++++++++----------------
4 files changed, 57 insertions(+), 22 deletions(-)
Comments
>>>>> "Andrew" == Andrew Burgess <andrew.burgess@embecosm.com> writes:
>> Would you mind seeing what happens if you use obstack_zalloc here
>> instead? It seems like that would give us some extra compile-time
>> safety. If it fails, that would be interesting.
Andrew> Unless I'm missing something (very possible) there's no suitable
Andrew> obstack_zalloc to use here.
Yeah, you're right, thanks for looking.
Andrew> - TYPE_FIELDS (new_type) = XCNEWVEC (struct field, nfields);
Andrew> + TYPE_FIELDS (new_type) = (struct field *)
Andrew> + TYPE_ALLOC (new_type, nfields * sizeof (struct field));
XCNEWVEC clears the memory so I think this should be TYPE_ZALLOC.
This is ok with this change.
thanks,
Tom
@@ -683,9 +683,13 @@ fake_method::fake_method (type_instance_flags flags,
}
}
+ /* We don't use TYPE_ZALLOC here to allocate space as TYPE is owned by
+ neither an objfile nor a gdbarch. As a result we must manually
+ allocate memory for auxiliary fields, and free the memory ourselves
+ when we are done with it. */
TYPE_NFIELDS (type) = num_types;
TYPE_FIELDS (type) = (struct field *)
- TYPE_ZALLOC (type, sizeof (struct field) * num_types);
+ xzalloc (sizeof (struct field) * num_types);
while (num_types-- > 0)
TYPE_FIELD_TYPE (type, num_types) = param_types[num_types];
@@ -233,10 +233,19 @@ alloc_type_copy (const struct type *type)
struct gdbarch *
get_type_arch (const struct type *type)
{
+ struct gdbarch *arch;
+
if (TYPE_OBJFILE_OWNED (type))
- return get_objfile_arch (TYPE_OWNER (type).objfile);
+ arch = get_objfile_arch (TYPE_OWNER (type).objfile);
else
- return TYPE_OWNER (type).gdbarch;
+ arch = TYPE_OWNER (type).gdbarch;
+
+ /* The ARCH can be NULL if TYPE is associated with neither an objfile nor
+ a gdbarch, however, this is very rare, and even then, in most cases
+ that get_type_arch is called, we assume that a non-NULL value is
+ returned. */
+ gdb_assert (arch != NULL);
+ return arch;
}
/* See gdbtypes.h. */
@@ -277,7 +286,7 @@ alloc_type_instance (struct type *oldtype)
/* Allocate the structure. */
if (! TYPE_OBJFILE_OWNED (oldtype))
- type = XCNEW (struct type);
+ type = GDBARCH_OBSTACK_ZALLOC (get_type_arch (oldtype), struct type);
else
type = OBSTACK_ZALLOC (&TYPE_OBJFILE (oldtype)->objfile_obstack,
struct type);
@@ -4903,7 +4912,8 @@ copy_type_recursive (struct objfile *objfile,
int i, nfields;
nfields = TYPE_NFIELDS (type);
- TYPE_FIELDS (new_type) = XCNEWVEC (struct field, nfields);
+ TYPE_FIELDS (new_type) = (struct field *)
+ TYPE_ALLOC (new_type, nfields * sizeof (struct field));
for (i = 0; i < nfields; i++)
{
TYPE_FIELD_ARTIFICIAL (new_type, i) =
@@ -4946,7 +4956,8 @@ copy_type_recursive (struct objfile *objfile,
/* For range types, copy the bounds information. */
if (TYPE_CODE (type) == TYPE_CODE_RANGE)
{
- TYPE_RANGE_DATA (new_type) = XNEW (struct range_bounds);
+ TYPE_RANGE_DATA (new_type) = (struct range_bounds *)
+ TYPE_ALLOC (new_type, sizeof (struct range_bounds));
*TYPE_RANGE_DATA (new_type) = *TYPE_RANGE_DATA (type);
}
@@ -49,6 +49,7 @@
#include "common/enum-flags.h"
#include "common/underlying.h"
#include "common/print-utils.h"
+#include "gdbarch.h"
/* Forward declarations for prototypes. */
struct field;
@@ -1717,26 +1718,30 @@ extern const struct floatformat *floatformats_vax_d[BFD_ENDIAN_UNKNOWN];
extern const struct floatformat *floatformats_ibm_long_double[BFD_ENDIAN_UNKNOWN];
-/* * Allocate space for storing data associated with a particular
+/* Allocate space for storing data associated with a particular
type. We ensure that the space is allocated using the same
mechanism that was used to allocate the space for the type
structure itself. I.e. if the type is on an objfile's
objfile_obstack, then the space for data associated with that type
- will also be allocated on the objfile_obstack. If the type is not
- associated with any particular objfile (such as builtin types),
- then the data space will be allocated with xmalloc, the same as for
- the type structure. */
-
-#define TYPE_ALLOC(t,size) \
- (TYPE_OBJFILE_OWNED (t) \
- ? obstack_alloc (&TYPE_OBJFILE (t) -> objfile_obstack, size) \
- : xmalloc (size))
-
-#define TYPE_ZALLOC(t,size) \
- (TYPE_OBJFILE_OWNED (t) \
- ? memset (obstack_alloc (&TYPE_OBJFILE (t)->objfile_obstack, size), \
- 0, size) \
- : xzalloc (size))
+ will also be allocated on the objfile_obstack. If the type is
+ associated with a gdbarch, then the space for data associated with that
+ type will also be allocated on the gdbarch_obstack.
+
+ If a type is not associated with neither an objfile or a gdbarch then
+ you should not use this macro to allocate space for data, instead you
+ should call xmalloc directly, and ensure the memory is correctly freed
+ when it is no longer needed. */
+
+#define TYPE_ALLOC(t,size) \
+ (obstack_alloc ((TYPE_OBJFILE_OWNED (t) \
+ ? &TYPE_OBJFILE (t)->objfile_obstack \
+ : gdbarch_obstack (TYPE_OWNER (t).gdbarch)), \
+ size))
+
+
+/* See comment on TYPE_ALLOC. */
+
+#define TYPE_ZALLOC(t,size) (memset (TYPE_ALLOC (t, size), 0, size))
/* Use alloc_type to allocate a type owned by an objfile. Use
alloc_type_arch to allocate a type owned by an architecture. Use