| OLD | NEW | 
|---|
| 1 #include "pthread_impl.h" | 1 #include "pthread_impl.h" | 
| 2 | 2 | 
| 3 static int pshared_barrier_wait(pthread_barrier_t *b) | 3 static int pshared_barrier_wait(pthread_barrier_t* b) { | 
| 4 { | 4   int limit = (b->_b_limit & INT_MAX) + 1; | 
| 5 »       int limit = (b->_b_limit & INT_MAX) + 1; | 5   int ret = 0; | 
| 6 »       int ret = 0; | 6   int v, w; | 
| 7 »       int v, w; |  | 
| 8 | 7 | 
| 9 »       if (limit==1) return PTHREAD_BARRIER_SERIAL_THREAD; | 8   if (limit == 1) | 
|  | 9     return PTHREAD_BARRIER_SERIAL_THREAD; | 
| 10 | 10 | 
| 11 »       while ((v=a_cas(&b->_b_lock, 0, limit))) | 11   while ((v = a_cas(&b->_b_lock, 0, limit))) | 
| 12 »       »       __wait(&b->_b_lock, &b->_b_waiters, v, 0); | 12     __wait(&b->_b_lock, &b->_b_waiters, v, 0); | 
| 13 | 13 | 
| 14 »       /* Wait for <limit> threads to get to the barrier */ | 14   /* Wait for <limit> threads to get to the barrier */ | 
| 15 »       if (++b->_b_count == limit) { | 15   if (++b->_b_count == limit) { | 
| 16 »       »       a_store(&b->_b_count, 0); | 16     a_store(&b->_b_count, 0); | 
| 17 »       »       ret = PTHREAD_BARRIER_SERIAL_THREAD; | 17     ret = PTHREAD_BARRIER_SERIAL_THREAD; | 
| 18 »       »       if (b->_b_waiters2) __wake(&b->_b_count, -1, 0); | 18     if (b->_b_waiters2) | 
| 19 »       } else { | 19       __wake(&b->_b_count, -1, 0); | 
| 20 »       »       a_store(&b->_b_lock, 0); | 20   } else { | 
| 21 »       »       if (b->_b_waiters) __wake(&b->_b_lock, 1, 0); | 21     a_store(&b->_b_lock, 0); | 
| 22 »       »       while ((v=b->_b_count)>0) | 22     if (b->_b_waiters) | 
| 23 »       »       »       __wait(&b->_b_count, &b->_b_waiters2, v, 0); | 23       __wake(&b->_b_lock, 1, 0); | 
| 24 »       } | 24     while ((v = b->_b_count) > 0) | 
|  | 25       __wait(&b->_b_count, &b->_b_waiters2, v, 0); | 
|  | 26   } | 
| 25 | 27 | 
| 26 »       __vm_lock(); | 28   __vm_lock(); | 
| 27 | 29 | 
| 28 »       /* Ensure all threads have a vm lock before proceeding */ | 30   /* Ensure all threads have a vm lock before proceeding */ | 
| 29 »       if (a_fetch_add(&b->_b_count, -1)==1-limit) { | 31   if (a_fetch_add(&b->_b_count, -1) == 1 - limit) { | 
| 30 »       »       a_store(&b->_b_count, 0); | 32     a_store(&b->_b_count, 0); | 
| 31 »       »       if (b->_b_waiters2) __wake(&b->_b_count, -1, 0); | 33     if (b->_b_waiters2) | 
| 32 »       } else { | 34       __wake(&b->_b_count, -1, 0); | 
| 33 »       »       while ((v=b->_b_count)) | 35   } else { | 
| 34 »       »       »       __wait(&b->_b_count, &b->_b_waiters2, v, 0); | 36     while ((v = b->_b_count)) | 
| 35 »       } | 37       __wait(&b->_b_count, &b->_b_waiters2, v, 0); | 
| 36 » | 38   } | 
| 37 »       /* Perform a recursive unlock suitable for self-sync'd destruction */ |  | 
| 38 »       do { |  | 
| 39 »       »       v = b->_b_lock; |  | 
| 40 »       »       w = b->_b_waiters; |  | 
| 41 »       } while (a_cas(&b->_b_lock, v, v==INT_MIN+1 ? 0 : v-1) != v); |  | 
| 42 | 39 | 
| 43 »       /* Wake a thread waiting to reuse or destroy the barrier */ | 40   /* Perform a recursive unlock suitable for self-sync'd destruction */ | 
| 44 »       if (v==INT_MIN+1 || (v==1 && w)) | 41   do { | 
| 45 »       »       __wake(&b->_b_lock, 1, 0); | 42     v = b->_b_lock; | 
|  | 43     w = b->_b_waiters; | 
|  | 44   } while (a_cas(&b->_b_lock, v, v == INT_MIN + 1 ? 0 : v - 1) != v); | 
| 46 | 45 | 
| 47 »       __vm_unlock(); | 46   /* Wake a thread waiting to reuse or destroy the barrier */ | 
|  | 47   if (v == INT_MIN + 1 || (v == 1 && w)) | 
|  | 48     __wake(&b->_b_lock, 1, 0); | 
| 48 | 49 | 
| 49 »       return ret; | 50   __vm_unlock(); | 
|  | 51 | 
|  | 52   return ret; | 
| 50 } | 53 } | 
| 51 | 54 | 
| 52 struct instance | 55 struct instance { | 
| 53 { | 56   volatile int count; | 
| 54 »       volatile int count; | 57   volatile int last; | 
| 55 »       volatile int last; | 58   volatile int waiters; | 
| 56 »       volatile int waiters; | 59   volatile int finished; | 
| 57 »       volatile int finished; |  | 
| 58 }; | 60 }; | 
| 59 | 61 | 
| 60 int pthread_barrier_wait(pthread_barrier_t *b) | 62 int pthread_barrier_wait(pthread_barrier_t* b) { | 
| 61 { | 63   int limit = b->_b_limit; | 
| 62 »       int limit = b->_b_limit; | 64   struct instance* inst; | 
| 63 »       struct instance *inst; |  | 
| 64 | 65 | 
| 65 »       /* Trivial case: count was set at 1 */ | 66   /* Trivial case: count was set at 1 */ | 
| 66 »       if (!limit) return PTHREAD_BARRIER_SERIAL_THREAD; | 67   if (!limit) | 
|  | 68     return PTHREAD_BARRIER_SERIAL_THREAD; | 
| 67 | 69 | 
| 68 »       /* Process-shared barriers require a separate, inefficient wait */ | 70   /* Process-shared barriers require a separate, inefficient wait */ | 
| 69 »       if (limit < 0) return pshared_barrier_wait(b); | 71   if (limit < 0) | 
|  | 72     return pshared_barrier_wait(b); | 
| 70 | 73 | 
| 71 »       /* Otherwise we need a lock on the barrier object */ | 74   /* Otherwise we need a lock on the barrier object */ | 
| 72 »       while (a_swap(&b->_b_lock, 1)) | 75   while (a_swap(&b->_b_lock, 1)) | 
| 73 »       »       __wait(&b->_b_lock, &b->_b_waiters, 1, 1); | 76     __wait(&b->_b_lock, &b->_b_waiters, 1, 1); | 
| 74 »       inst = b->_b_inst; | 77   inst = b->_b_inst; | 
| 75 | 78 | 
| 76 »       /* First thread to enter the barrier becomes the "instance owner" */ | 79   /* First thread to enter the barrier becomes the "instance owner" */ | 
| 77 »       if (!inst) { | 80   if (!inst) { | 
| 78 »       »       struct instance new_inst = { 0 }; | 81     struct instance new_inst = {0}; | 
| 79 »       »       int spins = 200; | 82     int spins = 200; | 
| 80 »       »       b->_b_inst = inst = &new_inst; | 83     b->_b_inst = inst = &new_inst; | 
| 81 »       »       a_store(&b->_b_lock, 0); | 84     a_store(&b->_b_lock, 0); | 
| 82 »       »       if (b->_b_waiters) __wake(&b->_b_lock, 1, 1); | 85     if (b->_b_waiters) | 
| 83 »       »       while (spins-- && !inst->finished) | 86       __wake(&b->_b_lock, 1, 1); | 
| 84 »       »       »       a_spin(); | 87     while (spins-- && !inst->finished) | 
| 85 »       »       a_inc(&inst->finished); | 88       a_spin(); | 
| 86 »       »       while (inst->finished == 1) | 89     a_inc(&inst->finished); | 
| 87 »       »       »       __syscall(SYS_futex,&inst->finished,FUTEX_WAIT|128,1,0) 
     != -ENOSYS | 90     while (inst->finished == 1) | 
| 88 »       »       »       || __syscall(SYS_futex,&inst->finished,FUTEX_WAIT,1,0); | 91       __syscall(SYS_futex, &inst->finished, FUTEX_WAIT | 128, 1, 0) != | 
| 89 »       »       return PTHREAD_BARRIER_SERIAL_THREAD; | 92               -ENOSYS || | 
| 90 »       } | 93           __syscall(SYS_futex, &inst->finished, FUTEX_WAIT, 1, 0); | 
|  | 94     return PTHREAD_BARRIER_SERIAL_THREAD; | 
|  | 95   } | 
| 91 | 96 | 
| 92 »       /* Last thread to enter the barrier wakes all non-instance-owners */ | 97   /* Last thread to enter the barrier wakes all non-instance-owners */ | 
| 93 »       if (++inst->count == limit) { | 98   if (++inst->count == limit) { | 
| 94 »       »       b->_b_inst = 0; | 99     b->_b_inst = 0; | 
| 95 »       »       a_store(&b->_b_lock, 0); | 100     a_store(&b->_b_lock, 0); | 
| 96 »       »       if (b->_b_waiters) __wake(&b->_b_lock, 1, 1); | 101     if (b->_b_waiters) | 
| 97 »       »       a_store(&inst->last, 1); | 102       __wake(&b->_b_lock, 1, 1); | 
| 98 »       »       if (inst->waiters) | 103     a_store(&inst->last, 1); | 
| 99 »       »       »       __wake(&inst->last, -1, 1); | 104     if (inst->waiters) | 
| 100 »       } else { | 105       __wake(&inst->last, -1, 1); | 
| 101 »       »       a_store(&b->_b_lock, 0); | 106   } else { | 
| 102 »       »       if (b->_b_waiters) __wake(&b->_b_lock, 1, 1); | 107     a_store(&b->_b_lock, 0); | 
| 103 »       »       __wait(&inst->last, &inst->waiters, 0, 1); | 108     if (b->_b_waiters) | 
| 104 »       } | 109       __wake(&b->_b_lock, 1, 1); | 
|  | 110     __wait(&inst->last, &inst->waiters, 0, 1); | 
|  | 111   } | 
| 105 | 112 | 
| 106 »       /* Last thread to exit the barrier wakes the instance owner */ | 113   /* Last thread to exit the barrier wakes the instance owner */ | 
| 107 »       if (a_fetch_add(&inst->count,-1)==1 && a_fetch_add(&inst->finished,1)) | 114   if (a_fetch_add(&inst->count, -1) == 1 && a_fetch_add(&inst->finished, 1)) | 
| 108 »       »       __wake(&inst->finished, 1, 1); | 115     __wake(&inst->finished, 1, 1); | 
| 109 | 116 | 
| 110 »       return 0; | 117   return 0; | 
| 111 } | 118 } | 
| OLD | NEW | 
|---|