| 1 |
.\" Copyright 2015, Michael Kerrisk <[email protected]> |
| 2 |
.\" Copyright, the authors of the Linux man-pages project |
| 3 |
.\" |
| 4 |
.\" %%%LICENSE_START(FREELY_REDISTRIBUTABLE) |
| 5 |
.\" may be freely modified and distributed |
| 6 |
.\" %%%LICENSE_END |
| 7 |
.\" |
| 8 |
.\" FIXME Still to integrate are some points from Torvald Riegel's mail of |
| 9 |
.\" 2015-01-23: |
| 10 |
.\" https://lore.kernel.org/lkml/[email protected] |
| 11 |
.\" |
| 12 |
.\" FIXME Do we need to add some text regarding Torvald Riegel's 2015-01-24 mail |
| 13 |
.\" https://lore.kernel.org/lkml/[email protected] |
| 14 |
.\" |
| 15 |
.TH futex 2 2026-02-14 "Linux man-pages 6.18" |
| 16 |
.SH NAME |
| 17 |
futex \- fast user-space locking |
| 18 |
.SH LIBRARY |
| 19 |
Standard C library |
| 20 |
.RI ( libc ,\~ \-lc ) |
| 21 |
.SH SYNOPSIS |
| 22 |
.nf |
| 23 |
.BR "#include <linux/futex.h>" " /* Definition of " FUTEX_* " constants */" |
| 24 |
.BR "#include <sys/syscall.h>" " /* Definition of " SYS_* " constants */" |
| 25 |
.B #include <unistd.h> |
| 26 |
.P |
| 27 |
.BI "long syscall(SYS_futex, uint32_t *" uaddr ", int " op ", ...);" |
| 28 |
.fi |
| 29 |
.SH DESCRIPTION |
| 30 |
The |
| 31 |
.BR futex () |
| 32 |
system call provides a method for waiting until a certain condition becomes |
| 33 |
true. |
| 34 |
It is typically used as a blocking construct in the context of |
| 35 |
shared-memory synchronization. |
| 36 |
When using futexes, the majority of |
| 37 |
the synchronization operations are performed in user space. |
| 38 |
A user-space program employs the |
| 39 |
.BR futex () |
| 40 |
system call only when it is likely that the program has to block for |
| 41 |
a longer time until the condition becomes true. |
| 42 |
Other |
| 43 |
.BR futex () |
| 44 |
operations can be used to wake any processes or threads waiting |
| 45 |
for a particular condition. |
| 46 |
.P |
| 47 |
A futex is a 32-bit value\[em]referred to below as a |
| 48 |
.IR "futex word" \[em]whose |
| 49 |
address is supplied to the |
| 50 |
.BR futex () |
| 51 |
system call. |
| 52 |
(Futexes are 32 bits in size on all platforms, including 64-bit systems.) |
| 53 |
All futex operations are governed by this value. |
| 54 |
In order to share a futex between processes, |
| 55 |
the futex is placed in a region of shared memory, |
| 56 |
created using (for example) |
| 57 |
.BR mmap (2) |
| 58 |
or |
| 59 |
.BR shmat (2). |
| 60 |
(Thus, the futex word may have different |
| 61 |
virtual addresses in different processes, |
| 62 |
but these addresses all refer to the same location in physical memory.) |
| 63 |
In a multithreaded program, it is sufficient to place the futex word |
| 64 |
in a global variable shared by all threads. |
| 65 |
.P |
| 66 |
When executing a futex operation that requests to block a thread, |
| 67 |
the kernel will block only if the futex word has the value that the |
| 68 |
calling thread supplied (as one of the arguments of the |
| 69 |
.BR futex () |
| 70 |
call) as the expected value of the futex word. |
| 71 |
The loading of the futex word's value, |
| 72 |
the comparison of that value with the expected value, |
| 73 |
and the actual blocking will happen atomically and will be totally ordered |
| 74 |
with respect to concurrent operations performed by other threads |
| 75 |
on the same futex word. |
| 76 |
.\" Notes from Darren Hart (Dec 2015): |
| 77 |
.\" Totally ordered with respect futex operations refers to semantics |
| 78 |
.\" of the ACQUIRE/RELEASE operations and how they impact ordering of |
| 79 |
.\" memory reads and writes. The kernel futex operations are protected |
| 80 |
.\" by spinlocks, which ensure that all operations are serialized |
| 81 |
.\" with respect to one another. |
| 82 |
.\" |
| 83 |
.\" This is a lot to attempt to define in this document. Perhaps a |
| 84 |
.\" reference to linux/Documentation/memory-barriers.txt as a footnote |
| 85 |
.\" would be sufficient? Or perhaps for this manual, "serialized" would |
| 86 |
.\" be sufficient, with a footnote regarding "totally ordered" and a |
| 87 |
.\" pointer to the memory-barrier documentation? |
| 88 |
Thus, the futex word is used to connect the synchronization in user space |
| 89 |
with the implementation of blocking by the kernel. |
| 90 |
Analogously to an atomic |
| 91 |
compare-and-exchange operation that potentially changes shared memory, |
| 92 |
blocking via a futex is an atomic compare-and-block operation. |
| 93 |
.\" FIXME(Torvald Riegel): |
| 94 |
.\" Eventually we want to have some text in NOTES to satisfy |
| 95 |
.\" the reference in the following sentence |
| 96 |
.\" See NOTES for a detailed specification of |
| 97 |
.\" the synchronization semantics. |
| 98 |
.P |
| 99 |
One use of futexes is for implementing locks. |
| 100 |
The state of the lock (i.e., acquired or not acquired) |
| 101 |
can be represented as an atomically accessed flag in shared memory. |
| 102 |
In the uncontended case, |
| 103 |
a thread can access or modify the lock state with atomic instructions, |
| 104 |
for example atomically changing it from not acquired to acquired |
| 105 |
using an atomic compare-and-exchange instruction. |
| 106 |
(Such instructions are performed entirely in user mode, |
| 107 |
and the kernel maintains no information about the lock state.) |
| 108 |
On the other hand, a thread may be unable to acquire a lock because |
| 109 |
it is already acquired by another thread. |
| 110 |
It then may pass the lock's flag as a futex word and the value |
| 111 |
representing the acquired state as the expected value to a |
| 112 |
.BR futex () |
| 113 |
wait operation. |
| 114 |
This |
| 115 |
.BR futex () |
| 116 |
operation will block if and only if the lock is still acquired |
| 117 |
(i.e., the value in the futex word still matches the "acquired state"). |
| 118 |
When releasing the lock, a thread has to first reset the |
| 119 |
lock state to not acquired and then execute a futex |
| 120 |
operation that wakes threads blocked on the lock flag used as a futex word |
| 121 |
(this can be further optimized to avoid unnecessary wake-ups). |
| 122 |
See |
| 123 |
.BR futex (7) |
| 124 |
for more detail on how to use futexes. |
| 125 |
.P |
| 126 |
Besides the basic wait and wake-up futex functionality, there are further |
| 127 |
futex operations aimed at supporting more complex use cases. |
| 128 |
.P |
| 129 |
Note that |
| 130 |
no explicit initialization or destruction is necessary to use futexes; |
| 131 |
the kernel maintains a futex |
| 132 |
(i.e., the kernel-internal implementation artifact) |
| 133 |
only while operations such as |
| 134 |
.BR FUTEX_WAIT (2const) |
| 135 |
are being performed on a particular futex word. |
| 136 |
.\" |
| 137 |
.SS Arguments |
| 138 |
The |
| 139 |
.I uaddr |
| 140 |
argument points to the futex word. |
| 141 |
On all platforms, futexes are four-byte |
| 142 |
integers that must be aligned on a four-byte boundary. |
| 143 |
The operation to perform on the futex is specified in the |
| 144 |
.I op |
| 145 |
argument. |
| 146 |
.\" |
| 147 |
.\"""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""" |
| 148 |
.\" |
| 149 |
.SS Futex operations |
| 150 |
The |
| 151 |
.I op |
| 152 |
argument consists of two parts: |
| 153 |
a command that specifies the operation to be performed, |
| 154 |
bitwise ORed with zero or more options that |
| 155 |
modify the behavior of the operation. |
| 156 |
The options that may be included in |
| 157 |
.I op |
| 158 |
are as follows: |
| 159 |
.TP |
| 160 |
.BR FUTEX_PRIVATE_FLAG " (since Linux 2.6.22)" |
| 161 |
.\" commit 34f01cc1f512fa783302982776895c73714ebbc2 |
| 162 |
This option bit can be employed with all futex operations. |
| 163 |
It tells the kernel that the futex is process-private and not shared |
| 164 |
with another process (i.e., it is being used for synchronization |
| 165 |
only between threads of the same process). |
| 166 |
This allows the kernel to make some additional performance optimizations. |
| 167 |
.\" I.e., It allows the kernel choose the fast path for validating |
| 168 |
.\" the user-space address and avoids expensive VMA lookups, |
| 169 |
.\" taking reference counts on file backing store, and so on. |
| 170 |
.IP |
| 171 |
As a convenience, |
| 172 |
.I <linux/futex.h> |
| 173 |
defines a set of constants with the suffix |
| 174 |
.B _PRIVATE |
| 175 |
that are equivalents of all of the operations listed below, |
| 176 |
.\" except the obsolete FUTEX_FD(2const), for which the "private" flag was |
| 177 |
.\" meaningless |
| 178 |
but with the |
| 179 |
.B FUTEX_PRIVATE_FLAG |
| 180 |
ORed into the constant value. |
| 181 |
Thus, there are |
| 182 |
.BR FUTEX_WAIT_PRIVATE , |
| 183 |
.BR FUTEX_WAKE_PRIVATE , |
| 184 |
and so on. |
| 185 |
.TP |
| 186 |
.BR FUTEX_CLOCK_REALTIME " (since Linux 2.6.28)" |
| 187 |
.\" commit 1acdac104668a0834cfa267de9946fac7764d486 |
| 188 |
This option bit can be employed only with the |
| 189 |
.BR FUTEX_WAIT_BITSET (2const), |
| 190 |
.BR FUTEX_WAIT_REQUEUE_PI (2const), |
| 191 |
(since Linux 4.5) |
| 192 |
.\" commit 337f13046ff03717a9e99675284a817527440a49 |
| 193 |
.BR FUTEX_WAIT (2const), |
| 194 |
and |
| 195 |
(since Linux 5.14) |
| 196 |
.\" commit bf22a6976897977b0a3f1aeba6823c959fc4fdae |
| 197 |
.BR FUTEX_LOCK_PI2 (2const) |
| 198 |
operations. |
| 199 |
.IP |
| 200 |
If this option is set, the kernel measures the |
| 201 |
.I timeout |
| 202 |
against the |
| 203 |
.B CLOCK_REALTIME |
| 204 |
clock. |
| 205 |
.IP |
| 206 |
If this option is not set, the kernel measures the |
| 207 |
.I timeout |
| 208 |
against the |
| 209 |
.B CLOCK_MONOTONIC |
| 210 |
clock. |
| 211 |
.P |
| 212 |
The operation specified in |
| 213 |
.I op |
| 214 |
is one of the following: |
| 215 |
.\" |
| 216 |
.\"""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""" |
| 217 |
.\" |
| 218 |
.TP |
| 219 |
.BR FUTEX_WAIT (2const) |
| 220 |
.TQ |
| 221 |
.BR FUTEX_WAKE (2const) |
| 222 |
.TQ |
| 223 |
.BR FUTEX_FD (2const) |
| 224 |
.TQ |
| 225 |
.BR FUTEX_REQUEUE (2const) |
| 226 |
.TQ |
| 227 |
.BR FUTEX_CMP_REQUEUE (2const) |
| 228 |
.TQ |
| 229 |
.BR FUTEX_WAKE_OP (2const) |
| 230 |
.TQ |
| 231 |
.BR FUTEX_WAIT_BITSET (2const) |
| 232 |
.TQ |
| 233 |
.BR FUTEX_WAKE_BITSET (2const) |
| 234 |
.\" |
| 235 |
.SS Priority-inheritance futexes |
| 236 |
Linux supports priority-inheritance (PI) futexes in order to handle |
| 237 |
priority-inversion problems that can be encountered with |
| 238 |
normal futex locks. |
| 239 |
Priority inversion is the problem that occurs when a high-priority |
| 240 |
task is blocked waiting to acquire a lock held by a low-priority task, |
| 241 |
while tasks at an intermediate priority continuously preempt |
| 242 |
the low-priority task from the CPU. |
| 243 |
Consequently, the low-priority task makes no progress toward |
| 244 |
releasing the lock, and the high-priority task remains blocked. |
| 245 |
.P |
| 246 |
Priority inheritance is a mechanism for dealing with |
| 247 |
the priority-inversion problem. |
| 248 |
With this mechanism, when a high-priority task becomes blocked |
| 249 |
by a lock held by a low-priority task, |
| 250 |
the priority of the low-priority task is temporarily raised |
| 251 |
to that of the high-priority task, |
| 252 |
so that it is not preempted by any intermediate level tasks, |
| 253 |
and can thus make progress toward releasing the lock. |
| 254 |
To be effective, priority inheritance must be transitive, |
| 255 |
meaning that if a high-priority task blocks on a lock |
| 256 |
held by a lower-priority task that is itself blocked by a lock |
| 257 |
held by another intermediate-priority task |
| 258 |
(and so on, for chains of arbitrary length), |
| 259 |
then both of those tasks |
| 260 |
(or more generally, all of the tasks in a lock chain) |
| 261 |
have their priorities raised to be the same as the high-priority task. |
| 262 |
.P |
| 263 |
From a user-space perspective, |
| 264 |
what makes a futex PI-aware is a policy agreement (described below) |
| 265 |
between user space and the kernel about the value of the futex word, |
| 266 |
coupled with the use of the PI-futex operations described below. |
| 267 |
(Unlike the other futex operations described above, |
| 268 |
the PI-futex operations are designed |
| 269 |
for the implementation of very specific IPC mechanisms.) |
| 270 |
.\" |
| 271 |
.\" Quoting Darren Hart: |
| 272 |
.\" These opcodes paired with the PI futex value policy (described below) |
| 273 |
.\" defines a "futex" as PI aware. These were created very specifically |
| 274 |
.\" in support of PI pthread_mutexes, so it makes a lot more sense to |
| 275 |
.\" talk about a PI aware pthread_mutex, than a PI aware futex, since |
| 276 |
.\" there is a lot of policy and scaffolding that has to be built up |
| 277 |
.\" around it to use it properly (this is what a PI pthread_mutex is). |
| 278 |
.P |
| 279 |
.\" mtk: The following text is drawn from the Hart/Guniguntala paper |
| 280 |
.\" (listed in SEE ALSO), but I have reworded some pieces |
| 281 |
.\" significantly. |
| 282 |
.\" |
| 283 |
The PI-futex operations described below differ from the other |
| 284 |
futex operations in that they impose policy on the use of the value of the |
| 285 |
futex word: |
| 286 |
.IP \[bu] 3 |
| 287 |
If the lock is not acquired, the futex word's value shall be 0. |
| 288 |
.IP \[bu] |
| 289 |
If the lock is acquired, the futex word's value shall |
| 290 |
be the thread ID (TID; |
| 291 |
see |
| 292 |
.BR gettid (2)) |
| 293 |
of the owning thread. |
| 294 |
.IP \[bu] |
| 295 |
If the lock is owned and there are threads contending for the lock, |
| 296 |
then the |
| 297 |
.B FUTEX_WAITERS |
| 298 |
bit shall be set in the futex word's value; |
| 299 |
in other words, |
| 300 |
this value is: |
| 301 |
.IP |
| 302 |
.in +4n |
| 303 |
.EX |
| 304 |
FUTEX_WAITERS | TID |
| 305 |
.EE |
| 306 |
.in |
| 307 |
.IP |
| 308 |
(Note that is invalid for a PI futex word to have no owner and |
| 309 |
.B FUTEX_WAITERS |
| 310 |
set.) |
| 311 |
.P |
| 312 |
With this policy in place, |
| 313 |
a user-space application can acquire an unacquired |
| 314 |
lock or release a lock using atomic instructions executed in user mode |
| 315 |
(e.g., a compare-and-swap operation such as |
| 316 |
.I cmpxchg |
| 317 |
on the x86 architecture). |
| 318 |
Acquiring a lock simply consists of using compare-and-swap to atomically |
| 319 |
set the futex word's value to the caller's TID if its previous value was 0. |
| 320 |
Releasing a lock requires using compare-and-swap to set the futex word's |
| 321 |
value to 0 if the previous value was the expected TID. |
| 322 |
.P |
| 323 |
If a futex is already acquired (i.e., has a nonzero value), |
| 324 |
waiters must employ the |
| 325 |
.BR FUTEX_LOCK_PI (2const) |
| 326 |
or |
| 327 |
.BR FUTEX_LOCK_PI2 (2const) |
| 328 |
operations to acquire the lock. |
| 329 |
If other threads are waiting for the lock, then the |
| 330 |
.B FUTEX_WAITERS |
| 331 |
bit is set in the futex value; |
| 332 |
in this case, the lock owner must employ the |
| 333 |
.BR FUTEX_UNLOCK_PI (2const) |
| 334 |
operation to release the lock. |
| 335 |
.P |
| 336 |
In the cases where callers are forced into the kernel |
| 337 |
(i.e., required to perform a |
| 338 |
.BR futex () |
| 339 |
call), |
| 340 |
they then deal directly with a so-called RT-mutex, |
| 341 |
a kernel locking mechanism which implements the required |
| 342 |
priority-inheritance semantics. |
| 343 |
After the RT-mutex is acquired, the futex value is updated accordingly, |
| 344 |
before the calling thread returns to user space. |
| 345 |
.P |
| 346 |
It is important to note |
| 347 |
.\" tglx (July 2015): |
| 348 |
.\" If there are multiple waiters on a pi futex then a wake pi operation |
| 349 |
.\" will wake the first waiter and hand over the lock to this waiter. This |
| 350 |
.\" includes handing over the rtmutex which represents the futex in the |
| 351 |
.\" kernel. The strict requirement is that the futex owner and the rtmutex |
| 352 |
.\" owner must be the same, except for the update period which is |
| 353 |
.\" serialized by the futex internal locking. That means the kernel must |
| 354 |
.\" update the user-space value prior to returning to user space |
| 355 |
that the kernel will update the futex word's value prior |
| 356 |
to returning to user space. |
| 357 |
(This prevents the possibility of the futex word's value ending |
| 358 |
up in an invalid state, such as having an owner but the value being 0, |
| 359 |
or having waiters but not having the |
| 360 |
.B FUTEX_WAITERS |
| 361 |
bit set.) |
| 362 |
.P |
| 363 |
If a futex has an associated RT-mutex in the kernel |
| 364 |
(i.e., there are blocked waiters) |
| 365 |
and the owner of the futex/RT-mutex dies unexpectedly, |
| 366 |
then the kernel cleans up the RT-mutex and hands it over to the next waiter. |
| 367 |
This in turn requires that the user-space value is updated accordingly. |
| 368 |
To indicate that this is required, the kernel sets the |
| 369 |
.B FUTEX_OWNER_DIED |
| 370 |
bit in the futex word along with the thread ID of the new owner. |
| 371 |
User space can detect this situation via the presence of the |
| 372 |
.B FUTEX_OWNER_DIED |
| 373 |
bit and is then responsible for cleaning up the stale state left over by |
| 374 |
the dead owner. |
| 375 |
.\" tglx (July 2015): |
| 376 |
.\" The FUTEX_OWNER_DIED bit can also be set on uncontended futexes, where |
| 377 |
.\" the kernel has no state associated. This happens via the robust futex |
| 378 |
.\" mechanism. In that case the futex value will be set to |
| 379 |
.\" FUTEX_OWNER_DIED. The robust futex mechanism is also available for non |
| 380 |
.\" PI futexes. |
| 381 |
.P |
| 382 |
PI futexes are operated on by specifying one of the values listed below in |
| 383 |
.IR op . |
| 384 |
Note that the PI futex operations must be used as paired operations |
| 385 |
and are subject to some additional requirements: |
| 386 |
.IP \[bu] 3 |
| 387 |
.BR FUTEX_LOCK_PI (2const), |
| 388 |
.BR FUTEX_LOCK_PI2 (2const), |
| 389 |
and |
| 390 |
.BR FUTEX_TRYLOCK_PI (2const) |
| 391 |
pair with |
| 392 |
.BR FUTEX_UNLOCK_PI (2const). |
| 393 |
.BR FUTEX_UNLOCK_PI (2const) |
| 394 |
must be called only on a futex owned by the calling thread, |
| 395 |
as defined by the value policy, otherwise the error |
| 396 |
.B EPERM |
| 397 |
results. |
| 398 |
.IP \[bu] |
| 399 |
.BR FUTEX_WAIT_REQUEUE_PI (2const) |
| 400 |
pairs with |
| 401 |
.BR FUTEX_CMP_REQUEUE_PI (2const). |
| 402 |
This must be performed from a non-PI futex to a distinct PI futex |
| 403 |
(or the error |
| 404 |
.B EINVAL |
| 405 |
results). |
| 406 |
Additionally, |
| 407 |
the number of waiters to be woken must be 1 |
| 408 |
(or the error |
| 409 |
.B EINVAL |
| 410 |
results). |
| 411 |
.P |
| 412 |
The PI futex operations are as follows: |
| 413 |
.\" |
| 414 |
.\"""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""" |
| 415 |
.\" |
| 416 |
.TP |
| 417 |
.BR FUTEX_LOCK_PI (2const) |
| 418 |
.TQ |
| 419 |
.BR FUTEX_LOCK_PI2 (2const) |
| 420 |
.TQ |
| 421 |
.BR FUTEX_TRYLOCK_PI (2const) |
| 422 |
.TQ |
| 423 |
.BR FUTEX_UNLOCK_PI (2const) |
| 424 |
.TQ |
| 425 |
.BR FUTEX_CMP_REQUEUE_PI (2const) |
| 426 |
.TQ |
| 427 |
.BR FUTEX_WAIT_REQUEUE_PI (2const) |
| 428 |
.P |
| 429 |
The |
| 430 |
.BR FUTEX_WAIT_REQUEUE_PI (2const) |
| 431 |
and |
| 432 |
.BR FUTEX_CMP_REQUEUE_PI (2const) |
| 433 |
were added to support a fairly specific use case: |
| 434 |
support for priority-inheritance-aware POSIX threads condition variables. |
| 435 |
The idea is that these operations should always be paired, |
| 436 |
in order to ensure that user space and the kernel remain in sync. |
| 437 |
Thus, in the |
| 438 |
.BR FUTEX_WAIT_REQUEUE_PI (2const) |
| 439 |
operation, the user-space application pre-specifies the target |
| 440 |
of the requeue that takes place in the |
| 441 |
.BR FUTEX_CMP_REQUEUE_PI (2const) |
| 442 |
operation. |
| 443 |
.\" |
| 444 |
.\" Darren Hart notes that a patch to allow glibc to fully support |
| 445 |
.\" PI-aware pthreads condition variables has not yet been accepted into |
| 446 |
.\" glibc. The story is complex, and can be found at |
| 447 |
.\" https://sourceware.org/bugzilla/show_bug.cgi?id=11588 |
| 448 |
.\" Darren notes that in the meantime, the patch is shipped with various |
| 449 |
.\" PREEMPT_RT-enabled Linux systems. |
| 450 |
.\" |
| 451 |
.\" Related to the preceding, Darren proposed that somewhere, man-pages |
| 452 |
.\" should document the following point: |
| 453 |
.\" |
| 454 |
.\" While the Linux kernel, since Linux 2.6.31, supports requeueing of |
| 455 |
.\" priority-inheritance (PI) aware mutexes via the |
| 456 |
.\" FUTEX_WAIT_REQUEUE_PI and FUTEX_CMP_REQUEUE_PI futex operations, |
| 457 |
.\" the glibc implementation does not yet take full advantage of this. |
| 458 |
.\" Specifically, the condvar internal data lock remains a non-PI aware |
| 459 |
.\" mutex, regardless of the type of the pthread_mutex associated with |
| 460 |
.\" the condvar. This can lead to an unbounded priority inversion on |
| 461 |
.\" the internal data lock even when associating a PI aware |
| 462 |
.\" pthread_mutex with a condvar during a pthread_cond*_wait |
| 463 |
.\" operation. For this reason, it is not recommended to rely on |
| 464 |
.\" priority inheritance when using pthread condition variables. |
| 465 |
.\" |
| 466 |
.\" The problem is that the obvious location for this text is |
| 467 |
.\" the pthread_cond*wait(3) man page. However, such a man page |
| 468 |
.\" does not currently exist. |
| 469 |
.\" |
| 470 |
.\"""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""" |
| 471 |
.\" |
| 472 |
.SH RETURN VALUE |
| 473 |
On error, |
| 474 |
\-1 is returned, |
| 475 |
and |
| 476 |
.I errno |
| 477 |
is set to indicate the error. |
| 478 |
.P |
| 479 |
The return value on success depends on the operation. |
| 480 |
.\" |
| 481 |
.\"""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""" |
| 482 |
.\" |
| 483 |
.SH ERRORS |
| 484 |
.TP |
| 485 |
.B EACCES |
| 486 |
No read access to the memory of a futex word. |
| 487 |
.TP |
| 488 |
.B EFAULT |
| 489 |
.I uaddr |
| 490 |
did not point to a valid user-space address. |
| 491 |
.TP |
| 492 |
.B EINVAL |
| 493 |
.I uaddr |
| 494 |
does not point to a valid object\[em]that is, |
| 495 |
the address is not four-byte-aligned. |
| 496 |
.TP |
| 497 |
.B EINVAL |
| 498 |
Invalid argument. |
| 499 |
.TP |
| 500 |
.B ENOSYS |
| 501 |
Invalid operation specified in |
| 502 |
.IR op . |
| 503 |
.TP |
| 504 |
.B ENOSYS |
| 505 |
The |
| 506 |
.B FUTEX_CLOCK_REALTIME |
| 507 |
option was specified in |
| 508 |
.IR op , |
| 509 |
but the accompanying operation was neither |
| 510 |
.BR FUTEX_WAIT_BITSET (2const), |
| 511 |
.BR FUTEX_WAIT_REQUEUE_PI (2const), |
| 512 |
nor |
| 513 |
.BR FUTEX_LOCK_PI2 (2const). |
| 514 |
.\" |
| 515 |
.\"""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""" |
| 516 |
.\" |
| 517 |
.SH STANDARDS |
| 518 |
Linux. |
| 519 |
.SH HISTORY |
| 520 |
Linux 2.6.0. |
| 521 |
.P |
| 522 |
Initial futex support was merged in Linux 2.5.7 but with different |
| 523 |
semantics from what was described above. |
| 524 |
A four-argument system call with the semantics |
| 525 |
described in this page was introduced in Linux 2.5.40. |
| 526 |
A fifth argument was added in Linux 2.5.70, |
| 527 |
and a sixth argument was added in Linux 2.6.7. |
| 528 |
.SH EXAMPLES |
| 529 |
The program below demonstrates use of futexes in a program where a parent |
| 530 |
process and a child process use a pair of futexes located inside a |
| 531 |
shared anonymous mapping to synchronize access to a shared resource: |
| 532 |
the terminal. |
| 533 |
The two processes each write |
| 534 |
.I nloops |
| 535 |
(a command-line argument that defaults to 5 if omitted) |
| 536 |
messages to the terminal and employ a synchronization protocol |
| 537 |
that ensures that they alternate in writing messages. |
| 538 |
Upon running this program we see output such as the following: |
| 539 |
.P |
| 540 |
.in +4n |
| 541 |
.EX |
| 542 |
.RB $ " ./futex_demo" ; |
| 543 |
Parent (18534) 0 |
| 544 |
Child (18535) 0 |
| 545 |
Parent (18534) 1 |
| 546 |
Child (18535) 1 |
| 547 |
Parent (18534) 2 |
| 548 |
Child (18535) 2 |
| 549 |
Parent (18534) 3 |
| 550 |
Child (18535) 3 |
| 551 |
Parent (18534) 4 |
| 552 |
Child (18535) 4 |
| 553 |
.EE |
| 554 |
.in |
| 555 |
.SS Program source |
| 556 |
\& |
| 557 |
.\" SRC BEGIN (futex.c) |
| 558 |
.EX |
| 559 |
/* futex_demo.c |
| 560 |
\& |
| 561 |
Usage: futex_demo [nloops] |
| 562 |
(Default: 5) |
| 563 |
\& |
| 564 |
Demonstrate the use of futexes in a program where parent and child |
| 565 |
use a pair of futexes located inside a shared anonymous mapping to |
| 566 |
synchronize access to a shared resource: the terminal. The two |
| 567 |
processes each write \[aq]num\-loops\[aq] messages to the terminal and employ |
| 568 |
a synchronization protocol that ensures that they alternate in |
| 569 |
writing messages. |
| 570 |
*/ |
| 571 |
#define _GNU_SOURCE |
| 572 |
#include <err.h> |
| 573 |
#include <errno.h> |
| 574 |
#include <linux/futex.h> |
| 575 |
#include <stdatomic.h> |
| 576 |
#include <stdint.h> |
| 577 |
#include <stdio.h> |
| 578 |
#include <stdlib.h> |
| 579 |
#include <sys/mman.h> |
| 580 |
#include <sys/syscall.h> |
| 581 |
#include <sys/time.h> |
| 582 |
#include <sys/wait.h> |
| 583 |
#include <unistd.h> |
| 584 |
\& |
| 585 |
static uint32_t *futex1, *futex2, *iaddr; |
| 586 |
\& |
| 587 |
static int |
| 588 |
futex(uint32_t *uaddr, int op, uint32_t val, |
| 589 |
const struct timespec *timeout, uint32_t *uaddr2, uint32_t val3) |
| 590 |
{ |
| 591 |
return syscall(SYS_futex, uaddr, op, val, |
| 592 |
timeout, uaddr2, val3); |
| 593 |
} |
| 594 |
\& |
| 595 |
/* Acquire the futex pointed to by \[aq]futexp\[aq]: wait for its value to |
| 596 |
become 1, and then set the value to 0. */ |
| 597 |
\& |
| 598 |
static void |
| 599 |
fwait(uint32_t *futexp) |
| 600 |
{ |
| 601 |
long s; |
| 602 |
const uint32_t one = 1; |
| 603 |
\& |
| 604 |
/* atomic_compare_exchange_strong(ptr, oldval, newval) |
| 605 |
atomically performs the equivalent of: |
| 606 |
\& |
| 607 |
if (*ptr == *oldval) |
| 608 |
*ptr = newval; |
| 609 |
\& |
| 610 |
It returns true if the test yielded true and *ptr was updated. */ |
| 611 |
\& |
| 612 |
for (;;) { |
| 613 |
\& |
| 614 |
/* Is the futex available? */ |
| 615 |
if (atomic_compare_exchange_strong(futexp, &one, 0)) |
| 616 |
break; /* Yes */ |
| 617 |
\& |
| 618 |
/* Futex is not available; wait. */ |
| 619 |
\& |
| 620 |
s = futex(futexp, FUTEX_WAIT, 0, NULL, NULL, 0); |
| 621 |
if (s == \-1 && errno != EAGAIN) |
| 622 |
err(EXIT_FAILURE, "futex\-FUTEX_WAIT"); |
| 623 |
} |
| 624 |
} |
| 625 |
\& |
| 626 |
/* Release the futex pointed to by \[aq]futexp\[aq]: if the futex currently |
| 627 |
has the value 0, set its value to 1 and then wake any futex waiters, |
| 628 |
so that if the peer is blocked in fwait(), it can proceed. */ |
| 629 |
\& |
| 630 |
static void |
| 631 |
fpost(uint32_t *futexp) |
| 632 |
{ |
| 633 |
long s; |
| 634 |
const uint32_t zero = 0; |
| 635 |
\& |
| 636 |
/* atomic_compare_exchange_strong() was described |
| 637 |
in comments above. */ |
| 638 |
\& |
| 639 |
if (atomic_compare_exchange_strong(futexp, &zero, 1)) { |
| 640 |
s = futex(futexp, FUTEX_WAKE, 1, NULL, NULL, 0); |
| 641 |
if (s == \-1) |
| 642 |
err(EXIT_FAILURE, "futex\-FUTEX_WAKE"); |
| 643 |
} |
| 644 |
} |
| 645 |
\& |
| 646 |
int |
| 647 |
main(int argc, char *argv[]) |
| 648 |
{ |
| 649 |
pid_t childPid; |
| 650 |
unsigned int nloops; |
| 651 |
\& |
| 652 |
setbuf(stdout, NULL); |
| 653 |
\& |
| 654 |
nloops = (argc > 1) ? atoi(argv[1]) : 5; |
| 655 |
\& |
| 656 |
/* Create a shared anonymous mapping that will hold the futexes. |
| 657 |
Since the futexes are being shared between processes, we |
| 658 |
subsequently use the "shared" futex operations (i.e., not the |
| 659 |
ones suffixed "_PRIVATE"). */ |
| 660 |
\& |
| 661 |
iaddr = mmap(NULL, sizeof(*iaddr) * 2, PROT_READ | PROT_WRITE, |
| 662 |
MAP_ANONYMOUS | MAP_SHARED, \-1, 0); |
| 663 |
if (iaddr == MAP_FAILED) |
| 664 |
err(EXIT_FAILURE, "mmap"); |
| 665 |
\& |
| 666 |
futex1 = &iaddr[0]; |
| 667 |
futex2 = &iaddr[1]; |
| 668 |
\& |
| 669 |
*futex1 = 0; /* State: unavailable */ |
| 670 |
*futex2 = 1; /* State: available */ |
| 671 |
\& |
| 672 |
/* Create a child process that inherits the shared anonymous |
| 673 |
mapping. */ |
| 674 |
\& |
| 675 |
childPid = fork(); |
| 676 |
if (childPid == \-1) |
| 677 |
err(EXIT_FAILURE, "fork"); |
| 678 |
\& |
| 679 |
if (childPid == 0) { /* Child */ |
| 680 |
for (unsigned int j = 0; j < nloops; j++) { |
| 681 |
fwait(futex1); |
| 682 |
printf("Child (%jd) %u\[rs]n", (intmax_t) getpid(), j); |
| 683 |
fpost(futex2); |
| 684 |
} |
| 685 |
\& |
| 686 |
exit(EXIT_SUCCESS); |
| 687 |
} |
| 688 |
\& |
| 689 |
/* Parent falls through to here. */ |
| 690 |
\& |
| 691 |
for (unsigned int j = 0; j < nloops; j++) { |
| 692 |
fwait(futex2); |
| 693 |
printf("Parent (%jd) %u\[rs]n", (intmax_t) getpid(), j); |
| 694 |
fpost(futex1); |
| 695 |
} |
| 696 |
\& |
| 697 |
wait(NULL); |
| 698 |
\& |
| 699 |
exit(EXIT_SUCCESS); |
| 700 |
} |
| 701 |
.EE |
| 702 |
.\" SRC END |
| 703 |
.SH SEE ALSO |
| 704 |
.ad l |
| 705 |
.BR get_robust_list (2), |
| 706 |
.BR restart_syscall (2), |
| 707 |
.BR pthread_mutexattr_getprotocol (3), |
| 708 |
.BR futex (7), |
| 709 |
.BR sched (7) |
| 710 |
.P |
| 711 |
The following kernel source files: |
| 712 |
.IP \[bu] 3 |
| 713 |
.I Documentation/locking/pi\-futex.rst |
| 714 |
.IP \[bu] |
| 715 |
.I Documentation/locking/futex\-requeue\-pi.rst |
| 716 |
.IP \[bu] |
| 717 |
.I Documentation/locking/rt\-mutex.rst |
| 718 |
.IP \[bu] |
| 719 |
.I Documentation/locking/rt\-mutex\-design.rst |
| 720 |
.IP \[bu] |
| 721 |
.I Documentation/robust\-futex\-ABI.rst |
| 722 |
.P |
| 723 |
Franke, H., Russell, R., and Kirwood, M., 2002. |
| 724 |
.br |
| 725 |
.UR http:\://kernel.org/\:doc/\:ols/\:2002/\:ols2002\-pages\-479\-495.pdf |
| 726 |
.I Fuss, Futexes and Furwocks: Fast Userlevel Locking in Linux |
| 727 |
.UE |
| 728 |
(from proceedings of the Ottawa Linux Symposium 2002). |
| 729 |
.P |
| 730 |
Hart, D., 2009. |
| 731 |
.UR http:\://lwn.net/\:Articles/\:360699/ |
| 732 |
.I A futex overview and update |
| 733 |
.UE . |
| 734 |
.P |
| 735 |
Hart, D.\& and Guniguntala, D., 2009. |
| 736 |
.UR http:\://lwn.net/\:images/\:conf/\:rtlws11/\:papers/\:proc/\:p10.pdf |
| 737 |
.I Requeue-PI: Making glibc Condvars PI-Aware |
| 738 |
.UE |
| 739 |
(from proceedings of the 2009 Real-Time Linux Workshop). |
| 740 |
.P |
| 741 |
Drepper, U., 2011. |
| 742 |
.UR http:\://www.akkadia.org/\:drepper/\:futex.pdf |
| 743 |
.I Futexes Are Tricky |
| 744 |
.UE . |
| 745 |
.P |
| 746 |
Futex example library, |
| 747 |
.UR https:\://mirrors.kernel.org/\:pub/\:linux/\:kernel/\:people/\:rusty/ |
| 748 |
futex\-*.tar.bz2 |
| 749 |
.UE . |
| 750 |
.\" |
| 751 |
.\" FIXME(Torvald) We should probably refer to the glibc code here, in |
| 752 |
.\" particular the glibc-internal futex wrapper functions that are |
| 753 |
.\" WIP, and the generic pthread_mutex_t and perhaps condvar |
| 754 |
.\" implementations. |