Merge git://git.kernel.org/pub/scm/linux/kernel/git/davem/sparc-next
[sfrench/cifs-2.6.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 #define ATOMIC_INIT(i)  { (i) }
21
22 #ifndef CONFIG_ARC_PLAT_EZNPS
23
24 #define atomic_read(v)  READ_ONCE((v)->counter)
25
26 #ifdef CONFIG_ARC_HAS_LLSC
27
28 #define atomic_set(v, i) WRITE_ONCE(((v)->counter), (i))
29
30 #define ATOMIC_OP(op, c_op, asm_op)                                     \
31 static inline void atomic_##op(int i, atomic_t *v)                      \
32 {                                                                       \
33         unsigned int val;                                               \
34                                                                         \
35         __asm__ __volatile__(                                           \
36         "1:     llock   %[val], [%[ctr]]                \n"             \
37         "       " #asm_op " %[val], %[val], %[i]        \n"             \
38         "       scond   %[val], [%[ctr]]                \n"             \
39         "       bnz     1b                              \n"             \
40         : [val] "=&r"   (val) /* Early clobber to prevent reg reuse */  \
41         : [ctr] "r"     (&v->counter), /* Not "m": llock only supports reg direct addr mode */  \
42           [i]   "ir"    (i)                                             \
43         : "cc");                                                        \
44 }                                                                       \
45
46 #define ATOMIC_OP_RETURN(op, c_op, asm_op)                              \
47 static inline int atomic_##op##_return(int i, atomic_t *v)              \
48 {                                                                       \
49         unsigned int val;                                               \
50                                                                         \
51         /*                                                              \
52          * Explicit full memory barrier needed before/after as          \
53          * LLOCK/SCOND thmeselves don't provide any such semantics      \
54          */                                                             \
55         smp_mb();                                                       \
56                                                                         \
57         __asm__ __volatile__(                                           \
58         "1:     llock   %[val], [%[ctr]]                \n"             \
59         "       " #asm_op " %[val], %[val], %[i]        \n"             \
60         "       scond   %[val], [%[ctr]]                \n"             \
61         "       bnz     1b                              \n"             \
62         : [val] "=&r"   (val)                                           \
63         : [ctr] "r"     (&v->counter),                                  \
64           [i]   "ir"    (i)                                             \
65         : "cc");                                                        \
66                                                                         \
67         smp_mb();                                                       \
68                                                                         \
69         return val;                                                     \
70 }
71
72 #define ATOMIC_FETCH_OP(op, c_op, asm_op)                               \
73 static inline int atomic_fetch_##op(int i, atomic_t *v)                 \
74 {                                                                       \
75         unsigned int val, orig;                                         \
76                                                                         \
77         /*                                                              \
78          * Explicit full memory barrier needed before/after as          \
79          * LLOCK/SCOND thmeselves don't provide any such semantics      \
80          */                                                             \
81         smp_mb();                                                       \
82                                                                         \
83         __asm__ __volatile__(                                           \
84         "1:     llock   %[orig], [%[ctr]]               \n"             \
85         "       " #asm_op " %[val], %[orig], %[i]       \n"             \
86         "       scond   %[val], [%[ctr]]                \n"             \
87         "                                               \n"             \
88         : [val] "=&r"   (val),                                          \
89           [orig] "=&r" (orig)                                           \
90         : [ctr] "r"     (&v->counter),                                  \
91           [i]   "ir"    (i)                                             \
92         : "cc");                                                        \
93                                                                         \
94         smp_mb();                                                       \
95                                                                         \
96         return orig;                                                    \
97 }
98
99 #else   /* !CONFIG_ARC_HAS_LLSC */
100
101 #ifndef CONFIG_SMP
102
103  /* violating atomic_xxx API locking protocol in UP for optimization sake */
104 #define atomic_set(v, i) WRITE_ONCE(((v)->counter), (i))
105
106 #else
107
108 static inline void atomic_set(atomic_t *v, int i)
109 {
110         /*
111          * Independent of hardware support, all of the atomic_xxx() APIs need
112          * to follow the same locking rules to make sure that a "hardware"
113          * atomic insn (e.g. LD) doesn't clobber an "emulated" atomic insn
114          * sequence
115          *
116          * Thus atomic_set() despite being 1 insn (and seemingly atomic)
117          * requires the locking.
118          */
119         unsigned long flags;
120
121         atomic_ops_lock(flags);
122         WRITE_ONCE(v->counter, i);
123         atomic_ops_unlock(flags);
124 }
125
126 #define atomic_set_release(v, i)        atomic_set((v), (i))
127
128 #endif
129
130 /*
131  * Non hardware assisted Atomic-R-M-W
132  * Locking would change to irq-disabling only (UP) and spinlocks (SMP)
133  */
134
135 #define ATOMIC_OP(op, c_op, asm_op)                                     \
136 static inline void atomic_##op(int i, atomic_t *v)                      \
137 {                                                                       \
138         unsigned long flags;                                            \
139                                                                         \
140         atomic_ops_lock(flags);                                         \
141         v->counter c_op i;                                              \
142         atomic_ops_unlock(flags);                                       \
143 }
144
145 #define ATOMIC_OP_RETURN(op, c_op, asm_op)                              \
146 static inline int atomic_##op##_return(int i, atomic_t *v)              \
147 {                                                                       \
148         unsigned long flags;                                            \
149         unsigned long temp;                                             \
150                                                                         \
151         /*                                                              \
152          * spin lock/unlock provides the needed smp_mb() before/after   \
153          */                                                             \
154         atomic_ops_lock(flags);                                         \
155         temp = v->counter;                                              \
156         temp c_op i;                                                    \
157         v->counter = temp;                                              \
158         atomic_ops_unlock(flags);                                       \
159                                                                         \
160         return temp;                                                    \
161 }
162
163 #define ATOMIC_FETCH_OP(op, c_op, asm_op)                               \
164 static inline int atomic_fetch_##op(int i, atomic_t *v)                 \
165 {                                                                       \
166         unsigned long flags;                                            \
167         unsigned long orig;                                             \
168                                                                         \
169         /*                                                              \
170          * spin lock/unlock provides the needed smp_mb() before/after   \
171          */                                                             \
172         atomic_ops_lock(flags);                                         \
173         orig = v->counter;                                              \
174         v->counter c_op i;                                              \
175         atomic_ops_unlock(flags);                                       \
176                                                                         \
177         return orig;                                                    \
178 }
179
180 #endif /* !CONFIG_ARC_HAS_LLSC */
181
182 #define ATOMIC_OPS(op, c_op, asm_op)                                    \
183         ATOMIC_OP(op, c_op, asm_op)                                     \
184         ATOMIC_OP_RETURN(op, c_op, asm_op)                              \
185         ATOMIC_FETCH_OP(op, c_op, asm_op)
186
187 ATOMIC_OPS(add, +=, add)
188 ATOMIC_OPS(sub, -=, sub)
189
190 #define atomic_andnot atomic_andnot
191
192 #undef ATOMIC_OPS
193 #define ATOMIC_OPS(op, c_op, asm_op)                                    \
194         ATOMIC_OP(op, c_op, asm_op)                                     \
195         ATOMIC_FETCH_OP(op, c_op, asm_op)
196
197 ATOMIC_OPS(and, &=, and)
198 ATOMIC_OPS(andnot, &= ~, bic)
199 ATOMIC_OPS(or, |=, or)
200 ATOMIC_OPS(xor, ^=, xor)
201
202 #else /* CONFIG_ARC_PLAT_EZNPS */
203
204 static inline int atomic_read(const atomic_t *v)
205 {
206         int temp;
207
208         __asm__ __volatile__(
209         "       ld.di %0, [%1]"
210         : "=r"(temp)
211         : "r"(&v->counter)
212         : "memory");
213         return temp;
214 }
215
216 static inline void atomic_set(atomic_t *v, int i)
217 {
218         __asm__ __volatile__(
219         "       st.di %0,[%1]"
220         :
221         : "r"(i), "r"(&v->counter)
222         : "memory");
223 }
224
225 #define ATOMIC_OP(op, c_op, asm_op)                                     \
226 static inline void atomic_##op(int i, atomic_t *v)                      \
227 {                                                                       \
228         __asm__ __volatile__(                                           \
229         "       mov r2, %0\n"                                           \
230         "       mov r3, %1\n"                                           \
231         "       .word %2\n"                                             \
232         :                                                               \
233         : "r"(i), "r"(&v->counter), "i"(asm_op)                         \
234         : "r2", "r3", "memory");                                        \
235 }                                                                       \
236
237 #define ATOMIC_OP_RETURN(op, c_op, asm_op)                              \
238 static inline int atomic_##op##_return(int i, atomic_t *v)              \
239 {                                                                       \
240         unsigned int temp = i;                                          \
241                                                                         \
242         /* Explicit full memory barrier needed before/after */          \
243         smp_mb();                                                       \
244                                                                         \
245         __asm__ __volatile__(                                           \
246         "       mov r2, %0\n"                                           \
247         "       mov r3, %1\n"                                           \
248         "       .word %2\n"                                             \
249         "       mov %0, r2"                                             \
250         : "+r"(temp)                                                    \
251         : "r"(&v->counter), "i"(asm_op)                                 \
252         : "r2", "r3", "memory");                                        \
253                                                                         \
254         smp_mb();                                                       \
255                                                                         \
256         temp c_op i;                                                    \
257                                                                         \
258         return temp;                                                    \
259 }
260
261 #define ATOMIC_FETCH_OP(op, c_op, asm_op)                               \
262 static inline int atomic_fetch_##op(int i, atomic_t *v)                 \
263 {                                                                       \
264         unsigned int temp = i;                                          \
265                                                                         \
266         /* Explicit full memory barrier needed before/after */          \
267         smp_mb();                                                       \
268                                                                         \
269         __asm__ __volatile__(                                           \
270         "       mov r2, %0\n"                                           \
271         "       mov r3, %1\n"                                           \
272         "       .word %2\n"                                             \
273         "       mov %0, r2"                                             \
274         : "+r"(temp)                                                    \
275         : "r"(&v->counter), "i"(asm_op)                                 \
276         : "r2", "r3", "memory");                                        \
277                                                                         \
278         smp_mb();                                                       \
279                                                                         \
280         return temp;                                                    \
281 }
282
283 #define ATOMIC_OPS(op, c_op, asm_op)                                    \
284         ATOMIC_OP(op, c_op, asm_op)                                     \
285         ATOMIC_OP_RETURN(op, c_op, asm_op)                              \
286         ATOMIC_FETCH_OP(op, c_op, asm_op)
287
288 ATOMIC_OPS(add, +=, CTOP_INST_AADD_DI_R2_R2_R3)
289 #define atomic_sub(i, v) atomic_add(-(i), (v))
290 #define atomic_sub_return(i, v) atomic_add_return(-(i), (v))
291 #define atomic_fetch_sub(i, v) atomic_fetch_add(-(i), (v))
292
293 #undef ATOMIC_OPS
294 #define ATOMIC_OPS(op, c_op, asm_op)                                    \
295         ATOMIC_OP(op, c_op, asm_op)                                     \
296         ATOMIC_FETCH_OP(op, c_op, asm_op)
297
298 ATOMIC_OPS(and, &=, CTOP_INST_AAND_DI_R2_R2_R3)
299 #define atomic_andnot(mask, v) atomic_and(~(mask), (v))
300 #define atomic_fetch_andnot(mask, v) atomic_fetch_and(~(mask), (v))
301 ATOMIC_OPS(or, |=, CTOP_INST_AOR_DI_R2_R2_R3)
302 ATOMIC_OPS(xor, ^=, CTOP_INST_AXOR_DI_R2_R2_R3)
303
304 #endif /* CONFIG_ARC_PLAT_EZNPS */
305
306 #undef ATOMIC_OPS
307 #undef ATOMIC_FETCH_OP
308 #undef ATOMIC_OP_RETURN
309 #undef ATOMIC_OP
310
311 /**
312  * __atomic_add_unless - add unless the number is a given value
313  * @v: pointer of type atomic_t
314  * @a: the amount to add to v...
315  * @u: ...unless v is equal to u.
316  *
317  * Atomically adds @a to @v, so long as it was not @u.
318  * Returns the old value of @v
319  */
320 #define __atomic_add_unless(v, a, u)                                    \
321 ({                                                                      \
322         int c, old;                                                     \
323                                                                         \
324         /*                                                              \
325          * Explicit full memory barrier needed before/after as          \
326          * LLOCK/SCOND thmeselves don't provide any such semantics      \
327          */                                                             \
328         smp_mb();                                                       \
329                                                                         \
330         c = atomic_read(v);                                             \
331         while (c != (u) && (old = atomic_cmpxchg((v), c, c + (a))) != c)\
332                 c = old;                                                \
333                                                                         \
334         smp_mb();                                                       \
335                                                                         \
336         c;                                                              \
337 })
338
339 #define atomic_inc_not_zero(v)          atomic_add_unless((v), 1, 0)
340
341 #define atomic_inc(v)                   atomic_add(1, v)
342 #define atomic_dec(v)                   atomic_sub(1, v)
343
344 #define atomic_inc_and_test(v)          (atomic_add_return(1, v) == 0)
345 #define atomic_dec_and_test(v)          (atomic_sub_return(1, v) == 0)
346 #define atomic_inc_return(v)            atomic_add_return(1, (v))
347 #define atomic_dec_return(v)            atomic_sub_return(1, (v))
348 #define atomic_sub_and_test(i, v)       (atomic_sub_return(i, v) == 0)
349
350 #define atomic_add_negative(i, v)       (atomic_add_return(i, v) < 0)
351
352
353 #ifdef CONFIG_GENERIC_ATOMIC64
354
355 #include <asm-generic/atomic64.h>
356
357 #else   /* Kconfig ensures this is only enabled with needed h/w assist */
358
359 /*
360  * ARCv2 supports 64-bit exclusive load (LLOCKD) / store (SCONDD)
361  *  - The address HAS to be 64-bit aligned
362  *  - There are 2 semantics involved here:
363  *    = exclusive implies no interim update between load/store to same addr
364  *    = both words are observed/updated together: this is guaranteed even
365  *      for regular 64-bit load (LDD) / store (STD). Thus atomic64_set()
366  *      is NOT required to use LLOCKD+SCONDD, STD suffices
367  */
368
369 typedef struct {
370         aligned_u64 counter;
371 } atomic64_t;
372
373 #define ATOMIC64_INIT(a) { (a) }
374
375 static inline long long atomic64_read(const atomic64_t *v)
376 {
377         unsigned long long val;
378
379         __asm__ __volatile__(
380         "       ldd   %0, [%1]  \n"
381         : "=r"(val)
382         : "r"(&v->counter));
383
384         return val;
385 }
386
387 static inline void atomic64_set(atomic64_t *v, long long a)
388 {
389         /*
390          * This could have been a simple assignment in "C" but would need
391          * explicit volatile. Otherwise gcc optimizers could elide the store
392          * which borked atomic64 self-test
393          * In the inline asm version, memory clobber needed for exact same
394          * reason, to tell gcc about the store.
395          *
396          * This however is not needed for sibling atomic64_add() etc since both
397          * load/store are explicitly done in inline asm. As long as API is used
398          * for each access, gcc has no way to optimize away any load/store
399          */
400         __asm__ __volatile__(
401         "       std   %0, [%1]  \n"
402         :
403         : "r"(a), "r"(&v->counter)
404         : "memory");
405 }
406
407 #define ATOMIC64_OP(op, op1, op2)                                       \
408 static inline void atomic64_##op(long long a, atomic64_t *v)            \
409 {                                                                       \
410         unsigned long long val;                                         \
411                                                                         \
412         __asm__ __volatile__(                                           \
413         "1:                             \n"                             \
414         "       llockd  %0, [%1]        \n"                             \
415         "       " #op1 " %L0, %L0, %L2  \n"                             \
416         "       " #op2 " %H0, %H0, %H2  \n"                             \
417         "       scondd   %0, [%1]       \n"                             \
418         "       bnz     1b              \n"                             \
419         : "=&r"(val)                                                    \
420         : "r"(&v->counter), "ir"(a)                                     \
421         : "cc");                                                \
422 }                                                                       \
423
424 #define ATOMIC64_OP_RETURN(op, op1, op2)                                \
425 static inline long long atomic64_##op##_return(long long a, atomic64_t *v)      \
426 {                                                                       \
427         unsigned long long val;                                         \
428                                                                         \
429         smp_mb();                                                       \
430                                                                         \
431         __asm__ __volatile__(                                           \
432         "1:                             \n"                             \
433         "       llockd   %0, [%1]       \n"                             \
434         "       " #op1 " %L0, %L0, %L2  \n"                             \
435         "       " #op2 " %H0, %H0, %H2  \n"                             \
436         "       scondd   %0, [%1]       \n"                             \
437         "       bnz     1b              \n"                             \
438         : [val] "=&r"(val)                                              \
439         : "r"(&v->counter), "ir"(a)                                     \
440         : "cc");        /* memory clobber comes from smp_mb() */        \
441                                                                         \
442         smp_mb();                                                       \
443                                                                         \
444         return val;                                                     \
445 }
446
447 #define ATOMIC64_FETCH_OP(op, op1, op2)                                 \
448 static inline long long atomic64_fetch_##op(long long a, atomic64_t *v) \
449 {                                                                       \
450         unsigned long long val, orig;                                   \
451                                                                         \
452         smp_mb();                                                       \
453                                                                         \
454         __asm__ __volatile__(                                           \
455         "1:                             \n"                             \
456         "       llockd   %0, [%2]       \n"                             \
457         "       " #op1 " %L1, %L0, %L3  \n"                             \
458         "       " #op2 " %H1, %H0, %H3  \n"                             \
459         "       scondd   %1, [%2]       \n"                             \
460         "       bnz     1b              \n"                             \
461         : "=&r"(orig), "=&r"(val)                                       \
462         : "r"(&v->counter), "ir"(a)                                     \
463         : "cc");        /* memory clobber comes from smp_mb() */        \
464                                                                         \
465         smp_mb();                                                       \
466                                                                         \
467         return orig;                                                    \
468 }
469
470 #define ATOMIC64_OPS(op, op1, op2)                                      \
471         ATOMIC64_OP(op, op1, op2)                                       \
472         ATOMIC64_OP_RETURN(op, op1, op2)                                \
473         ATOMIC64_FETCH_OP(op, op1, op2)
474
475 #define atomic64_andnot atomic64_andnot
476
477 ATOMIC64_OPS(add, add.f, adc)
478 ATOMIC64_OPS(sub, sub.f, sbc)
479 ATOMIC64_OPS(and, and, and)
480 ATOMIC64_OPS(andnot, bic, bic)
481 ATOMIC64_OPS(or, or, or)
482 ATOMIC64_OPS(xor, xor, xor)
483
484 #undef ATOMIC64_OPS
485 #undef ATOMIC64_FETCH_OP
486 #undef ATOMIC64_OP_RETURN
487 #undef ATOMIC64_OP
488
489 static inline long long
490 atomic64_cmpxchg(atomic64_t *ptr, long long expected, long long new)
491 {
492         long long prev;
493
494         smp_mb();
495
496         __asm__ __volatile__(
497         "1:     llockd  %0, [%1]        \n"
498         "       brne    %L0, %L2, 2f    \n"
499         "       brne    %H0, %H2, 2f    \n"
500         "       scondd  %3, [%1]        \n"
501         "       bnz     1b              \n"
502         "2:                             \n"
503         : "=&r"(prev)
504         : "r"(ptr), "ir"(expected), "r"(new)
505         : "cc");        /* memory clobber comes from smp_mb() */
506
507         smp_mb();
508
509         return prev;
510 }
511
512 static inline long long atomic64_xchg(atomic64_t *ptr, long long new)
513 {
514         long long prev;
515
516         smp_mb();
517
518         __asm__ __volatile__(
519         "1:     llockd  %0, [%1]        \n"
520         "       scondd  %2, [%1]        \n"
521         "       bnz     1b              \n"
522         "2:                             \n"
523         : "=&r"(prev)
524         : "r"(ptr), "r"(new)
525         : "cc");        /* memory clobber comes from smp_mb() */
526
527         smp_mb();
528
529         return prev;
530 }
531
532 /**
533  * atomic64_dec_if_positive - decrement by 1 if old value positive
534  * @v: pointer of type atomic64_t
535  *
536  * The function returns the old value of *v minus 1, even if
537  * the atomic variable, v, was not decremented.
538  */
539
540 static inline long long atomic64_dec_if_positive(atomic64_t *v)
541 {
542         long long val;
543
544         smp_mb();
545
546         __asm__ __volatile__(
547         "1:     llockd  %0, [%1]        \n"
548         "       sub.f   %L0, %L0, 1     # w0 - 1, set C on borrow\n"
549         "       sub.c   %H0, %H0, 1     # if C set, w1 - 1\n"
550         "       brlt    %H0, 0, 2f      \n"
551         "       scondd  %0, [%1]        \n"
552         "       bnz     1b              \n"
553         "2:                             \n"
554         : "=&r"(val)
555         : "r"(&v->counter)
556         : "cc");        /* memory clobber comes from smp_mb() */
557
558         smp_mb();
559
560         return val;
561 }
562
563 /**
564  * atomic64_add_unless - add unless the number is a given value
565  * @v: pointer of type atomic64_t
566  * @a: the amount to add to v...
567  * @u: ...unless v is equal to u.
568  *
569  * if (v != u) { v += a; ret = 1} else {ret = 0}
570  * Returns 1 iff @v was not @u (i.e. if add actually happened)
571  */
572 static inline int atomic64_add_unless(atomic64_t *v, long long a, long long u)
573 {
574         long long val;
575         int op_done;
576
577         smp_mb();
578
579         __asm__ __volatile__(
580         "1:     llockd  %0, [%2]        \n"
581         "       mov     %1, 1           \n"
582         "       brne    %L0, %L4, 2f    # continue to add since v != u \n"
583         "       breq.d  %H0, %H4, 3f    # return since v == u \n"
584         "       mov     %1, 0           \n"
585         "2:                             \n"
586         "       add.f   %L0, %L0, %L3   \n"
587         "       adc     %H0, %H0, %H3   \n"
588         "       scondd  %0, [%2]        \n"
589         "       bnz     1b              \n"
590         "3:                             \n"
591         : "=&r"(val), "=&r" (op_done)
592         : "r"(&v->counter), "r"(a), "r"(u)
593         : "cc");        /* memory clobber comes from smp_mb() */
594
595         smp_mb();
596
597         return op_done;
598 }
599
600 #define atomic64_add_negative(a, v)     (atomic64_add_return((a), (v)) < 0)
601 #define atomic64_inc(v)                 atomic64_add(1LL, (v))
602 #define atomic64_inc_return(v)          atomic64_add_return(1LL, (v))
603 #define atomic64_inc_and_test(v)        (atomic64_inc_return(v) == 0)
604 #define atomic64_sub_and_test(a, v)     (atomic64_sub_return((a), (v)) == 0)
605 #define atomic64_dec(v)                 atomic64_sub(1LL, (v))
606 #define atomic64_dec_return(v)          atomic64_sub_return(1LL, (v))
607 #define atomic64_dec_and_test(v)        (atomic64_dec_return((v)) == 0)
608 #define atomic64_inc_not_zero(v)        atomic64_add_unless((v), 1LL, 0LL)
609
610 #endif  /* !CONFIG_GENERIC_ATOMIC64 */
611
612 #endif  /* !__ASSEMBLY__ */
613
614 #endif