1 Userspace RCU Implementation
2 ============================
4 by Mathieu Desnoyers and Paul E. McKenney
10 ./bootstrap # skip if using tarball
18 - Forcing 32-bit build:
20 CFLAGS="-m32 -g -O2" ./configure
22 - Forcing 64-bit build:
24 CFLAGS="-m64 -g -O2" ./configure
26 - Forcing a 32-bit build with 386 backward compatibility:
28 CFLAGS="-m32 -g -O2" ./configure --host=i386-pc-linux-gnu
30 - Forcing a 32-bit build for Sparcv9 (typical for Sparc v9)
32 CFLAGS="-m32 -Wa,-Av9a -g -O2" ./configure
35 Architectures supported
36 -----------------------
38 Currently, the following architectures are supported:
40 - x86 (i386, i486, i586, i686)
57 - Linux all architectures
58 - FreeBSD 13 i386/amd64
69 (more testing needed before claiming support for these OS).
75 The C compiler used needs to support at least C99. The C++ compiler used needs
76 to support at least C++11. The oldest GCC version officialy supported and
79 Older GCC versions might still work with the following exceptions:
81 - GCC 3.3 and 3.4 have a bug that prevents them from generating volatile
82 accesses to offsets in a TLS structure on 32-bit x86. These versions are
83 therefore not compatible with `liburcu` on x86 32-bit
84 (i386, i486, i586, i686).
85 The problem has been reported to the GCC community:
86 <http://www.mail-archive.com/gcc-bugs@gcc.gnu.org/msg281255.html>
87 - GCC 3.3 cannot match the "xchg" instruction on 32-bit x86 build.
88 See <http://kerneltrap.org/node/7507>
89 - Alpha, ia64 and ARM architectures depend on GCC 4.x with atomic builtins
90 support. For ARM this was introduced with GCC 4.4:
91 <http://gcc.gnu.org/gcc-4.4/changes.html>.
92 - Linux aarch64 depends on GCC 5.1 or better because prior versions
93 perform unsafe access to deallocated stack.
95 Clang version 3.0 (based on LLVM 3.0) is supported.
97 Glibc >= 2.4 should work but the older version we test against is
104 For developers using the Git tree:
106 This source tree is based on the autotools suite from GNU to simplify
107 portability. Here are some things you should have on your system in order to
108 compile the git repository tree :
110 - GNU autotools (automake >=1.12, autoconf >=2.69)
111 (make sure your system wide `automake` points to a recent version!)
113 (for more information, go to <http://www.gnu.org/software/autoconf/>)
115 If you get the tree from the repository, you will need to use the `bootstrap`
116 script in the root of the tree. It calls all the GNU tools needed to prepare
117 the tree configuration.
119 Test scripts provided in the `tests/` directory of the source tree depend
120 on `bash` and the `seq` program.
126 See the relevant API documentation files in `doc/`. The APIs provided by
127 Userspace RCU are, by prefix:
129 - `rcu_`: Read-Copy Update (see [`doc/rcu-api.md`](doc/rcu-api.md))
130 - `cmm_`: Concurrent Memory Model
131 - `caa_`: Concurrent Architecture Abstraction
132 - `cds_`: Concurrent Data Structures
133 (see [`doc/cds-api.md`](doc/cds-api.md))
134 - `uatomic_`: Userspace Atomic
135 (see [`doc/uatomic-api.md`](doc/uatomic-api.md))
141 ### Usage of all urcu libraries:
143 - Define `_LGPL_SOURCE` (only) if your code is LGPL or GPL compatible
144 before including the `urcu.h` or `urcu-qsbr.h` header. If your application
145 is distributed under another license, function calls will be generated
146 instead of inlines, so your application can link with the library.
147 - Linking with one of the libraries below is always necessary even for
148 LGPL and GPL applications.
149 - Define `URCU_INLINE_SMALL_FUNCTIONS` before including Userspace RCU
150 headers if you want Userspace RCU to inline small functions (10
151 lines or less) into the application. It can be used by applications
152 distributed under any kind of license, and does *not* make the
153 application a derived work of Userspace RCU.
155 Those small inlined functions are guaranteed to match the library
156 content as long as the library major version is unchanged.
157 Therefore, the application *must* be compiled with headers matching
158 the library major version number. Applications using
159 `URCU_INLINE_SMALL_FUNCTIONS` may be unable to use debugging
160 features of Userspace RCU without being recompiled.
162 There are multiple flavors of liburcu available:
170 The API members start with the prefix `urcu_<flavor>_`, where
171 `<flavor>` is the chosen flavor name.
174 ### Usage of `liburcu-memb`
176 1. `#include <urcu/urcu-memb.h>`
177 2. Link the application with `-lurcu-memb`
179 This is the preferred version of the library, in terms of
180 grace-period detection speed, read-side speed and flexibility.
181 Dynamically detects kernel support for `sys_membarrier()`. Falls back
182 on `urcu-mb` scheme if support is not present, which has slower
183 read-side. Use the `--disable-sys-membarrier-fallback` configure option
184 to disable the fall back, thus requiring `sys_membarrier()` to be
185 available. This gives a small speedup when `sys_membarrier()` is
186 supported by the kernel, and aborts in the library constructor if not
190 ### Usage of `liburcu-qsbr`
192 1. `#include <urcu/urcu-qsbr.h>`
193 2. Link with `-lurcu-qsbr`
195 The QSBR flavor of RCU needs to have each reader thread executing
196 `rcu_quiescent_state()` periodically to progress. `rcu_thread_online()`
197 and `rcu_thread_offline()` can be used to mark long periods for which
198 the threads are not active. It provides the fastest read-side at the
199 expense of more intrusiveness in the application code.
202 ### Usage of `liburcu-mb`
204 1. `#include <urcu/urcu-mb.h>`
205 2. Link with `-lurcu-mb`
207 This version of the urcu library uses memory barriers on the writer
208 and reader sides. This results in faster grace-period detection, but
209 results in slower reads.
212 ### Usage of `liburcu-signal`
214 1. `#include <urcu/urcu-signal.h>`
215 2. Link the application with `-lurcu-signal`
217 Version of the library that requires a signal, typically `SIGUSR1`. Can
218 be overridden with `-DSIGRCU` by modifying `Makefile.build.inc`.
221 ### Usage of `liburcu-bp`
223 1. `#include <urcu/urcu-bp.h>`
224 2. Link with `-lurcu-bp`
226 The BP library flavor stands for "bulletproof". It is specifically
227 designed to help tracing library to hook on applications without
228 requiring to modify these applications. `urcu_bp_init()`, and
229 `urcu_bp_unregister_thread()` all become nops, whereas calling
230 `urcu_bp_register_thread()` becomes optional. The state is dealt with by
231 the library internally at the expense of read-side and write-side
237 Each thread that has reader critical sections (that uses
238 `urcu_<flavor>_read_lock()`/`urcu_<flavor>_read_unlock()` must first
239 register to the URCU library. This is done by calling
240 `urcu_<flavor>_register_thread()`. Unregistration must be performed
241 before exiting the thread by using `urcu_<flavor>_unregister_thread()`.
246 Reader critical sections must be protected by locating them between
247 calls to `urcu_<flavor>_read_lock()` and `urcu_<flavor>_read_unlock()`.
248 Inside that lock, `rcu_dereference()` may be called to read an RCU
254 `rcu_assign_pointer()` and `rcu_xchg_pointer()` may be called anywhere.
255 After, `urcu_<flavor>_synchronize_rcu()` must be called. When it
256 returns, the old values are not in usage anymore.
258 As an alternative to `urcu_<flavor>_synchronize_rcu()`,
259 it is also possible to use the urcu polling mechanism to wait for a
260 grace period to elapse. This can be done by using
261 `urcu_<flavor>_start_poll_synchronize_rcu()`
262 to start the grace period polling, and then invoke
263 `urcu_<flavor>_poll_state_synchronize_rcu()`, which returns true if
264 the grace period has completed, false otherwise.
267 ### Usage of `liburcu-defer`
269 - Follow instructions for either `liburcu-memb`, `liburcu-qsbr`,
270 `liburcu-mb`, `liburcu-signal`, or `liburcu-bp` above.
271 The `liburcu-defer` functionality is pulled into each of
272 those library modules.
273 - Provides `urcu_<flavor>_defer_rcu()` primitive to enqueue delayed
274 callbacks. Queued callbacks are executed in batch periodically after
275 a grace period. Do _not_ use `urcu_<flavor>_defer_rcu()` within a
276 read-side critical section, because it may call
277 `urcu_<flavor>_synchronize_rcu()` if the thread queue is full. This
278 can lead to deadlock or worse.
279 - Requires that `urcu_<flavor>_defer_barrier()` must be called in
280 library destructor if a library queues callbacks and is expected to
281 be unloaded with `dlclose()`.
283 Its API is currently experimental. It may change in future library releases.
286 ### Usage of `urcu-call-rcu`
288 - Follow instructions for either `liburcu-memb`, `liburcu-qsbr`,
289 `liburcu-mb`, `liburcu-signal`, or `liburcu-bp` above.
290 The `urcu-call-rcu` functionality is pulled into each of
291 those library modules.
292 - Provides the `urcu_<flavor>_call_rcu()` primitive to enqueue delayed
293 callbacks in a manner similar to `urcu_<flavor>_defer_rcu()`, but
294 without ever delaying for a grace period. On the other hand,
295 `urcu_<flavor>_call_rcu()`'s best-case overhead is not quite as good
296 as that of `urcu_<flavor>_defer_rcu()`.
297 - Provides `urcu_<flavor>_call_rcu()` to allow asynchronous handling
298 of RCU grace periods. A number of additional functions are provided
299 to manage the helper threads used by `urcu_<flavor>_call_rcu()`, but
300 reasonable defaults are used if these additional functions are not
301 invoked. See [`doc/rcu-api.md`](doc/rcu-api.md) in userspace-rcu
302 documentation for more details.
305 ### Being careful with signals
307 The `liburcu-signal` library uses signals internally. The signal handler is
308 registered with the `SA_RESTART` flag. However, these signals may cause
309 some non-restartable system calls to fail with `errno = EINTR`. Care
310 should be taken to restart system calls manually if they fail with this
311 error. A list of non-restartable system calls may be found in
314 Read-side critical sections are allowed in a signal handler,
315 except those setup with `sigaltstack(2)`, with `liburcu-memb` and
316 `liburcu-mb`. Be careful, however, to disable these signals
317 between thread creation and calls to `urcu_<flavor>_register_thread()`,
318 because a signal handler nesting on an unregistered thread would not be
319 allowed to call `urcu_<flavor>_read_lock()`.
321 Read-side critical sections are _not_ allowed in a signal handler with
322 `liburcu-qsbr`, unless signals are disabled explicitly around each
323 `urcu_qsbr_quiescent_state()` calls, when threads are put offline and around
324 calls to `urcu_qsbr_synchronize_rcu()`. Even then, we do not recommend it.
327 ### Interaction with mutexes
329 One must be careful to do not cause deadlocks due to interaction of
330 `urcu_<flavor>_synchronize_rcu()` and RCU read-side with mutexes. If
331 `urcu_<flavor>_synchronize_rcu()` is called with a mutex held, this
332 mutex (or any mutex which has this mutex in its dependency chain) should
333 not be acquired from within a RCU read-side critical section.
335 This is especially important to understand in the context of the
336 QSBR flavor: a registered reader thread being "online" by
337 default should be considered as within a RCU read-side critical
338 section unless explicitly put "offline". Therefore, if
339 `urcu_qsbr_synchronize_rcu()` is called with a mutex held, this mutex,
340 as well as any mutex which has this mutex in its dependency chain should
341 only be taken when the RCU reader thread is "offline" (this can be
342 performed by calling `urcu_qsbr_thread_offline()`).
345 ### Interaction with `fork()`
347 Special care must be taken for applications performing `fork()` without
348 any following `exec()`. This is caused by the fact that Linux only clones
349 the thread calling `fork()`, and thus never replicates any of the other
350 parent thread into the child process. Most `liburcu` implementations
351 require that all registrations (as reader, `defer_rcu` and `call_rcu`
352 threads) should be released before a `fork()` is performed, except for the
353 rather common scenario where `fork()` is immediately followed by `exec()` in
354 the child process. The only implementation not subject to that rule is
355 `liburcu-bp`, which is designed to handle `fork()` by calling
356 `urcu_bp_before_fork`, `urcu_bp_after_fork_parent` and
357 `urcu_bp_after_fork_child`.
359 Applications that use `urcu_<flavor>_call_rcu()` and that `fork()`
360 without doing an immediate `exec()` must take special action. The
361 parent must invoke `urcu_<flavor>_call_rcu_before_fork()` before the
362 `fork()` and `urcu_<flavor>_call_rcu_after_fork_parent()` after the
363 `fork()`. The child process must invoke
364 `urcu_<flavor>_call_rcu_after_fork_child()`. Even though these three
365 APIs are suitable for passing to `pthread_atfork()`, use of
366 `pthread_atfork()` is **STRONGLY DISCOURAGED** for programs calling the
367 glibc memory allocator (`malloc()`, `calloc()`, `free()`, ...) within
368 `urcu_<flavor>_call_rcu` callbacks. This is due to limitations in the
369 way glibc memory allocator handles calls to the memory allocator from
370 concurrent threads while the `pthread_atfork()` handlers are executing.
374 - call to `free()` from callbacks executed within
375 `urcu_<flavor>_call_rcu` worker threads,
376 - executing `urcu_<flavor>_call_rcu` atfork handlers within the glibc
377 pthread atfork mechanism,
379 will sometimes trigger interesting process hangs. This usually
380 hangs on a memory allocator lock within glibc.
383 ### Thread Local Storage (TLS)
385 Userspace RCU can fall back on `pthread_getspecific()` to emulate
386 TLS variables on systems where it is not available. This behavior
387 can be forced by specifying `--disable-compiler-tls` as configure
391 ### Usage of `DEBUG_RCU` & `--enable-rcu-debug`
393 By default the library is configured with internal debugging
394 self-checks disabled.
396 For always-on debugging self-checks:
398 ./configure --enable-rcu-debug
400 For fine grained enabling of debugging self-checks, build
401 userspace-rcu with `DEBUG_RCU` defined and compile dependent
402 applications with `DEBUG_RCU` defined when necessary.
404 Warning: Enabling this feature result in a performance penalty.
407 ### Usage of `DEBUG_YIELD`
409 `DEBUG_YIELD` is used to add random delays in the code for testing
415 By default the library is configured to use synchronization primitives
416 adequate for SMP systems. On uniprocessor systems, support for SMP
417 systems can be disabled with:
419 ./configure --disable-smp-support
421 theoretically yielding slightly better performance.
424 ### Usage of `--enable-cds-lfht-iter-debug`
426 By default the library is configured with extra debugging checks for
427 lock-free hash table iterator traversal disabled.
429 Building liburcu with `--enable-cds-lfht-iter-debug` and rebuilding
430 application to match the ABI change allows finding cases where the hash
431 table iterator is re-purposed to be used on a different hash table while
432 still being used to iterate on a hash table.
434 This option alters the rculfhash ABI. Make sure to compile both library
435 and application with matching configuration.
441 In addition to the usual `make check` target, Userspace RCU features
442 `make regtest`, `make short_bench` and `make long_bench` targets:
444 - `make check`: short tests, meant to be run when rebuilding or
445 porting Userspace RCU.
446 - `make regtest`: long (many hours) test, meant to be run when
447 modifying Userspace RCU or porting it to a new architecture or
449 - `make short_bench`: short benchmarks, 3 seconds per test.
450 - `make long_bench`: long (many hours) benchmarks, 30 seconds per test.
456 There is an application vs library compatibility issue between
457 applications built using Userspace RCU 0.10 headers linked against
458 Userspace RCU 0.11 or 0.12 shared objects. The problem occurs as
461 - An application executable is built with `_LGPL_SOURCE` defined, includes
462 any of the Userspace RCU 0.10 urcu flavor headers, and is built
463 without the `-fpic` compiler option.
465 - The Userspace RCU 0.10 library shared objects are updated to 0.11
466 or 0.12 without rebuilding the application.
468 - The application will hang, typically when RCU grace period
469 (synchronize_rcu) is invoked.
471 Some possible work-arounds for this are:
473 - Rebuild the application against Userspace RCU 0.11+.
475 - Rebuild the application with `-fpic`.
477 - Upgrade Userspace RCU to 0.13+ without installing 0.11 nor 0.12.
483 You can contact the maintainers on the following mailing list:
484 `lttng-dev@lists.lttng.org`.