| 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 down to 386 compatibility: |
| 19 | * CFLAGS="-m32 -g -O2" ./configure --target=i386-pc-linux-gnu |
| 20 | |
| 21 | ARCHITECTURES SUPPORTED |
| 22 | ----------------------- |
| 23 | |
| 24 | Currently, x86 (i386, i486, i586, i686), x86 64, PowerPC 32/64 and S390 are |
| 25 | supported. The current use of sys_futex() makes it Linux-dependent, although |
| 26 | this portability limitation might go away in a near future by using the pthread |
| 27 | cond vars. |
| 28 | |
| 29 | QUICK START GUIDE |
| 30 | ----------------- |
| 31 | |
| 32 | Usage of all urcu libraries |
| 33 | |
| 34 | * Define _LGPL_SOURCE (only) if your code is LGPL or GPL compatible |
| 35 | before including the urcu.h or urcu-qsbr.h header. If your application |
| 36 | is distributed under another license, function calls will be generated |
| 37 | instead of inlines, so your application can link with the library. |
| 38 | * Linking with one of the libraries below is always necessary even for |
| 39 | LGPL and GPL applications. |
| 40 | |
| 41 | Usage of liburcu |
| 42 | |
| 43 | * #include <urcu.h> |
| 44 | * Link the application with "-lurcu". |
| 45 | * This is the preferred version of the library, both in terms of speed |
| 46 | and flexibility. Requires a signal, typically SIGUSR1. Can be |
| 47 | overridden with -DSIGURCU by modifying Makefile.build.inc. |
| 48 | |
| 49 | Usage of liburcu-mb |
| 50 | |
| 51 | * #include <urcu.h> |
| 52 | * Compile any _LGPL_SOURCE code using this library with "-DURCU_MB". |
| 53 | * Link with "-lurcu-mb". |
| 54 | * This version of the urcu library does not need to |
| 55 | reserve a signal number. URCU_MB uses full memory barriers for |
| 56 | readers. This eliminates the need for signals but results in slower |
| 57 | reads. |
| 58 | |
| 59 | Usage of liburcu-qsbr |
| 60 | |
| 61 | * #include <urcu-qsbr.h> |
| 62 | * Link with "-lurcu-qsbr". |
| 63 | * The QSBR flavor of RCU needs to have each reader thread executing |
| 64 | rcu_quiescent_state() periodically to progress. rcu_thread_online() |
| 65 | and rcu_thread_offline() can be used to mark long periods for which |
| 66 | the threads are not active. It provides the fastest read-side at the |
| 67 | expense of more intrusiveness in the application code. |
| 68 | |
| 69 | Usage of liburcu-bp |
| 70 | |
| 71 | * #include <urcu-bp.h> |
| 72 | * Link with "-lurcu-bp". |
| 73 | * The BP library flavor stands for "bulletproof". It is specifically |
| 74 | designed to help tracing library to hook on applications without |
| 75 | requiring to modify these applications. urcu_init(), |
| 76 | rcu_register_thread() and rcu_unregister_thread() all become nops. |
| 77 | The state is dealt with by the library internally at the expense of |
| 78 | read-side and write-side performance. |
| 79 | |
| 80 | Initialization |
| 81 | |
| 82 | Each thread that has reader critical sections (that uses |
| 83 | rcu_read_lock()/rcu_read_unlock() must first register to the URCU |
| 84 | library. This is done by calling rcu_register_thread(). Unregistration |
| 85 | must be performed before exiting the thread by using |
| 86 | rcu_unregister_thread(). |
| 87 | |
| 88 | Reading |
| 89 | |
| 90 | Reader critical sections must be protected by locating them between |
| 91 | calls to rcu_read_lock() and rcu_read_unlock(). Inside that lock, |
| 92 | rcu_dereference() may be called to read an RCU protected pointer. |
| 93 | |
| 94 | Writing |
| 95 | |
| 96 | rcu_assign_pointer() and rcu_xchg_pointer() may be called anywhere. |
| 97 | After, synchronize_rcu() must be called. When it returns, the old |
| 98 | values are not in usage anymore. |
| 99 | |
| 100 | Usage of liburcu-defer |
| 101 | |
| 102 | * #include <urcu-defer.h> |
| 103 | * Link with "-lurcu-defer" |
| 104 | * Provides call_rcu() primitive to enqueue delayed callbacks. Queued |
| 105 | callbacks are executed in batch periodically after a grace period. |
| 106 | Do _not_ use call_rcu() within a read-side critical section, because |
| 107 | it may call synchronize_rcu() if the thread queue is full. |
| 108 | |
| 109 | Being careful with signals |
| 110 | |
| 111 | The liburcu library uses signals internally. The signal handler is |
| 112 | registered with the SA_RESTART flag. However, these signals may cause |
| 113 | some non-restartable system calls to fail with errno = EINTR. Care |
| 114 | should be taken to restart system calls manually if they fail with this |
| 115 | error. A list of non-restartable system calls may be found in |
| 116 | signal(7). The liburcu-mb and liburcu-qsbr versions of the Userspace RCU |
| 117 | library do not require any signal. |
| 118 | |
| 119 | Read-side critical sections are allowed in a signal handler with |
| 120 | liburcu and liburcu-mb. Be careful, however, to disable these signals |
| 121 | between thread creation and calls to rcu_register_thread(), because a |
| 122 | signal handler nesting on an unregistered thread would not be allowed to |
| 123 | call rcu_read_lock(). |
| 124 | |
| 125 | Read-side critical sections are _not_ allowed in a signal handler with |
| 126 | liburcu-qsbr, unless signals are disabled explicitly around each |
| 127 | rcu_quiescent_state() calls, when threads are put offline and around |
| 128 | calls to synchronize_rcu(). Even then, we do not recommend it. |
| 129 | |
| 130 | Usage of DEBUG_RCU |
| 131 | |
| 132 | DEBUG_RCU is used to add internal debugging self-checks to the |
| 133 | RCU library. This define adds a performance penalty when enabled. |
| 134 | Can be enabled by uncommenting the corresponding line in |
| 135 | Makefile.build.inc. |
| 136 | |
| 137 | Usage of DEBUG_YIELD |
| 138 | |
| 139 | DEBUG_YIELD is used to add random delays in the code for testing |
| 140 | purposes. |
| 141 | |
| 142 | SMP support |
| 143 | |
| 144 | By default the library is configured to use synchronization primitives |
| 145 | adequate for SMP systems. On uniprocessor systems, support for SMP |
| 146 | systems can be disabled with: |
| 147 | |
| 148 | ./configure --disable-smp-support |
| 149 | |
| 150 | theoretically yielding slightly better performance. |