2022-01-01 22:56:03 +08:00
|
|
|
# Copyright (C) 2021-2022 Free Software Foundation, Inc.
|
gdb: prevent an assertion when computing the frame_id for an inline frame
I ran into this assertion while GDB was trying to unwind the stack:
gdb/inline-frame.c:173: internal-error: void inline_frame_this_id(frame_info*, void**, frame_id*): Assertion `frame_id_p (*this_id)' failed.
That is, when building the frame_id for an inline frame, GDB asks for
the frame_id of the previous frame. Unfortunately, no valid frame_id
was returned for the previous frame, and so the assertion triggers.
What is happening is this, I had a stack that looked something like
this (the arrows '->' point from caller to callee):
normal_frame -> inline_frame
However, for whatever reason (e.g. broken debug information, or
corrupted stack contents in the inferior), when GDB tries to unwind
"normal_frame", it ends up getting back effectively the same frame,
thus the call stack looks like this to GDB:
.-> normal_frame -> inline_frame
| |
'-----'
Given such a situation we would expect GDB to terminate the stack with
an error like this:
Backtrace stopped: previous frame identical to this frame (corrupt stack?)
However, the inline_frame causes a problem, and here's why:
When unwinding we start from the sentinel frame and call
get_prev_frame. We eventually end up in get_prev_frame_if_no_cycle,
in here we create a raw frame, and as this is frame #0 we immediately
return.
However, eventually we will try to unwind the stack further. When we
do this we inevitably needing to know the frame_id for frame #0, and
so, eventually, we end up in compute_frame_id.
In compute_frame_id we first find the right unwinder for this frame,
in our case (i.e. for inline_frame) the $pc is within the function
normal_frame, but also within a block associated with the inlined
function inline_frame, as such the inline frame unwinder claims this
frame.
Back in compute_frame_id we next compute the frame_id, for our
inline_frame this means a call to inline_frame_this_id.
The ID of an inline frame is based on the id of the previous frame, so
from inline_frame_this_id we call get_prev_frame_always, this
eventually calls get_prev_frame_if_no_cycle again, which creates
another raw frame and calls compute_frame_id (for frames other than
frame 0 we immediately compute the frame_id).
In compute_frame_id we again identify the correct unwinder for this
frame. Our $pc is unchanged, however, the fact that the next frame is
of type INLINE_FRAME prevents the inline frame unwinder from claiming
this frame again, and so, the standard DWARF frame unwinder claims
normal_frame.
We return to compute_frame_id and call the standard DWARF function to
build the frame_id for normal_frame.
With the frame_id of normal_frame figured out we return to
compute_frame_id, and then to get_prev_frame_if_no_cycle, where we add
the ID for normal_frame into the frame_id cache, and return the frame
back to inline_frame_this_id.
From inline_frame_this_id we build a frame_id for inline_frame and
return to compute_frame_id, and then to get_prev_frame_if_no_cycle,
which adds the frame_id for inline_frame into the frame_id cache.
So far, so good.
However, as we are trying to unwind the complete stack, we eventually
ask for the previous frame of normal_frame, remember, at this point
GDB doesn't know the stack is corrupted (with a cycle), GDB still
needs to figure that out.
So, we eventually end up in get_prev_frame_if_no_cycle where we create
a raw frame and call compute_frame_id, remember, this is for the frame
before normal_frame.
The first task for compute_frame_id is to find the unwinder for this
frame, so all of the frame sniffers are tried in order, this includes
the inline frame sniffer.
The inline frame sniffer asks for the $pc, this request is sent up the
stack to normal_frame, which, due to its cyclic behaviour, tells GDB
that the $pc in the previous frame was the same as the $pc in
normal_frame.
GDB spots that this $pc corresponds to both the function normal_frame
and also the inline function inline_frame. As the next frame is not
an INLINE_FRAME then GDB figures that we have not yet built a frame to
cover inline_frame, and so the inline sniffer claims this new frame.
Our stack is now looking like this:
inline_frame -> normal_frame -> inline_frame
But, we have not yet computed the frame id for the outer most (on the
left) inline_frame. After the frame sniffer has claimed the inline
frame GDB returns to compute_frame_id and calls inline_frame_this_id.
In here GDB calls get_prev_frame_always, which eventually ends up
in get_prev_frame_if_no_cycle again, where we create a raw frame and
call compute_frame_id.
Just like before, compute_frame_id tries to find an unwinder for this
new frame, it sees that the $pc is within both normal_frame and
inline_frame, but the next frame is, again, an INLINE_FRAME, so, just
like before the standard DWARF unwinder claims this frame. Back in
compute_frame_id we again call the standard DWARF function to build
the frame_id for this new copy of normal_frame.
At this point the stack looks like this:
normal_frame -> inline_frame -> normal_frame -> inline_frame
After compute_frame_id we return to get_prev_frame_if_no_cycle, where
we try to add the frame_id for the new normal_frame into the frame_id
cache, however, unlike before, we fail to add this frame_id as it is
a duplicate of the previous normal_frame frame_id. Having found a
duplicate get_prev_frame_if_no_cycle unlinks the new frame from the
stack, and returns nullptr, the stack now looks like this:
inline_frame -> normal_frame -> inline_frame
The nullptr result from get_prev_frame_if_no_cycle is fed back to
inline_frame_this_id, which forwards this to get_frame_id, which
immediately returns null_frame_id. As null_frame_id is not considered
a valid frame_id, this is what triggers the assertion.
In summary then:
- inline_frame_this_id currently assumes that as the inline frame
exists, we will always get a valid frame back from
get_prev_frame_always,
- get_prev_frame_if_no_cycle currently assumes that it is safe to
return nullptr when it sees a cycle.
Notice that in frame.c:compute_frame_id, this code:
fi->this_id.value = outer_frame_id;
fi->unwind->this_id (fi, &fi->prologue_cache, &fi->this_id.value);
gdb_assert (frame_id_p (fi->this_id.value));
The assertion makes it clear that the this_id function must always
return a valid frame_id (e.g. null_frame_id is not a valid return
value), and similarly in inline_frame.c:inline_frame_this_id this
code:
*this_id = get_frame_id (get_prev_frame_always (this_frame));
/* snip comment */
gdb_assert (frame_id_p (*this_id));
Makes it clear that every inline frame expects to be able to get a
previous frame, which will have a valid frame_id.
As I have discussed above, these assumptions don't currently hold in
all cases.
One possibility would be to move the call to get_prev_frame_always
forward from inline_frame_this_id to inline_frame_sniffer, however,
this falls foul of (in frame.c:frame_cleanup_after_sniffer) this
assertion:
/* No sniffer should extend the frame chain; sniff based on what is
already certain. */
gdb_assert (!frame->prev_p);
This assert prohibits any sniffer from trying to get the previous
frame, as getting the previous frame is likely to depend on the next
frame, I can understand why this assertion is a good thing, and I'm in
no rush to alter this rule.
The solution proposed here takes onboard feedback from both Pedro, and
Simon (see the links below). The get_prev_frame_if_no_cycle function
is renamed to get_prev_frame_maybe_check_cycle, and will now not do
cycle detection for inline frames, even when we spot a duplicate frame
it is still returned. This is fine, as, if the normal frame has a
duplicate frame-id then the inline frame will also have a duplicate
frame-id. And so, when we reject the inline frame, the duplicate
normal frame, which is previous to the inline frame, will also be
rejected.
In inline-frame.c the call to get_prev_frame_always is no longer
nested inside the call to get_frame_id. There are reasons why
get_prev_frame_always can return nullptr, for example, if there is a
memory error while trying to get the previous frame, if this should
happen then we now give a more informative error message.
Historical Links:
Patch v2: https://sourceware.org/pipermail/gdb-patches/2021-June/180208.html
Feedback: https://sourceware.org/pipermail/gdb-patches/2021-July/180651.html
https://sourceware.org/pipermail/gdb-patches/2021-July/180663.html
Patch v3: https://sourceware.org/pipermail/gdb-patches/2021-July/181029.html
Feedback: https://sourceware.org/pipermail/gdb-patches/2021-July/181035.html
Additional input: https://sourceware.org/pipermail/gdb-patches/2021-September/182040.html
2021-05-27 05:03:23 +08:00
|
|
|
|
|
|
|
# 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/>.
|
|
|
|
|
|
|
|
import gdb
|
|
|
|
from gdb.unwinder import Unwinder
|
|
|
|
|
|
|
|
# Set this to the stack level the backtrace should be corrupted at.
|
|
|
|
# This will only work for frame 1, 3, or 5 in the test this unwinder
|
|
|
|
# was written for.
|
|
|
|
stop_at_level = None
|
|
|
|
|
|
|
|
# Set this to the stack frame size of frames 1, 3, and 5. These
|
|
|
|
# frames will all have the same stack frame size as they are the same
|
|
|
|
# function called recursively.
|
|
|
|
stack_adjust = None
|
|
|
|
|
|
|
|
|
|
|
|
class FrameId(object):
|
|
|
|
def __init__(self, sp, pc):
|
|
|
|
self._sp = sp
|
|
|
|
self._pc = pc
|
|
|
|
|
|
|
|
@property
|
|
|
|
def sp(self):
|
|
|
|
return self._sp
|
|
|
|
|
|
|
|
@property
|
|
|
|
def pc(self):
|
|
|
|
return self._pc
|
|
|
|
|
|
|
|
|
|
|
|
class TestUnwinder(Unwinder):
|
|
|
|
def __init__(self):
|
|
|
|
Unwinder.__init__(self, "stop at level")
|
|
|
|
|
|
|
|
def __call__(self, pending_frame):
|
|
|
|
global stop_at_level
|
|
|
|
global stack_adjust
|
|
|
|
|
|
|
|
if stop_at_level is None or pending_frame.level() != stop_at_level:
|
|
|
|
return None
|
|
|
|
|
|
|
|
if stack_adjust is None:
|
|
|
|
raise gdb.GdbError("invalid stack_adjust")
|
|
|
|
|
|
|
|
if not stop_at_level in [1, 3, 5]:
|
|
|
|
raise gdb.GdbError("invalid stop_at_level")
|
|
|
|
|
|
|
|
sp_desc = pending_frame.architecture().registers().find("sp")
|
|
|
|
sp = pending_frame.read_register(sp_desc) + stack_adjust
|
|
|
|
pc = (gdb.lookup_symbol("normal_func"))[0].value().address
|
|
|
|
unwinder = pending_frame.create_unwind_info(FrameId(sp, pc))
|
|
|
|
|
|
|
|
for reg in pending_frame.architecture().registers("general"):
|
|
|
|
val = pending_frame.read_register(reg)
|
|
|
|
unwinder.add_saved_register(reg, val)
|
|
|
|
return unwinder
|
|
|
|
|
|
|
|
|
|
|
|
gdb.unwinder.register_unwinder(None, TestUnwinder(), True)
|
|
|
|
|
|
|
|
# When loaded, it is expected that the stack looks like:
|
|
|
|
#
|
|
|
|
# main -> normal_func -> inline_func -> normal_func -> inline_func -> normal_func -> inline_func
|
|
|
|
#
|
|
|
|
# Compute the stack frame size of normal_func, which has inline_func
|
|
|
|
# inlined within it.
|
|
|
|
f0 = gdb.newest_frame()
|
|
|
|
f1 = f0.older()
|
|
|
|
f2 = f1.older()
|
|
|
|
f0_sp = f0.read_register("sp")
|
|
|
|
f2_sp = f2.read_register("sp")
|
|
|
|
stack_adjust = f2_sp - f0_sp
|