After spending a bit of time with simple UWP applications with C++/CX and ++/WinRT I have begun to enjoy some of the features of targeting that environment for Windows UI app development.
Now having to go back to a more familiar MFC application development I want to change my approach to something that is similar to UWP app development. The idea is to use asynchronous C++11 threads to generate content and modify content that is displayed in the MFC UI.
The main change I want to make is to use C++11 threads to offload some time-consuming tasks and have those threads communicate the results back to the main MFC UI.
Some of the tasks that I am looking to offload onto C++11 threads, which are similar to what I would use with asynchronous tasks with C++/CX and C++/WinRT in UWP apps are:
connect to and exchange data with another computer
open a data file and parse it to update the UI view
convert a data file to another format such as CSV and export to a file
read a file in a format such as CSV and convert the content into a data file
perform searches and filtering of the presentation of the data file in the UI
The problem I am running into is similar to the problem described in Can I have multiple GUI threads in MFC?, however, I am looking for a general approach rather than the specific progress bar update in that question.
I have been trying a simple test with an experimental MFC app using the Visual Studio template which has a tree control docked on the left to build the tree in a worker thread.
If I have a CViewTree, an MFC window that displays a tree view, which I want to update from a C++11 thread, I am currently using ::PostMessage() to request that the tree control in the docked pane is updated.
If I use a lambda with a global std::thread such as the following code:
std::thread t1;
void CClassView::FillClassView()
{
// ::SendMessage() seems to deadlock so ::PostMessage() is required.
t1 = std::thread([this]() { Sleep(5000); ::PostMessage(this->m_hWnd, WM_COMMAND, ID_NEW_FOLDER, 0); });
}
the message handler for the MFC docked pane which looks like:
void CClassView::OnNewFolder()
{
t1.join(); // this join seems to deadlock if ::SendMessage() is used.
AddItemsToPane(m_wndClassView);
}
does indeed update the MFC docked pane with the tree control content just as if the function AddItemsToPane(m_wndClassView); were called at the same place where the C++11 thread is created. The pane update is delayed by 5 seconds when the C++11 thread is used just to provide a visible indication that the thread approach is actually working.
My problem is that I want the C++11 thread to create the content for the tree control and provide it to the docked pane rather than having the docked pane generate the content.
Thus far the only approach I can think of is to develop my own class library that will provide C++11 thread analogues to the MFC library and controls using ::PostMessage() to send the appropriate Windows messages to the designated MFC window or control.
I am wondering if it is possible to have the C++11 threads have their own, shadowing MFC control which they update and then send a message to the UI asking the UI to update its displayed control with the contents of the shadow MFC control? Or is there some other approach that people are using?
I am looking for some other, less arduous approaches to solving this problem of updating an MFC UI from C++11 threads.
By the way #1 ::SendMessage() appears to deadlock on the join() in CClassView::OnNewFolder() which I assume means that some kind of synchronization between the C+11 thread and the UI thread is blocking the C++11 thread from reaching its side of the join()?
Yes there is a deadlock as the thread waits for SendMessage() to return while the message handler is waiting at the join() for the thread to finish. According to the Windows Dev Center SendMessage function:
Sends the specified message to a window or windows. The SendMessage
function calls the window procedure for the specified window and does
not return until the window procedure has processed the message.
To send a message and return immediately, use the SendMessageCallback
or SendNotifyMessage function. To post a message to a thread's message
queue and return immediately, use the PostMessage or PostThreadMessage
function.
By the way #2 It would also seem that using the actual Window handle rather than the this pointer in the lambda for the C++11 thread would be safer. Just in case the this pointer becomes undefined for some reason such as the control is removed?
By the way #3 The Microsoft concurrency namespace, provided via #include <ppltasks.h>, is an alternative to C++11 threads. The concurrency namespace functions are at a higher level of abstraction than C++11 threads and are easier to use.
For instance the above use of std:thread could be rewritten as:
void CClassView::FillClassView()
{
concurrency::create_task([this]() { Sleep(5000); ::SendMessage(this->m_hWnd, WM_COMMAND, ID_NEW_FOLDER, 0); });
}
and this does not require the use of a std::thread join() to cleanly terminate the thread. Also SendMessage() or PostMessage() may be used to send the Windows message since we do not have the same deadlock issue as we have with C++11 threads.
Notes
Note #1: About Messages and Message Queues as well as Using Messages and Message Queues.
For MFC specific content see Messages and Commands in the Framework.
Note #2: Multithreading with C++ and MFC and specifically Multithreading: Programming Tips which says.
If you have a multithreaded application that creates a thread in a way
other than using a CWinThread object, you cannot access other MFC
objects from that thread. In other words, if you want to access any
MFC object from a secondary thread, you must create that thread with
one of the methods described in Multithreading: Creating
User-Interface Threads or Multithreading: Creating Worker Threads.
These methods are the only ones that allow the class library to
initialize the internal variables necessary to handle multithreaded
applications.
Note #3: UWP APIs callable from a classic desktop app which says:
With some notable exceptions, the general rule is that a Universal
Windows Platform (UWP) API can be called from a classic desktop app.
The two main areas of APIs that are an exception to this general rule
are XAML UI APIs, and APIs that require the calling app to have a
package identity. UWP apps have a well-defined app model, and they
have a package identity. Classic desktop apps do not have a
well-defined app model, and they do not have a package identity. A
classic desktop app that has been converted to a UWP app does have a
package identity.
See as well the following blogs from Sep 2012 about WinRT with VS 2012 and Windows 8. Though C++/WinRT with VS 2017 seems more appropriate for Windows 10 than Windows Runtime template Library (WRL) used:
Accessing WinRT From Desktop apps (Part 1)
Accessing WinRT from Desktop Apps (Part 2)
Note #4: MFC Desktop Applications which is a jumping off point with lots of links. See also MFC COM which is a jumping off point with lots of links about MFC with COM along with this article Introduction to COM. See as well MFC Macros and Globals.
As for using AfxGetMainWnd() to get the main application window, the Microsoft developer center has this to say in the article AfxGetMainWnd:
If AfxGetMainWnd is called from the application's primary thread, it
returns the application's main window according to the above rules. If
the function is called from a secondary thread in the application, the
function returns the main window associated with the thread that made
the call.
After some experimentation, there are a few recommendations that I feel comfortable making.
the concurrency task functionality is easier to use than C++11 std:thread and is more flexible in using with coroutines however std::async is easier to use than std::thread and works with co_await as well
coroutines using co_await look to be a great addition with concurrency when using C++/WinRT and the Async type functions in WinRT (see C++ Coroutines: Understanding operator co_await for a technical explanation)
you can make your own async functions using the concurrency::task<> template as the return type of the function or use concurrency::create_task() and you can use co_await with such a task
you can also use co_await with std::async() since std::async() returns a std::future<> which has an Awaitable interface (see await/yield: C++ coroutines though it is dated Nov-2016)
you can also use co_await with a std::future<> as provided by the get_future() method of a std::packaged_task<> ( see also What is the difference between packaged_task and async )
you can make generator functions using std::experimental::generator<type> as a function return type along with the co_yield operator to return a value of the specified type in the generated series
to update the MFC UI requires that any code runs in the MFC UI thread the MFC object was created so Windows messages are needed to communicate with the MFC windows and window class objects from other threads or you must switch the thread context/affinity to the UI thread context for that object
winrt::apartment_context can be used to capture the current thread context and later resumed using co_await which may be used to capture the main UI thread context to be reused later (see Programming with thread affinity in mind in the article Concurrency and asynchronous operations with C++/WinRT)
co_await winrt::resume_background(); can be used to push the current thread's context to a background thread which can be useful for a lengthy task that may be in the main UI thread context and you want to make sure that it is not
when sending messages to a window, make sure that the window actually has been created and exists, during application startup the application must create windows before you can use them; just because the MFC class exists does not mean the underlying window has been created yet
::SendMessage() is synchronous in which the message is sent and a response is returned
::PostMessage() is asynchronous in which the message is sent and a response is not returned
using ::PostMessage() be careful that pointers sent in messages do not go out of scope before the receiver uses them as there is typically a delay between when ::PostMessage() returns and the message handle receiving the message actually does something with the message
probably the most straightforward approach is to use the ON_MESSAGE() macro in the message map with a message handler interface of afx_msg LRESULT OnMessageThing(WPARAM, LPARAM)
you can use Windows message identifiers in the space beginning with the defined constant of WM_APP and the same identifier can be used in different classes
you can use much of C++/WinRT quite easily with MFC with just a bit of care though admittedly I only have tried a few things and there are some limitations such as not using XAML according to the documentation
if you do use C++/WinRT within an MFC application, you are limiting your application to versions of Windows that has the Windows Runtime which pretty much means Windows 10 (this rules out using C++/WinRT with Windows 7, POS Ready 7, etc.)
using C++/WinRT requires adding compiler option /stdc++17 to enable the ISO C++17 Standard for the C++ Language Standard and using coroutines requires the /await compiler option
Here are a view resources.
Microsoft Build 2018
Effective C++/WinRT for UWP and Win32
May 06, 2018 at 3:27PM by Brent Rector, Kenny Kerr
CppCon 2017: Scott Jones & Kenny Kerr
C++/WinRT and the Future of C++ on Windows
Published on Nov 2, 2017
Using Visual Studio 2017 Community edition, I created a new MFC Single Document Interface (SDI) project using the Visual Studio style. After the application comes up, it looks like the following image.
Helper functions for messages
The first change I made was to provide a way to send a Windows message to one of the panes (ClassView or OutputWindow) that I would like to update. Since the CMainFrame class in MainFrm.h had the MFC objects for these windows as in:
protected: // control bar embedded members
CMFCMenuBar m_wndMenuBar;
CMFCToolBar m_wndToolBar;
CMFCStatusBar m_wndStatusBar;
CMFCToolBarImages m_UserImages;
CFileView m_wndFileView;
CClassView m_wndClassView;
COutputWnd m_wndOutput;
CPropertiesWnd m_wndProperties;
I modified the class to provide a way to send a message to these windows. I chose to use SendMessage() rather than PostMessage() to eliminate the pointer going out of scope issue. The concurrency class works fine with SendMessage().
LRESULT SendMessageToFileView(UINT msgId, WPARAM wParam, LPARAM lParam) { return m_wndFileView.SendMessage(msgId, wParam, lParam); }
LRESULT SendMessageToClassView(UINT msgId, WPARAM wParam, LPARAM lParam) { return m_wndClassView.SendMessage(msgId, wParam, lParam); }
LRESULT SendMessageToOutputWnd(UINT msgId, WPARAM wParam, LPARAM lParam) { return m_wndOutput.SendMessage(msgId, wParam, lParam); }
These are the raw, bare infrastructure for sending the messages to update the various MFC windows. I put these into CMainFrame class as that is a central point and the AfxGetMainWnd() function allows me to access the object of that class at any place within the MFC application. Additional classes to wrap these raw functions would be appropriate.
I then put message handlers into each of the classes within the BEGIN_MESSAGE_MAP and the END_MESSAGE_MAP macros. The output window update was the easiest and simplest and look like:
BEGIN_MESSAGE_MAP(COutputWnd, CDockablePane)
ON_WM_CREATE()
ON_WM_SIZE()
// ADD_ON: message handler for the WM_APP message containing an index as
// to which output window to write to along with a pointer to the
// text string to write.
// this->SendMessageToOutputWnd(WM_APP, COutputWnd::OutputBuild, (LPARAM)_T("some text"));
ON_MESSAGE(WM_APP, OnAddItemsToPane)
END_MESSAGE_MAP()
with the message handler looking like:
// ADD_ON: message handler for the WM_APP message containing an array of the
// struct ItemToInsert above. Uses method AddItemsToPane().
LRESULT COutputWnd::OnAddItemsToPane(WPARAM wParam, LPARAM lParam)
{
switch (wParam) {
case OutputBuild:
m_wndOutputBuild.AddString((TCHAR *)lParam);
break;
case OutputDebug:
m_wndOutputDebug.AddString((TCHAR *)lParam);
break;
case OutputFind:
m_wndOutputFind.AddString((TCHAR *)lParam);
break;
}
return 0;
}
I added the method prototype to the class along with this enumeration to make the functionality a bit easier to use.
enum WindowList { OutputBuild = 1, OutputDebug = 2, OutputFind = 3 };
With the above changes I was able to insert into the message handler for "New" in BOOL CMFCAppWinRTDoc::OnNewDocument() the following code to put a text string into the "Build" Output Window:
CMainFrame *p = dynamic_cast <CMainFrame *> (AfxGetMainWnd());
if (p) {
p->SendMessageToOutputWnd(WM_APP, COutputWnd::OutputBuild, (LPARAM)_T("this is a test from OnNewDocument()."));
}
Using C++/WinRT with MFC and concurrency
To test this along with testing using C++/WinRT with MFC, I added the following concurrency task to CMainFrame::OnCreate() which is called when the application is starting up. This source spins off a task which then uses the Syndication functionality of C++/WinRT to fetch an RSS feed list and displays the titles in the OutputWindow pane labeled "Build" as seen in the screen shot above.
concurrency::create_task([this]() {
winrt::init_apartment();
Sleep(5000);
winrt::Windows::Foundation::Uri uri(L"http://kennykerr.ca/feed");
winrt::Windows::Web::Syndication::SyndicationClient client;
winrt::Windows::Web::Syndication::SyndicationFeed feed = client.RetrieveFeedAsync(uri).get();
for (winrt::Windows::Web::Syndication::SyndicationItem item : feed.Items())
{
winrt::hstring title = item.Title().Text();
this->SendMessageToOutputWnd(WM_APP, COutputWnd::OutputBuild, (LPARAM)title.c_str()); // print a string to an output window in the output pane.
}
winrt::uninit_apartment();
});
To use the concurrency and the C++/WinRT functionality I had to add a couple of include files near the top of MainFrm.c source file.
// ADD_ON: include files for using the concurrency namespace.
#include <experimental\resumable>
#include <pplawait.h>
#pragma comment(lib, "windowsapp")
#include "winrt/Windows.Foundation.h"
#include "winrt/Windows.Web.Syndication.h"
In addition, I had to modify the Properties of the solution to specify C++17 and an additional compiler option of /await which are marked with blue arrows in the screen shot below.
Using co_await with MFC and C++/WinRT
From a helpful comment from #IInspectable I took a look at coroutines with Visual Studio 2017 and MFC. I have been curious about them however it seemed that I could not come up with anything that would compile without errors from using co_await.
However starting with a link in the comment from #IInspectable I found a link to this YouTube video, CppCon 2016: Kenny Kerr & James McNellis “Putting Coroutines to Work with the Windows Runtime", which had a source code sample around time 10:28 that finally I was able to come up with something that would compile and work.
I created the following function which I then used to replace the above source with concurrency::create_task() and the lambda with a function call to the following function. The function call was simple, myTaskMain(this); replacing the concurrency::create_task([this]() { lambda in the int CMainFrame::OnCreate(LPCREATESTRUCT lpCreateStruct) method and then adding the following source code above the OnCreate() function body.
winrt::Windows::Foundation::IAsyncAction myTaskMain(CMainFrame *p)
{
winrt::Windows::Foundation::Uri uri(L"http://kennykerr.ca/feed");
winrt::Windows::Web::Syndication::SyndicationClient client;
winrt::Windows::Web::Syndication::SyndicationFeed feed = co_await client.RetrieveFeedAsync(uri);
Sleep(5000);
for (winrt::Windows::Web::Syndication::SyndicationItem item : feed.Items())
{
winrt::hstring title = item.Title().Text();
p->SendMessageToOutputWnd(WM_APP, COutputWnd::OutputBuild, (LPARAM)title.c_str()); // print a string to an output window in the output pane.
}
}
There are two changes that I made from the source of the concurrency::create_task() being replaced:
removed the winrt::init_apartment(); and winrt::uninit_apartment(); since using them triggered an exception and removing them seems to make no difference
moved the Sleep(5000); to after the co_await since leaving it where it was caused the function to sleep for 5 seconds which meant the UI thread slept for 5 seconds
What I found with the debugger was that at the point that the function myTaskMain() was called, there was an immediate return from the function and the UI thread continued running while the coroutine executed in the background. The UI displayed promptly and then some five seconds later the additional actions, updating the Class View tree and the RSS feed list in the "Build" tab of the Output Window happened.
Note #1: One other thing I have run into with other testing is that the UI will freeze for a few seconds (menu is non-functional). This seems to be due to the Sleep(5000); indicating that the code after the co_await is being run on the main UI thread. This change in application behavior started after I began to explore using winrt::apartment_context ui_thread; to capture the main UI thread context in order to then use co_await ui_thread; to return my coroutine thread to the main UI thread context.
What may be happening is that client.RetrieveFeedAsync(uri) is immediately being satisfied with no delay, perhaps from a cache, so rather than pushing the task to another thread and then returning to the caller, the co_await is getting an immediate result back and the function myTaskMain() is able to immediately continue using the current thread which is the main UI thread?
I noticed that in Visual Studio 2017 the co_await used with client.RetrieveFeedAsync(uri) is colored green while the co_await used with co_await ui_thread; is blue. Doing a mouse hover over the green colored co_await I get a tool tip indicating this is a different version of co_await.
Note #2: There is a C++/WinRT function to move to a background thread context, winrt::resume_background() which can be used with co_await. If I modify the above function myTaskMain() by replacing the line of code of Sleep(5000); after the call to client.RetrieveFeedAsync(uri) with the following two lines of code to push the thread context to a background thread, I do not see the freeze (the UI is responsive to menu selections) and the RSS feed text lines are displayed in the "Build" tab of the Output Window after some 15 seconds.
co_await winrt::resume_background(); // switch context to background thread
Sleep(15000);
Rolling an async task using concurrency::task<> that works with co_await
One thing I was curious about was being able to create my own async task that I could used with co_await similar to the async type functions of C++/WinRT. I spent some time searching about until I finally found this article, Concurrency and asynchronous operations with C++/WinRT, with a section called Asychronously return a non-Windows-Runtime type.
Here is a simple demo function that creates a concurrency::task<> with a lambda and returns the task which is then used with co_await. This particular lambda is returning an int so the function is defined as a task that returns an int, concurrency::task<int>
concurrency::task<int> mySleepTaskAsync()
{
return concurrency::create_task([]() {
Sleep(15000);
return 5;
});
}
The above function is then used with the co_await operator in a statement such as:
int jj = co_await mySleepTaskAsync();
which will cause the variable jj to have a value of 5 after a wait of 15 seconds.
The above is used in a function that returns a winrt::Windows::Foundation::IAsyncAction such as the function myTaskMain() above.
If you like you can also just directly use the lambda with the co_await as in:
int jj = co_await concurrency::create_task([]() {
Sleep(15000);
return 5;
});
Or you could have a normal function such as:
int mySleepTaskAsyncInt()
{
Sleep(15000);
return 5;
}
and then use it with a co_await using concurrency::task<> as in:
int jj = co_await concurrency::create_task(mySleepTaskAsyncInt);
Rolling an async task using std::async that works with co_await
While std::thread does not work with co_await, causing a compilation error, you can use std::async with co_await. The reason is the kind of return value that co_await operator requires and the difference in the return value of std::thread, a std::thread, and the return value of std::async, a std::future<>.
The co_await operator requires that the variable it is operating on is a std::future<>, has a get() method to retrieve a result from the thread, and is Awaitable.
#include <future>
int mySleepTaskAsyncInt()
{
Sleep(7000);
return 5;
}
winrt::Windows::Foundation::IAsyncAction myTaskMain(CMainFrame *p)
{
auto t1 = co_await std::async (std::launch::async, mySleepTaskAsyncInt);
// do something with the variable t1
}
Rolling an async task with std::packaged_task<> and std::future<> with co_await
As the co_await requires an Awaitable object, another way to create such an object is to create a task with std::packaged_task<> then start the task and use the get_future() method of the task to obtain a std::future<> which is then usable with co_await.
For instance we can have the following simple function that will create a task package, start the task executing, and then return a std::future<>. We can then use this function as the target for the co_await operator to implement a coroutine.
#include <future>
std::future<int> mySleepTaskStdFutureInt()
{
// create the task to prepare it for running.
std::packaged_task<int()> task([]() {
Sleep(7000);
return 455; // return an int value
});
// start the task running and return the future
return task(), task.get_future();
}
and then in our source code we would use this function similar to:
int jkjk = co_await mySleepTaskStdFutureInt();
The return statement is using the comma operator to introduce a sequence point so that we start the task running and then call the get_future() method on the running task. The result of the get_future() method, a std::future<int> is what is actually returned by the function.
The task that is created with std::packaged_task() must be started with a function like call using the variable. If you do not start the task then the std::future<> returned by the function will never have a variable and the co_await that is waiting for the Awaitable to complete and to provide a value will never fire. The result is that the source after your co_await will not be executed because the co_await will never be triggered.
Generator with co_yield and std::experimental::generator<type>
While investigating co_await I came across co_yield which is used to return a value as part of a generator of a set of values. With Visual Studio 2017 using co_yield requires that the header file experimental/generator be included. Here is a simple example of a generator that generates a series of integers.
#include <experimental/generator>
std::experimental::generator<int> makeSomeInts(int kCount)
{
for (int i = 0; i < kCount; i++) {
co_yield i;
}
}
And this function can be used with a ranged for as in:
for (int kkk : makeSomeInts(10)) {
// code that uses the variable kkk which contains
// an int from the generated range 0 up to be not including 10.
}
The above loop will be executed for each integer value of 0 through and including 9.
More complex message: updating ClassView pan
I also did an experiment with the ClassView tree control to provide a simple way of doing the most elementary actions: create an initial tree control and add to it.
In CClassView class in the file ClassView.h, I added the following data structs. By the way after I was done, I realized this was probably the wrong place to put this since the CFileView class uses the same tree structure so the same approach would work for both of these panes. Anyway, I added the following:
// ADD_ON: enumeration listing the various types of tree control icons which
// correspond to the position of a control in the tree.
// choose either classview_hc.bmp or classview.bmp for the bitmap strip that
// contains the 7 icons we are using for the images in our tree control.
// icons are standard size in height and width (15x15 pixels) in the order of:
// - main root icon
// - tree node icon which can be opened to show nodes beneath it
// - folder icon which is used to indicate a folder
// - method icon indicating a method of a class
// - locked method icon
// - member variable icon
// - locked member variable icon
enum IconList { MainRoot = 0, TreeNode = 1, FolderNode = 2, MethodNode = 3, MethodLockedNode = 4, MemberNode = 5, MemberLockedNode = 6 };
// ADD_ON: struct used to contain the necessary data for a node in the tree control.
struct ItemToInsert {
std::wstring label; // text to be displayed with the node.
int nImage; // zero based offset of the node's icon in the image, one of enum IconList above.
int nSelectedImage; // zero based offset of the node's icon in the image, one of enum IconList above.
};
I created a message handler, and added it to the message map in ClassView.cpp
ON_MESSAGE(WM_APP, OnAddItemsToPane)
and added the actual message handler itself along with a helper function that does the actual processing.
// ADD_ON: function for filling in the ClassView pane using an array of the
// struct ItemToInsert above. array is terminated by an entry with
// all zeros as in { _T(""), 0, 0 }
void CClassView::AddItemsToPane(CViewTree &xwndClassView, void *xrayp)
{
if (xrayp == 0) return;
// the images are icons that are laid out in a line of icons within a single bitmap image.
// see class method OnChangeVisualStyle() for when the bitmap image is loaded and then
// divided up into sections, 0 through 6, of the single bitmap image loaded.
// see classview.bmp and classview_hc.bmp in the ResourceFiles list.
HTREEITEM hRoot = xwndClassView.GetRootItem();
HTREEITEM hClass = 0;
ItemToInsert *xray = (ItemToInsert *)xrayp;
for (int i = 0; xray[i].label.size() != 0; i++) {
switch (xray[i].nImage) {
case MainRoot:
hRoot = xwndClassView.InsertItem(xray[i].label.c_str(), xray[i].nImage, xray[i].nSelectedImage);
xwndClassView.SetItemState(hRoot, TVIS_BOLD, TVIS_BOLD);
xwndClassView.Expand(hRoot, TVE_EXPAND);
break;
case TreeNode:
hClass = xwndClassView.InsertItem(xray[i].label.c_str(), xray[i].nImage, xray[i].nSelectedImage, hRoot);
break;
case FolderNode:
hClass = xwndClassView.InsertItem(xray[i].label.c_str(), xray[i].nImage, xray[i].nSelectedImage, hRoot);
break;
case MethodNode:
case MethodLockedNode:
case MemberNode:
case MemberLockedNode:
xwndClassView.InsertItem(xray[i].label.c_str(), xray[i].nImage, xray[i].nSelectedImage, hClass);
break;
default:
break;
}
}
}
// ADD_ON: message handler for the WM_APP message containing an array of the
// struct ItemToInsert above. Uses method AddItemsToPane().
LRESULT CClassView::OnAddItemsToPane(WPARAM wParam, LPARAM lParam)
{
switch (wParam) {
case 1:
AddItemsToPane(m_wndClassView, (void *)lParam);
break;
}
return 0;
}
I then created some sample data for an initial tree and then an add on node.
// ADD_ON: this is the content to be put into the ClassView tree pane.
// this is a tree structure.
CClassView::ItemToInsert xray2[] = {
{ _T("CFakeMainProject"), CClassView::MainRoot, CClassView::MainRoot },
{ _T("CFakeAboutDlg"), CClassView::TreeNode, CClassView::TreeNode },
{ _T("CFakeAboutDlg()"), CClassView::MethodNode, CClassView::MethodNode },
{ _T("CFakeApp"), CClassView::TreeNode, CClassView::TreeNode },
{ _T("CFakeApp()"), CClassView::MethodNode, CClassView::MethodNode },
{ _T("InitInstance()"), CClassView::MethodNode, CClassView::MethodNode },
{ _T("OnAppAbout()"), CClassView::MethodNode, CClassView::MethodNode },
{ _T("CFakeAppDoc"), CClassView::TreeNode, CClassView::TreeNode },
{ _T("CFakeAppDoc()"), CClassView::MethodLockedNode, CClassView::MethodLockedNode },
{ _T("~CFakeAppDoc()"), CClassView::MethodNode, CClassView::MethodNode },
{ _T("OnNewDocument()"), CClassView::MethodNode, CClassView::MethodNode },
{ _T("CFakeAppView"), CClassView::TreeNode, CClassView::TreeNode },
{ _T("CFakeAppView()"), CClassView::MethodLockedNode, CClassView::MethodLockedNode },
{ _T("~CFakeAppView()"), CClassView::MethodNode, CClassView::MethodNode },
{ _T("GetDocument()"), CClassView::MethodNode, CClassView::MethodNode },
{ _T("CFakeAppFrame"), CClassView::TreeNode, CClassView::TreeNode },
{ _T("CFakeAppFrame()"), CClassView::MethodNode, CClassView::MethodNode },
{ _T("~CFakeAppFrame()"), CClassView::MethodNode, CClassView::MethodNode },
{ _T("m_wndMenuBar"), CClassView::MemberLockedNode, CClassView::MemberLockedNode },
{ _T("m_wndToolBar"), CClassView::MemberLockedNode, CClassView::MemberLockedNode },
{ _T("m_wndStatusBar"), CClassView::MemberLockedNode, CClassView::MemberLockedNode },
{ _T("Globals"), CClassView::FolderNode, CClassView::FolderNode },
{ _T("theFakeApp"), CClassView::MemberNode, CClassView::MemberNode },
{ _T(""), 0, 0 }
};
CClassView::ItemToInsert xray3[] = {
{ _T("CAdditionalDelay"), CClassView::TreeNode, CClassView::TreeNode },
{ _T("CAdditionalDelayMethod()"), CClassView::MethodNode, CClassView::MethodNode },
{ _T(""), 0, 0 }
};
I then exercises this message handler by spinning off two concurrency tasks in the CMainFrame::OnCreate() method which did a time delay and then updated the ClassView window tree contents.
concurrency::create_task([this]() { Sleep(5000); this->SendMessageToClassView(WM_APP, 1, (LPARAM)xray2); });
concurrency::create_task([this]() { Sleep(10000); this->SendMessageToClassView(WM_APP, 1, (LPARAM)xray3); });
Maybe I mix too many different technologies together and run in some roadblock; some advise would be much appreciated.
I have an app which connects to several server; each connection with one input and output socket stream. The connection goes to a defined port and is close to telnet protocol. Text input/output. quite simple.
First I have an openStream function as wrapper called from main thread which create a client-specific GDC queue and dispatch the input/output-stream creation within that queue asynchronously:
gcdQueue = dispatch_queue_create([self.client.hostName cStringUsingEncoding:NSASCIIStringEncoding], DISPATCH_QUEUE_CONCURRENT);
// possible priorities:
// DISPATCH_QUEUE_PRIORITY_HIGH
// DISPATCH_QUEUE_PRIORITY_DEFAULT
// DISPATCH_QUEUE_PRIORITY_LOW
if (self.runASync)
{
// dispatch_async(dispatch_get_global_queue(DISPATCH_QUEUE_PRIORITY_LOW, 0), ^{
dispatch_async(gcdQueue, ^{
[self openStreamsInternal];
});
}
Code for the technical open of steams
...
//
// in openStreamsInternal()
//
CFStreamCreatePairWithSocketToHost(NULL, (__bridge CFStringRef)self.client.hostName, [self.client.hostPort intValue], &_readStream, &_writeStream);
self.inputStream = (__bridge_transfer NSInputStream *)_readStream;
self.outputStream = (__bridge_transfer NSOutputStream *)_writeStream;
[self.inputStream setDelegate:self];
[self.outputStream setDelegate:self];
self.runLoop = [NSRunLoop currentRunLoop];
[self.inputStream scheduleInRunLoop:self.runLoop forMode:NSDefaultRunLoopMode];
[self.outputStream scheduleInRunLoop:self.runLoop forMode:NSDefaultRunLoopMode];
[self.inputStream open];
[self.outputStream open];
// ... some lines later
if (self.runASync && (self.inputStream || self.outputStream))
{
[self.runLoop run];
}
I open both socket streams and link them to a runloop within the GCD-queue (assuming it will indirectly create a thread; not sure if that is always guaranteed).
Then via the delegate (a member function of my connection class) for the streams in
Code:
- (void)stream:(NSStream *)stream handleEvent:(NSStreamEvent)eventCode
I do what I have to do with in data flooding in. No issues until I stay in foreground. I close the streams when I go into background to release the resources.
Now with iOS 7 I want to enable background refresh for the streams. For that i don't close the streams anymore when moving into background and have the notification code in
Code:
- (void)application:(UIApplication *)application performFetchWithCompletionHandler:(void (^)(UIBackgroundFetchResult))completionHandler
{
NSLog(#"called in background for data fetch");
dispatch_queue_t queue = dispatch_get_global_queue(DISPATCH_QUEUE_PRIORITY_DEFAULT, 0);
dispatch_group_t group = dispatch_group_create();
for (Connection *connection in self.document.clientList)
{
// Add a task to the group
[connection parseResponseInQueue:group];
}
dispatch_group_wait(group, DISPATCH_TIME_FOREVER);
NSLog(#"finished with background for data fetch");
completionHandler(UIBackgroundFetchResultNewData);
}
This is one of my variants of background processing; not working well. This one supposed to wait a second and check if the input stream has data copied. If thats the case the parser would be called and the method comes to and end; removing one item from the dispatch group created in the iOS7 background app notification.
I don't like the dispatch_after as it seems very brute; but without I run in an endless loop as the streams seems not be triggered at all times.
Code:
- (void)parseResponseInQueue:(dispatch_group_t)group
{
if (gcdQueue != nil)
{
dispatch_group_async(group, gcdQueue, ^{
while ([self.data length] > 0)
{
dispatch_after(dispatch_time(DISPATCH_TIME_NOW, 1 * NSEC_PER_SEC), gcdQueue, ^{
NSLog(#"%#, wait for parser in background", self.client.hostName);
});
}
#if 0
if ([self.data length] > 0)
{
NSLog(#"%#, start working on buffer %d from background", self.client.hostName, [self.data length]);
[self parseResponse];
NSLog(#"%#, finish working on buffer, left %d in background", self.client.hostName, [self.data length]);
// NSLog(#"data : %#", [[NSString alloc] initWithData:self.data encoding:NSUTF8StringEncoding]);
}
else
{
NSLog(#"%#, no data for background processing", self.client.hostName);
}
#endif
});
}
}
But somehow I don't get the refresh done. Sometimes the completionHandler finish without any update and sometime my dispatch group never finish.
So my question is mainly:
1) what is your suggestion to combine background app refresh with multiple streams in GCD queues.
2) does those runloops still be active when I trigger in background
3) are the GCD queues still active
4) should I better schedule in one runloop for all client connection in addition to one main runloop ?
Somehow I need to fresh thoughts on the way forward.
TIA
I have been working on a cross platform real time multiplayer game and ran into similar problems. The core issue I faced was with multiple threads competing for compute cycles and bottlenecks while trying to update common resources. I also had a bunch of scheduled threads that updated my game state.
I ended up switching to a single run loop for all my connections and background scheduled events. That helped in clearly identifying and properly isolating critical sections. This also simplified my control flow. I also saw an increase in performance since there were only 4 threads now that were competing for cycles instead of the earlier 20+
So for your Q4 I recommend that you switch to a single schedule.
For Q1 again, A single combined process for app refresh will not only speed up your app but make it easier to manage and debug.
For Q2 and Q3 It's best if you tested it out on your code by using logs as there may be factors involved that are not apparent in the code you shared.
My app is making heavy use of webservice calls. Lately, some of the calls got stuck. After playing around I figured out that it
happens mostly for release builds
happens in the Simulator AND on the device (iPad, iOS 4.3)
happens more often on iPad 1 than on iPad 2
it is not limited to web services an SOAP but also affects the System.Net.WebClient
does not affest [NSString stringWithContentsOfUrl:] if invoked manually, since not bound
The effect is that the CPU load of the device drops to zero. memory is stable (in my demo project 8.5MB). If I put Console.WriteLines() everywhere, I can see that the code is stuck inside one of the WebClient.Download*() methods.
The code below demonstrates that (if built RELEASE with MT 4.0.1, LLVM off or on does not matter) downloading a file from the web over and over again fails sometimes right away on the first try, sometimes after 10 times, sometimes after around 30 downloads.
It is totally random. If you think it works, kill the app and restart it and eventually it will hang.
When building the same using MT 3.2.6, the downloading goes on all day without issues. It is impossible to break it.
MONO installed is the latest available version.
Can somebody from the MT team comment on it?
using System;
using System.IO;
using System.Threading;
using System.Net;
using System.Collections.Generic;
using System.Linq;
using MonoTouch.Foundation;
using MonoTouch.UIKit;
namespace iOSTest
{
public class Application
{
static void Main (string[] args)
{
UIApplication.Main (args);
}
}
// The name AppDelegate is referenced in the MainWindow.xib file.
public partial class AppDelegate : UIApplicationDelegate
{
private Thread oThread;
// This method is invoked when the application has loaded its UI and its ready to run
public override bool FinishedLaunching (UIApplication app, NSDictionary options)
{
// Make a release build and run on iPad 1 with iOS 4.3.2.
// Fails after downloading between 1 time and 30 times on MT 4.0.1.
// It is possible that it seems to work. Then just kill the app and restart and suddenly the effect
// will become visible. If you watch it with Instruments, CPU suddenly drops to zero. The app then is
// stuck somewhere inside WebClient. After about 10 minutes, an exception will be thrown (timeout).
// Never fails on MT 3.2.6
Console.WriteLine(MonoTouch.Constants.Version);
// A label that counts how often we downloaded.
UILabel oLbl = new UILabel(new System.Drawing.RectangleF(40, 100, 150, 30));
window.AddSubview(oLbl);
// This thread downloads the same file over and over again.
// The thread is not required to demonstrate the issue. The same problem occurs
// if the download is running on the main thread.
this.oThread = new Thread(delegate()
{
using(var oPool = new NSAutoreleasePool())
{
int i = 0;
while(true)
{
// Setup webclient and download a file from my website (around 2.4 MB)
WebClient oClient = new WebClient();
// It would be nice to hange it to your own URL to save me from all the traffic.
oClient.DownloadFile(new Uri("http://www.wildsau.net/image.axd?picture=2011%2f4%2fDSC05178.JPG"), Path.Combine(Environment.GetFolderPath(Environment.SpecialFolder.Personal), "test.jpg"));
// Increase counter and update label.
i++;
this.InvokeOnMainThread(delegate { oLbl.Text = i.ToString(); });
Console.WriteLine("Done " + i + " times.");
}
}
});
// Have a button that starts the action.
UIButton oBtn = UIButton.FromType(UIButtonType.RoundedRect);
oBtn.SetTitle("Download", UIControlState.Normal);
oBtn.Frame = new System.Drawing.RectangleF(40, 40, 150, 30);
oBtn.TouchUpInside += delegate(object sender, System.EventArgs e)
{
this.oThread.Start();
};
window.AddSubview(oBtn);
window.MakeKeyAndVisible ();
return true;
}
// This method is required in iPhoneOS 3.0
public override void OnActivated (UIApplication application)
{
}
}
}
From Gonzalo-
When the problem occurs, "kicking" the threadpool by adding another
work item will make the problem go away.
Something like this (not tested or compiled ;-) should do:
Timer timer = new Timer (AddMe);
...
WebClient wc = new WebClient ();
Uri uri = new Uri(url);
timer.Change (0, 500); // Trigger it now and every 500ms
byte[] bytes = wc.DownloadData(uri);
timer.Change (Timeout.Infinite, Timeout.Infinite);
....
static void AddMe (object state)
{
// Empty.
}
#
Works 100% of the time - for me at least - YMMV. And it did, once we put the code under stress (Lots of files to download) it stalled again. Just heard from MT that 4.0.6 will have the fix in it. Should see it later this week!
Promised to be fixed by Xamarin in the next major release. Still does not work in 4.0.4 though.