[PATCH][gdb] Block SIGTERM in worker threads

[PATCH][gdb] Block SIGTERM in worker threads

[Tom de Vries]

Hi,

With gdb build with gcc-12 and -fsanitize=thread, and test-case
gdb.base/gdb-sigterm.exp, I run into:
...
WARNING: ThreadSanitizer: data race (pid=9722)^M
  Write of size 4 at 0x00000325bc68 by thread T1:^M
  #0 handle_sigterm(int) src/gdb/event-top.c:1211 (gdb+0x8ec01f)^M
  ...
  Previous read of size 4 at 0x00000325bc68 by main thread:^M
    [failed to restore the stack]^M
^M
  Location is global 'sync_quit_force_run' of size 4 at \
  0x00000325bc68 (gdb+0x325bc68)^M
  ...
SUMMARY: ThreadSanitizer: data race gdb/event-top.c:1211 in \
  handle_sigterm(int)^M
...
and 3 more data races involving handle_sigterm and locations:
- active_ext_lang
- quit_flag
- heap block of size 40
  (XNEW (async_signal_handler) in create_async_signal_handler)

This was reported in PR29297.

The testcase executes a "kill -TERM $gdb_pid", which generates a
process-directed signal.

A process-directed signal can be delivered to any thread, and what we see
here is the fallout of the signal being delivered to a worker thread rather
than the main thread.

Fix this by blocking SIGTERM in the worker threads.

[ I have not been able to reproduce this after it occurred for the first time,
so unfortunately I cannot confirm that the patch fixes the problem. ]

Tested on x86_64-linux, with and without -fsanitize=thread.

Any comments?

Thanks,
- Tom

[gdb] Block SIGTERM in worker threads

---
 gdbsupport/block-signals.h | 1 +
 1 file changed, 1 insertion(+)

diff --git a/gdbsupport/block-signals.h b/gdbsupport/block-signals.h
index cf97208c750..4703690a4ec 100644
--- a/gdbsupport/block-signals.h
+++ b/gdbsupport/block-signals.h
@@ -42,6 +42,7 @@ class block_signals
     sigaddset (&mask, SIGCHLD);
     sigaddset (&mask, SIGALRM);
     sigaddset (&mask, SIGWINCH);
+    sigaddset (&mask, SIGTERM);
     gdb_sigmask (SIG_BLOCK, &mask, &m_old_mask);
 #endif
   }

Re: [PATCH][gdb] Block SIGTERM in worker threads

[Pedro Alves]

On 2022-06-30 10:58, Tom de Vries wrote:
> Hi,
> 
> With gdb build with gcc-12 and -fsanitize=thread, and test-case
> gdb.base/gdb-sigterm.exp, I run into:
> ...
> WARNING: ThreadSanitizer: data race (pid=9722)^M
>   Write of size 4 at 0x00000325bc68 by thread T1:^M
>   #0 handle_sigterm(int) src/gdb/event-top.c:1211 (gdb+0x8ec01f)^M
>   ...
>   Previous read of size 4 at 0x00000325bc68 by main thread:^M
>     [failed to restore the stack]^M
> ^M
>   Location is global 'sync_quit_force_run' of size 4 at \
>   0x00000325bc68 (gdb+0x325bc68)^M
>   ...
> SUMMARY: ThreadSanitizer: data race gdb/event-top.c:1211 in \
>   handle_sigterm(int)^M
> ...
> and 3 more data races involving handle_sigterm and locations:
> - active_ext_lang
> - quit_flag
> - heap block of size 40
>   (XNEW (async_signal_handler) in create_async_signal_handler)
> 
> This was reported in PR29297.
> 
> The testcase executes a "kill -TERM $gdb_pid", which generates a
> process-directed signal.
> 
> A process-directed signal can be delivered to any thread, and what we see
> here is the fallout of the signal being delivered to a worker thread rather
> than the main thread.
> 
> Fix this by blocking SIGTERM in the worker threads.
> 
> [ I have not been able to reproduce this after it occurred for the first time,
> so unfortunately I cannot confirm that the patch fixes the problem. ]
> 
> Tested on x86_64-linux, with and without -fsanitize=thread.
> 
> Any comments?

OK.