Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1 | Semantics and Behavior of Atomic and |
| 2 | Bitmask Operations |
| 3 | |
| 4 | David S. Miller |
| 5 | |
| 6 | This document is intended to serve as a guide to Linux port |
| 7 | maintainers on how to implement atomic counter, bitops, and spinlock |
| 8 | interfaces properly. |
| 9 | |
| 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 |
| 13 | suffice: |
| 14 | |
| 15 | typedef struct { volatile int counter; } atomic_t; |
| 16 | |
| 17 | The first operations to implement for atomic_t's are the |
| 18 | initializers and plain reads. |
| 19 | |
| 20 | #define ATOMIC_INIT(i) { (i) } |
| 21 | #define atomic_set(v, i) ((v)->counter = (i)) |
| 22 | |
| 23 | The first macro is used in definitions, such as: |
| 24 | |
| 25 | static atomic_t my_counter = ATOMIC_INIT(1); |
| 26 | |
| 27 | The second interface can be used at runtime, as in: |
| 28 | |
| 29 | struct foo { atomic_t counter; }; |
| 30 | ... |
| 31 | |
| 32 | struct foo *k; |
| 33 | |
| 34 | k = kmalloc(sizeof(*k), GFP_KERNEL); |
| 35 | if (!k) |
| 36 | return -ENOMEM; |
| 37 | atomic_set(&k->counter, 0); |
| 38 | |
| 39 | Next, we have: |
| 40 | |
| 41 | #define atomic_read(v) ((v)->counter) |
| 42 | |
| 43 | which simply reads the current value of the counter. |
| 44 | |
| 45 | Now, we move onto the actual atomic operation interfaces. |
| 46 | |
| 47 | void atomic_add(int i, atomic_t *v); |
| 48 | void atomic_sub(int i, atomic_t *v); |
| 49 | void atomic_inc(atomic_t *v); |
| 50 | void atomic_dec(atomic_t *v); |
| 51 | |
| 52 | These four routines add and subtract integral values to/from the given |
| 53 | atomic_t value. The first two routines pass explicit integers by |
| 54 | which to make the adjustment, whereas the latter two use an implicit |
| 55 | adjustment value of "1". |
| 56 | |
| 57 | One very important aspect of these two routines is that they DO NOT |
| 58 | require any explicit memory barriers. They need only perform the |
| 59 | atomic_t counter update in an SMP safe manner. |
| 60 | |
| 61 | Next, we have: |
| 62 | |
| 63 | int atomic_inc_return(atomic_t *v); |
| 64 | int atomic_dec_return(atomic_t *v); |
| 65 | |
| 66 | These routines add 1 and subtract 1, respectively, from the given |
| 67 | atomic_t and return the new counter value after the operation is |
| 68 | performed. |
| 69 | |
| 70 | Unlike the above routines, it is required that explicit memory |
| 71 | barriers are performed before and after the operation. It must be |
| 72 | done such that all memory operations before and after the atomic |
| 73 | operation calls are strongly ordered with respect to the atomic |
| 74 | operation itself. |
| 75 | |
| 76 | For example, it should behave as if a smp_mb() call existed both |
| 77 | before and after the atomic operation. |
| 78 | |
| 79 | If the atomic instructions used in an implementation provide explicit |
| 80 | memory barrier semantics which satisfy the above requirements, that is |
| 81 | fine as well. |
| 82 | |
| 83 | Let's move on: |
| 84 | |
| 85 | int atomic_add_return(int i, atomic_t *v); |
| 86 | int atomic_sub_return(int i, atomic_t *v); |
| 87 | |
| 88 | These behave just like atomic_{inc,dec}_return() except that an |
| 89 | explicit counter adjustment is given instead of the implicit "1". |
| 90 | This means that like atomic_{inc,dec}_return(), the memory barrier |
| 91 | semantics are required. |
| 92 | |
| 93 | Next: |
| 94 | |
| 95 | int atomic_inc_and_test(atomic_t *v); |
| 96 | int atomic_dec_and_test(atomic_t *v); |
| 97 | |
| 98 | These two routines increment and decrement by 1, respectively, the |
| 99 | given atomic counter. They return a boolean indicating whether the |
| 100 | resulting counter value was zero or not. |
| 101 | |
| 102 | It requires explicit memory barrier semantics around the operation as |
| 103 | above. |
| 104 | |
| 105 | int atomic_sub_and_test(int i, atomic_t *v); |
| 106 | |
| 107 | This is identical to atomic_dec_and_test() except that an explicit |
| 108 | decrement is given instead of the implicit "1". It requires explicit |
| 109 | memory barrier semantics around the operation. |
| 110 | |
| 111 | int atomic_add_negative(int i, atomic_t *v); |
| 112 | |
| 113 | The given increment is added to the given atomic counter value. A |
| 114 | boolean is return which indicates whether the resulting counter value |
| 115 | is negative. It requires explicit memory barrier semantics around the |
| 116 | operation. |
| 117 | |
| 118 | If a caller requires memory barrier semantics around an atomic_t |
| 119 | operation which does not return a value, a set of interfaces are |
| 120 | defined which accomplish this: |
| 121 | |
| 122 | void smp_mb__before_atomic_dec(void); |
| 123 | void smp_mb__after_atomic_dec(void); |
| 124 | void smp_mb__before_atomic_inc(void); |
| 125 | void smp_mb__after_atomic_dec(void); |
| 126 | |
| 127 | For example, smp_mb__before_atomic_dec() can be used like so: |
| 128 | |
| 129 | obj->dead = 1; |
| 130 | smp_mb__before_atomic_dec(); |
| 131 | atomic_dec(&obj->ref_count); |
| 132 | |
| 133 | It makes sure that all memory operations preceeding the atomic_dec() |
| 134 | call are strongly ordered with respect to the atomic counter |
| 135 | operation. In the above example, it guarentees that the assignment of |
| 136 | "1" to obj->dead will be globally visible to other cpus before the |
| 137 | atomic counter decrement. |
| 138 | |
| 139 | Without the explicitl smp_mb__before_atomic_dec() call, the |
| 140 | implementation could legally allow the atomic counter update visible |
| 141 | to other cpus before the "obj->dead = 1;" assignment. |
| 142 | |
| 143 | The other three interfaces listed are used to provide explicit |
| 144 | ordering with respect to memory operations after an atomic_dec() call |
| 145 | (smp_mb__after_atomic_dec()) and around atomic_inc() calls |
| 146 | (smp_mb__{before,after}_atomic_inc()). |
| 147 | |
| 148 | A missing memory barrier in the cases where they are required by the |
| 149 | atomic_t implementation above can have disasterous results. Here is |
| 150 | an example, which follows a pattern occuring frequently in the Linux |
| 151 | kernel. It is the use of atomic counters to implement reference |
| 152 | counting, and it works such that once the counter falls to zero it can |
| 153 | be guarenteed that no other entity can be accessing the object: |
| 154 | |
| 155 | static void obj_list_add(struct obj *obj) |
| 156 | { |
| 157 | obj->active = 1; |
| 158 | list_add(&obj->list); |
| 159 | } |
| 160 | |
| 161 | static void obj_list_del(struct obj *obj) |
| 162 | { |
| 163 | list_del(&obj->list); |
| 164 | obj->active = 0; |
| 165 | } |
| 166 | |
| 167 | static void obj_destroy(struct obj *obj) |
| 168 | { |
| 169 | BUG_ON(obj->active); |
| 170 | kfree(obj); |
| 171 | } |
| 172 | |
| 173 | struct obj *obj_list_peek(struct list_head *head) |
| 174 | { |
| 175 | if (!list_empty(head)) { |
| 176 | struct obj *obj; |
| 177 | |
| 178 | obj = list_entry(head->next, struct obj, list); |
| 179 | atomic_inc(&obj->refcnt); |
| 180 | return obj; |
| 181 | } |
| 182 | return NULL; |
| 183 | } |
| 184 | |
| 185 | void obj_poke(void) |
| 186 | { |
| 187 | struct obj *obj; |
| 188 | |
| 189 | spin_lock(&global_list_lock); |
| 190 | obj = obj_list_peek(&global_list); |
| 191 | spin_unlock(&global_list_lock); |
| 192 | |
| 193 | if (obj) { |
| 194 | obj->ops->poke(obj); |
| 195 | if (atomic_dec_and_test(&obj->refcnt)) |
| 196 | obj_destroy(obj); |
| 197 | } |
| 198 | } |
| 199 | |
| 200 | void obj_timeout(struct obj *obj) |
| 201 | { |
| 202 | spin_lock(&global_list_lock); |
| 203 | obj_list_del(obj); |
| 204 | spin_unlock(&global_list_lock); |
| 205 | |
| 206 | if (atomic_dec_and_test(&obj->refcnt)) |
| 207 | obj_destroy(obj); |
| 208 | } |
| 209 | |
| 210 | (This is a simplification of the ARP queue management in the |
| 211 | generic neighbour discover code of the networking. Olaf Kirch |
| 212 | found a bug wrt. memory barriers in kfree_skb() that exposed |
| 213 | the atomic_t memory barrier requirements quite clearly.) |
| 214 | |
| 215 | Given the above scheme, it must be the case that the obj->active |
| 216 | update done by the obj list deletion be visible to other processors |
| 217 | before the atomic counter decrement is performed. |
| 218 | |
| 219 | Otherwise, the counter could fall to zero, yet obj->active would still |
| 220 | be set, thus triggering the assertion in obj_destroy(). The error |
| 221 | sequence looks like this: |
| 222 | |
| 223 | cpu 0 cpu 1 |
| 224 | obj_poke() obj_timeout() |
| 225 | obj = obj_list_peek(); |
| 226 | ... gains ref to obj, refcnt=2 |
| 227 | obj_list_del(obj); |
| 228 | obj->active = 0 ... |
| 229 | ... visibility delayed ... |
| 230 | atomic_dec_and_test() |
| 231 | ... refcnt drops to 1 ... |
| 232 | atomic_dec_and_test() |
| 233 | ... refcount drops to 0 ... |
| 234 | obj_destroy() |
| 235 | BUG() triggers since obj->active |
| 236 | still seen as one |
| 237 | obj->active update visibility occurs |
| 238 | |
| 239 | With the memory barrier semantics required of the atomic_t operations |
| 240 | which return values, the above sequence of memory visibility can never |
| 241 | happen. Specifically, in the above case the atomic_dec_and_test() |
| 242 | counter decrement would not become globally visible until the |
| 243 | obj->active update does. |
| 244 | |
| 245 | As a historical note, 32-bit Sparc used to only allow usage of |
| 246 | 24-bits of it's atomic_t type. This was because it used 8 bits |
| 247 | as a spinlock for SMP safety. Sparc32 lacked a "compare and swap" |
| 248 | type instruction. However, 32-bit Sparc has since been moved over |
| 249 | to a "hash table of spinlocks" scheme, that allows the full 32-bit |
| 250 | counter to be realized. Essentially, an array of spinlocks are |
| 251 | indexed into based upon the address of the atomic_t being operated |
| 252 | on, and that lock protects the atomic operation. Parisc uses the |
| 253 | same scheme. |
| 254 | |
| 255 | Another note is that the atomic_t operations returning values are |
| 256 | extremely slow on an old 386. |
| 257 | |
| 258 | We will now cover the atomic bitmask operations. You will find that |
| 259 | their SMP and memory barrier semantics are similar in shape and scope |
| 260 | to the atomic_t ops above. |
| 261 | |
| 262 | Native atomic bit operations are defined to operate on objects aligned |
| 263 | to the size of an "unsigned long" C data type, and are least of that |
| 264 | size. The endianness of the bits within each "unsigned long" are the |
| 265 | native endianness of the cpu. |
| 266 | |
| 267 | void set_bit(unsigned long nr, volatils unsigned long *addr); |
| 268 | void clear_bit(unsigned long nr, volatils unsigned long *addr); |
| 269 | void change_bit(unsigned long nr, volatils unsigned long *addr); |
| 270 | |
| 271 | These routines set, clear, and change, respectively, the bit number |
| 272 | indicated by "nr" on the bit mask pointed to by "ADDR". |
| 273 | |
| 274 | They must execute atomically, yet there are no implicit memory barrier |
| 275 | semantics required of these interfaces. |
| 276 | |
| 277 | int test_and_set_bit(unsigned long nr, volatils unsigned long *addr); |
| 278 | int test_and_clear_bit(unsigned long nr, volatils unsigned long *addr); |
| 279 | int test_and_change_bit(unsigned long nr, volatils unsigned long *addr); |
| 280 | |
| 281 | Like the above, except that these routines return a boolean which |
| 282 | indicates whether the changed bit was set _BEFORE_ the atomic bit |
| 283 | operation. |
| 284 | |
| 285 | WARNING! It is incredibly important that the value be a boolean, |
| 286 | ie. "0" or "1". Do not try to be fancy and save a few instructions by |
| 287 | declaring the above to return "long" and just returning something like |
| 288 | "old_val & mask" because that will not work. |
| 289 | |
| 290 | For one thing, this return value gets truncated to int in many code |
| 291 | paths using these interfaces, so on 64-bit if the bit is set in the |
| 292 | upper 32-bits then testers will never see that. |
| 293 | |
| 294 | One great example of where this problem crops up are the thread_info |
| 295 | flag operations. Routines such as test_and_set_ti_thread_flag() chop |
| 296 | the return value into an int. There are other places where things |
| 297 | like this occur as well. |
| 298 | |
| 299 | These routines, like the atomic_t counter operations returning values, |
| 300 | require explicit memory barrier semantics around their execution. All |
| 301 | memory operations before the atomic bit operation call must be made |
| 302 | visible globally before the atomic bit operation is made visible. |
| 303 | Likewise, the atomic bit operation must be visible globally before any |
| 304 | subsequent memory operation is made visible. For example: |
| 305 | |
| 306 | obj->dead = 1; |
| 307 | if (test_and_set_bit(0, &obj->flags)) |
| 308 | /* ... */; |
| 309 | obj->killed = 1; |
| 310 | |
| 311 | The implementation of test_and_set_bit() must guarentee that |
| 312 | "obj->dead = 1;" is visible to cpus before the atomic memory operation |
| 313 | done by test_and_set_bit() becomes visible. Likewise, the atomic |
| 314 | memory operation done by test_and_set_bit() must become visible before |
| 315 | "obj->killed = 1;" is visible. |
| 316 | |
| 317 | Finally there is the basic operation: |
| 318 | |
| 319 | int test_bit(unsigned long nr, __const__ volatile unsigned long *addr); |
| 320 | |
| 321 | Which returns a boolean indicating if bit "nr" is set in the bitmask |
| 322 | pointed to by "addr". |
| 323 | |
| 324 | If explicit memory barriers are required around clear_bit() (which |
| 325 | does not return a value, and thus does not need to provide memory |
| 326 | barrier semantics), two interfaces are provided: |
| 327 | |
| 328 | void smp_mb__before_clear_bit(void); |
| 329 | void smp_mb__after_clear_bit(void); |
| 330 | |
| 331 | They are used as follows, and are akin to their atomic_t operation |
| 332 | brothers: |
| 333 | |
| 334 | /* All memory operations before this call will |
| 335 | * be globally visible before the clear_bit(). |
| 336 | */ |
| 337 | smp_mb__before_clear_bit(); |
| 338 | clear_bit( ... ); |
| 339 | |
| 340 | /* The clear_bit() will be visible before all |
| 341 | * subsequent memory operations. |
| 342 | */ |
| 343 | smp_mb__after_clear_bit(); |
| 344 | |
| 345 | Finally, there are non-atomic versions of the bitmask operations |
| 346 | provided. They are used in contexts where some other higher-level SMP |
| 347 | locking scheme is being used to protect the bitmask, and thus less |
| 348 | expensive non-atomic operations may be used in the implementation. |
| 349 | They have names similar to the above bitmask operation interfaces, |
| 350 | except that two underscores are prefixed to the interface name. |
| 351 | |
| 352 | void __set_bit(unsigned long nr, volatile unsigned long *addr); |
| 353 | void __clear_bit(unsigned long nr, volatile unsigned long *addr); |
| 354 | void __change_bit(unsigned long nr, volatile unsigned long *addr); |
| 355 | int __test_and_set_bit(unsigned long nr, volatile unsigned long *addr); |
| 356 | int __test_and_clear_bit(unsigned long nr, volatile unsigned long *addr); |
| 357 | int __test_and_change_bit(unsigned long nr, volatile unsigned long *addr); |
| 358 | |
| 359 | These non-atomic variants also do not require any special memory |
| 360 | barrier semantics. |
| 361 | |
| 362 | The routines xchg() and cmpxchg() need the same exact memory barriers |
| 363 | as the atomic and bit operations returning values. |
| 364 | |
| 365 | Spinlocks and rwlocks have memory barrier expectations as well. |
| 366 | The rule to follow is simple: |
| 367 | |
| 368 | 1) When acquiring a lock, the implementation must make it globally |
| 369 | visible before any subsequent memory operation. |
| 370 | |
| 371 | 2) When releasing a lock, the implementation must make it such that |
| 372 | all previous memory operations are globally visible before the |
| 373 | lock release. |
| 374 | |
| 375 | Which finally brings us to _atomic_dec_and_lock(). There is an |
| 376 | architecture-neutral version implemented in lib/dec_and_lock.c, |
| 377 | but most platforms will wish to optimize this in assembler. |
| 378 | |
| 379 | int _atomic_dec_and_lock(atomic_t *atomic, spinlock_t *lock); |
| 380 | |
| 381 | Atomically decrement the given counter, and if will drop to zero |
| 382 | atomically acquire the given spinlock and perform the decrement |
| 383 | of the counter to zero. If it does not drop to zero, do nothing |
| 384 | with the spinlock. |
| 385 | |
| 386 | It is actually pretty simple to get the memory barrier correct. |
| 387 | Simply satisfy the spinlock grab requirements, which is make |
| 388 | sure the spinlock operation is globally visible before any |
| 389 | subsequent memory operation. |
| 390 | |
| 391 | We can demonstrate this operation more clearly if we define |
| 392 | an abstract atomic operation: |
| 393 | |
| 394 | long cas(long *mem, long old, long new); |
| 395 | |
| 396 | "cas" stands for "compare and swap". It atomically: |
| 397 | |
| 398 | 1) Compares "old" with the value currently at "mem". |
| 399 | 2) If they are equal, "new" is written to "mem". |
| 400 | 3) Regardless, the current value at "mem" is returned. |
| 401 | |
| 402 | As an example usage, here is what an atomic counter update |
| 403 | might look like: |
| 404 | |
| 405 | void example_atomic_inc(long *counter) |
| 406 | { |
| 407 | long old, new, ret; |
| 408 | |
| 409 | while (1) { |
| 410 | old = *counter; |
| 411 | new = old + 1; |
| 412 | |
| 413 | ret = cas(counter, old, new); |
| 414 | if (ret == old) |
| 415 | break; |
| 416 | } |
| 417 | } |
| 418 | |
| 419 | Let's use cas() in order to build a pseudo-C atomic_dec_and_lock(): |
| 420 | |
| 421 | int _atomic_dec_and_lock(atomic_t *atomic, spinlock_t *lock) |
| 422 | { |
| 423 | long old, new, ret; |
| 424 | int went_to_zero; |
| 425 | |
| 426 | went_to_zero = 0; |
| 427 | while (1) { |
| 428 | old = atomic_read(atomic); |
| 429 | new = old - 1; |
| 430 | if (new == 0) { |
| 431 | went_to_zero = 1; |
| 432 | spin_lock(lock); |
| 433 | } |
| 434 | ret = cas(atomic, old, new); |
| 435 | if (ret == old) |
| 436 | break; |
| 437 | if (went_to_zero) { |
| 438 | spin_unlock(lock); |
| 439 | went_to_zero = 0; |
| 440 | } |
| 441 | } |
| 442 | |
| 443 | return went_to_zero; |
| 444 | } |
| 445 | |
| 446 | Now, as far as memory barriers go, as long as spin_lock() |
| 447 | strictly orders all subsequent memory operations (including |
| 448 | the cas()) with respect to itself, things will be fine. |
| 449 | |
| 450 | Said another way, _atomic_dec_and_lock() must guarentee that |
| 451 | a counter dropping to zero is never made visible before the |
| 452 | spinlock being acquired. |
| 453 | |
| 454 | Note that this also means that for the case where the counter |
| 455 | is not dropping to zero, there are no memory ordering |
| 456 | requirements. |