| 1 |
.\" Page by b.hubert |
| 2 |
.\" and Copyright (C) 2015, Thomas Gleixner <[email protected]> |
| 3 |
.\" and Copyright (C) 2015, Michael Kerrisk <[email protected]> |
| 4 |
.\" |
| 5 |
.\" %%%LICENSE_START(FREELY_REDISTRIBUTABLE) |
| 6 |
.\" may be freely modified and distributed |
| 7 |
.\" %%%LICENSE_END |
| 8 |
.\" |
| 9 |
.\" Niki A. Rahimi (LTC Security Development, [email protected]) |
| 10 |
.\" added ERRORS section. |
| 11 |
.\" |
| 12 |
.\" Modified 2004-06-17 mtk |
| 13 |
.\" Modified 2004-10-07 aeb, added FUTEX_REQUEUE, FUTEX_CMP_REQUEUE |
| 14 |
.\" |
| 15 |
.\" FIXME Still to integrate are some points from Torvald Riegel's mail of |
| 16 |
.\" 2015-01-23: |
| 17 |
.\" http://thread.gmane.org/gmane.linux.kernel/1703405/focus=7977 |
| 18 |
.\" |
| 19 |
.\" FIXME Do we need to add some text regarding Torvald Riegel's 2015-01-24 mail |
| 20 |
.\" http://thread.gmane.org/gmane.linux.kernel/1703405/focus=1873242 |
| 21 |
.\" |
| 22 |
.TH futex 2 2024-06-15 "Linux man-pages 6.9.1" |
| 23 |
.SH NAME |
| 24 |
futex \- fast user-space locking |
| 25 |
.SH LIBRARY |
| 26 |
Standard C library |
| 27 |
.RI ( libc ", " \-lc ) |
| 28 |
.SH SYNOPSIS |
| 29 |
.nf |
| 30 |
.P |
| 31 |
.BR "#include <linux/futex.h>" " /* Definition of " FUTEX_* " constants */" |
| 32 |
.BR "#include <sys/syscall.h>" " /* Definition of " SYS_* " constants */" |
| 33 |
.B #include <unistd.h> |
| 34 |
.P |
| 35 |
.BI "long syscall(SYS_futex, uint32_t *" uaddr ", int " futex_op \ |
| 36 |
", uint32_t " val , |
| 37 |
.BI " const struct timespec *" timeout , \ |
| 38 |
" \fR /* or: \fBuint32_t \fIval2\fP */" |
| 39 |
.BI " uint32_t *" uaddr2 ", uint32_t " val3 ); |
| 40 |
.fi |
| 41 |
.P |
| 42 |
.IR Note : |
| 43 |
glibc provides no wrapper for |
| 44 |
.BR futex (), |
| 45 |
necessitating the use of |
| 46 |
.BR syscall (2). |
| 47 |
.SH DESCRIPTION |
| 48 |
The |
| 49 |
.BR futex () |
| 50 |
system call provides a method for waiting until a certain condition becomes |
| 51 |
true. |
| 52 |
It is typically used as a blocking construct in the context of |
| 53 |
shared-memory synchronization. |
| 54 |
When using futexes, the majority of |
| 55 |
the synchronization operations are performed in user space. |
| 56 |
A user-space program employs the |
| 57 |
.BR futex () |
| 58 |
system call only when it is likely that the program has to block for |
| 59 |
a longer time until the condition becomes true. |
| 60 |
Other |
| 61 |
.BR futex () |
| 62 |
operations can be used to wake any processes or threads waiting |
| 63 |
for a particular condition. |
| 64 |
.P |
| 65 |
A futex is a 32-bit value\[em]referred to below as a |
| 66 |
.IR "futex word" \[em]whose |
| 67 |
address is supplied to the |
| 68 |
.BR futex () |
| 69 |
system call. |
| 70 |
(Futexes are 32 bits in size on all platforms, including 64-bit systems.) |
| 71 |
All futex operations are governed by this value. |
| 72 |
In order to share a futex between processes, |
| 73 |
the futex is placed in a region of shared memory, |
| 74 |
created using (for example) |
| 75 |
.BR mmap (2) |
| 76 |
or |
| 77 |
.BR shmat (2). |
| 78 |
(Thus, the futex word may have different |
| 79 |
virtual addresses in different processes, |
| 80 |
but these addresses all refer to the same location in physical memory.) |
| 81 |
In a multithreaded program, it is sufficient to place the futex word |
| 82 |
in a global variable shared by all threads. |
| 83 |
.P |
| 84 |
When executing a futex operation that requests to block a thread, |
| 85 |
the kernel will block only if the futex word has the value that the |
| 86 |
calling thread supplied (as one of the arguments of the |
| 87 |
.BR futex () |
| 88 |
call) as the expected value of the futex word. |
| 89 |
The loading of the futex word's value, |
| 90 |
the comparison of that value with the expected value, |
| 91 |
and the actual blocking will happen atomically and will be totally ordered |
| 92 |
with respect to concurrent operations performed by other threads |
| 93 |
on the same futex word. |
| 94 |
.\" Notes from Darren Hart (Dec 2015): |
| 95 |
.\" Totally ordered with respect futex operations refers to semantics |
| 96 |
.\" of the ACQUIRE/RELEASE operations and how they impact ordering of |
| 97 |
.\" memory reads and writes. The kernel futex operations are protected |
| 98 |
.\" by spinlocks, which ensure that all operations are serialized |
| 99 |
.\" with respect to one another. |
| 100 |
.\" |
| 101 |
.\" This is a lot to attempt to define in this document. Perhaps a |
| 102 |
.\" reference to linux/Documentation/memory-barriers.txt as a footnote |
| 103 |
.\" would be sufficient? Or perhaps for this manual, "serialized" would |
| 104 |
.\" be sufficient, with a footnote regarding "totally ordered" and a |
| 105 |
.\" pointer to the memory-barrier documentation? |
| 106 |
Thus, the futex word is used to connect the synchronization in user space |
| 107 |
with the implementation of blocking by the kernel. |
| 108 |
Analogously to an atomic |
| 109 |
compare-and-exchange operation that potentially changes shared memory, |
| 110 |
blocking via a futex is an atomic compare-and-block operation. |
| 111 |
.\" FIXME(Torvald Riegel): |
| 112 |
.\" Eventually we want to have some text in NOTES to satisfy |
| 113 |
.\" the reference in the following sentence |
| 114 |
.\" See NOTES for a detailed specification of |
| 115 |
.\" the synchronization semantics. |
| 116 |
.P |
| 117 |
One use of futexes is for implementing locks. |
| 118 |
The state of the lock (i.e., acquired or not acquired) |
| 119 |
can be represented as an atomically accessed flag in shared memory. |
| 120 |
In the uncontended case, |
| 121 |
a thread can access or modify the lock state with atomic instructions, |
| 122 |
for example atomically changing it from not acquired to acquired |
| 123 |
using an atomic compare-and-exchange instruction. |
| 124 |
(Such instructions are performed entirely in user mode, |
| 125 |
and the kernel maintains no information about the lock state.) |
| 126 |
On the other hand, a thread may be unable to acquire a lock because |
| 127 |
it is already acquired by another thread. |
| 128 |
It then may pass the lock's flag as a futex word and the value |
| 129 |
representing the acquired state as the expected value to a |
| 130 |
.BR futex () |
| 131 |
wait operation. |
| 132 |
This |
| 133 |
.BR futex () |
| 134 |
operation will block if and only if the lock is still acquired |
| 135 |
(i.e., the value in the futex word still matches the "acquired state"). |
| 136 |
When releasing the lock, a thread has to first reset the |
| 137 |
lock state to not acquired and then execute a futex |
| 138 |
operation that wakes threads blocked on the lock flag used as a futex word |
| 139 |
(this can be further optimized to avoid unnecessary wake-ups). |
| 140 |
See |
| 141 |
.BR futex (7) |
| 142 |
for more detail on how to use futexes. |
| 143 |
.P |
| 144 |
Besides the basic wait and wake-up futex functionality, there are further |
| 145 |
futex operations aimed at supporting more complex use cases. |
| 146 |
.P |
| 147 |
Note that |
| 148 |
no explicit initialization or destruction is necessary to use futexes; |
| 149 |
the kernel maintains a futex |
| 150 |
(i.e., the kernel-internal implementation artifact) |
| 151 |
only while operations such as |
| 152 |
.BR FUTEX_WAIT , |
| 153 |
described below, are being performed on a particular futex word. |
| 154 |
.\" |
| 155 |
.SS Arguments |
| 156 |
The |
| 157 |
.I uaddr |
| 158 |
argument points to the futex word. |
| 159 |
On all platforms, futexes are four-byte |
| 160 |
integers that must be aligned on a four-byte boundary. |
| 161 |
The operation to perform on the futex is specified in the |
| 162 |
.I futex_op |
| 163 |
argument; |
| 164 |
.I val |
| 165 |
is a value whose meaning and purpose depends on |
| 166 |
.IR futex_op . |
| 167 |
.P |
| 168 |
The remaining arguments |
| 169 |
.RI ( timeout , |
| 170 |
.IR uaddr2 , |
| 171 |
and |
| 172 |
.IR val3 ) |
| 173 |
are required only for certain of the futex operations described below. |
| 174 |
Where one of these arguments is not required, it is ignored. |
| 175 |
.P |
| 176 |
For several blocking operations, the |
| 177 |
.I timeout |
| 178 |
argument is a pointer to a |
| 179 |
.I timespec |
| 180 |
structure that specifies a timeout for the operation. |
| 181 |
However, notwithstanding the prototype shown above, for some operations, |
| 182 |
the least significant four bytes of this argument are instead |
| 183 |
used as an integer whose meaning is determined by the operation. |
| 184 |
For these operations, the kernel casts the |
| 185 |
.I timeout |
| 186 |
value first to |
| 187 |
.IR "unsigned long", |
| 188 |
then to |
| 189 |
.IR uint32_t , |
| 190 |
and in the remainder of this page, this argument is referred to as |
| 191 |
.I val2 |
| 192 |
when interpreted in this fashion. |
| 193 |
.P |
| 194 |
Where it is required, the |
| 195 |
.I uaddr2 |
| 196 |
argument is a pointer to a second futex word that is employed |
| 197 |
by the operation. |
| 198 |
.P |
| 199 |
The interpretation of the final integer argument, |
| 200 |
.IR val3 , |
| 201 |
depends on the operation. |
| 202 |
.\" |
| 203 |
.\"""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""" |
| 204 |
.\" |
| 205 |
.SS Futex operations |
| 206 |
The |
| 207 |
.I futex_op |
| 208 |
argument consists of two parts: |
| 209 |
a command that specifies the operation to be performed, |
| 210 |
bitwise ORed with zero or more options that |
| 211 |
modify the behaviour of the operation. |
| 212 |
The options that may be included in |
| 213 |
.I futex_op |
| 214 |
are as follows: |
| 215 |
.TP |
| 216 |
.BR FUTEX_PRIVATE_FLAG " (since Linux 2.6.22)" |
| 217 |
.\" commit 34f01cc1f512fa783302982776895c73714ebbc2 |
| 218 |
This option bit can be employed with all futex operations. |
| 219 |
It tells the kernel that the futex is process-private and not shared |
| 220 |
with another process (i.e., it is being used for synchronization |
| 221 |
only between threads of the same process). |
| 222 |
This allows the kernel to make some additional performance optimizations. |
| 223 |
.\" I.e., It allows the kernel choose the fast path for validating |
| 224 |
.\" the user-space address and avoids expensive VMA lookups, |
| 225 |
.\" taking reference counts on file backing store, and so on. |
| 226 |
.IP |
| 227 |
As a convenience, |
| 228 |
.I <linux/futex.h> |
| 229 |
defines a set of constants with the suffix |
| 230 |
.B _PRIVATE |
| 231 |
that are equivalents of all of the operations listed below, |
| 232 |
.\" except the obsolete FUTEX_FD, for which the "private" flag was |
| 233 |
.\" meaningless |
| 234 |
but with the |
| 235 |
.B FUTEX_PRIVATE_FLAG |
| 236 |
ORed into the constant value. |
| 237 |
Thus, there are |
| 238 |
.BR FUTEX_WAIT_PRIVATE , |
| 239 |
.BR FUTEX_WAKE_PRIVATE , |
| 240 |
and so on. |
| 241 |
.TP |
| 242 |
.BR FUTEX_CLOCK_REALTIME " (since Linux 2.6.28)" |
| 243 |
.\" commit 1acdac104668a0834cfa267de9946fac7764d486 |
| 244 |
This option bit can be employed only with the |
| 245 |
.BR FUTEX_WAIT_BITSET , |
| 246 |
.BR FUTEX_WAIT_REQUEUE_PI , |
| 247 |
(since Linux 4.5) |
| 248 |
.\" commit 337f13046ff03717a9e99675284a817527440a49 |
| 249 |
.BR FUTEX_WAIT , |
| 250 |
and |
| 251 |
(since Linux 5.14) |
| 252 |
.\" commit bf22a6976897977b0a3f1aeba6823c959fc4fdae |
| 253 |
.B FUTEX_LOCK_PI2 |
| 254 |
operations. |
| 255 |
.IP |
| 256 |
If this option is set, the kernel measures the |
| 257 |
.I timeout |
| 258 |
against the |
| 259 |
.B CLOCK_REALTIME |
| 260 |
clock. |
| 261 |
.IP |
| 262 |
If this option is not set, the kernel measures the |
| 263 |
.I timeout |
| 264 |
against the |
| 265 |
.B CLOCK_MONOTONIC |
| 266 |
clock. |
| 267 |
.P |
| 268 |
The operation specified in |
| 269 |
.I futex_op |
| 270 |
is one of the following: |
| 271 |
.\" |
| 272 |
.\"""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""" |
| 273 |
.\" |
| 274 |
.TP |
| 275 |
.BR FUTEX_WAIT " (since Linux 2.6.0)" |
| 276 |
.\" Strictly speaking, since some time in Linux 2.5.x |
| 277 |
This operation tests that the value at the |
| 278 |
futex word pointed to by the address |
| 279 |
.I uaddr |
| 280 |
still contains the expected value |
| 281 |
.IR val , |
| 282 |
and if so, then sleeps waiting for a |
| 283 |
.B FUTEX_WAKE |
| 284 |
operation on the futex word. |
| 285 |
The load of the value of the futex word is an atomic memory |
| 286 |
access (i.e., using atomic machine instructions of the respective |
| 287 |
architecture). |
| 288 |
This load, the comparison with the expected value, and |
| 289 |
starting to sleep are performed atomically |
| 290 |
.\" FIXME: Torvald, I think we may need to add some explanation of |
| 291 |
.\" "totally ordered" here. |
| 292 |
and totally ordered |
| 293 |
with respect to other futex operations on the same futex word. |
| 294 |
If the thread starts to sleep, |
| 295 |
it is considered a waiter on this futex word. |
| 296 |
If the futex value does not match |
| 297 |
.IR val , |
| 298 |
then the call fails immediately with the error |
| 299 |
.BR EAGAIN . |
| 300 |
.IP |
| 301 |
The purpose of the comparison with the expected value is to prevent lost |
| 302 |
wake-ups. |
| 303 |
If another thread changed the value of the futex word after the |
| 304 |
calling thread decided to block based on the prior value, |
| 305 |
and if the other thread executed a |
| 306 |
.B FUTEX_WAKE |
| 307 |
operation (or similar wake-up) after the value change and before this |
| 308 |
.B FUTEX_WAIT |
| 309 |
operation, then the calling thread will observe the |
| 310 |
value change and will not start to sleep. |
| 311 |
.IP |
| 312 |
If the |
| 313 |
.I timeout |
| 314 |
is not NULL, the structure it points to specifies a |
| 315 |
timeout for the wait. |
| 316 |
(This interval will be rounded up to the system clock granularity, |
| 317 |
and is guaranteed not to expire early.) |
| 318 |
The timeout is by default measured according to the |
| 319 |
.B CLOCK_MONOTONIC |
| 320 |
clock, but, since Linux 4.5, the |
| 321 |
.B CLOCK_REALTIME |
| 322 |
clock can be selected by specifying |
| 323 |
.B FUTEX_CLOCK_REALTIME |
| 324 |
in |
| 325 |
.IR futex_op . |
| 326 |
If |
| 327 |
.I timeout |
| 328 |
is NULL, the call blocks indefinitely. |
| 329 |
.IP |
| 330 |
.IR Note : |
| 331 |
for |
| 332 |
.BR FUTEX_WAIT , |
| 333 |
.I timeout |
| 334 |
is interpreted as a |
| 335 |
.I relative |
| 336 |
value. |
| 337 |
This differs from other futex operations, where |
| 338 |
.I timeout |
| 339 |
is interpreted as an absolute value. |
| 340 |
To obtain the equivalent of |
| 341 |
.B FUTEX_WAIT |
| 342 |
with an absolute timeout, employ |
| 343 |
.B FUTEX_WAIT_BITSET |
| 344 |
with |
| 345 |
.I val3 |
| 346 |
specified as |
| 347 |
.BR FUTEX_BITSET_MATCH_ANY . |
| 348 |
.IP |
| 349 |
The arguments |
| 350 |
.I uaddr2 |
| 351 |
and |
| 352 |
.I val3 |
| 353 |
are ignored. |
| 354 |
.\" FIXME . (Torvald) I think we should remove this. Or maybe adapt to a |
| 355 |
.\" different example. |
| 356 |
.\" |
| 357 |
.\" For |
| 358 |
.\" .BR futex (7), |
| 359 |
.\" this call is executed if decrementing the count gave a negative value |
| 360 |
.\" (indicating contention), |
| 361 |
.\" and will sleep until another process or thread releases |
| 362 |
.\" the futex and executes the |
| 363 |
.\" .B FUTEX_WAKE |
| 364 |
.\" operation. |
| 365 |
.\" |
| 366 |
.\"""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""" |
| 367 |
.\" |
| 368 |
.TP |
| 369 |
.BR FUTEX_WAKE " (since Linux 2.6.0)" |
| 370 |
.\" Strictly speaking, since Linux 2.5.x |
| 371 |
This operation wakes at most |
| 372 |
.I val |
| 373 |
of the waiters that are waiting (e.g., inside |
| 374 |
.BR FUTEX_WAIT ) |
| 375 |
on the futex word at the address |
| 376 |
.IR uaddr . |
| 377 |
Most commonly, |
| 378 |
.I val |
| 379 |
is specified as either 1 (wake up a single waiter) or |
| 380 |
.B INT_MAX |
| 381 |
(wake up all waiters). |
| 382 |
No guarantee is provided about which waiters are awoken |
| 383 |
(e.g., a waiter with a higher scheduling priority is not guaranteed |
| 384 |
to be awoken in preference to a waiter with a lower priority). |
| 385 |
.IP |
| 386 |
The arguments |
| 387 |
.IR timeout , |
| 388 |
.IR uaddr2 , |
| 389 |
and |
| 390 |
.I val3 |
| 391 |
are ignored. |
| 392 |
.\" FIXME . (Torvald) I think we should remove this. Or maybe adapt to |
| 393 |
.\" a different example. |
| 394 |
.\" |
| 395 |
.\" For |
| 396 |
.\" .BR futex (7), |
| 397 |
.\" this is executed if incrementing the count showed that |
| 398 |
.\" there were waiters, |
| 399 |
.\" once the futex value has been set to 1 |
| 400 |
.\" (indicating that it is available). |
| 401 |
.\" |
| 402 |
.\" How does "incrementing the count show that there were waiters"? |
| 403 |
.\" |
| 404 |
.\"""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""" |
| 405 |
.\" |
| 406 |
.TP |
| 407 |
.BR FUTEX_FD " (from Linux 2.6.0 up to and including Linux 2.6.25)" |
| 408 |
.\" Strictly speaking, from Linux 2.5.x to Linux 2.6.25 |
| 409 |
This operation creates a file descriptor that is associated with |
| 410 |
the futex at |
| 411 |
.IR uaddr . |
| 412 |
The caller must close the returned file descriptor after use. |
| 413 |
When another process or thread performs a |
| 414 |
.B FUTEX_WAKE |
| 415 |
on the futex word, the file descriptor indicates as being readable with |
| 416 |
.BR select (2), |
| 417 |
.BR poll (2), |
| 418 |
and |
| 419 |
.BR epoll (7) |
| 420 |
.IP |
| 421 |
The file descriptor can be used to obtain asynchronous notifications: if |
| 422 |
.I val |
| 423 |
is nonzero, then, when another process or thread executes a |
| 424 |
.BR FUTEX_WAKE , |
| 425 |
the caller will receive the signal number that was passed in |
| 426 |
.IR val . |
| 427 |
.IP |
| 428 |
The arguments |
| 429 |
.IR timeout , |
| 430 |
.IR uaddr2 , |
| 431 |
and |
| 432 |
.I val3 |
| 433 |
are ignored. |
| 434 |
.IP |
| 435 |
Because it was inherently racy, |
| 436 |
.B FUTEX_FD |
| 437 |
has been removed |
| 438 |
.\" commit 82af7aca56c67061420d618cc5a30f0fd4106b80 |
| 439 |
from Linux 2.6.26 onward. |
| 440 |
.\" |
| 441 |
.\"""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""" |
| 442 |
.\" |
| 443 |
.TP |
| 444 |
.BR FUTEX_REQUEUE " (since Linux 2.6.0)" |
| 445 |
This operation performs the same task as |
| 446 |
.B FUTEX_CMP_REQUEUE |
| 447 |
(see below), except that no check is made using the value in |
| 448 |
.IR val3 . |
| 449 |
(The argument |
| 450 |
.I val3 |
| 451 |
is ignored.) |
| 452 |
.\" |
| 453 |
.\"""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""" |
| 454 |
.\" |
| 455 |
.TP |
| 456 |
.BR FUTEX_CMP_REQUEUE " (since Linux 2.6.7)" |
| 457 |
This operation first checks whether the location |
| 458 |
.I uaddr |
| 459 |
still contains the value |
| 460 |
.IR val3 . |
| 461 |
If not, the operation fails with the error |
| 462 |
.BR EAGAIN . |
| 463 |
Otherwise, the operation wakes up a maximum of |
| 464 |
.I val |
| 465 |
waiters that are waiting on the futex at |
| 466 |
.IR uaddr . |
| 467 |
If there are more than |
| 468 |
.I val |
| 469 |
waiters, then the remaining waiters are removed |
| 470 |
from the wait queue of the source futex at |
| 471 |
.I uaddr |
| 472 |
and added to the wait queue of the target futex at |
| 473 |
.IR uaddr2 . |
| 474 |
The |
| 475 |
.I val2 |
| 476 |
argument specifies an upper limit on the number of waiters |
| 477 |
that are requeued to the futex at |
| 478 |
.IR uaddr2 . |
| 479 |
.IP |
| 480 |
.\" FIXME(Torvald) Is the following correct? Or is just the decision |
| 481 |
.\" which threads to wake or requeue part of the atomic operation? |
| 482 |
The load from |
| 483 |
.I uaddr |
| 484 |
is an atomic memory access (i.e., using atomic machine instructions of |
| 485 |
the respective architecture). |
| 486 |
This load, the comparison with |
| 487 |
.IR val3 , |
| 488 |
and the requeueing of any waiters are performed atomically and totally |
| 489 |
ordered with respect to other operations on the same futex word. |
| 490 |
.\" Notes from a f2f conversation with Thomas Gleixner (Aug 2015): ### |
| 491 |
.\" The operation is serialized with respect to operations on both |
| 492 |
.\" source and target futex. No other waiter can enqueue itself |
| 493 |
.\" for waiting and no other waiter can dequeue itself because of |
| 494 |
.\" a timeout or signal. |
| 495 |
.IP |
| 496 |
Typical values to specify for |
| 497 |
.I val |
| 498 |
are 0 or 1. |
| 499 |
(Specifying |
| 500 |
.B INT_MAX |
| 501 |
is not useful, because it would make the |
| 502 |
.B FUTEX_CMP_REQUEUE |
| 503 |
operation equivalent to |
| 504 |
.BR FUTEX_WAKE .) |
| 505 |
The limit value specified via |
| 506 |
.I val2 |
| 507 |
is typically either 1 or |
| 508 |
.BR INT_MAX . |
| 509 |
(Specifying the argument as 0 is not useful, because it would make the |
| 510 |
.B FUTEX_CMP_REQUEUE |
| 511 |
operation equivalent to |
| 512 |
.BR FUTEX_WAIT .) |
| 513 |
.IP |
| 514 |
The |
| 515 |
.B FUTEX_CMP_REQUEUE |
| 516 |
operation was added as a replacement for the earlier |
| 517 |
.BR FUTEX_REQUEUE . |
| 518 |
The difference is that the check of the value at |
| 519 |
.I uaddr |
| 520 |
can be used to ensure that requeueing happens only under certain |
| 521 |
conditions, which allows race conditions to be avoided in certain use cases. |
| 522 |
.\" But, as Rich Felker points out, there remain valid use cases for |
| 523 |
.\" FUTEX_REQUEUE, for example, when the calling thread is requeuing |
| 524 |
.\" the target(s) to a lock that the calling thread owns |
| 525 |
.\" From: Rich Felker <[email protected]> |
| 526 |
.\" Date: Wed, 29 Oct 2014 22:43:17 -0400 |
| 527 |
.\" To: Darren Hart <[email protected]> |
| 528 |
.\" CC: [email protected], ... |
| 529 |
.\" Subject: Re: Add futex wrapper to glibc? |
| 530 |
.IP |
| 531 |
Both |
| 532 |
.B FUTEX_REQUEUE |
| 533 |
and |
| 534 |
.B FUTEX_CMP_REQUEUE |
| 535 |
can be used to avoid "thundering herd" wake-ups that could occur when using |
| 536 |
.B FUTEX_WAKE |
| 537 |
in cases where all of the waiters that are woken need to acquire |
| 538 |
another futex. |
| 539 |
Consider the following scenario, |
| 540 |
where multiple waiter threads are waiting on B, |
| 541 |
a wait queue implemented using a futex: |
| 542 |
.IP |
| 543 |
.in +4n |
| 544 |
.EX |
| 545 |
lock(A) |
| 546 |
while (!check_value(V)) { |
| 547 |
unlock(A); |
| 548 |
block_on(B); |
| 549 |
lock(A); |
| 550 |
}; |
| 551 |
unlock(A); |
| 552 |
.EE |
| 553 |
.in |
| 554 |
.IP |
| 555 |
If a waker thread used |
| 556 |
.BR FUTEX_WAKE , |
| 557 |
then all waiters waiting on B would be woken up, |
| 558 |
and they would all try to acquire lock A. |
| 559 |
However, waking all of the threads in this manner would be pointless because |
| 560 |
all except one of the threads would immediately block on lock A again. |
| 561 |
By contrast, a requeue operation wakes just one waiter and moves |
| 562 |
the other waiters to lock A, |
| 563 |
and when the woken waiter unlocks A then the next waiter can proceed. |
| 564 |
.\" |
| 565 |
.\"""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""" |
| 566 |
.\" |
| 567 |
.TP |
| 568 |
.BR FUTEX_WAKE_OP " (since Linux 2.6.14)" |
| 569 |
.\" commit 4732efbeb997189d9f9b04708dc26bf8613ed721 |
| 570 |
.\" Author: Jakub Jelinek <[email protected]> |
| 571 |
.\" Date: Tue Sep 6 15:16:25 2005 -0700 |
| 572 |
.\" FIXME. (Torvald) The glibc condvar implementation is currently being |
| 573 |
.\" revised (e.g., to not use an internal lock anymore). |
| 574 |
.\" It is probably more future-proof to remove this paragraph. |
| 575 |
.\" [Torvald, do you have an update here?] |
| 576 |
This operation was added to support some user-space use cases |
| 577 |
where more than one futex must be handled at the same time. |
| 578 |
The most notable example is the implementation of |
| 579 |
.BR pthread_cond_signal (3), |
| 580 |
which requires operations on two futexes, |
| 581 |
the one used to implement the mutex and the one used in the implementation |
| 582 |
of the wait queue associated with the condition variable. |
| 583 |
.B FUTEX_WAKE_OP |
| 584 |
allows such cases to be implemented without leading to |
| 585 |
high rates of contention and context switching. |
| 586 |
.IP |
| 587 |
The |
| 588 |
.B FUTEX_WAKE_OP |
| 589 |
operation is equivalent to executing the following code atomically |
| 590 |
and totally ordered with respect to other futex operations on |
| 591 |
any of the two supplied futex words: |
| 592 |
.IP |
| 593 |
.in +4n |
| 594 |
.EX |
| 595 |
uint32_t oldval = *(uint32_t *) uaddr2; |
| 596 |
*(uint32_t *) uaddr2 = oldval \fIop\fP \fIoparg\fP; |
| 597 |
futex(uaddr, FUTEX_WAKE, val, 0, 0, 0); |
| 598 |
if (oldval \fIcmp\fP \fIcmparg\fP) |
| 599 |
futex(uaddr2, FUTEX_WAKE, val2, 0, 0, 0); |
| 600 |
.EE |
| 601 |
.in |
| 602 |
.IP |
| 603 |
In other words, |
| 604 |
.B FUTEX_WAKE_OP |
| 605 |
does the following: |
| 606 |
.RS |
| 607 |
.IP \[bu] 3 |
| 608 |
saves the original value of the futex word at |
| 609 |
.I uaddr2 |
| 610 |
and performs an operation to modify the value of the futex at |
| 611 |
.IR uaddr2 ; |
| 612 |
this is an atomic read-modify-write memory access (i.e., using atomic |
| 613 |
machine instructions of the respective architecture) |
| 614 |
.IP \[bu] |
| 615 |
wakes up a maximum of |
| 616 |
.I val |
| 617 |
waiters on the futex for the futex word at |
| 618 |
.IR uaddr ; |
| 619 |
and |
| 620 |
.IP \[bu] |
| 621 |
dependent on the results of a test of the original value of the |
| 622 |
futex word at |
| 623 |
.IR uaddr2 , |
| 624 |
wakes up a maximum of |
| 625 |
.I val2 |
| 626 |
waiters on the futex for the futex word at |
| 627 |
.IR uaddr2 . |
| 628 |
.RE |
| 629 |
.IP |
| 630 |
The operation and comparison that are to be performed are encoded |
| 631 |
in the bits of the argument |
| 632 |
.IR val3 . |
| 633 |
Pictorially, the encoding is: |
| 634 |
.IP |
| 635 |
.in +4n |
| 636 |
.EX |
| 637 |
+---+---+-----------+-----------+ |
| 638 |
|op |cmp| oparg | cmparg | |
| 639 |
+---+---+-----------+-----------+ |
| 640 |
4 4 12 12 <== # of bits |
| 641 |
.EE |
| 642 |
.in |
| 643 |
.IP |
| 644 |
Expressed in code, the encoding is: |
| 645 |
.IP |
| 646 |
.in +4n |
| 647 |
.EX |
| 648 |
#define FUTEX_OP(op, oparg, cmp, cmparg) \[rs] |
| 649 |
(((op & 0xf) << 28) | \[rs] |
| 650 |
((cmp & 0xf) << 24) | \[rs] |
| 651 |
((oparg & 0xfff) << 12) | \[rs] |
| 652 |
(cmparg & 0xfff)) |
| 653 |
.EE |
| 654 |
.in |
| 655 |
.IP |
| 656 |
In the above, |
| 657 |
.I op |
| 658 |
and |
| 659 |
.I cmp |
| 660 |
are each one of the codes listed below. |
| 661 |
The |
| 662 |
.I oparg |
| 663 |
and |
| 664 |
.I cmparg |
| 665 |
components are literal numeric values, except as noted below. |
| 666 |
.IP |
| 667 |
The |
| 668 |
.I op |
| 669 |
component has one of the following values: |
| 670 |
.IP |
| 671 |
.in +4n |
| 672 |
.EX |
| 673 |
FUTEX_OP_SET 0 /* uaddr2 = oparg; */ |
| 674 |
FUTEX_OP_ADD 1 /* uaddr2 += oparg; */ |
| 675 |
FUTEX_OP_OR 2 /* uaddr2 |= oparg; */ |
| 676 |
FUTEX_OP_ANDN 3 /* uaddr2 &= \[ti]oparg; */ |
| 677 |
FUTEX_OP_XOR 4 /* uaddr2 \[ha]= oparg; */ |
| 678 |
.EE |
| 679 |
.in |
| 680 |
.IP |
| 681 |
In addition, bitwise ORing the following value into |
| 682 |
.I op |
| 683 |
causes |
| 684 |
.I (1\~<<\~oparg) |
| 685 |
to be used as the operand: |
| 686 |
.IP |
| 687 |
.in +4n |
| 688 |
.EX |
| 689 |
FUTEX_OP_ARG_SHIFT 8 /* Use (1 << oparg) as operand */ |
| 690 |
.EE |
| 691 |
.in |
| 692 |
.IP |
| 693 |
The |
| 694 |
.I cmp |
| 695 |
field is one of the following: |
| 696 |
.IP |
| 697 |
.in +4n |
| 698 |
.EX |
| 699 |
FUTEX_OP_CMP_EQ 0 /* if (oldval == cmparg) wake */ |
| 700 |
FUTEX_OP_CMP_NE 1 /* if (oldval != cmparg) wake */ |
| 701 |
FUTEX_OP_CMP_LT 2 /* if (oldval < cmparg) wake */ |
| 702 |
FUTEX_OP_CMP_LE 3 /* if (oldval <= cmparg) wake */ |
| 703 |
FUTEX_OP_CMP_GT 4 /* if (oldval > cmparg) wake */ |
| 704 |
FUTEX_OP_CMP_GE 5 /* if (oldval >= cmparg) wake */ |
| 705 |
.EE |
| 706 |
.in |
| 707 |
.IP |
| 708 |
The return value of |
| 709 |
.B FUTEX_WAKE_OP |
| 710 |
is the sum of the number of waiters woken on the futex |
| 711 |
.I uaddr |
| 712 |
plus the number of waiters woken on the futex |
| 713 |
.IR uaddr2 . |
| 714 |
.\" |
| 715 |
.\"""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""" |
| 716 |
.\" |
| 717 |
.TP |
| 718 |
.BR FUTEX_WAIT_BITSET " (since Linux 2.6.25)" |
| 719 |
.\" commit cd689985cf49f6ff5c8eddc48d98b9d581d9475d |
| 720 |
This operation is like |
| 721 |
.B FUTEX_WAIT |
| 722 |
except that |
| 723 |
.I val3 |
| 724 |
is used to provide a 32-bit bit mask to the kernel. |
| 725 |
This bit mask, in which at least one bit must be set, |
| 726 |
is stored in the kernel-internal state of the waiter. |
| 727 |
See the description of |
| 728 |
.B FUTEX_WAKE_BITSET |
| 729 |
for further details. |
| 730 |
.IP |
| 731 |
If |
| 732 |
.I timeout |
| 733 |
is not NULL, the structure it points to specifies |
| 734 |
an absolute timeout for the wait operation. |
| 735 |
If |
| 736 |
.I timeout |
| 737 |
is NULL, the operation can block indefinitely. |
| 738 |
.IP |
| 739 |
The |
| 740 |
.I uaddr2 |
| 741 |
argument is ignored. |
| 742 |
.\" |
| 743 |
.\"""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""" |
| 744 |
.\" |
| 745 |
.TP |
| 746 |
.BR FUTEX_WAKE_BITSET " (since Linux 2.6.25)" |
| 747 |
.\" commit cd689985cf49f6ff5c8eddc48d98b9d581d9475d |
| 748 |
This operation is the same as |
| 749 |
.B FUTEX_WAKE |
| 750 |
except that the |
| 751 |
.I val3 |
| 752 |
argument is used to provide a 32-bit bit mask to the kernel. |
| 753 |
This bit mask, in which at least one bit must be set, |
| 754 |
is used to select which waiters should be woken up. |
| 755 |
The selection is done by a bitwise AND of the "wake" bit mask |
| 756 |
(i.e., the value in |
| 757 |
.IR val3 ) |
| 758 |
and the bit mask which is stored in the kernel-internal |
| 759 |
state of the waiter (the "wait" bit mask that is set using |
| 760 |
.BR FUTEX_WAIT_BITSET ). |
| 761 |
All of the waiters for which the result of the AND is nonzero are woken up; |
| 762 |
the remaining waiters are left sleeping. |
| 763 |
.IP |
| 764 |
The effect of |
| 765 |
.B FUTEX_WAIT_BITSET |
| 766 |
and |
| 767 |
.B FUTEX_WAKE_BITSET |
| 768 |
is to allow selective wake-ups among multiple waiters that are blocked |
| 769 |
on the same futex. |
| 770 |
However, note that, depending on the use case, |
| 771 |
employing this bit-mask multiplexing feature on a |
| 772 |
futex can be less efficient than simply using multiple futexes, |
| 773 |
because employing bit-mask multiplexing requires the kernel |
| 774 |
to check all waiters on a futex, |
| 775 |
including those that are not interested in being woken up |
| 776 |
(i.e., they do not have the relevant bit set in their "wait" bit mask). |
| 777 |
.\" According to http://locklessinc.com/articles/futex_cheat_sheet/: |
| 778 |
.\" |
| 779 |
.\" "The original reason for the addition of these extensions |
| 780 |
.\" was to improve the performance of pthread read-write locks |
| 781 |
.\" in glibc. However, the pthreads library no longer uses the |
| 782 |
.\" same locking algorithm, and these extensions are not used |
| 783 |
.\" without the bitset parameter being all ones. |
| 784 |
.\" |
| 785 |
.\" The page goes on to note that the FUTEX_WAIT_BITSET operation |
| 786 |
.\" is nevertheless used (with a bit mask of all ones) in order to |
| 787 |
.\" obtain the absolute timeout functionality that is useful |
| 788 |
.\" for efficiently implementing Pthreads APIs (which use absolute |
| 789 |
.\" timeouts); FUTEX_WAIT provides only relative timeouts. |
| 790 |
.IP |
| 791 |
The constant |
| 792 |
.BR FUTEX_BITSET_MATCH_ANY , |
| 793 |
which corresponds to all 32 bits set in the bit mask, can be used as the |
| 794 |
.I val3 |
| 795 |
argument for |
| 796 |
.B FUTEX_WAIT_BITSET |
| 797 |
and |
| 798 |
.BR FUTEX_WAKE_BITSET . |
| 799 |
Other than differences in the handling of the |
| 800 |
.I timeout |
| 801 |
argument, the |
| 802 |
.B FUTEX_WAIT |
| 803 |
operation is equivalent to |
| 804 |
.B FUTEX_WAIT_BITSET |
| 805 |
with |
| 806 |
.I val3 |
| 807 |
specified as |
| 808 |
.BR FUTEX_BITSET_MATCH_ANY ; |
| 809 |
that is, allow a wake-up by any waker. |
| 810 |
The |
| 811 |
.B FUTEX_WAKE |
| 812 |
operation is equivalent to |
| 813 |
.B FUTEX_WAKE_BITSET |
| 814 |
with |
| 815 |
.I val3 |
| 816 |
specified as |
| 817 |
.BR FUTEX_BITSET_MATCH_ANY ; |
| 818 |
that is, wake up any waiter(s). |
| 819 |
.IP |
| 820 |
The |
| 821 |
.I uaddr2 |
| 822 |
and |
| 823 |
.I timeout |
| 824 |
arguments are ignored. |
| 825 |
.\" |
| 826 |
.\"""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""" |
| 827 |
.\" |
| 828 |
.SS Priority-inheritance futexes |
| 829 |
Linux supports priority-inheritance (PI) futexes in order to handle |
| 830 |
priority-inversion problems that can be encountered with |
| 831 |
normal futex locks. |
| 832 |
Priority inversion is the problem that occurs when a high-priority |
| 833 |
task is blocked waiting to acquire a lock held by a low-priority task, |
| 834 |
while tasks at an intermediate priority continuously preempt |
| 835 |
the low-priority task from the CPU. |
| 836 |
Consequently, the low-priority task makes no progress toward |
| 837 |
releasing the lock, and the high-priority task remains blocked. |
| 838 |
.P |
| 839 |
Priority inheritance is a mechanism for dealing with |
| 840 |
the priority-inversion problem. |
| 841 |
With this mechanism, when a high-priority task becomes blocked |
| 842 |
by a lock held by a low-priority task, |
| 843 |
the priority of the low-priority task is temporarily raised |
| 844 |
to that of the high-priority task, |
| 845 |
so that it is not preempted by any intermediate level tasks, |
| 846 |
and can thus make progress toward releasing the lock. |
| 847 |
To be effective, priority inheritance must be transitive, |
| 848 |
meaning that if a high-priority task blocks on a lock |
| 849 |
held by a lower-priority task that is itself blocked by a lock |
| 850 |
held by another intermediate-priority task |
| 851 |
(and so on, for chains of arbitrary length), |
| 852 |
then both of those tasks |
| 853 |
(or more generally, all of the tasks in a lock chain) |
| 854 |
have their priorities raised to be the same as the high-priority task. |
| 855 |
.P |
| 856 |
From a user-space perspective, |
| 857 |
what makes a futex PI-aware is a policy agreement (described below) |
| 858 |
between user space and the kernel about the value of the futex word, |
| 859 |
coupled with the use of the PI-futex operations described below. |
| 860 |
(Unlike the other futex operations described above, |
| 861 |
the PI-futex operations are designed |
| 862 |
for the implementation of very specific IPC mechanisms.) |
| 863 |
.\" |
| 864 |
.\" Quoting Darren Hart: |
| 865 |
.\" These opcodes paired with the PI futex value policy (described below) |
| 866 |
.\" defines a "futex" as PI aware. These were created very specifically |
| 867 |
.\" in support of PI pthread_mutexes, so it makes a lot more sense to |
| 868 |
.\" talk about a PI aware pthread_mutex, than a PI aware futex, since |
| 869 |
.\" there is a lot of policy and scaffolding that has to be built up |
| 870 |
.\" around it to use it properly (this is what a PI pthread_mutex is). |
| 871 |
.P |
| 872 |
.\" mtk: The following text is drawn from the Hart/Guniguntala paper |
| 873 |
.\" (listed in SEE ALSO), but I have reworded some pieces |
| 874 |
.\" significantly. |
| 875 |
.\" |
| 876 |
The PI-futex operations described below differ from the other |
| 877 |
futex operations in that they impose policy on the use of the value of the |
| 878 |
futex word: |
| 879 |
.IP \[bu] 3 |
| 880 |
If the lock is not acquired, the futex word's value shall be 0. |
| 881 |
.IP \[bu] |
| 882 |
If the lock is acquired, the futex word's value shall |
| 883 |
be the thread ID (TID; |
| 884 |
see |
| 885 |
.BR gettid (2)) |
| 886 |
of the owning thread. |
| 887 |
.IP \[bu] |
| 888 |
If the lock is owned and there are threads contending for the lock, |
| 889 |
then the |
| 890 |
.B FUTEX_WAITERS |
| 891 |
bit shall be set in the futex word's value; in other words, this value is: |
| 892 |
.IP |
| 893 |
.in +4n |
| 894 |
.EX |
| 895 |
FUTEX_WAITERS | TID |
| 896 |
.EE |
| 897 |
.in |
| 898 |
.IP |
| 899 |
(Note that is invalid for a PI futex word to have no owner and |
| 900 |
.B FUTEX_WAITERS |
| 901 |
set.) |
| 902 |
.P |
| 903 |
With this policy in place, |
| 904 |
a user-space application can acquire an unacquired |
| 905 |
lock or release a lock using atomic instructions executed in user mode |
| 906 |
(e.g., a compare-and-swap operation such as |
| 907 |
.I cmpxchg |
| 908 |
on the x86 architecture). |
| 909 |
Acquiring a lock simply consists of using compare-and-swap to atomically |
| 910 |
set the futex word's value to the caller's TID if its previous value was 0. |
| 911 |
Releasing a lock requires using compare-and-swap to set the futex word's |
| 912 |
value to 0 if the previous value was the expected TID. |
| 913 |
.P |
| 914 |
If a futex is already acquired (i.e., has a nonzero value), |
| 915 |
waiters must employ the |
| 916 |
.B FUTEX_LOCK_PI |
| 917 |
or |
| 918 |
.B FUTEX_LOCK_PI2 |
| 919 |
operations to acquire the lock. |
| 920 |
If other threads are waiting for the lock, then the |
| 921 |
.B FUTEX_WAITERS |
| 922 |
bit is set in the futex value; |
| 923 |
in this case, the lock owner must employ the |
| 924 |
.B FUTEX_UNLOCK_PI |
| 925 |
operation to release the lock. |
| 926 |
.P |
| 927 |
In the cases where callers are forced into the kernel |
| 928 |
(i.e., required to perform a |
| 929 |
.BR futex () |
| 930 |
call), |
| 931 |
they then deal directly with a so-called RT-mutex, |
| 932 |
a kernel locking mechanism which implements the required |
| 933 |
priority-inheritance semantics. |
| 934 |
After the RT-mutex is acquired, the futex value is updated accordingly, |
| 935 |
before the calling thread returns to user space. |
| 936 |
.P |
| 937 |
It is important to note |
| 938 |
.\" tglx (July 2015): |
| 939 |
.\" If there are multiple waiters on a pi futex then a wake pi operation |
| 940 |
.\" will wake the first waiter and hand over the lock to this waiter. This |
| 941 |
.\" includes handing over the rtmutex which represents the futex in the |
| 942 |
.\" kernel. The strict requirement is that the futex owner and the rtmutex |
| 943 |
.\" owner must be the same, except for the update period which is |
| 944 |
.\" serialized by the futex internal locking. That means the kernel must |
| 945 |
.\" update the user-space value prior to returning to user space |
| 946 |
that the kernel will update the futex word's value prior |
| 947 |
to returning to user space. |
| 948 |
(This prevents the possibility of the futex word's value ending |
| 949 |
up in an invalid state, such as having an owner but the value being 0, |
| 950 |
or having waiters but not having the |
| 951 |
.B FUTEX_WAITERS |
| 952 |
bit set.) |
| 953 |
.P |
| 954 |
If a futex has an associated RT-mutex in the kernel |
| 955 |
(i.e., there are blocked waiters) |
| 956 |
and the owner of the futex/RT-mutex dies unexpectedly, |
| 957 |
then the kernel cleans up the RT-mutex and hands it over to the next waiter. |
| 958 |
This in turn requires that the user-space value is updated accordingly. |
| 959 |
To indicate that this is required, the kernel sets the |
| 960 |
.B FUTEX_OWNER_DIED |
| 961 |
bit in the futex word along with the thread ID of the new owner. |
| 962 |
User space can detect this situation via the presence of the |
| 963 |
.B FUTEX_OWNER_DIED |
| 964 |
bit and is then responsible for cleaning up the stale state left over by |
| 965 |
the dead owner. |
| 966 |
.\" tglx (July 2015): |
| 967 |
.\" The FUTEX_OWNER_DIED bit can also be set on uncontended futexes, where |
| 968 |
.\" the kernel has no state associated. This happens via the robust futex |
| 969 |
.\" mechanism. In that case the futex value will be set to |
| 970 |
.\" FUTEX_OWNER_DIED. The robust futex mechanism is also available for non |
| 971 |
.\" PI futexes. |
| 972 |
.P |
| 973 |
PI futexes are operated on by specifying one of the values listed below in |
| 974 |
.IR futex_op . |
| 975 |
Note that the PI futex operations must be used as paired operations |
| 976 |
and are subject to some additional requirements: |
| 977 |
.IP \[bu] 3 |
| 978 |
.BR FUTEX_LOCK_PI , |
| 979 |
.BR FUTEX_LOCK_PI2 , |
| 980 |
and |
| 981 |
.B FUTEX_TRYLOCK_PI |
| 982 |
pair with |
| 983 |
.BR FUTEX_UNLOCK_PI . |
| 984 |
.B FUTEX_UNLOCK_PI |
| 985 |
must be called only on a futex owned by the calling thread, |
| 986 |
as defined by the value policy, otherwise the error |
| 987 |
.B EPERM |
| 988 |
results. |
| 989 |
.IP \[bu] |
| 990 |
.B FUTEX_WAIT_REQUEUE_PI |
| 991 |
pairs with |
| 992 |
.BR FUTEX_CMP_REQUEUE_PI . |
| 993 |
This must be performed from a non-PI futex to a distinct PI futex |
| 994 |
(or the error |
| 995 |
.B EINVAL |
| 996 |
results). |
| 997 |
Additionally, |
| 998 |
.I val |
| 999 |
(the number of waiters to be woken) must be 1 |
| 1000 |
(or the error |
| 1001 |
.B EINVAL |
| 1002 |
results). |
| 1003 |
.P |
| 1004 |
The PI futex operations are as follows: |
| 1005 |
.\" |
| 1006 |
.\"""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""" |
| 1007 |
.\" |
| 1008 |
.TP |
| 1009 |
.BR FUTEX_LOCK_PI " (since Linux 2.6.18)" |
| 1010 |
.\" commit c87e2837be82df479a6bae9f155c43516d2feebc |
| 1011 |
This operation is used after an attempt to acquire |
| 1012 |
the lock via an atomic user-mode instruction failed |
| 1013 |
because the futex word has a nonzero value\[em]specifically, |
| 1014 |
because it contained the (PID-namespace-specific) TID of the lock owner. |
| 1015 |
.IP |
| 1016 |
The operation checks the value of the futex word at the address |
| 1017 |
.IR uaddr . |
| 1018 |
If the value is 0, then the kernel tries to atomically set |
| 1019 |
the futex value to the caller's TID. |
| 1020 |
If the futex word's value is nonzero, |
| 1021 |
the kernel atomically sets the |
| 1022 |
.B FUTEX_WAITERS |
| 1023 |
bit, which signals the futex owner that it cannot unlock the futex in |
| 1024 |
user space atomically by setting the futex value to 0. |
| 1025 |
.\" tglx (July 2015): |
| 1026 |
.\" The operation here is similar to the FUTEX_WAIT logic. When the user |
| 1027 |
.\" space atomic acquire does not succeed because the futex value was non |
| 1028 |
.\" zero, then the waiter goes into the kernel, takes the kernel internal |
| 1029 |
.\" lock and retries the acquisition under the lock. If the acquisition |
| 1030 |
.\" does not succeed either, then it sets the FUTEX_WAITERS bit, to signal |
| 1031 |
.\" the lock owner that it needs to go into the kernel. Here is the pseudo |
| 1032 |
.\" code: |
| 1033 |
.\" |
| 1034 |
.\" lock(kernel_lock); |
| 1035 |
.\" retry: |
| 1036 |
.\" |
| 1037 |
.\" /* |
| 1038 |
.\" * Owner might have unlocked in user space before we |
| 1039 |
.\" * were able to set the waiter bit. |
| 1040 |
.\" */ |
| 1041 |
.\" if (atomic_acquire(futex) == SUCCESS) { |
| 1042 |
.\" unlock(kernel_lock()); |
| 1043 |
.\" return 0; |
| 1044 |
.\" } |
| 1045 |
.\" |
| 1046 |
.\" /* |
| 1047 |
.\" * Owner might have unlocked after the above atomic_acquire() |
| 1048 |
.\" * attempt. |
| 1049 |
.\" */ |
| 1050 |
.\" if (atomic_set_waiters_bit(futex) != SUCCESS) |
| 1051 |
.\" goto retry; |
| 1052 |
.\" |
| 1053 |
.\" queue_waiter(); |
| 1054 |
.\" unlock(kernel_lock); |
| 1055 |
.\" block(); |
| 1056 |
.\" |
| 1057 |
After that, the kernel: |
| 1058 |
.RS |
| 1059 |
.IP (1) 5 |
| 1060 |
Tries to find the thread which is associated with the owner TID. |
| 1061 |
.IP (2) |
| 1062 |
Creates or reuses kernel state on behalf of the owner. |
| 1063 |
(If this is the first waiter, there is no kernel state for this |
| 1064 |
futex, so kernel state is created by locking the RT-mutex |
| 1065 |
and the futex owner is made the owner of the RT-mutex. |
| 1066 |
If there are existing waiters, then the existing state is reused.) |
| 1067 |
.IP (3) |
| 1068 |
Attaches the waiter to the futex |
| 1069 |
(i.e., the waiter is enqueued on the RT-mutex waiter list). |
| 1070 |
.RE |
| 1071 |
.IP |
| 1072 |
If more than one waiter exists, |
| 1073 |
the enqueueing of the waiter is in descending priority order. |
| 1074 |
(For information on priority ordering, see the discussion of the |
| 1075 |
.BR SCHED_DEADLINE , |
| 1076 |
.BR SCHED_FIFO , |
| 1077 |
and |
| 1078 |
.B SCHED_RR |
| 1079 |
scheduling policies in |
| 1080 |
.BR sched (7).) |
| 1081 |
The owner inherits either the waiter's CPU bandwidth |
| 1082 |
(if the waiter is scheduled under the |
| 1083 |
.B SCHED_DEADLINE |
| 1084 |
policy) or the waiter's priority (if the waiter is scheduled under the |
| 1085 |
.B SCHED_RR |
| 1086 |
or |
| 1087 |
.B SCHED_FIFO |
| 1088 |
policy). |
| 1089 |
.\" August 2015: |
| 1090 |
.\" mtk: If the realm is restricted purely to SCHED_OTHER (SCHED_NORMAL) |
| 1091 |
.\" processes, does the nice value come into play also? |
| 1092 |
.\" |
| 1093 |
.\" tglx: No. SCHED_OTHER/NORMAL tasks are handled in FIFO order |
| 1094 |
This inheritance follows the lock chain in the case of nested locking |
| 1095 |
.\" (i.e., task 1 blocks on lock A, held by task 2, |
| 1096 |
.\" while task 2 blocks on lock B, held by task 3) |
| 1097 |
and performs deadlock detection. |
| 1098 |
.IP |
| 1099 |
The |
| 1100 |
.I timeout |
| 1101 |
argument provides a timeout for the lock attempt. |
| 1102 |
If |
| 1103 |
.I timeout |
| 1104 |
is not NULL, the structure it points to specifies |
| 1105 |
an absolute timeout, measured against the |
| 1106 |
.B CLOCK_REALTIME |
| 1107 |
clock. |
| 1108 |
.\" 2016-07-07 response from Thomas Gleixner on LKML: |
| 1109 |
.\" From: Thomas Gleixner <[email protected]> |
| 1110 |
.\" Date: 6 July 2016 at 20:57 |
| 1111 |
.\" Subject: Re: futex: Allow FUTEX_CLOCK_REALTIME with FUTEX_WAIT op |
| 1112 |
.\" |
| 1113 |
.\" On Thu, 23 Jun 2016, Michael Kerrisk (man-pages) wrote: |
| 1114 |
.\" > On 06/23/2016 08:28 PM, Darren Hart wrote: |
| 1115 |
.\" > > And as a follow-on, what is the reason for FUTEX_LOCK_PI only using |
| 1116 |
.\" > > CLOCK_REALTIME? It seems reasonable to me that a user may want to wait a |
| 1117 |
.\" > > specific amount of time, regardless of wall time. |
| 1118 |
.\" > |
| 1119 |
.\" > Yes, that's another weird inconsistency. |
| 1120 |
.\" |
| 1121 |
.\" The reason is that phtread_mutex_timedlock() uses absolute timeouts based on |
| 1122 |
.\" CLOCK_REALTIME. glibc folks asked to make that the default behaviour back |
| 1123 |
.\" then when we added LOCK_PI. |
| 1124 |
If |
| 1125 |
.I timeout |
| 1126 |
is NULL, the operation will block indefinitely. |
| 1127 |
.IP |
| 1128 |
The |
| 1129 |
.IR uaddr2 , |
| 1130 |
.IR val , |
| 1131 |
and |
| 1132 |
.I val3 |
| 1133 |
arguments are ignored. |
| 1134 |
.\" |
| 1135 |
.\"""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""" |
| 1136 |
.\" |
| 1137 |
.TP |
| 1138 |
.BR FUTEX_LOCK_PI2 " (since Linux 5.14)" |
| 1139 |
.\" commit bf22a6976897977b0a3f1aeba6823c959fc4fdae |
| 1140 |
This operation is the same as |
| 1141 |
.BR FUTEX_LOCK_PI , |
| 1142 |
except that the clock against which |
| 1143 |
.I timeout |
| 1144 |
is measured is selectable. |
| 1145 |
By default, the (absolute) timeout specified in |
| 1146 |
.I timeout |
| 1147 |
is measured against the |
| 1148 |
.B CLOCK_MONOTONIC |
| 1149 |
clock, but if the |
| 1150 |
.B FUTEX_CLOCK_REALTIME |
| 1151 |
flag is specified in |
| 1152 |
.IR futex_op , |
| 1153 |
then the timeout is measured against the |
| 1154 |
.B CLOCK_REALTIME |
| 1155 |
clock. |
| 1156 |
.\" |
| 1157 |
.\"""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""" |
| 1158 |
.\" |
| 1159 |
.TP |
| 1160 |
.BR FUTEX_TRYLOCK_PI " (since Linux 2.6.18)" |
| 1161 |
.\" commit c87e2837be82df479a6bae9f155c43516d2feebc |
| 1162 |
This operation tries to acquire the lock at |
| 1163 |
.IR uaddr . |
| 1164 |
It is invoked when a user-space atomic acquire did not |
| 1165 |
succeed because the futex word was not 0. |
| 1166 |
.IP |
| 1167 |
Because the kernel has access to more state information than user space, |
| 1168 |
acquisition of the lock might succeed if performed by the |
| 1169 |
kernel in cases where the futex word |
| 1170 |
(i.e., the state information accessible to use-space) contains stale state |
| 1171 |
.RB ( FUTEX_WAITERS |
| 1172 |
and/or |
| 1173 |
.BR FUTEX_OWNER_DIED ). |
| 1174 |
This can happen when the owner of the futex died. |
| 1175 |
User space cannot handle this condition in a race-free manner, |
| 1176 |
but the kernel can fix this up and acquire the futex. |
| 1177 |
.\" Paraphrasing a f2f conversation with Thomas Gleixner about the |
| 1178 |
.\" above point (Aug 2015): ### |
| 1179 |
.\" There is a rare possibility of a race condition involving an |
| 1180 |
.\" uncontended futex with no owner, but with waiters. The |
| 1181 |
.\" kernel-user-space contract is that if a futex is nonzero, you must |
| 1182 |
.\" go into kernel. The futex was owned by a task, and that task dies |
| 1183 |
.\" but there are no waiters, so the futex value is non zero. |
| 1184 |
.\" Therefore, the next locker has to go into the kernel, |
| 1185 |
.\" so that the kernel has a chance to clean up. (CMXCH on zero |
| 1186 |
.\" in user space would fail, so kernel has to clean up.) |
| 1187 |
.\" Darren Hart (Oct 2015): |
| 1188 |
.\" The trylock in the kernel has more state, so it can independently |
| 1189 |
.\" verify the flags that user space must trust implicitly. |
| 1190 |
.IP |
| 1191 |
The |
| 1192 |
.IR uaddr2 , |
| 1193 |
.IR val , |
| 1194 |
.IR timeout , |
| 1195 |
and |
| 1196 |
.I val3 |
| 1197 |
arguments are ignored. |
| 1198 |
.\" |
| 1199 |
.\"""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""" |
| 1200 |
.\" |
| 1201 |
.TP |
| 1202 |
.BR FUTEX_UNLOCK_PI " (since Linux 2.6.18)" |
| 1203 |
.\" commit c87e2837be82df479a6bae9f155c43516d2feebc |
| 1204 |
This operation wakes the top priority waiter that is waiting in |
| 1205 |
.B FUTEX_LOCK_PI |
| 1206 |
or |
| 1207 |
.B FUTEX_LOCK_PI2 |
| 1208 |
on the futex address provided by the |
| 1209 |
.I uaddr |
| 1210 |
argument. |
| 1211 |
.IP |
| 1212 |
This is called when the user-space value at |
| 1213 |
.I uaddr |
| 1214 |
cannot be changed atomically from a TID (of the owner) to 0. |
| 1215 |
.IP |
| 1216 |
The |
| 1217 |
.IR uaddr2 , |
| 1218 |
.IR val , |
| 1219 |
.IR timeout , |
| 1220 |
and |
| 1221 |
.I val3 |
| 1222 |
arguments are ignored. |
| 1223 |
.\" |
| 1224 |
.\"""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""" |
| 1225 |
.\" |
| 1226 |
.TP |
| 1227 |
.BR FUTEX_CMP_REQUEUE_PI " (since Linux 2.6.31)" |
| 1228 |
.\" commit 52400ba946759af28442dee6265c5c0180ac7122 |
| 1229 |
This operation is a PI-aware variant of |
| 1230 |
.BR FUTEX_CMP_REQUEUE . |
| 1231 |
It requeues waiters that are blocked via |
| 1232 |
.B FUTEX_WAIT_REQUEUE_PI |
| 1233 |
on |
| 1234 |
.I uaddr |
| 1235 |
from a non-PI source futex |
| 1236 |
.RI ( uaddr ) |
| 1237 |
to a PI target futex |
| 1238 |
.RI ( uaddr2 ). |
| 1239 |
.IP |
| 1240 |
As with |
| 1241 |
.BR FUTEX_CMP_REQUEUE , |
| 1242 |
this operation wakes up a maximum of |
| 1243 |
.I val |
| 1244 |
waiters that are waiting on the futex at |
| 1245 |
.IR uaddr . |
| 1246 |
However, for |
| 1247 |
.BR FUTEX_CMP_REQUEUE_PI , |
| 1248 |
.I val |
| 1249 |
is required to be 1 |
| 1250 |
(since the main point is to avoid a thundering herd). |
| 1251 |
The remaining waiters are removed from the wait queue of the source futex at |
| 1252 |
.I uaddr |
| 1253 |
and added to the wait queue of the target futex at |
| 1254 |
.IR uaddr2 . |
| 1255 |
.IP |
| 1256 |
The |
| 1257 |
.I val2 |
| 1258 |
.\" val2 is the cap on the number of requeued waiters. |
| 1259 |
.\" In the glibc pthread_cond_broadcast() implementation, this argument |
| 1260 |
.\" is specified as INT_MAX, and for pthread_cond_signal() it is 0. |
| 1261 |
and |
| 1262 |
.I val3 |
| 1263 |
arguments serve the same purposes as for |
| 1264 |
.BR FUTEX_CMP_REQUEUE . |
| 1265 |
.\" |
| 1266 |
.\" The page at http://locklessinc.com/articles/futex_cheat_sheet/ |
| 1267 |
.\" notes that "priority-inheritance Futex to priority-inheritance |
| 1268 |
.\" Futex requeues are currently unsupported". However, probably |
| 1269 |
.\" the page does not need to say nothing about this, since |
| 1270 |
.\" Thomas Gleixner commented (July 2015): "they never will be |
| 1271 |
.\" supported because they make no sense at all" |
| 1272 |
.\" |
| 1273 |
.\"""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""" |
| 1274 |
.\" |
| 1275 |
.TP |
| 1276 |
.BR FUTEX_WAIT_REQUEUE_PI " (since Linux 2.6.31)" |
| 1277 |
.\" commit 52400ba946759af28442dee6265c5c0180ac7122 |
| 1278 |
.\" |
| 1279 |
Wait on a non-PI futex at |
| 1280 |
.I uaddr |
| 1281 |
and potentially be requeued (via a |
| 1282 |
.B FUTEX_CMP_REQUEUE_PI |
| 1283 |
operation in another task) onto a PI futex at |
| 1284 |
.IR uaddr2 . |
| 1285 |
The wait operation on |
| 1286 |
.I uaddr |
| 1287 |
is the same as for |
| 1288 |
.BR FUTEX_WAIT . |
| 1289 |
.IP |
| 1290 |
The waiter can be removed from the wait on |
| 1291 |
.I uaddr |
| 1292 |
without requeueing on |
| 1293 |
.I uaddr2 |
| 1294 |
via a |
| 1295 |
.B FUTEX_WAKE |
| 1296 |
operation in another task. |
| 1297 |
In this case, the |
| 1298 |
.B FUTEX_WAIT_REQUEUE_PI |
| 1299 |
operation fails with the error |
| 1300 |
.BR EAGAIN . |
| 1301 |
.IP |
| 1302 |
If |
| 1303 |
.I timeout |
| 1304 |
is not NULL, the structure it points to specifies |
| 1305 |
an absolute timeout for the wait operation. |
| 1306 |
If |
| 1307 |
.I timeout |
| 1308 |
is NULL, the operation can block indefinitely. |
| 1309 |
.IP |
| 1310 |
The |
| 1311 |
.I val3 |
| 1312 |
argument is ignored. |
| 1313 |
.IP |
| 1314 |
The |
| 1315 |
.B FUTEX_WAIT_REQUEUE_PI |
| 1316 |
and |
| 1317 |
.B FUTEX_CMP_REQUEUE_PI |
| 1318 |
were added to support a fairly specific use case: |
| 1319 |
support for priority-inheritance-aware POSIX threads condition variables. |
| 1320 |
The idea is that these operations should always be paired, |
| 1321 |
in order to ensure that user space and the kernel remain in sync. |
| 1322 |
Thus, in the |
| 1323 |
.B FUTEX_WAIT_REQUEUE_PI |
| 1324 |
operation, the user-space application pre-specifies the target |
| 1325 |
of the requeue that takes place in the |
| 1326 |
.B FUTEX_CMP_REQUEUE_PI |
| 1327 |
operation. |
| 1328 |
.\" |
| 1329 |
.\" Darren Hart notes that a patch to allow glibc to fully support |
| 1330 |
.\" PI-aware pthreads condition variables has not yet been accepted into |
| 1331 |
.\" glibc. The story is complex, and can be found at |
| 1332 |
.\" https://sourceware.org/bugzilla/show_bug.cgi?id=11588 |
| 1333 |
.\" Darren notes that in the meantime, the patch is shipped with various |
| 1334 |
.\" PREEMPT_RT-enabled Linux systems. |
| 1335 |
.\" |
| 1336 |
.\" Related to the preceding, Darren proposed that somewhere, man-pages |
| 1337 |
.\" should document the following point: |
| 1338 |
.\" |
| 1339 |
.\" While the Linux kernel, since Linux 2.6.31, supports requeueing of |
| 1340 |
.\" priority-inheritance (PI) aware mutexes via the |
| 1341 |
.\" FUTEX_WAIT_REQUEUE_PI and FUTEX_CMP_REQUEUE_PI futex operations, |
| 1342 |
.\" the glibc implementation does not yet take full advantage of this. |
| 1343 |
.\" Specifically, the condvar internal data lock remains a non-PI aware |
| 1344 |
.\" mutex, regardless of the type of the pthread_mutex associated with |
| 1345 |
.\" the condvar. This can lead to an unbounded priority inversion on |
| 1346 |
.\" the internal data lock even when associating a PI aware |
| 1347 |
.\" pthread_mutex with a condvar during a pthread_cond*_wait |
| 1348 |
.\" operation. For this reason, it is not recommended to rely on |
| 1349 |
.\" priority inheritance when using pthread condition variables. |
| 1350 |
.\" |
| 1351 |
.\" The problem is that the obvious location for this text is |
| 1352 |
.\" the pthread_cond*wait(3) man page. However, such a man page |
| 1353 |
.\" does not currently exist. |
| 1354 |
.\" |
| 1355 |
.\"""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""" |
| 1356 |
.\" |
| 1357 |
.SH RETURN VALUE |
| 1358 |
In the event of an error (and assuming that |
| 1359 |
.BR futex () |
| 1360 |
was invoked via |
| 1361 |
.BR syscall (2)), |
| 1362 |
all operations return \-1 and set |
| 1363 |
.I errno |
| 1364 |
to indicate the error. |
| 1365 |
.P |
| 1366 |
The return value on success depends on the operation, |
| 1367 |
as described in the following list: |
| 1368 |
.TP |
| 1369 |
.B FUTEX_WAIT |
| 1370 |
Returns 0 if the caller was woken up. |
| 1371 |
Note that a wake-up can also be caused by common futex usage patterns |
| 1372 |
in unrelated code that happened to have previously used the futex word's |
| 1373 |
memory location (e.g., typical futex-based implementations of |
| 1374 |
Pthreads mutexes can cause this under some conditions). |
| 1375 |
Therefore, callers should always conservatively assume that a return |
| 1376 |
value of 0 can mean a spurious wake-up, and use the futex word's value |
| 1377 |
(i.e., the user-space synchronization scheme) |
| 1378 |
to decide whether to continue to block or not. |
| 1379 |
.TP |
| 1380 |
.B FUTEX_WAKE |
| 1381 |
Returns the number of waiters that were woken up. |
| 1382 |
.TP |
| 1383 |
.B FUTEX_FD |
| 1384 |
Returns the new file descriptor associated with the futex. |
| 1385 |
.TP |
| 1386 |
.B FUTEX_REQUEUE |
| 1387 |
Returns the number of waiters that were woken up. |
| 1388 |
.TP |
| 1389 |
.B FUTEX_CMP_REQUEUE |
| 1390 |
Returns the total number of waiters that were woken up or |
| 1391 |
requeued to the futex for the futex word at |
| 1392 |
.IR uaddr2 . |
| 1393 |
If this value is greater than |
| 1394 |
.IR val , |
| 1395 |
then the difference is the number of waiters requeued to the futex for the |
| 1396 |
futex word at |
| 1397 |
.IR uaddr2 . |
| 1398 |
.TP |
| 1399 |
.B FUTEX_WAKE_OP |
| 1400 |
Returns the total number of waiters that were woken up. |
| 1401 |
This is the sum of the woken waiters on the two futexes for |
| 1402 |
the futex words at |
| 1403 |
.I uaddr |
| 1404 |
and |
| 1405 |
.IR uaddr2 . |
| 1406 |
.TP |
| 1407 |
.B FUTEX_WAIT_BITSET |
| 1408 |
Returns 0 if the caller was woken up. |
| 1409 |
See |
| 1410 |
.B FUTEX_WAIT |
| 1411 |
for how to interpret this correctly in practice. |
| 1412 |
.TP |
| 1413 |
.B FUTEX_WAKE_BITSET |
| 1414 |
Returns the number of waiters that were woken up. |
| 1415 |
.TP |
| 1416 |
.B FUTEX_LOCK_PI |
| 1417 |
Returns 0 if the futex was successfully locked. |
| 1418 |
.TP |
| 1419 |
.B FUTEX_LOCK_PI2 |
| 1420 |
Returns 0 if the futex was successfully locked. |
| 1421 |
.TP |
| 1422 |
.B FUTEX_TRYLOCK_PI |
| 1423 |
Returns 0 if the futex was successfully locked. |
| 1424 |
.TP |
| 1425 |
.B FUTEX_UNLOCK_PI |
| 1426 |
Returns 0 if the futex was successfully unlocked. |
| 1427 |
.TP |
| 1428 |
.B FUTEX_CMP_REQUEUE_PI |
| 1429 |
Returns the total number of waiters that were woken up or |
| 1430 |
requeued to the futex for the futex word at |
| 1431 |
.IR uaddr2 . |
| 1432 |
If this value is greater than |
| 1433 |
.IR val , |
| 1434 |
then difference is the number of waiters requeued to the futex for |
| 1435 |
the futex word at |
| 1436 |
.IR uaddr2 . |
| 1437 |
.TP |
| 1438 |
.B FUTEX_WAIT_REQUEUE_PI |
| 1439 |
Returns 0 if the caller was successfully requeued to the futex for |
| 1440 |
the futex word at |
| 1441 |
.IR uaddr2 . |
| 1442 |
.\" |
| 1443 |
.\"""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""" |
| 1444 |
.\" |
| 1445 |
.SH ERRORS |
| 1446 |
.TP |
| 1447 |
.B EACCES |
| 1448 |
No read access to the memory of a futex word. |
| 1449 |
.TP |
| 1450 |
.B EAGAIN |
| 1451 |
.RB ( FUTEX_WAIT , |
| 1452 |
.BR FUTEX_WAIT_BITSET , |
| 1453 |
.BR FUTEX_WAIT_REQUEUE_PI ) |
| 1454 |
The value pointed to by |
| 1455 |
.I uaddr |
| 1456 |
was not equal to the expected value |
| 1457 |
.I val |
| 1458 |
at the time of the call. |
| 1459 |
.IP |
| 1460 |
.BR Note : |
| 1461 |
on Linux, the symbolic names |
| 1462 |
.B EAGAIN |
| 1463 |
and |
| 1464 |
.B EWOULDBLOCK |
| 1465 |
(both of which appear in different parts of the kernel futex code) |
| 1466 |
have the same value. |
| 1467 |
.TP |
| 1468 |
.B EAGAIN |
| 1469 |
.RB ( FUTEX_CMP_REQUEUE , |
| 1470 |
.BR FUTEX_CMP_REQUEUE_PI ) |
| 1471 |
The value pointed to by |
| 1472 |
.I uaddr |
| 1473 |
is not equal to the expected value |
| 1474 |
.IR val3 . |
| 1475 |
.TP |
| 1476 |
.B EAGAIN |
| 1477 |
.RB ( FUTEX_LOCK_PI , |
| 1478 |
.BR FUTEX_LOCK_PI2 , |
| 1479 |
.BR FUTEX_TRYLOCK_PI , |
| 1480 |
.BR FUTEX_CMP_REQUEUE_PI ) |
| 1481 |
The futex owner thread ID of |
| 1482 |
.I uaddr |
| 1483 |
(for |
| 1484 |
.BR FUTEX_CMP_REQUEUE_PI : |
| 1485 |
.IR uaddr2 ) |
| 1486 |
is about to exit, |
| 1487 |
but has not yet handled the internal state cleanup. |
| 1488 |
Try again. |
| 1489 |
.TP |
| 1490 |
.B EDEADLK |
| 1491 |
.RB ( FUTEX_LOCK_PI , |
| 1492 |
.BR FUTEX_LOCK_PI2 , |
| 1493 |
.BR FUTEX_TRYLOCK_PI , |
| 1494 |
.BR FUTEX_CMP_REQUEUE_PI ) |
| 1495 |
The futex word at |
| 1496 |
.I uaddr |
| 1497 |
is already locked by the caller. |
| 1498 |
.TP |
| 1499 |
.B EDEADLK |
| 1500 |
.\" FIXME . I see that kernel/locking/rtmutex.c uses EDEADLK in some |
| 1501 |
.\" places, and EDEADLOCK in others. On almost all architectures |
| 1502 |
.\" these constants are synonymous. Is there a reason that both |
| 1503 |
.\" names are used? |
| 1504 |
.\" |
| 1505 |
.\" tglx (July 2015): "No. We should probably fix that." |
| 1506 |
.\" |
| 1507 |
.RB ( FUTEX_CMP_REQUEUE_PI ) |
| 1508 |
While requeueing a waiter to the PI futex for the futex word at |
| 1509 |
.IR uaddr2 , |
| 1510 |
the kernel detected a deadlock. |
| 1511 |
.TP |
| 1512 |
.B EFAULT |
| 1513 |
A required pointer argument (i.e., |
| 1514 |
.IR uaddr , |
| 1515 |
.IR uaddr2 , |
| 1516 |
or |
| 1517 |
.IR timeout ) |
| 1518 |
did not point to a valid user-space address. |
| 1519 |
.TP |
| 1520 |
.B EINTR |
| 1521 |
A |
| 1522 |
.B FUTEX_WAIT |
| 1523 |
or |
| 1524 |
.B FUTEX_WAIT_BITSET |
| 1525 |
operation was interrupted by a signal (see |
| 1526 |
.BR signal (7)). |
| 1527 |
Before Linux 2.6.22, this error could also be returned for |
| 1528 |
a spurious wakeup; since Linux 2.6.22, this no longer happens. |
| 1529 |
.TP |
| 1530 |
.B EINVAL |
| 1531 |
The operation in |
| 1532 |
.I futex_op |
| 1533 |
is one of those that employs a timeout, but the supplied |
| 1534 |
.I timeout |
| 1535 |
argument was invalid |
| 1536 |
.RI ( tv_sec |
| 1537 |
was less than zero, or |
| 1538 |
.I tv_nsec |
| 1539 |
was not less than 1,000,000,000). |
| 1540 |
.TP |
| 1541 |
.B EINVAL |
| 1542 |
The operation specified in |
| 1543 |
.I futex_op |
| 1544 |
employs one or both of the pointers |
| 1545 |
.I uaddr |
| 1546 |
and |
| 1547 |
.IR uaddr2 , |
| 1548 |
but one of these does not point to a valid object\[em]that is, |
| 1549 |
the address is not four-byte-aligned. |
| 1550 |
.TP |
| 1551 |
.B EINVAL |
| 1552 |
.RB ( FUTEX_WAIT_BITSET , |
| 1553 |
.BR FUTEX_WAKE_BITSET ) |
| 1554 |
The bit mask supplied in |
| 1555 |
.I val3 |
| 1556 |
is zero. |
| 1557 |
.TP |
| 1558 |
.B EINVAL |
| 1559 |
.RB ( FUTEX_CMP_REQUEUE_PI ) |
| 1560 |
.I uaddr |
| 1561 |
equals |
| 1562 |
.I uaddr2 |
| 1563 |
(i.e., an attempt was made to requeue to the same futex). |
| 1564 |
.TP |
| 1565 |
.B EINVAL |
| 1566 |
.RB ( FUTEX_FD ) |
| 1567 |
The signal number supplied in |
| 1568 |
.I val |
| 1569 |
is invalid. |
| 1570 |
.TP |
| 1571 |
.B EINVAL |
| 1572 |
.RB ( FUTEX_WAKE , |
| 1573 |
.BR FUTEX_WAKE_OP , |
| 1574 |
.BR FUTEX_WAKE_BITSET , |
| 1575 |
.BR FUTEX_REQUEUE , |
| 1576 |
.BR FUTEX_CMP_REQUEUE ) |
| 1577 |
The kernel detected an inconsistency between the user-space state at |
| 1578 |
.I uaddr |
| 1579 |
and the kernel state\[em]that is, it detected a waiter which waits in |
| 1580 |
.B FUTEX_LOCK_PI |
| 1581 |
or |
| 1582 |
.B FUTEX_LOCK_PI2 |
| 1583 |
on |
| 1584 |
.IR uaddr . |
| 1585 |
.TP |
| 1586 |
.B EINVAL |
| 1587 |
.RB ( FUTEX_LOCK_PI , |
| 1588 |
.BR FUTEX_LOCK_PI2 , |
| 1589 |
.BR FUTEX_TRYLOCK_PI , |
| 1590 |
.BR FUTEX_UNLOCK_PI ) |
| 1591 |
The kernel detected an inconsistency between the user-space state at |
| 1592 |
.I uaddr |
| 1593 |
and the kernel state. |
| 1594 |
This indicates either state corruption |
| 1595 |
or that the kernel found a waiter on |
| 1596 |
.I uaddr |
| 1597 |
which is waiting via |
| 1598 |
.B FUTEX_WAIT |
| 1599 |
or |
| 1600 |
.BR FUTEX_WAIT_BITSET . |
| 1601 |
.TP |
| 1602 |
.B EINVAL |
| 1603 |
.RB ( FUTEX_CMP_REQUEUE_PI ) |
| 1604 |
The kernel detected an inconsistency between the user-space state at |
| 1605 |
.I uaddr2 |
| 1606 |
and the kernel state; |
| 1607 |
.\" From a conversation with Thomas Gleixner (Aug 2015): ### |
| 1608 |
.\" The kernel sees: I have non PI state for a futex you tried to |
| 1609 |
.\" tell me was PI |
| 1610 |
that is, the kernel detected a waiter which waits via |
| 1611 |
.B FUTEX_WAIT |
| 1612 |
or |
| 1613 |
.B FUTEX_WAIT_BITSET |
| 1614 |
on |
| 1615 |
.IR uaddr2 . |
| 1616 |
.TP |
| 1617 |
.B EINVAL |
| 1618 |
.RB ( FUTEX_CMP_REQUEUE_PI ) |
| 1619 |
The kernel detected an inconsistency between the user-space state at |
| 1620 |
.I uaddr |
| 1621 |
and the kernel state; |
| 1622 |
that is, the kernel detected a waiter which waits via |
| 1623 |
.B FUTEX_WAIT |
| 1624 |
or |
| 1625 |
.B FUTEX_WAIT_BITSET |
| 1626 |
on |
| 1627 |
.IR uaddr . |
| 1628 |
.TP |
| 1629 |
.B EINVAL |
| 1630 |
.RB ( FUTEX_CMP_REQUEUE_PI ) |
| 1631 |
The kernel detected an inconsistency between the user-space state at |
| 1632 |
.I uaddr |
| 1633 |
and the kernel state; |
| 1634 |
that is, the kernel detected a waiter which waits on |
| 1635 |
.I uaddr |
| 1636 |
via |
| 1637 |
.B FUTEX_LOCK_PI |
| 1638 |
or |
| 1639 |
.B FUTEX_LOCK_PI2 |
| 1640 |
(instead of |
| 1641 |
.BR FUTEX_WAIT_REQUEUE_PI ). |
| 1642 |
.TP |
| 1643 |
.B EINVAL |
| 1644 |
.RB ( FUTEX_CMP_REQUEUE_PI ) |
| 1645 |
.\" This deals with the case: |
| 1646 |
.\" wait_requeue_pi(A, B); |
| 1647 |
.\" requeue_pi(A, C); |
| 1648 |
An attempt was made to requeue a waiter to a futex other than that |
| 1649 |
specified by the matching |
| 1650 |
.B FUTEX_WAIT_REQUEUE_PI |
| 1651 |
call for that waiter. |
| 1652 |
.TP |
| 1653 |
.B EINVAL |
| 1654 |
.RB ( FUTEX_CMP_REQUEUE_PI ) |
| 1655 |
The |
| 1656 |
.I val |
| 1657 |
argument is not 1. |
| 1658 |
.TP |
| 1659 |
.B EINVAL |
| 1660 |
Invalid argument. |
| 1661 |
.TP |
| 1662 |
.B ENFILE |
| 1663 |
.RB ( FUTEX_FD ) |
| 1664 |
The system-wide limit on the total number of open files has been reached. |
| 1665 |
.TP |
| 1666 |
.B ENOMEM |
| 1667 |
.RB ( FUTEX_LOCK_PI , |
| 1668 |
.BR FUTEX_LOCK_PI2 , |
| 1669 |
.BR FUTEX_TRYLOCK_PI , |
| 1670 |
.BR FUTEX_CMP_REQUEUE_PI ) |
| 1671 |
The kernel could not allocate memory to hold state information. |
| 1672 |
.TP |
| 1673 |
.B ENOSYS |
| 1674 |
Invalid operation specified in |
| 1675 |
.IR futex_op . |
| 1676 |
.TP |
| 1677 |
.B ENOSYS |
| 1678 |
The |
| 1679 |
.B FUTEX_CLOCK_REALTIME |
| 1680 |
option was specified in |
| 1681 |
.IR futex_op , |
| 1682 |
but the accompanying operation was neither |
| 1683 |
.BR FUTEX_WAIT , |
| 1684 |
.BR FUTEX_WAIT_BITSET , |
| 1685 |
.BR FUTEX_WAIT_REQUEUE_PI , |
| 1686 |
nor |
| 1687 |
.BR FUTEX_LOCK_PI2 . |
| 1688 |
.TP |
| 1689 |
.B ENOSYS |
| 1690 |
.RB ( FUTEX_LOCK_PI , |
| 1691 |
.BR FUTEX_LOCK_PI2 , |
| 1692 |
.BR FUTEX_TRYLOCK_PI , |
| 1693 |
.BR FUTEX_UNLOCK_PI , |
| 1694 |
.BR FUTEX_CMP_REQUEUE_PI , |
| 1695 |
.BR FUTEX_WAIT_REQUEUE_PI ) |
| 1696 |
A run-time check determined that the operation is not available. |
| 1697 |
The PI-futex operations are not implemented on all architectures and |
| 1698 |
are not supported on some CPU variants. |
| 1699 |
.TP |
| 1700 |
.B EPERM |
| 1701 |
.RB ( FUTEX_LOCK_PI , |
| 1702 |
.BR FUTEX_LOCK_PI2 , |
| 1703 |
.BR FUTEX_TRYLOCK_PI , |
| 1704 |
.BR FUTEX_CMP_REQUEUE_PI ) |
| 1705 |
The caller is not allowed to attach itself to the futex at |
| 1706 |
.I uaddr |
| 1707 |
(for |
| 1708 |
.BR FUTEX_CMP_REQUEUE_PI : |
| 1709 |
the futex at |
| 1710 |
.IR uaddr2 ). |
| 1711 |
(This may be caused by a state corruption in user space.) |
| 1712 |
.TP |
| 1713 |
.B EPERM |
| 1714 |
.RB ( FUTEX_UNLOCK_PI ) |
| 1715 |
The caller does not own the lock represented by the futex word. |
| 1716 |
.TP |
| 1717 |
.B ESRCH |
| 1718 |
.RB ( FUTEX_LOCK_PI , |
| 1719 |
.BR FUTEX_LOCK_PI2 , |
| 1720 |
.BR FUTEX_TRYLOCK_PI , |
| 1721 |
.BR FUTEX_CMP_REQUEUE_PI ) |
| 1722 |
The thread ID in the futex word at |
| 1723 |
.I uaddr |
| 1724 |
does not exist. |
| 1725 |
.TP |
| 1726 |
.B ESRCH |
| 1727 |
.RB ( FUTEX_CMP_REQUEUE_PI ) |
| 1728 |
The thread ID in the futex word at |
| 1729 |
.I uaddr2 |
| 1730 |
does not exist. |
| 1731 |
.TP |
| 1732 |
.B ETIMEDOUT |
| 1733 |
The operation in |
| 1734 |
.I futex_op |
| 1735 |
employed the timeout specified in |
| 1736 |
.IR timeout , |
| 1737 |
and the timeout expired before the operation completed. |
| 1738 |
.\" |
| 1739 |
.\"""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""" |
| 1740 |
.\" |
| 1741 |
.SH STANDARDS |
| 1742 |
Linux. |
| 1743 |
.SH HISTORY |
| 1744 |
Linux 2.6.0. |
| 1745 |
.P |
| 1746 |
Initial futex support was merged in Linux 2.5.7 but with different |
| 1747 |
semantics from what was described above. |
| 1748 |
A four-argument system call with the semantics |
| 1749 |
described in this page was introduced in Linux 2.5.40. |
| 1750 |
A fifth argument was added in Linux 2.5.70, |
| 1751 |
and a sixth argument was added in Linux 2.6.7. |
| 1752 |
.SH EXAMPLES |
| 1753 |
The program below demonstrates use of futexes in a program where a parent |
| 1754 |
process and a child process use a pair of futexes located inside a |
| 1755 |
shared anonymous mapping to synchronize access to a shared resource: |
| 1756 |
the terminal. |
| 1757 |
The two processes each write |
| 1758 |
.I nloops |
| 1759 |
(a command-line argument that defaults to 5 if omitted) |
| 1760 |
messages to the terminal and employ a synchronization protocol |
| 1761 |
that ensures that they alternate in writing messages. |
| 1762 |
Upon running this program we see output such as the following: |
| 1763 |
.P |
| 1764 |
.in +4n |
| 1765 |
.EX |
| 1766 |
$ \fB./futex_demo\fP |
| 1767 |
Parent (18534) 0 |
| 1768 |
Child (18535) 0 |
| 1769 |
Parent (18534) 1 |
| 1770 |
Child (18535) 1 |
| 1771 |
Parent (18534) 2 |
| 1772 |
Child (18535) 2 |
| 1773 |
Parent (18534) 3 |
| 1774 |
Child (18535) 3 |
| 1775 |
Parent (18534) 4 |
| 1776 |
Child (18535) 4 |
| 1777 |
.EE |
| 1778 |
.in |
| 1779 |
.SS Program source |
| 1780 |
\& |
| 1781 |
.\" SRC BEGIN (futex.c) |
| 1782 |
.EX |
| 1783 |
/* futex_demo.c |
| 1784 |
\& |
| 1785 |
Usage: futex_demo [nloops] |
| 1786 |
(Default: 5) |
| 1787 |
\& |
| 1788 |
Demonstrate the use of futexes in a program where parent and child |
| 1789 |
use a pair of futexes located inside a shared anonymous mapping to |
| 1790 |
synchronize access to a shared resource: the terminal. The two |
| 1791 |
processes each write \[aq]num\-loops\[aq] messages to the terminal and employ |
| 1792 |
a synchronization protocol that ensures that they alternate in |
| 1793 |
writing messages. |
| 1794 |
*/ |
| 1795 |
#define _GNU_SOURCE |
| 1796 |
#include <err.h> |
| 1797 |
#include <errno.h> |
| 1798 |
#include <linux/futex.h> |
| 1799 |
#include <stdatomic.h> |
| 1800 |
#include <stdint.h> |
| 1801 |
#include <stdio.h> |
| 1802 |
#include <stdlib.h> |
| 1803 |
#include <sys/mman.h> |
| 1804 |
#include <sys/syscall.h> |
| 1805 |
#include <sys/time.h> |
| 1806 |
#include <sys/wait.h> |
| 1807 |
#include <unistd.h> |
| 1808 |
\& |
| 1809 |
static uint32_t *futex1, *futex2, *iaddr; |
| 1810 |
\& |
| 1811 |
static int |
| 1812 |
futex(uint32_t *uaddr, int futex_op, uint32_t val, |
| 1813 |
const struct timespec *timeout, uint32_t *uaddr2, uint32_t val3) |
| 1814 |
{ |
| 1815 |
return syscall(SYS_futex, uaddr, futex_op, val, |
| 1816 |
timeout, uaddr2, val3); |
| 1817 |
} |
| 1818 |
\& |
| 1819 |
/* Acquire the futex pointed to by \[aq]futexp\[aq]: wait for its value to |
| 1820 |
become 1, and then set the value to 0. */ |
| 1821 |
\& |
| 1822 |
static void |
| 1823 |
fwait(uint32_t *futexp) |
| 1824 |
{ |
| 1825 |
long s; |
| 1826 |
const uint32_t one = 1; |
| 1827 |
\& |
| 1828 |
/* atomic_compare_exchange_strong(ptr, oldval, newval) |
| 1829 |
atomically performs the equivalent of: |
| 1830 |
\& |
| 1831 |
if (*ptr == *oldval) |
| 1832 |
*ptr = newval; |
| 1833 |
\& |
| 1834 |
It returns true if the test yielded true and *ptr was updated. */ |
| 1835 |
\& |
| 1836 |
while (1) { |
| 1837 |
\& |
| 1838 |
/* Is the futex available? */ |
| 1839 |
if (atomic_compare_exchange_strong(futexp, &one, 0)) |
| 1840 |
break; /* Yes */ |
| 1841 |
\& |
| 1842 |
/* Futex is not available; wait. */ |
| 1843 |
\& |
| 1844 |
s = futex(futexp, FUTEX_WAIT, 0, NULL, NULL, 0); |
| 1845 |
if (s == \-1 && errno != EAGAIN) |
| 1846 |
err(EXIT_FAILURE, "futex\-FUTEX_WAIT"); |
| 1847 |
} |
| 1848 |
} |
| 1849 |
\& |
| 1850 |
/* Release the futex pointed to by \[aq]futexp\[aq]: if the futex currently |
| 1851 |
has the value 0, set its value to 1 and then wake any futex waiters, |
| 1852 |
so that if the peer is blocked in fwait(), it can proceed. */ |
| 1853 |
\& |
| 1854 |
static void |
| 1855 |
fpost(uint32_t *futexp) |
| 1856 |
{ |
| 1857 |
long s; |
| 1858 |
const uint32_t zero = 0; |
| 1859 |
\& |
| 1860 |
/* atomic_compare_exchange_strong() was described |
| 1861 |
in comments above. */ |
| 1862 |
\& |
| 1863 |
if (atomic_compare_exchange_strong(futexp, &zero, 1)) { |
| 1864 |
s = futex(futexp, FUTEX_WAKE, 1, NULL, NULL, 0); |
| 1865 |
if (s == \-1) |
| 1866 |
err(EXIT_FAILURE, "futex\-FUTEX_WAKE"); |
| 1867 |
} |
| 1868 |
} |
| 1869 |
\& |
| 1870 |
int |
| 1871 |
main(int argc, char *argv[]) |
| 1872 |
{ |
| 1873 |
pid_t childPid; |
| 1874 |
unsigned int nloops; |
| 1875 |
\& |
| 1876 |
setbuf(stdout, NULL); |
| 1877 |
\& |
| 1878 |
nloops = (argc > 1) ? atoi(argv[1]) : 5; |
| 1879 |
\& |
| 1880 |
/* Create a shared anonymous mapping that will hold the futexes. |
| 1881 |
Since the futexes are being shared between processes, we |
| 1882 |
subsequently use the "shared" futex operations (i.e., not the |
| 1883 |
ones suffixed "_PRIVATE"). */ |
| 1884 |
\& |
| 1885 |
iaddr = mmap(NULL, sizeof(*iaddr) * 2, PROT_READ | PROT_WRITE, |
| 1886 |
MAP_ANONYMOUS | MAP_SHARED, \-1, 0); |
| 1887 |
if (iaddr == MAP_FAILED) |
| 1888 |
err(EXIT_FAILURE, "mmap"); |
| 1889 |
\& |
| 1890 |
futex1 = &iaddr[0]; |
| 1891 |
futex2 = &iaddr[1]; |
| 1892 |
\& |
| 1893 |
*futex1 = 0; /* State: unavailable */ |
| 1894 |
*futex2 = 1; /* State: available */ |
| 1895 |
\& |
| 1896 |
/* Create a child process that inherits the shared anonymous |
| 1897 |
mapping. */ |
| 1898 |
\& |
| 1899 |
childPid = fork(); |
| 1900 |
if (childPid == \-1) |
| 1901 |
err(EXIT_FAILURE, "fork"); |
| 1902 |
\& |
| 1903 |
if (childPid == 0) { /* Child */ |
| 1904 |
for (unsigned int j = 0; j < nloops; j++) { |
| 1905 |
fwait(futex1); |
| 1906 |
printf("Child (%jd) %u\[rs]n", (intmax_t) getpid(), j); |
| 1907 |
fpost(futex2); |
| 1908 |
} |
| 1909 |
\& |
| 1910 |
exit(EXIT_SUCCESS); |
| 1911 |
} |
| 1912 |
\& |
| 1913 |
/* Parent falls through to here. */ |
| 1914 |
\& |
| 1915 |
for (unsigned int j = 0; j < nloops; j++) { |
| 1916 |
fwait(futex2); |
| 1917 |
printf("Parent (%jd) %u\[rs]n", (intmax_t) getpid(), j); |
| 1918 |
fpost(futex1); |
| 1919 |
} |
| 1920 |
\& |
| 1921 |
wait(NULL); |
| 1922 |
\& |
| 1923 |
exit(EXIT_SUCCESS); |
| 1924 |
} |
| 1925 |
.EE |
| 1926 |
.\" SRC END |
| 1927 |
.SH SEE ALSO |
| 1928 |
.ad l |
| 1929 |
.BR get_robust_list (2), |
| 1930 |
.BR restart_syscall (2), |
| 1931 |
.BR pthread_mutexattr_getprotocol (3), |
| 1932 |
.BR futex (7), |
| 1933 |
.BR sched (7) |
| 1934 |
.P |
| 1935 |
The following kernel source files: |
| 1936 |
.IP \[bu] 3 |
| 1937 |
.I Documentation/pi\-futex.txt |
| 1938 |
.IP \[bu] |
| 1939 |
.I Documentation/futex\-requeue\-pi.txt |
| 1940 |
.IP \[bu] |
| 1941 |
.I Documentation/locking/rt\-mutex.txt |
| 1942 |
.IP \[bu] |
| 1943 |
.I Documentation/locking/rt\-mutex\-design.txt |
| 1944 |
.IP \[bu] |
| 1945 |
.I Documentation/robust\-futex\-ABI.txt |
| 1946 |
.P |
| 1947 |
Franke, H., Russell, R., and Kirwood, M., 2002. |
| 1948 |
\fIFuss, Futexes and Furwocks: Fast Userlevel Locking in Linux\fP |
| 1949 |
(from proceedings of the Ottawa Linux Symposium 2002), |
| 1950 |
.br |
| 1951 |
.UR http://kernel.org\:/doc\:/ols\:/2002\:/ols2002\-pages\-479\-495.pdf |
| 1952 |
.UE |
| 1953 |
.P |
| 1954 |
Hart, D., 2009. \fIA futex overview and update\fP, |
| 1955 |
.UR http://lwn.net/Articles/360699/ |
| 1956 |
.UE |
| 1957 |
.P |
| 1958 |
Hart, D.\& and Guniguntala, D., 2009. |
| 1959 |
\fIRequeue-PI: Making glibc Condvars PI-Aware\fP |
| 1960 |
(from proceedings of the 2009 Real-Time Linux Workshop), |
| 1961 |
.UR http://lwn.net/images/conf/rtlws11/papers/proc/p10.pdf |
| 1962 |
.UE |
| 1963 |
.P |
| 1964 |
Drepper, U., 2011. \fIFutexes Are Tricky\fP, |
| 1965 |
.UR http://www.akkadia.org/drepper/futex.pdf |
| 1966 |
.UE |
| 1967 |
.P |
| 1968 |
Futex example library, futex\-*.tar.bz2 at |
| 1969 |
.br |
| 1970 |
.UR https://mirrors.kernel.org\:/pub\:/linux\:/kernel\:/people\:/rusty/ |
| 1971 |
.UE |
| 1972 |
.\" |
| 1973 |
.\" FIXME(Torvald) We should probably refer to the glibc code here, in |
| 1974 |
.\" particular the glibc-internal futex wrapper functions that are |
| 1975 |
.\" WIP, and the generic pthread_mutex_t and perhaps condvar |
| 1976 |
.\" implementations. |