1
Semantics and Behavior of Atomic and
6
This document is intended to serve as a guide to Linux port
7
maintainers on how to implement atomic counter, bitops, and spinlock
10
The atomic_t type should be defined as a signed integer.
11
Also, it should be made opaque such that any kind of cast to a normal
12
C integer type will fail. Something like the following should
15
typedef struct { volatile int counter; } atomic_t;
17
Historically, counter has been declared volatile. This is now discouraged.
18
See Documentation/volatile-considered-harmful.txt for the complete rationale.
20
local_t is very similar to atomic_t. If the counter is per CPU and only
21
updated by one CPU, local_t is probably more appropriate. Please see
22
Documentation/local_ops.txt for the semantics of local_t.
24
The first operations to implement for atomic_t's are the initializers and
27
#define ATOMIC_INIT(i) { (i) }
28
#define atomic_set(v, i) ((v)->counter = (i))
30
The first macro is used in definitions, such as:
32
static atomic_t my_counter = ATOMIC_INIT(1);
34
The initializer is atomic in that the return values of the atomic operations
35
are guaranteed to be correct reflecting the initialized value if the
36
initializer is used before runtime. If the initializer is used at runtime, a
37
proper implicit or explicit read memory barrier is needed before reading the
38
value with atomic_read from another thread.
40
The second interface can be used at runtime, as in:
42
struct foo { atomic_t counter; };
47
k = kmalloc(sizeof(*k), GFP_KERNEL);
50
atomic_set(&k->counter, 0);
52
The setting is atomic in that the return values of the atomic operations by
53
all threads are guaranteed to be correct reflecting either the value that has
54
been set with this operation or set with another operation. A proper implicit
55
or explicit memory barrier is needed before the value set with the operation
56
is guaranteed to be readable with atomic_read from another thread.
60
#define atomic_read(v) ((v)->counter)
62
which simply reads the counter value currently visible to the calling thread.
63
The read is atomic in that the return value is guaranteed to be one of the
64
values initialized or modified with the interface operations if a proper
65
implicit or explicit memory barrier is used after possible runtime
66
initialization by any other thread and the value is modified only with the
67
interface operations. atomic_read does not guarantee that the runtime
68
initialization by any other thread is visible yet, so the user of the
69
interface must take care of that with a proper implicit or explicit memory
72
*** WARNING: atomic_read() and atomic_set() DO NOT IMPLY BARRIERS! ***
74
Some architectures may choose to use the volatile keyword, barriers, or inline
75
assembly to guarantee some degree of immediacy for atomic_read() and
76
atomic_set(). This is not uniformly guaranteed, and may change in the future,
77
so all users of atomic_t should treat atomic_read() and atomic_set() as simple
78
C statements that may be reordered or optimized away entirely by the compiler
79
or processor, and explicitly invoke the appropriate compiler and/or memory
80
barrier for each use case. Failure to do so will result in code that may
81
suddenly break when used with different architectures or compiler
82
optimizations, or even changes in unrelated code which changes how the
83
compiler optimizes the section accessing atomic_t variables.
85
*** YOU HAVE BEEN WARNED! ***
87
Now, we move onto the atomic operation interfaces typically implemented with
88
the help of assembly code.
90
void atomic_add(int i, atomic_t *v);
91
void atomic_sub(int i, atomic_t *v);
92
void atomic_inc(atomic_t *v);
93
void atomic_dec(atomic_t *v);
95
These four routines add and subtract integral values to/from the given
96
atomic_t value. The first two routines pass explicit integers by
97
which to make the adjustment, whereas the latter two use an implicit
98
adjustment value of "1".
100
One very important aspect of these two routines is that they DO NOT
101
require any explicit memory barriers. They need only perform the
102
atomic_t counter update in an SMP safe manner.
106
int atomic_inc_return(atomic_t *v);
107
int atomic_dec_return(atomic_t *v);
109
These routines add 1 and subtract 1, respectively, from the given
110
atomic_t and return the new counter value after the operation is
113
Unlike the above routines, it is required that explicit memory
114
barriers are performed before and after the operation. It must be
115
done such that all memory operations before and after the atomic
116
operation calls are strongly ordered with respect to the atomic
119
For example, it should behave as if a smp_mb() call existed both
120
before and after the atomic operation.
122
If the atomic instructions used in an implementation provide explicit
123
memory barrier semantics which satisfy the above requirements, that is
128
int atomic_add_return(int i, atomic_t *v);
129
int atomic_sub_return(int i, atomic_t *v);
131
These behave just like atomic_{inc,dec}_return() except that an
132
explicit counter adjustment is given instead of the implicit "1".
133
This means that like atomic_{inc,dec}_return(), the memory barrier
134
semantics are required.
138
int atomic_inc_and_test(atomic_t *v);
139
int atomic_dec_and_test(atomic_t *v);
141
These two routines increment and decrement by 1, respectively, the
142
given atomic counter. They return a boolean indicating whether the
143
resulting counter value was zero or not.
145
It requires explicit memory barrier semantics around the operation as
148
int atomic_sub_and_test(int i, atomic_t *v);
150
This is identical to atomic_dec_and_test() except that an explicit
151
decrement is given instead of the implicit "1". It requires explicit
152
memory barrier semantics around the operation.
154
int atomic_add_negative(int i, atomic_t *v);
156
The given increment is added to the given atomic counter value. A
157
boolean is return which indicates whether the resulting counter value
158
is negative. It requires explicit memory barrier semantics around the
163
int atomic_xchg(atomic_t *v, int new);
165
This performs an atomic exchange operation on the atomic variable v, setting
166
the given new value. It returns the old value that the atomic variable v had
167
just before the operation.
169
int atomic_cmpxchg(atomic_t *v, int old, int new);
171
This performs an atomic compare exchange operation on the atomic value v,
172
with the given old and new values. Like all atomic_xxx operations,
173
atomic_cmpxchg will only satisfy its atomicity semantics as long as all
174
other accesses of *v are performed through atomic_xxx operations.
176
atomic_cmpxchg requires explicit memory barriers around the operation.
178
The semantics for atomic_cmpxchg are the same as those defined for 'cas'
183
int atomic_add_unless(atomic_t *v, int a, int u);
185
If the atomic value v is not equal to u, this function adds a to v, and
186
returns non zero. If v is equal to u then it returns zero. This is done as
189
atomic_add_unless requires explicit memory barriers around the operation
190
unless it fails (returns 0).
192
atomic_inc_not_zero, equivalent to atomic_add_unless(v, 1, 0)
195
If a caller requires memory barrier semantics around an atomic_t
196
operation which does not return a value, a set of interfaces are
197
defined which accomplish this:
199
void smp_mb__before_atomic_dec(void);
200
void smp_mb__after_atomic_dec(void);
201
void smp_mb__before_atomic_inc(void);
202
void smp_mb__after_atomic_inc(void);
204
For example, smp_mb__before_atomic_dec() can be used like so:
207
smp_mb__before_atomic_dec();
208
atomic_dec(&obj->ref_count);
210
It makes sure that all memory operations preceding the atomic_dec()
211
call are strongly ordered with respect to the atomic counter
212
operation. In the above example, it guarantees that the assignment of
213
"1" to obj->dead will be globally visible to other cpus before the
214
atomic counter decrement.
216
Without the explicit smp_mb__before_atomic_dec() call, the
217
implementation could legally allow the atomic counter update visible
218
to other cpus before the "obj->dead = 1;" assignment.
220
The other three interfaces listed are used to provide explicit
221
ordering with respect to memory operations after an atomic_dec() call
222
(smp_mb__after_atomic_dec()) and around atomic_inc() calls
223
(smp_mb__{before,after}_atomic_inc()).
225
A missing memory barrier in the cases where they are required by the
226
atomic_t implementation above can have disastrous results. Here is
227
an example, which follows a pattern occurring frequently in the Linux
228
kernel. It is the use of atomic counters to implement reference
229
counting, and it works such that once the counter falls to zero it can
230
be guaranteed that no other entity can be accessing the object:
232
static void obj_list_add(struct obj *obj, struct list_head *head)
235
list_add(&obj->list, head);
238
static void obj_list_del(struct obj *obj)
240
list_del(&obj->list);
244
static void obj_destroy(struct obj *obj)
250
struct obj *obj_list_peek(struct list_head *head)
252
if (!list_empty(head)) {
255
obj = list_entry(head->next, struct obj, list);
256
atomic_inc(&obj->refcnt);
266
spin_lock(&global_list_lock);
267
obj = obj_list_peek(&global_list);
268
spin_unlock(&global_list_lock);
272
if (atomic_dec_and_test(&obj->refcnt))
277
void obj_timeout(struct obj *obj)
279
spin_lock(&global_list_lock);
281
spin_unlock(&global_list_lock);
283
if (atomic_dec_and_test(&obj->refcnt))
287
(This is a simplification of the ARP queue management in the
288
generic neighbour discover code of the networking. Olaf Kirch
289
found a bug wrt. memory barriers in kfree_skb() that exposed
290
the atomic_t memory barrier requirements quite clearly.)
292
Given the above scheme, it must be the case that the obj->active
293
update done by the obj list deletion be visible to other processors
294
before the atomic counter decrement is performed.
296
Otherwise, the counter could fall to zero, yet obj->active would still
297
be set, thus triggering the assertion in obj_destroy(). The error
298
sequence looks like this:
301
obj_poke() obj_timeout()
302
obj = obj_list_peek();
303
... gains ref to obj, refcnt=2
306
... visibility delayed ...
307
atomic_dec_and_test()
308
... refcnt drops to 1 ...
309
atomic_dec_and_test()
310
... refcount drops to 0 ...
312
BUG() triggers since obj->active
314
obj->active update visibility occurs
316
With the memory barrier semantics required of the atomic_t operations
317
which return values, the above sequence of memory visibility can never
318
happen. Specifically, in the above case the atomic_dec_and_test()
319
counter decrement would not become globally visible until the
320
obj->active update does.
322
As a historical note, 32-bit Sparc used to only allow usage of
323
24-bits of it's atomic_t type. This was because it used 8 bits
324
as a spinlock for SMP safety. Sparc32 lacked a "compare and swap"
325
type instruction. However, 32-bit Sparc has since been moved over
326
to a "hash table of spinlocks" scheme, that allows the full 32-bit
327
counter to be realized. Essentially, an array of spinlocks are
328
indexed into based upon the address of the atomic_t being operated
329
on, and that lock protects the atomic operation. Parisc uses the
332
Another note is that the atomic_t operations returning values are
333
extremely slow on an old 386.
335
We will now cover the atomic bitmask operations. You will find that
336
their SMP and memory barrier semantics are similar in shape and scope
337
to the atomic_t ops above.
339
Native atomic bit operations are defined to operate on objects aligned
340
to the size of an "unsigned long" C data type, and are least of that
341
size. The endianness of the bits within each "unsigned long" are the
342
native endianness of the cpu.
344
void set_bit(unsigned long nr, volatile unsigned long *addr);
345
void clear_bit(unsigned long nr, volatile unsigned long *addr);
346
void change_bit(unsigned long nr, volatile unsigned long *addr);
348
These routines set, clear, and change, respectively, the bit number
349
indicated by "nr" on the bit mask pointed to by "ADDR".
351
They must execute atomically, yet there are no implicit memory barrier
352
semantics required of these interfaces.
354
int test_and_set_bit(unsigned long nr, volatile unsigned long *addr);
355
int test_and_clear_bit(unsigned long nr, volatile unsigned long *addr);
356
int test_and_change_bit(unsigned long nr, volatile unsigned long *addr);
358
Like the above, except that these routines return a boolean which
359
indicates whether the changed bit was set _BEFORE_ the atomic bit
362
WARNING! It is incredibly important that the value be a boolean,
363
ie. "0" or "1". Do not try to be fancy and save a few instructions by
364
declaring the above to return "long" and just returning something like
365
"old_val & mask" because that will not work.
367
For one thing, this return value gets truncated to int in many code
368
paths using these interfaces, so on 64-bit if the bit is set in the
369
upper 32-bits then testers will never see that.
371
One great example of where this problem crops up are the thread_info
372
flag operations. Routines such as test_and_set_ti_thread_flag() chop
373
the return value into an int. There are other places where things
374
like this occur as well.
376
These routines, like the atomic_t counter operations returning values,
377
require explicit memory barrier semantics around their execution. All
378
memory operations before the atomic bit operation call must be made
379
visible globally before the atomic bit operation is made visible.
380
Likewise, the atomic bit operation must be visible globally before any
381
subsequent memory operation is made visible. For example:
384
if (test_and_set_bit(0, &obj->flags))
388
The implementation of test_and_set_bit() must guarantee that
389
"obj->dead = 1;" is visible to cpus before the atomic memory operation
390
done by test_and_set_bit() becomes visible. Likewise, the atomic
391
memory operation done by test_and_set_bit() must become visible before
392
"obj->killed = 1;" is visible.
394
Finally there is the basic operation:
396
int test_bit(unsigned long nr, __const__ volatile unsigned long *addr);
398
Which returns a boolean indicating if bit "nr" is set in the bitmask
399
pointed to by "addr".
401
If explicit memory barriers are required around clear_bit() (which
402
does not return a value, and thus does not need to provide memory
403
barrier semantics), two interfaces are provided:
405
void smp_mb__before_clear_bit(void);
406
void smp_mb__after_clear_bit(void);
408
They are used as follows, and are akin to their atomic_t operation
411
/* All memory operations before this call will
412
* be globally visible before the clear_bit().
414
smp_mb__before_clear_bit();
417
/* The clear_bit() will be visible before all
418
* subsequent memory operations.
420
smp_mb__after_clear_bit();
422
There are two special bitops with lock barrier semantics (acquire/release,
423
same as spinlocks). These operate in the same way as their non-_lock/unlock
424
postfixed variants, except that they are to provide acquire/release semantics,
425
respectively. This means they can be used for bit_spin_trylock and
426
bit_spin_unlock type operations without specifying any more barriers.
428
int test_and_set_bit_lock(unsigned long nr, unsigned long *addr);
429
void clear_bit_unlock(unsigned long nr, unsigned long *addr);
430
void __clear_bit_unlock(unsigned long nr, unsigned long *addr);
432
The __clear_bit_unlock version is non-atomic, however it still implements
433
unlock barrier semantics. This can be useful if the lock itself is protecting
434
the other bits in the word.
436
Finally, there are non-atomic versions of the bitmask operations
437
provided. They are used in contexts where some other higher-level SMP
438
locking scheme is being used to protect the bitmask, and thus less
439
expensive non-atomic operations may be used in the implementation.
440
They have names similar to the above bitmask operation interfaces,
441
except that two underscores are prefixed to the interface name.
443
void __set_bit(unsigned long nr, volatile unsigned long *addr);
444
void __clear_bit(unsigned long nr, volatile unsigned long *addr);
445
void __change_bit(unsigned long nr, volatile unsigned long *addr);
446
int __test_and_set_bit(unsigned long nr, volatile unsigned long *addr);
447
int __test_and_clear_bit(unsigned long nr, volatile unsigned long *addr);
448
int __test_and_change_bit(unsigned long nr, volatile unsigned long *addr);
450
These non-atomic variants also do not require any special memory
453
The routines xchg() and cmpxchg() need the same exact memory barriers
454
as the atomic and bit operations returning values.
456
Spinlocks and rwlocks have memory barrier expectations as well.
457
The rule to follow is simple:
459
1) When acquiring a lock, the implementation must make it globally
460
visible before any subsequent memory operation.
462
2) When releasing a lock, the implementation must make it such that
463
all previous memory operations are globally visible before the
466
Which finally brings us to _atomic_dec_and_lock(). There is an
467
architecture-neutral version implemented in lib/dec_and_lock.c,
468
but most platforms will wish to optimize this in assembler.
470
int _atomic_dec_and_lock(atomic_t *atomic, spinlock_t *lock);
472
Atomically decrement the given counter, and if will drop to zero
473
atomically acquire the given spinlock and perform the decrement
474
of the counter to zero. If it does not drop to zero, do nothing
477
It is actually pretty simple to get the memory barrier correct.
478
Simply satisfy the spinlock grab requirements, which is make
479
sure the spinlock operation is globally visible before any
480
subsequent memory operation.
482
We can demonstrate this operation more clearly if we define
483
an abstract atomic operation:
485
long cas(long *mem, long old, long new);
487
"cas" stands for "compare and swap". It atomically:
489
1) Compares "old" with the value currently at "mem".
490
2) If they are equal, "new" is written to "mem".
491
3) Regardless, the current value at "mem" is returned.
493
As an example usage, here is what an atomic counter update
496
void example_atomic_inc(long *counter)
504
ret = cas(counter, old, new);
510
Let's use cas() in order to build a pseudo-C atomic_dec_and_lock():
512
int _atomic_dec_and_lock(atomic_t *atomic, spinlock_t *lock)
519
old = atomic_read(atomic);
525
ret = cas(atomic, old, new);
537
Now, as far as memory barriers go, as long as spin_lock()
538
strictly orders all subsequent memory operations (including
539
the cas()) with respect to itself, things will be fine.
541
Said another way, _atomic_dec_and_lock() must guarantee that
542
a counter dropping to zero is never made visible before the
543
spinlock being acquired.
545
Note that this also means that for the case where the counter
546
is not dropping to zero, there are no memory ordering