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_acquire(&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         /*
  96          * We need both ACQUIRE (pairs with corresponding RELEASE in
  97          * unlock() uncontended, or fastpath) and RELEASE (to publish
  98          * the node fields we just initialised) semantics when updating
  99          * the lock tail.
 100          */
 101         old = atomic_xchg(&lock->tail, curr);
 102         if (old == OSQ_UNLOCKED_VAL)
 103                 return true;
 104
 105         prev = decode_cpu(old);
 106         node->prev = prev;
 107
 108         /*
 109          * osq_lock()                   unqueue
 110          *
 111          * node->prev = prev            osq_wait_next()
 112          * WMB                          MB
 113          * prev->next = node            next->prev = prev // unqueue-C
 114          *
 115          * Here 'node->prev' and 'next->prev' are the same variable and we need
 116          * to ensure these stores happen in-order to avoid corrupting the list.
 117          */
 118         smp_wmb();
 119
 120         WRITE_ONCE(prev->next, node);
 121
 122         /*
 123          * Normally @prev is untouchable after the above store; because at that
 124          * moment unlock can proceed and wipe the node element from stack.
 125          *
 126          * However, since our nodes are static per-cpu storage, we're
 127          * guaranteed their existence -- this allows us to apply
 128          * cmpxchg in an attempt to undo our queueing.
 129          */
 130
 131         while (!READ_ONCE(node->locked)) {
 132                 /*
 133                  * If we need to reschedule bail... so we can block.
 134                  */
 135                 if (need_resched())
 136                         goto unqueue;
 137
 138                 cpu_relax_lowlatency();
 139         }
 140         return true;
 141
 142 unqueue:
 143         /*
 144          * Step - A  -- stabilize @prev
 145          *
 146          * Undo our @prev->next assignment; this will make @prev's
 147          * unlock()/unqueue() wait for a next pointer since @lock points to us
 148          * (or later).
 149          */
 150
 151         for (;;) {
 152                 if (prev->next == node &&
 153                     cmpxchg(&prev->next, node, NULL) == node)
 154                         break;
 155
 156                 /*
 157                  * We can only fail the cmpxchg() racing against an unlock(),
 158                  * in which case we should observe @node->locked becomming
 159                  * true.
 160                  */
 161                 if (smp_load_acquire(&node->locked))
 162                         return true;
 163
 164                 cpu_relax_lowlatency();
 165
 166                 /*
 167                  * Or we race against a concurrent unqueue()'s step-B, in which
 168                  * case its step-C will write us a new @node->prev pointer.
 169                  */
 170                 prev = READ_ONCE(node->prev);
 171         }
 172
 173         /*
 174          * Step - B -- stabilize @next
 175          *
 176          * Similar to unlock(), wait for @node->next or move @lock from @node
 177          * back to @prev.
 178          */
 179
 180         next = osq_wait_next(lock, node, prev);
 181         if (!next)
 182                 return false;
 183
 184         /*
 185          * Step - C -- unlink
 186          *
 187          * @prev is stable because its still waiting for a new @prev->next
 188          * pointer, @next is stable because our @node->next pointer is NULL and
 189          * it will wait in Step-A.
 190          */
 191
 192         WRITE_ONCE(next->prev, prev);
 193         WRITE_ONCE(prev->next, next);
 194
 195         return false;
 196 }
 197
 198 void osq_unlock(struct optimistic_spin_queue *lock)
 199 {
 200         struct optimistic_spin_node *node, *next;
 201         int curr = encode_cpu(smp_processor_id());
 202
 203         /*
 204          * Fast path for the uncontended case.
 205          */
 206         if (likely(atomic_cmpxchg_release(&lock->tail, curr,
 207                                           OSQ_UNLOCKED_VAL) == curr))
 208                 return;
 209
 210         /*
 211          * Second most likely case.
 212          */
 213         node = this_cpu_ptr(&osq_node);
 214         next = xchg(&node->next, NULL);
 215         if (next) {
 216                 WRITE_ONCE(next->locked, 1);
 217                 return;
 218         }
 219
 220         next = osq_wait_next(lock, node, NULL);
 221         if (next)
 222                 WRITE_ONCE(next->locked, 1);
 223 }