| 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 | |
| 12 | Hints: Forcing 32-bit build: |
| 13 | * CFLAGS="-m32 -g -O2" ./configure |
| 14 | |
| 15 | Forcing 64-bit build: |
| 16 | * CFLAGS="-m64 -g -O2" ./configure |
| 17 | |
| 18 | Forcing a 32-bit build with 386 backward compatibility: |
| 19 | * CFLAGS="-m32 -g -O2" ./configure --host=i386-pc-linux-gnu |
| 20 | |
| 21 | Forcing a 32-bit build for Sparcv9 (typical for Sparc v9) |
| 22 | * CFLAGS="-m32 -Wa,-Av9a -g -O2" ./configure |
| 23 | |
| 24 | ARCHITECTURES SUPPORTED |
| 25 | ----------------------- |
| 26 | |
| 27 | Currently, x86 (i386, i486, i586, i686), x86 64-bit, PowerPC 32/64, S390, S390x, |
| 28 | ARM, Alpha, ia64 and Sparcv9 32/64 are supported. Only tested on Linux so |
| 29 | far, but should theoretically work on other operating systems. |
| 30 | |
| 31 | ARM depends on running a Linux kernel 2.6.15 or better. |
| 32 | |
| 33 | The gcc compiler versions 3.3, 3.4, 4.0, 4.1, 4.2, 4.3, 4.4 and 4.5 are |
| 34 | supported, with the following exceptions: |
| 35 | |
| 36 | - gcc 3.3 and 3.4 have a bug that prevents them from generating volatile |
| 37 | accesses to offsets in a TLS structure on 32-bit x86. These versions are |
| 38 | therefore not compatible with liburcu on x86 32-bit (i386, i486, i586, i686). |
| 39 | The problem has been reported to the gcc community: |
| 40 | http://www.mail-archive.com/gcc-bugs@gcc.gnu.org/msg281255.html |
| 41 | - Alpha, ia64 and ARM architectures depend on 4.x gcc with atomic builtins |
| 42 | support. |
| 43 | |
| 44 | |
| 45 | QUICK START GUIDE |
| 46 | ----------------- |
| 47 | |
| 48 | Usage of all urcu libraries |
| 49 | |
| 50 | * Define _LGPL_SOURCE (only) if your code is LGPL or GPL compatible |
| 51 | before including the urcu.h or urcu-qsbr.h header. If your application |
| 52 | is distributed under another license, function calls will be generated |
| 53 | instead of inlines, so your application can link with the library. |
| 54 | * Linking with one of the libraries below is always necessary even for |
| 55 | LGPL and GPL applications. |
| 56 | |
| 57 | Usage of liburcu |
| 58 | |
| 59 | * #include <urcu.h> |
| 60 | * Link the application with "-lurcu". |
| 61 | * This is the preferred version of the library, in terms of |
| 62 | grace-period detection speed, read-side speed and flexibility. |
| 63 | Dynamically detects kernel support for sys_membarrier(). Falls back |
| 64 | on urcu-mb scheme if support is not present, which has slower |
| 65 | read-side. |
| 66 | |
| 67 | Usage of liburcu-qsbr |
| 68 | |
| 69 | * #include <urcu-qsbr.h> |
| 70 | * Link with "-lurcu-qsbr". |
| 71 | * The QSBR flavor of RCU needs to have each reader thread executing |
| 72 | rcu_quiescent_state() periodically to progress. rcu_thread_online() |
| 73 | and rcu_thread_offline() can be used to mark long periods for which |
| 74 | the threads are not active. It provides the fastest read-side at the |
| 75 | expense of more intrusiveness in the application code. |
| 76 | |
| 77 | Usage of liburcu-mb |
| 78 | |
| 79 | * #include <urcu.h> |
| 80 | * Compile any _LGPL_SOURCE code using this library with "-DRCU_MB". |
| 81 | * Link with "-lurcu-mb". |
| 82 | * This version of the urcu library uses memory barriers on the writer |
| 83 | and reader sides. This results in faster grace-period detection, but |
| 84 | results in slower reads. |
| 85 | |
| 86 | Usage of liburcu-signal |
| 87 | |
| 88 | * #include <urcu.h> |
| 89 | * Compile any _LGPL_SOURCE code using this library with "-DRCU_SIGNAL". |
| 90 | * Link the application with "-lurcu-signal". |
| 91 | * Version of the library that requires a signal, typically SIGUSR1. Can |
| 92 | be overridden with -DSIGRCU by modifying Makefile.build.inc. |
| 93 | |
| 94 | Usage of liburcu-bp |
| 95 | |
| 96 | * #include <urcu-bp.h> |
| 97 | * Link with "-lurcu-bp". |
| 98 | * The BP library flavor stands for "bulletproof". It is specifically |
| 99 | designed to help tracing library to hook on applications without |
| 100 | requiring to modify these applications. rcu_init(), |
| 101 | rcu_register_thread() and rcu_unregister_thread() all become nops. |
| 102 | The state is dealt with by the library internally at the expense of |
| 103 | read-side and write-side performance. |
| 104 | |
| 105 | Initialization |
| 106 | |
| 107 | Each thread that has reader critical sections (that uses |
| 108 | rcu_read_lock()/rcu_read_unlock() must first register to the URCU |
| 109 | library. This is done by calling rcu_register_thread(). Unregistration |
| 110 | must be performed before exiting the thread by using |
| 111 | rcu_unregister_thread(). |
| 112 | |
| 113 | Reading |
| 114 | |
| 115 | Reader critical sections must be protected by locating them between |
| 116 | calls to rcu_read_lock() and rcu_read_unlock(). Inside that lock, |
| 117 | rcu_dereference() may be called to read an RCU protected pointer. |
| 118 | |
| 119 | Writing |
| 120 | |
| 121 | rcu_assign_pointer() and rcu_xchg_pointer() may be called anywhere. |
| 122 | After, synchronize_rcu() must be called. When it returns, the old |
| 123 | values are not in usage anymore. |
| 124 | |
| 125 | Usage of liburcu-defer |
| 126 | |
| 127 | * Follow instructions for either liburcu, liburcu-qsbr, |
| 128 | liburcu-mb, liburcu-signal, or liburcu-bp above. |
| 129 | The liburcu-defer functionality is pulled into each of |
| 130 | those library modules. |
| 131 | * Provides defer_rcu() primitive to enqueue delayed callbacks. Queued |
| 132 | callbacks are executed in batch periodically after a grace period. |
| 133 | Do _not_ use defer_rcu() within a read-side critical section, because |
| 134 | it may call synchronize_rcu() if the thread queue is full. |
| 135 | This can lead to deadlock or worse. |
| 136 | * Requires that rcu_defer_barrier() must be called in library destructor |
| 137 | if a library queues callbacks and is expected to be unloaded with |
| 138 | dlclose(). |
| 139 | * Its API is currently experimental. It may change in future library |
| 140 | releases. |
| 141 | |
| 142 | Usage of urcu-call-rcu |
| 143 | |
| 144 | * Follow instructions for either liburcu, liburcu-qsbr, |
| 145 | liburcu-mb, liburcu-signal, or liburcu-bp above. |
| 146 | The urcu-call-rcu functionality is provided for each of |
| 147 | these library modules. |
| 148 | * Provides the call_rcu() primitive to enqueue delayed callbacks |
| 149 | in a manner similar to defer_rcu(), but without ever delaying |
| 150 | for a grace period. On the other hand, call_rcu()'s best-case |
| 151 | overhead is not quite as good as that of defer_rcu(). |
| 152 | * Provides call_rcu() to allow asynchronous handling of RCU |
| 153 | grace periods. A number of additional functions are provided |
| 154 | to manage the helper threads used by call_rcu(), but reasonable |
| 155 | defaults are used if these additional functions are not invoked. |
| 156 | See API.txt for more details. |
| 157 | |
| 158 | Being careful with signals |
| 159 | |
| 160 | The liburcu library uses signals internally. The signal handler is |
| 161 | registered with the SA_RESTART flag. However, these signals may cause |
| 162 | some non-restartable system calls to fail with errno = EINTR. Care |
| 163 | should be taken to restart system calls manually if they fail with this |
| 164 | error. A list of non-restartable system calls may be found in |
| 165 | signal(7). The liburcu-mb and liburcu-qsbr versions of the Userspace RCU |
| 166 | library do not require any signal. |
| 167 | |
| 168 | Read-side critical sections are allowed in a signal handler with |
| 169 | liburcu and liburcu-mb. Be careful, however, to disable these signals |
| 170 | between thread creation and calls to rcu_register_thread(), because a |
| 171 | signal handler nesting on an unregistered thread would not be allowed to |
| 172 | call rcu_read_lock(). |
| 173 | |
| 174 | Read-side critical sections are _not_ allowed in a signal handler with |
| 175 | liburcu-qsbr, unless signals are disabled explicitly around each |
| 176 | rcu_quiescent_state() calls, when threads are put offline and around |
| 177 | calls to synchronize_rcu(). Even then, we do not recommend it. |
| 178 | |
| 179 | Interaction with mutexes |
| 180 | |
| 181 | One must be careful to do not cause deadlocks due to interaction of |
| 182 | synchronize_rcu() and RCU read-side with mutexes. If synchronize_rcu() |
| 183 | is called with a mutex held, this mutex (or any mutex which has this |
| 184 | mutex in its dependency chain) should not be acquired from within a RCU |
| 185 | read-side critical section. |
| 186 | |
| 187 | Usage of DEBUG_RCU |
| 188 | |
| 189 | DEBUG_RCU is used to add internal debugging self-checks to the |
| 190 | RCU library. This define adds a performance penalty when enabled. |
| 191 | Can be enabled by uncommenting the corresponding line in |
| 192 | Makefile.build.inc. |
| 193 | |
| 194 | Usage of DEBUG_YIELD |
| 195 | |
| 196 | DEBUG_YIELD is used to add random delays in the code for testing |
| 197 | purposes. |
| 198 | |
| 199 | SMP support |
| 200 | |
| 201 | By default the library is configured to use synchronization primitives |
| 202 | adequate for SMP systems. On uniprocessor systems, support for SMP |
| 203 | systems can be disabled with: |
| 204 | |
| 205 | ./configure --disable-smp-support |
| 206 | |
| 207 | theoretically yielding slightly better performance. |
| 208 | |
| 209 | Interaction with fork() |
| 210 | |
| 211 | Special care must be taken for applications performing fork() without |
| 212 | any following exec(). This is caused by the fact that Linux only clones |
| 213 | the thread calling fork(), and thus never replicates any of the other |
| 214 | parent thread into the child process. Most liburcu implementations |
| 215 | require that all registrations (as reader, defer_rcu and call_rcu |
| 216 | threads) should be released before a fork() is performed, except for the |
| 217 | rather common scenario where fork() is immediately followed by exec() in |
| 218 | the child process. The only implementation not subject to that rule is |
| 219 | liburcu-bp, which is designed to handle fork() by calling |
| 220 | rcu_bp_before_fork, rcu_bp_after_fork_parent and |
| 221 | rcu_bp_after_fork_child. |
| 222 | |
| 223 | Applications that use call_rcu() and that fork() without |
| 224 | doing an immediate exec() must take special action. The parent |
| 225 | must invoke call_rcu_before_fork() before the fork() and |
| 226 | call_rcu_after_fork_parent() after the fork(). The child |
| 227 | process must invoke call_rcu_after_fork_child(). |
| 228 | These three APIs are suitable for passing to pthread_atfork(). |