kernel/locking/mcs_spinlock.c

   1
   2 #include <linux/percpu.h>
   3 #include <linux/mutex.h>
   4 #include <linux/sched.h>
   5 #include "mcs_spinlock.h"
   6
   7 #ifdef CONFIG_SMP
   8
   9 /*
  10  * An MCS like lock especially tailored for optimistic spinning for sleeping
  11  * lock implementations (mutex, rwsem, etc).
  12  *
  13  * Using a single mcs node per CPU is safe because sleeping locks should not be
  14  * called from interrupt context and we have preemption disabled while
  15  * spinning.
  16  */
  17 static DEFINE_PER_CPU_SHARED_ALIGNED(struct optimistic_spin_queue, osq_node);
  18
  19 /*
  20  * Get a stable @node->next pointer, either for unlock() or unqueue() purposes.
  21  * Can return NULL in case we were the last queued and we updated @lock instead.
  22  */
  23 static inline struct optimistic_spin_queue *
  24 osq_wait_next(struct optimistic_spin_queue **lock,
  25               struct optimistic_spin_queue *node,
  26               struct optimistic_spin_queue *prev)
  27 {
  28         struct optimistic_spin_queue *next = NULL;
  29
  30         for (;;) {
  31                 if (*lock == node && cmpxchg(lock, node, prev) == node) {
  32                         /*
  33                          * We were the last queued, we moved @lock back. @prev
  34                          * will now observe @lock and will complete its
  35                          * unlock()/unqueue().
  36                          */
  37                         break;
  38                 }
  39
  40                 /*
  41                  * We must xchg() the @node->next value, because if we were to
  42                  * leave it in, a concurrent unlock()/unqueue() from
  43                  * @node->next might complete Step-A and think its @prev is
  44                  * still valid.
  45                  *
  46                  * If the concurrent unlock()/unqueue() wins the race, we'll
  47                  * wait for either @lock to point to us, through its Step-B, or
  48                  * wait for a new @node->next from its Step-C.
  49                  */
  50                 if (node->next) {
  51                         next = xchg(&node->next, NULL);
  52                         if (next)
  53                                 break;
  54                 }
  55
  56                 arch_mutex_cpu_relax();
  57         }
  58
  59         return next;
  60 }
  61
  62 bool osq_lock(struct optimistic_spin_queue **lock)
  63 {
  64         struct optimistic_spin_queue *node = this_cpu_ptr(&osq_node);
  65         struct optimistic_spin_queue *prev, *next;
  66
  67         node->locked = 0;
  68         node->next = NULL;
  69
  70         node->prev = prev = xchg(lock, node);
  71         if (likely(prev == NULL))
  72                 return true;
  73
  74         ACCESS_ONCE(prev->next) = node;
  75
  76         /*
  77          * Normally @prev is untouchable after the above store; because at that
  78          * moment unlock can proceed and wipe the node element from stack.
  79          *
  80          * However, since our nodes are static per-cpu storage, we're
  81          * guaranteed their existence -- this allows us to apply
  82          * cmpxchg in an attempt to undo our queueing.
  83          */
  84
  85         while (!smp_load_acquire(&node->locked)) {
  86                 /*
  87                  * If we need to reschedule bail... so we can block.
  88                  */
  89                 if (need_resched())
  90                         goto unqueue;
  91
  92                 arch_mutex_cpu_relax();
  93         }
  94         return true;
  95
  96 unqueue:
  97         /*
  98          * Step - A  -- stabilize @prev
  99          *
 100          * Undo our @prev->next assignment; this will make @prev's
 101          * unlock()/unqueue() wait for a next pointer since @lock points to us
 102          * (or later).
 103          */
 104
 105         for (;;) {
 106                 if (prev->next == node &&
 107                     cmpxchg(&prev->next, node, NULL) == node)
 108                         break;
 109
 110                 /*
 111                  * We can only fail the cmpxchg() racing against an unlock(),
 112                  * in which case we should observe @node->locked becomming
 113                  * true.
 114                  */
 115                 if (smp_load_acquire(&node->locked))
 116                         return true;
 117
 118                 arch_mutex_cpu_relax();
 119
 120                 /*
 121                  * Or we race against a concurrent unqueue()'s step-B, in which
 122                  * case its step-C will write us a new @node->prev pointer.
 123                  */
 124                 prev = ACCESS_ONCE(node->prev);
 125         }
 126
 127         /*
 128          * Step - B -- stabilize @next
 129          *
 130          * Similar to unlock(), wait for @node->next or move @lock from @node
 131          * back to @prev.
 132          */
 133
 134         next = osq_wait_next(lock, node, prev);
 135         if (!next)
 136                 return false;
 137
 138         /*
 139          * Step - C -- unlink
 140          *
 141          * @prev is stable because its still waiting for a new @prev->next
 142          * pointer, @next is stable because our @node->next pointer is NULL and
 143          * it will wait in Step-A.
 144          */
 145
 146         ACCESS_ONCE(next->prev) = prev;
 147         ACCESS_ONCE(prev->next) = next;
 148
 149         return false;
 150 }
 151
 152 void osq_unlock(struct optimistic_spin_queue **lock)
 153 {
 154         struct optimistic_spin_queue *node = this_cpu_ptr(&osq_node);
 155         struct optimistic_spin_queue *next;
 156
 157         /*
 158          * Fast path for the uncontended case.
 159          */
 160         if (likely(cmpxchg(lock, node, NULL) == node))
 161                 return;
 162
 163         /*
 164          * Second most likely case.
 165          */
 166         next = xchg(&node->next, NULL);
 167         if (next) {
 168                 ACCESS_ONCE(next->locked) = 1;
 169                 return;
 170         }
 171
 172         next = osq_wait_next(lock, node, NULL);
 173         if (next)
 174                 ACCESS_ONCE(next->locked) = 1;
 175 }
 176
 177 #endif
 178