blob: d17c91981da2f1c94a5fb045067f821ece61b4fe [file] [log] [blame]
H. Peter Anvin1965aae2008-10-22 22:26:29 -07001#ifndef _ASM_X86_SPINLOCK_H
2#define _ASM_X86_SPINLOCK_H
Glauber de Oliveira Costa2fed0c52008-01-30 13:30:33 +01003
Thomas Gleixner1075cf72008-01-30 13:30:34 +01004#include <asm/atomic.h>
5#include <asm/rwlock.h>
6#include <asm/page.h>
7#include <asm/processor.h>
Nick Piggin314cdbe2008-01-30 13:31:21 +01008#include <linux/compiler.h>
Jeremy Fitzhardinge74d4aff2008-07-07 12:07:50 -07009#include <asm/paravirt.h>
Thomas Gleixner1075cf72008-01-30 13:30:34 +010010/*
11 * Your basic SMP spinlocks, allowing only a single CPU anywhere
12 *
13 * Simple spin lock operations. There are two variants, one clears IRQ's
14 * on the local processor, one does not.
15 *
Nick Piggin314cdbe2008-01-30 13:31:21 +010016 * These are fair FIFO ticket locks, which are currently limited to 256
17 * CPUs.
Thomas Gleixner1075cf72008-01-30 13:30:34 +010018 *
19 * (the type definitions are in asm/spinlock_types.h)
20 */
21
Thomas Gleixner96a388d2007-10-11 11:20:03 +020022#ifdef CONFIG_X86_32
Thomas Gleixner1075cf72008-01-30 13:30:34 +010023# define LOCK_PTR_REG "a"
Jan Beulich74e91602008-09-05 13:27:45 +010024# define REG_PTR_MODE "k"
Thomas Gleixner96a388d2007-10-11 11:20:03 +020025#else
Thomas Gleixner1075cf72008-01-30 13:30:34 +010026# define LOCK_PTR_REG "D"
Jan Beulich74e91602008-09-05 13:27:45 +010027# define REG_PTR_MODE "q"
Thomas Gleixner96a388d2007-10-11 11:20:03 +020028#endif
Glauber de Oliveira Costa2fed0c52008-01-30 13:30:33 +010029
Nick Piggin3a556b22008-01-30 13:33:00 +010030#if defined(CONFIG_X86_32) && \
31 (defined(CONFIG_X86_OOSTORE) || defined(CONFIG_X86_PPRO_FENCE))
32/*
33 * On PPro SMP or if we are using OOSTORE, we use a locked operation to unlock
34 * (PPro errata 66, 92)
35 */
36# define UNLOCK_LOCK_PREFIX LOCK_PREFIX
37#else
38# define UNLOCK_LOCK_PREFIX
Nick Piggin314cdbe2008-01-30 13:31:21 +010039#endif
40
Nick Piggin3a556b22008-01-30 13:33:00 +010041/*
42 * Ticket locks are conceptually two parts, one indicating the current head of
43 * the queue, and the other indicating the current tail. The lock is acquired
44 * by atomically noting the tail and incrementing it by one (thus adding
45 * ourself to the queue and noting our position), then waiting until the head
46 * becomes equal to the the initial value of the tail.
47 *
48 * We use an xadd covering *both* parts of the lock, to increment the tail and
49 * also load the position of the head, which takes care of memory ordering
50 * issues and should be optimal for the uncontended case. Note the tail must be
51 * in the high part, because a wide xadd increment of the low part would carry
52 * up and contaminate the high part.
53 *
54 * With fewer than 2^8 possible CPUs, we can use x86's partial registers to
55 * save some instructions and make the code more elegant. There really isn't
56 * much between them in performance though, especially as locks are out of line.
57 */
58#if (NR_CPUS < 256)
Jan Beulich08f5fcb2008-09-05 13:26:39 +010059#define TICKET_SHIFT 8
Thomas Gleixner1075cf72008-01-30 13:30:34 +010060
Jeremy Fitzhardinge74d4aff2008-07-07 12:07:50 -070061static __always_inline void __ticket_spin_lock(raw_spinlock_t *lock)
Thomas Gleixner1075cf72008-01-30 13:30:34 +010062{
Nick Piggin314cdbe2008-01-30 13:31:21 +010063 short inc = 0x0100;
64
Joe Perchesd3bf60a2008-03-23 01:03:31 -070065 asm volatile (
Nick Piggin314cdbe2008-01-30 13:31:21 +010066 LOCK_PREFIX "xaddw %w0, %1\n"
67 "1:\t"
68 "cmpb %h0, %b0\n\t"
69 "je 2f\n\t"
70 "rep ; nop\n\t"
71 "movb %1, %b0\n\t"
72 /* don't need lfence here, because loads are in-order */
Thomas Gleixner1075cf72008-01-30 13:30:34 +010073 "jmp 1b\n"
Nick Piggin314cdbe2008-01-30 13:31:21 +010074 "2:"
Joe Perchesd3bf60a2008-03-23 01:03:31 -070075 : "+Q" (inc), "+m" (lock->slock)
Nick Piggin314cdbe2008-01-30 13:31:21 +010076 :
Joe Perchesd3bf60a2008-03-23 01:03:31 -070077 : "memory", "cc");
Thomas Gleixner1075cf72008-01-30 13:30:34 +010078}
79
Jeremy Fitzhardinge74d4aff2008-07-07 12:07:50 -070080static __always_inline int __ticket_spin_trylock(raw_spinlock_t *lock)
Thomas Gleixner1075cf72008-01-30 13:30:34 +010081{
Jan Beulich74e91602008-09-05 13:27:45 +010082 int tmp, new;
Thomas Gleixner1075cf72008-01-30 13:30:34 +010083
Jan Beulich74e91602008-09-05 13:27:45 +010084 asm volatile("movzwl %2, %0\n\t"
Joe Perchesd3bf60a2008-03-23 01:03:31 -070085 "cmpb %h0,%b0\n\t"
Jan Beulich74e91602008-09-05 13:27:45 +010086 "leal 0x100(%" REG_PTR_MODE "0), %1\n\t"
Joe Perchesd3bf60a2008-03-23 01:03:31 -070087 "jne 1f\n\t"
Mathieu Desnoyers5bbd4c32008-08-15 12:56:59 -040088 LOCK_PREFIX "cmpxchgw %w1,%2\n\t"
Joe Perchesd3bf60a2008-03-23 01:03:31 -070089 "1:"
90 "sete %b1\n\t"
91 "movzbl %b1,%0\n\t"
Jan Beulich74e91602008-09-05 13:27:45 +010092 : "=&a" (tmp), "=&q" (new), "+m" (lock->slock)
Joe Perchesd3bf60a2008-03-23 01:03:31 -070093 :
94 : "memory", "cc");
Thomas Gleixner1075cf72008-01-30 13:30:34 +010095
Nick Piggin314cdbe2008-01-30 13:31:21 +010096 return tmp;
Thomas Gleixner1075cf72008-01-30 13:30:34 +010097}
98
Jeremy Fitzhardinge74d4aff2008-07-07 12:07:50 -070099static __always_inline void __ticket_spin_unlock(raw_spinlock_t *lock)
Thomas Gleixner1075cf72008-01-30 13:30:34 +0100100{
Joe Perchesd3bf60a2008-03-23 01:03:31 -0700101 asm volatile(UNLOCK_LOCK_PREFIX "incb %0"
102 : "+m" (lock->slock)
103 :
104 : "memory", "cc");
Thomas Gleixner1075cf72008-01-30 13:30:34 +0100105}
Nick Piggin3a556b22008-01-30 13:33:00 +0100106#else
Jan Beulich08f5fcb2008-09-05 13:26:39 +0100107#define TICKET_SHIFT 16
Nick Piggin3a556b22008-01-30 13:33:00 +0100108
Jeremy Fitzhardinge74d4aff2008-07-07 12:07:50 -0700109static __always_inline void __ticket_spin_lock(raw_spinlock_t *lock)
Nick Piggin3a556b22008-01-30 13:33:00 +0100110{
111 int inc = 0x00010000;
112 int tmp;
113
Mathieu Desnoyers5bbd4c32008-08-15 12:56:59 -0400114 asm volatile(LOCK_PREFIX "xaddl %0, %1\n"
Joe Perchesd3bf60a2008-03-23 01:03:31 -0700115 "movzwl %w0, %2\n\t"
116 "shrl $16, %0\n\t"
117 "1:\t"
118 "cmpl %0, %2\n\t"
119 "je 2f\n\t"
120 "rep ; nop\n\t"
121 "movzwl %1, %2\n\t"
122 /* don't need lfence here, because loads are in-order */
123 "jmp 1b\n"
124 "2:"
Jan Beulichef1f3412008-09-05 13:26:39 +0100125 : "+r" (inc), "+m" (lock->slock), "=&r" (tmp)
Joe Perchesd3bf60a2008-03-23 01:03:31 -0700126 :
127 : "memory", "cc");
Nick Piggin3a556b22008-01-30 13:33:00 +0100128}
129
Jeremy Fitzhardinge74d4aff2008-07-07 12:07:50 -0700130static __always_inline int __ticket_spin_trylock(raw_spinlock_t *lock)
Nick Piggin3a556b22008-01-30 13:33:00 +0100131{
132 int tmp;
133 int new;
134
Joe Perchesd3bf60a2008-03-23 01:03:31 -0700135 asm volatile("movl %2,%0\n\t"
136 "movl %0,%1\n\t"
137 "roll $16, %0\n\t"
138 "cmpl %0,%1\n\t"
Jan Beulich74e91602008-09-05 13:27:45 +0100139 "leal 0x00010000(%" REG_PTR_MODE "0), %1\n\t"
Joe Perchesd3bf60a2008-03-23 01:03:31 -0700140 "jne 1f\n\t"
Mathieu Desnoyers5bbd4c32008-08-15 12:56:59 -0400141 LOCK_PREFIX "cmpxchgl %1,%2\n\t"
Joe Perchesd3bf60a2008-03-23 01:03:31 -0700142 "1:"
143 "sete %b1\n\t"
144 "movzbl %b1,%0\n\t"
Jan Beulichef1f3412008-09-05 13:26:39 +0100145 : "=&a" (tmp), "=&q" (new), "+m" (lock->slock)
Joe Perchesd3bf60a2008-03-23 01:03:31 -0700146 :
147 : "memory", "cc");
Nick Piggin3a556b22008-01-30 13:33:00 +0100148
149 return tmp;
150}
151
Jeremy Fitzhardinge74d4aff2008-07-07 12:07:50 -0700152static __always_inline void __ticket_spin_unlock(raw_spinlock_t *lock)
Nick Piggin3a556b22008-01-30 13:33:00 +0100153{
Joe Perchesd3bf60a2008-03-23 01:03:31 -0700154 asm volatile(UNLOCK_LOCK_PREFIX "incw %0"
155 : "+m" (lock->slock)
156 :
157 : "memory", "cc");
Nick Piggin3a556b22008-01-30 13:33:00 +0100158}
159#endif
Thomas Gleixner1075cf72008-01-30 13:30:34 +0100160
Jan Beulich08f5fcb2008-09-05 13:26:39 +0100161static inline int __ticket_spin_is_locked(raw_spinlock_t *lock)
162{
163 int tmp = ACCESS_ONCE(lock->slock);
164
165 return !!(((tmp >> TICKET_SHIFT) ^ tmp) & ((1 << TICKET_SHIFT) - 1));
166}
167
168static inline int __ticket_spin_is_contended(raw_spinlock_t *lock)
169{
170 int tmp = ACCESS_ONCE(lock->slock);
171
172 return (((tmp >> TICKET_SHIFT) - tmp) & ((1 << TICKET_SHIFT) - 1)) > 1;
173}
Jeremy Fitzhardinge74d4aff2008-07-07 12:07:50 -0700174
Jeremy Fitzhardinge8efcbab2008-07-07 12:07:51 -0700175#ifdef CONFIG_PARAVIRT
176/*
177 * Define virtualization-friendly old-style lock byte lock, for use in
178 * pv_lock_ops if desired.
179 *
180 * This differs from the pre-2.6.24 spinlock by always using xchgb
181 * rather than decb to take the lock; this allows it to use a
182 * zero-initialized lock structure. It also maintains a 1-byte
183 * contention counter, so that we can implement
184 * __byte_spin_is_contended.
185 */
186struct __byte_spinlock {
187 s8 lock;
188 s8 spinners;
189};
190
191static inline int __byte_spin_is_locked(raw_spinlock_t *lock)
192{
193 struct __byte_spinlock *bl = (struct __byte_spinlock *)lock;
194 return bl->lock != 0;
195}
196
197static inline int __byte_spin_is_contended(raw_spinlock_t *lock)
198{
199 struct __byte_spinlock *bl = (struct __byte_spinlock *)lock;
200 return bl->spinners != 0;
201}
202
203static inline void __byte_spin_lock(raw_spinlock_t *lock)
204{
205 struct __byte_spinlock *bl = (struct __byte_spinlock *)lock;
206 s8 val = 1;
207
208 asm("1: xchgb %1, %0\n"
209 " test %1,%1\n"
210 " jz 3f\n"
211 " " LOCK_PREFIX "incb %2\n"
212 "2: rep;nop\n"
213 " cmpb $1, %0\n"
214 " je 2b\n"
215 " " LOCK_PREFIX "decb %2\n"
216 " jmp 1b\n"
217 "3:"
218 : "+m" (bl->lock), "+q" (val), "+m" (bl->spinners): : "memory");
219}
220
221static inline int __byte_spin_trylock(raw_spinlock_t *lock)
222{
223 struct __byte_spinlock *bl = (struct __byte_spinlock *)lock;
224 u8 old = 1;
225
226 asm("xchgb %1,%0"
227 : "+m" (bl->lock), "+q" (old) : : "memory");
228
229 return old == 0;
230}
231
232static inline void __byte_spin_unlock(raw_spinlock_t *lock)
233{
234 struct __byte_spinlock *bl = (struct __byte_spinlock *)lock;
235 smp_wmb();
236 bl->lock = 0;
237}
238#else /* !CONFIG_PARAVIRT */
Jeremy Fitzhardinge74d4aff2008-07-07 12:07:50 -0700239static inline int __raw_spin_is_locked(raw_spinlock_t *lock)
240{
241 return __ticket_spin_is_locked(lock);
242}
243
244static inline int __raw_spin_is_contended(raw_spinlock_t *lock)
245{
246 return __ticket_spin_is_contended(lock);
247}
248
249static __always_inline void __raw_spin_lock(raw_spinlock_t *lock)
250{
251 __ticket_spin_lock(lock);
252}
253
254static __always_inline int __raw_spin_trylock(raw_spinlock_t *lock)
255{
256 return __ticket_spin_trylock(lock);
257}
258
259static __always_inline void __raw_spin_unlock(raw_spinlock_t *lock)
260{
261 __ticket_spin_unlock(lock);
262}
Jeremy Fitzhardinge63d3a752008-08-19 13:19:36 -0700263
264static __always_inline void __raw_spin_lock_flags(raw_spinlock_t *lock,
265 unsigned long flags)
266{
267 __raw_spin_lock(lock);
268}
269
Jeremy Fitzhardinge74d4aff2008-07-07 12:07:50 -0700270#endif /* CONFIG_PARAVIRT */
271
Thomas Gleixner1075cf72008-01-30 13:30:34 +0100272static inline void __raw_spin_unlock_wait(raw_spinlock_t *lock)
273{
274 while (__raw_spin_is_locked(lock))
275 cpu_relax();
276}
277
278/*
279 * Read-write spinlocks, allowing multiple readers
280 * but only one writer.
281 *
282 * NOTE! it is quite common to have readers in interrupts
283 * but no interrupt writers. For those circumstances we
284 * can "mix" irq-safe locks - any writer needs to get a
285 * irq-safe write-lock, but readers can get non-irqsafe
286 * read-locks.
287 *
288 * On x86, we implement read-write locks as a 32-bit counter
289 * with the high bit (sign) being the "contended" bit.
290 */
291
Nick Piggin314cdbe2008-01-30 13:31:21 +0100292/**
293 * read_can_lock - would read_trylock() succeed?
294 * @lock: the rwlock in question.
295 */
Thomas Gleixner1075cf72008-01-30 13:30:34 +0100296static inline int __raw_read_can_lock(raw_rwlock_t *lock)
297{
298 return (int)(lock)->lock > 0;
299}
300
Nick Piggin314cdbe2008-01-30 13:31:21 +0100301/**
302 * write_can_lock - would write_trylock() succeed?
303 * @lock: the rwlock in question.
304 */
Thomas Gleixner1075cf72008-01-30 13:30:34 +0100305static inline int __raw_write_can_lock(raw_rwlock_t *lock)
306{
307 return (lock)->lock == RW_LOCK_BIAS;
308}
309
310static inline void __raw_read_lock(raw_rwlock_t *rw)
311{
312 asm volatile(LOCK_PREFIX " subl $1,(%0)\n\t"
313 "jns 1f\n"
314 "call __read_lock_failed\n\t"
315 "1:\n"
316 ::LOCK_PTR_REG (rw) : "memory");
317}
318
319static inline void __raw_write_lock(raw_rwlock_t *rw)
320{
321 asm volatile(LOCK_PREFIX " subl %1,(%0)\n\t"
322 "jz 1f\n"
323 "call __write_lock_failed\n\t"
324 "1:\n"
325 ::LOCK_PTR_REG (rw), "i" (RW_LOCK_BIAS) : "memory");
326}
327
328static inline int __raw_read_trylock(raw_rwlock_t *lock)
329{
330 atomic_t *count = (atomic_t *)lock;
331
332 atomic_dec(count);
333 if (atomic_read(count) >= 0)
334 return 1;
335 atomic_inc(count);
336 return 0;
337}
338
339static inline int __raw_write_trylock(raw_rwlock_t *lock)
340{
341 atomic_t *count = (atomic_t *)lock;
342
343 if (atomic_sub_and_test(RW_LOCK_BIAS, count))
344 return 1;
345 atomic_add(RW_LOCK_BIAS, count);
346 return 0;
347}
348
349static inline void __raw_read_unlock(raw_rwlock_t *rw)
350{
351 asm volatile(LOCK_PREFIX "incl %0" :"+m" (rw->lock) : : "memory");
352}
353
354static inline void __raw_write_unlock(raw_rwlock_t *rw)
355{
356 asm volatile(LOCK_PREFIX "addl %1, %0"
357 : "+m" (rw->lock) : "i" (RW_LOCK_BIAS) : "memory");
358}
359
360#define _raw_spin_relax(lock) cpu_relax()
361#define _raw_read_relax(lock) cpu_relax()
362#define _raw_write_relax(lock) cpu_relax()
363
H. Peter Anvin1965aae2008-10-22 22:26:29 -0700364#endif /* _ASM_X86_SPINLOCK_H */