Author Topic: unnecessory locker in CClogger  (Read 10217 times)

Offline ollydbg

  • Developer
  • Lives here!
  • *****
  • Posts: 5247
  • OpenCV and Robotics
    • Chinese OpenCV forum moderator
unnecessory locker in CClogger
« on: June 19, 2013, 09:59:37 am »
Hi, when reading the code:
Code: [Select]
void CCLogger::AddToken(const wxString& msg)
{
    wxMutexLocker l(m_CCLoggerMutex);

    if (!m_Parent || m_AddTokenId<1) return;

    wxCommandEvent evt(wxEVT_COMMAND_MENU_SELECTED, m_AddTokenId);
    evt.SetString(msg);
#if CC_PROCESS_LOG_EVENT_TO_PARENT
    m_Parent->ProcessEvent(evt);
#else
    wxPostEvent(m_Parent, evt);
#endif
}

void CCLogger::Log(const wxString& msg)
{
    wxMutexLocker l(m_CCLoggerMutex);

    if (!m_Parent || m_LogId<1) return;

    wxCommandEvent evt(wxEVT_COMMAND_MENU_SELECTED, m_LogId);
    evt.SetString(msg);
#if CC_PROCESS_LOG_EVENT_TO_PARENT
    m_Parent->ProcessEvent(evt);
#else
    wxPostEvent(m_Parent, evt);
#endif
}

void CCLogger::DebugLog(const wxString& msg)
{
    wxMutexLocker l(m_CCLoggerMutex);

    if (!m_Parent || m_DebugLogId<1) return;

    wxCommandEvent evt(wxEVT_COMMAND_MENU_SELECTED, m_DebugLogId);
    evt.SetString(msg);
#if CC_PROCESS_LOG_EVENT_TO_PARENT
    m_Parent->ProcessEvent(evt);
#else
    wxPostEvent(m_Parent, evt);
#endif
}

Here, in the case of CC_PROCESS_LOG_EVENT_TO_PARENT is 0.

Refer to the link below, I think no wxMutexLocker should be used, because wxPostEvent is safe to send an event from a working thread to main GUI thread.

http://wiki.wxwidgets.org/Custom_Events#wxEvtHandler::ProcessEvent_vs._wxPostEvent_and_wxEvtHandler::AddPendingEvent


Any ideas/comments ?
If some piece of memory should be reused, turn them to variables (or const variables).
If some piece of operations should be reused, turn them to functions.
If they happened together, then turn them to classes.

Offline dmoore

  • Developer
  • Lives here!
  • *****
  • Posts: 1576
Re: unnecessory locker in CClogger
« Reply #1 on: June 19, 2013, 08:54:24 pm »
One issue (i.e. crash) I've had with wxPostEvent (or AddPendingEvent) in the past is passing wxString between threads in a wxEvent object. Any time I passed a string, I had to make a copy of it with c_str() to avoid issues. I don't know if this is what the Mutex Locker in that code was trying to solve.
« Last Edit: June 20, 2013, 04:28:34 am by dmoore »

Offline ollydbg

  • Developer
  • Lives here!
  • *****
  • Posts: 5247
  • OpenCV and Robotics
    • Chinese OpenCV forum moderator
Re: unnecessory locker in CClogger
« Reply #2 on: June 20, 2013, 03:51:26 am »
Hi, dmoore, you are definitely right. I just debug a simple wx multiply thread app which link to the debug wx 2.8.12 library , so I can step into all the wx sources. I see that "wxPostEvent(m_Parent, evt);" will clone a new event:
Code: [Select]
virtual wxEvent *Clone() const { return new wxCommandEvent(*this); }
Code: [Select]
   wxCommandEvent(const wxCommandEvent& event)
        : wxEvent(event),
#if WXWIN_COMPATIBILITY_2_4
          m_commandString(this),
#endif
          m_cmdString(event.m_cmdString),
          m_commandInt(event.m_commandInt),
          m_extraLong(event.m_extraLong),
          m_clientData(event.m_clientData),
          m_clientObject(event.m_clientObject)
        { }
But the bad thing is here:
Code: [Select]
m_cmdString(event.m_cmdString)A wxString is construct from another wxString, so reference counting is used. Too bad, wxString's reference counting is not atomic variables (see my post several months ago wxString and the locker issue in our CC code)

Here, in the code:
Code: [Select]
void CCLogger::AddToken(const wxString& msg)
{
    wxMutexLocker l(m_CCLoggerMutex);

    if (!m_Parent || m_AddTokenId<1) return;

    wxCommandEvent evt(wxEVT_COMMAND_MENU_SELECTED, m_AddTokenId);
    evt.SetString(msg);
#if CC_PROCESS_LOG_EVENT_TO_PARENT
    m_Parent->ProcessEvent(evt);
#else
    wxPostEvent(m_Parent, evt);
#endif
}
I think the wxMutexLocker can NOT solve the thread-safe issue of wxString, as you said, here we must do a deep copy of wxString by ourselves, the only way use: evt.SetString(msg.c_str()); (See this post as in wx forum as a reference


EDIT:
wx since 2.9.0 has a nice event type: wxThreadEvent, which did a deep copy of every thing (in-fact, it did a deep copy of it's wxString)
see event.h in wx 2.9.0, there is also a "SetString(GetString().c_str());"
Code: [Select]
// Thread event

class WXDLLIMPEXP_BASE wxThreadEvent : public wxEvent,
                                       public wxEventAnyPayloadMixin
{
public:
    wxThreadEvent(wxEventType eventType = wxEVT_THREAD, int id = wxID_ANY)
        : wxEvent(id, eventType)
        { }

    wxThreadEvent(const wxThreadEvent& event)
        : wxEvent(event),
          wxEventAnyPayloadMixin(event)
    {
        // make sure our string member (which uses COW, aka refcounting) is not
        // shared by other wxString instances:
        SetString(GetString().c_str());
    }

    virtual wxEvent *Clone() const
    {
        return new wxThreadEvent(*this);
    }

    // this is important to avoid that calling wxEventLoopBase::YieldFor thread events
    // gets processed when this is unwanted:
    virtual wxEventCategory GetEventCategory() const
        { return wxEVT_CATEGORY_THREAD; }

private:
    DECLARE_DYNAMIC_CLASS_NO_ASSIGN(wxThreadEvent)
};

EDIT2:
In wx2.8.12, when build with GCC, the wxString is COW(copy on write, and reference counted), but the reference counter is not thread-safe(it does not use an atomic counter).
In wx2.9.x, when wxString build with GCC use GCC's native std::string, which is also COW, and it uses the atomic reference counter, so it is safe to use "evt.SetString(msg);".
But COW with atomic reference counter has performance issue(reported by Herb Sutter in his article Optimizations That Aren't (In a Multithreaded World)), so I see somewhere that GCC 5.0 or GCC 4.9 in the future may totally remove the COW feature of std::string. I once hear that Visual C++'s std::string is already NOT COW for years.

« Last Edit: June 20, 2013, 04:33:06 am by ollydbg »
If some piece of memory should be reused, turn them to variables (or const variables).
If some piece of operations should be reused, turn them to functions.
If they happened together, then turn them to classes.

Offline golgepapaz

  • Multiple posting newcomer
  • *
  • Posts: 44
Re: unnecessory locker in CClogger
« Reply #3 on: June 20, 2013, 07:24:47 am »

EDIT2:
In wx2.8.12, when build with GCC, the wxString is COW(copy on write, and reference counted), but the reference counter is not thread-safe(it does not use an atomic counter).
In wx2.9.x, when wxString build with GCC use GCC's native std::string, which is also COW, and it uses the atomic reference counter, so it is safe to use "evt.SetString(msg);".
But COW with atomic reference counter has performance issue(reported by Herb Sutter in his article Optimizations That Aren't (In a Multithreaded World)), so I see somewhere that GCC 5.0 or GCC 4.9 in the future may totally remove the COW feature of std::string. I once hear that Visual C++'s std::string is already NOT COW for years.


 Heh I've burned because of that issue too. Mind you c++11 does not allow COW semantics for strings. But  gcc still uses COW strings in 4.8.0 even with the -std=c++11 flag.Likely ABI change is the factor for delaying the transition here.
 

Offline ollydbg

  • Developer
  • Lives here!
  • *****
  • Posts: 5247
  • OpenCV and Robotics
    • Chinese OpenCV forum moderator
Re: unnecessory locker in CClogger
« Reply #4 on: June 20, 2013, 10:04:39 am »
I believe the using "c_str()" still have the thread safety issue, see my post in thread safety issue about wxCommandEvent in wx2.8.12
If some piece of memory should be reused, turn them to variables (or const variables).
If some piece of operations should be reused, turn them to functions.
If they happened together, then turn them to classes.

Offline ollydbg

  • Developer
  • Lives here!
  • *****
  • Posts: 5247
  • OpenCV and Robotics
    • Chinese OpenCV forum moderator
Re: unnecessory locker in CClogger
« Reply #5 on: June 20, 2013, 05:12:56 pm »
Hi, dmoore, I'm trying to explain that using c_str before the calling of wxPostEvent() is not safe, instead we should derive a new event class, and override it's Clone() function. See my detailed explanation in this post: Re: thread safety issue about wxCommandEvent in wx2.8.12
If some piece of memory should be reused, turn them to variables (or const variables).
If some piece of operations should be reused, turn them to functions.
If they happened together, then turn them to classes.

Offline dmoore

  • Developer
  • Lives here!
  • *****
  • Posts: 1576
Re: unnecessory locker in CClogger
« Reply #6 on: June 21, 2013, 03:47:03 am »
Hi, dmoore, I'm trying to explain that using c_str before the calling of wxPostEvent() is not safe, instead we should derive a new event class, and override it's Clone() function. See my detailed explanation in this post: Re: thread safety issue about wxCommandEvent in wx2.8.12

I didn't look closely at what you wrote, but I agree in principle that c_str() is not enough to make the operation safe because we are still making a new wxString on worker thread and accessing it on the GUI thread and can't guarantee the wxString is (still) valid when it gets accessed on the GUI. It does seem to be stable at least some of the time...

Curious to see what you come up with as a solution

Offline ollydbg

  • Developer
  • Lives here!
  • *****
  • Posts: 5247
  • OpenCV and Robotics
    • Chinese OpenCV forum moderator
Re: unnecessory locker in CClogger
« Reply #7 on: September 11, 2013, 11:37:57 am »
FYI: I commit the fix in rev9311.
If some piece of memory should be reused, turn them to variables (or const variables).
If some piece of operations should be reused, turn them to functions.
If they happened together, then turn them to classes.

Offline dmoore

  • Developer
  • Lives here!
  • *****
  • Posts: 1576
Re: unnecessory locker in CClogger
« Reply #8 on: September 11, 2013, 03:25:06 pm »
FYI: I commit the fix in rev9311.

Nice. I plan to migrate to your new CodeBlocksThreadEvent class with the FileManager plugin (for reasons unknown to me, it occasionally causes C::B to crash and I have long suspected that the cause is related to passing event data across threads).

Offline killerbot

  • Administrator
  • Lives here!
  • *****
  • Posts: 5193
Re: unnecessory locker in CClogger
« Reply #9 on: September 11, 2013, 09:27:54 pm »
and this plugin causes on OPenSuse to have CB not shutdown when it is started from app menu, only when started from a console terminal CB can shutdown. Maybe it is related ...

Offline dmoore

  • Developer
  • Lives here!
  • *****
  • Posts: 1576
Re: unnecessory locker in CClogger
« Reply #10 on: September 11, 2013, 11:06:11 pm »
and this plugin causes on OPenSuse to have CB not shutdown when it is started from app menu, only when started from a console terminal CB can shutdown. Maybe it is related ...

Good to know. This could be related to using a unix pipe to communicate between threads. The was another issue related to GAMIN that jens solved.