| 1 | #ifndef JEMALLOC_INTERNAL_MUTEX_H |
| 2 | #define JEMALLOC_INTERNAL_MUTEX_H |
| 3 | |
| 4 | #include "jemalloc/internal/atomic.h" |
| 5 | #include "jemalloc/internal/mutex_prof.h" |
| 6 | #include "jemalloc/internal/tsd.h" |
| 7 | #include "jemalloc/internal/witness.h" |
| 8 | |
| 9 | typedef enum { |
| 10 | /* Can only acquire one mutex of a given witness rank at a time. */ |
| 11 | malloc_mutex_rank_exclusive, |
| 12 | /* |
| 13 | * Can acquire multiple mutexes of the same witness rank, but in |
| 14 | * address-ascending order only. |
| 15 | */ |
| 16 | malloc_mutex_address_ordered |
| 17 | } malloc_mutex_lock_order_t; |
| 18 | |
| 19 | typedef struct malloc_mutex_s malloc_mutex_t; |
| 20 | struct malloc_mutex_s { |
| 21 | union { |
| 22 | struct { |
| 23 | /* |
| 24 | * prof_data is defined first to reduce cacheline |
| 25 | * bouncing: the data is not touched by the mutex holder |
| 26 | * during unlocking, while might be modified by |
| 27 | * contenders. Having it before the mutex itself could |
| 28 | * avoid prefetching a modified cacheline (for the |
| 29 | * unlocking thread). |
| 30 | */ |
| 31 | mutex_prof_data_t prof_data; |
| 32 | #ifdef _WIN32 |
| 33 | # if _WIN32_WINNT >= 0x0600 |
| 34 | SRWLOCK lock; |
| 35 | # else |
| 36 | CRITICAL_SECTION lock; |
| 37 | # endif |
| 38 | #elif (defined(JEMALLOC_OS_UNFAIR_LOCK)) |
| 39 | os_unfair_lock lock; |
| 40 | #elif (defined(JEMALLOC_OSSPIN)) |
| 41 | OSSpinLock lock; |
| 42 | #elif (defined(JEMALLOC_MUTEX_INIT_CB)) |
| 43 | pthread_mutex_t lock; |
| 44 | malloc_mutex_t *postponed_next; |
| 45 | #else |
| 46 | pthread_mutex_t lock; |
| 47 | #endif |
| 48 | }; |
| 49 | /* |
| 50 | * We only touch witness when configured w/ debug. However we |
| 51 | * keep the field in a union when !debug so that we don't have |
| 52 | * to pollute the code base with #ifdefs, while avoid paying the |
| 53 | * memory cost. |
| 54 | */ |
| 55 | #if !defined(JEMALLOC_DEBUG) |
| 56 | witness_t witness; |
| 57 | malloc_mutex_lock_order_t lock_order; |
| 58 | #endif |
| 59 | }; |
| 60 | |
| 61 | #if defined(JEMALLOC_DEBUG) |
| 62 | witness_t witness; |
| 63 | malloc_mutex_lock_order_t lock_order; |
| 64 | #endif |
| 65 | }; |
| 66 | |
| 67 | /* |
| 68 | * Based on benchmark results, a fixed spin with this amount of retries works |
| 69 | * well for our critical sections. |
| 70 | */ |
| 71 | #define MALLOC_MUTEX_MAX_SPIN 250 |
| 72 | |
| 73 | #ifdef _WIN32 |
| 74 | # if _WIN32_WINNT >= 0x0600 |
| 75 | # define MALLOC_MUTEX_LOCK(m) AcquireSRWLockExclusive(&(m)->lock) |
| 76 | # define MALLOC_MUTEX_UNLOCK(m) ReleaseSRWLockExclusive(&(m)->lock) |
| 77 | # define MALLOC_MUTEX_TRYLOCK(m) (!TryAcquireSRWLockExclusive(&(m)->lock)) |
| 78 | # else |
| 79 | # define MALLOC_MUTEX_LOCK(m) EnterCriticalSection(&(m)->lock) |
| 80 | # define MALLOC_MUTEX_UNLOCK(m) LeaveCriticalSection(&(m)->lock) |
| 81 | # define MALLOC_MUTEX_TRYLOCK(m) (!TryEnterCriticalSection(&(m)->lock)) |
| 82 | # endif |
| 83 | #elif (defined(JEMALLOC_OS_UNFAIR_LOCK)) |
| 84 | # define MALLOC_MUTEX_LOCK(m) os_unfair_lock_lock(&(m)->lock) |
| 85 | # define MALLOC_MUTEX_UNLOCK(m) os_unfair_lock_unlock(&(m)->lock) |
| 86 | # define MALLOC_MUTEX_TRYLOCK(m) (!os_unfair_lock_trylock(&(m)->lock)) |
| 87 | #elif (defined(JEMALLOC_OSSPIN)) |
| 88 | # define MALLOC_MUTEX_LOCK(m) OSSpinLockLock(&(m)->lock) |
| 89 | # define MALLOC_MUTEX_UNLOCK(m) OSSpinLockUnlock(&(m)->lock) |
| 90 | # define MALLOC_MUTEX_TRYLOCK(m) (!OSSpinLockTry(&(m)->lock)) |
| 91 | #else |
| 92 | # define MALLOC_MUTEX_LOCK(m) pthread_mutex_lock(&(m)->lock) |
| 93 | # define MALLOC_MUTEX_UNLOCK(m) pthread_mutex_unlock(&(m)->lock) |
| 94 | # define MALLOC_MUTEX_TRYLOCK(m) (pthread_mutex_trylock(&(m)->lock) != 0) |
| 95 | #endif |
| 96 | |
| 97 | #define LOCK_PROF_DATA_INITIALIZER \ |
| 98 | {NSTIME_ZERO_INITIALIZER, NSTIME_ZERO_INITIALIZER, 0, 0, 0, \ |
| 99 | ATOMIC_INIT(0), 0, NULL, 0} |
| 100 | |
| 101 | #ifdef _WIN32 |
| 102 | # define MALLOC_MUTEX_INITIALIZER |
| 103 | #elif (defined(JEMALLOC_OS_UNFAIR_LOCK)) |
| 104 | # if defined(JEMALLOC_DEBUG) |
| 105 | # define MALLOC_MUTEX_INITIALIZER \ |
| 106 | {{{LOCK_PROF_DATA_INITIALIZER, OS_UNFAIR_LOCK_INIT}}, \ |
| 107 | WITNESS_INITIALIZER("mutex", WITNESS_RANK_OMIT), 0} |
| 108 | # else |
| 109 | # define MALLOC_MUTEX_INITIALIZER \ |
| 110 | {{{LOCK_PROF_DATA_INITIALIZER, OS_UNFAIR_LOCK_INIT}}, \ |
| 111 | WITNESS_INITIALIZER("mutex", WITNESS_RANK_OMIT)} |
| 112 | # endif |
| 113 | #elif (defined(JEMALLOC_OSSPIN)) |
| 114 | # define MALLOC_MUTEX_INITIALIZER \ |
| 115 | {{{LOCK_PROF_DATA_INITIALIZER, 0}}, \ |
| 116 | WITNESS_INITIALIZER("mutex", WITNESS_RANK_OMIT)} |
| 117 | #elif (defined(JEMALLOC_MUTEX_INIT_CB)) |
| 118 | # if (defined(JEMALLOC_DEBUG)) |
| 119 | # define MALLOC_MUTEX_INITIALIZER \ |
| 120 | {{{LOCK_PROF_DATA_INITIALIZER, PTHREAD_MUTEX_INITIALIZER, NULL}}, \ |
| 121 | WITNESS_INITIALIZER("mutex", WITNESS_RANK_OMIT), 0} |
| 122 | # else |
| 123 | # define MALLOC_MUTEX_INITIALIZER \ |
| 124 | {{{LOCK_PROF_DATA_INITIALIZER, PTHREAD_MUTEX_INITIALIZER, NULL}}, \ |
| 125 | WITNESS_INITIALIZER("mutex", WITNESS_RANK_OMIT)} |
| 126 | # endif |
| 127 | |
| 128 | #else |
| 129 | # define MALLOC_MUTEX_TYPE PTHREAD_MUTEX_DEFAULT |
| 130 | # if defined(JEMALLOC_DEBUG) |
| 131 | # define MALLOC_MUTEX_INITIALIZER \ |
| 132 | {{{LOCK_PROF_DATA_INITIALIZER, PTHREAD_MUTEX_INITIALIZER}}, \ |
| 133 | WITNESS_INITIALIZER("mutex", WITNESS_RANK_OMIT), 0} |
| 134 | # else |
| 135 | # define MALLOC_MUTEX_INITIALIZER \ |
| 136 | {{{LOCK_PROF_DATA_INITIALIZER, PTHREAD_MUTEX_INITIALIZER}}, \ |
| 137 | WITNESS_INITIALIZER("mutex", WITNESS_RANK_OMIT)} |
| 138 | # endif |
| 139 | #endif |
| 140 | |
| 141 | #ifdef JEMALLOC_LAZY_LOCK |
| 142 | extern bool isthreaded; |
| 143 | #else |
| 144 | # undef isthreaded /* Undo private_namespace.h definition. */ |
| 145 | # define isthreaded true |
| 146 | #endif |
| 147 | |
| 148 | bool malloc_mutex_init(malloc_mutex_t *mutex, const char *name, |
| 149 | witness_rank_t rank, malloc_mutex_lock_order_t lock_order); |
| 150 | void malloc_mutex_prefork(tsdn_t *tsdn, malloc_mutex_t *mutex); |
| 151 | void malloc_mutex_postfork_parent(tsdn_t *tsdn, malloc_mutex_t *mutex); |
| 152 | void malloc_mutex_postfork_child(tsdn_t *tsdn, malloc_mutex_t *mutex); |
| 153 | bool malloc_mutex_boot(void); |
| 154 | void malloc_mutex_prof_data_reset(tsdn_t *tsdn, malloc_mutex_t *mutex); |
| 155 | |
| 156 | void malloc_mutex_lock_slow(malloc_mutex_t *mutex); |
| 157 | |
| 158 | static inline void |
| 159 | malloc_mutex_lock_final(malloc_mutex_t *mutex) { |
| 160 | MALLOC_MUTEX_LOCK(mutex); |
| 161 | } |
| 162 | |
| 163 | static inline bool |
| 164 | malloc_mutex_trylock_final(malloc_mutex_t *mutex) { |
| 165 | return MALLOC_MUTEX_TRYLOCK(mutex); |
| 166 | } |
| 167 | |
| 168 | static inline void |
| 169 | mutex_owner_stats_update(tsdn_t *tsdn, malloc_mutex_t *mutex) { |
| 170 | if (config_stats) { |
| 171 | mutex_prof_data_t *data = &mutex->prof_data; |
| 172 | data->n_lock_ops++; |
| 173 | if (data->prev_owner != tsdn) { |
| 174 | data->prev_owner = tsdn; |
| 175 | data->n_owner_switches++; |
| 176 | } |
| 177 | } |
| 178 | } |
| 179 | |
| 180 | /* Trylock: return false if the lock is successfully acquired. */ |
| 181 | static inline bool |
| 182 | malloc_mutex_trylock(tsdn_t *tsdn, malloc_mutex_t *mutex) { |
| 183 | witness_assert_not_owner(tsdn_witness_tsdp_get(tsdn), &mutex->witness); |
| 184 | if (isthreaded) { |
| 185 | if (malloc_mutex_trylock_final(mutex)) { |
| 186 | return true; |
| 187 | } |
| 188 | mutex_owner_stats_update(tsdn, mutex); |
| 189 | } |
| 190 | witness_lock(tsdn_witness_tsdp_get(tsdn), &mutex->witness); |
| 191 | |
| 192 | return false; |
| 193 | } |
| 194 | |
| 195 | /* Aggregate lock prof data. */ |
| 196 | static inline void |
| 197 | malloc_mutex_prof_merge(mutex_prof_data_t *sum, mutex_prof_data_t *data) { |
| 198 | nstime_add(&sum->tot_wait_time, &data->tot_wait_time); |
| 199 | if (nstime_compare(&sum->max_wait_time, &data->max_wait_time) < 0) { |
| 200 | nstime_copy(&sum->max_wait_time, &data->max_wait_time); |
| 201 | } |
| 202 | |
| 203 | sum->n_wait_times += data->n_wait_times; |
| 204 | sum->n_spin_acquired += data->n_spin_acquired; |
| 205 | |
| 206 | if (sum->max_n_thds < data->max_n_thds) { |
| 207 | sum->max_n_thds = data->max_n_thds; |
| 208 | } |
| 209 | uint32_t cur_n_waiting_thds = atomic_load_u32(&sum->n_waiting_thds, |
| 210 | ATOMIC_RELAXED); |
| 211 | uint32_t new_n_waiting_thds = cur_n_waiting_thds + atomic_load_u32( |
| 212 | &data->n_waiting_thds, ATOMIC_RELAXED); |
| 213 | atomic_store_u32(&sum->n_waiting_thds, new_n_waiting_thds, |
| 214 | ATOMIC_RELAXED); |
| 215 | sum->n_owner_switches += data->n_owner_switches; |
| 216 | sum->n_lock_ops += data->n_lock_ops; |
| 217 | } |
| 218 | |
| 219 | static inline void |
| 220 | malloc_mutex_lock(tsdn_t *tsdn, malloc_mutex_t *mutex) { |
| 221 | witness_assert_not_owner(tsdn_witness_tsdp_get(tsdn), &mutex->witness); |
| 222 | if (isthreaded) { |
| 223 | if (malloc_mutex_trylock_final(mutex)) { |
| 224 | malloc_mutex_lock_slow(mutex); |
| 225 | } |
| 226 | mutex_owner_stats_update(tsdn, mutex); |
| 227 | } |
| 228 | witness_lock(tsdn_witness_tsdp_get(tsdn), &mutex->witness); |
| 229 | } |
| 230 | |
| 231 | static inline void |
| 232 | malloc_mutex_unlock(tsdn_t *tsdn, malloc_mutex_t *mutex) { |
| 233 | witness_unlock(tsdn_witness_tsdp_get(tsdn), &mutex->witness); |
| 234 | if (isthreaded) { |
| 235 | MALLOC_MUTEX_UNLOCK(mutex); |
| 236 | } |
| 237 | } |
| 238 | |
| 239 | static inline void |
| 240 | malloc_mutex_assert_owner(tsdn_t *tsdn, malloc_mutex_t *mutex) { |
| 241 | witness_assert_owner(tsdn_witness_tsdp_get(tsdn), &mutex->witness); |
| 242 | } |
| 243 | |
| 244 | static inline void |
| 245 | malloc_mutex_assert_not_owner(tsdn_t *tsdn, malloc_mutex_t *mutex) { |
| 246 | witness_assert_not_owner(tsdn_witness_tsdp_get(tsdn), &mutex->witness); |
| 247 | } |
| 248 | |
| 249 | /* Copy the prof data from mutex for processing. */ |
| 250 | static inline void |
| 251 | malloc_mutex_prof_read(tsdn_t *tsdn, mutex_prof_data_t *data, |
| 252 | malloc_mutex_t *mutex) { |
| 253 | mutex_prof_data_t *source = &mutex->prof_data; |
| 254 | /* Can only read holding the mutex. */ |
| 255 | malloc_mutex_assert_owner(tsdn, mutex); |
| 256 | |
| 257 | /* |
| 258 | * Not *really* allowed (we shouldn't be doing non-atomic loads of |
| 259 | * atomic data), but the mutex protection makes this safe, and writing |
| 260 | * a member-for-member copy is tedious for this situation. |
| 261 | */ |
| 262 | *data = *source; |
| 263 | /* n_wait_thds is not reported (modified w/o locking). */ |
| 264 | atomic_store_u32(&data->n_waiting_thds, 0, ATOMIC_RELAXED); |
| 265 | } |
| 266 | |
| 267 | #endif /* JEMALLOC_INTERNAL_MUTEX_H */ |
| 268 | |