Commit | Line | Data |
---|---|---|
d589a916 PP |
1 | Userspace RCU Implementation |
2 | ============================ | |
3 | ||
4 | by Mathieu Desnoyers and Paul E. McKenney | |
5 | ||
6 | ||
7 | Building | |
8 | -------- | |
9 | ||
10 | ./bootstrap # skip if using tarball | |
11 | ./configure | |
12 | make | |
13 | make install | |
14 | ldconfig | |
15 | ||
16 | Hints: | |
17 | ||
18 | - Forcing 32-bit build: | |
19 | ||
20 | CFLAGS="-m32 -g -O2" ./configure | |
21 | ||
22 | - Forcing 64-bit build: | |
23 | ||
24 | CFLAGS="-m64 -g -O2" ./configure | |
25 | ||
26 | - Forcing a 32-bit build with 386 backward compatibility: | |
27 | ||
28 | CFLAGS="-m32 -g -O2" ./configure --host=i386-pc-linux-gnu | |
29 | ||
30 | - Forcing a 32-bit build for Sparcv9 (typical for Sparc v9) | |
31 | ||
32 | CFLAGS="-m32 -Wa,-Av9a -g -O2" ./configure | |
33 | ||
34 | ||
35 | Architectures supported | |
36 | ----------------------- | |
37 | ||
38 | Currently, the following architectures are supported: | |
39 | ||
40 | - Linux x86 (i386, i486, i586, i686) | |
41 | - x86 64-bit | |
42 | - PowerPC 32/64 | |
43 | - S390, S390x | |
44 | - ARM 32/64 | |
45 | - MIPS | |
859050b3 | 46 | - NIOS2 |
d589a916 PP |
47 | - Alpha |
48 | - ia64 | |
49 | - Sparcv9 32/64 | |
50 | - Tilera | |
51 | - hppa/PA-RISC | |
52 | ||
53 | Tested on Linux, FreeBSD 8.2/8.3/9.0/9.1/10.0 i386/amd64, and Cygwin. | |
54 | Should also work on: | |
55 | ||
56 | - Android | |
57 | - NetBSD 5 | |
58 | - OpenBSD | |
59 | - Darwin | |
60 | ||
61 | (more testing needed before claiming support for these OS). | |
62 | ||
63 | Linux ARM depends on running a Linux kernel 2.6.15 or better, GCC 4.4 or | |
64 | better. | |
65 | ||
66 | The GCC compiler versions 3.3, 3.4, 4.0, 4.1, 4.2, 4.3, 4.4 and 4.5 are | |
67 | supported, with the following exceptions: | |
68 | ||
69 | - GCC 3.3 and 3.4 have a bug that prevents them from generating volatile | |
70 | accesses to offsets in a TLS structure on 32-bit x86. These versions are | |
71 | therefore not compatible with `liburcu` on x86 32-bit | |
72 | (i386, i486, i586, i686). | |
73 | The problem has been reported to the GCC community: | |
74 | http://www.mail-archive.com/gcc-bugs@gcc.gnu.org/msg281255.html | |
75 | - GCC 3.3 cannot match the "xchg" instruction on 32-bit x86 build. | |
76 | See http://kerneltrap.org/node/7507 | |
77 | - Alpha, ia64 and ARM architectures depend on GCC 4.x with atomic builtins | |
78 | support. For ARM this was introduced with GCC 4.4: | |
79 | http://gcc.gnu.org/gcc-4.4/changes.html. | |
80 | ||
81 | Clang version 3.0 (based on LLVM 3.0) is supported. | |
82 | ||
83 | Building on MacOS X (Darwin) requires a work-around for processor | |
84 | detection: | |
85 | ||
86 | - 32-bit: | |
87 | ||
88 | ./configure --build=i686-apple-darwin11 | |
89 | ||
90 | - 64-bit: | |
91 | ||
92 | ./configure --build=x86_64-apple-darwin11 | |
93 | ||
94 | For developers using the Git tree: | |
95 | ||
96 | This source tree is based on the autotools suite from GNU to simplify | |
97 | portability. Here are some things you should have on your system in order to | |
98 | compile the git repository tree : | |
99 | ||
100 | - GNU autotools (automake >=1.10, autoconf >=2.50, autoheader >=2.50) | |
101 | (make sure your system wide `automake` points to a recent version!) | |
102 | - GNU Libtool >=2.2 | |
103 | (for more information, go to http://www.gnu.org/software/autoconf/) | |
104 | ||
105 | If you get the tree from the repository, you will need to use the `bootstrap` | |
106 | script in the root of the tree. It calls all the GNU tools needed to prepare | |
107 | the tree configuration. | |
108 | ||
109 | Test scripts provided in the `tests/` directory of the source tree depend | |
110 | on `bash` and the `seq` program. | |
111 | ||
112 | ||
113 | API | |
114 | --- | |
115 | ||
116 | See the relevant API documentation files in `doc/`. The APIs provided by | |
117 | Userspace RCU are, by prefix: | |
118 | ||
dcb9c05a | 119 | - `rcu_`: Read-Copy Update (see [`doc/rcu-api.md`](doc/rcu-api.md)) |
d589a916 PP |
120 | - `cmm_`: Concurrent Memory Model |
121 | - `caa_`: Concurrent Architecture Abstraction | |
122 | - `cds_`: Concurrent Data Structures | |
dcb9c05a | 123 | (see [`doc/cds-api.md`](doc/cds-api.md)) |
d589a916 | 124 | - `uatomic_`: Userspace Atomic |
dcb9c05a | 125 | (see [`doc/uatomic-api.md`](doc/uatomic-api.md)) |
d589a916 PP |
126 | |
127 | ||
128 | Quick start guide | |
129 | ----------------- | |
130 | ||
131 | ### Usage of all urcu libraries: | |
132 | ||
133 | - Define `_LGPL_SOURCE` (only) if your code is LGPL or GPL compatible | |
134 | before including the `urcu.h` or `urcu-qsbr.h` header. If your application | |
135 | is distributed under another license, function calls will be generated | |
136 | instead of inlines, so your application can link with the library. | |
137 | - Linking with one of the libraries below is always necessary even for | |
138 | LGPL and GPL applications. | |
139 | - Define `URCU_INLINE_SMALL_FUNCTIONS` before including Userspace RCU | |
140 | headers if you want Userspace RCU to inline small functions (10 | |
141 | lines or less) into the application. It can be used by applications | |
142 | distributed under any kind of license, and does *not* make the | |
143 | application a derived work of Userspace RCU. | |
144 | ||
145 | Those small inlined functions are guaranteed to match the library | |
146 | content as long as the library major version is unchanged. | |
147 | Therefore, the application *must* be compiled with headers matching | |
148 | the library major version number. Applications using | |
149 | `URCU_INLINE_SMALL_FUNCTIONS` may be unable to use debugging | |
150 | features of Userspace RCU without being recompiled. | |
151 | ||
152 | ||
153 | ### Usage of `liburcu` | |
154 | ||
155 | 1. `#include <urcu.h>` | |
156 | 2. Link the application with `-lurcu` | |
157 | ||
158 | This is the preferred version of the library, in terms of | |
159 | grace-period detection speed, read-side speed and flexibility. | |
160 | Dynamically detects kernel support for `sys_membarrier()`. Falls back | |
161 | on `urcu-mb` scheme if support is not present, which has slower | |
162 | read-side. | |
163 | ||
164 | ||
165 | ### Usage of `liburcu-qsbr` | |
166 | ||
167 | 1. `#include <urcu-qsbr.h>` | |
168 | 2. Link with `-lurcu-qsbr` | |
169 | ||
170 | The QSBR flavor of RCU needs to have each reader thread executing | |
171 | `rcu_quiescent_state()` periodically to progress. `rcu_thread_online()` | |
172 | and `rcu_thread_offline()` can be used to mark long periods for which | |
173 | the threads are not active. It provides the fastest read-side at the | |
174 | expense of more intrusiveness in the application code. | |
175 | ||
176 | ||
177 | ### Usage of `liburcu-mb` | |
178 | ||
179 | 1. `#include <urcu.h>` | |
180 | 2. Compile any `_LGPL_SOURCE` code using this library with `-DRCU_MB` | |
181 | 3. Link with `-lurcu-mb` | |
182 | ||
183 | This version of the urcu library uses memory barriers on the writer | |
184 | and reader sides. This results in faster grace-period detection, but | |
185 | results in slower reads. | |
186 | ||
187 | ||
188 | ### Usage of `liburcu-signal` | |
189 | ||
190 | 1. `#include <urcu.h>` | |
191 | 2. Compile any `_LGPL_SOURCE` code using this library with `-DRCU_SIGNAL` | |
192 | 3. Link the application with `-lurcu-signal` | |
193 | ||
194 | Version of the library that requires a signal, typically `SIGUSR1`. Can | |
195 | be overridden with `-DSIGRCU` by modifying `Makefile.build.inc`. | |
196 | ||
197 | ||
198 | ### Usage of `liburcu-bp` | |
199 | ||
200 | 1. `#include <urcu-bp.h>` | |
201 | 2. Link with `-lurcu-bp` | |
202 | ||
203 | The BP library flavor stands for "bulletproof". It is specifically | |
204 | designed to help tracing library to hook on applications without | |
205 | requiring to modify these applications. `rcu_init()`, | |
206 | `rcu_register_thread()` and `rcu_unregister_thread()` all become nops. | |
207 | The state is dealt with by the library internally at the expense of | |
208 | read-side and write-side performance. | |
209 | ||
210 | ||
211 | ### Initialization | |
212 | ||
213 | Each thread that has reader critical sections (that uses | |
214 | `rcu_read_lock()`/`rcu_read_unlock()` must first register to the URCU | |
215 | library. This is done by calling `rcu_register_thread()`. Unregistration | |
216 | must be performed before exiting the thread by using | |
217 | `rcu_unregister_thread()`. | |
218 | ||
219 | ||
220 | ### Reading | |
221 | ||
222 | Reader critical sections must be protected by locating them between | |
223 | calls to `rcu_read_lock()` and `rcu_read_unlock()`. Inside that lock, | |
224 | `rcu_dereference()` may be called to read an RCU protected pointer. | |
225 | ||
226 | ||
227 | ### Writing | |
228 | ||
229 | `rcu_assign_pointer()` and `rcu_xchg_pointer()` may be called anywhere. | |
230 | After, `synchronize_rcu()` must be called. When it returns, the old | |
231 | values are not in usage anymore. | |
232 | ||
233 | ||
234 | ### Usage of `liburcu-defer` | |
235 | ||
236 | - Follow instructions for either `liburcu`, `liburcu-qsbr`, | |
237 | `liburcu-mb`, `liburcu-signal`, or `liburcu-bp` above. | |
238 | The `liburcu-defer` functionality is pulled into each of | |
239 | those library modules. | |
240 | - Provides `defer_rcu()` primitive to enqueue delayed callbacks. Queued | |
241 | callbacks are executed in batch periodically after a grace period. | |
242 | Do _not_ use `defer_rcu()` within a read-side critical section, because | |
243 | it may call `synchronize_rcu()` if the thread queue is full. | |
244 | This can lead to deadlock or worse. | |
245 | - Requires that `rcu_defer_barrier()` must be called in library destructor | |
246 | if a library queues callbacks and is expected to be unloaded with | |
247 | `dlclose()`. | |
248 | ||
249 | Its API is currently experimental. It may change in future library releases. | |
250 | ||
251 | ||
252 | ### Usage of `urcu-call-rcu` | |
253 | ||
254 | - Follow instructions for either `liburcu`, `liburcu-qsbr`, | |
255 | `liburcu-mb`, `liburcu-signal`, or `liburcu-bp` above. | |
256 | The `urcu-call-rcu` functionality is pulled into each of | |
257 | those library modules. | |
258 | - Provides the `call_rcu()` primitive to enqueue delayed callbacks | |
259 | in a manner similar to `defer_rcu()`, but without ever delaying | |
260 | for a grace period. On the other hand, `call_rcu()`'s best-case | |
261 | overhead is not quite as good as that of `defer_rcu()`. | |
262 | - Provides `call_rcu()` to allow asynchronous handling of RCU | |
263 | grace periods. A number of additional functions are provided | |
264 | to manage the helper threads used by `call_rcu()`, but reasonable | |
265 | defaults are used if these additional functions are not invoked. | |
dcb9c05a | 266 | See [`doc/rcu-api.md`](doc/rcu-api.md) in userspace-rcu documentation |
d589a916 PP |
267 | for more details. |
268 | ||
269 | ||
270 | ### Being careful with signals | |
271 | ||
272 | The `liburcu` library uses signals internally. The signal handler is | |
273 | registered with the `SA_RESTART` flag. However, these signals may cause | |
274 | some non-restartable system calls to fail with `errno = EINTR`. Care | |
275 | should be taken to restart system calls manually if they fail with this | |
276 | error. A list of non-restartable system calls may be found in | |
277 | `signal(7)`. The `liburcu-mb` and `liburcu-qsbr` versions of the Userspace RCU | |
278 | library do not require any signal. | |
279 | ||
280 | Read-side critical sections are allowed in a signal handler, | |
281 | except those setup with `sigaltstack(2)`, with `liburcu` and | |
282 | `liburcu-mb`. Be careful, however, to disable these signals | |
283 | between thread creation and calls to `rcu_register_thread()`, because a | |
284 | signal handler nesting on an unregistered thread would not be | |
285 | allowed to call `rcu_read_lock()`. | |
286 | ||
287 | Read-side critical sections are _not_ allowed in a signal handler with | |
288 | `liburcu-qsbr`, unless signals are disabled explicitly around each | |
289 | `rcu_quiescent_state()` calls, when threads are put offline and around | |
290 | calls to `synchronize_rcu()`. Even then, we do not recommend it. | |
291 | ||
292 | ||
293 | ### Interaction with mutexes | |
294 | ||
295 | One must be careful to do not cause deadlocks due to interaction of | |
296 | `synchronize_rcu()` and RCU read-side with mutexes. If `synchronize_rcu()` | |
297 | is called with a mutex held, this mutex (or any mutex which has this | |
298 | mutex in its dependency chain) should not be acquired from within a RCU | |
299 | read-side critical section. | |
300 | ||
301 | This is especially important to understand in the context of the | |
302 | QSBR flavor: a registered reader thread being "online" by | |
303 | default should be considered as within a RCU read-side critical | |
304 | section unless explicitly put "offline". Therefore, if | |
305 | `synchronize_rcu()` is called with a mutex held, this mutex, as | |
306 | well as any mutex which has this mutex in its dependency chain | |
307 | should only be taken when the RCU reader thread is "offline" | |
308 | (this can be performed by calling `rcu_thread_offline()`). | |
309 | ||
310 | ||
311 | ### Interaction with `fork()` | |
312 | ||
313 | Special care must be taken for applications performing `fork()` without | |
314 | any following `exec()`. This is caused by the fact that Linux only clones | |
315 | the thread calling `fork()`, and thus never replicates any of the other | |
316 | parent thread into the child process. Most `liburcu` implementations | |
317 | require that all registrations (as reader, `defer_rcu` and `call_rcu` | |
318 | threads) should be released before a `fork()` is performed, except for the | |
319 | rather common scenario where `fork()` is immediately followed by `exec()` in | |
320 | the child process. The only implementation not subject to that rule is | |
321 | `liburcu-bp`, which is designed to handle `fork()` by calling | |
322 | `rcu_bp_before_fork`, `rcu_bp_after_fork_parent` and | |
323 | `rcu_bp_after_fork_child`. | |
324 | ||
325 | Applications that use `call_rcu()` and that `fork()` without | |
326 | doing an immediate `exec()` must take special action. The parent | |
327 | must invoke `call_rcu_before_fork()` before the `fork()` and | |
328 | `call_rcu_after_fork_parent()` after the `fork()`. The child | |
329 | process must invoke `call_rcu_after_fork_child()`. | |
330 | Even though these three APIs are suitable for passing to | |
331 | `pthread_atfork()`, use of `pthread_atfork()` is **STRONGLY | |
332 | DISCOURAGED** for programs calling the glibc memory allocator | |
333 | (`malloc()`, `calloc()`, `free()`, ...) within `call_rcu` callbacks. | |
334 | This is due to limitations in the way glibc memory allocator | |
335 | handles calls to the memory allocator from concurrent threads | |
336 | while the `pthread_atfork()` handlers are executing. | |
337 | ||
338 | Combining e.g.: | |
339 | ||
340 | - call to `free()` from callbacks executed within `call_rcu` worker | |
341 | threads, | |
342 | - executing `call_rcu` atfork handlers within the glibc pthread | |
343 | atfork mechanism, | |
344 | ||
345 | will sometimes trigger interesting process hangs. This usually | |
346 | hangs on a memory allocator lock within glibc. | |
347 | ||
348 | ||
349 | ### Thread Local Storage (TLS) | |
350 | ||
351 | Userspace RCU can fall back on `pthread_getspecific()` to emulate | |
352 | TLS variables on systems where it is not available. This behavior | |
353 | can be forced by specifying `--disable-compiler-tls` as configure | |
354 | argument. | |
355 | ||
356 | ||
357 | ### Usage of `DEBUG_RCU` | |
358 | ||
359 | `DEBUG_RCU` is used to add internal debugging self-checks to the | |
360 | RCU library. This define adds a performance penalty when enabled. | |
361 | Can be enabled by uncommenting the corresponding line in | |
362 | `Makefile.build.inc`. | |
363 | ||
364 | ||
365 | ### Usage of `DEBUG_YIELD` | |
366 | ||
367 | `DEBUG_YIELD` is used to add random delays in the code for testing | |
368 | purposes. | |
369 | ||
370 | ||
371 | ### SMP support | |
372 | ||
373 | By default the library is configured to use synchronization primitives | |
374 | adequate for SMP systems. On uniprocessor systems, support for SMP | |
375 | systems can be disabled with: | |
376 | ||
377 | ./configure --disable-smp-support | |
378 | ||
379 | theoretically yielding slightly better performance. | |
380 | ||
381 | ||
382 | Make targets | |
383 | ------------ | |
384 | ||
385 | In addition to the usual `make check` target, Userspace RCU features | |
386 | `make regtest` and `make bench` targets: | |
387 | ||
388 | - `make check`: short tests, meant to be run when rebuilding or | |
389 | porting Userspace RCU. | |
390 | - `make regtest`: long (many hours) test, meant to be run when | |
391 | modifying Userspace RCU or porting it to a new architecture or | |
392 | operating system. | |
393 | - `make bench`: long (many hours) benchmarks. | |
394 | ||
395 | ||
396 | Contacts | |
397 | -------- | |
398 | ||
399 | You can contact the maintainers on the following mailing list: | |
dcb9c05a | 400 | `lttng-dev@lists.lttng.org`. |