1FUTEX(2) Linux Programmer's Manual FUTEX(2)
2
6 futex - fast user-space locking
7
9 #include <linux/futex.h>
10 #include <sys/time.h>
11 int futex(int *uaddr, int futex_op, int val,
13 const struct timespec *timeout, /* or: uint32_t val2 */
14 int *uaddr2, int val3);
15 Note: There is no glibc wrapper for this system call; see NOTES.
17
19 The futex() system call provides a method for waiting until a certain
20 condition becomes true. It is typically used as a blocking construct
21 in the context of shared-memory synchronization. When using futexes,
22 the majority of the synchronization operations are performed in user
23 space. A user-space program employs the futex() system call only when
24 it is likely that the program has to block for a longer time until the
25 condition becomes true. Other futex() operations can be used to wake
26 any processes or threads waiting for a particular condition.
27 A futex is a 32-bit value—referred to below as a futex word—whose
29 address is supplied to the futex() system call. (Futexes are 32 bits
30 in size on all platforms, including 64-bit systems.) All futex opera‐
31 tions are governed by this value. In order to share a futex between
32 processes, the futex is placed in a region of shared memory, created
33 using (for example) mmap(2) or shmat(2). (Thus, the futex word may
34 have different virtual addresses in different processes, but these
35 addresses all refer to the same location in physical memory.) In a
36 multithreaded program, it is sufficient to place the futex word in a
37 global variable shared by all threads.
38 When executing a futex operation that requests to block a thread, the
40 kernel will block only if the futex word has the value that the calling
41 thread supplied (as one of the arguments of the futex() call) as the
42 expected value of the futex word. The loading of the futex word's
43 value, the comparison of that value with the expected value, and the
44 actual blocking will happen atomically and will be totally ordered with
45 respect to concurrent operations performed by other threads on the same
46 futex word. Thus, the futex word is used to connect the synchroniza‐
47 tion in user space with the implementation of blocking by the kernel.
48 Analogously to an atomic compare-and-exchange operation that poten‐
49 tially changes shared memory, blocking via a futex is an atomic com‐
50 pare-and-block operation.
51 One use of futexes is for implementing locks. The state of the lock
53 (i.e., acquired or not acquired) can be represented as an atomically
54 accessed flag in shared memory. In the uncontended case, a thread can
55 access or modify the lock state with atomic instructions, for example
56 atomically changing it from not acquired to acquired using an atomic
57 compare-and-exchange instruction. (Such instructions are performed
58 entirely in user mode, and the kernel maintains no information about
59 the lock state.) On the other hand, a thread may be unable to acquire
60 a lock because it is already acquired by another thread. It then may
61 pass the lock's flag as a futex word and the value representing the
62 acquired state as the expected value to a futex() wait operation. This
63 futex() operation will block if and only if the lock is still acquired
64 (i.e., the value in the futex word still matches the "acquired state").
65 When releasing the lock, a thread has to first reset the lock state to
66 not acquired and then execute a futex operation that wakes threads
67 blocked on the lock flag used as a futex word (this can be further
68 optimized to avoid unnecessary wake-ups). See futex(7) for more detail
69 on how to use futexes.
70 Besides the basic wait and wake-up futex functionality, there are fur‐
72 ther futex operations aimed at supporting more complex use cases.
73 Note that no explicit initialization or destruction is necessary to use
75 futexes; the kernel maintains a futex (i.e., the kernel-internal imple‐
76 mentation artifact) only while operations such as FUTEX_WAIT, described
77 below, are being performed on a particular futex word.
78 Arguments
80 The uaddr argument points to the futex word. On all platforms, futexes
81 are four-byte integers that must be aligned on a four-byte boundary.
82 The operation to perform on the futex is specified in the futex_op
83 argument; val is a value whose meaning and purpose depends on futex_op.
84 The remaining arguments (timeout, uaddr2, and val3) are required only
86 for certain of the futex operations described below. Where one of
87 these arguments is not required, it is ignored.
88 For several blocking operations, the timeout argument is a pointer to a
90 timespec structure that specifies a timeout for the operation. How‐
91 ever, notwithstanding the prototype shown above, for some operations,
92 the least significant four bytes of this argument are instead used as
93 an integer whose meaning is determined by the operation. For these
94 operations, the kernel casts the timeout value first to unsigned long,
95 then to uint32_t, and in the remainder of this page, this argument is
96 referred to as val2 when interpreted in this fashion.
97 Where it is required, the uaddr2 argument is a pointer to a second
99 futex word that is employed by the operation.
100 The interpretation of the final integer argument, val3, depends on the
102 operation.
103 Futex operations
105 The futex_op argument consists of two parts: a command that specifies
106 the operation to be performed, bit-wise ORed with zero or more options
107 that modify the behaviour of the operation. The options that may be
108 included in futex_op are as follows:
109 FUTEX_PRIVATE_FLAG (since Linux 2.6.22)
111 This option bit can be employed with all futex operations. It
112 tells the kernel that the futex is process-private and not
113 shared with another process (i.e., it is being used for synchro‐
114 nization only between threads of the same process). This allows
115 the kernel to make some additional performance optimizations.
116 As a convenience, <linux/futex.h> defines a set of constants
118 with the suffix _PRIVATE that are equivalents of all of the
119 operations listed below, but with the FUTEX_PRIVATE_FLAG ORed
120 into the constant value. Thus, there are FUTEX_WAIT_PRIVATE,
121 FUTEX_WAKE_PRIVATE, and so on.
122 FUTEX_CLOCK_REALTIME (since Linux 2.6.28)
124 This option bit can be employed only with the FUTEX_WAIT_BITSET,
125 FUTEX_WAIT_REQUEUE_PI, and (since Linux 4.5) FUTEX_WAIT opera‐
126 tions.
127 If this option is set, the kernel measures the timeout against
129 the CLOCK_REALTIME clock.
130 If this option is not set, the kernel measures the timeout
132 against the CLOCK_MONOTONIC clock.
133 The operation specified in futex_op is one of the following:
135 FUTEX_WAIT (since Linux 2.6.0)
137 This operation tests that the value at the futex word pointed to
138 by the address uaddr still contains the expected value val, and
139 if so, then sleeps waiting for a FUTEX_WAKE operation on the
140 futex word. The load of the value of the futex word is an
141 atomic memory access (i.e., using atomic machine instructions of
142 the respective architecture). This load, the comparison with
143 the expected value, and starting to sleep are performed atomi‐
144 cally and totally ordered with respect to other futex operations
145 on the same futex word. If the thread starts to sleep, it is
146 considered a waiter on this futex word. If the futex value does
147 not match val, then the call fails immediately with the error
148 EAGAIN.
149 The purpose of the comparison with the expected value is to pre‐
151 vent lost wake-ups. If another thread changed the value of the
152 futex word after the calling thread decided to block based on
153 the prior value, and if the other thread executed a FUTEX_WAKE
154 operation (or similar wake-up) after the value change and before
155 this FUTEX_WAIT operation, then the calling thread will observe
156 the value change and will not start to sleep.
157 If the timeout is not NULL, the structure it points to specifies
159 a timeout for the wait. (This interval will be rounded up to
160 the system clock granularity, and is guaranteed not to expire
161 early.) The timeout is by default measured according to the
162 CLOCK_MONOTONIC clock, but, since Linux 4.5, the CLOCK_REALTIME
163 clock can be selected by specifying FUTEX_CLOCK_REALTIME in
164 futex_op. If timeout is NULL, the call blocks indefinitely.
165 Note: for FUTEX_WAIT, timeout is interpreted as a relative
167 value. This differs from other futex operations, where timeout
168 is interpreted as an absolute value. To obtain the equivalent
169 of FUTEX_WAIT with an absolute timeout, employ FUTEX_WAIT_BITSET
170 with val3 specified as FUTEX_BITSET_MATCH_ANY.
171 The arguments uaddr2 and val3 are ignored.
173 FUTEX_WAKE (since Linux 2.6.0)
175 This operation wakes at most val of the waiters that are waiting
176 (e.g., inside FUTEX_WAIT) on the futex word at the address
177 uaddr. Most commonly, val is specified as either 1 (wake up a
178 single waiter) or INT_MAX (wake up all waiters). No guarantee
179 is provided about which waiters are awoken (e.g., a waiter with
180 a higher scheduling priority is not guaranteed to be awoken in
181 preference to a waiter with a lower priority).
182 The arguments timeout, uaddr2, and val3 are ignored.
184 FUTEX_FD (from Linux 2.6.0 up to and including Linux 2.6.25)
186 This operation creates a file descriptor that is associated with
187 the futex at uaddr. The caller must close the returned file
188 descriptor after use. When another process or thread performs a
189 FUTEX_WAKE on the futex word, the file descriptor indicates as
190 being readable with select(2), poll(2), and epoll(7)
191 The file descriptor can be used to obtain asynchronous notifica‐
193 tions: if val is nonzero, then, when another process or thread
194 executes a FUTEX_WAKE, the caller will receive the signal number
195 that was passed in val.
196 The arguments timeout, uaddr2 and val3 are ignored.
198 Because it was inherently racy, FUTEX_FD has been removed from
200 Linux 2.6.26 onward.
201 FUTEX_REQUEUE (since Linux 2.6.0)
203 This operation performs the same task as FUTEX_CMP_REQUEUE (see
204 below), except that no check is made using the value in val3.
205 (The argument val3 is ignored.)
206 FUTEX_CMP_REQUEUE (since Linux 2.6.7)
208 This operation first checks whether the location uaddr still
209 contains the value val3. If not, the operation fails with the
210 error EAGAIN. Otherwise, the operation wakes up a maximum of
211 val waiters that are waiting on the futex at uaddr. If there
212 are more than val waiters, then the remaining waiters are
213 removed from the wait queue of the source futex at uaddr and
214 added to the wait queue of the target futex at uaddr2. The val2
215 argument specifies an upper limit on the number of waiters that
216 are requeued to the futex at uaddr2.
217 The load from uaddr is an atomic memory access (i.e., using
219 atomic machine instructions of the respective architecture).
220 This load, the comparison with val3, and the requeueing of any
221 waiters are performed atomically and totally ordered with
222 respect to other operations on the same futex word.
223 Typical values to specify for val are 0 or 1. (Specifying
225 INT_MAX is not useful, because it would make the
226 FUTEX_CMP_REQUEUE operation equivalent to FUTEX_WAKE.) The
227 limit value specified via val2 is typically either 1 or INT_MAX.
228 (Specifying the argument as 0 is not useful, because it would
229 make the FUTEX_CMP_REQUEUE operation equivalent to FUTEX_WAIT.)
230 The FUTEX_CMP_REQUEUE operation was added as a replacement for
232 the earlier FUTEX_REQUEUE. The difference is that the check of
233 the value at uaddr can be used to ensure that requeueing happens
234 only under certain conditions, which allows race conditions to
235 be avoided in certain use cases.
236 Both FUTEX_REQUEUE and FUTEX_CMP_REQUEUE can be used to avoid
238 "thundering herd" wake-ups that could occur when using
239 FUTEX_WAKE in cases where all of the waiters that are woken need
240 to acquire another futex. Consider the following scenario,
241 where multiple waiter threads are waiting on B, a wait queue
242 implemented using a futex:
243 lock(A)
245 while (!check_value(V)) {
246 unlock(A);
247 block_on(B);
248 lock(A);
249 };
250 unlock(A);
251 If a waker thread used FUTEX_WAKE, then all waiters waiting on B
253 would be woken up, and they would all try to acquire lock A.
254 However, waking all of the threads in this manner would be
255 pointless because all except one of the threads would immedi‐
256 ately block on lock A again. By contrast, a requeue operation
257 wakes just one waiter and moves the other waiters to lock A, and
258 when the woken waiter unlocks A then the next waiter can pro‐
259 ceed.
260 FUTEX_WAKE_OP (since Linux 2.6.14)
262 This operation was added to support some user-space use cases
263 where more than one futex must be handled at the same time. The
264 most notable example is the implementation of pthread_cond_sig‐
265 nal(3), which requires operations on two futexes, the one used
266 to implement the mutex and the one used in the implementation of
267 the wait queue associated with the condition variable.
268 FUTEX_WAKE_OP allows such cases to be implemented without lead‐
269 ing to high rates of contention and context switching.
270 The FUTEX_WAKE_OP operation is equivalent to executing the fol‐
272 lowing code atomically and totally ordered with respect to other
273 futex operations on any of the two supplied futex words:
274 int oldval = *(int *) uaddr2;
276 *(int *) uaddr2 = oldval op oparg;
277 futex(uaddr, FUTEX_WAKE, val, 0, 0, 0);
278 if (oldval cmp cmparg)
279 futex(uaddr2, FUTEX_WAKE, val2, 0, 0, 0);
280 In other words, FUTEX_WAKE_OP does the following:
282 * saves the original value of the futex word at uaddr2 and per‐
284 forms an operation to modify the value of the futex at
285 uaddr2; this is an atomic read-modify-write memory access
286 (i.e., using atomic machine instructions of the respective
287 architecture)
288 * wakes up a maximum of val waiters on the futex for the futex
290 word at uaddr; and
291 * dependent on the results of a test of the original value of
293 the futex word at uaddr2, wakes up a maximum of val2 waiters
294 on the futex for the futex word at uaddr2.
295 The operation and comparison that are to be performed are
297 encoded in the bits of the argument val3. Pictorially, the
298 encoding is:
299 +---+---+-----------+-----------+
301 |op |cmp| oparg | cmparg |
302 +---+---+-----------+-----------+
303 4 4 12 12 <== # of bits
304 Expressed in code, the encoding is:
306 #define FUTEX_OP(op, oparg, cmp, cmparg) \
308 (((op & 0xf) << 28) | \
309 ((cmp & 0xf) << 24) | \
310 ((oparg & 0xfff) << 12) | \
311 (cmparg & 0xfff))
312 In the above, op and cmp are each one of the codes listed below.
314 The oparg and cmparg components are literal numeric values,
315 except as noted below.
316 The op component has one of the following values:
318 FUTEX_OP_SET 0 /* uaddr2 = oparg; */
320 FUTEX_OP_ADD 1 /* uaddr2 += oparg; */
321 FUTEX_OP_OR 2 /* uaddr2 |= oparg; */
322 FUTEX_OP_ANDN 3 /* uaddr2 &= ~oparg; */
323 FUTEX_OP_XOR 4 /* uaddr2 ^= oparg; */
324 In addition, bit-wise ORing the following value into op causes
326 (1 << oparg) to be used as the operand:
327 FUTEX_OP_ARG_SHIFT 8 /* Use (1 << oparg) as operand */
329 The cmp field is one of the following:
331 FUTEX_OP_CMP_EQ 0 /* if (oldval == cmparg) wake */
333 FUTEX_OP_CMP_NE 1 /* if (oldval != cmparg) wake */
334 FUTEX_OP_CMP_LT 2 /* if (oldval < cmparg) wake */
335 FUTEX_OP_CMP_LE 3 /* if (oldval <= cmparg) wake */
336 FUTEX_OP_CMP_GT 4 /* if (oldval > cmparg) wake */
337 FUTEX_OP_CMP_GE 5 /* if (oldval >= cmparg) wake */
338 The return value of FUTEX_WAKE_OP is the sum of the number of
340 waiters woken on the futex uaddr plus the number of waiters
341 woken on the futex uaddr2.
342 FUTEX_WAIT_BITSET (since Linux 2.6.25)
344 This operation is like FUTEX_WAIT except that val3 is used to
345 provide a 32-bit bit mask to the kernel. This bit mask, in
346 which at least one bit must be set, is stored in the kernel-
347 internal state of the waiter. See the description of
348 FUTEX_WAKE_BITSET for further details.
349 If timeout is not NULL, the structure it points to specifies an
351 absolute timeout for the wait operation. If timeout is NULL,
352 the operation can block indefinitely.
353 The uaddr2 argument is ignored.
355 FUTEX_WAKE_BITSET (since Linux 2.6.25)
357 This operation is the same as FUTEX_WAKE except that the val3
358 argument is used to provide a 32-bit bit mask to the kernel.
359 This bit mask, in which at least one bit must be set, is used to
360 select which waiters should be woken up. The selection is done
361 by a bit-wise AND of the "wake" bit mask (i.e., the value in
362 val3) and the bit mask which is stored in the kernel-internal
363 state of the waiter (the "wait" bit mask that is set using
364 FUTEX_WAIT_BITSET). All of the waiters for which the result of
365 the AND is nonzero are woken up; the remaining waiters are left
366 sleeping.
367 The effect of FUTEX_WAIT_BITSET and FUTEX_WAKE_BITSET is to
369 allow selective wake-ups among multiple waiters that are blocked
370 on the same futex. However, note that, depending on the use
371 case, employing this bit-mask multiplexing feature on a futex
372 can be less efficient than simply using multiple futexes,
373 because employing bit-mask multiplexing requires the kernel to
374 check all waiters on a futex, including those that are not
375 interested in being woken up (i.e., they do not have the rele‐
376 vant bit set in their "wait" bit mask).
377 The constant FUTEX_BITSET_MATCH_ANY, which corresponds to all 32
379 bits set in the bit mask, can be used as the val3 argument for
380 FUTEX_WAIT_BITSET and FUTEX_WAKE_BITSET. Other than differences
381 in the handling of the timeout argument, the FUTEX_WAIT opera‐
382 tion is equivalent to FUTEX_WAIT_BITSET with val3 specified as
383 FUTEX_BITSET_MATCH_ANY; that is, allow a wake-up by any waker.
384 The FUTEX_WAKE operation is equivalent to FUTEX_WAKE_BITSET with
385 val3 specified as FUTEX_BITSET_MATCH_ANY; that is, wake up any
386 waiter(s).
387 The uaddr2 and timeout arguments are ignored.
389 Priority-inheritance futexes
391 Linux supports priority-inheritance (PI) futexes in order to handle
392 priority-inversion problems that can be encountered with normal futex
393 locks. Priority inversion is the problem that occurs when a high-pri‐
394 ority task is blocked waiting to acquire a lock held by a low-priority
395 task, while tasks at an intermediate priority continuously preempt the
396 low-priority task from the CPU. Consequently, the low-priority task
397 makes no progress toward releasing the lock, and the high-priority task
398 remains blocked.
399 Priority inheritance is a mechanism for dealing with the priority-
401 inversion problem. With this mechanism, when a high-priority task
402 becomes blocked by a lock held by a low-priority task, the priority of
403 the low-priority task is temporarily raised to that of the high-prior‐
404 ity task, so that it is not preempted by any intermediate level tasks,
405 and can thus make progress toward releasing the lock. To be effective,
406 priority inheritance must be transitive, meaning that if a high-prior‐
407 ity task blocks on a lock held by a lower-priority task that is itself
408 blocked by a lock held by another intermediate-priority task (and so
409 on, for chains of arbitrary length), then both of those tasks (or more
410 generally, all of the tasks in a lock chain) have their priorities
411 raised to be the same as the high-priority task.
412 From a user-space perspective, what makes a futex PI-aware is a policy
414 agreement (described below) between user space and the kernel about the
415 value of the futex word, coupled with the use of the PI-futex opera‐
416 tions described below. (Unlike the other futex operations described
417 above, the PI-futex operations are designed for the implementation of
418 very specific IPC mechanisms.)
419 The PI-futex operations described below differ from the other futex
421 operations in that they impose policy on the use of the value of the
422 futex word:
423 * If the lock is not acquired, the futex word's value shall be 0.
425 * If the lock is acquired, the futex word's value shall be the thread
427 ID (TID; see gettid(2)) of the owning thread.
428 * If the lock is owned and there are threads contending for the lock,
430 then the FUTEX_WAITERS bit shall be set in the futex word's value;
431 in other words, this value is:
432 FUTEX_WAITERS | TID
434 (Note that is invalid for a PI futex word to have no owner and
436 FUTEX_WAITERS set.)
437 With this policy in place, a user-space application can acquire an
439 unacquired lock or release a lock using atomic instructions executed in
440 user mode (e.g., a compare-and-swap operation such as cmpxchg on the
441 x86 architecture). Acquiring a lock simply consists of using compare-
442 and-swap to atomically set the futex word's value to the caller's TID
443 if its previous value was 0. Releasing a lock requires using compare-
444 and-swap to set the futex word's value to 0 if the previous value was
445 the expected TID.
446 If a futex is already acquired (i.e., has a nonzero value), waiters
448 must employ the FUTEX_LOCK_PI operation to acquire the lock. If other
449 threads are waiting for the lock, then the FUTEX_WAITERS bit is set in
450 the futex value; in this case, the lock owner must employ the
451 FUTEX_UNLOCK_PI operation to release the lock.
452 In the cases where callers are forced into the kernel (i.e., required
454 to perform a futex() call), they then deal directly with a so-called
455 RT-mutex, a kernel locking mechanism which implements the required pri‐
456 ority-inheritance semantics. After the RT-mutex is acquired, the futex
457 value is updated accordingly, before the calling thread returns to user
458 space.
459 It is important to note that the kernel will update the futex word's
461 value prior to returning to user space. (This prevents the possibility
462 of the futex word's value ending up in an invalid state, such as having
463 an owner but the value being 0, or having waiters but not having the
464 FUTEX_WAITERS bit set.)
465 If a futex has an associated RT-mutex in the kernel (i.e., there are
467 blocked waiters) and the owner of the futex/RT-mutex dies unexpectedly,
468 then the kernel cleans up the RT-mutex and hands it over to the next
469 waiter. This in turn requires that the user-space value is updated
470 accordingly. To indicate that this is required, the kernel sets the
471 FUTEX_OWNER_DIED bit in the futex word along with the thread ID of the
472 new owner. User space can detect this situation via the presence of
473 the FUTEX_OWNER_DIED bit and is then responsible for cleaning up the
474 stale state left over by the dead owner.
475 PI futexes are operated on by specifying one of the values listed below
477 in futex_op. Note that the PI futex operations must be used as paired
478 operations and are subject to some additional requirements:
479 * FUTEX_LOCK_PI and FUTEX_TRYLOCK_PI pair with FUTEX_UNLOCK_PI.
481 FUTEX_UNLOCK_PI must be called only on a futex owned by the calling
482 thread, as defined by the value policy, otherwise the error EPERM
483 results.
484 * FUTEX_WAIT_REQUEUE_PI pairs with FUTEX_CMP_REQUEUE_PI. This must be
486 performed from a non-PI futex to a distinct PI futex (or the error
487 EINVAL results). Additionally, val (the number of waiters to be
488 woken) must be 1 (or the error EINVAL results).
489 The PI futex operations are as follows:
491 FUTEX_LOCK_PI (since Linux 2.6.18)
493 This operation is used after an attempt to acquire the lock via
494 an atomic user-mode instruction failed because the futex word
495 has a nonzero value—specifically, because it contained the (PID-
496 namespace-specific) TID of the lock owner.
497 The operation checks the value of the futex word at the address
499 uaddr. If the value is 0, then the kernel tries to atomically
500 set the futex value to the caller's TID. If the futex word's
501 value is nonzero, the kernel atomically sets the FUTEX_WAITERS
502 bit, which signals the futex owner that it cannot unlock the
503 futex in user space atomically by setting the futex value to 0.
504 After that, the kernel:
505 1. Tries to find the thread which is associated with the owner
507 TID.
508 2. Creates or reuses kernel state on behalf of the owner. (If
510 this is the first waiter, there is no kernel state for this
511 futex, so kernel state is created by locking the RT-mutex and
512 the futex owner is made the owner of the RT-mutex. If there
513 are existing waiters, then the existing state is reused.)
514 3. Attaches the waiter to the futex (i.e., the waiter is
516 enqueued on the RT-mutex waiter list).
517 If more than one waiter exists, the enqueueing of the waiter is
519 in descending priority order. (For information on priority
520 ordering, see the discussion of the SCHED_DEADLINE, SCHED_FIFO,
521 and SCHED_RR scheduling policies in sched(7).) The owner inher‐
522 its either the waiter's CPU bandwidth (if the waiter is sched‐
523 uled under the SCHED_DEADLINE policy) or the waiter's priority
524 (if the waiter is scheduled under the SCHED_RR or SCHED_FIFO
525 policy). This inheritance follows the lock chain in the case of
526 nested locking and performs deadlock detection.
527 The timeout argument provides a timeout for the lock attempt.
529 If timeout is not NULL, the structure it points to specifies an
530 absolute timeout, measured against the CLOCK_REALTIME clock. If
531 timeout is NULL, the operation will block indefinitely.
532 The uaddr2, val, and val3 arguments are ignored.
534 FUTEX_TRYLOCK_PI (since Linux 2.6.18)
536 This operation tries to acquire the lock at uaddr. It is
537 invoked when a user-space atomic acquire did not succeed because
538 the futex word was not 0.
539 Because the kernel has access to more state information than
541 user space, acquisition of the lock might succeed if performed
542 by the kernel in cases where the futex word (i.e., the state
543 information accessible to use-space) contains stale state
544 (FUTEX_WAITERS and/or FUTEX_OWNER_DIED). This can happen when
545 the owner of the futex died. User space cannot handle this con‐
546 dition in a race-free manner, but the kernel can fix this up and
547 acquire the futex.
548 The uaddr2, val, timeout, and val3 arguments are ignored.
550 FUTEX_UNLOCK_PI (since Linux 2.6.18)
552 This operation wakes the top priority waiter that is waiting in
553 FUTEX_LOCK_PI on the futex address provided by the uaddr argu‐
554 ment.
555 This is called when the user-space value at uaddr cannot be
557 changed atomically from a TID (of the owner) to 0.
558 The uaddr2, val, timeout, and val3 arguments are ignored.
560 FUTEX_CMP_REQUEUE_PI (since Linux 2.6.31)
562 This operation is a PI-aware variant of FUTEX_CMP_REQUEUE. It
563 requeues waiters that are blocked via FUTEX_WAIT_REQUEUE_PI on
564 uaddr from a non-PI source futex (uaddr) to a PI target futex
565 (uaddr2).
566 As with FUTEX_CMP_REQUEUE, this operation wakes up a maximum of
568 val waiters that are waiting on the futex at uaddr. However,
569 for FUTEX_CMP_REQUEUE_PI, val is required to be 1 (since the
570 main point is to avoid a thundering herd). The remaining wait‐
571 ers are removed from the wait queue of the source futex at uaddr
572 and added to the wait queue of the target futex at uaddr2.
573 The val2 and val3 arguments serve the same purposes as for
575 FUTEX_CMP_REQUEUE.
576 FUTEX_WAIT_REQUEUE_PI (since Linux 2.6.31)
578 Wait on a non-PI futex at uaddr and potentially be requeued (via
579 a FUTEX_CMP_REQUEUE_PI operation in another task) onto a PI
580 futex at uaddr2. The wait operation on uaddr is the same as for
581 FUTEX_WAIT.
582 The waiter can be removed from the wait on uaddr without
584 requeueing on uaddr2 via a FUTEX_WAKE operation in another task.
585 In this case, the FUTEX_WAIT_REQUEUE_PI operation fails with the
586 error EAGAIN.
587 If timeout is not NULL, the structure it points to specifies an
589 absolute timeout for the wait operation. If timeout is NULL,
590 the operation can block indefinitely.
591 The val3 argument is ignored.
593 The FUTEX_WAIT_REQUEUE_PI and FUTEX_CMP_REQUEUE_PI were added to
595 support a fairly specific use case: support for priority-inheri‐
596 tance-aware POSIX threads condition variables. The idea is that
597 these operations should always be paired, in order to ensure
598 that user space and the kernel remain in sync. Thus, in the
599 FUTEX_WAIT_REQUEUE_PI operation, the user-space application pre-
600 specifies the target of the requeue that takes place in the
601 FUTEX_CMP_REQUEUE_PI operation.
602
604 In the event of an error (and assuming that futex() was invoked via
605 syscall(2)), all operations return -1 and set errno to indicate the
606 cause of the error.
607 The return value on success depends on the operation, as described in
609 the following list:
610 FUTEX_WAIT
612 Returns 0 if the caller was woken up. Note that a wake-up can
613 also be caused by common futex usage patterns in unrelated code
614 that happened to have previously used the futex word's memory
615 location (e.g., typical futex-based implementations of Pthreads
616 mutexes can cause this under some conditions). Therefore, call‐
617 ers should always conservatively assume that a return value of 0
618 can mean a spurious wake-up, and use the futex word's value
619 (i.e., the user-space synchronization scheme) to decide whether
620 to continue to block or not.
621 FUTEX_WAKE
623 Returns the number of waiters that were woken up.
624 FUTEX_FD
626 Returns the new file descriptor associated with the futex.
627 FUTEX_REQUEUE
629 Returns the number of waiters that were woken up.
630 FUTEX_CMP_REQUEUE
632 Returns the total number of waiters that were woken up or
633 requeued to the futex for the futex word at uaddr2. If this
634 value is greater than val, then the difference is the number of
635 waiters requeued to the futex for the futex word at uaddr2.
636 FUTEX_WAKE_OP
638 Returns the total number of waiters that were woken up. This is
639 the sum of the woken waiters on the two futexes for the futex
640 words at uaddr and uaddr2.
641 FUTEX_WAIT_BITSET
643 Returns 0 if the caller was woken up. See FUTEX_WAIT for how to
644 interpret this correctly in practice.
645 FUTEX_WAKE_BITSET
647 Returns the number of waiters that were woken up.
648 FUTEX_LOCK_PI
650 Returns 0 if the futex was successfully locked.
651 FUTEX_TRYLOCK_PI
653 Returns 0 if the futex was successfully locked.
654 FUTEX_UNLOCK_PI
656 Returns 0 if the futex was successfully unlocked.
657 FUTEX_CMP_REQUEUE_PI
659 Returns the total number of waiters that were woken up or
660 requeued to the futex for the futex word at uaddr2. If this
661 value is greater than val, then difference is the number of
662 waiters requeued to the futex for the futex word at uaddr2.
663 FUTEX_WAIT_REQUEUE_PI
665 Returns 0 if the caller was successfully requeued to the futex
666 for the futex word at uaddr2.
667
669 EACCES No read access to the memory of a futex word.
670 EAGAIN (FUTEX_WAIT, FUTEX_WAIT_BITSET, FUTEX_WAIT_REQUEUE_PI) The value
672 pointed to by uaddr was not equal to the expected value val at
673 the time of the call.
674 Note: on Linux, the symbolic names EAGAIN and EWOULDBLOCK (both
676 of which appear in different parts of the kernel futex code)
677 have the same value.
678 EAGAIN (FUTEX_CMP_REQUEUE, FUTEX_CMP_REQUEUE_PI) The value pointed to
680 by uaddr is not equal to the expected value val3.
681 EAGAIN (FUTEX_LOCK_PI, FUTEX_TRYLOCK_PI, FUTEX_CMP_REQUEUE_PI) The
683 futex owner thread ID of uaddr (for FUTEX_CMP_REQUEUE_PI:
684 uaddr2) is about to exit, but has not yet handled the internal
685 state cleanup. Try again.
686 EDEADLK
688 (FUTEX_LOCK_PI, FUTEX_TRYLOCK_PI, FUTEX_CMP_REQUEUE_PI) The
689 futex word at uaddr is already locked by the caller.
690 EDEADLK
692 (FUTEX_CMP_REQUEUE_PI) While requeueing a waiter to the PI futex
693 for the futex word at uaddr2, the kernel detected a deadlock.
694 EFAULT A required pointer argument (i.e., uaddr, uaddr2, or timeout)
696 did not point to a valid user-space address.
697 EINTR A FUTEX_WAIT or FUTEX_WAIT_BITSET operation was interrupted by a
699 signal (see signal(7)). In kernels before Linux 2.6.22, this
700 error could also be returned for a spurious wakeup; since Linux
701 2.6.22, this no longer happens.
702 EINVAL The operation in futex_op is one of those that employs a time‐
704 out, but the supplied timeout argument was invalid (tv_sec was
705 less than zero, or tv_nsec was not less than 1,000,000,000).
706 EINVAL The operation specified in futex_op employs one or both of the
708 pointers uaddr and uaddr2, but one of these does not point to a
709 valid object—that is, the address is not four-byte-aligned.
710 EINVAL (FUTEX_WAIT_BITSET, FUTEX_WAKE_BITSET) The bit mask supplied in
712 val3 is zero.
713 EINVAL (FUTEX_CMP_REQUEUE_PI) uaddr equals uaddr2 (i.e., an attempt was
715 made to requeue to the same futex).
716 EINVAL (FUTEX_FD) The signal number supplied in val is invalid.
718 EINVAL (FUTEX_WAKE, FUTEX_WAKE_OP, FUTEX_WAKE_BITSET, FUTEX_REQUEUE,
720 FUTEX_CMP_REQUEUE) The kernel detected an inconsistency between
721 the user-space state at uaddr and the kernel state—that is, it
722 detected a waiter which waits in FUTEX_LOCK_PI on uaddr.
723 EINVAL (FUTEX_LOCK_PI, FUTEX_TRYLOCK_PI, FUTEX_UNLOCK_PI) The kernel
725 detected an inconsistency between the user-space state at uaddr
726 and the kernel state. This indicates either state corruption or
727 that the kernel found a waiter on uaddr which is waiting via
728 FUTEX_WAIT or FUTEX_WAIT_BITSET.
729 EINVAL (FUTEX_CMP_REQUEUE_PI) The kernel detected an inconsistency
731 between the user-space state at uaddr2 and the kernel state;
732 that is, the kernel detected a waiter which waits via FUTEX_WAIT
733 or FUTEX_WAIT_BITSET on uaddr2.
734 EINVAL (FUTEX_CMP_REQUEUE_PI) The kernel detected an inconsistency
736 between the user-space state at uaddr and the kernel state; that
737 is, the kernel detected a waiter which waits via FUTEX_WAIT or
738 FUTEX_WAIT_BITESET on uaddr.
739 EINVAL (FUTEX_CMP_REQUEUE_PI) The kernel detected an inconsistency
741 between the user-space state at uaddr and the kernel state; that
742 is, the kernel detected a waiter which waits on uaddr via
743 FUTEX_LOCK_PI (instead of FUTEX_WAIT_REQUEUE_PI).
744 EINVAL (FUTEX_CMP_REQUEUE_PI) An attempt was made to requeue a waiter
746 to a futex other than that specified by the matching
747 FUTEX_WAIT_REQUEUE_PI call for that waiter.
748 EINVAL (FUTEX_CMP_REQUEUE_PI) The val argument is not 1.
750 EINVAL Invalid argument.
752 ENFILE (FUTEX_FD) The system-wide limit on the total number of open
754 files has been reached.
755 ENOMEM (FUTEX_LOCK_PI, FUTEX_TRYLOCK_PI, FUTEX_CMP_REQUEUE_PI) The ker‐
757 nel could not allocate memory to hold state information.
758 ENOSYS Invalid operation specified in futex_op.
760 ENOSYS The FUTEX_CLOCK_REALTIME option was specified in futex_op, but
762 the accompanying operation was neither FUTEX_WAIT,
763 FUTEX_WAIT_BITSET, nor FUTEX_WAIT_REQUEUE_PI.
764 ENOSYS (FUTEX_LOCK_PI, FUTEX_TRYLOCK_PI, FUTEX_UNLOCK_PI,
766 FUTEX_CMP_REQUEUE_PI, FUTEX_WAIT_REQUEUE_PI) A run-time check
767 determined that the operation is not available. The PI-futex
768 operations are not implemented on all architectures and are not
769 supported on some CPU variants.
770 EPERM (FUTEX_LOCK_PI, FUTEX_TRYLOCK_PI, FUTEX_CMP_REQUEUE_PI) The
772 caller is not allowed to attach itself to the futex at uaddr
773 (for FUTEX_CMP_REQUEUE_PI: the futex at uaddr2). (This may be
774 caused by a state corruption in user space.)
775 EPERM (FUTEX_UNLOCK_PI) The caller does not own the lock represented
777 by the futex word.
778 ESRCH (FUTEX_LOCK_PI, FUTEX_TRYLOCK_PI, FUTEX_CMP_REQUEUE_PI) The
780 thread ID in the futex word at uaddr does not exist.
781 ESRCH (FUTEX_CMP_REQUEUE_PI) The thread ID in the futex word at uaddr2
783 does not exist.
784 ETIMEDOUT
786 The operation in futex_op employed the timeout specified in
787 timeout, and the timeout expired before the operation completed.
788
790 Futexes were first made available in a stable kernel release with Linux
791 2.6.0.
792 Initial futex support was merged in Linux 2.5.7 but with different
794 semantics from what was described above. A four-argument system call
795 with the semantics described in this page was introduced in Linux
796 2.5.40. A fifth argument was added in Linux 2.5.70, and a sixth argu‐
797 ment was added in Linux 2.6.7.
798
800 This system call is Linux-specific.
801
803 Glibc does not provide a wrapper for this system call; call it using
804 syscall(2).
805 Several higher-level programming abstractions are implemented via
807 futexes, including POSIX semaphores and various POSIX threads synchro‐
808 nization mechanisms (mutexes, condition variables, read-write locks,
809 and barriers).
810
812 The program below demonstrates use of futexes in a program where a par‐
813 ent process and a child process use a pair of futexes located inside a
814 shared anonymous mapping to synchronize access to a shared resource:
815 the terminal. The two processes each write nloops (a command-line
816 argument that defaults to 5 if omitted) messages to the terminal and
817 employ a synchronization protocol that ensures that they alternate in
818 writing messages. Upon running this program we see output such as the
819 following:
820 $ ./futex_demo
822 Parent (18534) 0
823 Child (18535) 0
824 Parent (18534) 1
825 Child (18535) 1
826 Parent (18534) 2
827 Child (18535) 2
828 Parent (18534) 3
829 Child (18535) 3
830 Parent (18534) 4
831 Child (18535) 4
832 Program source
834 /* futex_demo.c
836 Usage: futex_demo [nloops]
838 (Default: 5)
839 Demonstrate the use of futexes in a program where parent and child
841 use a pair of futexes located inside a shared anonymous mapping to
842 synchronize access to a shared resource: the terminal. The two
843 processes each write 'num-loops' messages to the terminal and employ
844 a synchronization protocol that ensures that they alternate in
845 writing messages.
846 */
847 #define _GNU_SOURCE
848 #include <stdio.h>
849 #include <errno.h>
850 #include <stdatomic.h>
851 #include <stdlib.h>
852 #include <unistd.h>
853 #include <sys/wait.h>
854 #include <sys/mman.h>
855 #include <sys/syscall.h>
856 #include <linux/futex.h>
857 #include <sys/time.h>
858 #define errExit(msg) do { perror(msg); exit(EXIT_FAILURE); \
860 } while (0)
861 static int *futex1, *futex2, *iaddr;
863 static int
865 futex(int *uaddr, int futex_op, int val,
866 const struct timespec *timeout, int *uaddr2, int val3)
867 {
868 return syscall(SYS_futex, uaddr, futex_op, val,
869 timeout, uaddr2, val3);
870 }
871 /* Acquire the futex pointed to by 'futexp': wait for its value to
873 become 1, and then set the value to 0. */
874 static void
876 fwait(int *futexp)
877 {
878 int s;
879 /* atomic_compare_exchange_strong(ptr, oldval, newval)
881 atomically performs the equivalent of:
882 if (*ptr == *oldval)
884 *ptr = newval;
885 It returns true if the test yielded true and *ptr was updated. */
887 while (1) {
889 /* Is the futex available? */
891 const int one = 1;
892 if (atomic_compare_exchange_strong(futexp, &one, 0))
893 break; /* Yes */
894 /* Futex is not available; wait */
896 s = futex(futexp, FUTEX_WAIT, 0, NULL, NULL, 0);
898 if (s == -1 && errno != EAGAIN)
899 errExit("futex-FUTEX_WAIT");
900 }
901 }
902 /* Release the futex pointed to by 'futexp': if the futex currently
904 has the value 0, set its value to 1 and the wake any futex waiters,
905 so that if the peer is blocked in fpost(), it can proceed. */
906 static void
908 fpost(int *futexp)
909 {
910 int s;
911 /* atomic_compare_exchange_strong() was described in comments above */
913 const int zero = 0;
915 if (atomic_compare_exchange_strong(futexp, &zero, 1)) {
916 s = futex(futexp, FUTEX_WAKE, 1, NULL, NULL, 0);
917 if (s == -1)
918 errExit("futex-FUTEX_WAKE");
919 }
920 }
921 int
923 main(int argc, char *argv[])
924 {
925 pid_t childPid;
926 int j, nloops;
927 setbuf(stdout, NULL);
929 nloops = (argc > 1) ? atoi(argv[1]) : 5;
931 /* Create a shared anonymous mapping that will hold the futexes.
933 Since the futexes are being shared between processes, we
934 subsequently use the "shared" futex operations (i.e., not the
935 ones suffixed "_PRIVATE") */
936 iaddr = mmap(NULL, sizeof(int) * 2, PROT_READ | PROT_WRITE,
938 MAP_ANONYMOUS | MAP_SHARED, -1, 0);
939 if (iaddr == MAP_FAILED)
940 errExit("mmap");
941 futex1 = &iaddr[0];
943 futex2 = &iaddr[1];
944 *futex1 = 0; /* State: unavailable */
946 *futex2 = 1; /* State: available */
947 /* Create a child process that inherits the shared anonymous
949 mapping */
950 childPid = fork();
952 if (childPid == -1)
953 errExit("fork");
954 if (childPid == 0) { /* Child */
956 for (j = 0; j < nloops; j++) {
957 fwait(futex1);
958 printf("Child (%ld) %d\n", (long) getpid(), j);
959 fpost(futex2);
960 }
961 exit(EXIT_SUCCESS);
963 }
964 /* Parent falls through to here */
966 for (j = 0; j < nloops; j++) {
968 fwait(futex2);
969 printf("Parent (%ld) %d\n", (long) getpid(), j);
970 fpost(futex1);
971 }
972 wait(NULL);
974 exit(EXIT_SUCCESS);
976 }
977
979 get_robust_list(2), restart_syscall(2), pthread_mutexattr_getproto‐
980 col(3), futex(7), sched(7)
981 The following kernel source files:
983 * Documentation/pi-futex.txt
985 * Documentation/futex-requeue-pi.txt
987 * Documentation/locking/rt-mutex.txt
989 * Documentation/locking/rt-mutex-design.txt
991 * Documentation/robust-futex-ABI.txt
993 Franke, H., Russell, R., and Kirwood, M., 2002. Fuss, Futexes and Fur‐
995 wocks: Fast Userlevel Locking in Linux (from proceedings of the Ottawa
996 Linux Symposium 2002),
997 ⟨http://kernel.org/doc/ols/2002/ols2002-pages-479-495.pdf⟩
998 Hart, D., 2009. A futex overview and update,
1000 ⟨http://lwn.net/Articles/360699/⟩
1001 Hart, D. and Guniguntala, D., 2009. Requeue-PI: Making Glibc Condvars
1003 PI-Aware (from proceedings of the 2009 Real-Time Linux Workshop),
1004 ⟨http://lwn.net/images/conf/rtlws11/papers/proc/p10.pdf⟩
1005 Drepper, U., 2011. Futexes Are Tricky,
1007 ⟨http://www.akkadia.org/drepper/futex.pdf⟩
1008 Futex example library, futex-*.tar.bz2 at
1010 ⟨ftp://ftp.kernel.org/pub/linux/kernel/people/rusty/⟩
1011
1013 This page is part of release 5.07 of the Linux man-pages project. A
1014 description of the project, information about reporting bugs, and the
1015 latest version of this page, can be found at
1016 https://www.kernel.org/doc/man-pages/.
1017Linux 2020-06-09 FUTEX(2)