README

   1 Userspace RCU Implementation
   2 by Mathieu Desnoyers and Paul E. McKenney
   3
   4 BUILDING
   5 --------
   6
   7         ./bootstrap (skip if using tarball)
   8         ./configure
   9         make
  10         make install
  11         ldconfig
  12
  13         Hints:  Forcing 32-bit build:
  14                 * CFLAGS="-m32 -g -O2" ./configure
  15
  16                 Forcing 64-bit build:
  17                 * CFLAGS="-m64 -g -O2" ./configure
  18
  19                 Forcing a 32-bit build with 386 backward compatibility:
  20                 * CFLAGS="-m32 -g -O2" ./configure --host=i386-pc-linux-gnu
  21
  22                 Forcing a 32-bit build for Sparcv9 (typical for Sparc v9)
  23                 * CFLAGS="-m32 -Wa,-Av9a -g -O2" ./configure
  24
  25 ARCHITECTURES SUPPORTED
  26 -----------------------
  27
  28 Currently, Linux x86 (i386, i486, i586, i686), x86 64-bit, PowerPC 32/64,
  29 S390, S390x, ARM, Alpha, ia64 and Sparcv9 32/64 are supported. Tested on
  30 Linux, FreeBSD 8.2/9.0, and Cygwin. Should also work on: Android, NetBSD 5,
  31 OpenBSD, Darwin (more testing needed before claiming support for these OS).
  32
  33 Linux ARM depends on running a Linux kernel 2.6.15 or better, GCC 4.4 or
  34 better.
  35
  36 The gcc compiler versions 3.3, 3.4, 4.0, 4.1, 4.2, 4.3, 4.4 and 4.5 are
  37 supported, with the following exceptions:
  38
  39 - gcc 3.3 and 3.4 have a bug that prevents them from generating volatile
  40   accesses to offsets in a TLS structure on 32-bit x86. These versions are
  41   therefore not compatible with liburcu on x86 32-bit (i386, i486, i586, i686).
  42   The problem has been reported to the gcc community:
  43     http://www.mail-archive.com/gcc-bugs@gcc.gnu.org/msg281255.html
  44 - gcc 3.3 cannot match the "xchg" instruction on 32-bit x86 build.
  45   See: http://kerneltrap.org/node/7507
  46 - Alpha, ia64 and ARM architectures depend on gcc 4.x with atomic builtins
  47   support. For ARM this was introduced with gcc 4.4:
  48     http://gcc.gnu.org/gcc-4.4/changes.html
  49
  50 Clang version 3.0 (based on LLVM 3.0) is supported.
  51
  52 For developers using the git tree:
  53
  54 This source tree is based on the autotools suite from GNU to simplify
  55 portability. Here are some things you should have on your system in order to
  56 compile the git repository tree :
  57
  58 - GNU autotools (automake >=1.10, autoconf >=2.50, autoheader >=2.50)
  59   (make sure your system wide "automake" points to a recent version!)
  60 - GNU Libtool >=2.2
  61   (for more information, go to http://www.gnu.org/software/autoconf/)
  62
  63 If you get the tree from the repository, you will need to use the "bootstrap"
  64 script in the root of the tree. It calls all the GNU tools needed to prepare the
  65 tree configuration.
  66
  67 Test scripts provided in the tests/ directory of the source tree depend
  68 on "bash" and the "seq" program.
  69
  70
  71 QUICK START GUIDE
  72 -----------------
  73
  74 Usage of all urcu libraries
  75
  76         * Define _LGPL_SOURCE (only) if your code is LGPL or GPL compatible
  77           before including the urcu.h or urcu-qsbr.h header. If your application
  78           is distributed under another license, function calls will be generated
  79           instead of inlines, so your application can link with the library.
  80         * Linking with one of the libraries below is always necessary even for
  81           LGPL and GPL applications.
  82
  83 Usage of liburcu
  84
  85         * #include <urcu.h>
  86         * Link the application with "-lurcu".
  87         * This is the preferred version of the library, in terms of
  88           grace-period detection speed, read-side speed and flexibility.
  89           Dynamically detects kernel support for sys_membarrier(). Falls back
  90           on urcu-mb scheme if support is not present, which has slower
  91           read-side.
  92
  93 Usage of liburcu-qsbr
  94
  95         * #include <urcu-qsbr.h>
  96         * Link with "-lurcu-qsbr".
  97         * The QSBR flavor of RCU needs to have each reader thread executing
  98           rcu_quiescent_state() periodically to progress. rcu_thread_online()
  99           and rcu_thread_offline() can be used to mark long periods for which
 100           the threads are not active. It provides the fastest read-side at the
 101           expense of more intrusiveness in the application code.
 102
 103 Usage of liburcu-mb
 104
 105         * #include <urcu.h>
 106         * Compile any _LGPL_SOURCE code using this library with "-DRCU_MB".
 107         * Link with "-lurcu-mb".
 108         * This version of the urcu library uses memory barriers on the writer
 109           and reader sides. This results in faster grace-period detection, but
 110           results in slower reads.
 111
 112 Usage of liburcu-signal
 113
 114         * #include <urcu.h>
 115         * Compile any _LGPL_SOURCE code using this library with "-DRCU_SIGNAL".
 116         * Link the application with "-lurcu-signal".
 117         * Version of the library that requires a signal, typically SIGUSR1. Can
 118           be overridden with -DSIGRCU by modifying Makefile.build.inc.
 119
 120 Usage of liburcu-bp
 121
 122         * #include <urcu-bp.h>
 123         * Link with "-lurcu-bp".
 124         * The BP library flavor stands for "bulletproof". It is specifically
 125           designed to help tracing library to hook on applications without
 126           requiring to modify these applications. rcu_init(),
 127           rcu_register_thread() and rcu_unregister_thread() all become nops.
 128           The state is dealt with by the library internally at the expense of
 129           read-side and write-side performance.
 130
 131 Initialization
 132
 133         Each thread that has reader critical sections (that uses
 134         rcu_read_lock()/rcu_read_unlock() must first register to the URCU
 135         library. This is done by calling rcu_register_thread(). Unregistration
 136         must be performed before exiting the thread by using
 137         rcu_unregister_thread().
 138
 139 Reading
 140
 141         Reader critical sections must be protected by locating them between
 142         calls to rcu_read_lock() and rcu_read_unlock(). Inside that lock,
 143         rcu_dereference() may be called to read an RCU protected pointer.
 144
 145 Writing
 146
 147         rcu_assign_pointer() and rcu_xchg_pointer() may be called anywhere.
 148         After, synchronize_rcu() must be called. When it returns, the old
 149         values are not in usage anymore.
 150
 151 Usage of liburcu-defer
 152
 153         * Follow instructions for either liburcu, liburcu-qsbr,
 154           liburcu-mb, liburcu-signal, or liburcu-bp above.
 155           The liburcu-defer functionality is pulled into each of
 156           those library modules.
 157         * Provides defer_rcu() primitive to enqueue delayed callbacks. Queued
 158           callbacks are executed in batch periodically after a grace period.
 159           Do _not_ use defer_rcu() within a read-side critical section, because
 160           it may call synchronize_rcu() if the thread queue is full.
 161           This can lead to deadlock or worse.
 162         * Requires that rcu_defer_barrier() must be called in library destructor
 163           if a library queues callbacks and is expected to be unloaded with
 164           dlclose().
 165         * Its API is currently experimental. It may change in future library
 166           releases.
 167
 168 Usage of urcu-call-rcu
 169
 170         * Follow instructions for either liburcu, liburcu-qsbr,
 171           liburcu-mb, liburcu-signal, or liburcu-bp above.
 172           The urcu-call-rcu functionality is provided for each of
 173           these library modules.
 174         * Provides the call_rcu() primitive to enqueue delayed callbacks
 175           in a manner similar to defer_rcu(), but without ever delaying
 176           for a grace period.  On the other hand, call_rcu()'s best-case
 177           overhead is not quite as good as that of defer_rcu().
 178         * Provides call_rcu() to allow asynchronous handling of RCU
 179           grace periods.  A number of additional functions are provided
 180           to manage the helper threads used by call_rcu(), but reasonable
 181           defaults are used if these additional functions are not invoked.
 182           See rcu-api.txt in userspace-rcu documentation for more details.
 183
 184 Being careful with signals
 185
 186         The liburcu library uses signals internally. The signal handler is
 187         registered with the SA_RESTART flag. However, these signals may cause
 188         some non-restartable system calls to fail with errno = EINTR. Care
 189         should be taken to restart system calls manually if they fail with this
 190         error. A list of non-restartable system calls may be found in
 191         signal(7). The liburcu-mb and liburcu-qsbr versions of the Userspace RCU
 192         library do not require any signal.
 193
 194         Read-side critical sections are allowed in a signal handler,
 195         except those setup with sigaltstack(2), with liburcu and
 196         liburcu-mb. Be careful, however, to disable these signals
 197         between thread creation and calls to rcu_register_thread(), because a
 198         signal handler nesting on an unregistered thread would not be
 199         allowed to call rcu_read_lock().
 200
 201         Read-side critical sections are _not_ allowed in a signal handler with
 202         liburcu-qsbr, unless signals are disabled explicitly around each
 203         rcu_quiescent_state() calls, when threads are put offline and around
 204         calls to synchronize_rcu(). Even then, we do not recommend it.
 205
 206 Interaction with mutexes
 207
 208         One must be careful to do not cause deadlocks due to interaction of
 209         synchronize_rcu() and RCU read-side with mutexes. If synchronize_rcu()
 210         is called with a mutex held, this mutex (or any mutex which has this
 211         mutex in its dependency chain) should not be acquired from within a RCU
 212         read-side critical section.
 213
 214         This is especially important to understand in the context of the
 215         QSBR flavor: a registered reader thread being "online" by
 216         default should be considered as within a RCU read-side critical
 217         section unless explicitly put "offline". Therefore, if
 218         synchronize_rcu() is called with a mutex held, this mutex, as
 219         well as any mutex which has this mutex in its dependency chain
 220         should only be taken when the RCU reader thread is "offline"
 221         (this can be performed by calling rcu_thread_offline()).
 222
 223 Usage of DEBUG_RCU
 224
 225         DEBUG_RCU is used to add internal debugging self-checks to the
 226         RCU library. This define adds a performance penalty when enabled.
 227         Can be enabled by uncommenting the corresponding line in
 228         Makefile.build.inc.
 229
 230 Usage of DEBUG_YIELD
 231
 232         DEBUG_YIELD is used to add random delays in the code for testing
 233         purposes.
 234
 235 SMP support
 236
 237         By default the library is configured to use synchronization primitives
 238         adequate for SMP systems. On uniprocessor systems, support for SMP
 239         systems can be disabled with:
 240
 241                 ./configure --disable-smp-support
 242
 243         theoretically yielding slightly better performance.
 244
 245 Interaction with fork()
 246
 247         Special care must be taken for applications performing fork() without
 248         any following exec(). This is caused by the fact that Linux only clones
 249         the thread calling fork(), and thus never replicates any of the other
 250         parent thread into the child process. Most liburcu implementations
 251         require that all registrations (as reader, defer_rcu and call_rcu
 252         threads) should be released before a fork() is performed, except for the
 253         rather common scenario where fork() is immediately followed by exec() in
 254         the child process. The only implementation not subject to that rule is
 255         liburcu-bp, which is designed to handle fork() by calling
 256         rcu_bp_before_fork, rcu_bp_after_fork_parent and
 257         rcu_bp_after_fork_child.
 258
 259         Applications that use call_rcu() and that fork() without
 260         doing an immediate exec() must take special action.  The parent
 261         must invoke call_rcu_before_fork() before the fork() and
 262         call_rcu_after_fork_parent() after the fork().  The child
 263         process must invoke call_rcu_after_fork_child().
 264         Even though these three APIs are suitable for passing to
 265         pthread_atfork(), use of pthread_atfork() is *STRONGLY
 266         DISCOURAGED* for programs calling the glibc memory allocator
 267         (malloc(), calloc(), free(), ...) within call_rcu callbacks.
 268         This is due to limitations in the way glibc memory allocator
 269         handles calls to the memory allocator from concurrent threads
 270         while the pthread_atfork() handlers are executing.
 271         Combining e.g.:
 272         * call to free() from callbacks executed within call_rcu worker
 273           threads,
 274         * executing call_rcu atfork handlers within the glibc pthread
 275           atfork mechanism,
 276         will sometimes trigger interesting process hangs. This usually
 277         hangs on a memory allocator lock within glibc.
 278
 279 Thread Local Storage (TLS)
 280
 281         Userspace RCU can fall back on pthread_getspecific() to emulate
 282         TLS variables on systems where it is not available. This behavior
 283         can be forced by specifying --disable-compiler-tls as configure
 284         argument.