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_node, osq_node);
  18
  19 /*
  20  * We use the value 0 to represent "no CPU", thus the encoded value
  21  * will be the CPU number incremented by 1.
  22  */
  23 static inline int encode_cpu(int cpu_nr)
  24 {
  25         return cpu_nr + 1;
  26 }
  27
  28 static inline struct optimistic_spin_node *decode_cpu(int encoded_cpu_val)
  29 {
  30         int cpu_nr = encoded_cpu_val - 1;
  31
  32         return per_cpu_ptr(&osq_node, cpu_nr);
  33 }
  34
  35 /*
  36  * Get a stable @node->next pointer, either for unlock() or unqueue() purposes.
  37  * Can return NULL in case we were the last queued and we updated @lock instead.
  38  */
  39 static inline struct optimistic_spin_node *
  40 osq_wait_next(struct optimistic_spin_queue *lock,
  41               struct optimistic_spin_node *node,
  42               struct optimistic_spin_node *prev)
  43 {
  44         struct optimistic_spin_node *next = NULL;
  45         int curr = encode_cpu(smp_processor_id());
  46         int old;
  47
  48         /*
  49          * If there is a prev node in queue, then the 'old' value will be
  50          * the prev node's CPU #, else it's set to OSQ_UNLOCKED_VAL since if
  51          * we're currently last in queue, then the queue will then become empty.
  52          */
  53         old = prev ? prev->cpu : OSQ_UNLOCKED_VAL;
  54
  55         for (;;) {
  56                 if (atomic_read(&lock->tail) == curr &&
  57                     atomic_cmpxchg(&lock->tail, curr, old) == curr) {
  58                         /*
  59                          * We were the last queued, we moved @lock back. @prev
  60                          * will now observe @lock and will complete its
  61                          * unlock()/unqueue().
  62                          */
  63                         break;
  64                 }
  65
  66                 /*
  67                  * We must xchg() the @node->next value, because if we were to
  68                  * leave it in, a concurrent unlock()/unqueue() from
  69                  * @node->next might complete Step-A and think its @prev is
  70                  * still valid.
  71                  *
  72                  * If the concurrent unlock()/unqueue() wins the race, we'll
  73                  * wait for either @lock to point to us, through its Step-B, or
  74                  * wait for a new @node->next from its Step-C.
  75                  */
  76                 if (node->next) {
  77                         next = xchg(&node->next, NULL);
  78                         if (next)
  79                                 break;
  80                 }
  81
  82                 arch_mutex_cpu_relax();
  83         }
  84
  85         return next;
  86 }
  87
  88 bool osq_lock(struct optimistic_spin_queue *lock)
  89 {
  90         struct optimistic_spin_node *node = this_cpu_ptr(&osq_node);
  91         struct optimistic_spin_node *prev, *next;
  92         int curr = encode_cpu(smp_processor_id());
  93         int old;
  94
  95         node->locked = 0;
  96         node->next = NULL;
  97         node->cpu = curr;
  98
  99         old = atomic_xchg(&lock->tail, curr);
 100         if (old == OSQ_UNLOCKED_VAL)
 101                 return true;
 102
 103         prev = decode_cpu(old);
 104         node->prev = prev;
 105         ACCESS_ONCE(prev->next) = node;
 106
 107         /*
 108          * Normally @prev is untouchable after the above store; because at that
 109          * moment unlock can proceed and wipe the node element from stack.
 110          *
 111          * However, since our nodes are static per-cpu storage, we're
 112          * guaranteed their existence -- this allows us to apply
 113          * cmpxchg in an attempt to undo our queueing.
 114          */
 115
 116         while (!smp_load_acquire(&node->locked)) {
 117                 /*
 118                  * If we need to reschedule bail... so we can block.
 119                  */
 120                 if (need_resched())
 121                         goto unqueue;
 122
 123                 arch_mutex_cpu_relax();
 124         }
 125         return true;
 126
 127 unqueue:
 128         /*
 129          * Step - A  -- stabilize @prev
 130          *
 131          * Undo our @prev->next assignment; this will make @prev's
 132          * unlock()/unqueue() wait for a next pointer since @lock points to us
 133          * (or later).
 134          */
 135
 136         for (;;) {
 137                 if (prev->next == node &&
 138                     cmpxchg(&prev->next, node, NULL) == node)
 139                         break;
 140
 141                 /*
 142                  * We can only fail the cmpxchg() racing against an unlock(),
 143                  * in which case we should observe @node->locked becomming
 144                  * true.
 145                  */
 146                 if (smp_load_acquire(&node->locked))
 147                         return true;
 148
 149                 arch_mutex_cpu_relax();
 150
 151                 /*
 152                  * Or we race against a concurrent unqueue()'s step-B, in which
 153                  * case its step-C will write us a new @node->prev pointer.
 154                  */
 155                 prev = ACCESS_ONCE(node->prev);
 156         }
 157
 158         /*
 159          * Step - B -- stabilize @next
 160          *
 161          * Similar to unlock(), wait for @node->next or move @lock from @node
 162          * back to @prev.
 163          */
 164
 165         next = osq_wait_next(lock, node, prev);
 166         if (!next)
 167                 return false;
 168
 169         /*
 170          * Step - C -- unlink
 171          *
 172          * @prev is stable because its still waiting for a new @prev->next
 173          * pointer, @next is stable because our @node->next pointer is NULL and
 174          * it will wait in Step-A.
 175          */
 176
 177         ACCESS_ONCE(next->prev) = prev;
 178         ACCESS_ONCE(prev->next) = next;
 179
 180         return false;
 181 }
 182
 183 void osq_unlock(struct optimistic_spin_queue *lock)
 184 {
 185         struct optimistic_spin_node *node, *next;
 186         int curr = encode_cpu(smp_processor_id());
 187
 188         /*
 189          * Fast path for the uncontended case.
 190          */
 191         if (likely(atomic_cmpxchg(&lock->tail, curr, OSQ_UNLOCKED_VAL) == curr))
 192                 return;
 193
 194         /*
 195          * Second most likely case.
 196          */
 197         node = this_cpu_ptr(&osq_node);
 198         next = xchg(&node->next, NULL);
 199         if (next) {
 200                 ACCESS_ONCE(next->locked) = 1;
 201                 return;
 202         }
 203
 204         next = osq_wait_next(lock, node, NULL);
 205         if (next)
 206                 ACCESS_ONCE(next->locked) = 1;
 207 }
 208
 209 #endif
 210