1 /*-*- Mode: C; c-basic-offset: 8; indent-tabs-mode: nil -*-*/
4 This file is part of systemd.
6 Copyright 2014 David Herrmann <dh.herrmann@gmail.com>
8 systemd is free software; you can redistribute it and/or modify it
9 under the terms of the GNU Lesser General Public License as published by
10 the Free Software Foundation; either version 2.1 of the License, or
11 (at your option) any later version.
13 systemd is distributed in the hope that it will be useful, but
14 WITHOUT ANY WARRANTY; without even the implied warranty of
15 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
16 Lesser General Public License for more details.
18 You should have received a copy of the GNU Lesser General Public License
19 along with systemd; If not, see <http://www.gnu.org/licenses/>.
31 #include <sys/eventfd.h>
32 #include <sys/types.h>
41 * This barrier implementation provides a simple synchronization method based
42 * on file-descriptors that can safely be used between threads and processes. A
43 * barrier object contains 2 shared counters based on eventfd. Both processes
44 * can now place barriers and wait for the other end to reach a random or
46 * Barriers are numbered, so you can either wait for the other end to reach any
47 * barrier or the last barrier that you placed. This way, you can use barriers
48 * for one-way *and* full synchronization. Note that even-though barriers are
49 * numbered, these numbers are internal and recycled once both sides reached the
50 * same barrier (implemented as a simple signed counter). It is thus not
51 * possible to address barriers by their ID.
53 * Barrier-API: Both ends can place as many barriers via barrier_place() as
54 * they want and each pair of barriers on both sides will be implicitly linked.
55 * Each side can use the barrier_wait/sync_*() family of calls to wait for the
56 * other side to place a specific barrier. barrier_wait_next() waits until the
57 * other side calls barrier_place(). No links between the barriers are
58 * considered and this simply serves as most basic asynchronous barrier.
59 * barrier_sync_next() is like barrier_wait_next() and waits for the other side
60 * to place their next barrier via barrier_place(). However, it only waits for
61 * barriers that are linked to a barrier we already placed. If the other side
62 * already placed more barriers than we did, barrier_sync_next() returns
64 * barrier_sync() extends barrier_sync_next() and waits until the other end
65 * placed as many barriers via barrier_place() as we did. If they already placed
66 * as many as we did (or more), it returns immediately.
68 * Additionally to basic barriers, an abortion event is available.
69 * barrier_abort() places an abortion event that cannot be undone. An abortion
70 * immediately cancels all placed barriers and replaces them. Any running and
71 * following wait/sync call besides barrier_wait_abortion() will immediately
72 * return false on both sides (otherwise, they always return true).
73 * barrier_abort() can be called multiple times on both ends and will be a
74 * no-op if already called on this side.
75 * barrier_wait_abortion() can be used to wait for the other side to call
76 * barrier_abort() and is the only wait/sync call that does not return
77 * immediately if we aborted outself. It only returns once the other side
78 * called barrier_abort().
80 * Barriers can be used for in-process and inter-process synchronization.
81 * However, for in-process synchronization you could just use mutexes.
82 * Therefore, main target is IPC and we require both sides to *not* share the FD
83 * table. If that's given, barriers provide target tracking: If the remote side
84 * exit()s, an abortion event is implicitly queued on the other side. This way,
85 * a sync/wait call will be woken up if the remote side crashed or exited
86 * unexpectedly. However, note that these abortion events are only queued if the
87 * barrier-queue has been drained. Therefore, it is safe to place a barrier and
88 * exit. The other side can safely wait on the barrier even though the exit
89 * queued an abortion event. Usually, the abortion event would overwrite the
90 * barrier, however, that's not true for exit-abortion events. Those are only
91 * queued if the barrier-queue is drained (thus, the receiving side has placed
92 * more barriers than the remote side).
96 * barrier_create() - Initialize a barrier object
97 * @obj: barrier to initialize
99 * This initializes a barrier object. The caller is responsible of allocating
100 * the memory and keeping it valid. The memory does not have to be zeroed
102 * Two eventfd objects are allocated for each barrier. If allocation fails, an
105 * If this function fails, the barrier is reset to an invalid state so it is
106 * safe to call barrier_destroy() on the object regardless whether the
107 * initialization succeeded or not.
109 * The caller is responsible to destroy the object via barrier_destroy() before
110 * releasing the underlying memory.
112 * Returns: 0 on success, negative error code on failure.
114 int barrier_create(Barrier *b) {
115 _cleanup_(barrier_destroyp) Barrier *staging = b;
120 b->me = eventfd(0, EFD_CLOEXEC | EFD_NONBLOCK);
124 b->them = eventfd(0, EFD_CLOEXEC | EFD_NONBLOCK);
128 r = pipe2(b->pipe, O_CLOEXEC | O_NONBLOCK);
137 * barrier_destroy() - Destroy a barrier object
138 * @b: barrier to destroy or NULL
140 * This destroys a barrier object that has previously been passed to
141 * barrier_create(). The object is released and reset to invalid
142 * state. Therefore, it is safe to call barrier_destroy() multiple
143 * times or even if barrier_create() failed. However, barrier must be
144 * always initalized with BARRIER_NULL.
146 * If @b is NULL, this is a no-op.
148 void barrier_destroy(Barrier *b) {
152 b->me = safe_close(b->me);
153 b->them = safe_close(b->them);
154 safe_close_pair(b->pipe);
159 * barrier_set_role() - Set the local role of the barrier
160 * @b: barrier to operate on
161 * @role: role to set on the barrier
163 * This sets the roles on a barrier object. This is needed to know
164 * which side of the barrier you're on. Usually, the parent creates
165 * the barrier via barrier_create() and then calls fork() or clone().
166 * Therefore, the FDs are duplicated and the child retains the same
169 * Both sides need to call barrier_set_role() after fork() or clone()
170 * are done. If this is not done, barriers will not work correctly.
172 * Note that barriers could be supported without fork() or clone(). However,
173 * this is currently not needed so it hasn't been implemented.
175 void barrier_set_role(Barrier *b, unsigned int role) {
179 assert(role == BARRIER_PARENT || role == BARRIER_CHILD);
180 /* make sure this is only called once */
181 assert(b->pipe[1] >= 0 && b->pipe[1] >= 0);
183 if (role == BARRIER_PARENT)
184 b->pipe[1] = safe_close(b->pipe[1]);
186 b->pipe[0] = safe_close(b->pipe[0]);
188 /* swap me/them for children */
195 /* places barrier; returns false if we aborted, otherwise true */
196 static bool barrier_write(Barrier *b, uint64_t buf) {
199 /* prevent new sync-points if we already aborted */
200 if (barrier_i_aborted(b))
204 len = write(b->me, &buf, sizeof(buf));
205 } while (len < 0 && IN_SET(errno, EAGAIN, EINTR));
207 if (len != sizeof(buf))
210 /* lock if we aborted */
211 if (buf >= (uint64_t)BARRIER_ABORTION) {
212 if (barrier_they_aborted(b))
213 b->barriers = BARRIER_WE_ABORTED;
215 b->barriers = BARRIER_I_ABORTED;
216 } else if (!barrier_is_aborted(b))
219 return !barrier_i_aborted(b);
222 /* If there is an unexpected error, we have to make this fatal. There
223 * is no way we can recover from sync-errors. Therefore, we close the
224 * pipe-ends and treat this as abortion. The other end will notice the
225 * pipe-close and treat it as abortion, too. */
227 safe_close_pair(b->pipe);
228 b->barriers = BARRIER_WE_ABORTED;
232 /* waits for barriers; returns false if they aborted, otherwise true */
233 static bool barrier_read(Barrier *b, int64_t comp) {
234 if (barrier_they_aborted(b))
237 while (b->barriers > comp) {
238 struct pollfd pfd[2] = {
239 { .fd = b->pipe[0] >= 0 ? b->pipe[0] : b->pipe[1],
246 r = poll(pfd, 2, -1);
247 if (r < 0 && IN_SET(errno, EAGAIN, EINTR))
252 if (pfd[1].revents) {
255 /* events on @them signal new data for us */
256 len = read(b->them, &buf, sizeof(buf));
257 if (len < 0 && IN_SET(errno, EAGAIN, EINTR))
260 if (len != sizeof(buf))
262 } else if (pfd[0].revents & (POLLHUP | POLLERR | POLLNVAL))
263 /* POLLHUP on the pipe tells us the other side exited.
264 * We treat this as implicit abortion. But we only
265 * handle it if there's no event on the eventfd. This
266 * guarantees that exit-abortions do not overwrite real
268 buf = BARRIER_ABORTION;
272 /* lock if they aborted */
273 if (buf >= (uint64_t)BARRIER_ABORTION) {
274 if (barrier_i_aborted(b))
275 b->barriers = BARRIER_WE_ABORTED;
277 b->barriers = BARRIER_THEY_ABORTED;
278 } else if (!barrier_is_aborted(b))
282 return !barrier_they_aborted(b);
285 /* If there is an unexpected error, we have to make this fatal. There
286 * is no way we can recover from sync-errors. Therefore, we close the
287 * pipe-ends and treat this as abortion. The other end will notice the
288 * pipe-close and treat it as abortion, too. */
290 safe_close_pair(b->pipe);
291 b->barriers = BARRIER_WE_ABORTED;
296 * barrier_place() - Place a new barrier
299 * This places a new barrier on the barrier object. If either side already
300 * aborted, this is a no-op and returns "false". Otherwise, the barrier is
301 * placed and this returns "true".
303 * Returns: true if barrier was placed, false if either side aborted.
305 bool barrier_place(Barrier *b) {
308 if (barrier_is_aborted(b))
311 barrier_write(b, BARRIER_SINGLE);
316 * barrier_abort() - Abort the synchronization
317 * @b: barrier object to abort
319 * This aborts the barrier-synchronization. If barrier_abort() was already
320 * called on this side, this is a no-op. Otherwise, the barrier is put into the
321 * ABORT-state and will stay there. The other side is notified about the
322 * abortion. Any following attempt to place normal barriers or to wait on normal
323 * barriers will return immediately as "false".
325 * You can wait for the other side to call barrier_abort(), too. Use
326 * barrier_wait_abortion() for that.
328 * Returns: false if the other side already aborted, true otherwise.
330 bool barrier_abort(Barrier *b) {
333 barrier_write(b, BARRIER_ABORTION);
334 return !barrier_they_aborted(b);
338 * barrier_wait_next() - Wait for the next barrier of the other side
339 * @b: barrier to operate on
341 * This waits until the other side places its next barrier. This is independent
342 * of any barrier-links and just waits for any next barrier of the other side.
344 * If either side aborted, this returns false.
346 * Returns: false if either side aborted, true otherwise.
348 bool barrier_wait_next(Barrier *b) {
351 if (barrier_is_aborted(b))
354 barrier_read(b, b->barriers - 1);
355 return !barrier_is_aborted(b);
359 * barrier_wait_abortion() - Wait for the other side to abort
360 * @b: barrier to operate on
362 * This waits until the other side called barrier_abort(). This can be called
363 * regardless whether the local side already called barrier_abort() or not.
365 * If the other side has already aborted, this returns immediately.
367 * Returns: false if the local side aborted, true otherwise.
369 bool barrier_wait_abortion(Barrier *b) {
372 barrier_read(b, BARRIER_THEY_ABORTED);
373 return !barrier_i_aborted(b);
377 * barrier_sync_next() - Wait for the other side to place a next linked barrier
378 * @b: barrier to operate on
380 * This is like barrier_wait_next() and waits for the other side to call
381 * barrier_place(). However, this only waits for linked barriers. That means, if
382 * the other side already placed more barriers than (or as much as) we did, this
383 * returns immediately instead of waiting.
385 * If either side aborted, this returns false.
387 * Returns: false if either side aborted, true otherwise.
389 bool barrier_sync_next(Barrier *b) {
392 if (barrier_is_aborted(b))
395 barrier_read(b, MAX((int64_t)0, b->barriers - 1));
396 return !barrier_is_aborted(b);
400 * barrier_sync() - Wait for the other side to place as many barriers as we did
401 * @b: barrier to operate on
403 * This is like barrier_sync_next() but waits for the other side to call
404 * barrier_place() as often as we did (in total). If they already placed as much
405 * as we did (or more), this returns immediately instead of waiting.
407 * If either side aborted, this returns false.
409 * Returns: false if either side aborted, true otherwise.
411 bool barrier_sync(Barrier *b) {
414 if (barrier_is_aborted(b))
418 return !barrier_is_aborted(b);