1 //===-- atomic.c - Implement support functions for atomic operations.------===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 //  atomic.c defines a set of functions for performing atomic accesses on
10 //  arbitrary-sized memory locations.  This design uses locks that should
11 //  be fast in the uncontended case, for two reasons:
12 //
13 //  1) This code must work with C programs that do not link to anything
14 //     (including pthreads) and so it should not depend on any pthread
15 //     functions.
16 //  2) Atomic operations, rather than explicit mutexes, are most commonly used
17 //     on code where contended operations are rate.
18 //
19 //  To avoid needing a per-object lock, this code allocates an array of
20 //  locks and hashes the object pointers to find the one that it should use.
21 //  For operations that must be atomic on two locations, the lower lock is
22 //  always acquired first, to avoid deadlock.
23 //
24 //===----------------------------------------------------------------------===//
25 
26 #include <stdbool.h>
27 #include <stdint.h>
28 #include <string.h>
29 
30 #include "assembly.h"
31 
32 // Clang objects if you redefine a builtin.  This little hack allows us to
33 // define a function with the same name as an intrinsic.
34 #pragma redefine_extname __atomic_load_c SYMBOL_NAME(__atomic_load)
35 #pragma redefine_extname __atomic_store_c SYMBOL_NAME(__atomic_store)
36 #pragma redefine_extname __atomic_exchange_c SYMBOL_NAME(__atomic_exchange)
37 #pragma redefine_extname __atomic_compare_exchange_c SYMBOL_NAME(              \
38     __atomic_compare_exchange)
39 
40 /// Number of locks.  This allocates one page on 32-bit platforms, two on
41 /// 64-bit.  This can be specified externally if a different trade between
42 /// memory usage and contention probability is required for a given platform.
43 #ifndef SPINLOCK_COUNT
44 #define SPINLOCK_COUNT (1 << 10)
45 #endif
46 static const long SPINLOCK_MASK = SPINLOCK_COUNT - 1;
47 
48 ////////////////////////////////////////////////////////////////////////////////
49 // Platform-specific lock implementation.  Falls back to spinlocks if none is
50 // defined.  Each platform should define the Lock type, and corresponding
51 // lock() and unlock() functions.
52 ////////////////////////////////////////////////////////////////////////////////
53 #ifdef __FreeBSD__
54 #include <errno.h>
55 // clang-format off
56 #include <sys/types.h>
57 #include <machine/atomic.h>
58 #include <sys/umtx.h>
59 // clang-format on
60 typedef struct _usem Lock;
unlock(Lock * l)61 __inline static void unlock(Lock *l) {
62   __c11_atomic_store((_Atomic(uint32_t) *)&l->_count, 1, __ATOMIC_RELEASE);
63   __c11_atomic_thread_fence(__ATOMIC_SEQ_CST);
64   if (l->_has_waiters)
65     _umtx_op(l, UMTX_OP_SEM_WAKE, 1, 0, 0);
66 }
lock(Lock * l)67 __inline static void lock(Lock *l) {
68   uint32_t old = 1;
69   while (!__c11_atomic_compare_exchange_weak((_Atomic(uint32_t) *)&l->_count,
70                                              &old, 0, __ATOMIC_ACQUIRE,
71                                              __ATOMIC_RELAXED)) {
72     _umtx_op(l, UMTX_OP_SEM_WAIT, 0, 0, 0);
73     old = 1;
74   }
75 }
76 /// locks for atomic operations
77 static Lock locks[SPINLOCK_COUNT] = {[0 ... SPINLOCK_COUNT - 1] = {0, 1, 0}};
78 
79 #elif defined(__APPLE__)
80 #include <libkern/OSAtomic.h>
81 typedef OSSpinLock Lock;
unlock(Lock * l)82 __inline static void unlock(Lock *l) { OSSpinLockUnlock(l); }
83 /// Locks a lock.  In the current implementation, this is potentially
84 /// unbounded in the contended case.
lock(Lock * l)85 __inline static void lock(Lock *l) { OSSpinLockLock(l); }
86 static Lock locks[SPINLOCK_COUNT]; // initialized to OS_SPINLOCK_INIT which is 0
87 
88 #else
89 typedef _Atomic(uintptr_t) Lock;
90 /// Unlock a lock.  This is a release operation.
unlock(Lock * l)91 __inline static void unlock(Lock *l) {
92   __c11_atomic_store(l, 0, __ATOMIC_RELEASE);
93 }
94 /// Locks a lock.  In the current implementation, this is potentially
95 /// unbounded in the contended case.
lock(Lock * l)96 __inline static void lock(Lock *l) {
97   uintptr_t old = 0;
98   while (!__c11_atomic_compare_exchange_weak(l, &old, 1, __ATOMIC_ACQUIRE,
99                                              __ATOMIC_RELAXED))
100     old = 0;
101 }
102 /// locks for atomic operations
103 static Lock locks[SPINLOCK_COUNT];
104 #endif
105 
106 /// Returns a lock to use for a given pointer.
lock_for_pointer(void * ptr)107 static __inline Lock *lock_for_pointer(void *ptr) {
108   intptr_t hash = (intptr_t)ptr;
109   // Disregard the lowest 4 bits.  We want all values that may be part of the
110   // same memory operation to hash to the same value and therefore use the same
111   // lock.
112   hash >>= 4;
113   // Use the next bits as the basis for the hash
114   intptr_t low = hash & SPINLOCK_MASK;
115   // Now use the high(er) set of bits to perturb the hash, so that we don't
116   // get collisions from atomic fields in a single object
117   hash >>= 16;
118   hash ^= low;
119   // Return a pointer to the word to use
120   return locks + (hash & SPINLOCK_MASK);
121 }
122 
123 /// Macros for determining whether a size is lock free.
124 #define ATOMIC_ALWAYS_LOCK_FREE_OR_ALIGNED_LOCK_FREE(size, p)                  \
125   (__atomic_always_lock_free(size, p) ||                                       \
126    (__atomic_always_lock_free(size, 0) && ((uintptr_t)p % size) == 0))
127 #define IS_LOCK_FREE_1(p) ATOMIC_ALWAYS_LOCK_FREE_OR_ALIGNED_LOCK_FREE(1, p)
128 #define IS_LOCK_FREE_2(p) ATOMIC_ALWAYS_LOCK_FREE_OR_ALIGNED_LOCK_FREE(2, p)
129 #define IS_LOCK_FREE_4(p) ATOMIC_ALWAYS_LOCK_FREE_OR_ALIGNED_LOCK_FREE(4, p)
130 #define IS_LOCK_FREE_8(p) ATOMIC_ALWAYS_LOCK_FREE_OR_ALIGNED_LOCK_FREE(8, p)
131 #define IS_LOCK_FREE_16(p) ATOMIC_ALWAYS_LOCK_FREE_OR_ALIGNED_LOCK_FREE(16, p)
132 
133 /// Macro that calls the compiler-generated lock-free versions of functions
134 /// when they exist.
135 #define TRY_LOCK_FREE_CASE(n, type, ptr)                                       \
136   case n:                                                                      \
137     if (IS_LOCK_FREE_##n(ptr)) {                                               \
138       LOCK_FREE_ACTION(type);                                                  \
139     }                                                                          \
140     break;
141 #ifdef __SIZEOF_INT128__
142 #define TRY_LOCK_FREE_CASE_16(p) TRY_LOCK_FREE_CASE(16, __uint128_t, p)
143 #else
144 #define TRY_LOCK_FREE_CASE_16(p) /* __uint128_t not available */
145 #endif
146 
147 #define LOCK_FREE_CASES(ptr)                                                   \
148   do {                                                                         \
149     switch (size) {                                                            \
150       TRY_LOCK_FREE_CASE(1, uint8_t, ptr)                                      \
151       TRY_LOCK_FREE_CASE(2, uint16_t, ptr)                                     \
152       TRY_LOCK_FREE_CASE(4, uint32_t, ptr)                                     \
153       TRY_LOCK_FREE_CASE(8, uint64_t, ptr)                                     \
154       TRY_LOCK_FREE_CASE_16(ptr) /* __uint128_t may not be supported */        \
155     default:                                                                   \
156       break;                                                                   \
157     }                                                                          \
158   } while (0)
159 
160 /// An atomic load operation.  This is atomic with respect to the source
161 /// pointer only.
__atomic_load_c(int size,void * src,void * dest,int model)162 void __atomic_load_c(int size, void *src, void *dest, int model) {
163 #define LOCK_FREE_ACTION(type)                                                 \
164   *((type *)dest) = __c11_atomic_load((_Atomic(type) *)src, model);            \
165   return;
166   LOCK_FREE_CASES(src);
167 #undef LOCK_FREE_ACTION
168   Lock *l = lock_for_pointer(src);
169   lock(l);
170   memcpy(dest, src, size);
171   unlock(l);
172 }
173 
174 /// An atomic store operation.  This is atomic with respect to the destination
175 /// pointer only.
__atomic_store_c(int size,void * dest,void * src,int model)176 void __atomic_store_c(int size, void *dest, void *src, int model) {
177 #define LOCK_FREE_ACTION(type)                                                 \
178   __c11_atomic_store((_Atomic(type) *)dest, *(type *)src, model);              \
179   return;
180   LOCK_FREE_CASES(dest);
181 #undef LOCK_FREE_ACTION
182   Lock *l = lock_for_pointer(dest);
183   lock(l);
184   memcpy(dest, src, size);
185   unlock(l);
186 }
187 
188 /// Atomic compare and exchange operation.  If the value at *ptr is identical
189 /// to the value at *expected, then this copies value at *desired to *ptr.  If
190 /// they  are not, then this stores the current value from *ptr in *expected.
191 ///
192 /// This function returns 1 if the exchange takes place or 0 if it fails.
__atomic_compare_exchange_c(int size,void * ptr,void * expected,void * desired,int success,int failure)193 int __atomic_compare_exchange_c(int size, void *ptr, void *expected,
194                                 void *desired, int success, int failure) {
195 #define LOCK_FREE_ACTION(type)                                                 \
196   return __c11_atomic_compare_exchange_strong(                                 \
197       (_Atomic(type) *)ptr, (type *)expected, *(type *)desired, success,       \
198       failure)
199   LOCK_FREE_CASES(ptr);
200 #undef LOCK_FREE_ACTION
201   Lock *l = lock_for_pointer(ptr);
202   lock(l);
203   if (memcmp(ptr, expected, size) == 0) {
204     memcpy(ptr, desired, size);
205     unlock(l);
206     return 1;
207   }
208   memcpy(expected, ptr, size);
209   unlock(l);
210   return 0;
211 }
212 
213 /// Performs an atomic exchange operation between two pointers.  This is atomic
214 /// with respect to the target address.
__atomic_exchange_c(int size,void * ptr,void * val,void * old,int model)215 void __atomic_exchange_c(int size, void *ptr, void *val, void *old, int model) {
216 #define LOCK_FREE_ACTION(type)                                                 \
217   *(type *)old =                                                               \
218       __c11_atomic_exchange((_Atomic(type) *)ptr, *(type *)val, model);        \
219   return;
220   LOCK_FREE_CASES(ptr);
221 #undef LOCK_FREE_ACTION
222   Lock *l = lock_for_pointer(ptr);
223   lock(l);
224   memcpy(old, ptr, size);
225   memcpy(ptr, val, size);
226   unlock(l);
227 }
228 
229 ////////////////////////////////////////////////////////////////////////////////
230 // Where the size is known at compile time, the compiler may emit calls to
231 // specialised versions of the above functions.
232 ////////////////////////////////////////////////////////////////////////////////
233 #ifdef __SIZEOF_INT128__
234 #define OPTIMISED_CASES                                                        \
235   OPTIMISED_CASE(1, IS_LOCK_FREE_1, uint8_t)                                   \
236   OPTIMISED_CASE(2, IS_LOCK_FREE_2, uint16_t)                                  \
237   OPTIMISED_CASE(4, IS_LOCK_FREE_4, uint32_t)                                  \
238   OPTIMISED_CASE(8, IS_LOCK_FREE_8, uint64_t)                                  \
239   OPTIMISED_CASE(16, IS_LOCK_FREE_16, __uint128_t)
240 #else
241 #define OPTIMISED_CASES                                                        \
242   OPTIMISED_CASE(1, IS_LOCK_FREE_1, uint8_t)                                   \
243   OPTIMISED_CASE(2, IS_LOCK_FREE_2, uint16_t)                                  \
244   OPTIMISED_CASE(4, IS_LOCK_FREE_4, uint32_t)                                  \
245   OPTIMISED_CASE(8, IS_LOCK_FREE_8, uint64_t)
246 #endif
247 
248 #define OPTIMISED_CASE(n, lockfree, type)                                      \
249   type __atomic_load_##n(type *src, int model) {                               \
250     if (lockfree(src))                                                         \
251       return __c11_atomic_load((_Atomic(type) *)src, model);                   \
252     Lock *l = lock_for_pointer(src);                                           \
253     lock(l);                                                                   \
254     type val = *src;                                                           \
255     unlock(l);                                                                 \
256     return val;                                                                \
257   }
258 OPTIMISED_CASES
259 #undef OPTIMISED_CASE
260 
261 #define OPTIMISED_CASE(n, lockfree, type)                                      \
262   void __atomic_store_##n(type *dest, type val, int model) {                   \
263     if (lockfree(dest)) {                                                      \
264       __c11_atomic_store((_Atomic(type) *)dest, val, model);                   \
265       return;                                                                  \
266     }                                                                          \
267     Lock *l = lock_for_pointer(dest);                                          \
268     lock(l);                                                                   \
269     *dest = val;                                                               \
270     unlock(l);                                                                 \
271     return;                                                                    \
272   }
273 OPTIMISED_CASES
274 #undef OPTIMISED_CASE
275 
276 #define OPTIMISED_CASE(n, lockfree, type)                                      \
277   type __atomic_exchange_##n(type *dest, type val, int model) {                \
278     if (lockfree(dest))                                                        \
279       return __c11_atomic_exchange((_Atomic(type) *)dest, val, model);         \
280     Lock *l = lock_for_pointer(dest);                                          \
281     lock(l);                                                                   \
282     type tmp = *dest;                                                          \
283     *dest = val;                                                               \
284     unlock(l);                                                                 \
285     return tmp;                                                                \
286   }
287 OPTIMISED_CASES
288 #undef OPTIMISED_CASE
289 
290 #define OPTIMISED_CASE(n, lockfree, type)                                      \
291   bool __atomic_compare_exchange_##n(type *ptr, type *expected, type desired,  \
292                                      int success, int failure) {               \
293     if (lockfree(ptr))                                                         \
294       return __c11_atomic_compare_exchange_strong(                             \
295           (_Atomic(type) *)ptr, expected, desired, success, failure);          \
296     Lock *l = lock_for_pointer(ptr);                                           \
297     lock(l);                                                                   \
298     if (*ptr == *expected) {                                                   \
299       *ptr = desired;                                                          \
300       unlock(l);                                                               \
301       return true;                                                             \
302     }                                                                          \
303     *expected = *ptr;                                                          \
304     unlock(l);                                                                 \
305     return false;                                                              \
306   }
307 OPTIMISED_CASES
308 #undef OPTIMISED_CASE
309 
310 ////////////////////////////////////////////////////////////////////////////////
311 // Atomic read-modify-write operations for integers of various sizes.
312 ////////////////////////////////////////////////////////////////////////////////
313 #define ATOMIC_RMW(n, lockfree, type, opname, op)                              \
314   type __atomic_fetch_##opname##_##n(type *ptr, type val, int model) {         \
315     if (lockfree(ptr))                                                         \
316       return __c11_atomic_fetch_##opname((_Atomic(type) *)ptr, val, model);    \
317     Lock *l = lock_for_pointer(ptr);                                           \
318     lock(l);                                                                   \
319     type tmp = *ptr;                                                           \
320     *ptr = tmp op val;                                                         \
321     unlock(l);                                                                 \
322     return tmp;                                                                \
323   }
324 
325 #define OPTIMISED_CASE(n, lockfree, type) ATOMIC_RMW(n, lockfree, type, add, +)
326 OPTIMISED_CASES
327 #undef OPTIMISED_CASE
328 #define OPTIMISED_CASE(n, lockfree, type) ATOMIC_RMW(n, lockfree, type, sub, -)
329 OPTIMISED_CASES
330 #undef OPTIMISED_CASE
331 #define OPTIMISED_CASE(n, lockfree, type) ATOMIC_RMW(n, lockfree, type, and, &)
332 OPTIMISED_CASES
333 #undef OPTIMISED_CASE
334 #define OPTIMISED_CASE(n, lockfree, type) ATOMIC_RMW(n, lockfree, type, or, |)
335 OPTIMISED_CASES
336 #undef OPTIMISED_CASE
337 #define OPTIMISED_CASE(n, lockfree, type) ATOMIC_RMW(n, lockfree, type, xor, ^)
338 OPTIMISED_CASES
339 #undef OPTIMISED_CASE
340