Re: [PATCH 0/4] Modernize frame unwinders and add disable feature

[Guinevere Larsen <blarsen@redhat.com>]

On 11/03/2024 16:10, Luis Machado wrote:
> On 3/11/24 15:00, Guinevere Larsen wrote:
>> On 11/03/2024 15:56, Luis Machado wrote:
>>> Hi,
>>>
>>> On 3/6/24 12:51, Guinevere Larsen wrote:
>>>> This patch series started with me trying to make it easier to test GDB's
>>>> ability to unwind using CFI data, to improve a previous patch I sent to
>>>> the list. However, once I finished these changes, I realized there was
>>>> an unrelated bug I should fix before proposing the CFI test. Since these
>>>> changes are significant enough already, and I think would be interesting
>>>> on their own, I figured I shoudl submit this patch series as is right
>>>> now while I figure out the other bug.
>>>>
>>>> The first patch is just a minor change, storing frame unwinders in a
>>>> vector instead of through an unwinder table accessible using the
>>>> registry system. This isn't required (like I originally thought it was),
>>>> but it does make the whole system more readable in my opinion.
>>>>
>>>> Patch 3 has the real meat of the modernization, making GDB use
>>>> polymorphism to handle frame unwinders. This is slightly slower than
>>>> using function pointers, but much more readable in my opinion.
>>>>
>>>> As for the unwinder classes, they were chosen somewhat arbitrarily,
>>>> mostly based on where I found an unwinder and its name. I almost expect
>>>> some unwinders to be mis-categorized, but that should be easy to fix.
>>>>
>>>> The changes up to patch 3 have been tested with a try-branch, no
>>>> regressions as far as I could see.
>>>>
>>>> Guinevere Larsen (4):
>>>>     gdb: make gdbarch store a vector of frame unwinders
>>>>     gdb: add "unwinder class" to frame unwinders
>>>>     gdb: Migrate frame unwinders to use C++ classes
>>>>     GDB: introduce ability to disable frame unwinders
>>> I haven't gone through the series in detail, but I thought I'd give it a try on one of the
>>> aarch64 machines I have access to. I didn't look particularly healthy:
>>>
>>>
>>> # of unexpected core files      47
>>> # of expected passes            116521
>>> # of unexpected failures        581
>>> # of expected failures          77
>>> # of known failures             116
>>> # of untested testcases         128
>>> # of unresolved testcases       1102
>>> # of unsupported tests          458
>>> # of duplicate test names       10
>>>
>>> I see a number of internal errors going on. Mostly like these:
>>>
>>>           ../../../repos/binutils-gdb/gdbsupport/errors.cc:58
>>> 0xaaaad27e7877 check_ptrace_stopped_lwp_gone
>>>           ../../../repos/binutils-gdb/gdb/linux-nat.c:1634
>>> 0xaaaad27e7877 check_ptrace_stopped_lwp_gone
>>>           ../../../repos/binutils-gdb/gdb/linux-nat.c:1630
>>> 0xaaaad2ae0fa3 linux_resume_one_lwp
>>>
>> Oh no! Linaro CI had showed something was wrong, but I wasn't able to grab an aarch64 machine to test yet. I'll check it out when I can, thanks for narrowing it down for me
>>
> You're welcome. Do let me know if you can't easily find an aarch64 machine to try things on, and I can provide more info
> and take a deeper look.
>
I think I found the reason why aarch64 started failing so many cases. 
Can you try the following change and see if it works?

I'm not 100% sure because the machine I'm using seems to have some 
rather unstable results, but I think it should have solved the problem.

-- 
Cheers,
Guinevere Larsen
She/Her/Hers



diff --git a/gdb/tramp-frame.c b/gdb/tramp-frame.c
index 722d24ee129..a6f23e2b1b6 100644
--- a/gdb/tramp-frame.c
+++ b/gdb/tramp-frame.c
@@ -178,6 +178,9 @@ class frame_unwind_trampoline : public frame_unwind {
    {
      return prev_arch_p (this_frame, this_prologue_cache);
    }
+
+  void dealloc_cache (frame_info *self, void *this_cache) const override
+  { }
  };

  void