kernel/locking/osq_lock.c

   1 #include <linux/percpu.h>
   2 #include <linux/sched.h>
   3 #include <linux/osq_lock.h>
   4
   5 /*
   6  * An MCS like lock especially tailored for optimistic spinning for sleeping
   7  * lock implementations (mutex, rwsem, etc).
   8  *
   9  * Using a single mcs node per CPU is safe because sleeping locks should not be
  10  * called from interrupt context and we have preemption disabled while
  11  * spinning.
  12  */
  13 static DEFINE_PER_CPU_SHARED_ALIGNED(struct optimistic_spin_node, osq_node);
  14
  15 /*
  16  * We use the value 0 to represent "no CPU", thus the encoded value
  17  * will be the CPU number incremented by 1.
  18  */
  19 static inline int encode_cpu(int cpu_nr)
  20 {
  21         return cpu_nr + 1;
  22 }
  23
  24 static inline struct optimistic_spin_node *decode_cpu(int encoded_cpu_val)
  25 {
  26         int cpu_nr = encoded_cpu_val - 1;
  27
  28         return per_cpu_ptr(&osq_node, cpu_nr);
  29 }
  30
  31 /*
  32  * Get a stable @node->next pointer, either for unlock() or unqueue() purposes.
  33  * Can return NULL in case we were the last queued and we updated @lock instead.
  34  */
  35 static inline struct optimistic_spin_node *
  36 osq_wait_next(struct optimistic_spin_queue *lock,
  37               struct optimistic_spin_node *node,
  38               struct optimistic_spin_node *prev)
  39 {
  40         struct optimistic_spin_node *next = NULL;
  41         int curr = encode_cpu(smp_processor_id());
  42         int old;
  43
  44         /*
  45          * If there is a prev node in queue, then the 'old' value will be
  46          * the prev node's CPU #, else it's set to OSQ_UNLOCKED_VAL since if
  47          * we're currently last in queue, then the queue will then become empty.
  48          */
  49         old = prev ? prev->cpu : OSQ_UNLOCKED_VAL;
  50
  51         for (;;) {
  52                 if (atomic_read(&lock->tail) == curr &&
  53                     atomic_cmpxchg(&lock->tail, curr, old) == curr) {
  54                         /*
  55                          * We were the last queued, we moved @lock back. @prev
  56                          * will now observe @lock and will complete its
  57                          * unlock()/unqueue().
  58                          */
  59                         break;
  60                 }
  61
  62                 /*
  63                  * We must xchg() the @node->next value, because if we were to
  64                  * leave it in, a concurrent unlock()/unqueue() from
  65                  * @node->next might complete Step-A and think its @prev is
  66                  * still valid.
  67                  *
  68                  * If the concurrent unlock()/unqueue() wins the race, we'll
  69                  * wait for either @lock to point to us, through its Step-B, or
  70                  * wait for a new @node->next from its Step-C.
  71                  */
  72                 if (node->next) {
  73                         next = xchg(&node->next, NULL);
  74                         if (next)
  75                                 break;
  76                 }
  77
  78                 cpu_relax_lowlatency();
  79         }
  80
  81         return next;
  82 }
  83
  84 bool osq_lock(struct optimistic_spin_queue *lock)
  85 {
  86         struct optimistic_spin_node *node = this_cpu_ptr(&osq_node);
  87         struct optimistic_spin_node *prev, *next;
  88         int curr = encode_cpu(smp_processor_id());
  89         int old;
  90
  91         node->locked = 0;
  92         node->next = NULL;
  93         node->cpu = curr;
  94
  95         old = atomic_xchg(&lock->tail, curr);
  96         if (old == OSQ_UNLOCKED_VAL)
  97                 return true;
  98
  99         prev = decode_cpu(old);
 100         node->prev = prev;
 101         WRITE_ONCE(prev->next, node);
 102
 103         /*
 104          * Normally @prev is untouchable after the above store; because at that
 105          * moment unlock can proceed and wipe the node element from stack.
 106          *
 107          * However, since our nodes are static per-cpu storage, we're
 108          * guaranteed their existence -- this allows us to apply
 109          * cmpxchg in an attempt to undo our queueing.
 110          */
 111
 112         while (!READ_ONCE(node->locked)) {
 113                 /*
 114                  * If we need to reschedule bail... so we can block.
 115                  */
 116                 if (need_resched())
 117                         goto unqueue;
 118
 119                 cpu_relax_lowlatency();
 120         }
 121         return true;
 122
 123 unqueue:
 124         /*
 125          * Step - A  -- stabilize @prev
 126          *
 127          * Undo our @prev->next assignment; this will make @prev's
 128          * unlock()/unqueue() wait for a next pointer since @lock points to us
 129          * (or later).
 130          */
 131
 132         for (;;) {
 133                 if (prev->next == node &&
 134                     cmpxchg(&prev->next, node, NULL) == node)
 135                         break;
 136
 137                 /*
 138                  * We can only fail the cmpxchg() racing against an unlock(),
 139                  * in which case we should observe @node->locked becomming
 140                  * true.
 141                  */
 142                 if (smp_load_acquire(&node->locked))
 143                         return true;
 144
 145                 cpu_relax_lowlatency();
 146
 147                 /*
 148                  * Or we race against a concurrent unqueue()'s step-B, in which
 149                  * case its step-C will write us a new @node->prev pointer.
 150                  */
 151                 prev = READ_ONCE(node->prev);
 152         }
 153
 154         /*
 155          * Step - B -- stabilize @next
 156          *
 157          * Similar to unlock(), wait for @node->next or move @lock from @node
 158          * back to @prev.
 159          */
 160
 161         next = osq_wait_next(lock, node, prev);
 162         if (!next)
 163                 return false;
 164
 165         /*
 166          * Step - C -- unlink
 167          *
 168          * @prev is stable because its still waiting for a new @prev->next
 169          * pointer, @next is stable because our @node->next pointer is NULL and
 170          * it will wait in Step-A.
 171          */
 172
 173         WRITE_ONCE(next->prev, prev);
 174         WRITE_ONCE(prev->next, next);
 175
 176         return false;
 177 }
 178
 179 void osq_unlock(struct optimistic_spin_queue *lock)
 180 {
 181         struct optimistic_spin_node *node, *next;
 182         int curr = encode_cpu(smp_processor_id());
 183
 184         /*
 185          * Fast path for the uncontended case.
 186          */
 187         if (likely(atomic_cmpxchg(&lock->tail, curr, OSQ_UNLOCKED_VAL) == curr))
 188                 return;
 189
 190         /*
 191          * Second most likely case.
 192          */
 193         node = this_cpu_ptr(&osq_node);
 194         next = xchg(&node->next, NULL);
 195         if (next) {
 196                 WRITE_ONCE(next->locked, 1);
 197                 return;
 198         }
 199
 200         next = osq_wait_next(lock, node, NULL);
 201         if (next)
 202                 WRITE_ONCE(next->locked, 1);
 203 }