]> git.proxmox.com Git - mirror_qemu.git/blob - include/qemu/atomic.h
machine: Refactor smp-related call chains to pass MachineState
[mirror_qemu.git] / include / qemu / atomic.h
1 /*
2 * Simple interface for atomic operations.
3 *
4 * Copyright (C) 2013 Red Hat, Inc.
5 *
6 * Author: Paolo Bonzini <pbonzini@redhat.com>
7 *
8 * This work is licensed under the terms of the GNU GPL, version 2 or later.
9 * See the COPYING file in the top-level directory.
10 *
11 * See docs/devel/atomics.txt for discussion about the guarantees each
12 * atomic primitive is meant to provide.
13 */
14
15 #ifndef QEMU_ATOMIC_H
16 #define QEMU_ATOMIC_H
17
18 /* Compiler barrier */
19 #define barrier() ({ asm volatile("" ::: "memory"); (void)0; })
20
21 /* The variable that receives the old value of an atomically-accessed
22 * variable must be non-qualified, because atomic builtins return values
23 * through a pointer-type argument as in __atomic_load(&var, &old, MODEL).
24 *
25 * This macro has to handle types smaller than int manually, because of
26 * implicit promotion. int and larger types, as well as pointers, can be
27 * converted to a non-qualified type just by applying a binary operator.
28 */
29 #define typeof_strip_qual(expr) \
30 typeof( \
31 __builtin_choose_expr( \
32 __builtin_types_compatible_p(typeof(expr), bool) || \
33 __builtin_types_compatible_p(typeof(expr), const bool) || \
34 __builtin_types_compatible_p(typeof(expr), volatile bool) || \
35 __builtin_types_compatible_p(typeof(expr), const volatile bool), \
36 (bool)1, \
37 __builtin_choose_expr( \
38 __builtin_types_compatible_p(typeof(expr), signed char) || \
39 __builtin_types_compatible_p(typeof(expr), const signed char) || \
40 __builtin_types_compatible_p(typeof(expr), volatile signed char) || \
41 __builtin_types_compatible_p(typeof(expr), const volatile signed char), \
42 (signed char)1, \
43 __builtin_choose_expr( \
44 __builtin_types_compatible_p(typeof(expr), unsigned char) || \
45 __builtin_types_compatible_p(typeof(expr), const unsigned char) || \
46 __builtin_types_compatible_p(typeof(expr), volatile unsigned char) || \
47 __builtin_types_compatible_p(typeof(expr), const volatile unsigned char), \
48 (unsigned char)1, \
49 __builtin_choose_expr( \
50 __builtin_types_compatible_p(typeof(expr), signed short) || \
51 __builtin_types_compatible_p(typeof(expr), const signed short) || \
52 __builtin_types_compatible_p(typeof(expr), volatile signed short) || \
53 __builtin_types_compatible_p(typeof(expr), const volatile signed short), \
54 (signed short)1, \
55 __builtin_choose_expr( \
56 __builtin_types_compatible_p(typeof(expr), unsigned short) || \
57 __builtin_types_compatible_p(typeof(expr), const unsigned short) || \
58 __builtin_types_compatible_p(typeof(expr), volatile unsigned short) || \
59 __builtin_types_compatible_p(typeof(expr), const volatile unsigned short), \
60 (unsigned short)1, \
61 (expr)+0))))))
62
63 #ifdef __ATOMIC_RELAXED
64 /* For C11 atomic ops */
65
66 /* Manual memory barriers
67 *
68 *__atomic_thread_fence does not include a compiler barrier; instead,
69 * the barrier is part of __atomic_load/__atomic_store's "volatile-like"
70 * semantics. If smp_wmb() is a no-op, absence of the barrier means that
71 * the compiler is free to reorder stores on each side of the barrier.
72 * Add one here, and similarly in smp_rmb() and smp_read_barrier_depends().
73 */
74
75 #define smp_mb() ({ barrier(); __atomic_thread_fence(__ATOMIC_SEQ_CST); })
76 #define smp_mb_release() ({ barrier(); __atomic_thread_fence(__ATOMIC_RELEASE); })
77 #define smp_mb_acquire() ({ barrier(); __atomic_thread_fence(__ATOMIC_ACQUIRE); })
78
79 /* Most compilers currently treat consume and acquire the same, but really
80 * no processors except Alpha need a barrier here. Leave it in if
81 * using Thread Sanitizer to avoid warnings, otherwise optimize it away.
82 */
83 #if defined(__SANITIZE_THREAD__)
84 #define smp_read_barrier_depends() ({ barrier(); __atomic_thread_fence(__ATOMIC_CONSUME); })
85 #elif defined(__alpha__)
86 #define smp_read_barrier_depends() asm volatile("mb":::"memory")
87 #else
88 #define smp_read_barrier_depends() barrier()
89 #endif
90
91 /* Sanity check that the size of an atomic operation isn't "overly large".
92 * Despite the fact that e.g. i686 has 64-bit atomic operations, we do not
93 * want to use them because we ought not need them, and this lets us do a
94 * bit of sanity checking that other 32-bit hosts might build.
95 *
96 * That said, we have a problem on 64-bit ILP32 hosts in that in order to
97 * sync with TCG_OVERSIZED_GUEST, this must match TCG_TARGET_REG_BITS.
98 * We'd prefer not want to pull in everything else TCG related, so handle
99 * those few cases by hand.
100 *
101 * Note that x32 is fully detected with __x86_64__ + _ILP32, and that for
102 * Sparc we always force the use of sparcv9 in configure. MIPS n32 (ILP32) &
103 * n64 (LP64) ABIs are both detected using __mips64.
104 */
105 #if defined(__x86_64__) || defined(__sparc__) || defined(__mips64)
106 # define ATOMIC_REG_SIZE 8
107 #else
108 # define ATOMIC_REG_SIZE sizeof(void *)
109 #endif
110
111 /* Weak atomic operations prevent the compiler moving other
112 * loads/stores past the atomic operation load/store. However there is
113 * no explicit memory barrier for the processor.
114 *
115 * The C11 memory model says that variables that are accessed from
116 * different threads should at least be done with __ATOMIC_RELAXED
117 * primitives or the result is undefined. Generally this has little to
118 * no effect on the generated code but not using the atomic primitives
119 * will get flagged by sanitizers as a violation.
120 */
121 #define atomic_read__nocheck(ptr) \
122 __atomic_load_n(ptr, __ATOMIC_RELAXED)
123
124 #define atomic_read(ptr) \
125 ({ \
126 QEMU_BUILD_BUG_ON(sizeof(*ptr) > ATOMIC_REG_SIZE); \
127 atomic_read__nocheck(ptr); \
128 })
129
130 #define atomic_set__nocheck(ptr, i) \
131 __atomic_store_n(ptr, i, __ATOMIC_RELAXED)
132
133 #define atomic_set(ptr, i) do { \
134 QEMU_BUILD_BUG_ON(sizeof(*ptr) > ATOMIC_REG_SIZE); \
135 atomic_set__nocheck(ptr, i); \
136 } while(0)
137
138 /* See above: most compilers currently treat consume and acquire the
139 * same, but this slows down atomic_rcu_read unnecessarily.
140 */
141 #ifdef __SANITIZE_THREAD__
142 #define atomic_rcu_read__nocheck(ptr, valptr) \
143 __atomic_load(ptr, valptr, __ATOMIC_CONSUME);
144 #else
145 #define atomic_rcu_read__nocheck(ptr, valptr) \
146 __atomic_load(ptr, valptr, __ATOMIC_RELAXED); \
147 smp_read_barrier_depends();
148 #endif
149
150 #define atomic_rcu_read(ptr) \
151 ({ \
152 QEMU_BUILD_BUG_ON(sizeof(*ptr) > ATOMIC_REG_SIZE); \
153 typeof_strip_qual(*ptr) _val; \
154 atomic_rcu_read__nocheck(ptr, &_val); \
155 _val; \
156 })
157
158 #define atomic_rcu_set(ptr, i) do { \
159 QEMU_BUILD_BUG_ON(sizeof(*ptr) > ATOMIC_REG_SIZE); \
160 __atomic_store_n(ptr, i, __ATOMIC_RELEASE); \
161 } while(0)
162
163 #define atomic_load_acquire(ptr) \
164 ({ \
165 QEMU_BUILD_BUG_ON(sizeof(*ptr) > ATOMIC_REG_SIZE); \
166 typeof_strip_qual(*ptr) _val; \
167 __atomic_load(ptr, &_val, __ATOMIC_ACQUIRE); \
168 _val; \
169 })
170
171 #define atomic_store_release(ptr, i) do { \
172 QEMU_BUILD_BUG_ON(sizeof(*ptr) > ATOMIC_REG_SIZE); \
173 __atomic_store_n(ptr, i, __ATOMIC_RELEASE); \
174 } while(0)
175
176
177 /* All the remaining operations are fully sequentially consistent */
178
179 #define atomic_xchg__nocheck(ptr, i) ({ \
180 __atomic_exchange_n(ptr, (i), __ATOMIC_SEQ_CST); \
181 })
182
183 #define atomic_xchg(ptr, i) ({ \
184 QEMU_BUILD_BUG_ON(sizeof(*ptr) > ATOMIC_REG_SIZE); \
185 atomic_xchg__nocheck(ptr, i); \
186 })
187
188 /* Returns the eventual value, failed or not */
189 #define atomic_cmpxchg__nocheck(ptr, old, new) ({ \
190 typeof_strip_qual(*ptr) _old = (old); \
191 (void)__atomic_compare_exchange_n(ptr, &_old, new, false, \
192 __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST); \
193 _old; \
194 })
195
196 #define atomic_cmpxchg(ptr, old, new) ({ \
197 QEMU_BUILD_BUG_ON(sizeof(*ptr) > ATOMIC_REG_SIZE); \
198 atomic_cmpxchg__nocheck(ptr, old, new); \
199 })
200
201 /* Provide shorter names for GCC atomic builtins, return old value */
202 #define atomic_fetch_inc(ptr) __atomic_fetch_add(ptr, 1, __ATOMIC_SEQ_CST)
203 #define atomic_fetch_dec(ptr) __atomic_fetch_sub(ptr, 1, __ATOMIC_SEQ_CST)
204 #define atomic_fetch_add(ptr, n) __atomic_fetch_add(ptr, n, __ATOMIC_SEQ_CST)
205 #define atomic_fetch_sub(ptr, n) __atomic_fetch_sub(ptr, n, __ATOMIC_SEQ_CST)
206 #define atomic_fetch_and(ptr, n) __atomic_fetch_and(ptr, n, __ATOMIC_SEQ_CST)
207 #define atomic_fetch_or(ptr, n) __atomic_fetch_or(ptr, n, __ATOMIC_SEQ_CST)
208 #define atomic_fetch_xor(ptr, n) __atomic_fetch_xor(ptr, n, __ATOMIC_SEQ_CST)
209
210 #define atomic_inc_fetch(ptr) __atomic_add_fetch(ptr, 1, __ATOMIC_SEQ_CST)
211 #define atomic_dec_fetch(ptr) __atomic_sub_fetch(ptr, 1, __ATOMIC_SEQ_CST)
212 #define atomic_add_fetch(ptr, n) __atomic_add_fetch(ptr, n, __ATOMIC_SEQ_CST)
213 #define atomic_sub_fetch(ptr, n) __atomic_sub_fetch(ptr, n, __ATOMIC_SEQ_CST)
214 #define atomic_and_fetch(ptr, n) __atomic_and_fetch(ptr, n, __ATOMIC_SEQ_CST)
215 #define atomic_or_fetch(ptr, n) __atomic_or_fetch(ptr, n, __ATOMIC_SEQ_CST)
216 #define atomic_xor_fetch(ptr, n) __atomic_xor_fetch(ptr, n, __ATOMIC_SEQ_CST)
217
218 /* And even shorter names that return void. */
219 #define atomic_inc(ptr) ((void) __atomic_fetch_add(ptr, 1, __ATOMIC_SEQ_CST))
220 #define atomic_dec(ptr) ((void) __atomic_fetch_sub(ptr, 1, __ATOMIC_SEQ_CST))
221 #define atomic_add(ptr, n) ((void) __atomic_fetch_add(ptr, n, __ATOMIC_SEQ_CST))
222 #define atomic_sub(ptr, n) ((void) __atomic_fetch_sub(ptr, n, __ATOMIC_SEQ_CST))
223 #define atomic_and(ptr, n) ((void) __atomic_fetch_and(ptr, n, __ATOMIC_SEQ_CST))
224 #define atomic_or(ptr, n) ((void) __atomic_fetch_or(ptr, n, __ATOMIC_SEQ_CST))
225 #define atomic_xor(ptr, n) ((void) __atomic_fetch_xor(ptr, n, __ATOMIC_SEQ_CST))
226
227 #else /* __ATOMIC_RELAXED */
228
229 /*
230 * We use GCC builtin if it's available, as that can use mfence on
231 * 32-bit as well, e.g. if built with -march=pentium-m. However, on
232 * i386 the spec is buggy, and the implementation followed it until
233 * 4.3 (http://gcc.gnu.org/bugzilla/show_bug.cgi?id=36793).
234 */
235 #if defined(__i386__) || defined(__x86_64__)
236 #if !QEMU_GNUC_PREREQ(4, 4)
237 #if defined __x86_64__
238 #define smp_mb() ({ asm volatile("mfence" ::: "memory"); (void)0; })
239 #else
240 #define smp_mb() ({ asm volatile("lock; addl $0,0(%%esp) " ::: "memory"); (void)0; })
241 #endif
242 #endif
243 #endif
244
245
246 #ifdef __alpha__
247 #define smp_read_barrier_depends() asm volatile("mb":::"memory")
248 #endif
249
250 #if defined(__i386__) || defined(__x86_64__) || defined(__s390x__)
251
252 /*
253 * Because of the strongly ordered storage model, wmb() and rmb() are nops
254 * here (a compiler barrier only). QEMU doesn't do accesses to write-combining
255 * qemu memory or non-temporal load/stores from C code.
256 */
257 #define smp_mb_release() barrier()
258 #define smp_mb_acquire() barrier()
259
260 /*
261 * __sync_lock_test_and_set() is documented to be an acquire barrier only,
262 * but it is a full barrier at the hardware level. Add a compiler barrier
263 * to make it a full barrier also at the compiler level.
264 */
265 #define atomic_xchg(ptr, i) (barrier(), __sync_lock_test_and_set(ptr, i))
266
267 #elif defined(_ARCH_PPC)
268
269 /*
270 * We use an eieio() for wmb() on powerpc. This assumes we don't
271 * need to order cacheable and non-cacheable stores with respect to
272 * each other.
273 *
274 * smp_mb has the same problem as on x86 for not-very-new GCC
275 * (http://patchwork.ozlabs.org/patch/126184/, Nov 2011).
276 */
277 #define smp_wmb() ({ asm volatile("eieio" ::: "memory"); (void)0; })
278 #if defined(__powerpc64__)
279 #define smp_mb_release() ({ asm volatile("lwsync" ::: "memory"); (void)0; })
280 #define smp_mb_acquire() ({ asm volatile("lwsync" ::: "memory"); (void)0; })
281 #else
282 #define smp_mb_release() ({ asm volatile("sync" ::: "memory"); (void)0; })
283 #define smp_mb_acquire() ({ asm volatile("sync" ::: "memory"); (void)0; })
284 #endif
285 #define smp_mb() ({ asm volatile("sync" ::: "memory"); (void)0; })
286
287 #endif /* _ARCH_PPC */
288
289 /*
290 * For (host) platforms we don't have explicit barrier definitions
291 * for, we use the gcc __sync_synchronize() primitive to generate a
292 * full barrier. This should be safe on all platforms, though it may
293 * be overkill for smp_mb_acquire() and smp_mb_release().
294 */
295 #ifndef smp_mb
296 #define smp_mb() __sync_synchronize()
297 #endif
298
299 #ifndef smp_mb_acquire
300 #define smp_mb_acquire() __sync_synchronize()
301 #endif
302
303 #ifndef smp_mb_release
304 #define smp_mb_release() __sync_synchronize()
305 #endif
306
307 #ifndef smp_read_barrier_depends
308 #define smp_read_barrier_depends() barrier()
309 #endif
310
311 /* These will only be atomic if the processor does the fetch or store
312 * in a single issue memory operation
313 */
314 #define atomic_read__nocheck(p) (*(__typeof__(*(p)) volatile*) (p))
315 #define atomic_set__nocheck(p, i) ((*(__typeof__(*(p)) volatile*) (p)) = (i))
316
317 #define atomic_read(ptr) atomic_read__nocheck(ptr)
318 #define atomic_set(ptr, i) atomic_set__nocheck(ptr,i)
319
320 /**
321 * atomic_rcu_read - reads a RCU-protected pointer to a local variable
322 * into a RCU read-side critical section. The pointer can later be safely
323 * dereferenced within the critical section.
324 *
325 * This ensures that the pointer copy is invariant thorough the whole critical
326 * section.
327 *
328 * Inserts memory barriers on architectures that require them (currently only
329 * Alpha) and documents which pointers are protected by RCU.
330 *
331 * atomic_rcu_read also includes a compiler barrier to ensure that
332 * value-speculative optimizations (e.g. VSS: Value Speculation
333 * Scheduling) does not perform the data read before the pointer read
334 * by speculating the value of the pointer.
335 *
336 * Should match atomic_rcu_set(), atomic_xchg(), atomic_cmpxchg().
337 */
338 #define atomic_rcu_read(ptr) ({ \
339 typeof(*ptr) _val = atomic_read(ptr); \
340 smp_read_barrier_depends(); \
341 _val; \
342 })
343
344 /**
345 * atomic_rcu_set - assigns (publicizes) a pointer to a new data structure
346 * meant to be read by RCU read-side critical sections.
347 *
348 * Documents which pointers will be dereferenced by RCU read-side critical
349 * sections and adds the required memory barriers on architectures requiring
350 * them. It also makes sure the compiler does not reorder code initializing the
351 * data structure before its publication.
352 *
353 * Should match atomic_rcu_read().
354 */
355 #define atomic_rcu_set(ptr, i) do { \
356 smp_wmb(); \
357 atomic_set(ptr, i); \
358 } while (0)
359
360 #define atomic_load_acquire(ptr) ({ \
361 typeof(*ptr) _val = atomic_read(ptr); \
362 smp_mb_acquire(); \
363 _val; \
364 })
365
366 #define atomic_store_release(ptr, i) do { \
367 smp_mb_release(); \
368 atomic_set(ptr, i); \
369 } while (0)
370
371 #ifndef atomic_xchg
372 #if defined(__clang__)
373 #define atomic_xchg(ptr, i) __sync_swap(ptr, i)
374 #else
375 /* __sync_lock_test_and_set() is documented to be an acquire barrier only. */
376 #define atomic_xchg(ptr, i) (smp_mb(), __sync_lock_test_and_set(ptr, i))
377 #endif
378 #endif
379 #define atomic_xchg__nocheck atomic_xchg
380
381 /* Provide shorter names for GCC atomic builtins. */
382 #define atomic_fetch_inc(ptr) __sync_fetch_and_add(ptr, 1)
383 #define atomic_fetch_dec(ptr) __sync_fetch_and_add(ptr, -1)
384 #define atomic_fetch_add(ptr, n) __sync_fetch_and_add(ptr, n)
385 #define atomic_fetch_sub(ptr, n) __sync_fetch_and_sub(ptr, n)
386 #define atomic_fetch_and(ptr, n) __sync_fetch_and_and(ptr, n)
387 #define atomic_fetch_or(ptr, n) __sync_fetch_and_or(ptr, n)
388 #define atomic_fetch_xor(ptr, n) __sync_fetch_and_xor(ptr, n)
389
390 #define atomic_inc_fetch(ptr) __sync_add_and_fetch(ptr, 1)
391 #define atomic_dec_fetch(ptr) __sync_add_and_fetch(ptr, -1)
392 #define atomic_add_fetch(ptr, n) __sync_add_and_fetch(ptr, n)
393 #define atomic_sub_fetch(ptr, n) __sync_sub_and_fetch(ptr, n)
394 #define atomic_and_fetch(ptr, n) __sync_and_and_fetch(ptr, n)
395 #define atomic_or_fetch(ptr, n) __sync_or_and_fetch(ptr, n)
396 #define atomic_xor_fetch(ptr, n) __sync_xor_and_fetch(ptr, n)
397
398 #define atomic_cmpxchg(ptr, old, new) __sync_val_compare_and_swap(ptr, old, new)
399 #define atomic_cmpxchg__nocheck(ptr, old, new) atomic_cmpxchg(ptr, old, new)
400
401 /* And even shorter names that return void. */
402 #define atomic_inc(ptr) ((void) __sync_fetch_and_add(ptr, 1))
403 #define atomic_dec(ptr) ((void) __sync_fetch_and_add(ptr, -1))
404 #define atomic_add(ptr, n) ((void) __sync_fetch_and_add(ptr, n))
405 #define atomic_sub(ptr, n) ((void) __sync_fetch_and_sub(ptr, n))
406 #define atomic_and(ptr, n) ((void) __sync_fetch_and_and(ptr, n))
407 #define atomic_or(ptr, n) ((void) __sync_fetch_and_or(ptr, n))
408 #define atomic_xor(ptr, n) ((void) __sync_fetch_and_xor(ptr, n))
409
410 #endif /* __ATOMIC_RELAXED */
411
412 #ifndef smp_wmb
413 #define smp_wmb() smp_mb_release()
414 #endif
415 #ifndef smp_rmb
416 #define smp_rmb() smp_mb_acquire()
417 #endif
418
419 /* This is more efficient than a store plus a fence. */
420 #if !defined(__SANITIZE_THREAD__)
421 #if defined(__i386__) || defined(__x86_64__) || defined(__s390x__)
422 #define atomic_mb_set(ptr, i) ((void)atomic_xchg(ptr, i))
423 #endif
424 #endif
425
426 /* atomic_mb_read/set semantics map Java volatile variables. They are
427 * less expensive on some platforms (notably POWER) than fully
428 * sequentially consistent operations.
429 *
430 * As long as they are used as paired operations they are safe to
431 * use. See docs/devel/atomics.txt for more discussion.
432 */
433
434 #ifndef atomic_mb_read
435 #define atomic_mb_read(ptr) \
436 atomic_load_acquire(ptr)
437 #endif
438
439 #ifndef atomic_mb_set
440 #define atomic_mb_set(ptr, i) do { \
441 atomic_store_release(ptr, i); \
442 smp_mb(); \
443 } while(0)
444 #endif
445
446 #define atomic_fetch_inc_nonzero(ptr) ({ \
447 typeof_strip_qual(*ptr) _oldn = atomic_read(ptr); \
448 while (_oldn && atomic_cmpxchg(ptr, _oldn, _oldn + 1) != _oldn) { \
449 _oldn = atomic_read(ptr); \
450 } \
451 _oldn; \
452 })
453
454 /* Abstractions to access atomically (i.e. "once") i64/u64 variables */
455 #ifdef CONFIG_ATOMIC64
456 static inline int64_t atomic_read_i64(const int64_t *ptr)
457 {
458 /* use __nocheck because sizeof(void *) might be < sizeof(u64) */
459 return atomic_read__nocheck(ptr);
460 }
461
462 static inline uint64_t atomic_read_u64(const uint64_t *ptr)
463 {
464 return atomic_read__nocheck(ptr);
465 }
466
467 static inline void atomic_set_i64(int64_t *ptr, int64_t val)
468 {
469 atomic_set__nocheck(ptr, val);
470 }
471
472 static inline void atomic_set_u64(uint64_t *ptr, uint64_t val)
473 {
474 atomic_set__nocheck(ptr, val);
475 }
476
477 static inline void atomic64_init(void)
478 {
479 }
480 #else /* !CONFIG_ATOMIC64 */
481 int64_t atomic_read_i64(const int64_t *ptr);
482 uint64_t atomic_read_u64(const uint64_t *ptr);
483 void atomic_set_i64(int64_t *ptr, int64_t val);
484 void atomic_set_u64(uint64_t *ptr, uint64_t val);
485 void atomic64_init(void);
486 #endif /* !CONFIG_ATOMIC64 */
487
488 #endif /* QEMU_ATOMIC_H */