gdb: avoid resolving dynamic properties for non-allocated arrays

In PR gdb/27059 an issue was discovered where GDB would sometimes
trigger undefined behaviour in the form of signed integer overflow.
The problem here is that GDB was reading random garbage from the
inferior memory space, assuming this data was valid, and performing
arithmetic on it.

This bug raises an interesting general problem with GDB's DWARF
expression evaluator, which is this:

We currently assume that the DWARF expressions being evaluated are
well formed, and well behaving.  As an example, this is the expression
that the bug was running into problems on, this was used as the
expression for a DW_AT_byte_stride of a DW_TAG_subrange_type:

	DW_OP_push_object_address;
	DW_OP_plus_uconst: 88;
	DW_OP_deref;
	DW_OP_push_object_address;
	DW_OP_plus_uconst: 32;
	DW_OP_deref;
	DW_OP_mul

Two values are read from the inferior and multiplied together.  GDB
should not assume that any value read from the inferior is in any way
sane, as such the implementation of DW_OP_mul should be guarding
against overflow and doing something semi-sane here.

However, it turns out that the original bug PR gdb/27059, is hitting a
more specific case, which doesn't require changes to the DWARF
expression evaluator, so I'm going to leave the above issue for
another day.

In the test mentioned in the bug GDB is actually trying to resolve the
dynamic type of a Fortran array that is NOT allocated.  A
non-allocated Fortran array is one that does not have any data
allocated for it yet, and even the upper and lower bounds of the array
are not yet known.

It turns out that, at least for gfortran compiled code, the data
fields that describe the byte-stride are not initialised until the
array is allocated.

This leads me to the following conclusion: GDB should not try to
resolve the bounds, or stride information for an array that is not
allocated (or not associated, a similar, but slightly different
Fortran feature).  Instead, each of these properties should be set to
undefined if the array is not allocated (or associated).

That is what this commit does.  There's a new flag that is passed
around during the dynamic array resolution.  When this flag is true
the dynamic properties are resolved using the DWARF expressions as
they currently are, but when this flag is false the expressions are
not evaluated, and instead the properties are set to undefined.

gdb/ChangeLog:

	PR gdb/27059
	* eval.c (evaluate_subexp_for_sizeof): Handle not allocated and
	not associated arrays.
	* f-lang.c (fortran_adjust_dynamic_array_base_address_hack): Don't
	adjust arrays that are not allocated/associated.
	* gdbtypes.c (resolve_dynamic_range): Update header comment.  Add
	new parameter which is used to sometimes set dynamic properties to
	undefined.
	(resolve_dynamic_array_or_string): Update header comment.  Add new
	parameter which is used to guard evaluating dynamic properties.
	Resolve allocated/associated properties first.

gdb/testsuite/ChangeLog:

	PR gdb/27059
	* gdb.dwarf2/dyn-type-unallocated.c: New file.
	* gdb.dwarf2/dyn-type-unallocated.exp: New file.
This commit is contained in:
Andrew Burgess 2020-12-18 11:59:54 +00:00
parent 5ba3b20ec2
commit b7874836c3
7 changed files with 240 additions and 24 deletions

View file

@ -1,3 +1,17 @@
2020-12-24 Andrew Burgess <andrew.burgess@embecosm.com>
PR gdb/27059
* eval.c (evaluate_subexp_for_sizeof): Handle not allocated and
not associated arrays.
* f-lang.c (fortran_adjust_dynamic_array_base_address_hack): Don't
adjust arrays that are not allocated/associated.
* gdbtypes.c (resolve_dynamic_range): Update header comment. Add
new parameter which is used to sometimes set dynamic properties to
undefined.
(resolve_dynamic_array_or_string): Update header comment. Add new
parameter which is used to guard evaluating dynamic properties.
Resolve allocated/associated properties first.
2020-12-24 Andrew Burgess <andrew.burgess@embecosm.com>
* gdbtypes (recursive_dump_type): Include allocated and associated

View file

@ -2961,10 +2961,14 @@ evaluate_subexp_for_sizeof (struct expression *exp, int *pos,
{
val = evaluate_subexp (nullptr, exp, pos, EVAL_NORMAL);
type = value_type (val);
if (type->code () == TYPE_CODE_ARRAY
&& is_dynamic_type (type->index_type ())
&& type->bounds ()->high.kind () == PROP_UNDEFINED)
return allocate_optimized_out_value (size_type);
if (type->code () == TYPE_CODE_ARRAY)
{
if (type_not_allocated (type) || type_not_associated (type))
return value_zero (size_type, not_lval);
else if (is_dynamic_type (type->index_type ())
&& type->bounds ()->high.kind () == PROP_UNDEFINED)
return allocate_optimized_out_value (size_type);
}
}
else
(*pos) += 4;

View file

@ -1391,6 +1391,10 @@ fortran_adjust_dynamic_array_base_address_hack (struct type *type,
{
gdb_assert (type->code () == TYPE_CODE_ARRAY);
/* We can't adjust the base address for arrays that have no content. */
if (type_not_allocated (type) || type_not_associated (type))
return address;
int ndimensions = calc_f77_array_dims (type);
LONGEST total_offset = 0;

View file

@ -2177,11 +2177,20 @@ static struct type *resolve_dynamic_type_internal
/* Given a dynamic range type (dyn_range_type) and a stack of
struct property_addr_info elements, return a static version
of that type. */
of that type.
When RESOLVE_P is true then the returned static range is created by
actually evaluating any dynamic properties within the range type, while
when RESOLVE_P is false the returned static range has all of the bounds
and stride information set to undefined. The RESOLVE_P set to false
case will be used when evaluating a dynamic array that is not
allocated, or not associated, i.e. the bounds information might not be
initialized yet. */
static struct type *
resolve_dynamic_range (struct type *dyn_range_type,
struct property_addr_info *addr_stack)
struct property_addr_info *addr_stack,
bool resolve_p = true)
{
CORE_ADDR value;
struct type *static_range_type, *static_target_type;
@ -2190,13 +2199,13 @@ resolve_dynamic_range (struct type *dyn_range_type,
gdb_assert (dyn_range_type->code () == TYPE_CODE_RANGE);
const struct dynamic_prop *prop = &dyn_range_type->bounds ()->low;
if (dwarf2_evaluate_property (prop, NULL, addr_stack, &value))
if (resolve_p && dwarf2_evaluate_property (prop, NULL, addr_stack, &value))
low_bound.set_const_val (value);
else
low_bound.set_undefined ();
prop = &dyn_range_type->bounds ()->high;
if (dwarf2_evaluate_property (prop, NULL, addr_stack, &value))
if (resolve_p && dwarf2_evaluate_property (prop, NULL, addr_stack, &value))
{
high_bound.set_const_val (value);
@ -2209,7 +2218,7 @@ resolve_dynamic_range (struct type *dyn_range_type,
bool byte_stride_p = dyn_range_type->bounds ()->flag_is_byte_stride;
prop = &dyn_range_type->bounds ()->stride;
if (dwarf2_evaluate_property (prop, NULL, addr_stack, &value))
if (resolve_p && dwarf2_evaluate_property (prop, NULL, addr_stack, &value))
{
stride.set_const_val (value);
@ -2242,11 +2251,16 @@ resolve_dynamic_range (struct type *dyn_range_type,
/* Resolves dynamic bound values of an array or string type TYPE to static
ones. ADDR_STACK is a stack of struct property_addr_info to be used if
needed during the dynamic resolution. */
needed during the dynamic resolution.
When RESOLVE_P is true then the dynamic properties of TYPE are
evaluated, otherwise the dynamic properties of TYPE are not evaluated,
instead we assume the array is not allocated/associated yet. */
static struct type *
resolve_dynamic_array_or_string (struct type *type,
struct property_addr_info *addr_stack)
struct property_addr_info *addr_stack,
bool resolve_p = true)
{
CORE_ADDR value;
struct type *elt_type;
@ -2262,29 +2276,44 @@ resolve_dynamic_array_or_string (struct type *type,
type = copy_type (type);
elt_type = type;
range_type = check_typedef (elt_type->index_type ());
range_type = resolve_dynamic_range (range_type, addr_stack);
/* Resolve allocated/associated here before creating a new array type, which
will update the length of the array accordingly. */
/* Resolve the allocated and associated properties before doing anything
else. If an array is not allocated or not associated then (at least
for Fortran) there is no guarantee that the data to define the upper
bound, lower bound, or stride will be correct. If RESOLVE_P is
already false at this point then this is not the first dimension of
the array and a more outer dimension has already marked this array as
not allocated/associated, as such we just ignore this property. This
is fine as GDB only checks the allocated/associated on the outer most
dimension of the array. */
prop = TYPE_ALLOCATED_PROP (type);
if (prop != NULL && dwarf2_evaluate_property (prop, NULL, addr_stack, &value))
prop->set_const_val (value);
if (prop != NULL && resolve_p
&& dwarf2_evaluate_property (prop, NULL, addr_stack, &value))
{
prop->set_const_val (value);
if (value == 0)
resolve_p = false;
}
prop = TYPE_ASSOCIATED_PROP (type);
if (prop != NULL && dwarf2_evaluate_property (prop, NULL, addr_stack, &value))
prop->set_const_val (value);
if (prop != NULL && resolve_p
&& dwarf2_evaluate_property (prop, NULL, addr_stack, &value))
{
prop->set_const_val (value);
if (value == 0)
resolve_p = false;
}
ary_dim = check_typedef (TYPE_TARGET_TYPE (elt_type));
range_type = check_typedef (type->index_type ());
range_type = resolve_dynamic_range (range_type, addr_stack, resolve_p);
ary_dim = check_typedef (TYPE_TARGET_TYPE (type));
if (ary_dim != NULL && ary_dim->code () == TYPE_CODE_ARRAY)
elt_type = resolve_dynamic_array_or_string (ary_dim, addr_stack);
elt_type = resolve_dynamic_array_or_string (ary_dim, addr_stack, resolve_p);
else
elt_type = TYPE_TARGET_TYPE (type);
prop = type->dyn_prop (DYN_PROP_BYTE_STRIDE);
if (prop != NULL)
if (prop != NULL && resolve_p)
{
if (dwarf2_evaluate_property (prop, NULL, addr_stack, &value))
{

View file

@ -1,3 +1,9 @@
2020-12-24 Andrew Burgess <andrew.burgess@embecosm.com>
PR gdb/27059
* gdb.dwarf2/dyn-type-unallocated.c: New file.
* gdb.dwarf2/dyn-type-unallocated.exp: New file.
2020-12-21 Peter Waller <p@pwaller.net>
* gdb.base/style-interp-exec-mi.exp: New.

View file

@ -0,0 +1,40 @@
/* This testcase is part of GDB, the GNU debugger.
Copyright 2020 Free Software Foundation, Inc.
This program is free software; you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation; either version 3 of the License, or
(at your option) any later version.
This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License
along with this program. If not, see <http://www.gnu.org/licenses/>. */
#include "../lib/attributes.h"
/* Our fake dynamic object. */
void *dyn_object;
void __attribute__((noinline)) ATTRIBUTE_NOCLONE
marker ()
{ /* Nothing. */ }
int
main ()
{
asm ("main_label: .globl main_label");
/* Initialise the dynamic object. */
dyn_object = 0;
asm ("marker_label: .globl marker_label");
marker (); /* Break here. */
return 0;
}

View file

@ -0,0 +1,119 @@
# Copyright 2020 Free Software Foundation, Inc.
# This program is free software; you can redistribute it and/or modify
# it under the terms of the GNU General Public License as published by
# the Free Software Foundation; either version 3 of the License, or
# (at your option) any later version.
#
# This program is distributed in the hope that it will be useful,
# but WITHOUT ANY WARRANTY; without even the implied warranty of
# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
# GNU General Public License for more details.
#
# You should have received a copy of the GNU General Public License
# along with this program. If not, see <http://www.gnu.org/licenses/>.
# Test for issue PR gdb/27059. The problem was that when resolving a
# dynamic type that was not-allocated GDB would still try to execute
# the DWARF expressions for the upper, lower, and byte-stride values.
#
# The problem is that, at least in some gfortran compiled programs,
# these values are undefined until the array is allocated.
#
# As a result, executing the dwarf expressions was triggering integer
# overflow in some cases.
#
# This test aims to make the sometimes occurring integer overflow a
# more noticeable error by creating an array that is always marked as
# not-allocated.
#
# The dwarf expressions for the various attributes then contains an
# infinite loop. If GDB ever tries to execute these expressions we
# will get a test timeout. With this issue fixed the expressions are
# never executed and the test completes as we'd expect.
load_lib dwarf.exp
if {![dwarf2_support]} {
return 0
}
standard_testfile .c -dw.S
if { [prepare_for_testing "failed to prepare" ${testfile} ${srcfile}] } {
return -1
}
set asm_file [standard_output_file $srcfile2]
Dwarf::assemble $asm_file {
cu {} {
global srcfile
compile_unit {
{producer "gcc" }
{language @DW_LANG_Fortran90}
{name ${srcfile}}
{low_pc 0 addr}
} {
declare_labels array_type_label integer_type_label
set int_size [get_sizeof "int" "UNKNOWN"]
set voidp_size [get_sizeof "void *" "UNKNOWN"]
integer_type_label: DW_TAG_base_type {
{DW_AT_byte_size $int_size DW_FORM_sdata}
{DW_AT_encoding @DW_ATE_signed}
{DW_AT_name integer}
}
array_type_label: DW_TAG_array_type {
{DW_AT_type :$integer_type_label}
{DW_AT_data_location {
DW_OP_push_object_address
DW_OP_deref
} SPECIAL_expr}
{DW_AT_allocated {
DW_OP_lit0
} SPECIAL_expr}
} {
DW_TAG_subrange_type {
{DW_AT_type :$integer_type_label}
{DW_AT_lower_bound {
DW_OP_skip -3
} SPECIAL_expr}
{DW_AT_upper_bound {
DW_OP_skip -3
} SPECIAL_expr}
{DW_AT_byte_stride {
DW_OP_skip -3
} SPECIAL_expr}
}
}
DW_TAG_variable {
{DW_AT_location {
DW_OP_addr [gdb_target_symbol dyn_object]
} SPECIAL_expr}
{name "dyn_object"}
{type :$array_type_label}
}
subprogram {
{external 1 flag}
{MACRO_AT_func {main}}
}
}
}
}
if { [prepare_for_testing "failed to prepare" "${testfile}" \
[list $srcfile $asm_file] {nodebug}] } {
return -1
}
if ![runto_main] {
return -1
}
gdb_breakpoint "marker_label"
gdb_continue_to_breakpoint "stop at marker_label"
gdb_test "ptype dyn_object" "type = integer, allocatable \\(:\\)"