[Bug 49301] New: pthread_exit hangs on unwind through call_thread_exit_func on arm64

[WineHQ Bugzilla]

https://bugs.winehq.org/show_bug.cgi?id=49301

Bug ID: 49301
           Summary: pthread_exit hangs on unwind through
                    call_thread_exit_func on arm64
           Product: Wine
           Version: unspecified
          Hardware: aarch64
                OS: Linux
            Status: NEW
          Severity: normal
          Priority: P2
         Component: ntdll
          Assignee: [e-mail]
          Reporter: [e-mail]
      Distribution: ---

Created attachment 67309
  --> https://bugs.winehq.org/attachment.cgi?id=67309
Workaround

My arm64 Wine build spins in services.exe after starting up anything. The
backtrace points into pthread_exit() - see attached file backtrace.txt.

I placed pthread_exit calls throughout the thread creation / shutdown code and
it works up to the point where signal_exit_thread() calls
call_thread_exit_func(). After call_thread_exit_func() passes control to the C
function exit_thread() pthread_exit() will hang.

I previously had this problem only in a gcc wine build (with 8fb8cc03
reverted), but recently, presumably after an update to clang-10, the problem
also happens when building Wine with clang. The attached backtrace is from a
build with clang 10.

Calling exit_thread() directly from signal_exit_thread() without going through
call_thread_exit_func works around the problem. See attached file noasm.diff

[Bug 49301] pthread_exit hangs on unwind through call_thread_exit_func on arm64

[WineHQ Bugzilla]

https://bugs.winehq.org/show_bug.cgi?id=49301

--- Comment #1 from Stefan Dösinger  ---
Created attachment 67310
  --> https://bugs.winehq.org/attachment.cgi?id=67310
Backtrace

This file shows a backtrace of the hung thread. I did not notice this before,
but gdb also struggles with the asm code.

[Bug 49301] pthread_exit hangs on unwind through call_thread_exit_func on arm64

[WineHQ Bugzilla]

https://bugs.winehq.org/show_bug.cgi?id=49301

André H.  changed:

What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |[e-mail]

[Bug 49301] pthread_exit hangs on unwind through call_thread_exit_func on arm64

[WineHQ Bugzilla]

https://bugs.winehq.org/show_bug.cgi?id=49301

--- Comment #2 from Stefan Dösinger  ---
Created attachment 67312
  --> https://bugs.winehq.org/attachment.cgi?id=67312
Add CFI info to call_thread_exit_func

This patch makes threads exit properly for me. I think it goes in the right
direction, but I would be highly surprised if all the numbers in it are
correct.

It specifies the call frame setup done at the start of start_thread(), because
after the mov %sp points to the 'old' stack as set up in start_thread, not
whatever happened before call_thread_exit_func was called.

I am worried that gdb still can't unwind the backtrace through
call_thread_exit_func and start_thread. I would expect to see
signal_start_thread, thread.c:start_thread and then the pthread creation
functions below call_thread_exit_func.

[Bug 49301] pthread_exit hangs on unwind through call_thread_exit_func on arm64

[WineHQ Bugzilla]

https://bugs.winehq.org/show_bug.cgi?id=49301

François Gouget  changed:

What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |[e-mail]
           Keywords|                            |patch