2 * Copyright (c) 2013, 2014 Nicira, Inc.
4 * Licensed under the Apache License, Version 2.0 (the "License");
5 * you may not use this file except in compliance with the License.
6 * You may obtain a copy of the License at:
8 * http://www.apache.org/licenses/LICENSE-2.0
10 * Unless required by applicable law or agreed to in writing, software
11 * distributed under the License is distributed on an "AS IS" BASIS,
12 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13 * See the License for the specific language governing permissions and
14 * limitations under the License.
18 #define OVS_ATOMIC_H 1
22 * This library implements atomic operations with an API based on the one
23 * defined in C11. It includes multiple implementations for compilers and
24 * libraries with varying degrees of built-in support for C11, including a
25 * fallback implementation for systems that have pthreads but no other support
28 * This comment describes the common features of all the implementations.
34 * The following atomic types are supported as typedefs for atomic versions of
35 * the listed ordinary types:
37 * ordinary type atomic version
38 * ------------------- ----------------------
42 * signed char atomic_schar
43 * unsigned char atomic_uchar
46 * unsigned short atomic_ushort
49 * unsigned int atomic_uint
52 * unsigned long atomic_ulong
54 * long long atomic_llong
55 * unsigned long long atomic_ullong
57 * size_t atomic_size_t
58 * ptrdiff_t atomic_ptrdiff_t
60 * intmax_t atomic_intmax_t
61 * uintmax_t atomic_uintmax_t
63 * intptr_t atomic_intptr_t
64 * uintptr_t atomic_uintptr_t
66 * uint8_t atomic_uint8_t (*)
67 * uint16_t atomic_uint16_t (*)
68 * uint32_t atomic_uint32_t (*)
69 * int8_t atomic_int8_t (*)
70 * int16_t atomic_int16_t (*)
71 * int32_t atomic_int32_t (*)
73 * (*) Not specified by C11.
75 * Atomic types may also be obtained via ATOMIC(TYPE), e.g. ATOMIC(void *).
76 * Only basic integer types and pointer types can be made atomic this way,
77 * e.g. atomic structs are not supported.
79 * The atomic version of a type doesn't necessarily have the same size or
80 * representation as the ordinary version; for example, atomic_int might be a
81 * typedef for a struct. The range of an atomic type does match the range of
82 * the corresponding ordinary type.
84 * C11 says that one may use the _Atomic keyword in place of the typedef name,
85 * e.g. "_Atomic int" instead of "atomic_int". This library doesn't support
92 * To initialize an atomic variable at its point of definition, use
95 * static atomic_int ai = ATOMIC_VAR_INIT(123);
97 * To initialize an atomic variable in code, use atomic_init():
99 * static atomic_int ai;
101 * atomic_init(&ai, 123);
107 * enum memory_order specifies the strictness of a memory barrier. It has the
110 * memory_order_relaxed:
112 * Only atomicity is provided, does not imply any memory ordering with
113 * respect to any other variable (atomic or not). Relaxed accesses to
114 * the same atomic variable will be performed in the program order.
115 * The compiler and CPU are free to move memory accesses to other
116 * variables past the atomic operation.
118 * memory_order_consume:
120 * Memory accesses with data dependency on the result of the consume
121 * operation (atomic_read_explicit, or a load operation preceding a
122 * atomic_thread_fence) will not be moved prior to the consume
123 * barrier. Non-data-dependent loads and stores can be reordered to
124 * happen before the the consume barrier.
126 * RCU is the prime example of the use of the consume barrier: The
127 * consume barrier guarantees that reads from a RCU protected object
128 * are performed after the RCU protected pointer is read. A
129 * corresponding release barrier is used to store the modified RCU
130 * protected pointer after the RCU protected object has been fully
131 * constructed. The synchronization between these barriers prevents
132 * the RCU "consumer" from seeing uninitialized data.
134 * May not be used with atomic_store_explicit(), as consume semantics
135 * applies only to atomic loads.
137 * memory_order_acquire:
139 * Memory accesses after an acquire barrier cannot be moved before the
140 * barrier. Memory accesses before an acquire barrier *can* be moved
143 * An atomic_thread_fence with memory_order_acquire does not have a
144 * load operation by itself; it prevents all following memory accesses
145 * from moving prior to preceding loads.
147 * May not be used with atomic_store_explicit(), as acquire semantics
148 * applies only to atomic loads.
150 * memory_order_release:
152 * Memory accesses before a release barrier cannot be moved after the
153 * barrier. Memory accesses after a release barrier *can* be moved
156 * An atomic_thread_fence with memory_order_release does not have a
157 * store operation by itself; it prevents all preceding memory accesses
158 * from moving past subsequent stores.
160 * May not be used with atomic_read_explicit(), as release semantics
161 * applies only to atomic stores.
163 * memory_order_acq_rel:
165 * Memory accesses cannot be moved across an acquire-release barrier in
168 * May only be used with atomic read-modify-write operations, as both
169 * load and store operation is required for acquire-release semantics.
171 * An atomic_thread_fence with memory_order_acq_rel does not have
172 * either load or store operation by itself; it prevents all following
173 * memory accesses from moving prior to preceding loads and all
174 * preceding memory accesses from moving past subsequent stores.
176 * memory_order_seq_cst:
178 * Prevents movement of memory accesses like an acquire-release barrier,
179 * but whereas acquire-release synchronizes cooperating threads (using
180 * the same atomic variable), sequential-consistency synchronizes the
181 * whole system, providing a total order for stores on all atomic
184 * OVS atomics require the memory_order to be passed as a compile-time constant
185 * value, as some compiler implementations may perform poorly if the memory
186 * order parameter is passed in as a run-time value.
188 * The following functions insert explicit barriers. Most of the other atomic
189 * functions also include barriers.
191 * void atomic_thread_fence(memory_order order);
193 * Inserts a barrier of the specified type.
195 * For memory_order_relaxed, this is a no-op.
197 * void atomic_signal_fence(memory_order order);
199 * Inserts a barrier of the specified type, but only with respect to
200 * signal handlers in the same thread as the barrier. This is
201 * basically a compiler optimization barrier, except for
202 * memory_order_relaxed, which is a no-op.
208 * In this section, A is an atomic type and C is the corresponding non-atomic
211 * The "store" and "compare_exchange" primitives match C11:
213 * void atomic_store(A *object, C value);
214 * void atomic_store_explicit(A *object, C value, memory_order);
216 * Atomically stores 'value' into '*object', respecting the given
217 * memory order (or memory_order_seq_cst for atomic_store()).
219 * bool atomic_compare_exchange_strong(A *object, C *expected, C desired);
220 * bool atomic_compare_exchange_weak(A *object, C *expected, C desired);
221 * bool atomic_compare_exchange_strong_explicit(A *object, C *expected,
223 * memory_order success,
224 * memory_order failure);
225 * bool atomic_compare_exchange_weak_explicit(A *object, C *expected,
227 * memory_order success,
228 * memory_order failure);
230 * Atomically loads '*object' and compares it with '*expected' and if
231 * equal, stores 'desired' into '*object' (an atomic read-modify-write
232 * operation) and returns true, and if non-equal, stores the actual
233 * value of '*object' into '*expected' (an atomic load operation) and
234 * returns false. The memory order for the successful case (atomic
235 * read-modify-write operation) is 'success', and for the unsuccessful
236 * case (atomic load operation) 'failure'. 'failure' may not be
237 * stronger than 'success'.
239 * The weak forms may fail (returning false) also when '*object' equals
240 * '*expected'. The strong form can be implemented by the weak form in
241 * a loop. Some platforms can implement the weak form more
242 * efficiently, so it should be used if the application will need to
245 * The following primitives differ from the C11 ones (and have different names)
246 * because there does not appear to be a way to implement the standard
247 * primitives in standard C:
249 * void atomic_read(A *src, C *dst);
250 * void atomic_read_explicit(A *src, C *dst, memory_order);
252 * Atomically loads a value from 'src', writing the value read into
253 * '*dst', respecting the given memory order (or memory_order_seq_cst
254 * for atomic_read()).
256 * void atomic_add(A *rmw, C arg, C *orig);
257 * void atomic_sub(A *rmw, C arg, C *orig);
258 * void atomic_or(A *rmw, C arg, C *orig);
259 * void atomic_xor(A *rmw, C arg, C *orig);
260 * void atomic_and(A *rmw, C arg, C *orig);
261 * void atomic_add_explicit(A *rmw, C arg, C *orig, memory_order);
262 * void atomic_sub_explicit(A *rmw, C arg, C *orig, memory_order);
263 * void atomic_or_explicit(A *rmw, C arg, C *orig, memory_order);
264 * void atomic_xor_explicit(A *rmw, C arg, C *orig, memory_order);
265 * void atomic_and_explicit(A *rmw, C arg, C *orig, memory_order);
267 * Atomically applies the given operation, with 'arg' as the second
268 * operand, to '*rmw', and stores the original value of '*rmw' into
269 * '*orig', respecting the given memory order (or memory_order_seq_cst
270 * if none is specified).
272 * The results are similar to those that would be obtained with +=, -=,
273 * |=, ^=, or |= on non-atomic types.
279 * atomic_flag is a typedef for a type with two states, set and clear, that
280 * provides atomic test-and-set functionality.
286 * ATOMIC_FLAG_INIT is an initializer for atomic_flag. The initial state is
289 * An atomic_flag may also be initialized at runtime with atomic_flag_clear().
295 * The following functions are available.
297 * bool atomic_flag_test_and_set(atomic_flag *object)
298 * bool atomic_flag_test_and_set_explicit(atomic_flag *object,
301 * Atomically sets '*object', respsecting the given memory order (or
302 * memory_order_seq_cst for atomic_flag_test_and_set()). Returns the
303 * previous value of the flag (false for clear, true for set).
305 * void atomic_flag_clear(atomic_flag *object);
306 * void atomic_flag_clear_explicit(atomic_flag *object, memory_order);
308 * Atomically clears '*object', respecting the given memory order (or
309 * memory_order_seq_cst for atomic_flag_clear()).
317 #include "compiler.h"
320 #define IN_OVS_ATOMIC_H
322 /* sparse doesn't understand some GCC extensions we use. */
323 #include "ovs-atomic-pthreads.h"
324 #elif HAVE_STDATOMIC_H
325 #include "ovs-atomic-c11.h"
326 #elif __has_extension(c_atomic)
327 #include "ovs-atomic-clang.h"
328 #elif __GNUC__ >= 4 && __GNUC_MINOR__ >= 7
329 #include "ovs-atomic-gcc4.7+.h"
330 #elif __GNUC__ && defined(__x86_64__)
331 #include "ovs-atomic-x86_64.h"
332 #elif __GNUC__ && defined(__i386__)
333 #include "ovs-atomic-i586.h"
334 #elif HAVE_GCC4_ATOMICS
335 #include "ovs-atomic-gcc4+.h"
337 /* ovs-atomic-pthreads implementation is provided for portability.
338 * It might be too slow for real use because Open vSwitch is
339 * optimized for platforms where real atomic ops are available. */
340 #include "ovs-atomic-pthreads.h"
342 #undef IN_OVS_ATOMIC_H
344 #ifndef OMIT_STANDARD_ATOMIC_TYPES
345 typedef ATOMIC(bool) atomic_bool;
347 typedef ATOMIC(char) atomic_char;
348 typedef ATOMIC(signed char) atomic_schar;
349 typedef ATOMIC(unsigned char) atomic_uchar;
351 typedef ATOMIC(short) atomic_short;
352 typedef ATOMIC(unsigned short) atomic_ushort;
354 typedef ATOMIC(int) atomic_int;
355 typedef ATOMIC(unsigned int) atomic_uint;
357 typedef ATOMIC(long) atomic_long;
358 typedef ATOMIC(unsigned long) atomic_ulong;
360 typedef ATOMIC(long long) atomic_llong;
361 typedef ATOMIC(unsigned long long) atomic_ullong;
363 typedef ATOMIC(size_t) atomic_size_t;
364 typedef ATOMIC(ptrdiff_t) atomic_ptrdiff_t;
366 typedef ATOMIC(intmax_t) atomic_intmax_t;
367 typedef ATOMIC(uintmax_t) atomic_uintmax_t;
369 typedef ATOMIC(intptr_t) atomic_intptr_t;
370 typedef ATOMIC(uintptr_t) atomic_uintptr_t;
371 #endif /* !OMIT_STANDARD_ATOMIC_TYPES */
373 /* Nonstandard atomic types. */
374 typedef ATOMIC(uint8_t) atomic_uint8_t;
375 typedef ATOMIC(uint16_t) atomic_uint16_t;
376 typedef ATOMIC(uint32_t) atomic_uint32_t;
378 typedef ATOMIC(int8_t) atomic_int8_t;
379 typedef ATOMIC(int16_t) atomic_int16_t;
380 typedef ATOMIC(int32_t) atomic_int32_t;
382 /* Relaxed atomic operations.
384 * When an operation on an atomic variable is not expected to synchronize
385 * with operations on other (atomic or non-atomic) variables, no memory
386 * barriers are needed and the relaxed memory ordering can be used. These
387 * macros make such uses less daunting, but not invisible. */
388 #define atomic_store_relaxed(VAR, VALUE) \
389 atomic_store_explicit(VAR, VALUE, memory_order_relaxed)
390 #define atomic_read_relaxed(VAR, DST) \
391 atomic_read_explicit(VAR, DST, memory_order_relaxed)
392 #define atomic_compare_exchange_strong_relaxed(DST, EXP, SRC) \
393 atomic_compare_exchange_strong_explicit(DST, EXP, SRC, \
394 memory_order_relaxed, \
395 memory_order_relaxed)
396 #define atomic_compare_exchange_weak_relaxed(DST, EXP, SRC) \
397 atomic_compare_exchange_weak_explicit(DST, EXP, SRC, \
398 memory_order_relaxed, \
399 memory_order_relaxed)
400 #define atomic_add_relaxed(RMW, ARG, ORIG) \
401 atomic_add_explicit(RMW, ARG, ORIG, memory_order_relaxed)
402 #define atomic_sub_relaxed(RMW, ARG, ORIG) \
403 atomic_sub_explicit(RMW, ARG, ORIG, memory_order_relaxed)
404 #define atomic_or_relaxed(RMW, ARG, ORIG) \
405 atomic_or_explicit(RMW, ARG, ORIG, memory_order_relaxed)
406 #define atomic_xor_relaxed(RMW, ARG, ORIG) \
407 atomic_xor_explicit(RMW, ARG, ORIG, memory_order_relaxed)
408 #define atomic_and_relaxed(RMW, ARG, ORIG) \
409 atomic_and_explicit(RMW, ARG, ORIG, memory_order_relaxed)
410 #define atomic_flag_test_and_set_relaxed(FLAG) \
411 atomic_flag_test_and_set_explicit(FLAG, memory_order_relaxed)
412 #define atomic_flag_clear_relaxed(FLAG) \
413 atomic_flag_clear_explicit(FLAG, memory_order_relaxed)
415 /* A simplified atomic count. Does not provide any synchronization with any
418 * Typically a counter is not used to synchronize the state of any other
419 * variables (with the notable exception of reference count, below).
420 * This abstraction releaves the user from the memory order considerations,
421 * and may make the code easier to read.
423 * We only support the unsigned int counters, as those are the most common. */
424 typedef struct atomic_count {
428 #define ATOMIC_COUNT_INIT(VALUE) { VALUE }
431 atomic_count_init(atomic_count *count, unsigned int value)
433 atomic_init(&count->count, value);
436 static inline unsigned int
437 atomic_count_inc(atomic_count *count)
441 atomic_add_relaxed(&count->count, 1, &old);
446 static inline unsigned int
447 atomic_count_dec(atomic_count *count)
451 atomic_sub_relaxed(&count->count, 1, &old);
456 static inline unsigned int
457 atomic_count_get(atomic_count *count)
461 atomic_read_relaxed(&count->count, &value);
467 atomic_count_set(atomic_count *count, unsigned int value)
469 atomic_store_relaxed(&count->count, value);
472 /* Reference count. */
473 struct ovs_refcount {
477 /* Initializes 'refcount'. The reference count is initially 1. */
479 ovs_refcount_init(struct ovs_refcount *refcount)
481 atomic_init(&refcount->count, 1);
484 /* Increments 'refcount'.
486 * Does not provide a memory barrier, as the calling thread must have
487 * protected access to the object already. */
489 ovs_refcount_ref(struct ovs_refcount *refcount)
491 unsigned int old_refcount;
493 atomic_add_explicit(&refcount->count, 1, &old_refcount,
494 memory_order_relaxed);
495 ovs_assert(old_refcount > 0);
498 /* Decrements 'refcount' and returns the previous reference count. Often used
501 * if (ovs_refcount_unref(&object->ref_cnt) == 1) {
502 * // ...uninitialize object...
506 * Provides a release barrier making the preceding loads and stores to not be
507 * reordered after the unref, and in case of the last reference provides also
508 * an acquire barrier to keep all the following uninitialization from being
509 * reordered before the atomic decrement operation. Together these synchronize
510 * any concurrent unref operations between each other. */
511 static inline unsigned int
512 ovs_refcount_unref(struct ovs_refcount *refcount)
514 unsigned int old_refcount;
516 atomic_sub_explicit(&refcount->count, 1, &old_refcount,
517 memory_order_release);
518 ovs_assert(old_refcount > 0);
519 if (old_refcount == 1) {
520 /* 'memory_order_release' above means that there are no (reordered)
521 * accesses to the protected object from any thread at this point.
522 * An acquire barrier is needed to keep all subsequent access to the
523 * object's memory from being reordered before the atomic operation
525 atomic_thread_fence(memory_order_acquire);
530 /* Reads and returns 'refcount_''s current reference count.
532 * Does not provide a memory barrier.
535 static inline unsigned int
536 ovs_refcount_read(const struct ovs_refcount *refcount_)
538 struct ovs_refcount *refcount
539 = CONST_CAST(struct ovs_refcount *, refcount_);
542 atomic_read_explicit(&refcount->count, &count, memory_order_relaxed);
546 /* Increments 'refcount', but only if it is non-zero.
548 * This may only be called for an object which is RCU protected during
549 * this call. This implies that its possible destruction is postponed
550 * until all current RCU threads quiesce.
552 * Returns false if the refcount was zero. In this case the object may
553 * be safely accessed until the current thread quiesces, but no additional
554 * references to the object may be taken.
556 * Does not provide a memory barrier, as the calling thread must have
557 * RCU protected access to the object already.
559 * It is critical that we never increment a zero refcount to a
560 * non-zero value, as whenever a refcount reaches the zero value, the
561 * protected object may be irrevocably scheduled for deletion. */
563 ovs_refcount_try_ref_rcu(struct ovs_refcount *refcount)
567 atomic_read_explicit(&refcount->count, &count, memory_order_relaxed);
572 } while (!atomic_compare_exchange_weak_explicit(&refcount->count, &count,
574 memory_order_relaxed,
575 memory_order_relaxed));
579 /* Decrements 'refcount' and returns the previous reference count. To
580 * be used only when a memory barrier is already provided for the
581 * protected object independently.
585 * if (ovs_refcount_unref_relaxed(&object->ref_cnt) == 1) {
586 * // Schedule uninitialization and freeing of the object:
587 * ovsrcu_postpone(destructor_function, object);
590 * Here RCU quiescing already provides a full memory barrier. No additional
591 * barriers are needed here.
595 * if (stp && ovs_refcount_unref_relaxed(&stp->ref_cnt) == 1) {
596 * ovs_mutex_lock(&mutex);
597 * list_remove(&stp->node);
598 * ovs_mutex_unlock(&mutex);
603 * Here a mutex is used to guard access to all of 'stp' apart from
604 * 'ref_cnt'. Hence all changes to 'stp' by other threads must be
605 * visible when we get the mutex, and no access after the unlock can
606 * be reordered to happen prior the lock operation. No additional
607 * barriers are needed here.
609 static inline unsigned int
610 ovs_refcount_unref_relaxed(struct ovs_refcount *refcount)
612 unsigned int old_refcount;
614 atomic_sub_explicit(&refcount->count, 1, &old_refcount,
615 memory_order_relaxed);
616 ovs_assert(old_refcount > 0);
620 #endif /* ovs-atomic.h */