| 1 | /* Low-level lock implementation. Generic futex-based version. |
| 2 | Copyright (C) 2005-2023 Free Software Foundation, Inc. |
| 3 | This file is part of the GNU C Library. |
| 4 | |
| 5 | The GNU C Library is free software; you can redistribute it and/or |
| 6 | modify it under the terms of the GNU Lesser General Public |
| 7 | License as published by the Free Software Foundation; either |
| 8 | version 2.1 of the License, or (at your option) any later version. |
| 9 | |
| 10 | The GNU C Library is distributed in the hope that it will be useful, |
| 11 | but WITHOUT ANY WARRANTY; without even the implied warranty of |
| 12 | MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU |
| 13 | Lesser General Public License for more details. |
| 14 | |
| 15 | You should have received a copy of the GNU Lesser General Public |
| 16 | License along with the GNU C Library. If not, see |
| 17 | <https://www.gnu.org/licenses/>. */ |
| 18 | |
| 19 | #ifndef _LOWLEVELLOCK_H |
| 20 | #define _LOWLEVELLOCK_H 1 |
| 21 | |
| 22 | #include <atomic.h> |
| 23 | #include <elision-conf.h> |
| 24 | #include <lowlevellock-futex.h> |
| 25 | #include <time.h> |
| 26 | |
| 27 | /* Low-level locks use a combination of atomic operations (to acquire and |
| 28 | release lock ownership) and futex operations (to block until the state |
| 29 | of a lock changes). A lock can be in one of three states: |
| 30 | 0: not acquired, |
| 31 | 1: acquired with no waiters; no other threads are blocked or about to block |
| 32 | for changes to the lock state, |
| 33 | >1: acquired, possibly with waiters; there may be other threads blocked or |
| 34 | about to block for changes to the lock state. |
| 35 | |
| 36 | We expect that the common case is an uncontended lock, so we just need |
| 37 | to transition the lock between states 0 and 1; releasing the lock does |
| 38 | not need to wake any other blocked threads. If the lock is contended |
| 39 | and a thread decides to block using a futex operation, then this thread |
| 40 | needs to first change the state to >1; if this state is observed during |
| 41 | lock release, the releasing thread will wake one of the potentially |
| 42 | blocked threads. |
| 43 | |
| 44 | Much of this code takes a 'private' parameter. This may be: |
| 45 | LLL_PRIVATE: lock only shared within a process |
| 46 | LLL_SHARED: lock may be shared across processes. |
| 47 | |
| 48 | Condition variables contain an optimization for broadcasts that requeues |
| 49 | waiting threads on a lock's futex. Therefore, there is a special |
| 50 | variant of the locks (whose name contains "cond") that makes sure to |
| 51 | always set the lock state to >1 and not just 1. |
| 52 | |
| 53 | Robust locks set the lock to the id of the owner. This allows detection |
| 54 | of the case where the owner exits without releasing the lock. Flags are |
| 55 | OR'd with the owner id to record additional information about lock state. |
| 56 | Therefore the states of robust locks are: |
| 57 | 0: not acquired |
| 58 | id: acquired (by user identified by id & FUTEX_TID_MASK) |
| 59 | |
| 60 | The following flags may be set in the robust lock value: |
| 61 | FUTEX_WAITERS - possibly has waiters |
| 62 | FUTEX_OWNER_DIED - owning user has exited without releasing the futex. */ |
| 63 | |
| 64 | |
| 65 | /* If LOCK is 0 (not acquired), set to 1 (acquired with no waiters) and return |
| 66 | 0. Otherwise leave lock unchanged and return non-zero to indicate that the |
| 67 | lock was not acquired. */ |
| 68 | #define __lll_trylock(lock) \ |
| 69 | __glibc_unlikely (atomic_compare_and_exchange_bool_acq ((lock), 1, 0)) |
| 70 | #define lll_trylock(lock) \ |
| 71 | __lll_trylock (&(lock)) |
| 72 | |
| 73 | /* If LOCK is 0 (not acquired), set to 2 (acquired, possibly with waiters) and |
| 74 | return 0. Otherwise leave lock unchanged and return non-zero to indicate |
| 75 | that the lock was not acquired. */ |
| 76 | #define lll_cond_trylock(lock) \ |
| 77 | __glibc_unlikely (atomic_compare_and_exchange_bool_acq (&(lock), 2, 0)) |
| 78 | |
| 79 | extern void __lll_lock_wait_private (int *futex); |
| 80 | libc_hidden_proto (__lll_lock_wait_private) |
| 81 | extern void __lll_lock_wait (int *futex, int private); |
| 82 | libc_hidden_proto (__lll_lock_wait) |
| 83 | |
| 84 | /* This is an expression rather than a statement even though its value is |
| 85 | void, so that it can be used in a comma expression or as an expression |
| 86 | that's cast to void. */ |
| 87 | /* The inner conditional compiles to a call to __lll_lock_wait_private if |
| 88 | private is known at compile time to be LLL_PRIVATE, and to a call to |
| 89 | __lll_lock_wait otherwise. */ |
| 90 | /* If FUTEX is 0 (not acquired), set to 1 (acquired with no waiters) and |
| 91 | return. Otherwise, ensure that it is >1 (acquired, possibly with waiters) |
| 92 | and then block until we acquire the lock, at which point FUTEX will still be |
| 93 | >1. The lock is always acquired on return. */ |
| 94 | #define __lll_lock(futex, private) \ |
| 95 | ((void) \ |
| 96 | ({ \ |
| 97 | int *__futex = (futex); \ |
| 98 | if (__glibc_unlikely \ |
| 99 | (atomic_compare_and_exchange_bool_acq (__futex, 1, 0))) \ |
| 100 | { \ |
| 101 | if (__builtin_constant_p (private) && (private) == LLL_PRIVATE) \ |
| 102 | __lll_lock_wait_private (__futex); \ |
| 103 | else \ |
| 104 | __lll_lock_wait (__futex, private); \ |
| 105 | } \ |
| 106 | })) |
| 107 | #define lll_lock(futex, private) \ |
| 108 | __lll_lock (&(futex), private) |
| 109 | |
| 110 | |
| 111 | /* This is an expression rather than a statement even though its value is |
| 112 | void, so that it can be used in a comma expression or as an expression |
| 113 | that's cast to void. */ |
| 114 | /* Unconditionally set FUTEX to 2 (acquired, possibly with waiters). If FUTEX |
| 115 | was 0 (not acquired) then return. Otherwise, block until the lock is |
| 116 | acquired, at which point FUTEX is 2 (acquired, possibly with waiters). The |
| 117 | lock is always acquired on return. */ |
| 118 | #define __lll_cond_lock(futex, private) \ |
| 119 | ((void) \ |
| 120 | ({ \ |
| 121 | int *__futex = (futex); \ |
| 122 | if (__glibc_unlikely (atomic_exchange_acquire (__futex, 2) != 0)) \ |
| 123 | __lll_lock_wait (__futex, private); \ |
| 124 | })) |
| 125 | #define lll_cond_lock(futex, private) __lll_cond_lock (&(futex), private) |
| 126 | |
| 127 | |
| 128 | extern void __lll_lock_wake_private (int *futex); |
| 129 | libc_hidden_proto (__lll_lock_wake_private) |
| 130 | extern void __lll_lock_wake (int *futex, int private); |
| 131 | libc_hidden_proto (__lll_lock_wake) |
| 132 | |
| 133 | /* This is an expression rather than a statement even though its value is |
| 134 | void, so that it can be used in a comma expression or as an expression |
| 135 | that's cast to void. */ |
| 136 | /* Unconditionally set FUTEX to 0 (not acquired), releasing the lock. If FUTEX |
| 137 | was >1 (acquired, possibly with waiters), then wake any waiters. The waiter |
| 138 | that acquires the lock will set FUTEX to >1. |
| 139 | Evaluate PRIVATE before releasing the lock so that we do not violate the |
| 140 | mutex destruction requirements. Specifically, we need to ensure that |
| 141 | another thread can destroy the mutex (and reuse its memory) once it |
| 142 | acquires the lock and when there will be no further lock acquisitions; |
| 143 | thus, we must not access the lock after releasing it, or those accesses |
| 144 | could be concurrent with mutex destruction or reuse of the memory. */ |
| 145 | #define __lll_unlock(futex, private) \ |
| 146 | ((void) \ |
| 147 | ({ \ |
| 148 | int *__futex = (futex); \ |
| 149 | int __private = (private); \ |
| 150 | int __oldval = atomic_exchange_release (__futex, 0); \ |
| 151 | if (__glibc_unlikely (__oldval > 1)) \ |
| 152 | { \ |
| 153 | if (__builtin_constant_p (private) && (private) == LLL_PRIVATE) \ |
| 154 | __lll_lock_wake_private (__futex); \ |
| 155 | else \ |
| 156 | __lll_lock_wake (__futex, __private); \ |
| 157 | } \ |
| 158 | })) |
| 159 | #define lll_unlock(futex, private) \ |
| 160 | __lll_unlock (&(futex), private) |
| 161 | |
| 162 | |
| 163 | #define lll_islocked(futex) \ |
| 164 | ((futex) != LLL_LOCK_INITIALIZER) |
| 165 | |
| 166 | |
| 167 | /* Our internal lock implementation is identical to the binary-compatible |
| 168 | mutex implementation. */ |
| 169 | |
| 170 | /* Initializers for lock. */ |
| 171 | #define LLL_LOCK_INITIALIZER (0) |
| 172 | #define LLL_LOCK_INITIALIZER_LOCKED (1) |
| 173 | |
| 174 | /* Elision support. */ |
| 175 | |
| 176 | #if ENABLE_ELISION_SUPPORT |
| 177 | /* Force elision for all new locks. This is used to decide whether |
| 178 | existing DEFAULT locks should be automatically upgraded to elision |
| 179 | in pthread_mutex_lock. Disabled for suid programs. Only used when |
| 180 | elision is available. */ |
| 181 | extern int __pthread_force_elision; |
| 182 | libc_hidden_proto (__pthread_force_elision) |
| 183 | |
| 184 | extern void __lll_elision_init (void) attribute_hidden; |
| 185 | extern int __lll_clocklock_elision (int *futex, short *adapt_count, |
| 186 | clockid_t clockid, |
| 187 | const struct __timespec64 *timeout, |
| 188 | int private); |
| 189 | libc_hidden_proto (__lll_clocklock_elision) |
| 190 | |
| 191 | extern int __lll_lock_elision (int *futex, short *adapt_count, int private); |
| 192 | libc_hidden_proto (__lll_lock_elision) |
| 193 | |
| 194 | # if ELISION_UNLOCK_NEEDS_ADAPT_COUNT |
| 195 | extern int __lll_unlock_elision (int *lock, short *adapt_count, int private); |
| 196 | # else |
| 197 | extern int __lll_unlock_elision (int *lock, int private); |
| 198 | # endif |
| 199 | libc_hidden_proto (__lll_unlock_elision) |
| 200 | |
| 201 | extern int __lll_trylock_elision (int *lock, short *adapt_count); |
| 202 | libc_hidden_proto (__lll_trylock_elision) |
| 203 | |
| 204 | # define lll_clocklock_elision(futex, adapt_count, clockid, timeout, private) \ |
| 205 | __lll_clocklock_elision (&(futex), &(adapt_count), clockid, timeout, private) |
| 206 | # define lll_lock_elision(futex, adapt_count, private) \ |
| 207 | __lll_lock_elision (&(futex), &(adapt_count), private) |
| 208 | # define lll_trylock_elision(futex, adapt_count) \ |
| 209 | __lll_trylock_elision (&(futex), &(adapt_count)) |
| 210 | # if ELISION_UNLOCK_NEEDS_ADAPT_COUNT |
| 211 | # define lll_unlock_elision(futex, adapt_count, private) \ |
| 212 | __lll_unlock_elision (&(futex), &(adapt_count), private) |
| 213 | # else |
| 214 | # define lll_unlock_elision(futex, adapt_count, private) \ |
| 215 | __lll_unlock_elision (&(futex), private) |
| 216 | # endif |
| 217 | |
| 218 | /* Automatically enable elision for existing user lock kinds. */ |
| 219 | # define FORCE_ELISION(m, s) \ |
| 220 | if (__pthread_force_elision) \ |
| 221 | { \ |
| 222 | /* See concurrency notes regarding __kind in \ |
| 223 | struct __pthread_mutex_s in \ |
| 224 | sysdeps/nptl/bits/thread-shared-types.h. \ |
| 225 | \ |
| 226 | There are the following cases for the kind of a mutex \ |
| 227 | (The mask PTHREAD_MUTEX_ELISION_FLAGS_NP covers the flags \ |
| 228 | PTHREAD_MUTEX_ELISION_NP and PTHREAD_MUTEX_NO_ELISION_NP where \ |
| 229 | only one of both flags can be set): \ |
| 230 | - both flags are not set: \ |
| 231 | This is the first lock operation for this mutex. Enable \ |
| 232 | elision as it is not enabled so far. \ |
| 233 | Note: It can happen that multiple threads are calling e.g. \ |
| 234 | pthread_mutex_lock at the same time as the first lock \ |
| 235 | operation for this mutex. Then elision is enabled for this \ |
| 236 | mutex by multiple threads. Storing with relaxed MO is enough \ |
| 237 | as all threads will store the same new value for the kind of \ |
| 238 | the mutex. But we have to ensure that we always use the \ |
| 239 | elision path regardless if this thread has enabled elision or \ |
| 240 | another one. \ |
| 241 | \ |
| 242 | - PTHREAD_MUTEX_ELISION_NP flag is set: \ |
| 243 | Elision was already enabled for this mutex by a previous lock \ |
| 244 | operation. See case above. Just use the elision path. \ |
| 245 | \ |
| 246 | - PTHREAD_MUTEX_NO_ELISION_NP flag is set: \ |
| 247 | Elision was explicitly disabled by pthread_mutexattr_settype. \ |
| 248 | Do not use the elision path. \ |
| 249 | Note: The flag PTHREAD_MUTEX_NO_ELISION_NP will never be \ |
| 250 | changed after mutex initialization. */ \ |
| 251 | int mutex_kind = atomic_load_relaxed (&((m)->__data.__kind)); \ |
| 252 | if ((mutex_kind & PTHREAD_MUTEX_ELISION_FLAGS_NP) == 0) \ |
| 253 | { \ |
| 254 | mutex_kind |= PTHREAD_MUTEX_ELISION_NP; \ |
| 255 | atomic_store_relaxed (&((m)->__data.__kind), mutex_kind); \ |
| 256 | } \ |
| 257 | if ((mutex_kind & PTHREAD_MUTEX_ELISION_NP) != 0) \ |
| 258 | { \ |
| 259 | s; \ |
| 260 | } \ |
| 261 | } |
| 262 | |
| 263 | #else /* !ENABLE_ELISION_SUPPORT */ |
| 264 | |
| 265 | # define lll_clocklock_elision(futex, adapt_count, clockid, abstime, private) \ |
| 266 | __futex_clocklock64 (&(futex), clockid, abstime, private) |
| 267 | # define lll_lock_elision(lock, try_lock, private) \ |
| 268 | ({ lll_lock (lock, private); 0; }) |
| 269 | # define lll_trylock_elision(a,t) lll_trylock(a) |
| 270 | # define lll_unlock_elision(a,b,c) ({ lll_unlock (a,c); 0; }) |
| 271 | # define FORCE_ELISION(m, s) |
| 272 | |
| 273 | #endif /* !ENABLE_ELISION_SUPPORT */ |
| 274 | |
| 275 | #endif /* lowlevellock.h */ |
| 276 | |