libstdc++-v3/doc/html/manual/using_concurrency.html

   1 <?xml version="1.0" encoding="UTF-8" standalone="no"?>
   2 <!DOCTYPE html PUBLIC "-//W3C//DTD XHTML 1.1//EN" "http://www.w3.org/TR/xhtml11/DTD/xhtml11.dtd">
   3 <html xmlns="http://www.w3.org/1999/xhtml"><head><title>Concurrency</title><meta name="generator" content="DocBook XSL-NS Stylesheets V1.76.1"/><meta name="keywords" content="&#10;      ISO C++&#10;    , &#10;      library&#10;    "/><meta name="keywords" content="&#10;      ISO C++&#10;    , &#10;      runtime&#10;    , &#10;      library&#10;    "/><link rel="home" href="../index.html" title="The GNU C++ Library"/><link rel="up" href="using.html" title="Chapter 3. Using"/><link rel="prev" href="using_dynamic_or_shared.html" title="Linking"/><link rel="next" href="using_exceptions.html" title="Exceptions"/></head><body><div class="navheader"><table width="100%" summary="Navigation header"><tr><th colspan="3" align="center">Concurrency</th></tr><tr><td align="left"><a accesskey="p" href="using_dynamic_or_shared.html">Prev</a> </td><th width="60%" align="center">Chapter 3. Using</th><td align="right"> <a accesskey="n" href="using_exceptions.html">Next</a></td></tr></table><hr/></div><div class="section" title="Concurrency"><div class="titlepage"><div><div><h2 class="title"><a id="manual.intro.using.concurrency"/>Concurrency</h2></div></div></div><p>This section discusses issues surrounding the proper compilation
   4       of multithreaded applications which use the Standard C++
   5       library.  This information is GCC-specific since the C++
   6       standard does not address matters of multithreaded applications.
   7    </p><div class="section" title="Prerequisites"><div class="titlepage"><div><div><h3 class="title"><a id="manual.intro.using.concurrency.prereq"/>Prerequisites</h3></div></div></div><p>All normal disclaimers aside, multithreaded C++ application are
   8       only supported when libstdc++ and all user code was built with
   9       compilers which report (via <code class="code"> gcc/g++ -v </code>) the same thread
  10       model and that model is not <span class="emphasis"><em>single</em></span>.  As long as your
  11       final application is actually single-threaded, then it should be
  12       safe to mix user code built with a thread model of
  13       <span class="emphasis"><em>single</em></span> with a libstdc++ and other C++ libraries built
  14       with another thread model useful on the platform.  Other mixes
  15       may or may not work but are not considered supported.  (Thus, if
  16       you distribute a shared C++ library in binary form only, it may
  17       be best to compile it with a GCC configured with
  18       --enable-threads for maximal interchangeability and usefulness
  19       with a user population that may have built GCC with either
  20       --enable-threads or --disable-threads.)
  21    </p><p>When you link a multithreaded application, you will probably
  22       need to add a library or flag to g++.  This is a very
  23       non-standardized area of GCC across ports.  Some ports support a
  24       special flag (the spelling isn't even standardized yet) to add
  25       all required macros to a compilation (if any such flags are
  26       required then you must provide the flag for all compilations not
  27       just linking) and link-library additions and/or replacements at
  28       link time.  The documentation is weak.  Here is a quick summary
  29       to display how ad hoc this is: On Solaris, both -pthreads and
  30       -threads (with subtly different meanings) are honored.  On OSF,
  31       -pthread and -threads (with subtly different meanings) are
  32       honored.  On Linux/i386, -pthread is honored.  On FreeBSD,
  33       -pthread is honored.  Some other ports use other switches.
  34       AFAIK, none of this is properly documented anywhere other than
  35       in ``gcc -dumpspecs'' (look at lib and cpp entries).
  36    </p></div><div class="section" title="Thread Safety"><div class="titlepage"><div><div><h3 class="title"><a id="manual.intro.using.concurrency.thread_safety"/>Thread Safety</h3></div></div></div><p>
  37 In the terms of the 2011 C++ standard a thread-safe program is one which
  38 does not perform any conflicting non-atomic operations on memory locations
  39 and so does not contain any data races.
  40 The standard places requirements on the library to ensure that no data
  41 races are caused by the library itself or by programs which use the
  42 library correctly (as described below).
  43 The C++11 memory model and library requirements are a more formal version
  44 of the <a class="link" href="http://www.sgi.com/tech/stl/thread_safety.html">SGI STL</a> definition of thread safety, which the library used
  45 prior to the 2011 standard.
  46 </p><p>The library strives to be thread-safe when all of the following
  47          conditions are met:
  48       </p><div class="itemizedlist"><ul class="itemizedlist"><li class="listitem"><p>The system's libc is itself thread-safe,
  49        </p></li><li class="listitem"><p>
  50            The compiler in use reports a thread model other than
  51            'single'. This can be tested via output from <code class="code">gcc
  52            -v</code>. Multi-thread capable versions of gcc output
  53            something like this:
  54          </p><pre class="programlisting">
  55 %gcc -v
  56 Using built-in specs.
  57 ...
  58 Thread model: posix
  59 gcc version 4.1.2 20070925 (Red Hat 4.1.2-33)
  60 </pre><p>Look for "Thread model" lines that aren't equal to "single."</p></li><li class="listitem"><p>
  61          Requisite command-line flags are used for atomic operations
  62          and threading. Examples of this include <code class="code">-pthread</code>
  63          and <code class="code">-march=native</code>, although specifics vary
  64          depending on the host environment. See <a class="link" href="http://gcc.gnu.org/onlinedocs/gcc/Option-Summary.html">Machine
  65          Dependent Options</a>.
  66        </p></li><li class="listitem"><p>
  67            An implementation of atomicity.h functions
  68            exists for the architecture in question. See the internals documentation for more <a class="link" href="internals.html#internals.thread_safety" title="Thread Safety">details</a>.
  69        </p></li></ul></div><p>The user code must guard against concurrent function calls which
  70          access any particular library object's state when one or more of
  71          those accesses modifies the state. An object will be modified by
  72          invoking a non-const member function on it or passing it as a
  73          non-const argument to a library function. An object will not be
  74          modified by invoking a const member function on it or passing it to
  75          a function as a pointer- or reference-to-const.
  76          Typically, the application
  77          programmer may infer what object locks must be held based on the
  78          objects referenced in a function call and whether the objects are
  79          accessed as const or non-const.  Without getting
  80          into great detail, here is an example which requires user-level
  81          locks:
  82       </p><pre class="programlisting">
  83      library_class_a shared_object_a;
  84
  85      void thread_main () {
  86        library_class_b *object_b = new library_class_b;
  87        shared_object_a.add_b (object_b);   // must hold lock for shared_object_a
  88        shared_object_a.mutate ();          // must hold lock for shared_object_a
  89      }
  90
  91      // Multiple copies of thread_main() are started in independent threads.</pre><p>Under the assumption that object_a and object_b are never exposed to
  92          another thread, here is an example that does not require any
  93          user-level locks:
  94       </p><pre class="programlisting">
  95      void thread_main () {
  96        library_class_a object_a;
  97        library_class_b *object_b = new library_class_b;
  98        object_a.add_b (object_b);
  99        object_a.mutate ();
 100      } </pre><p>All library types are safe to use in a multithreaded program
 101          if objects are not shared between threads or as
 102          long each thread carefully locks out access by any other
 103          thread while it modifies any object visible to another thread.
 104          Unless otherwise documented, the only exceptions to these rules
 105          are atomic operations on the types in
 106          <code class="filename">&lt;atomic&gt;</code>
 107          and lock/unlock operations on the standard mutex types in
 108          <code class="filename">&lt;mutex&gt;</code>. These
 109          atomic operations allow concurrent accesses to the same object
 110          without introducing data races.
 111       </p><p>The following member functions of standard containers can be
 112          considered to be const for the purposes of avoiding data races:
 113          <code class="code">begin</code>, <code class="code">end</code>, <code class="code">rbegin</code>, <code class="code">rend</code>,
 114          <code class="code">front</code>, <code class="code">back</code>, <code class="code">data</code>,
 115          <code class="code">find</code>, <code class="code">lower_bound</code>, <code class="code">upper_bound</code>,
 116          <code class="code">equal_range</code>, <code class="code">at</code>
 117          and, except in associative or unordered associative containers,
 118          <code class="code">operator[]</code>. In other words, although they are non-const
 119          so that they can return mutable iterators, those member functions
 120          will not modify the container.
 121          Accessing an iterator might cause a non-modifying access to
 122          the container the iterator refers to (for example incrementing a
 123          list iterator must access the pointers between nodes, which are part
 124          of the container and so conflict with other accesses to the container).
 125       </p><p>Programs which follow the rules above will not encounter data
 126          races in library code, even when using library types which share
 127          state between distinct objects.  In the example below the
 128          <code class="code">shared_ptr</code> objects share a reference count, but
 129          because the code does not perform any non-const operations on the
 130          globally-visible object, the library ensures that the reference
 131          count updates are atomic and do not introduce data races:
 132       </p><pre class="programlisting">
 133     std::shared_ptr&lt;int&gt; global_sp;
 134
 135     void thread_main() {
 136       auto local_sp = global_sp;  // OK, copy constructor's parameter is reference-to-const
 137
 138       int i = *global_sp;         // OK, operator* is const
 139       int j = *local_sp;          // OK, does not operate on global_sp
 140
 141       // *global_sp = 2;          // NOT OK, modifies int visible to other threads
 142       // *local_sp = 2;           // NOT OK, modifies int visible to other threads
 143
 144       // global_sp.reset();       // NOT OK, reset is non-const
 145       local_sp.reset();           // OK, does not operate on global_sp
 146     }
 147
 148     int main() {
 149       global_sp.reset(new int(1));
 150       std::thread t1(thread_main);
 151       std::thread t2(thread_main);
 152       t1.join();
 153       t2.join();
 154     }
 155       </pre><p>For further details of the C++11 memory model see Hans-J. Boehm's
 156       <a class="link" href="http://www.hpl.hp.com/personal/Hans_Boehm/c++mm/user-faq.html">Threads
 157       and memory model for C++</a> pages, particularly the <a class="link" href="http://www.hpl.hp.com/personal/Hans_Boehm/c++mm/threadsintro.html">introduction</a>
 158       and <a class="link" href="http://www.hpl.hp.com/personal/Hans_Boehm/c++mm/user-faq.html">FAQ</a>.
 159       </p></div><div class="section" title="Atomics"><div class="titlepage"><div><div><h3 class="title"><a id="manual.intro.using.concurrency.atomics"/>Atomics</h3></div></div></div><p>
 160     </p></div><div class="section" title="IO"><div class="titlepage"><div><div><h3 class="title"><a id="manual.intro.using.concurrency.io"/>IO</h3></div></div></div><p>This gets a bit tricky.  Please read carefully, and bear with me.
 161    </p><div class="section" title="Structure"><div class="titlepage"><div><div><h4 class="title"><a id="concurrency.io.structure"/>Structure</h4></div></div></div><p>A wrapper
 162       type called <code class="code">__basic_file</code> provides our abstraction layer
 163       for the <code class="code">std::filebuf</code> classes.  Nearly all decisions dealing
 164       with actual input and output must be made in <code class="code">__basic_file</code>.
 165    </p><p>A generic locking mechanism is somewhat in place at the filebuf layer,
 166       but is not used in the current code.  Providing locking at any higher
 167       level is akin to providing locking within containers, and is not done
 168       for the same reasons (see the links above).
 169    </p></div><div class="section" title="Defaults"><div class="titlepage"><div><div><h4 class="title"><a id="concurrency.io.defaults"/>Defaults</h4></div></div></div><p>The __basic_file type is simply a collection of small wrappers around
 170       the C stdio layer (again, see the link under Structure).  We do no
 171       locking ourselves, but simply pass through to calls to <code class="code">fopen</code>,
 172       <code class="code">fwrite</code>, and so forth.
 173    </p><p>So, for 3.0, the question of "is multithreading safe for I/O"
 174       must be answered with, "is your platform's C library threadsafe
 175       for I/O?"  Some are by default, some are not; many offer multiple
 176       implementations of the C library with varying tradeoffs of threadsafety
 177       and efficiency.  You, the programmer, are always required to take care
 178       with multiple threads.
 179    </p><p>(As an example, the POSIX standard requires that C stdio FILE*
 180        operations are atomic.  POSIX-conforming C libraries (e.g, on Solaris
 181        and GNU/Linux) have an internal mutex to serialize operations on
 182        FILE*s.  However, you still need to not do stupid things like calling
 183        <code class="code">fclose(fs)</code> in one thread followed by an access of
 184        <code class="code">fs</code> in another.)
 185    </p><p>So, if your platform's C library is threadsafe, then your
 186       <code class="code">fstream</code> I/O operations will be threadsafe at the lowest
 187       level.  For higher-level operations, such as manipulating the data
 188       contained in the stream formatting classes (e.g., setting up callbacks
 189       inside an <code class="code">std::ofstream</code>), you need to guard such accesses
 190       like any other critical shared resource.
 191    </p></div><div class="section" title="Future"><div class="titlepage"><div><div><h4 class="title"><a id="concurrency.io.future"/>Future</h4></div></div></div><p> A
 192       second choice may be available for I/O implementations:  libio.  This is
 193       disabled by default, and in fact will not currently work due to other
 194       issues.  It will be revisited, however.
 195    </p><p>The libio code is a subset of the guts of the GNU libc (glibc) I/O
 196       implementation.  When libio is in use, the <code class="code">__basic_file</code>
 197       type is basically derived from FILE.  (The real situation is more
 198       complex than that... it's derived from an internal type used to
 199       implement FILE.  See libio/libioP.h to see scary things done with
 200       vtbls.)  The result is that there is no "layer" of C stdio
 201       to go through; the filebuf makes calls directly into the same
 202       functions used to implement <code class="code">fread</code>, <code class="code">fwrite</code>,
 203       and so forth, using internal data structures.  (And when I say
 204       "makes calls directly," I mean the function is literally
 205       replaced by a jump into an internal function.  Fast but frightening.
 206       *grin*)
 207    </p><p>Also, the libio internal locks are used.  This requires pulling in
 208       large chunks of glibc, such as a pthreads implementation, and is one
 209       of the issues preventing widespread use of libio as the libstdc++
 210       cstdio implementation.
 211    </p><p>But we plan to make this work, at least as an option if not a future
 212       default.  Platforms running a copy of glibc with a recent-enough
 213       version will see calls from libstdc++ directly into the glibc already
 214       installed.  For other platforms, a copy of the libio subsection will
 215       be built and included in libstdc++.
 216    </p></div><div class="section" title="Alternatives"><div class="titlepage"><div><div><h4 class="title"><a id="concurrency.io.alt"/>Alternatives</h4></div></div></div><p>Don't forget that other cstdio implementations are possible.  You could
 217       easily write one to perform your own forms of locking, to solve your
 218       "interesting" problems.
 219    </p></div></div><div class="section" title="Containers"><div class="titlepage"><div><div><h3 class="title"><a id="manual.intro.using.concurrency.containers"/>Containers</h3></div></div></div><p>This section discusses issues surrounding the design of
 220       multithreaded applications which use Standard C++ containers.
 221       All information in this section is current as of the gcc 3.0
 222       release and all later point releases.  Although earlier gcc
 223       releases had a different approach to threading configuration and
 224       proper compilation, the basic code design rules presented here
 225       were similar.  For information on all other aspects of
 226       multithreading as it relates to libstdc++, including details on
 227       the proper compilation of threaded code (and compatibility between
 228       threaded and non-threaded code), see Chapter 17.
 229    </p><p>Two excellent pages to read when working with the Standard C++
 230       containers and threads are
 231       <a class="link" href="http://www.sgi.com/tech/stl/thread_safety.html">SGI's
 232       http://www.sgi.com/tech/stl/thread_safety.html</a> and
 233       <a class="link" href="http://www.sgi.com/tech/stl/Allocators.html">SGI's
 234       http://www.sgi.com/tech/stl/Allocators.html</a>.
 235    </p><p><span class="emphasis"><em>However, please ignore all discussions about the user-level
 236       configuration of the lock implementation inside the STL
 237       container-memory allocator on those pages.  For the sake of this
 238       discussion, libstdc++ configures the SGI STL implementation,
 239       not you.  This is quite different from how gcc pre-3.0 worked.
 240       In particular, past advice was for people using g++ to
 241       explicitly define _PTHREADS or other macros or port-specific
 242       compilation options on the command line to get a thread-safe
 243       STL.  This is no longer required for any port and should no
 244       longer be done unless you really know what you are doing and
 245       assume all responsibility.</em></span>
 246    </p><p>Since the container implementation of libstdc++ uses the SGI
 247       code, we use the same definition of thread safety as SGI when
 248       discussing design.  A key point that beginners may miss is the
 249       fourth major paragraph of the first page mentioned above
 250       (<span class="emphasis"><em>For most clients...</em></span>), which points out that
 251       locking must nearly always be done outside the container, by
 252       client code (that'd be you, not us).  There is a notable
 253       exceptions to this rule.  Allocators called while a container or
 254       element is constructed uses an internal lock obtained and
 255       released solely within libstdc++ code (in fact, this is the
 256       reason STL requires any knowledge of the thread configuration).
 257    </p><p>For implementing a container which does its own locking, it is
 258       trivial to provide a wrapper class which obtains the lock (as
 259       SGI suggests), performs the container operation, and then
 260       releases the lock.  This could be templatized <span class="emphasis"><em>to a certain
 261       extent</em></span>, on the underlying container and/or a locking
 262       mechanism.  Trying to provide a catch-all general template
 263       solution would probably be more trouble than it's worth.
 264    </p><p>The library implementation may be configured to use the
 265       high-speed caching memory allocator, which complicates thread
 266       safety issues. For all details about how to globally override
 267       this at application run-time
 268       see <a class="link" href="using_macros.html" title="Macros">here</a>. Also
 269       useful are details
 270       on <a class="link" href="memory.html#std.util.memory.allocator" title="Allocators">allocator</a>
 271       options and capabilities.
 272    </p></div></div><div class="navfooter"><hr/><table width="100%" summary="Navigation footer"><tr><td align="left"><a accesskey="p" href="using_dynamic_or_shared.html">Prev</a> </td><td align="center"><a accesskey="u" href="using.html">Up</a></td><td align="right"> <a accesskey="n" href="using_exceptions.html">Next</a></td></tr><tr><td align="left" valign="top">Linking </td><td align="center"><a accesskey="h" href="../index.html">Home</a></td><td align="right" valign="top"> Exceptions</td></tr></table></div></body></html>