arch/x86/include/asm/spinlock.h

   1 #ifndef _ASM_X86_SPINLOCK_H
   2 #define _ASM_X86_SPINLOCK_H
   3
   4 #include <linux/jump_label.h>
   5 #include <linux/atomic.h>
   6 #include <asm/page.h>
   7 #include <asm/processor.h>
   8 #include <linux/compiler.h>
   9 #include <asm/paravirt.h>
  10 #include <asm/bitops.h>
  11
  12 /*
  13  * Your basic SMP spinlocks, allowing only a single CPU anywhere
  14  *
  15  * Simple spin lock operations.  There are two variants, one clears IRQ's
  16  * on the local processor, one does not.
  17  *
  18  * These are fair FIFO ticket locks, which support up to 2^16 CPUs.
  19  *
  20  * (the type definitions are in asm/spinlock_types.h)
  21  */
  22
  23 #ifdef CONFIG_X86_32
  24 # define LOCK_PTR_REG "a"
  25 #else
  26 # define LOCK_PTR_REG "D"
  27 #endif
  28
  29 #if defined(CONFIG_X86_32) && \
  30         (defined(CONFIG_X86_OOSTORE) || defined(CONFIG_X86_PPRO_FENCE))
  31 /*
  32  * On PPro SMP or if we are using OOSTORE, we use a locked operation to unlock
  33  * (PPro errata 66, 92)
  34  */
  35 # define UNLOCK_LOCK_PREFIX LOCK_PREFIX
  36 #else
  37 # define UNLOCK_LOCK_PREFIX
  38 #endif
  39
  40 /* How long a lock should spin before we consider blocking */
  41 #define SPIN_THRESHOLD  (1 << 15)
  42
  43 extern struct static_key paravirt_ticketlocks_enabled;
  44 static __always_inline bool static_key_false(struct static_key *key);
  45
  46 #ifdef CONFIG_PARAVIRT_SPINLOCKS
  47
  48 static inline void __ticket_enter_slowpath(arch_spinlock_t *lock)
  49 {
  50         set_bit(0, (volatile unsigned long *)&lock->tickets.tail);
  51 }
  52
  53 #else  /* !CONFIG_PARAVIRT_SPINLOCKS */
  54 static __always_inline void __ticket_lock_spinning(arch_spinlock_t *lock,
  55                                                         __ticket_t ticket)
  56 {
  57 }
  58 static inline void __ticket_unlock_kick(arch_spinlock_t *lock,
  59                                                         __ticket_t ticket)
  60 {
  61 }
  62
  63 #endif /* CONFIG_PARAVIRT_SPINLOCKS */
  64
  65 static __always_inline int arch_spin_value_unlocked(arch_spinlock_t lock)
  66 {
  67         return lock.tickets.head == lock.tickets.tail;
  68 }
  69
  70 /*
  71  * Ticket locks are conceptually two parts, one indicating the current head of
  72  * the queue, and the other indicating the current tail. The lock is acquired
  73  * by atomically noting the tail and incrementing it by one (thus adding
  74  * ourself to the queue and noting our position), then waiting until the head
  75  * becomes equal to the the initial value of the tail.
  76  *
  77  * We use an xadd covering *both* parts of the lock, to increment the tail and
  78  * also load the position of the head, which takes care of memory ordering
  79  * issues and should be optimal for the uncontended case. Note the tail must be
  80  * in the high part, because a wide xadd increment of the low part would carry
  81  * up and contaminate the high part.
  82  */
  83 static __always_inline void arch_spin_lock(arch_spinlock_t *lock)
  84 {
  85         register struct __raw_tickets inc = { .tail = TICKET_LOCK_INC };
  86
  87         inc = xadd(&lock->tickets, inc);
  88         if (likely(inc.head == inc.tail))
  89                 goto out;
  90
  91         inc.tail &= ~TICKET_SLOWPATH_FLAG;
  92         for (;;) {
  93                 unsigned count = SPIN_THRESHOLD;
  94
  95                 do {
  96                         if (ACCESS_ONCE(lock->tickets.head) == inc.tail)
  97                                 goto out;
  98                         cpu_relax();
  99                 } while (--count);
 100                 __ticket_lock_spinning(lock, inc.tail);
 101         }
 102 out:    barrier();      /* make sure nothing creeps before the lock is taken */
 103 }
 104
 105 static __always_inline int arch_spin_trylock(arch_spinlock_t *lock)
 106 {
 107         arch_spinlock_t old, new;
 108
 109         old.tickets = ACCESS_ONCE(lock->tickets);
 110         if (old.tickets.head != (old.tickets.tail & ~TICKET_SLOWPATH_FLAG))
 111                 return 0;
 112
 113         new.head_tail = old.head_tail + (TICKET_LOCK_INC << TICKET_SHIFT);
 114
 115         /* cmpxchg is a full barrier, so nothing can move before it */
 116         return cmpxchg(&lock->head_tail, old.head_tail, new.head_tail) == old.head_tail;
 117 }
 118
 119 static inline void __ticket_unlock_slowpath(arch_spinlock_t *lock,
 120                                             arch_spinlock_t old)
 121 {
 122         arch_spinlock_t new;
 123
 124         BUILD_BUG_ON(((__ticket_t)NR_CPUS) != NR_CPUS);
 125
 126         /* Perform the unlock on the "before" copy */
 127         old.tickets.head += TICKET_LOCK_INC;
 128
 129         /* Clear the slowpath flag */
 130         new.head_tail = old.head_tail & ~(TICKET_SLOWPATH_FLAG << TICKET_SHIFT);
 131
 132         /*
 133          * If the lock is uncontended, clear the flag - use cmpxchg in
 134          * case it changes behind our back though.
 135          */
 136         if (new.tickets.head != new.tickets.tail ||
 137             cmpxchg(&lock->head_tail, old.head_tail,
 138                                         new.head_tail) != old.head_tail) {
 139                 /*
 140                  * Lock still has someone queued for it, so wake up an
 141                  * appropriate waiter.
 142                  */
 143                 __ticket_unlock_kick(lock, old.tickets.head);
 144         }
 145 }
 146
 147 static __always_inline void arch_spin_unlock(arch_spinlock_t *lock)
 148 {
 149         if (TICKET_SLOWPATH_FLAG &&
 150             static_key_false(&paravirt_ticketlocks_enabled)) {
 151                 arch_spinlock_t prev;
 152
 153                 prev = *lock;
 154                 add_smp(&lock->tickets.head, TICKET_LOCK_INC);
 155
 156                 /* add_smp() is a full mb() */
 157
 158                 if (unlikely(lock->tickets.tail & TICKET_SLOWPATH_FLAG))
 159                         __ticket_unlock_slowpath(lock, prev);
 160         } else
 161                 __add(&lock->tickets.head, TICKET_LOCK_INC, UNLOCK_LOCK_PREFIX);
 162 }
 163
 164 static inline int arch_spin_is_locked(arch_spinlock_t *lock)
 165 {
 166         struct __raw_tickets tmp = ACCESS_ONCE(lock->tickets);
 167
 168         return tmp.tail != tmp.head;
 169 }
 170
 171 static inline int arch_spin_is_contended(arch_spinlock_t *lock)
 172 {
 173         struct __raw_tickets tmp = ACCESS_ONCE(lock->tickets);
 174
 175         return (__ticket_t)(tmp.tail - tmp.head) > TICKET_LOCK_INC;
 176 }
 177 #define arch_spin_is_contended  arch_spin_is_contended
 178
 179 static __always_inline void arch_spin_lock_flags(arch_spinlock_t *lock,
 180                                                   unsigned long flags)
 181 {
 182         arch_spin_lock(lock);
 183 }
 184
 185 static inline void arch_spin_unlock_wait(arch_spinlock_t *lock)
 186 {
 187         __ticket_t head = ACCESS_ONCE(lock->tickets.head);
 188
 189         for (;;) {
 190                 struct __raw_tickets tmp = ACCESS_ONCE(lock->tickets);
 191                 /*
 192                  * We need to check "unlocked" in a loop, tmp.head == head
 193                  * can be false positive because of overflow.
 194                  */
 195                 if (tmp.head == (tmp.tail & ~TICKET_SLOWPATH_FLAG) ||
 196                     tmp.head != head)
 197                         break;
 198
 199                 cpu_relax();
 200         }
 201 }
 202
 203 /*
 204  * Read-write spinlocks, allowing multiple readers
 205  * but only one writer.
 206  *
 207  * NOTE! it is quite common to have readers in interrupts
 208  * but no interrupt writers. For those circumstances we
 209  * can "mix" irq-safe locks - any writer needs to get a
 210  * irq-safe write-lock, but readers can get non-irqsafe
 211  * read-locks.
 212  *
 213  * On x86, we implement read-write locks as a 32-bit counter
 214  * with the high bit (sign) being the "contended" bit.
 215  */
 216
 217 /**
 218  * read_can_lock - would read_trylock() succeed?
 219  * @lock: the rwlock in question.
 220  */
 221 static inline int arch_read_can_lock(arch_rwlock_t *lock)
 222 {
 223         return lock->lock > 0;
 224 }
 225
 226 /**
 227  * write_can_lock - would write_trylock() succeed?
 228  * @lock: the rwlock in question.
 229  */
 230 static inline int arch_write_can_lock(arch_rwlock_t *lock)
 231 {
 232         return lock->write == WRITE_LOCK_CMP;
 233 }
 234
 235 static inline void arch_read_lock(arch_rwlock_t *rw)
 236 {
 237         asm volatile(LOCK_PREFIX READ_LOCK_SIZE(dec) " (%0)\n\t"
 238                      "jns 1f\n"
 239                      "call __read_lock_failed\n\t"
 240                      "1:\n"
 241                      ::LOCK_PTR_REG (rw) : "memory");
 242 }
 243
 244 static inline void arch_write_lock(arch_rwlock_t *rw)
 245 {
 246         asm volatile(LOCK_PREFIX WRITE_LOCK_SUB(%1) "(%0)\n\t"
 247                      "jz 1f\n"
 248                      "call __write_lock_failed\n\t"
 249                      "1:\n"
 250                      ::LOCK_PTR_REG (&rw->write), "i" (RW_LOCK_BIAS)
 251                      : "memory");
 252 }
 253
 254 static inline int arch_read_trylock(arch_rwlock_t *lock)
 255 {
 256         READ_LOCK_ATOMIC(t) *count = (READ_LOCK_ATOMIC(t) *)lock;
 257
 258         if (READ_LOCK_ATOMIC(dec_return)(count) >= 0)
 259                 return 1;
 260         READ_LOCK_ATOMIC(inc)(count);
 261         return 0;
 262 }
 263
 264 static inline int arch_write_trylock(arch_rwlock_t *lock)
 265 {
 266         atomic_t *count = (atomic_t *)&lock->write;
 267
 268         if (atomic_sub_and_test(WRITE_LOCK_CMP, count))
 269                 return 1;
 270         atomic_add(WRITE_LOCK_CMP, count);
 271         return 0;
 272 }
 273
 274 static inline void arch_read_unlock(arch_rwlock_t *rw)
 275 {
 276         asm volatile(LOCK_PREFIX READ_LOCK_SIZE(inc) " %0"
 277                      :"+m" (rw->lock) : : "memory");
 278 }
 279
 280 static inline void arch_write_unlock(arch_rwlock_t *rw)
 281 {
 282         asm volatile(LOCK_PREFIX WRITE_LOCK_ADD(%1) "%0"
 283                      : "+m" (rw->write) : "i" (RW_LOCK_BIAS) : "memory");
 284 }
 285
 286 #define arch_read_lock_flags(lock, flags) arch_read_lock(lock)
 287 #define arch_write_lock_flags(lock, flags) arch_write_lock(lock)
 288
 289 #undef READ_LOCK_SIZE
 290 #undef READ_LOCK_ATOMIC
 291 #undef WRITE_LOCK_ADD
 292 #undef WRITE_LOCK_SUB
 293 #undef WRITE_LOCK_CMP
 294
 295 #define arch_spin_relax(lock)   cpu_relax()
 296 #define arch_read_relax(lock)   cpu_relax()
 297 #define arch_write_relax(lock)  cpu_relax()
 298
 299 #endif /* _ASM_X86_SPINLOCK_H */