1 #include <linux/atomic.h> 2 #include <linux/rwsem.h> 3 #include <linux/percpu.h> 4 #include <linux/wait.h> 5 #include <linux/lockdep.h> 6 #include <linux/percpu-rwsem.h> 7 #include <linux/rcupdate.h> 8 #include <linux/sched.h> 9 #include <linux/errno.h> 10 11 int __percpu_init_rwsem(struct percpu_rw_semaphore *brw, 12 const char *name, struct lock_class_key *rwsem_key) 13 { 14 brw->fast_read_ctr = alloc_percpu(int); 15 if (unlikely(!brw->fast_read_ctr)) 16 return -ENOMEM; 17 18 /* ->rw_sem represents the whole percpu_rw_semaphore for lockdep */ 19 __init_rwsem(&brw->rw_sem, name, rwsem_key); 20 rcu_sync_init(&brw->rss, RCU_SCHED_SYNC); 21 atomic_set(&brw->slow_read_ctr, 0); 22 init_waitqueue_head(&brw->write_waitq); 23 return 0; 24 } 25 EXPORT_SYMBOL_GPL(__percpu_init_rwsem); 26 27 void percpu_free_rwsem(struct percpu_rw_semaphore *brw) 28 { 29 /* 30 * XXX: temporary kludge. The error path in alloc_super() 31 * assumes that percpu_free_rwsem() is safe after kzalloc(). 32 */ 33 if (!brw->fast_read_ctr) 34 return; 35 36 rcu_sync_dtor(&brw->rss); 37 free_percpu(brw->fast_read_ctr); 38 brw->fast_read_ctr = NULL; /* catch use after free bugs */ 39 } 40 41 /* 42 * This is the fast-path for down_read/up_read. If it succeeds we rely 43 * on the barriers provided by rcu_sync_enter/exit; see the comments in 44 * percpu_down_write() and percpu_up_write(). 45 * 46 * If this helper fails the callers rely on the normal rw_semaphore and 47 * atomic_dec_and_test(), so in this case we have the necessary barriers. 48 */ 49 static bool update_fast_ctr(struct percpu_rw_semaphore *brw, unsigned int val) 50 { 51 bool success; 52 53 preempt_disable(); 54 success = rcu_sync_is_idle(&brw->rss); 55 if (likely(success)) 56 __this_cpu_add(*brw->fast_read_ctr, val); 57 preempt_enable(); 58 59 return success; 60 } 61 62 /* 63 * Like the normal down_read() this is not recursive, the writer can 64 * come after the first percpu_down_read() and create the deadlock. 65 * 66 * Note: returns with lock_is_held(brw->rw_sem) == T for lockdep, 67 * percpu_up_read() does rwsem_release(). This pairs with the usage 68 * of ->rw_sem in percpu_down/up_write(). 69 */ 70 void percpu_down_read(struct percpu_rw_semaphore *brw) 71 { 72 might_sleep(); 73 rwsem_acquire_read(&brw->rw_sem.dep_map, 0, 0, _RET_IP_); 74 75 if (likely(update_fast_ctr(brw, +1))) 76 return; 77 78 /* Avoid rwsem_acquire_read() and rwsem_release() */ 79 __down_read(&brw->rw_sem); 80 atomic_inc(&brw->slow_read_ctr); 81 __up_read(&brw->rw_sem); 82 } 83 EXPORT_SYMBOL_GPL(percpu_down_read); 84 85 int percpu_down_read_trylock(struct percpu_rw_semaphore *brw) 86 { 87 if (unlikely(!update_fast_ctr(brw, +1))) { 88 if (!__down_read_trylock(&brw->rw_sem)) 89 return 0; 90 atomic_inc(&brw->slow_read_ctr); 91 __up_read(&brw->rw_sem); 92 } 93 94 rwsem_acquire_read(&brw->rw_sem.dep_map, 0, 1, _RET_IP_); 95 return 1; 96 } 97 98 void percpu_up_read(struct percpu_rw_semaphore *brw) 99 { 100 rwsem_release(&brw->rw_sem.dep_map, 1, _RET_IP_); 101 102 if (likely(update_fast_ctr(brw, -1))) 103 return; 104 105 /* false-positive is possible but harmless */ 106 if (atomic_dec_and_test(&brw->slow_read_ctr)) 107 wake_up_all(&brw->write_waitq); 108 } 109 EXPORT_SYMBOL_GPL(percpu_up_read); 110 111 static int clear_fast_ctr(struct percpu_rw_semaphore *brw) 112 { 113 unsigned int sum = 0; 114 int cpu; 115 116 for_each_possible_cpu(cpu) { 117 sum += per_cpu(*brw->fast_read_ctr, cpu); 118 per_cpu(*brw->fast_read_ctr, cpu) = 0; 119 } 120 121 return sum; 122 } 123 124 void percpu_down_write(struct percpu_rw_semaphore *brw) 125 { 126 /* 127 * Make rcu_sync_is_idle() == F and thus disable the fast-path in 128 * percpu_down_read() and percpu_up_read(), and wait for gp pass. 129 * 130 * The latter synchronises us with the preceding readers which used 131 * the fast-past, so we can not miss the result of __this_cpu_add() 132 * or anything else inside their criticial sections. 133 */ 134 rcu_sync_enter(&brw->rss); 135 136 /* exclude other writers, and block the new readers completely */ 137 down_write(&brw->rw_sem); 138 139 /* nobody can use fast_read_ctr, move its sum into slow_read_ctr */ 140 atomic_add(clear_fast_ctr(brw), &brw->slow_read_ctr); 141 142 /* wait for all readers to complete their percpu_up_read() */ 143 wait_event(brw->write_waitq, !atomic_read(&brw->slow_read_ctr)); 144 } 145 EXPORT_SYMBOL_GPL(percpu_down_write); 146 147 void percpu_up_write(struct percpu_rw_semaphore *brw) 148 { 149 /* release the lock, but the readers can't use the fast-path */ 150 up_write(&brw->rw_sem); 151 /* 152 * Enable the fast-path in percpu_down_read() and percpu_up_read() 153 * but only after another gp pass; this adds the necessary barrier 154 * to ensure the reader can't miss the changes done by us. 155 */ 156 rcu_sync_exit(&brw->rss); 157 } 158 EXPORT_SYMBOL_GPL(percpu_up_write); 159