1 Userspace RCU Implementation
2 by Mathieu Desnoyers and Paul E. McKenney
7 ./bootstrap (skip if using tarball)
13 Hints: Forcing 32-bit build:
14 * CFLAGS="-m32 -g -O2" ./configure
17 * CFLAGS="-m64 -g -O2" ./configure
19 Forcing a 32-bit build with 386 backward compatibility:
20 * CFLAGS="-m32 -g -O2" ./configure --host=i386-pc-linux-gnu
22 Forcing a 32-bit build for Sparcv9 (typical for Sparc v9)
23 * CFLAGS="-m32 -Wa,-Av9a -g -O2" ./configure
26 ARCHITECTURES SUPPORTED
27 -----------------------
29 Currently, Linux x86 (i386, i486, i586, i686), x86 64-bit, PowerPC 32/64,
30 S390, S390x, ARM, MIPS, Alpha, ia64 and Sparcv9 32/64 are supported.
31 Tested on Linux, FreeBSD 8.2/8.3/9.0/9.1/10.0 i386/amd64, and Cygwin.
32 Should also work on: Android, NetBSD 5, OpenBSD, Darwin (more testing
33 needed before claiming support for these OS).
35 Linux ARM depends on running a Linux kernel 2.6.15 or better, GCC 4.4 or
38 The gcc compiler versions 3.3, 3.4, 4.0, 4.1, 4.2, 4.3, 4.4 and 4.5 are
39 supported, with the following exceptions:
41 - gcc 3.3 and 3.4 have a bug that prevents them from generating volatile
42 accesses to offsets in a TLS structure on 32-bit x86. These versions are
43 therefore not compatible with liburcu on x86 32-bit (i386, i486, i586, i686).
44 The problem has been reported to the gcc community:
45 http://www.mail-archive.com/gcc-bugs@gcc.gnu.org/msg281255.html
46 - gcc 3.3 cannot match the "xchg" instruction on 32-bit x86 build.
47 See: http://kerneltrap.org/node/7507
48 - Alpha, ia64 and ARM architectures depend on gcc 4.x with atomic builtins
49 support. For ARM this was introduced with gcc 4.4:
50 http://gcc.gnu.org/gcc-4.4/changes.html
52 Clang version 3.0 (based on LLVM 3.0) is supported.
54 For developers using the git tree:
56 This source tree is based on the autotools suite from GNU to simplify
57 portability. Here are some things you should have on your system in order to
58 compile the git repository tree :
60 - GNU autotools (automake >=1.10, autoconf >=2.50, autoheader >=2.50)
61 (make sure your system wide "automake" points to a recent version!)
63 (for more information, go to http://www.gnu.org/software/autoconf/)
65 If you get the tree from the repository, you will need to use the "bootstrap"
66 script in the root of the tree. It calls all the GNU tools needed to prepare the
69 Test scripts provided in the tests/ directory of the source tree depend
70 on "bash" and the "seq" program.
76 See the relevant API documentation files in doc/. The APIs provided by
77 Userspace RCU are, by prefix:
79 - rcu_ : Read-Copy Update (see doc/rcu-api.txt)
80 - cmm_ : Concurrent Memory Model
81 - caa_ : Concurrent Architecture Abstraction
82 - cds_ : Concurrent Data Structures (see doc/cds-api.txt)
83 - uatomic_: Userspace Atomic (see doc/uatomic-api.txt)
89 Usage of all urcu libraries
91 * Define _LGPL_SOURCE (only) if your code is LGPL or GPL compatible
92 before including the urcu.h or urcu-qsbr.h header. If your application
93 is distributed under another license, function calls will be generated
94 instead of inlines, so your application can link with the library.
95 * Linking with one of the libraries below is always necessary even for
96 LGPL and GPL applications.
101 * Link the application with "-lurcu".
102 * This is the preferred version of the library, in terms of
103 grace-period detection speed, read-side speed and flexibility.
104 Dynamically detects kernel support for sys_membarrier(). Falls back
105 on urcu-mb scheme if support is not present, which has slower
108 Usage of liburcu-qsbr
110 * #include <urcu-qsbr.h>
111 * Link with "-lurcu-qsbr".
112 * The QSBR flavor of RCU needs to have each reader thread executing
113 rcu_quiescent_state() periodically to progress. rcu_thread_online()
114 and rcu_thread_offline() can be used to mark long periods for which
115 the threads are not active. It provides the fastest read-side at the
116 expense of more intrusiveness in the application code.
121 * Compile any _LGPL_SOURCE code using this library with "-DRCU_MB".
122 * Link with "-lurcu-mb".
123 * This version of the urcu library uses memory barriers on the writer
124 and reader sides. This results in faster grace-period detection, but
125 results in slower reads.
127 Usage of liburcu-signal
130 * Compile any _LGPL_SOURCE code using this library with "-DRCU_SIGNAL".
131 * Link the application with "-lurcu-signal".
132 * Version of the library that requires a signal, typically SIGUSR1. Can
133 be overridden with -DSIGRCU by modifying Makefile.build.inc.
137 * #include <urcu-bp.h>
138 * Link with "-lurcu-bp".
139 * The BP library flavor stands for "bulletproof". It is specifically
140 designed to help tracing library to hook on applications without
141 requiring to modify these applications. rcu_init(),
142 rcu_register_thread() and rcu_unregister_thread() all become nops.
143 The state is dealt with by the library internally at the expense of
144 read-side and write-side performance.
148 Each thread that has reader critical sections (that uses
149 rcu_read_lock()/rcu_read_unlock() must first register to the URCU
150 library. This is done by calling rcu_register_thread(). Unregistration
151 must be performed before exiting the thread by using
152 rcu_unregister_thread().
156 Reader critical sections must be protected by locating them between
157 calls to rcu_read_lock() and rcu_read_unlock(). Inside that lock,
158 rcu_dereference() may be called to read an RCU protected pointer.
162 rcu_assign_pointer() and rcu_xchg_pointer() may be called anywhere.
163 After, synchronize_rcu() must be called. When it returns, the old
164 values are not in usage anymore.
166 Usage of liburcu-defer
168 * Follow instructions for either liburcu, liburcu-qsbr,
169 liburcu-mb, liburcu-signal, or liburcu-bp above.
170 The liburcu-defer functionality is pulled into each of
171 those library modules.
172 * Provides defer_rcu() primitive to enqueue delayed callbacks. Queued
173 callbacks are executed in batch periodically after a grace period.
174 Do _not_ use defer_rcu() within a read-side critical section, because
175 it may call synchronize_rcu() if the thread queue is full.
176 This can lead to deadlock or worse.
177 * Requires that rcu_defer_barrier() must be called in library destructor
178 if a library queues callbacks and is expected to be unloaded with
180 * Its API is currently experimental. It may change in future library
183 Usage of urcu-call-rcu
185 * Follow instructions for either liburcu, liburcu-qsbr,
186 liburcu-mb, liburcu-signal, or liburcu-bp above.
187 The urcu-call-rcu functionality is provided for each of
188 these library modules.
189 * Provides the call_rcu() primitive to enqueue delayed callbacks
190 in a manner similar to defer_rcu(), but without ever delaying
191 for a grace period. On the other hand, call_rcu()'s best-case
192 overhead is not quite as good as that of defer_rcu().
193 * Provides call_rcu() to allow asynchronous handling of RCU
194 grace periods. A number of additional functions are provided
195 to manage the helper threads used by call_rcu(), but reasonable
196 defaults are used if these additional functions are not invoked.
197 See rcu-api.txt in userspace-rcu documentation for more details.
199 Being careful with signals
201 The liburcu library uses signals internally. The signal handler is
202 registered with the SA_RESTART flag. However, these signals may cause
203 some non-restartable system calls to fail with errno = EINTR. Care
204 should be taken to restart system calls manually if they fail with this
205 error. A list of non-restartable system calls may be found in
206 signal(7). The liburcu-mb and liburcu-qsbr versions of the Userspace RCU
207 library do not require any signal.
209 Read-side critical sections are allowed in a signal handler,
210 except those setup with sigaltstack(2), with liburcu and
211 liburcu-mb. Be careful, however, to disable these signals
212 between thread creation and calls to rcu_register_thread(), because a
213 signal handler nesting on an unregistered thread would not be
214 allowed to call rcu_read_lock().
216 Read-side critical sections are _not_ allowed in a signal handler with
217 liburcu-qsbr, unless signals are disabled explicitly around each
218 rcu_quiescent_state() calls, when threads are put offline and around
219 calls to synchronize_rcu(). Even then, we do not recommend it.
221 Interaction with mutexes
223 One must be careful to do not cause deadlocks due to interaction of
224 synchronize_rcu() and RCU read-side with mutexes. If synchronize_rcu()
225 is called with a mutex held, this mutex (or any mutex which has this
226 mutex in its dependency chain) should not be acquired from within a RCU
227 read-side critical section.
229 This is especially important to understand in the context of the
230 QSBR flavor: a registered reader thread being "online" by
231 default should be considered as within a RCU read-side critical
232 section unless explicitly put "offline". Therefore, if
233 synchronize_rcu() is called with a mutex held, this mutex, as
234 well as any mutex which has this mutex in its dependency chain
235 should only be taken when the RCU reader thread is "offline"
236 (this can be performed by calling rcu_thread_offline()).
238 Interaction with fork()
240 Special care must be taken for applications performing fork() without
241 any following exec(). This is caused by the fact that Linux only clones
242 the thread calling fork(), and thus never replicates any of the other
243 parent thread into the child process. Most liburcu implementations
244 require that all registrations (as reader, defer_rcu and call_rcu
245 threads) should be released before a fork() is performed, except for the
246 rather common scenario where fork() is immediately followed by exec() in
247 the child process. The only implementation not subject to that rule is
248 liburcu-bp, which is designed to handle fork() by calling
249 rcu_bp_before_fork, rcu_bp_after_fork_parent and
250 rcu_bp_after_fork_child.
252 Applications that use call_rcu() and that fork() without
253 doing an immediate exec() must take special action. The parent
254 must invoke call_rcu_before_fork() before the fork() and
255 call_rcu_after_fork_parent() after the fork(). The child
256 process must invoke call_rcu_after_fork_child().
257 Even though these three APIs are suitable for passing to
258 pthread_atfork(), use of pthread_atfork() is *STRONGLY
259 DISCOURAGED* for programs calling the glibc memory allocator
260 (malloc(), calloc(), free(), ...) within call_rcu callbacks.
261 This is due to limitations in the way glibc memory allocator
262 handles calls to the memory allocator from concurrent threads
263 while the pthread_atfork() handlers are executing.
265 * call to free() from callbacks executed within call_rcu worker
267 * executing call_rcu atfork handlers within the glibc pthread
269 will sometimes trigger interesting process hangs. This usually
270 hangs on a memory allocator lock within glibc.
272 Thread Local Storage (TLS)
274 Userspace RCU can fall back on pthread_getspecific() to emulate
275 TLS variables on systems where it is not available. This behavior
276 can be forced by specifying --disable-compiler-tls as configure
281 DEBUG_RCU is used to add internal debugging self-checks to the
282 RCU library. This define adds a performance penalty when enabled.
283 Can be enabled by uncommenting the corresponding line in
288 DEBUG_YIELD is used to add random delays in the code for testing
293 By default the library is configured to use synchronization primitives
294 adequate for SMP systems. On uniprocessor systems, support for SMP
295 systems can be disabled with:
297 ./configure --disable-smp-support
299 theoretically yielding slightly better performance.