[PATCH] Fix race in DAP startup

[PATCH] Fix race in DAP startup

[Tom Tromey]

Internal AdaCore DAP testing on Windows has had occasional failures
that show:

    assert threading.current_thread() is _dap_thread

I think this is a race in DAP startup: the _dap_thread global is only
set on return from start_thread, but it seems possible that the thread
itself could already run and encounter a @in_dap_thread decorator.

This patch fixes the problem by setting the global before running any
of the code in the new thread.  This also lets us remove a FIXME.
---
 gdb/python/lib/gdb/dap/server.py  |  4 +---
 gdb/python/lib/gdb/dap/startup.py | 12 +++++++++---
 2 files changed, 10 insertions(+), 6 deletions(-)

diff --git a/gdb/python/lib/gdb/dap/server.py b/gdb/python/lib/gdb/dap/server.py
index 92b4eee1c5e..ff88282049f 100644
--- a/gdb/python/lib/gdb/dap/server.py
+++ b/gdb/python/lib/gdb/dap/server.py
@@ -100,9 +100,7 @@ class Server:
         log("WROTE: <<<" + json.dumps(obj) + ">>>")
         self.write_queue.put(obj)
 
-    # This must be run in the DAP thread, but we can't use
-    # @in_dap_thread here because the global isn't set until after
-    # this starts running.  FIXME.
+    @in_dap_thread
     def main_loop(self):
         """The main loop of the DAP server."""
         # Before looping, start the thread that writes JSON to the
diff --git a/gdb/python/lib/gdb/dap/startup.py b/gdb/python/lib/gdb/dap/startup.py
index 22246a937ce..aa834cdb14c 100644
--- a/gdb/python/lib/gdb/dap/startup.py
+++ b/gdb/python/lib/gdb/dap/startup.py
@@ -59,14 +59,20 @@ def start_thread(name, target, args=()):
     with blocked_signals():
         result = threading.Thread(target=target, args=args, daemon=True)
         result.start()
-        return result
 
 
 def start_dap(target):
     """Start the DAP thread and invoke TARGET there."""
-    global _dap_thread
     exec_and_log("set breakpoint pending on")
-    _dap_thread = start_thread("DAP", target)
+
+    # Functions in this thread contain assertions that check for this
+    # global, so we must set it before letting these functions run.
+    def really_start_dap():
+        global _dap_thread
+        _dap_thread = threading.current_thread()
+        target()
+
+    start_thread("DAP", really_start_dap)
 
 
 def in_gdb_thread(func):
-- 
2.39.1

Re: [PATCH] Fix race in DAP startup

[Tom de Vries]

On 3/22/23 15:34, Tom Tromey via Gdb-patches wrote:
> Internal AdaCore DAP testing on Windows has had occasional failures
> that show:
> 
>      assert threading.current_thread() is _dap_thread
> 
> I think this is a race in DAP startup: the _dap_thread global is only
> set on return from start_thread, but it seems possible that the thread
> itself could already run and encounter a @in_dap_thread decorator.
> 
> This patch fixes the problem by setting the global before running any
> of the code in the new thread.  This also lets us remove a FIXME.

Hi,

I tried this patch out, hoping it would fix the stability issues I have 
with the gdb.dap tests, but I'm still seeing those, so I filed a PR: 
https://sourceware.org/bugzilla/show_bug.cgi?id=30268 .

Thanks,
- Tom

> ---
>   gdb/python/lib/gdb/dap/server.py  |  4 +---
>   gdb/python/lib/gdb/dap/startup.py | 12 +++++++++---
>   2 files changed, 10 insertions(+), 6 deletions(-)
> 
> diff --git a/gdb/python/lib/gdb/dap/server.py b/gdb/python/lib/gdb/dap/server.py
> index 92b4eee1c5e..ff88282049f 100644
> --- a/gdb/python/lib/gdb/dap/server.py
> +++ b/gdb/python/lib/gdb/dap/server.py
> @@ -100,9 +100,7 @@ class Server:
>           log("WROTE: <<<" + json.dumps(obj) + ">>>")
>           self.write_queue.put(obj)
>   
> -    # This must be run in the DAP thread, but we can't use
> -    # @in_dap_thread here because the global isn't set until after
> -    # this starts running.  FIXME.
> +    @in_dap_thread
>       def main_loop(self):
>           """The main loop of the DAP server."""
>           # Before looping, start the thread that writes JSON to the
> diff --git a/gdb/python/lib/gdb/dap/startup.py b/gdb/python/lib/gdb/dap/startup.py
> index 22246a937ce..aa834cdb14c 100644
> --- a/gdb/python/lib/gdb/dap/startup.py
> +++ b/gdb/python/lib/gdb/dap/startup.py
> @@ -59,14 +59,20 @@ def start_thread(name, target, args=()):
>       with blocked_signals():
>           result = threading.Thread(target=target, args=args, daemon=True)
>           result.start()
> -        return result
>   
>   
>   def start_dap(target):
>       """Start the DAP thread and invoke TARGET there."""
> -    global _dap_thread
>       exec_and_log("set breakpoint pending on")
> -    _dap_thread = start_thread("DAP", target)
> +
> +    # Functions in this thread contain assertions that check for this
> +    # global, so we must set it before letting these functions run.
> +    def really_start_dap():
> +        global _dap_thread
> +        _dap_thread = threading.current_thread()
> +        target()
> +
> +    start_thread("DAP", really_start_dap)
>   
>   
>   def in_gdb_thread(func):

Re: [PATCH] Fix race in DAP startup

[Tom de Vries]

On 3/23/23 17:04, Tom de Vries wrote:
> On 3/22/23 15:34, Tom Tromey via Gdb-patches wrote:
>> Internal AdaCore DAP testing on Windows has had occasional failures
>> that show:
>>
>>      assert threading.current_thread() is _dap_thread
>>
>> I think this is a race in DAP startup: the _dap_thread global is only
>> set on return from start_thread, but it seems possible that the thread
>> itself could already run and encounter a @in_dap_thread decorator.
>>
>> This patch fixes the problem by setting the global before running any
>> of the code in the new thread.  This also lets us remove a FIXME.
> 
> Hi,
> 
> I tried this patch out, hoping it would fix the stability issues I have 
> with the gdb.dap tests, but I'm still seeing those, so I filed a PR: 
> https://sourceware.org/bugzilla/show_bug.cgi?id=30268 .
> 

Hi,

as you pointed out in the PR, that was due to me not rebuilding.

LGTM.

Thanks,
- Tom

> Thanks,
> - Tom
> 
>> ---
>>   gdb/python/lib/gdb/dap/server.py  |  4 +---
>>   gdb/python/lib/gdb/dap/startup.py | 12 +++++++++---
>>   2 files changed, 10 insertions(+), 6 deletions(-)
>>
>> diff --git a/gdb/python/lib/gdb/dap/server.py 
>> b/gdb/python/lib/gdb/dap/server.py
>> index 92b4eee1c5e..ff88282049f 100644
>> --- a/gdb/python/lib/gdb/dap/server.py
>> +++ b/gdb/python/lib/gdb/dap/server.py
>> @@ -100,9 +100,7 @@ class Server:
>>           log("WROTE: <<<" + json.dumps(obj) + ">>>")
>>           self.write_queue.put(obj)
>> -    # This must be run in the DAP thread, but we can't use
>> -    # @in_dap_thread here because the global isn't set until after
>> -    # this starts running.  FIXME.
>> +    @in_dap_thread
>>       def main_loop(self):
>>           """The main loop of the DAP server."""
>>           # Before looping, start the thread that writes JSON to the
>> diff --git a/gdb/python/lib/gdb/dap/startup.py 
>> b/gdb/python/lib/gdb/dap/startup.py
>> index 22246a937ce..aa834cdb14c 100644
>> --- a/gdb/python/lib/gdb/dap/startup.py
>> +++ b/gdb/python/lib/gdb/dap/startup.py
>> @@ -59,14 +59,20 @@ def start_thread(name, target, args=()):
>>       with blocked_signals():
>>           result = threading.Thread(target=target, args=args, 
>> daemon=True)
>>           result.start()
>> -        return result
>>   def start_dap(target):
>>       """Start the DAP thread and invoke TARGET there."""
>> -    global _dap_thread
>>       exec_and_log("set breakpoint pending on")
>> -    _dap_thread = start_thread("DAP", target)
>> +
>> +    # Functions in this thread contain assertions that check for this
>> +    # global, so we must set it before letting these functions run.
>> +    def really_start_dap():
>> +        global _dap_thread
>> +        _dap_thread = threading.current_thread()
>> +        target()
>> +
>> +    start_thread("DAP", really_start_dap)
>>   def in_gdb_thread(func):
> 

Re: [PATCH] Fix race in DAP startup

[Tom Tromey]

Tom> as you pointed out in the PR, that was due to me not rebuilding.

Tom> LGTM.

Thanks for trying it out.  I'm checking it in now.

Tom