Sweden-Number/documentation/threading.sgml

<chapter id="threading">
  <title>Multi-threading in Wine</title>
  
  <para>
    This section will assume you understand the basics of multithreading. If not there are plenty of
    good tutorials available on the net to get you started.
  </para>

  <para>
    Threading in Wine is somewhat complex due to several factors. The first is the advanced level of
    multithreading support provided by Windows - there are far more threading related constructs available
    in Win32 than the Linux equivalent (pthreads). The second is the need to be able to map Win32 threads
    to native Linux threads which provides us with benefits like having the kernel schedule them without
    our intervention. While it's possible to implement threading entirely without kernel support, doing so
    is not desirable on most platforms that Wine runs on.
  </para>

  <sect1>
    <title> Threading support in Win32 </title>

    <para>
      Win32 is an unusually thread friendly API. Not only is it entirely thread safe, but it provides
      many different facilities for working with threads. These range from the basics such as starting
      and stopping threads, to the extremely complex such as injecting threads into other processes and
      COM inter-thread marshalling.
    </para>

    <para>
      One of the primary challenges of writing Wine code therefore is ensuring that all our DLLs are
      thread safe, free of race conditions and so on. This isn't simple - don't be afraid to ask if
      you aren't sure whether a piece of code is thread safe or not!
    </para>

    <para>
      Win32 provides many different ways you can make your code thread safe however the most common
      are <emphasis>critical section</emphasis> and the <emphasis>interlocked functions</emphasis>.
      Critical sections are a type of mutex designed to protect a geographic area of code. If you don't
      want multiple threads running in a piece of code at once, you can protect them with calls to
      EnterCriticalSection and LeaveCriticalSection. The first call to EnterCriticalSection by a thread
      will lock the section and continue without stopping. If another thread calls it then it will block
      until the original thread calls LeaveCriticalSection again.
    </para>

    <para>
      It is therefore vitally important that if you use critical sections to make some code thread-safe,
      that you check every possible codepath out of the code to ensure that any held sections are left.
      Code like this:
    </para>

    <programlisting> if (res != ERROR_SUCCESS) return res;  </programlisting>

    <para>
      is extremely suspect in a function that also contains a call to EnterCriticalSection. Be careful.
    </para>

    <para>
      If a thread blocks while waiting for another thread to leave a critical section, you will
      see an error from the RtlpWaitForCriticalSection function, along with a note of which
      thread is holding the lock. This only appears after a certain timeout, normally a few
      seconds. It's possible the thread holding the lock is just being really slow which is why
      Wine won't terminate the app like a non-checked build of Windows would, but the most
      common cause is that for some reason a thread forgot to call LeaveCriticalSection, or died
      while holding the lock (perhaps because it was in turn waiting for another lock). This
      doesn't just happen in Wine code: a deadlock while waiting for a critical section could
      be due to a bug in the app triggered by a slight difference in the emulation.
    </para>
    
    <para>
      Another popular mechanism available is the use of functions like InterlockedIncrement and
      InterlockedExchange. These make use of native CPU abilities to execute a single
      instruction while ensuring any other processors on the system cannot access memory, and
      allow you to do common operations like add/remove/check a variable in thread-safe code
      without holding a mutex. These are useful for reference counting especially in
      free-threaded (thread safe) COM objects.
    </para>

    <para>
      Finally, the usage of TLS slots are also popular. TLS stands for thread-local storage, and is
      a set of slots scoped local to a thread which you can store pointers in. Look on MSDN for the
      TlsAlloc function to learn more about the Win32 implementation of this. Essentially, the
      contents of a given slot will be different in each thread, so you can use this to store data
      that is only meaningful in the context of a single thread. On recent versions of Linux the
      __thread keyword provides a convenient interface to this functionality - a more portable API
      is exposed in the pthread library. However, these facilities is not used by Wine, rather, we
      implement Win32 TLS entirely ourselves.
    </para>
  </sect1>

  <sect1>
    <title> SysLevels </title>

    <para>
      SysLevels are an undocumented Windows-internal thread-safety system. They are basically
      critical sections which must be taken in a particular order. The mechanism is generic but
      there are always three syslevels: level 1 is the Win16 mutex, level 2 is the USER mutex
      and level 3 is the GDI mutex.
    </para>

    <para>
      When entering a syslevel, the code (in dlls/kernel/syslevel.c) will check that a
      higher syslevel is not already held and produce an error if so. This is because it's not
      legal to enter level 2 while holding level 3 - first, you must leave level 3.
    </para>

    <para>
      Throughout the code you may see calls to _ConfirmSysLevel() and _CheckNotSysLevel(). These
      functions are essentially assertions about the syslevel states and can be used to check
      that the rules have not been accidentally violated. In particular, _CheckNotSysLevel()
      will break (probably into the debugger) if the check fails. If this happens the solution
      is to get a backtrace and find out, by reading the source of the wine functions called
      along the way, how Wine got into the invalid state.
    </para>
    
  </sect1>

  <sect1>
    <title> POSIX threading vs kernel threading </title>

    <para>
      Wine runs in one of two modes: either pthreads (posix threading) or kthreads (kernel
      threading). This section explains the differences between them. The one that is used is
      automatically selected on startup by a small test program which then execs the correct
      binary, either wine-kthread or wine-pthread. On NPTL-enabled systems pthreads will be
      used, and on older non-NPTL systems kthreads is selected.
    </para>

    <para>
      Let's start with a bit of history. Back in the dark ages when Wines threading support was
      first implemented a problem was faced - Windows had much more capable threading APIs than
      Linux did. This presented a problem - Wine works either by reimplementing an API entirely
      or by mapping it onto the underlying systems equivalent. How could Win32 threading be
      implemented using a library which did not have all the neeed features? The answer, of
      course, was that it couldn't be.
    </para>

    <para>
      On Linux the pthreads interface is used to start, stop and control threads. The pthreads
      library in turn is based on top of so-called "kernel threads" which are created using the
      clone(2) syscall. Pthreads provides a nicer (more portable) interface to this
      functionality and also provides APIs for controlling mutexes. There is a
      <ulink url="http://www.llnl.gov/computing/tutorials/pthreads/">
        good tutorial on pthreads </ulink> available if you want to learn more.
    </para>

    <para>
      As pthreads did not provide the necessary semantics to implement Win32 threading, the
      decision was made to implement Win32 threading on top of the underlying kernel threads by
      using syscalls like clone directly. This provided maximum flexibility and allowed a
      correct implementation but caused some bad side effects. Most notably, all the userland
      Linux APIs assumed that the user was utilising the pthreads library. Some only enabled
      thread safety when they detected that pthreads was in use - this is true of glibc, for
      instance. Worse, pthreads and pure kernel threads had strange interactions when run in
      the same process yet some libraries used by Wine used pthreads internally. Throw in
      source code porting using WineLib - where you have both UNIX and Win32 code in the same
      process - and chaos was the result.
    </para>

    <para>
      The solution was simple yet ingenius: Wine would provide its own implementation of the pthread
      library <emphasis>inside</emphasis> its own binary. Due to the semantics of ELF symbol
      scoping, this would cause Wines own implementations to override any implementation loaded
      later on (like the real libpthread.so). Therefore, any calls to the pthread APIs in
      external libraries would be linked to Wines instead of the systems pthreads library, and
      Wine implemented pthreads by using the standard Windows threading APIs it in turn
      implemented itself.
    </para>

    <para>
      As a result, libraries that only became thread-safe in the presence of a loaded pthreads
      implementation would now do so, and any external code that used pthreads would actually
      end up creating Win32 threads that Wine was aware of and controlled. This worked quite
      nicely for a long time, even though it required doing some extremely un-kosher things like
      overriding internal libc structures and functions. That is, it worked until NPTL was
      developed at which point the underlying thread implementation on Linux changed
      dramatically.
    </para>

    <para>
      The fake pthread implementation can be found in loader/kthread.c, which is used to
      produce to wine-kthread binary. In contrast, loader/pthread.c produces the wine-pthread
      binary which is used on newer NPTL systems.
    </para>

    <para>
      NPTL is a new threading subsystem for Linux that hugely improves its performance and
      flexibility. By allowing threads to become much more scalable and adding new pthread
      APIs, NPTL made Linux competitive with Windows in the multi-threaded world. Unfortunately
      it also broke many assumptions made by Wine (as well as other applications such as the
      Sun JVM and RealPlayer) in the process.
    </para>

    <para>
      There was, however, some good news. NPTL made Linux threading powerful enough
      that Win32 threads could now be implemented on top of pthreads like any other normal
      application. There would no longer be problems with mixing win32-kthreads and pthreads
      created by external libraries, and no need to override glibc internals. As you can see
      from the relative sizes of the loader/kthread.c and loader/pthread.c files, the
      difference in code complexity is considerable. NPTL also made several other semantic
      changes to things such as signal delivery so changes were required in many different
      places in Wine.
    </para>

    <para>
      On non-Linux systems the threading interface is typically not powerful enough to
      replicate the semantics Win32 applications expect and so kthreads with the
      pthread overrides are used.
    </para>
  </sect1>

  <sect1>
    <title> The Win32 thread environment </title>

    <para>
      All Win32 code, whether from a native EXE/DLL or in Wine itself, expects certain constructs to
      be present in its environment. This section explores what those constructs are and how Wine
      sets them up. The lack of this environment is one thing that makes it hard to use Wine code
      directly from standard Linux applications - in order to interact with Win32 code a thread
      must first be "adopted" by Wine.
    </para>

    <para>
      The first thing Win32 code requires is the <emphasis>TEB</emphasis> or "Thread Environment
      Block". This is an internal (undocumented) Windows structure associated with every thread
      which stores a variety of things such as TLS slots, a pointer to the threads message queue,
      the last error code and so on. You can see the definition of the TEB in include/thread.h, or
      at least what we know of it so far. Being internal and subject to change, the layout of the
      TEB has had to be reverse engineered from scratch.
    </para>

    <para>
      A pointer to the TEB is stored in the %fs register and can be accessed using NtCurrentTeb()
      from within Wine code. %fs actually stores a selector, and setting it therefore requires
      modifying the processes local descriptor table (LDT) - the code to do this is in lib/wine/ldt.c.
    </para>

    <para>
      The TEB is required by nearly all Win32 code run in the Wine environment, as any wineserver
      RPC will use it, which in turn implies that any code which could possibly block (for instance
      by using a critical section) needs it. The TEB also holds the SEH exception handler chain as
      the first element, so if when disassembling you see code like this:
    </para>

    <programlisting> movl %esp, %fs:0 </programlisting>

    <para>
      ... then you are seeing the program set up an SEH handler frame. All threads must have at
      least one SEH entry, which normally points to the backstop handler which is ultimately
      responsible for popping up the all-too-familiar "This program has performed an illegal
      operation and will be terminated" message. On Wine we just drop straight into the debugger.
      A full description of SEH is out of the scope of this section, however there are some good
      articles in MSJ if you are interested.
    </para>

    <para>
      All Win32-aware threads must have a wineserver connection. Many different APIs
      require the ability to communicate with the wineserver. In turn, the wineserver must be aware
      of Win32 threads in order to be able to accurately report information to other parts of the
      program and do things like route inter-thread messages, dispatch APCs (asynchronous procedure
      calls) and so on. Therefore a part of thread initialization is initializing the thread
      serverside. The result is not only correct information in the server, but a set of file
      descriptors the thread can use to communicate with the server - the request fd, reply fd and
      wait fd (used for blocking).
    </para>
    
  </sect1>
</chapter>