ARCv2: Implement atomic64 based on LLOCKD/SCONDD instructions
[cascardo/linux.git] / arch / arc / include / asm / atomic.h
1 /*
2  * Copyright (C) 2004, 2007-2010, 2011-2012 Synopsys, Inc. (www.synopsys.com)
3  *
4  * This program is free software; you can redistribute it and/or modify
5  * it under the terms of the GNU General Public License version 2 as
6  * published by the Free Software Foundation.
7  */
8
9 #ifndef _ASM_ARC_ATOMIC_H
10 #define _ASM_ARC_ATOMIC_H
11
12 #ifndef __ASSEMBLY__
13
14 #include <linux/types.h>
15 #include <linux/compiler.h>
16 #include <asm/cmpxchg.h>
17 #include <asm/barrier.h>
18 #include <asm/smp.h>
19
20 #ifndef CONFIG_ARC_PLAT_EZNPS
21
22 #define atomic_read(v)  READ_ONCE((v)->counter)
23 #define ATOMIC_INIT(i)  { (i) }
24
25 #ifdef CONFIG_ARC_HAS_LLSC
26
27 #define atomic_set(v, i) WRITE_ONCE(((v)->counter), (i))
28
29 #define ATOMIC_OP(op, c_op, asm_op)                                     \
30 static inline void atomic_##op(int i, atomic_t *v)                      \
31 {                                                                       \
32         unsigned int val;                                               \
33                                                                         \
34         __asm__ __volatile__(                                           \
35         "1:     llock   %[val], [%[ctr]]                \n"             \
36         "       " #asm_op " %[val], %[val], %[i]        \n"             \
37         "       scond   %[val], [%[ctr]]                \n"             \
38         "       bnz     1b                              \n"             \
39         : [val] "=&r"   (val) /* Early clobber to prevent reg reuse */  \
40         : [ctr] "r"     (&v->counter), /* Not "m": llock only supports reg direct addr mode */  \
41           [i]   "ir"    (i)                                             \
42         : "cc");                                                        \
43 }                                                                       \
44
45 #define ATOMIC_OP_RETURN(op, c_op, asm_op)                              \
46 static inline int atomic_##op##_return(int i, atomic_t *v)              \
47 {                                                                       \
48         unsigned int val;                                               \
49                                                                         \
50         /*                                                              \
51          * Explicit full memory barrier needed before/after as          \
52          * LLOCK/SCOND thmeselves don't provide any such semantics      \
53          */                                                             \
54         smp_mb();                                                       \
55                                                                         \
56         __asm__ __volatile__(                                           \
57         "1:     llock   %[val], [%[ctr]]                \n"             \
58         "       " #asm_op " %[val], %[val], %[i]        \n"             \
59         "       scond   %[val], [%[ctr]]                \n"             \
60         "       bnz     1b                              \n"             \
61         : [val] "=&r"   (val)                                           \
62         : [ctr] "r"     (&v->counter),                                  \
63           [i]   "ir"    (i)                                             \
64         : "cc");                                                        \
65                                                                         \
66         smp_mb();                                                       \
67                                                                         \
68         return val;                                                     \
69 }
70
71 #define ATOMIC_FETCH_OP(op, c_op, asm_op)                               \
72 static inline int atomic_fetch_##op(int i, atomic_t *v)                 \
73 {                                                                       \
74         unsigned int val, orig;                                         \
75                                                                         \
76         /*                                                              \
77          * Explicit full memory barrier needed before/after as          \
78          * LLOCK/SCOND thmeselves don't provide any such semantics      \
79          */                                                             \
80         smp_mb();                                                       \
81                                                                         \
82         __asm__ __volatile__(                                           \
83         "1:     llock   %[orig], [%[ctr]]               \n"             \
84         "       " #asm_op " %[val], %[orig], %[i]       \n"             \
85         "       scond   %[val], [%[ctr]]                \n"             \
86         "                                               \n"             \
87         : [val] "=&r"   (val),                                          \
88           [orig] "=&r" (orig)                                           \
89         : [ctr] "r"     (&v->counter),                                  \
90           [i]   "ir"    (i)                                             \
91         : "cc");                                                        \
92                                                                         \
93         smp_mb();                                                       \
94                                                                         \
95         return orig;                                                    \
96 }
97
98 #else   /* !CONFIG_ARC_HAS_LLSC */
99
100 #ifndef CONFIG_SMP
101
102  /* violating atomic_xxx API locking protocol in UP for optimization sake */
103 #define atomic_set(v, i) WRITE_ONCE(((v)->counter), (i))
104
105 #else
106
107 static inline void atomic_set(atomic_t *v, int i)
108 {
109         /*
110          * Independent of hardware support, all of the atomic_xxx() APIs need
111          * to follow the same locking rules to make sure that a "hardware"
112          * atomic insn (e.g. LD) doesn't clobber an "emulated" atomic insn
113          * sequence
114          *
115          * Thus atomic_set() despite being 1 insn (and seemingly atomic)
116          * requires the locking.
117          */
118         unsigned long flags;
119
120         atomic_ops_lock(flags);
121         WRITE_ONCE(v->counter, i);
122         atomic_ops_unlock(flags);
123 }
124
125 #endif
126
127 /*
128  * Non hardware assisted Atomic-R-M-W
129  * Locking would change to irq-disabling only (UP) and spinlocks (SMP)
130  */
131
132 #define ATOMIC_OP(op, c_op, asm_op)                                     \
133 static inline void atomic_##op(int i, atomic_t *v)                      \
134 {                                                                       \
135         unsigned long flags;                                            \
136                                                                         \
137         atomic_ops_lock(flags);                                         \
138         v->counter c_op i;                                              \
139         atomic_ops_unlock(flags);                                       \
140 }
141
142 #define ATOMIC_OP_RETURN(op, c_op, asm_op)                              \
143 static inline int atomic_##op##_return(int i, atomic_t *v)              \
144 {                                                                       \
145         unsigned long flags;                                            \
146         unsigned long temp;                                             \
147                                                                         \
148         /*                                                              \
149          * spin lock/unlock provides the needed smp_mb() before/after   \
150          */                                                             \
151         atomic_ops_lock(flags);                                         \
152         temp = v->counter;                                              \
153         temp c_op i;                                                    \
154         v->counter = temp;                                              \
155         atomic_ops_unlock(flags);                                       \
156                                                                         \
157         return temp;                                                    \
158 }
159
160 #define ATOMIC_FETCH_OP(op, c_op, asm_op)                               \
161 static inline int atomic_fetch_##op(int i, atomic_t *v)                 \
162 {                                                                       \
163         unsigned long flags;                                            \
164         unsigned long orig;                                             \
165                                                                         \
166         /*                                                              \
167          * spin lock/unlock provides the needed smp_mb() before/after   \
168          */                                                             \
169         atomic_ops_lock(flags);                                         \
170         orig = v->counter;                                              \
171         v->counter c_op i;                                              \
172         atomic_ops_unlock(flags);                                       \
173                                                                         \
174         return orig;                                                    \
175 }
176
177 #endif /* !CONFIG_ARC_HAS_LLSC */
178
179 #define ATOMIC_OPS(op, c_op, asm_op)                                    \
180         ATOMIC_OP(op, c_op, asm_op)                                     \
181         ATOMIC_OP_RETURN(op, c_op, asm_op)                              \
182         ATOMIC_FETCH_OP(op, c_op, asm_op)
183
184 ATOMIC_OPS(add, +=, add)
185 ATOMIC_OPS(sub, -=, sub)
186
187 #define atomic_andnot atomic_andnot
188
189 #undef ATOMIC_OPS
190 #define ATOMIC_OPS(op, c_op, asm_op)                                    \
191         ATOMIC_OP(op, c_op, asm_op)                                     \
192         ATOMIC_FETCH_OP(op, c_op, asm_op)
193
194 ATOMIC_OPS(and, &=, and)
195 ATOMIC_OPS(andnot, &= ~, bic)
196 ATOMIC_OPS(or, |=, or)
197 ATOMIC_OPS(xor, ^=, xor)
198
199 #else /* CONFIG_ARC_PLAT_EZNPS */
200
201 static inline int atomic_read(const atomic_t *v)
202 {
203         int temp;
204
205         __asm__ __volatile__(
206         "       ld.di %0, [%1]"
207         : "=r"(temp)
208         : "r"(&v->counter)
209         : "memory");
210         return temp;
211 }
212
213 static inline void atomic_set(atomic_t *v, int i)
214 {
215         __asm__ __volatile__(
216         "       st.di %0,[%1]"
217         :
218         : "r"(i), "r"(&v->counter)
219         : "memory");
220 }
221
222 #define ATOMIC_OP(op, c_op, asm_op)                                     \
223 static inline void atomic_##op(int i, atomic_t *v)                      \
224 {                                                                       \
225         __asm__ __volatile__(                                           \
226         "       mov r2, %0\n"                                           \
227         "       mov r3, %1\n"                                           \
228         "       .word %2\n"                                             \
229         :                                                               \
230         : "r"(i), "r"(&v->counter), "i"(asm_op)                         \
231         : "r2", "r3", "memory");                                        \
232 }                                                                       \
233
234 #define ATOMIC_OP_RETURN(op, c_op, asm_op)                              \
235 static inline int atomic_##op##_return(int i, atomic_t *v)              \
236 {                                                                       \
237         unsigned int temp = i;                                          \
238                                                                         \
239         /* Explicit full memory barrier needed before/after */          \
240         smp_mb();                                                       \
241                                                                         \
242         __asm__ __volatile__(                                           \
243         "       mov r2, %0\n"                                           \
244         "       mov r3, %1\n"                                           \
245         "       .word %2\n"                                             \
246         "       mov %0, r2"                                             \
247         : "+r"(temp)                                                    \
248         : "r"(&v->counter), "i"(asm_op)                                 \
249         : "r2", "r3", "memory");                                        \
250                                                                         \
251         smp_mb();                                                       \
252                                                                         \
253         temp c_op i;                                                    \
254                                                                         \
255         return temp;                                                    \
256 }
257
258 #define ATOMIC_FETCH_OP(op, c_op, asm_op)                               \
259 static inline int atomic_fetch_##op(int i, atomic_t *v)                 \
260 {                                                                       \
261         unsigned int temp = i;                                          \
262                                                                         \
263         /* Explicit full memory barrier needed before/after */          \
264         smp_mb();                                                       \
265                                                                         \
266         __asm__ __volatile__(                                           \
267         "       mov r2, %0\n"                                           \
268         "       mov r3, %1\n"                                           \
269         "       .word %2\n"                                             \
270         "       mov %0, r2"                                             \
271         : "+r"(temp)                                                    \
272         : "r"(&v->counter), "i"(asm_op)                                 \
273         : "r2", "r3", "memory");                                        \
274                                                                         \
275         smp_mb();                                                       \
276                                                                         \
277         return temp;                                                    \
278 }
279
280 #define ATOMIC_OPS(op, c_op, asm_op)                                    \
281         ATOMIC_OP(op, c_op, asm_op)                                     \
282         ATOMIC_OP_RETURN(op, c_op, asm_op)                              \
283         ATOMIC_FETCH_OP(op, c_op, asm_op)
284
285 ATOMIC_OPS(add, +=, CTOP_INST_AADD_DI_R2_R2_R3)
286 #define atomic_sub(i, v) atomic_add(-(i), (v))
287 #define atomic_sub_return(i, v) atomic_add_return(-(i), (v))
288
289 #undef ATOMIC_OPS
290 #define ATOMIC_OPS(op, c_op, asm_op)                                    \
291         ATOMIC_OP(op, c_op, asm_op)                                     \
292         ATOMIC_FETCH_OP(op, c_op, asm_op)
293
294 ATOMIC_OPS(and, &=, CTOP_INST_AAND_DI_R2_R2_R3)
295 #define atomic_andnot(mask, v) atomic_and(~(mask), (v))
296 ATOMIC_OPS(or, |=, CTOP_INST_AOR_DI_R2_R2_R3)
297 ATOMIC_OPS(xor, ^=, CTOP_INST_AXOR_DI_R2_R2_R3)
298
299 #endif /* CONFIG_ARC_PLAT_EZNPS */
300
301 #undef ATOMIC_OPS
302 #undef ATOMIC_FETCH_OP
303 #undef ATOMIC_OP_RETURN
304 #undef ATOMIC_OP
305
306 /**
307  * __atomic_add_unless - add unless the number is a given value
308  * @v: pointer of type atomic_t
309  * @a: the amount to add to v...
310  * @u: ...unless v is equal to u.
311  *
312  * Atomically adds @a to @v, so long as it was not @u.
313  * Returns the old value of @v
314  */
315 #define __atomic_add_unless(v, a, u)                                    \
316 ({                                                                      \
317         int c, old;                                                     \
318                                                                         \
319         /*                                                              \
320          * Explicit full memory barrier needed before/after as          \
321          * LLOCK/SCOND thmeselves don't provide any such semantics      \
322          */                                                             \
323         smp_mb();                                                       \
324                                                                         \
325         c = atomic_read(v);                                             \
326         while (c != (u) && (old = atomic_cmpxchg((v), c, c + (a))) != c)\
327                 c = old;                                                \
328                                                                         \
329         smp_mb();                                                       \
330                                                                         \
331         c;                                                              \
332 })
333
334 #define atomic_inc_not_zero(v)          atomic_add_unless((v), 1, 0)
335
336 #define atomic_inc(v)                   atomic_add(1, v)
337 #define atomic_dec(v)                   atomic_sub(1, v)
338
339 #define atomic_inc_and_test(v)          (atomic_add_return(1, v) == 0)
340 #define atomic_dec_and_test(v)          (atomic_sub_return(1, v) == 0)
341 #define atomic_inc_return(v)            atomic_add_return(1, (v))
342 #define atomic_dec_return(v)            atomic_sub_return(1, (v))
343 #define atomic_sub_and_test(i, v)       (atomic_sub_return(i, v) == 0)
344
345 #define atomic_add_negative(i, v)       (atomic_add_return(i, v) < 0)
346
347
348 #ifdef CONFIG_GENERIC_ATOMIC64
349
350 #include <asm-generic/atomic64.h>
351
352 #else   /* Kconfig ensures this is only enabled with needed h/w assist */
353
354 /*
355  * ARCv2 supports 64-bit exclusive load (LLOCKD) / store (SCONDD)
356  *  - The address HAS to be 64-bit aligned
357  *  - There are 2 semantics involved here:
358  *    = exclusive implies no interim update between load/store to same addr
359  *    = both words are observed/updated together: this is guaranteed even
360  *      for regular 64-bit load (LDD) / store (STD). Thus atomic64_set()
361  *      is NOT required to use LLOCKD+SCONDD, STD suffices
362  */
363
364 typedef struct {
365         aligned_u64 counter;
366 } atomic64_t;
367
368 #define ATOMIC64_INIT(a) { (a) }
369
370 static inline long long atomic64_read(const atomic64_t *v)
371 {
372         unsigned long long val;
373
374         __asm__ __volatile__(
375         "       ldd   %0, [%1]  \n"
376         : "=r"(val)
377         : "r"(&v->counter));
378
379         return val;
380 }
381
382 static inline void atomic64_set(atomic64_t *v, long long a)
383 {
384         /*
385          * This could have been a simple assignment in "C" but would need
386          * explicit volatile. Otherwise gcc optimizers could elide the store
387          * which borked atomic64 self-test
388          * In the inline asm version, memory clobber needed for exact same
389          * reason, to tell gcc about the store.
390          *
391          * This however is not needed for sibling atomic64_add() etc since both
392          * load/store are explicitly done in inline asm. As long as API is used
393          * for each access, gcc has no way to optimize away any load/store
394          */
395         __asm__ __volatile__(
396         "       std   %0, [%1]  \n"
397         :
398         : "r"(a), "r"(&v->counter)
399         : "memory");
400 }
401
402 #define ATOMIC64_OP(op, op1, op2)                                       \
403 static inline void atomic64_##op(long long a, atomic64_t *v)            \
404 {                                                                       \
405         unsigned long long val;                                         \
406                                                                         \
407         __asm__ __volatile__(                                           \
408         "1:                             \n"                             \
409         "       llockd  %0, [%1]        \n"                             \
410         "       " #op1 " %L0, %L0, %L2  \n"                             \
411         "       " #op2 " %H0, %H0, %H2  \n"                             \
412         "       scondd   %0, [%1]       \n"                             \
413         "       bnz     1b              \n"                             \
414         : "=&r"(val)                                                    \
415         : "r"(&v->counter), "ir"(a)                                     \
416         : "cc");                                                \
417 }                                                                       \
418
419 #define ATOMIC64_OP_RETURN(op, op1, op2)                                \
420 static inline long long atomic64_##op##_return(long long a, atomic64_t *v)      \
421 {                                                                       \
422         unsigned long long val;                                         \
423                                                                         \
424         smp_mb();                                                       \
425                                                                         \
426         __asm__ __volatile__(                                           \
427         "1:                             \n"                             \
428         "       llockd   %0, [%1]       \n"                             \
429         "       " #op1 " %L0, %L0, %L2  \n"                             \
430         "       " #op2 " %H0, %H0, %H2  \n"                             \
431         "       scondd   %0, [%1]       \n"                             \
432         "       bnz     1b              \n"                             \
433         : [val] "=&r"(val)                                              \
434         : "r"(&v->counter), "ir"(a)                                     \
435         : "cc");        /* memory clobber comes from smp_mb() */        \
436                                                                         \
437         smp_mb();                                                       \
438                                                                         \
439         return val;                                                     \
440 }
441
442 #define ATOMIC64_FETCH_OP(op, op1, op2)                                 \
443 static inline long long atomic64_fetch_##op(long long a, atomic64_t *v) \
444 {                                                                       \
445         unsigned long long val, orig;                                   \
446                                                                         \
447         smp_mb();                                                       \
448                                                                         \
449         __asm__ __volatile__(                                           \
450         "1:                             \n"                             \
451         "       llockd   %0, [%2]       \n"                             \
452         "       " #op1 " %L1, %L0, %L3  \n"                             \
453         "       " #op2 " %H1, %H0, %H3  \n"                             \
454         "       scondd   %1, [%2]       \n"                             \
455         "       bnz     1b              \n"                             \
456         : "=&r"(orig), "=&r"(val)                                       \
457         : "r"(&v->counter), "ir"(a)                                     \
458         : "cc");        /* memory clobber comes from smp_mb() */        \
459                                                                         \
460         smp_mb();                                                       \
461                                                                         \
462         return orig;                                                    \
463 }
464
465 #define ATOMIC64_OPS(op, op1, op2)                                      \
466         ATOMIC64_OP(op, op1, op2)                                       \
467         ATOMIC64_OP_RETURN(op, op1, op2)                                \
468         ATOMIC64_FETCH_OP(op, op1, op2)
469
470 #define atomic64_andnot atomic64_andnot
471
472 ATOMIC64_OPS(add, add.f, adc)
473 ATOMIC64_OPS(sub, sub.f, sbc)
474 ATOMIC64_OPS(and, and, and)
475 ATOMIC64_OPS(andnot, bic, bic)
476 ATOMIC64_OPS(or, or, or)
477 ATOMIC64_OPS(xor, xor, xor)
478
479 #undef ATOMIC64_OPS
480 #undef ATOMIC64_FETCH_OP
481 #undef ATOMIC64_OP_RETURN
482 #undef ATOMIC64_OP
483
484 static inline long long
485 atomic64_cmpxchg(atomic64_t *ptr, long long expected, long long new)
486 {
487         long long prev;
488
489         smp_mb();
490
491         __asm__ __volatile__(
492         "1:     llockd  %0, [%1]        \n"
493         "       brne    %L0, %L2, 2f    \n"
494         "       brne    %H0, %H2, 2f    \n"
495         "       scondd  %3, [%1]        \n"
496         "       bnz     1b              \n"
497         "2:                             \n"
498         : "=&r"(prev)
499         : "r"(ptr), "ir"(expected), "r"(new)
500         : "cc");        /* memory clobber comes from smp_mb() */
501
502         smp_mb();
503
504         return prev;
505 }
506
507 static inline long long atomic64_xchg(atomic64_t *ptr, long long new)
508 {
509         long long prev;
510
511         smp_mb();
512
513         __asm__ __volatile__(
514         "1:     llockd  %0, [%1]        \n"
515         "       scondd  %2, [%1]        \n"
516         "       bnz     1b              \n"
517         "2:                             \n"
518         : "=&r"(prev)
519         : "r"(ptr), "r"(new)
520         : "cc");        /* memory clobber comes from smp_mb() */
521
522         smp_mb();
523
524         return prev;
525 }
526
527 /**
528  * atomic64_dec_if_positive - decrement by 1 if old value positive
529  * @v: pointer of type atomic64_t
530  *
531  * The function returns the old value of *v minus 1, even if
532  * the atomic variable, v, was not decremented.
533  */
534
535 static inline long long atomic64_dec_if_positive(atomic64_t *v)
536 {
537         long long val;
538
539         smp_mb();
540
541         __asm__ __volatile__(
542         "1:     llockd  %0, [%1]        \n"
543         "       sub.f   %L0, %L0, 1     # w0 - 1, set C on borrow\n"
544         "       sub.c   %H0, %H0, 1     # if C set, w1 - 1\n"
545         "       brlt    %H0, 0, 2f      \n"
546         "       scondd  %0, [%1]        \n"
547         "       bnz     1b              \n"
548         "2:                             \n"
549         : "=&r"(val)
550         : "r"(&v->counter)
551         : "cc");        /* memory clobber comes from smp_mb() */
552
553         smp_mb();
554
555         return val;
556 }
557
558 /**
559  * atomic64_add_unless - add unless the number is a given value
560  * @v: pointer of type atomic64_t
561  * @a: the amount to add to v...
562  * @u: ...unless v is equal to u.
563  *
564  * if (v != u) { v += a; ret = 1} else {ret = 0}
565  * Returns 1 iff @v was not @u (i.e. if add actually happened)
566  */
567 static inline int atomic64_add_unless(atomic64_t *v, long long a, long long u)
568 {
569         long long val;
570         int op_done;
571
572         smp_mb();
573
574         __asm__ __volatile__(
575         "1:     llockd  %0, [%2]        \n"
576         "       mov     %1, 1           \n"
577         "       brne    %L0, %L4, 2f    # continue to add since v != u \n"
578         "       breq.d  %H0, %H4, 3f    # return since v == u \n"
579         "       mov     %1, 0           \n"
580         "2:                             \n"
581         "       add.f   %L0, %L0, %L3   \n"
582         "       adc     %H0, %H0, %H3   \n"
583         "       scondd  %0, [%2]        \n"
584         "       bnz     1b              \n"
585         "3:                             \n"
586         : "=&r"(val), "=&r" (op_done)
587         : "r"(&v->counter), "r"(a), "r"(u)
588         : "cc");        /* memory clobber comes from smp_mb() */
589
590         smp_mb();
591
592         return op_done;
593 }
594
595 #define atomic64_add_negative(a, v)     (atomic64_add_return((a), (v)) < 0)
596 #define atomic64_inc(v)                 atomic64_add(1LL, (v))
597 #define atomic64_inc_return(v)          atomic64_add_return(1LL, (v))
598 #define atomic64_inc_and_test(v)        (atomic64_inc_return(v) == 0)
599 #define atomic64_sub_and_test(a, v)     (atomic64_sub_return((a), (v)) == 0)
600 #define atomic64_dec(v)                 atomic64_sub(1LL, (v))
601 #define atomic64_dec_return(v)          atomic64_sub_return(1LL, (v))
602 #define atomic64_dec_and_test(v)        (atomic64_dec_return((v)) == 0)
603 #define atomic64_inc_not_zero(v)        atomic64_add_unless((v), 1LL, 0LL)
604
605 #endif  /* !CONFIG_GENERIC_ATOMIC64 */
606
607 #endif  /* !__ASSEMBLY__ */
608
609 #endif