1 Userspace RCU Implementation
2 by Mathieu Desnoyers and Paul E. McKenney
7 ./bootstrap (skip if using tarball)
12 Hints: Forcing 32-bit build:
13 * CFLAGS="-m32 -g -O2" ./configure
16 * CFLAGS="-m64 -g -O2" ./configure
18 Forcing a 32-bit build with 386 backward compatibility:
19 * CFLAGS="-m32 -g -O2" ./configure --host=i386-pc-linux-gnu
21 Forcing a 32-bit build for Sparcv9 (typical for Sparc v9)
22 * CFLAGS="-m32 -Wa,-Av9a -g -O2" ./configure
24 ARCHITECTURES SUPPORTED
25 -----------------------
27 Currently, x86 (i386, i486, i586, i686), x86 64-bit, PowerPC 32/64, S390, S390x
28 and Sparcv9 32/64 are supported. Only tested on Linux so far, but should
29 theoretically work on other operating systems.
34 Usage of all urcu libraries
36 * Define _LGPL_SOURCE (only) if your code is LGPL or GPL compatible
37 before including the urcu.h or urcu-qsbr.h header. If your application
38 is distributed under another license, function calls will be generated
39 instead of inlines, so your application can link with the library.
40 * Linking with one of the libraries below is always necessary even for
41 LGPL and GPL applications.
46 * Link the application with "-lurcu".
47 * This is the preferred version of the library, in terms of
48 grace-period detection speed, read-side speed and flexibility.
49 Dynamically detects kernel support for sys_membarrier(). Falls back
50 on urcu-mb scheme if support is not present, which has slower
55 * #include <urcu-qsbr.h>
56 * Link with "-lurcu-qsbr".
57 * The QSBR flavor of RCU needs to have each reader thread executing
58 rcu_quiescent_state() periodically to progress. rcu_thread_online()
59 and rcu_thread_offline() can be used to mark long periods for which
60 the threads are not active. It provides the fastest read-side at the
61 expense of more intrusiveness in the application code.
66 * Compile any _LGPL_SOURCE code using this library with "-DRCU_MB".
67 * Link with "-lurcu-mb".
68 * This version of the urcu library uses memory barriers on the writer
69 and reader sides. This results in faster grace-period detection, but
70 results in slower reads.
72 Usage of liburcu-signal
74 * #include <urcu-signal.h>
75 * Link the application with "-lurcu-signal".
76 * Version of the library that requires a signal, typically SIGUSR1. Can
77 be overridden with -DSIGRCU by modifying Makefile.build.inc.
81 * #include <urcu-bp.h>
82 * Link with "-lurcu-bp".
83 * The BP library flavor stands for "bulletproof". It is specifically
84 designed to help tracing library to hook on applications without
85 requiring to modify these applications. rcu_init(),
86 rcu_register_thread() and rcu_unregister_thread() all become nops.
87 The state is dealt with by the library internally at the expense of
88 read-side and write-side performance.
92 Each thread that has reader critical sections (that uses
93 rcu_read_lock()/rcu_read_unlock() must first register to the URCU
94 library. This is done by calling rcu_register_thread(). Unregistration
95 must be performed before exiting the thread by using
96 rcu_unregister_thread().
100 Reader critical sections must be protected by locating them between
101 calls to rcu_read_lock() and rcu_read_unlock(). Inside that lock,
102 rcu_dereference() may be called to read an RCU protected pointer.
106 rcu_assign_pointer() and rcu_xchg_pointer() may be called anywhere.
107 After, synchronize_rcu() must be called. When it returns, the old
108 values are not in usage anymore.
110 Usage of liburcu-defer
112 * #include <urcu-defer.h>
113 * Link with "-lurcu-defer", and also with one of the urcu library
114 (either urcu, urcu-bp, urcu-mb or urcu-qsbr).
115 * Provides defer_rcu() primitive to enqueue delayed callbacks. Queued
116 callbacks are executed in batch periodically after a grace period.
117 Do _not_ use defer_rcu() within a read-side critical section, because
118 it may call synchronize_rcu() if the thread queue is full.
119 * Provides defer_rcu_ratelimit() primitive, which acts just like
120 defer_rcu(), but takes an additional rate limiter callback forcing
121 synchronized callback execution of the limiter returns non-zero.
122 * Requires that rcu_defer_barrier() must be called in library destructor
123 if a library queues callbacks and is expected to be unloaded with
125 * Its API is currently experimental. It may change in future library
128 Being careful with signals
130 The liburcu library uses signals internally. The signal handler is
131 registered with the SA_RESTART flag. However, these signals may cause
132 some non-restartable system calls to fail with errno = EINTR. Care
133 should be taken to restart system calls manually if they fail with this
134 error. A list of non-restartable system calls may be found in
135 signal(7). The liburcu-mb and liburcu-qsbr versions of the Userspace RCU
136 library do not require any signal.
138 Read-side critical sections are allowed in a signal handler with
139 liburcu and liburcu-mb. Be careful, however, to disable these signals
140 between thread creation and calls to rcu_register_thread(), because a
141 signal handler nesting on an unregistered thread would not be allowed to
142 call rcu_read_lock().
144 Read-side critical sections are _not_ allowed in a signal handler with
145 liburcu-qsbr, unless signals are disabled explicitly around each
146 rcu_quiescent_state() calls, when threads are put offline and around
147 calls to synchronize_rcu(). Even then, we do not recommend it.
149 Interaction with mutexes
151 One must be careful to do not cause deadlocks due to interaction of
152 synchronize_rcu() and RCU read-side with mutexes. If synchronize_rcu()
153 is called with a mutex held, this mutex (or any mutex which has this
154 mutex in its dependency chain) should not be acquired from within a RCU
155 read-side critical section.
159 DEBUG_RCU is used to add internal debugging self-checks to the
160 RCU library. This define adds a performance penalty when enabled.
161 Can be enabled by uncommenting the corresponding line in
166 DEBUG_YIELD is used to add random delays in the code for testing
171 By default the library is configured to use synchronization primitives
172 adequate for SMP systems. On uniprocessor systems, support for SMP
173 systems can be disabled with:
175 ./configure --disable-smp-support
177 theoretically yielding slightly better performance.