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 int node_cpu(struct optimistic_spin_node *node)
  25 {
  26         return node->cpu - 1;
  27 }
  28
  29 static inline struct optimistic_spin_node *decode_cpu(int encoded_cpu_val)
  30 {
  31         int cpu_nr = encoded_cpu_val - 1;
  32
  33         return per_cpu_ptr(&osq_node, cpu_nr);
  34 }
  35
  36 /*
  37  * Get a stable @node->next pointer, either for unlock() or unqueue() purposes.
  38  * Can return NULL in case we were the last queued and we updated @lock instead.
  39  */
  40 static inline struct optimistic_spin_node *
  41 osq_wait_next(struct optimistic_spin_queue *lock,
  42               struct optimistic_spin_node *node,
  43               struct optimistic_spin_node *prev)
  44 {
  45         struct optimistic_spin_node *next = NULL;
  46         int curr = encode_cpu(smp_processor_id());
  47         int old;
  48
  49         /*
  50          * If there is a prev node in queue, then the 'old' value will be
  51          * the prev node's CPU #, else it's set to OSQ_UNLOCKED_VAL since if
  52          * we're currently last in queue, then the queue will then become empty.
  53          */
  54         old = prev ? prev->cpu : OSQ_UNLOCKED_VAL;
  55
  56         for (;;) {
  57                 if (atomic_read(&lock->tail) == curr &&
  58                     atomic_cmpxchg_acquire(&lock->tail, curr, old) == curr) {
  59                         /*
  60                          * We were the last queued, we moved @lock back. @prev
  61                          * will now observe @lock and will complete its
  62                          * unlock()/unqueue().
  63                          */
  64                         break;
  65                 }
  66
  67                 /*
  68                  * We must xchg() the @node->next value, because if we were to
  69                  * leave it in, a concurrent unlock()/unqueue() from
  70                  * @node->next might complete Step-A and think its @prev is
  71                  * still valid.
  72                  *
  73                  * If the concurrent unlock()/unqueue() wins the race, we'll
  74                  * wait for either @lock to point to us, through its Step-B, or
  75                  * wait for a new @node->next from its Step-C.
  76                  */
  77                 if (node->next) {
  78                         next = xchg(&node->next, NULL);
  79                         if (next)
  80                                 break;
  81                 }
  82
  83                 cpu_relax();
  84         }
  85
  86         return next;
  87 }
  88
  89 bool osq_lock(struct optimistic_spin_queue *lock)
  90 {
  91         struct optimistic_spin_node *node = this_cpu_ptr(&osq_node);
  92         struct optimistic_spin_node *prev, *next;
  93         int curr = encode_cpu(smp_processor_id());
  94         int old;
  95
  96         node->locked = 0;
  97         node->next = NULL;
  98         node->cpu = curr;
  99
 100         /*
 101          * We need both ACQUIRE (pairs with corresponding RELEASE in
 102          * unlock() uncontended, or fastpath) and RELEASE (to publish
 103          * the node fields we just initialised) semantics when updating
 104          * the lock tail.
 105          */
 106         old = atomic_xchg(&lock->tail, curr);
 107         if (old == OSQ_UNLOCKED_VAL)
 108                 return true;
 109
 110         prev = decode_cpu(old);
 111         node->prev = prev;
 112
 113         /*
 114          * osq_lock()                   unqueue
 115          *
 116          * node->prev = prev            osq_wait_next()
 117          * WMB                          MB
 118          * prev->next = node            next->prev = prev // unqueue-C
 119          *
 120          * Here 'node->prev' and 'next->prev' are the same variable and we need
 121          * to ensure these stores happen in-order to avoid corrupting the list.
 122          */
 123         smp_wmb();
 124
 125         WRITE_ONCE(prev->next, node);
 126
 127         /*
 128          * Normally @prev is untouchable after the above store; because at that
 129          * moment unlock can proceed and wipe the node element from stack.
 130          *
 131          * However, since our nodes are static per-cpu storage, we're
 132          * guaranteed their existence -- this allows us to apply
 133          * cmpxchg in an attempt to undo our queueing.
 134          */
 135
 136         while (!READ_ONCE(node->locked)) {
 137                 /*
 138                  * If we need to reschedule bail... so we can block.
 139                  * Use vcpu_is_preempted() to avoid waiting for a preempted
 140                  * lock holder:
 141                  */
 142                 if (need_resched() || vcpu_is_preempted(node_cpu(node->prev)))
 143                         goto unqueue;
 144
 145                 cpu_relax();
 146         }
 147         return true;
 148
 149 unqueue:
 150         /*
 151          * Step - A  -- stabilize @prev
 152          *
 153          * Undo our @prev->next assignment; this will make @prev's
 154          * unlock()/unqueue() wait for a next pointer since @lock points to us
 155          * (or later).
 156          */
 157
 158         for (;;) {
 159                 if (prev->next == node &&
 160                     cmpxchg(&prev->next, node, NULL) == node)
 161                         break;
 162
 163                 /*
 164                  * We can only fail the cmpxchg() racing against an unlock(),
 165                  * in which case we should observe @node->locked becomming
 166                  * true.
 167                  */
 168                 if (smp_load_acquire(&node->locked))
 169                         return true;
 170
 171                 cpu_relax();
 172
 173                 /*
 174                  * Or we race against a concurrent unqueue()'s step-B, in which
 175                  * case its step-C will write us a new @node->prev pointer.
 176                  */
 177                 prev = READ_ONCE(node->prev);
 178         }
 179
 180         /*
 181          * Step - B -- stabilize @next
 182          *
 183          * Similar to unlock(), wait for @node->next or move @lock from @node
 184          * back to @prev.
 185          */
 186
 187         next = osq_wait_next(lock, node, prev);
 188         if (!next)
 189                 return false;
 190
 191         /*
 192          * Step - C -- unlink
 193          *
 194          * @prev is stable because its still waiting for a new @prev->next
 195          * pointer, @next is stable because our @node->next pointer is NULL and
 196          * it will wait in Step-A.
 197          */
 198
 199         WRITE_ONCE(next->prev, prev);
 200         WRITE_ONCE(prev->next, next);
 201
 202         return false;
 203 }
 204
 205 void osq_unlock(struct optimistic_spin_queue *lock)
 206 {
 207         struct optimistic_spin_node *node, *next;
 208         int curr = encode_cpu(smp_processor_id());
 209
 210         /*
 211          * Fast path for the uncontended case.
 212          */
 213         if (likely(atomic_cmpxchg_release(&lock->tail, curr,
 214                                           OSQ_UNLOCKED_VAL) == curr))
 215                 return;
 216
 217         /*
 218          * Second most likely case.
 219          */
 220         node = this_cpu_ptr(&osq_node);
 221         next = xchg(&node->next, NULL);
 222         if (next) {
 223                 WRITE_ONCE(next->locked, 1);
 224                 return;
 225         }
 226
 227         next = osq_wait_next(lock, node, NULL);
 228         if (next)
 229                 WRITE_ONCE(next->locked, 1);
 230 }