Paolo Bonzini | 51dee5e | 2017-01-12 19:07:52 +0100 | [diff] [blame] | 1 | /* |
| 2 | * QemuLockCnt implementation |
| 3 | * |
| 4 | * Copyright Red Hat, Inc. 2017 |
| 5 | * |
| 6 | * Author: |
| 7 | * Paolo Bonzini <pbonzini@redhat.com> |
| 8 | */ |
| 9 | #include "qemu/osdep.h" |
| 10 | #include "qemu/thread.h" |
| 11 | #include "qemu/atomic.h" |
Paolo Bonzini | fbcc3e5 | 2017-01-12 19:07:54 +0100 | [diff] [blame] | 12 | #include "trace.h" |
Paolo Bonzini | 51dee5e | 2017-01-12 19:07:52 +0100 | [diff] [blame] | 13 | |
Paolo Bonzini | fbcc3e5 | 2017-01-12 19:07:54 +0100 | [diff] [blame] | 14 | #ifdef CONFIG_LINUX |
| 15 | #include "qemu/futex.h" |
| 16 | |
| 17 | /* On Linux, bits 0-1 are a futex-based lock, bits 2-31 are the counter. |
| 18 | * For the mutex algorithm see Ulrich Drepper's "Futexes Are Tricky" (ok, |
| 19 | * this is not the most relaxing citation I could make...). It is similar |
| 20 | * to mutex2 in the paper. |
| 21 | */ |
| 22 | |
| 23 | #define QEMU_LOCKCNT_STATE_MASK 3 |
| 24 | #define QEMU_LOCKCNT_STATE_FREE 0 /* free, uncontended */ |
| 25 | #define QEMU_LOCKCNT_STATE_LOCKED 1 /* locked, uncontended */ |
| 26 | #define QEMU_LOCKCNT_STATE_WAITING 2 /* locked, contended */ |
| 27 | |
| 28 | #define QEMU_LOCKCNT_COUNT_STEP 4 |
| 29 | #define QEMU_LOCKCNT_COUNT_SHIFT 2 |
| 30 | |
| 31 | void qemu_lockcnt_init(QemuLockCnt *lockcnt) |
| 32 | { |
| 33 | lockcnt->count = 0; |
| 34 | } |
| 35 | |
| 36 | void qemu_lockcnt_destroy(QemuLockCnt *lockcnt) |
| 37 | { |
| 38 | } |
| 39 | |
| 40 | /* *val is the current value of lockcnt->count. |
| 41 | * |
| 42 | * If the lock is free, try a cmpxchg from *val to new_if_free; return |
| 43 | * true and set *val to the old value found by the cmpxchg in |
| 44 | * lockcnt->count. |
| 45 | * |
| 46 | * If the lock is taken, wait for it to be released and return false |
| 47 | * *without trying again to take the lock*. Again, set *val to the |
| 48 | * new value of lockcnt->count. |
| 49 | * |
| 50 | * If *waited is true on return, new_if_free's bottom two bits must not |
| 51 | * be QEMU_LOCKCNT_STATE_LOCKED on subsequent calls, because the caller |
| 52 | * does not know if there are other waiters. Furthermore, after *waited |
| 53 | * is set the caller has effectively acquired the lock. If it returns |
| 54 | * with the lock not taken, it must wake another futex waiter. |
| 55 | */ |
| 56 | static bool qemu_lockcnt_cmpxchg_or_wait(QemuLockCnt *lockcnt, int *val, |
| 57 | int new_if_free, bool *waited) |
| 58 | { |
| 59 | /* Fast path for when the lock is free. */ |
| 60 | if ((*val & QEMU_LOCKCNT_STATE_MASK) == QEMU_LOCKCNT_STATE_FREE) { |
| 61 | int expected = *val; |
| 62 | |
| 63 | trace_lockcnt_fast_path_attempt(lockcnt, expected, new_if_free); |
| 64 | *val = atomic_cmpxchg(&lockcnt->count, expected, new_if_free); |
| 65 | if (*val == expected) { |
| 66 | trace_lockcnt_fast_path_success(lockcnt, expected, new_if_free); |
| 67 | *val = new_if_free; |
| 68 | return true; |
| 69 | } |
| 70 | } |
| 71 | |
| 72 | /* The slow path moves from locked to waiting if necessary, then |
| 73 | * does a futex wait. Both steps can be repeated ad nauseam, |
| 74 | * only getting out of the loop if we can have another shot at the |
| 75 | * fast path. Once we can, get out to compute the new destination |
| 76 | * value for the fast path. |
| 77 | */ |
| 78 | while ((*val & QEMU_LOCKCNT_STATE_MASK) != QEMU_LOCKCNT_STATE_FREE) { |
| 79 | if ((*val & QEMU_LOCKCNT_STATE_MASK) == QEMU_LOCKCNT_STATE_LOCKED) { |
| 80 | int expected = *val; |
| 81 | int new = expected - QEMU_LOCKCNT_STATE_LOCKED + QEMU_LOCKCNT_STATE_WAITING; |
| 82 | |
| 83 | trace_lockcnt_futex_wait_prepare(lockcnt, expected, new); |
| 84 | *val = atomic_cmpxchg(&lockcnt->count, expected, new); |
| 85 | if (*val == expected) { |
| 86 | *val = new; |
| 87 | } |
| 88 | continue; |
| 89 | } |
| 90 | |
| 91 | if ((*val & QEMU_LOCKCNT_STATE_MASK) == QEMU_LOCKCNT_STATE_WAITING) { |
| 92 | *waited = true; |
| 93 | trace_lockcnt_futex_wait(lockcnt, *val); |
| 94 | qemu_futex_wait(&lockcnt->count, *val); |
| 95 | *val = atomic_read(&lockcnt->count); |
| 96 | trace_lockcnt_futex_wait_resume(lockcnt, *val); |
| 97 | continue; |
| 98 | } |
| 99 | |
| 100 | abort(); |
| 101 | } |
| 102 | return false; |
| 103 | } |
| 104 | |
| 105 | static void lockcnt_wake(QemuLockCnt *lockcnt) |
| 106 | { |
| 107 | trace_lockcnt_futex_wake(lockcnt); |
| 108 | qemu_futex_wake(&lockcnt->count, 1); |
| 109 | } |
| 110 | |
| 111 | void qemu_lockcnt_inc(QemuLockCnt *lockcnt) |
| 112 | { |
| 113 | int val = atomic_read(&lockcnt->count); |
| 114 | bool waited = false; |
| 115 | |
| 116 | for (;;) { |
| 117 | if (val >= QEMU_LOCKCNT_COUNT_STEP) { |
| 118 | int expected = val; |
| 119 | val = atomic_cmpxchg(&lockcnt->count, val, val + QEMU_LOCKCNT_COUNT_STEP); |
| 120 | if (val == expected) { |
| 121 | break; |
| 122 | } |
| 123 | } else { |
| 124 | /* The fast path is (0, unlocked)->(1, unlocked). */ |
| 125 | if (qemu_lockcnt_cmpxchg_or_wait(lockcnt, &val, QEMU_LOCKCNT_COUNT_STEP, |
| 126 | &waited)) { |
| 127 | break; |
| 128 | } |
| 129 | } |
| 130 | } |
| 131 | |
| 132 | /* If we were woken by another thread, we should also wake one because |
| 133 | * we are effectively releasing the lock that was given to us. This is |
| 134 | * the case where qemu_lockcnt_lock would leave QEMU_LOCKCNT_STATE_WAITING |
| 135 | * in the low bits, and qemu_lockcnt_inc_and_unlock would find it and |
| 136 | * wake someone. |
| 137 | */ |
| 138 | if (waited) { |
| 139 | lockcnt_wake(lockcnt); |
| 140 | } |
| 141 | } |
| 142 | |
| 143 | void qemu_lockcnt_dec(QemuLockCnt *lockcnt) |
| 144 | { |
| 145 | atomic_sub(&lockcnt->count, QEMU_LOCKCNT_COUNT_STEP); |
| 146 | } |
| 147 | |
| 148 | /* Decrement a counter, and return locked if it is decremented to zero. |
| 149 | * If the function returns true, it is impossible for the counter to |
| 150 | * become nonzero until the next qemu_lockcnt_unlock. |
| 151 | */ |
| 152 | bool qemu_lockcnt_dec_and_lock(QemuLockCnt *lockcnt) |
| 153 | { |
| 154 | int val = atomic_read(&lockcnt->count); |
| 155 | int locked_state = QEMU_LOCKCNT_STATE_LOCKED; |
| 156 | bool waited = false; |
| 157 | |
| 158 | for (;;) { |
| 159 | if (val >= 2 * QEMU_LOCKCNT_COUNT_STEP) { |
| 160 | int expected = val; |
| 161 | val = atomic_cmpxchg(&lockcnt->count, val, val - QEMU_LOCKCNT_COUNT_STEP); |
| 162 | if (val == expected) { |
| 163 | break; |
| 164 | } |
| 165 | } else { |
| 166 | /* If count is going 1->0, take the lock. The fast path is |
| 167 | * (1, unlocked)->(0, locked) or (1, unlocked)->(0, waiting). |
| 168 | */ |
| 169 | if (qemu_lockcnt_cmpxchg_or_wait(lockcnt, &val, locked_state, &waited)) { |
| 170 | return true; |
| 171 | } |
| 172 | |
| 173 | if (waited) { |
| 174 | /* At this point we do not know if there are more waiters. Assume |
| 175 | * there are. |
| 176 | */ |
| 177 | locked_state = QEMU_LOCKCNT_STATE_WAITING; |
| 178 | } |
| 179 | } |
| 180 | } |
| 181 | |
| 182 | /* If we were woken by another thread, but we're returning in unlocked |
| 183 | * state, we should also wake a thread because we are effectively |
| 184 | * releasing the lock that was given to us. This is the case where |
| 185 | * qemu_lockcnt_lock would leave QEMU_LOCKCNT_STATE_WAITING in the low |
| 186 | * bits, and qemu_lockcnt_unlock would find it and wake someone. |
| 187 | */ |
| 188 | if (waited) { |
| 189 | lockcnt_wake(lockcnt); |
| 190 | } |
| 191 | return false; |
| 192 | } |
| 193 | |
| 194 | /* If the counter is one, decrement it and return locked. Otherwise do |
| 195 | * nothing. |
| 196 | * |
| 197 | * If the function returns true, it is impossible for the counter to |
| 198 | * become nonzero until the next qemu_lockcnt_unlock. |
| 199 | */ |
| 200 | bool qemu_lockcnt_dec_if_lock(QemuLockCnt *lockcnt) |
| 201 | { |
| 202 | int val = atomic_read(&lockcnt->count); |
| 203 | int locked_state = QEMU_LOCKCNT_STATE_LOCKED; |
| 204 | bool waited = false; |
| 205 | |
| 206 | while (val < 2 * QEMU_LOCKCNT_COUNT_STEP) { |
| 207 | /* If count is going 1->0, take the lock. The fast path is |
| 208 | * (1, unlocked)->(0, locked) or (1, unlocked)->(0, waiting). |
| 209 | */ |
| 210 | if (qemu_lockcnt_cmpxchg_or_wait(lockcnt, &val, locked_state, &waited)) { |
| 211 | return true; |
| 212 | } |
| 213 | |
| 214 | if (waited) { |
| 215 | /* At this point we do not know if there are more waiters. Assume |
| 216 | * there are. |
| 217 | */ |
| 218 | locked_state = QEMU_LOCKCNT_STATE_WAITING; |
| 219 | } |
| 220 | } |
| 221 | |
| 222 | /* If we were woken by another thread, but we're returning in unlocked |
| 223 | * state, we should also wake a thread because we are effectively |
| 224 | * releasing the lock that was given to us. This is the case where |
| 225 | * qemu_lockcnt_lock would leave QEMU_LOCKCNT_STATE_WAITING in the low |
| 226 | * bits, and qemu_lockcnt_inc_and_unlock would find it and wake someone. |
| 227 | */ |
| 228 | if (waited) { |
| 229 | lockcnt_wake(lockcnt); |
| 230 | } |
| 231 | return false; |
| 232 | } |
| 233 | |
| 234 | void qemu_lockcnt_lock(QemuLockCnt *lockcnt) |
| 235 | { |
| 236 | int val = atomic_read(&lockcnt->count); |
| 237 | int step = QEMU_LOCKCNT_STATE_LOCKED; |
| 238 | bool waited = false; |
| 239 | |
| 240 | /* The third argument is only used if the low bits of val are 0 |
| 241 | * (QEMU_LOCKCNT_STATE_FREE), so just blindly mix in the desired |
| 242 | * state. |
| 243 | */ |
| 244 | while (!qemu_lockcnt_cmpxchg_or_wait(lockcnt, &val, val + step, &waited)) { |
| 245 | if (waited) { |
| 246 | /* At this point we do not know if there are more waiters. Assume |
| 247 | * there are. |
| 248 | */ |
| 249 | step = QEMU_LOCKCNT_STATE_WAITING; |
| 250 | } |
| 251 | } |
| 252 | } |
| 253 | |
| 254 | void qemu_lockcnt_inc_and_unlock(QemuLockCnt *lockcnt) |
| 255 | { |
| 256 | int expected, new, val; |
| 257 | |
| 258 | val = atomic_read(&lockcnt->count); |
| 259 | do { |
| 260 | expected = val; |
| 261 | new = (val + QEMU_LOCKCNT_COUNT_STEP) & ~QEMU_LOCKCNT_STATE_MASK; |
| 262 | trace_lockcnt_unlock_attempt(lockcnt, val, new); |
| 263 | val = atomic_cmpxchg(&lockcnt->count, val, new); |
| 264 | } while (val != expected); |
| 265 | |
| 266 | trace_lockcnt_unlock_success(lockcnt, val, new); |
| 267 | if (val & QEMU_LOCKCNT_STATE_WAITING) { |
| 268 | lockcnt_wake(lockcnt); |
| 269 | } |
| 270 | } |
| 271 | |
| 272 | void qemu_lockcnt_unlock(QemuLockCnt *lockcnt) |
| 273 | { |
| 274 | int expected, new, val; |
| 275 | |
| 276 | val = atomic_read(&lockcnt->count); |
| 277 | do { |
| 278 | expected = val; |
| 279 | new = val & ~QEMU_LOCKCNT_STATE_MASK; |
| 280 | trace_lockcnt_unlock_attempt(lockcnt, val, new); |
| 281 | val = atomic_cmpxchg(&lockcnt->count, val, new); |
| 282 | } while (val != expected); |
| 283 | |
| 284 | trace_lockcnt_unlock_success(lockcnt, val, new); |
| 285 | if (val & QEMU_LOCKCNT_STATE_WAITING) { |
| 286 | lockcnt_wake(lockcnt); |
| 287 | } |
| 288 | } |
| 289 | |
| 290 | unsigned qemu_lockcnt_count(QemuLockCnt *lockcnt) |
| 291 | { |
| 292 | return atomic_read(&lockcnt->count) >> QEMU_LOCKCNT_COUNT_SHIFT; |
| 293 | } |
| 294 | #else |
Paolo Bonzini | 51dee5e | 2017-01-12 19:07:52 +0100 | [diff] [blame] | 295 | void qemu_lockcnt_init(QemuLockCnt *lockcnt) |
| 296 | { |
| 297 | qemu_mutex_init(&lockcnt->mutex); |
| 298 | lockcnt->count = 0; |
| 299 | } |
| 300 | |
| 301 | void qemu_lockcnt_destroy(QemuLockCnt *lockcnt) |
| 302 | { |
| 303 | qemu_mutex_destroy(&lockcnt->mutex); |
| 304 | } |
| 305 | |
| 306 | void qemu_lockcnt_inc(QemuLockCnt *lockcnt) |
| 307 | { |
| 308 | int old; |
| 309 | for (;;) { |
| 310 | old = atomic_read(&lockcnt->count); |
| 311 | if (old == 0) { |
| 312 | qemu_lockcnt_lock(lockcnt); |
| 313 | qemu_lockcnt_inc_and_unlock(lockcnt); |
| 314 | return; |
| 315 | } else { |
| 316 | if (atomic_cmpxchg(&lockcnt->count, old, old + 1) == old) { |
| 317 | return; |
| 318 | } |
| 319 | } |
| 320 | } |
| 321 | } |
| 322 | |
| 323 | void qemu_lockcnt_dec(QemuLockCnt *lockcnt) |
| 324 | { |
| 325 | atomic_dec(&lockcnt->count); |
| 326 | } |
| 327 | |
| 328 | /* Decrement a counter, and return locked if it is decremented to zero. |
| 329 | * It is impossible for the counter to become nonzero while the mutex |
| 330 | * is taken. |
| 331 | */ |
| 332 | bool qemu_lockcnt_dec_and_lock(QemuLockCnt *lockcnt) |
| 333 | { |
| 334 | int val = atomic_read(&lockcnt->count); |
| 335 | while (val > 1) { |
| 336 | int old = atomic_cmpxchg(&lockcnt->count, val, val - 1); |
| 337 | if (old != val) { |
| 338 | val = old; |
| 339 | continue; |
| 340 | } |
| 341 | |
| 342 | return false; |
| 343 | } |
| 344 | |
| 345 | qemu_lockcnt_lock(lockcnt); |
| 346 | if (atomic_fetch_dec(&lockcnt->count) == 1) { |
| 347 | return true; |
| 348 | } |
| 349 | |
| 350 | qemu_lockcnt_unlock(lockcnt); |
| 351 | return false; |
| 352 | } |
| 353 | |
| 354 | /* Decrement a counter and return locked if it is decremented to zero. |
| 355 | * Otherwise do nothing. |
| 356 | * |
| 357 | * It is impossible for the counter to become nonzero while the mutex |
| 358 | * is taken. |
| 359 | */ |
| 360 | bool qemu_lockcnt_dec_if_lock(QemuLockCnt *lockcnt) |
| 361 | { |
| 362 | /* No need for acquire semantics if we return false. */ |
| 363 | int val = atomic_read(&lockcnt->count); |
| 364 | if (val > 1) { |
| 365 | return false; |
| 366 | } |
| 367 | |
| 368 | qemu_lockcnt_lock(lockcnt); |
| 369 | if (atomic_fetch_dec(&lockcnt->count) == 1) { |
| 370 | return true; |
| 371 | } |
| 372 | |
| 373 | qemu_lockcnt_inc_and_unlock(lockcnt); |
| 374 | return false; |
| 375 | } |
| 376 | |
| 377 | void qemu_lockcnt_lock(QemuLockCnt *lockcnt) |
| 378 | { |
| 379 | qemu_mutex_lock(&lockcnt->mutex); |
| 380 | } |
| 381 | |
| 382 | void qemu_lockcnt_inc_and_unlock(QemuLockCnt *lockcnt) |
| 383 | { |
| 384 | atomic_inc(&lockcnt->count); |
| 385 | qemu_mutex_unlock(&lockcnt->mutex); |
| 386 | } |
| 387 | |
| 388 | void qemu_lockcnt_unlock(QemuLockCnt *lockcnt) |
| 389 | { |
| 390 | qemu_mutex_unlock(&lockcnt->mutex); |
| 391 | } |
| 392 | |
| 393 | unsigned qemu_lockcnt_count(QemuLockCnt *lockcnt) |
| 394 | { |
| 395 | return atomic_read(&lockcnt->count); |
| 396 | } |
Paolo Bonzini | fbcc3e5 | 2017-01-12 19:07:54 +0100 | [diff] [blame] | 397 | #endif |