[PATCH] Fix inline frame unwinding breakage
Andrew Burgess
andrew.burgess@embecosm.com
Thu Jun 18 17:29:22 GMT 2020
* Andrew Burgess <andrew.burgess@embecosm.com> [2020-06-18 17:58:55 +0100]:
> * Luis Machado via Gdb-patches <gdb-patches@sourceware.org> [2020-04-14 18:38:36 -0300]:
>
> > *** re-sending due to the poor choice of characters for the backtrace
> > annotations. GIT swallowed parts of it.
> >
> > There has been some breakage for aarch64-linux, arm-linux and s390-linux in
> > terms of inline frame unwinding. There may be other targets, but these are
> > the ones i'm aware of.
> >
> > The following testcases started to show numerous failures and trigger internal
> > errors in GDB after commit 1009d92fc621bc4d017029b90a5bfab16e17fde5,
> > "Find tailcall frames before inline frames".
> >
> > gdb.opt/inline-break.exp
> > gdb.opt/inline-cmds.exp
> > gdb.python/py-frame-inline.exp
> > gdb.reverse/insn-reverse.exp
> >
> > The internal errors were of this kind:
> >
> > binutils-gdb/gdb/frame.c:579: internal-error: frame_id get_frame_id(frame_info*): Assertion `fi->level == 0' failed.
> >
> > After a lengthy investigation to try and find the cause of these assertions,
> > it seems we're dealing with some fragile/poorly documented code to handle inline
> > frames and we are attempting to unwind from this fragile section of code.
> >
> > Before commit 1009d92fc621bc4d017029b90a5bfab16e17fde5, the tailcall sniffer
> > was invoked from dwarf2_frame_prev_register. By the time we invoke the
> > dwarf2_frame_prev_register function, we've probably already calculated the
> > frame id (via compute_frame_id).
> >
> > After said commit, the call to dwarf2_tailcall_sniffer_first was moved to
> > dwarf2_frame_cache. This is very early in a frame creation process, and
> > we're still calculating the frame ID (so compute_frame_id is in the call
> > stack).
> >
> > This would be fine for regular frames, but the above testcases all deal
> > with some inline frames.
> >
> > The particularity of inline frames is that their frame ID's depend on
> > the previous frame's ID, and the previous frame's ID relies in the inline
> > frame's registers. So it is a bit of a messy situation.
> >
> > We have comments in various parts of the code warning about some of these
> > particularities.
> >
> > In the case of dwarf2_tailcall_sniffer_first, we attempt to unwind the PC,
> > which goes through various functions until we eventually invoke
> > frame_unwind_got_register. This function will eventually attempt to create
> > a lazy value for a particular register, and this lazy value will require
> > a valid frame ID. Since the inline frame doesn't have a valid frame ID
> > yet (remember we're still calculating the previous frame's ID so we can tell
> > what the inline frame ID is) we will call compute_frame_id for the inline
> > frame (level 0).
> >
> > We'll eventually hit the assertion above, inside get_frame_id:
> >
> > --
> > /* If we haven't computed the frame id yet, then it must be that
> > this is the current frame. Compute it now, and stash the
> > result. The IDs of other frames are computed as soon as
> > they're created, in order to detect cycles. See
> > get_prev_frame_if_no_cycle. */
> > gdb_assert (fi->level == 0);
> > --
> >
> > It seems to me we shouldn't have reached this assertion without having the
> > inline frame ID already calculated. In fact, it seems we even start recursing
> > a bit when we invoke get_prev_frame_always within inline_frame_this_id. But
> > a check makes us quit the recursion and proceed to compute the id.
> >
> > Here's the call stack for context:
> >
> > #0 get_prev_frame_always_1 (this_frame=0xaaaaab85a670) at ../../../repos/binutils-gdb/gdb/frame.c:2109
> > RECURSION - #1 0x0000aaaaaae1d098 in get_prev_frame_always (this_frame=0xaaaaab85a670) at ../../../repos/binutils-gdb/gdb/frame.c:2124
> > #2 0x0000aaaaaae95768 in inline_frame_this_id (this_frame=0xaaaaab85a670, this_cache=0xaaaaab85a688, this_id=0xaaaaab85a6d0)
> > at ../../../repos/binutils-gdb/gdb/inline-frame.c:165
> > #3 0x0000aaaaaae1916c in compute_frame_id (fi=0xaaaaab85a670) at ../../../repos/binutils-gdb/gdb/frame.c:550
> > #4 0x0000aaaaaae19318 in get_frame_id (fi=0xaaaaab85a670) at ../../../repos/binutils-gdb/gdb/frame.c:582
> > #5 0x0000aaaaaae13480 in value_of_register_lazy (frame=0xaaaaab85a730, regnum=30) at ../../../repos/binutils-gdb/gdb/findvar.c:296
> > #6 0x0000aaaaaae16c00 in frame_unwind_got_register (frame=0xaaaaab85a730, regnum=30, new_regnum=30) at ../../../repos/binutils-gdb/gdb/frame-unwind.c:268
> > #7 0x0000aaaaaad52604 in dwarf2_frame_prev_register (this_frame=0xaaaaab85a730, this_cache=0xaaaaab85a748, regnum=30)
> > at ../../../repos/binutils-gdb/gdb/dwarf2/frame.c:1296
> > #8 0x0000aaaaaae1ae68 in frame_unwind_register_value (next_frame=0xaaaaab85a730, regnum=30) at ../../../repos/binutils-gdb/gdb/frame.c:1229
> > #9 0x0000aaaaaae1b304 in frame_unwind_register_unsigned (next_frame=0xaaaaab85a730, regnum=30) at ../../../repos/binutils-gdb/gdb/frame.c:1320
> > #10 0x0000aaaaaab76574 in aarch64_dwarf2_prev_register (this_frame=0xaaaaab85a730, this_cache=0xaaaaab85a748, regnum=32)
> > at ../../../repos/binutils-gdb/gdb/aarch64-tdep.c:1114
> > #11 0x0000aaaaaad52724 in dwarf2_frame_prev_register (this_frame=0xaaaaab85a730, this_cache=0xaaaaab85a748, regnum=32)
> > at ../../../repos/binutils-gdb/gdb/dwarf2/frame.c:1316
> > #12 0x0000aaaaaae1ae68 in frame_unwind_register_value (next_frame=0xaaaaab85a730, regnum=32) at ../../../repos/binutils-gdb/gdb/frame.c:1229
> > #13 0x0000aaaaaae1b304 in frame_unwind_register_unsigned (next_frame=0xaaaaab85a730, regnum=32) at ../../../repos/binutils-gdb/gdb/frame.c:1320
> > #14 0x0000aaaaaae16a84 in default_unwind_pc (gdbarch=0xaaaaab81edc0, next_frame=0xaaaaab85a730) at ../../../repos/binutils-gdb/gdb/frame-unwind.c:223
> > #15 0x0000aaaaaae32124 in gdbarch_unwind_pc (gdbarch=0xaaaaab81edc0, next_frame=0xaaaaab85a730) at ../../../repos/binutils-gdb/gdb/gdbarch.c:3074
> > #16 0x0000aaaaaad4f15c in dwarf2_tailcall_sniffer_first (this_frame=0xaaaaab85a730, tailcall_cachep=0xaaaaab85a830, entry_cfa_sp_offsetp=0x0)
> > at ../../../repos/binutils-gdb/gdb/dwarf2/frame-tailcall.c:388
> > #17 0x0000aaaaaad520c0 in dwarf2_frame_cache (this_frame=0xaaaaab85a730, this_cache=0xaaaaab85a748) at ../../../repos/binutils-gdb/gdb/dwarf2/frame.c:1190
> > #18 0x0000aaaaaad52204 in dwarf2_frame_this_id (this_frame=0xaaaaab85a730, this_cache=0xaaaaab85a748, this_id=0xaaaaab85a790)
> > at ../../../repos/binutils-gdb/gdb/dwarf2/frame.c:1218
> > #19 0x0000aaaaaae1916c in compute_frame_id (fi=0xaaaaab85a730) at ../../../repos/binutils-gdb/gdb/frame.c:550
> > #20 0x0000aaaaaae1c958 in get_prev_frame_if_no_cycle (this_frame=0xaaaaab85a670) at ../../../repos/binutils-gdb/gdb/frame.c:1927
> > #21 0x0000aaaaaae1cc44 in get_prev_frame_always_1 (this_frame=0xaaaaab85a670) at ../../../repos/binutils-gdb/gdb/frame.c:2006
> > FIRST CALL - #22 0x0000aaaaaae1d098 in get_prev_frame_always (this_frame=0xaaaaab85a670) at ../../../repos/binutils-gdb/gdb/frame.c:2124
> > #23 0x0000aaaaaae18f68 in skip_artificial_frames (frame=0xaaaaab85a670) at ../../../repos/binutils-gdb/gdb/frame.c:495
> > #24 0x0000aaaaaae193e8 in get_stack_frame_id (next_frame=0xaaaaab85a670) at ../../../repos/binutils-gdb/gdb/frame.c:596
> > #25 0x0000aaaaaae87a54 in process_event_stop_test (ecs=0xffffffffefc8) at ../../../repos/binutils-gdb/gdb/infrun.c:6857
> > #26 0x0000aaaaaae86bdc in handle_signal_stop (ecs=0xffffffffefc8) at ../../../repos/binutils-gdb/gdb/infrun.c:6381
> > #27 0x0000aaaaaae84fd0 in handle_inferior_event (ecs=0xffffffffefc8) at ../../../repos/binutils-gdb/gdb/infrun.c:5578
> > #28 0x0000aaaaaae81588 in fetch_inferior_event (client_data=0x0) at ../../../repos/binutils-gdb/gdb/infrun.c:4020
> > #29 0x0000aaaaaae5f7fc in inferior_event_handler (event_type=INF_REG_EVENT, client_data=0x0) at ../../../repos/binutils-gdb/gdb/inf-loop.c:43
> > #30 0x0000aaaaaae8d768 in infrun_async_inferior_event_handler (data=0x0) at ../../../repos/binutils-gdb/gdb/infrun.c:9377
> > #31 0x0000aaaaaabff970 in check_async_event_handlers () at ../../../repos/binutils-gdb/gdb/async-event.c:291
> > #32 0x0000aaaaab27cbec in gdb_do_one_event () at ../../../repos/binutils-gdb/gdbsupport/event-loop.cc:194
> > #33 0x0000aaaaaaef1894 in start_event_loop () at ../../../repos/binutils-gdb/gdb/main.c:356
> > #34 0x0000aaaaaaef1a04 in captured_command_loop () at ../../../repos/binutils-gdb/gdb/main.c:416
> > #35 0x0000aaaaaaef3338 in captured_main (data=0xfffffffff1f0) at ../../../repos/binutils-gdb/gdb/main.c:1254
> > #36 0x0000aaaaaaef33a0 in gdb_main (args=0xfffffffff1f0) at ../../../repos/binutils-gdb/gdb/main.c:1269
> > #37 0x0000aaaaaab6e0dc in main (argc=6, argv=0xfffffffff348) at ../../../repos/binutils-gdb/gdb/gdb.c:32
> >
> > The following patch addresses this by using a function that unwinds the PC
> > from the next (inline) frame directly as opposed to creating a lazy value
> > that is bound to the next frame's ID (still not computed).
> >
> > I've validated this for aarch64-linux and x86_64-linux by running the
> > testsuite.
> >
> > Tromey, would you mind checking if this suits your problematic core file
> > tailcall scenario?
> >
> > gdb/ChangeLog:
> >
> > 2020-04-14 Luis Machado <luis.machado@linaro.org>
> >
> > * dwarf2/frame-tailcall.c (dwarf2_tailcall_sniffer_first): Use
> > get_frame_register instead of gdbarch_unwind_pc.
> > ---
> > gdb/dwarf2/frame-tailcall.c | 4 +++-
> > 1 file changed, 3 insertions(+), 1 deletion(-)
> >
> > diff --git a/gdb/dwarf2/frame-tailcall.c b/gdb/dwarf2/frame-tailcall.c
> > index 2d219f13f9..01bb134a5c 100644
> > --- a/gdb/dwarf2/frame-tailcall.c
> > +++ b/gdb/dwarf2/frame-tailcall.c
> > @@ -385,7 +385,9 @@ dwarf2_tailcall_sniffer_first (struct frame_info *this_frame,
> > prev_gdbarch = frame_unwind_arch (this_frame);
> >
> > /* Simulate frame_unwind_pc without setting this_frame->prev_pc.p. */
> > - prev_pc = gdbarch_unwind_pc (prev_gdbarch, this_frame);
> > + get_frame_register (this_frame, gdbarch_pc_regnum (prev_gdbarch),
> > + (gdb_byte *) &prev_pc);
> > + prev_pc = gdbarch_addr_bits_remove (prev_gdbarch, prev_pc);
> >
> > /* call_site_find_chain can throw an exception. */
> > chain = call_site_find_chain (prev_gdbarch, prev_pc, this_pc);
>
> I'm now no longer convinced that this patch is correct, and I'd like
> to reopen the discussion.
>
> Here's what concerns me, we used to make the following call-chain:
>
> gdbarch_unwind_pc --> frame_unwind_register_unsigned --> frame_unwind_register_value
>
> Now we do this:
>
> get_frame_register --> frame_unwind_register --> frame_register_unwind --> frame_unwind_register_value
>
> The problem is that gdbarch_unwind_pc' takes an argument 'next_frame',
> while, get_frame_register takes an argument called frame', but is
> really 'this_frame', it then passes 'frame->next' to
> 'frame_unwind_register'.
>
> What this means is that if we have a call stack like this:
>
> #3 --> #2 --> #1 --> #0
>
> And we invoke the tail-call sniffer in frame #1, previously we figured
> out the $pc value in frame #2, while now we figure out the $pc value
> in frame #1.
>
> I'm even more convinced that this is an error based on the fix patch
> you applied later:
>
> commit 991a3e2e9944a4b3a27bd989ac03c18285bd545d
> Date: Sat Apr 25 00:32:44 2020 -0300
>
> Fix remaining inline/tailcall unwinding breakage for x86_64
>
> This basically sent all but a select few cases down the old code path,
> while restricting just a few cases to the new path.
>
> I ran the testsuite (on x86-64/Linux) looking for cases where the new
> code actually triggers and there are just 2. Remember that this code
> is use the $pc value to identify tail-call chains.
>
> In both of the cases I found, both _before_ and _after_ your change, a
> tail-call chain was not identified. What this means is that even if
> your code is returning the wrong value, it's not going to cause a test
> regression.
>
> Finally, if you catch the cases where your new code triggers, and then
> step into call_site_find_chain (which is called later in the sniffer),
> you'll see that this function is passed a caller address and a callee
> address. The callee address passed in is 'this_pc', in our example
> above, this is the current address in #1. We previously used to
> compute the address is #2, which makes sense, we're looking for a
> chain of tail-calls that gets us from #2 to #1.
>
> However, after your change we're now simply passing in the address in
> #1 as both the caller and the callee address, which makes no sense (to
> me, right now).
>
> I'm still investigating at the moment. Right now I have more
> questions than answer, but I wanted to raise my concerns in case I'm
> just totally missing something obvious and you can set me straight.
Patch below is a test case that reveals the issue. You'll notice that
if you revert this patch then there's an extra frame in the backtrace
that is missing with this patch.
Still looking into what the right fix here is, but would welcome
discussion.
[ It just occurred to me that the test case could end up being target
and compiler version dependent. I'm on X86-64/Linux with GCC version
'gcc (GCC) 9.3.1 20200408 (Red Hat 9.3.1-2)'. ]
Thanks,
Andrew
---
commit 566b2b1da20e461cee2798f3eda741c1e31bdff6
Author: Andrew Burgess <andrew.burgess@embecosm.com>
Date: Thu Jun 18 18:25:00 2020 +0100
gdb/testsuite: Test case for inline func, tailcall bug
diff --git a/gdb/testsuite/gdb.opt/inline-frame-tailcall.c b/gdb/testsuite/gdb.opt/inline-frame-tailcall.c
new file mode 100644
index 00000000000..2513c257a29
--- /dev/null
+++ b/gdb/testsuite/gdb.opt/inline-frame-tailcall.c
@@ -0,0 +1,52 @@
+/* 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/>. */
+
+#ifdef __GNUC__
+# define ATTR_INLINE __attribute__((gnu_inline)) __attribute__((always_inline)) __attribute__((noclone))
+#else
+# define ATTR_INLINE
+#endif
+
+volatile int global;
+
+volatile int counter;
+
+static inline ATTR_INLINE int
+bar ()
+{
+ /* Just some filler. */
+ for (counter = 0; counter < 10; ++counter)
+ global = 0;
+ return 0;
+}
+
+__attribute__ ((noinline)) int
+foo ()
+{
+ return bar ();
+}
+
+__attribute__ ((noinline)) int
+test_func ()
+{
+ return foo ();
+}
+
+int
+main ()
+{
+ global = test_func ();
+ return (global * 2);
+}
diff --git a/gdb/testsuite/gdb.opt/inline-frame-tailcall.exp b/gdb/testsuite/gdb.opt/inline-frame-tailcall.exp
new file mode 100644
index 00000000000..bac96835d12
--- /dev/null
+++ b/gdb/testsuite/gdb.opt/inline-frame-tailcall.exp
@@ -0,0 +1,47 @@
+# 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/>.
+
+# Tests a specific combination, a tailcall into a function, which then
+# has another function inlined within it. So:
+#
+# main --> test_func --> foo --> bar
+#
+# main makes a normal call to test_func.
+#
+# test_func makes a tail call to foo.
+#
+# bar is inlined within foo.
+#
+# We should still see test_func in the call stack.
+
+standard_testfile
+
+if { [prepare_for_testing "failed to prepare" ${testfile} \
+ [list $srcfile] {debug optimize=-O2}] } {
+ return -1
+}
+
+if ![runto_main] {
+ return -1
+}
+
+gdb_breakpoint "bar"
+gdb_continue_to_breakpoint "bar"
+
+gdb_test "bt" \
+ [multi_line "#0 bar \\(\\).*" \
+ "#1 foo \\(\\).*" \
+ "#2 $hex in test_func \\(\\).*" \
+ "#3 $hex in main \\(\\).*" ]
More information about the Gdb-patches
mailing list