kernel/locking/percpu-rwsem.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 #include <linux/atomic.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/sched/task.h>
  10 #include <linux/errno.h>
  11
  12 int __percpu_init_rwsem(struct percpu_rw_semaphore *sem,
  13                         const char *name, struct lock_class_key *key)
  14 {
  15         sem->read_count = alloc_percpu(int);
  16         if (unlikely(!sem->read_count))
  17                 return -ENOMEM;
  18
  19         rcu_sync_init(&sem->rss);
  20         rcuwait_init(&sem->writer);
  21         init_waitqueue_head(&sem->waiters);
  22         atomic_set(&sem->block, 0);
  23 #ifdef CONFIG_DEBUG_LOCK_ALLOC
  24         debug_check_no_locks_freed((void *)sem, sizeof(*sem));
  25         lockdep_init_map(&sem->dep_map, name, key, 0);
  26 #endif
  27         return 0;
  28 }
  29 EXPORT_SYMBOL_GPL(__percpu_init_rwsem);
  30
  31 void percpu_free_rwsem(struct percpu_rw_semaphore *sem)
  32 {
  33         /*
  34          * XXX: temporary kludge. The error path in alloc_super()
  35          * assumes that percpu_free_rwsem() is safe after kzalloc().
  36          */
  37         if (!sem->read_count)
  38                 return;
  39
  40         rcu_sync_dtor(&sem->rss);
  41         free_percpu(sem->read_count);
  42         sem->read_count = NULL; /* catch use after free bugs */
  43 }
  44 EXPORT_SYMBOL_GPL(percpu_free_rwsem);
  45
  46 static bool __percpu_down_read_trylock(struct percpu_rw_semaphore *sem)
  47 {
  48         this_cpu_inc(*sem->read_count);
  49
  50         /*
  51          * Due to having preemption disabled the decrement happens on
  52          * the same CPU as the increment, avoiding the
  53          * increment-on-one-CPU-and-decrement-on-another problem.
  54          *
  55          * If the reader misses the writer's assignment of sem->block, then the
  56          * writer is guaranteed to see the reader's increment.
  57          *
  58          * Conversely, any readers that increment their sem->read_count after
  59          * the writer looks are guaranteed to see the sem->block value, which
  60          * in turn means that they are guaranteed to immediately decrement
  61          * their sem->read_count, so that it doesn't matter that the writer
  62          * missed them.
  63          */
  64
  65         smp_mb(); /* A matches D */
  66
  67         /*
  68          * If !sem->block the critical section starts here, matched by the
  69          * release in percpu_up_write().
  70          */
  71         if (likely(!atomic_read_acquire(&sem->block)))
  72                 return true;
  73
  74         this_cpu_dec(*sem->read_count);
  75
  76         /* Prod writer to re-evaluate readers_active_check() */
  77         rcuwait_wake_up(&sem->writer);
  78
  79         return false;
  80 }
  81
  82 static inline bool __percpu_down_write_trylock(struct percpu_rw_semaphore *sem)
  83 {
  84         if (atomic_read(&sem->block))
  85                 return false;
  86
  87         return atomic_xchg(&sem->block, 1) == 0;
  88 }
  89
  90 static bool __percpu_rwsem_trylock(struct percpu_rw_semaphore *sem, bool reader)
  91 {
  92         if (reader) {
  93                 bool ret;
  94
  95                 preempt_disable();
  96                 ret = __percpu_down_read_trylock(sem);
  97                 preempt_enable();
  98
  99                 return ret;
 100         }
 101         return __percpu_down_write_trylock(sem);
 102 }
 103
 104 /*
 105  * The return value of wait_queue_entry::func means:
 106  *
 107  *  <0 - error, wakeup is terminated and the error is returned
 108  *   0 - no wakeup, a next waiter is tried
 109  *  >0 - woken, if EXCLUSIVE, counted towards @nr_exclusive.
 110  *
 111  * We use EXCLUSIVE for both readers and writers to preserve FIFO order,
 112  * and play games with the return value to allow waking multiple readers.
 113  *
 114  * Specifically, we wake readers until we've woken a single writer, or until a
 115  * trylock fails.
 116  */
 117 static int percpu_rwsem_wake_function(struct wait_queue_entry *wq_entry,
 118                                       unsigned int mode, int wake_flags,
 119                                       void *key)
 120 {
 121         bool reader = wq_entry->flags & WQ_FLAG_CUSTOM;
 122         struct percpu_rw_semaphore *sem = key;
 123         struct task_struct *p;
 124
 125         /* concurrent against percpu_down_write(), can get stolen */
 126         if (!__percpu_rwsem_trylock(sem, reader))
 127                 return 1;
 128
 129         p = get_task_struct(wq_entry->private);
 130         list_del_init(&wq_entry->entry);
 131         smp_store_release(&wq_entry->private, NULL);
 132
 133         wake_up_process(p);
 134         put_task_struct(p);
 135
 136         return !reader; /* wake (readers until) 1 writer */
 137 }
 138
 139 static void percpu_rwsem_wait(struct percpu_rw_semaphore *sem, bool reader)
 140 {
 141         DEFINE_WAIT_FUNC(wq_entry, percpu_rwsem_wake_function);
 142         bool wait;
 143
 144         spin_lock_irq(&sem->waiters.lock);
 145         /*
 146          * Serialize against the wakeup in percpu_up_write(), if we fail
 147          * the trylock, the wakeup must see us on the list.
 148          */
 149         wait = !__percpu_rwsem_trylock(sem, reader);
 150         if (wait) {
 151                 wq_entry.flags |= WQ_FLAG_EXCLUSIVE | reader * WQ_FLAG_CUSTOM;
 152                 __add_wait_queue_entry_tail(&sem->waiters, &wq_entry);
 153         }
 154         spin_unlock_irq(&sem->waiters.lock);
 155
 156         while (wait) {
 157                 set_current_state(TASK_UNINTERRUPTIBLE);
 158                 if (!smp_load_acquire(&wq_entry.private))
 159                         break;
 160                 schedule();
 161         }
 162         __set_current_state(TASK_RUNNING);
 163 }
 164
 165 bool __percpu_down_read(struct percpu_rw_semaphore *sem, bool try)
 166 {
 167         if (__percpu_down_read_trylock(sem))
 168                 return true;
 169
 170         if (try)
 171                 return false;
 172
 173         preempt_enable();
 174         percpu_rwsem_wait(sem, /* .reader = */ true);
 175         preempt_disable();
 176
 177         return true;
 178 }
 179 EXPORT_SYMBOL_GPL(__percpu_down_read);
 180
 181 #define per_cpu_sum(var)                                                \
 182 ({                                                                      \
 183         typeof(var) __sum = 0;                                          \
 184         int cpu;                                                        \
 185         compiletime_assert_atomic_type(__sum);                          \
 186         for_each_possible_cpu(cpu)                                      \
 187                 __sum += per_cpu(var, cpu);                             \
 188         __sum;                                                          \
 189 })
 190
 191 /*
 192  * Return true if the modular sum of the sem->read_count per-CPU variable is
 193  * zero.  If this sum is zero, then it is stable due to the fact that if any
 194  * newly arriving readers increment a given counter, they will immediately
 195  * decrement that same counter.
 196  *
 197  * Assumes sem->block is set.
 198  */
 199 static bool readers_active_check(struct percpu_rw_semaphore *sem)
 200 {
 201         if (per_cpu_sum(*sem->read_count) != 0)
 202                 return false;
 203
 204         /*
 205          * If we observed the decrement; ensure we see the entire critical
 206          * section.
 207          */
 208
 209         smp_mb(); /* C matches B */
 210
 211         return true;
 212 }
 213
 214 void percpu_down_write(struct percpu_rw_semaphore *sem)
 215 {
 216         might_sleep();
 217         rwsem_acquire(&sem->dep_map, 0, 0, _RET_IP_);
 218
 219         /* Notify readers to take the slow path. */
 220         rcu_sync_enter(&sem->rss);
 221
 222         /*
 223          * Try set sem->block; this provides writer-writer exclusion.
 224          * Having sem->block set makes new readers block.
 225          */
 226         if (!__percpu_down_write_trylock(sem))
 227                 percpu_rwsem_wait(sem, /* .reader = */ false);
 228
 229         /* smp_mb() implied by __percpu_down_write_trylock() on success -- D matches A */
 230
 231         /*
 232          * If they don't see our store of sem->block, then we are guaranteed to
 233          * see their sem->read_count increment, and therefore will wait for
 234          * them.
 235          */
 236
 237         /* Wait for all active readers to complete. */
 238         rcuwait_wait_event(&sem->writer, readers_active_check(sem), TASK_UNINTERRUPTIBLE);
 239 }
 240 EXPORT_SYMBOL_GPL(percpu_down_write);
 241
 242 void percpu_up_write(struct percpu_rw_semaphore *sem)
 243 {
 244         rwsem_release(&sem->dep_map, _RET_IP_);
 245
 246         /*
 247          * Signal the writer is done, no fast path yet.
 248          *
 249          * One reason that we cannot just immediately flip to readers_fast is
 250          * that new readers might fail to see the results of this writer's
 251          * critical section.
 252          *
 253          * Therefore we force it through the slow path which guarantees an
 254          * acquire and thereby guarantees the critical section's consistency.
 255          */
 256         atomic_set_release(&sem->block, 0);
 257
 258         /*
 259          * Prod any pending reader/writer to make progress.
 260          */
 261         __wake_up(&sem->waiters, TASK_NORMAL, 1, sem);
 262
 263         /*
 264          * Once this completes (at least one RCU-sched grace period hence) the
 265          * reader fast path will be available again. Safe to use outside the
 266          * exclusive write lock because its counting.
 267          */
 268         rcu_sync_exit(&sem->rss);
 269 }
 270 EXPORT_SYMBOL_GPL(percpu_up_write);