Branch data     Line data    Source code

       1                 :            : // SPDX-License-Identifier: GPL-2.0-only
       2                 :            : #include <linux/atomic.h>
       3                 :            : #include <linux/rwsem.h>
       4                 :            : #include <linux/percpu.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                 :            : #include "rwsem.h"
      12                 :            : 
      13                 :          3 : int __percpu_init_rwsem(struct percpu_rw_semaphore *sem,
      14                 :            :                         const char *name, struct lock_class_key *rwsem_key)
      15                 :            : {
      16                 :          3 :         sem->read_count = alloc_percpu(int);
      17                 :          3 :         if (unlikely(!sem->read_count))
      18                 :            :                 return -ENOMEM;
      19                 :            : 
      20                 :            :         /* ->rw_sem represents the whole percpu_rw_semaphore for lockdep */
      21                 :          3 :         rcu_sync_init(&sem->rss);
      22                 :          3 :         __init_rwsem(&sem->rw_sem, name, rwsem_key);
      23                 :            :         rcuwait_init(&sem->writer);
      24                 :            :         sem->readers_block = 0;
      25                 :          3 :         return 0;
      26                 :            : }
      27                 :            : EXPORT_SYMBOL_GPL(__percpu_init_rwsem);
      28                 :            : 
      29                 :          3 : void percpu_free_rwsem(struct percpu_rw_semaphore *sem)
      30                 :            : {
      31                 :            :         /*
      32                 :            :          * XXX: temporary kludge. The error path in alloc_super()
      33                 :            :          * assumes that percpu_free_rwsem() is safe after kzalloc().
      34                 :            :          */
      35                 :          3 :         if (!sem->read_count)
      36                 :          3 :                 return;
      37                 :            : 
      38                 :          3 :         rcu_sync_dtor(&sem->rss);
      39                 :          3 :         free_percpu(sem->read_count);
      40                 :          3 :         sem->read_count = NULL; /* catch use after free bugs */
      41                 :            : }
      42                 :            : EXPORT_SYMBOL_GPL(percpu_free_rwsem);
      43                 :            : 
      44                 :          3 : int __percpu_down_read(struct percpu_rw_semaphore *sem, int try)
      45                 :            : {
      46                 :            :         /*
      47                 :            :          * Due to having preemption disabled the decrement happens on
      48                 :            :          * the same CPU as the increment, avoiding the
      49                 :            :          * increment-on-one-CPU-and-decrement-on-another problem.
      50                 :            :          *
      51                 :            :          * If the reader misses the writer's assignment of readers_block, then
      52                 :            :          * the writer is guaranteed to see the reader's increment.
      53                 :            :          *
      54                 :            :          * Conversely, any readers that increment their sem->read_count after
      55                 :            :          * the writer looks are guaranteed to see the readers_block value,
      56                 :            :          * which in turn means that they are guaranteed to immediately
      57                 :            :          * decrement their sem->read_count, so that it doesn't matter that the
      58                 :            :          * writer missed them.
      59                 :            :          */
      60                 :            : 
      61                 :          3 :         smp_mb(); /* A matches D */
      62                 :            : 
      63                 :            :         /*
      64                 :            :          * If !readers_block the critical section starts here, matched by the
      65                 :            :          * release in percpu_up_write().
      66                 :            :          */
      67                 :          3 :         if (likely(!smp_load_acquire(&sem->readers_block)))
      68                 :            :                 return 1;
      69                 :            : 
      70                 :            :         /*
      71                 :            :          * Per the above comment; we still have preemption disabled and
      72                 :            :          * will thus decrement on the same CPU as we incremented.
      73                 :            :          */
      74                 :          3 :         __percpu_up_read(sem);
      75                 :            : 
      76                 :          3 :         if (try)
      77                 :            :                 return 0;
      78                 :            : 
      79                 :            :         /*
      80                 :            :          * We either call schedule() in the wait, or we'll fall through
      81                 :            :          * and reschedule on the preempt_enable() in percpu_down_read().
      82                 :            :          */
      83                 :          3 :         preempt_enable_no_resched();
      84                 :            : 
      85                 :            :         /*
      86                 :            :          * Avoid lockdep for the down/up_read() we already have them.
      87                 :            :          */
      88                 :          3 :         __down_read(&sem->rw_sem);
      89                 :          3 :         this_cpu_inc(*sem->read_count);
      90                 :          3 :         __up_read(&sem->rw_sem);
      91                 :            : 
      92                 :          3 :         preempt_disable();
      93                 :          3 :         return 1;
      94                 :            : }
      95                 :            : EXPORT_SYMBOL_GPL(__percpu_down_read);
      96                 :            : 
      97                 :          3 : void __percpu_up_read(struct percpu_rw_semaphore *sem)
      98                 :            : {
      99                 :          3 :         smp_mb(); /* B matches C */
     100                 :            :         /*
     101                 :            :          * In other words, if they see our decrement (presumably to aggregate
     102                 :            :          * zero, as that is the only time it matters) they will also see our
     103                 :            :          * critical section.
     104                 :            :          */
     105                 :          3 :         __this_cpu_dec(*sem->read_count);
     106                 :            : 
     107                 :            :         /* Prod writer to recheck readers_active */
     108                 :          3 :         rcuwait_wake_up(&sem->writer);
     109                 :          3 : }
     110                 :            : EXPORT_SYMBOL_GPL(__percpu_up_read);
     111                 :            : 
     112                 :            : #define per_cpu_sum(var)                                                \
     113                 :            : ({                                                                      \
     114                 :            :         typeof(var) __sum = 0;                                          \
     115                 :            :         int cpu;                                                        \
     116                 :            :         compiletime_assert_atomic_type(__sum);                          \
     117                 :            :         for_each_possible_cpu(cpu)                                      \
     118                 :            :                 __sum += per_cpu(var, cpu);                             \
     119                 :            :         __sum;                                                          \
     120                 :            : })
     121                 :            : 
     122                 :            : /*
     123                 :            :  * Return true if the modular sum of the sem->read_count per-CPU variable is
     124                 :            :  * zero.  If this sum is zero, then it is stable due to the fact that if any
     125                 :            :  * newly arriving readers increment a given counter, they will immediately
     126                 :            :  * decrement that same counter.
     127                 :            :  */
     128                 :          3 : static bool readers_active_check(struct percpu_rw_semaphore *sem)
     129                 :            : {
     130                 :          3 :         if (per_cpu_sum(*sem->read_count) != 0)
     131                 :            :                 return false;
     132                 :            : 
     133                 :            :         /*
     134                 :            :          * If we observed the decrement; ensure we see the entire critical
     135                 :            :          * section.
     136                 :            :          */
     137                 :            : 
     138                 :          3 :         smp_mb(); /* C matches B */
     139                 :            : 
     140                 :          3 :         return true;
     141                 :            : }
     142                 :            : 
     143                 :          3 : void percpu_down_write(struct percpu_rw_semaphore *sem)
     144                 :            : {
     145                 :            :         /* Notify readers to take the slow path. */
     146                 :          3 :         rcu_sync_enter(&sem->rss);
     147                 :            : 
     148                 :          3 :         down_write(&sem->rw_sem);
     149                 :            : 
     150                 :            :         /*
     151                 :            :          * Notify new readers to block; up until now, and thus throughout the
     152                 :            :          * longish rcu_sync_enter() above, new readers could still come in.
     153                 :            :          */
     154                 :            :         WRITE_ONCE(sem->readers_block, 1);
     155                 :            : 
     156                 :          3 :         smp_mb(); /* D matches A */
     157                 :            : 
     158                 :            :         /*
     159                 :            :          * If they don't see our writer of readers_block, then we are
     160                 :            :          * guaranteed to see their sem->read_count increment, and therefore
     161                 :            :          * will wait for them.
     162                 :            :          */
     163                 :            : 
     164                 :            :         /* Wait for all now active readers to complete. */
     165                 :          3 :         rcuwait_wait_event(&sem->writer, readers_active_check(sem));
     166                 :          3 : }
     167                 :            : EXPORT_SYMBOL_GPL(percpu_down_write);
     168                 :            : 
     169                 :          3 : void percpu_up_write(struct percpu_rw_semaphore *sem)
     170                 :            : {
     171                 :            :         /*
     172                 :            :          * Signal the writer is done, no fast path yet.
     173                 :            :          *
     174                 :            :          * One reason that we cannot just immediately flip to readers_fast is
     175                 :            :          * that new readers might fail to see the results of this writer's
     176                 :            :          * critical section.
     177                 :            :          *
     178                 :            :          * Therefore we force it through the slow path which guarantees an
     179                 :            :          * acquire and thereby guarantees the critical section's consistency.
     180                 :            :          */
     181                 :          3 :         smp_store_release(&sem->readers_block, 0);
     182                 :            : 
     183                 :            :         /*
     184                 :            :          * Release the write lock, this will allow readers back in the game.
     185                 :            :          */
     186                 :          3 :         up_write(&sem->rw_sem);
     187                 :            : 
     188                 :            :         /*
     189                 :            :          * Once this completes (at least one RCU-sched grace period hence) the
     190                 :            :          * reader fast path will be available again. Safe to use outside the
     191                 :            :          * exclusive write lock because its counting.
     192                 :            :          */
     193                 :          3 :         rcu_sync_exit(&sem->rss);
     194                 :          3 : }
     195                 :            : EXPORT_SYMBOL_GPL(percpu_up_write);