Commit | Line | Data |
---|---|---|
fdee2e6d MD |
1 | /* |
2 | * urcu-bp.c | |
3 | * | |
4 | * Userspace RCU library, "bulletproof" version. | |
5 | * | |
6982d6d7 | 6 | * Copyright (c) 2009 Mathieu Desnoyers <mathieu.desnoyers@efficios.com> |
fdee2e6d MD |
7 | * Copyright (c) 2009 Paul E. McKenney, IBM Corporation. |
8 | * | |
9 | * This library is free software; you can redistribute it and/or | |
10 | * modify it under the terms of the GNU Lesser General Public | |
11 | * License as published by the Free Software Foundation; either | |
12 | * version 2.1 of the License, or (at your option) any later version. | |
13 | * | |
14 | * This library is distributed in the hope that it will be useful, | |
15 | * but WITHOUT ANY WARRANTY; without even the implied warranty of | |
16 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU | |
17 | * Lesser General Public License for more details. | |
18 | * | |
19 | * You should have received a copy of the GNU Lesser General Public | |
20 | * License along with this library; if not, write to the Free Software | |
21 | * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA | |
22 | * | |
23 | * IBM's contributions to this file may be relicensed under LGPLv2 or later. | |
24 | */ | |
25 | ||
0617bf4c | 26 | #define _GNU_SOURCE |
fdee2e6d MD |
27 | #include <stdio.h> |
28 | #include <pthread.h> | |
29 | #include <signal.h> | |
30 | #include <assert.h> | |
31 | #include <stdlib.h> | |
32 | #include <string.h> | |
33 | #include <errno.h> | |
34 | #include <poll.h> | |
35 | #include <unistd.h> | |
36 | #include <sys/mman.h> | |
37 | ||
38 | #include "urcu-bp-static.h" | |
39 | /* Do not #define _LGPL_SOURCE to ensure we can emit the wrapper symbols */ | |
40 | #include "urcu-bp.h" | |
41 | ||
42 | /* Sleep delay in us */ | |
43 | #define RCU_SLEEP_DELAY 1000 | |
44 | #define ARENA_INIT_ALLOC 16 | |
45 | ||
02be5561 | 46 | void __attribute__((destructor)) rcu_bp_exit(void); |
fdee2e6d | 47 | |
6abb4bd5 | 48 | static pthread_mutex_t rcu_gp_lock = PTHREAD_MUTEX_INITIALIZER; |
fdee2e6d MD |
49 | |
50 | #ifdef DEBUG_YIELD | |
51 | unsigned int yield_active; | |
52 | unsigned int __thread rand_yield; | |
53 | #endif | |
54 | ||
55 | /* | |
56 | * Global grace period counter. | |
02be5561 | 57 | * Contains the current RCU_GP_CTR_PHASE. |
fdee2e6d MD |
58 | * Also has a RCU_GP_COUNT of 1, to accelerate the reader fast path. |
59 | * Written to only by writer with mutex taken. Read by both writer and readers. | |
60 | */ | |
02be5561 | 61 | long rcu_gp_ctr = RCU_GP_COUNT; |
fdee2e6d MD |
62 | |
63 | /* | |
64 | * Pointer to registry elements. Written to only by each individual reader. Read | |
65 | * by both the reader and the writers. | |
66 | */ | |
02be5561 | 67 | struct rcu_reader __thread *rcu_reader; |
fdee2e6d | 68 | |
16aa9ee8 | 69 | static CDS_LIST_HEAD(registry); |
fdee2e6d MD |
70 | |
71 | struct registry_arena { | |
72 | void *p; | |
73 | size_t len; | |
74 | size_t used; | |
75 | }; | |
76 | ||
77 | static struct registry_arena registry_arena; | |
78 | ||
4cf1675f MD |
79 | /* Saved fork signal mask, protected by rcu_gp_lock */ |
80 | static sigset_t saved_fork_signal_mask; | |
81 | ||
fdee2e6d MD |
82 | static void rcu_gc_registry(void); |
83 | ||
6abb4bd5 | 84 | static void mutex_lock(pthread_mutex_t *mutex) |
fdee2e6d MD |
85 | { |
86 | int ret; | |
87 | ||
88 | #ifndef DISTRUST_SIGNALS_EXTREME | |
6abb4bd5 | 89 | ret = pthread_mutex_lock(mutex); |
fdee2e6d MD |
90 | if (ret) { |
91 | perror("Error in pthread mutex lock"); | |
92 | exit(-1); | |
93 | } | |
94 | #else /* #ifndef DISTRUST_SIGNALS_EXTREME */ | |
6abb4bd5 | 95 | while ((ret = pthread_mutex_trylock(mutex)) != 0) { |
fdee2e6d MD |
96 | if (ret != EBUSY && ret != EINTR) { |
97 | printf("ret = %d, errno = %d\n", ret, errno); | |
98 | perror("Error in pthread mutex lock"); | |
99 | exit(-1); | |
100 | } | |
fdee2e6d MD |
101 | poll(NULL,0,10); |
102 | } | |
103 | #endif /* #else #ifndef DISTRUST_SIGNALS_EXTREME */ | |
104 | } | |
105 | ||
6abb4bd5 | 106 | static void mutex_unlock(pthread_mutex_t *mutex) |
fdee2e6d MD |
107 | { |
108 | int ret; | |
109 | ||
6abb4bd5 | 110 | ret = pthread_mutex_unlock(mutex); |
fdee2e6d MD |
111 | if (ret) { |
112 | perror("Error in pthread mutex unlock"); | |
113 | exit(-1); | |
114 | } | |
115 | } | |
116 | ||
2dfb8b5e | 117 | void update_counter_and_wait(void) |
fdee2e6d | 118 | { |
16aa9ee8 | 119 | CDS_LIST_HEAD(qsreaders); |
fdee2e6d | 120 | int wait_loops = 0; |
02be5561 | 121 | struct rcu_reader *index, *tmp; |
fdee2e6d | 122 | |
32c15e4e | 123 | /* Switch parity: 0 -> 1, 1 -> 0 */ |
6cf3827c | 124 | CMM_STORE_SHARED(rcu_gp_ctr, rcu_gp_ctr ^ RCU_GP_CTR_PHASE); |
2dfb8b5e MD |
125 | |
126 | /* | |
127 | * Must commit qparity update to memory before waiting for other parity | |
128 | * quiescent state. Failure to do so could result in the writer waiting | |
129 | * forever while new readers are always accessing data (no progress). | |
6cf3827c | 130 | * Ensured by CMM_STORE_SHARED and CMM_LOAD_SHARED. |
2dfb8b5e MD |
131 | */ |
132 | ||
133 | /* | |
5481ddb3 | 134 | * Adding a cmm_smp_mb() which is _not_ formally required, but makes the |
2dfb8b5e MD |
135 | * model easier to understand. It does not have a big performance impact |
136 | * anyway, given this is the write-side. | |
137 | */ | |
5481ddb3 | 138 | cmm_smp_mb(); |
2dfb8b5e | 139 | |
fdee2e6d | 140 | /* |
02be5561 | 141 | * Wait for each thread rcu_reader.ctr count to become 0. |
fdee2e6d MD |
142 | */ |
143 | for (;;) { | |
144 | wait_loops++; | |
16aa9ee8 | 145 | cds_list_for_each_entry_safe(index, tmp, ®istry, node) { |
fdee2e6d | 146 | if (!rcu_old_gp_ongoing(&index->ctr)) |
16aa9ee8 | 147 | cds_list_move(&index->node, &qsreaders); |
fdee2e6d MD |
148 | } |
149 | ||
16aa9ee8 | 150 | if (cds_list_empty(®istry)) { |
fdee2e6d MD |
151 | break; |
152 | } else { | |
153 | if (wait_loops == RCU_QS_ACTIVE_ATTEMPTS) | |
154 | usleep(RCU_SLEEP_DELAY); | |
155 | else | |
06f22bdb | 156 | caa_cpu_relax(); |
fdee2e6d MD |
157 | } |
158 | } | |
159 | /* put back the reader list in the registry */ | |
16aa9ee8 | 160 | cds_list_splice(&qsreaders, ®istry); |
fdee2e6d MD |
161 | } |
162 | ||
163 | void synchronize_rcu(void) | |
164 | { | |
165 | sigset_t newmask, oldmask; | |
166 | int ret; | |
167 | ||
168 | ret = sigemptyset(&newmask); | |
169 | assert(!ret); | |
170 | ret = pthread_sigmask(SIG_SETMASK, &newmask, &oldmask); | |
171 | assert(!ret); | |
172 | ||
6abb4bd5 | 173 | mutex_lock(&rcu_gp_lock); |
fdee2e6d | 174 | |
16aa9ee8 | 175 | if (cds_list_empty(®istry)) |
2dfb8b5e | 176 | goto out; |
fdee2e6d MD |
177 | |
178 | /* All threads should read qparity before accessing data structure | |
2dfb8b5e | 179 | * where new ptr points to. */ |
fdee2e6d | 180 | /* Write new ptr before changing the qparity */ |
5481ddb3 | 181 | cmm_smp_mb(); |
fdee2e6d | 182 | |
2dfb8b5e MD |
183 | /* Remove old registry elements */ |
184 | rcu_gc_registry(); | |
fdee2e6d MD |
185 | |
186 | /* | |
187 | * Wait for previous parity to be empty of readers. | |
188 | */ | |
2dfb8b5e | 189 | update_counter_and_wait(); /* 0 -> 1, wait readers in parity 0 */ |
fdee2e6d MD |
190 | |
191 | /* | |
5481ddb3 | 192 | * Adding a cmm_smp_mb() which is _not_ formally required, but makes the |
fdee2e6d MD |
193 | * model easier to understand. It does not have a big performance impact |
194 | * anyway, given this is the write-side. | |
195 | */ | |
5481ddb3 | 196 | cmm_smp_mb(); |
fdee2e6d | 197 | |
fdee2e6d | 198 | /* |
2dfb8b5e | 199 | * Wait for previous parity to be empty of readers. |
fdee2e6d | 200 | */ |
2dfb8b5e | 201 | update_counter_and_wait(); /* 1 -> 0, wait readers in parity 1 */ |
fdee2e6d MD |
202 | |
203 | /* | |
2dfb8b5e MD |
204 | * Finish waiting for reader threads before letting the old ptr being |
205 | * freed. | |
fdee2e6d | 206 | */ |
5481ddb3 | 207 | cmm_smp_mb(); |
2dfb8b5e | 208 | out: |
6abb4bd5 | 209 | mutex_unlock(&rcu_gp_lock); |
fdee2e6d MD |
210 | ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL); |
211 | assert(!ret); | |
212 | } | |
213 | ||
214 | /* | |
215 | * library wrappers to be used by non-LGPL compatible source code. | |
216 | */ | |
217 | ||
218 | void rcu_read_lock(void) | |
219 | { | |
220 | _rcu_read_lock(); | |
221 | } | |
222 | ||
223 | void rcu_read_unlock(void) | |
224 | { | |
225 | _rcu_read_unlock(); | |
226 | } | |
227 | ||
228 | /* | |
229 | * only grow for now. | |
230 | */ | |
231 | static void resize_arena(struct registry_arena *arena, size_t len) | |
232 | { | |
233 | void *new_arena; | |
234 | ||
0617bf4c MD |
235 | if (!arena->p) |
236 | new_arena = mmap(arena->p, len, | |
237 | PROT_READ | PROT_WRITE, | |
238 | MAP_ANONYMOUS | MAP_PRIVATE, | |
239 | -1, 0); | |
240 | else | |
241 | new_arena = mremap(arena->p, arena->len, | |
242 | len, MREMAP_MAYMOVE); | |
243 | assert(new_arena != MAP_FAILED); | |
244 | ||
fdee2e6d MD |
245 | /* |
246 | * re-used the same region ? | |
247 | */ | |
248 | if (new_arena == arena->p) | |
249 | return; | |
250 | ||
251 | memcpy(new_arena, arena->p, arena->len); | |
252 | bzero(new_arena + arena->len, len - arena->len); | |
253 | arena->p = new_arena; | |
254 | } | |
255 | ||
256 | /* Called with signals off and mutex locked */ | |
257 | static void add_thread(void) | |
258 | { | |
02be5561 | 259 | struct rcu_reader *rcu_reader_reg; |
fdee2e6d MD |
260 | |
261 | if (registry_arena.len | |
02be5561 | 262 | < registry_arena.used + sizeof(struct rcu_reader)) |
fdee2e6d MD |
263 | resize_arena(®istry_arena, |
264 | max(registry_arena.len << 1, ARENA_INIT_ALLOC)); | |
265 | /* | |
266 | * Find a free spot. | |
267 | */ | |
02be5561 MD |
268 | for (rcu_reader_reg = registry_arena.p; |
269 | (void *)rcu_reader_reg < registry_arena.p + registry_arena.len; | |
270 | rcu_reader_reg++) { | |
271 | if (!rcu_reader_reg->alloc) | |
fdee2e6d MD |
272 | break; |
273 | } | |
02be5561 MD |
274 | rcu_reader_reg->alloc = 1; |
275 | registry_arena.used += sizeof(struct rcu_reader); | |
fdee2e6d MD |
276 | |
277 | /* Add to registry */ | |
02be5561 MD |
278 | rcu_reader_reg->tid = pthread_self(); |
279 | assert(rcu_reader_reg->ctr == 0); | |
16aa9ee8 | 280 | cds_list_add(&rcu_reader_reg->node, ®istry); |
02be5561 | 281 | rcu_reader = rcu_reader_reg; |
fdee2e6d MD |
282 | } |
283 | ||
284 | /* Called with signals off and mutex locked */ | |
285 | static void rcu_gc_registry(void) | |
286 | { | |
02be5561 | 287 | struct rcu_reader *rcu_reader_reg; |
fdee2e6d MD |
288 | pthread_t tid; |
289 | int ret; | |
290 | ||
02be5561 MD |
291 | for (rcu_reader_reg = registry_arena.p; |
292 | (void *)rcu_reader_reg < registry_arena.p + registry_arena.len; | |
293 | rcu_reader_reg++) { | |
294 | if (!rcu_reader_reg->alloc) | |
fdee2e6d | 295 | continue; |
02be5561 | 296 | tid = rcu_reader_reg->tid; |
fdee2e6d MD |
297 | ret = pthread_kill(tid, 0); |
298 | assert(ret != EINVAL); | |
299 | if (ret == ESRCH) { | |
16aa9ee8 | 300 | cds_list_del(&rcu_reader_reg->node); |
79266051 | 301 | rcu_reader_reg->ctr = 0; |
02be5561 MD |
302 | rcu_reader_reg->alloc = 0; |
303 | registry_arena.used -= sizeof(struct rcu_reader); | |
fdee2e6d MD |
304 | } |
305 | } | |
306 | } | |
307 | ||
308 | /* Disable signals, take mutex, add to registry */ | |
309 | void rcu_bp_register(void) | |
310 | { | |
311 | sigset_t newmask, oldmask; | |
312 | int ret; | |
313 | ||
314 | ret = sigemptyset(&newmask); | |
315 | assert(!ret); | |
316 | ret = pthread_sigmask(SIG_SETMASK, &newmask, &oldmask); | |
317 | assert(!ret); | |
318 | ||
319 | /* | |
320 | * Check if a signal concurrently registered our thread since | |
321 | * the check in rcu_read_lock(). */ | |
02be5561 | 322 | if (rcu_reader) |
fdee2e6d MD |
323 | goto end; |
324 | ||
6abb4bd5 | 325 | mutex_lock(&rcu_gp_lock); |
fdee2e6d | 326 | add_thread(); |
6abb4bd5 | 327 | mutex_unlock(&rcu_gp_lock); |
fdee2e6d MD |
328 | end: |
329 | ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL); | |
330 | assert(!ret); | |
331 | } | |
332 | ||
02be5561 | 333 | void rcu_bp_exit() |
fdee2e6d MD |
334 | { |
335 | munmap(registry_arena.p, registry_arena.len); | |
336 | } | |
4cf1675f MD |
337 | |
338 | /* | |
339 | * Holding the rcu_gp_lock across fork will make sure we fork() don't race with | |
340 | * a concurrent thread executing with this same lock held. This ensures that the | |
341 | * registry is in a coherent state in the child. | |
342 | */ | |
343 | void rcu_bp_before_fork(void) | |
344 | { | |
345 | sigset_t newmask, oldmask; | |
346 | int ret; | |
347 | ||
348 | ret = sigemptyset(&newmask); | |
349 | assert(!ret); | |
350 | ret = pthread_sigmask(SIG_SETMASK, &newmask, &oldmask); | |
351 | assert(!ret); | |
352 | mutex_lock(&rcu_gp_lock); | |
353 | saved_fork_signal_mask = oldmask; | |
354 | } | |
355 | ||
356 | void rcu_bp_after_fork_parent(void) | |
357 | { | |
358 | sigset_t oldmask; | |
359 | int ret; | |
360 | ||
361 | oldmask = saved_fork_signal_mask; | |
362 | mutex_unlock(&rcu_gp_lock); | |
363 | ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL); | |
364 | assert(!ret); | |
365 | } | |
366 | ||
367 | void rcu_bp_after_fork_child(void) | |
368 | { | |
369 | sigset_t oldmask; | |
370 | int ret; | |
371 | ||
372 | rcu_gc_registry(); | |
373 | oldmask = saved_fork_signal_mask; | |
374 | mutex_unlock(&rcu_gp_lock); | |
375 | ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL); | |
376 | assert(!ret); | |
377 | } |