diff options
author | ian <ian@FreeBSD.org> | 2014-08-11 02:20:24 +0000 |
---|---|---|
committer | ian <ian@FreeBSD.org> | 2014-08-11 02:20:24 +0000 |
commit | 6b999dff3e159c3cb1559a30c663435c3265807b (patch) | |
tree | d732f48bb7ec4608a3f5ca242d03181e7f091bde | |
parent | 8067096dad03c7d9e39793e2c65fd2b67a2b1727 (diff) | |
download | FreeBSD-src-6b999dff3e159c3cb1559a30c663435c3265807b.zip FreeBSD-src-6b999dff3e159c3cb1559a30c663435c3265807b.tar.gz |
MFC r269403, r269405, r269410, r269414:
Add 64-bit atomic ops for armv6, and also for armv4 only in kernel code.
Use the new ops in the cddl code (and avoid defining functions with the
same names locally).
-rw-r--r-- | sys/arm/include/atomic.h | 321 | ||||
-rw-r--r-- | sys/cddl/compat/opensolaris/kern/opensolaris_atomic.c | 2 | ||||
-rw-r--r-- | sys/cddl/compat/opensolaris/sys/atomic.h | 4 |
3 files changed, 324 insertions, 3 deletions
diff --git a/sys/arm/include/atomic.h b/sys/arm/include/atomic.h index 02be1bd..a62ca00 100644 --- a/sys/arm/include/atomic.h +++ b/sys/arm/include/atomic.h @@ -88,6 +88,8 @@ defined (__ARM_ARCH_6T2__) || \ defined (__ARM_ARCH_6Z__) || \ defined (__ARM_ARCH_6ZK__) +#define ARM_HAVE_ATOMIC64 + static __inline void __do_dmb(void) { @@ -146,6 +148,28 @@ atomic_set_32(volatile uint32_t *address, uint32_t setmask) } static __inline void +atomic_set_64(volatile uint64_t *p, uint64_t val) +{ + uint64_t tmp; + uint32_t exflag; + + __asm __volatile( + "1: \n" + " ldrexd %[tmp], [%[ptr]]\n" + " orr %Q[tmp], %Q[val]\n" + " orr %R[tmp], %R[val]\n" + " strexd %[exf], %[tmp], [%[ptr]]\n" + " teq %[exf], #0\n" + " it ne \n" + " bne 1b\n" + : [exf] "=&r" (exflag), + [tmp] "=&r" (tmp) + : [ptr] "r" (p), + [val] "r" (val) + : "cc", "memory"); +} + +static __inline void atomic_set_long(volatile u_long *address, u_long setmask) { u_long tmp = 0, tmp2 = 0; @@ -177,6 +201,28 @@ atomic_clear_32(volatile uint32_t *address, uint32_t setmask) } static __inline void +atomic_clear_64(volatile uint64_t *p, uint64_t val) +{ + uint64_t tmp; + uint32_t exflag; + + __asm __volatile( + "1: \n" + " ldrexd %[tmp], [%[ptr]]\n" + " bic %Q[tmp], %Q[val]\n" + " bic %R[tmp], %R[val]\n" + " strexd %[exf], %[tmp], [%[ptr]]\n" + " teq %[exf], #0\n" + " it ne \n" + " bne 1b\n" + : [exf] "=&r" (exflag), + [tmp] "=&r" (tmp) + : [ptr] "r" (p), + [val] "r" (val) + : "cc", "memory"); +} + +static __inline void atomic_clear_long(volatile u_long *address, u_long setmask) { u_long tmp = 0, tmp2 = 0; @@ -213,6 +259,35 @@ atomic_cmpset_32(volatile u_int32_t *p, volatile u_int32_t cmpval, volatile u_in return (ret); } +static __inline int +atomic_cmpset_64(volatile uint64_t *p, uint64_t cmpval, uint64_t newval) +{ + uint64_t tmp; + uint32_t ret; + + __asm __volatile( + "1: \n" + " ldrexd %[tmp], [%[ptr]]\n" + " teq %Q[tmp], %Q[cmp]\n" + " itee eq \n" + " teqeq %R[tmp], %R[cmp]\n" + " movne %[ret], #0\n" + " bne 2f\n" + " strexd %[ret], %[new], [%[ptr]]\n" + " teq %[ret], #0\n" + " it ne \n" + " bne 1b\n" + " mov %[ret], #1\n" + "2: \n" + : [ret] "=&r" (ret), + [tmp] "=&r" (tmp) + : [ptr] "r" (p), + [cmp] "r" (cmpval), + [new] "r" (newval) + : "cc", "memory"); + return (ret); +} + static __inline u_long atomic_cmpset_long(volatile u_long *p, volatile u_long cmpval, volatile u_long newval) { @@ -244,6 +319,15 @@ atomic_cmpset_acq_32(volatile u_int32_t *p, volatile u_int32_t cmpval, volatile return (ret); } +static __inline uint64_t +atomic_cmpset_acq_64(volatile uint64_t *p, volatile uint64_t cmpval, volatile uint64_t newval) +{ + uint64_t ret = atomic_cmpset_64(p, cmpval, newval); + + __do_dmb(); + return (ret); +} + static __inline u_long atomic_cmpset_acq_long(volatile u_long *p, volatile u_long cmpval, volatile u_long newval) { @@ -261,6 +345,14 @@ atomic_cmpset_rel_32(volatile u_int32_t *p, volatile u_int32_t cmpval, volatile return (atomic_cmpset_32(p, cmpval, newval)); } +static __inline uint64_t +atomic_cmpset_rel_64(volatile uint64_t *p, volatile uint64_t cmpval, volatile uint64_t newval) +{ + + __do_dmb(); + return (atomic_cmpset_64(p, cmpval, newval)); +} + static __inline u_long atomic_cmpset_rel_long(volatile u_long *p, volatile u_long cmpval, volatile u_long newval) { @@ -286,6 +378,28 @@ atomic_add_32(volatile u_int32_t *p, u_int32_t val) } static __inline void +atomic_add_64(volatile uint64_t *p, uint64_t val) +{ + uint64_t tmp; + uint32_t exflag; + + __asm __volatile( + "1: \n" + " ldrexd %[tmp], [%[ptr]]\n" + " adds %Q[tmp], %Q[val]\n" + " adc %R[tmp], %R[val]\n" + " strexd %[exf], %[tmp], [%[ptr]]\n" + " teq %[exf], #0\n" + " it ne \n" + " bne 1b\n" + : [exf] "=&r" (exflag), + [tmp] "=&r" (tmp) + : [ptr] "r" (p), + [val] "r" (val) + : "cc", "memory"); +} + +static __inline void atomic_add_long(volatile u_long *p, u_long val) { u_long tmp = 0, tmp2 = 0; @@ -316,6 +430,28 @@ atomic_subtract_32(volatile u_int32_t *p, u_int32_t val) } static __inline void +atomic_subtract_64(volatile uint64_t *p, uint64_t val) +{ + uint64_t tmp; + uint32_t exflag; + + __asm __volatile( + "1: \n" + " ldrexd %[tmp], [%[ptr]]\n" + " subs %Q[tmp], %Q[val]\n" + " sbc %R[tmp], %R[val]\n" + " strexd %[exf], %[tmp], [%[ptr]]\n" + " teq %[exf], #0\n" + " it ne \n" + " bne 1b\n" + : [exf] "=&r" (exflag), + [tmp] "=&r" (tmp) + : [ptr] "r" (p), + [val] "r" (val) + : "cc", "memory"); +} + +static __inline void atomic_subtract_long(volatile u_long *p, u_long val) { u_long tmp = 0, tmp2 = 0; @@ -334,6 +470,10 @@ ATOMIC_ACQ_REL(clear, 32) ATOMIC_ACQ_REL(add, 32) ATOMIC_ACQ_REL(subtract, 32) ATOMIC_ACQ_REL(set, 32) +ATOMIC_ACQ_REL(clear, 64) +ATOMIC_ACQ_REL(add, 64) +ATOMIC_ACQ_REL(subtract, 64) +ATOMIC_ACQ_REL(set, 64) ATOMIC_ACQ_REL_LONG(clear) ATOMIC_ACQ_REL_LONG(add) ATOMIC_ACQ_REL_LONG(subtract) @@ -392,6 +532,116 @@ atomic_store_rel_32(volatile uint32_t *p, uint32_t v) *p = v; } +static __inline uint64_t +atomic_fetchadd_64(volatile uint64_t *p, uint64_t val) +{ + uint64_t ret, tmp; + uint32_t exflag; + + __asm __volatile( + "1: \n" + " ldrexd %[ret], [%[ptr]]\n" + " adds %Q[tmp], %Q[ret], %Q[val]\n" + " adc %R[tmp], %R[ret], %R[val]\n" + " strexd %[exf], %[tmp], [%[ptr]]\n" + " teq %[exf], #0\n" + " it ne \n" + " bne 1b\n" + : [ret] "=&r" (ret), + [exf] "=&r" (exflag), + [tmp] "=&r" (tmp) + : [ptr] "r" (p), + [val] "r" (val) + : "cc", "memory"); + return (ret); +} + +static __inline uint64_t +atomic_readandclear_64(volatile uint64_t *p) +{ + uint64_t ret, tmp; + uint32_t exflag; + + __asm __volatile( + "1: \n" + " ldrexd %[ret], [%[ptr]]\n" + " mov %Q[tmp], #0\n" + " mov %R[tmp], #0\n" + " strexd %[exf], %[tmp], [%[ptr]]\n" + " teq %[exf], #0\n" + " it ne \n" + " bne 1b\n" + : [ret] "=&r" (ret), + [exf] "=&r" (exflag), + [tmp] "=&r" (tmp) + : [ptr] "r" (p) + : "cc", "memory"); + return (ret); +} + +static __inline uint64_t +atomic_load_64(volatile uint64_t *p) +{ + uint64_t ret; + + /* + * The only way to atomically load 64 bits is with LDREXD which puts the + * exclusive monitor into the open state, so reset it with CLREX because + * we don't actually need to store anything. + */ + __asm __volatile( + "1: \n" + " ldrexd %[ret], [%[ptr]]\n" + " clrex \n" + : [ret] "=&r" (ret) + : [ptr] "r" (p) + : "cc", "memory"); + return (ret); +} + +static __inline uint64_t +atomic_load_acq_64(volatile uint64_t *p) +{ + uint64_t ret; + + ret = atomic_load_64(p); + __do_dmb(); + return (ret); +} + +static __inline void +atomic_store_64(volatile uint64_t *p, uint64_t val) +{ + uint64_t tmp; + uint32_t exflag; + + /* + * The only way to atomically store 64 bits is with STREXD, which will + * succeed only if paired up with a preceeding LDREXD using the same + * address, so we read and discard the existing value before storing. + */ + __asm __volatile( + "1: \n" + " ldrexd %[tmp], [%[ptr]]\n" + " strexd %[exf], %[val], [%[ptr]]\n" + " teq %[exf], #0\n" + " it ne \n" + " bne 1b\n" + : [tmp] "=&r" (tmp), + [exf] "=&r" (exflag) + : [ptr] "r" (p), + [val] "r" (val) + : "cc", "memory"); +} + +static __inline void +atomic_store_rel_64(volatile uint64_t *p, uint64_t val) +{ + + __do_dmb(); + atomic_store_64(p, val); +} + static __inline u_long atomic_fetchadd_long(volatile u_long *p, u_long val) { @@ -474,6 +724,8 @@ __swp(uint32_t val, volatile uint32_t *ptr) #ifdef _KERNEL +#define ARM_HAVE_ATOMIC64 + static __inline void atomic_set_32(volatile uint32_t *address, uint32_t setmask) { @@ -481,11 +733,23 @@ atomic_set_32(volatile uint32_t *address, uint32_t setmask) } static __inline void +atomic_set_64(volatile uint64_t *address, uint64_t setmask) +{ + __with_interrupts_disabled(*address |= setmask); +} + +static __inline void atomic_clear_32(volatile uint32_t *address, uint32_t clearmask) { __with_interrupts_disabled(*address &= ~clearmask); } +static __inline void +atomic_clear_64(volatile uint64_t *address, uint64_t clearmask) +{ + __with_interrupts_disabled(*address &= ~clearmask); +} + static __inline u_int32_t atomic_cmpset_32(volatile u_int32_t *p, volatile u_int32_t cmpval, volatile u_int32_t newval) { @@ -503,6 +767,23 @@ atomic_cmpset_32(volatile u_int32_t *p, volatile u_int32_t cmpval, volatile u_in return (ret); } +static __inline u_int64_t +atomic_cmpset_64(volatile u_int64_t *p, volatile u_int64_t cmpval, volatile u_int64_t newval) +{ + int ret; + + __with_interrupts_disabled( + { + if (*p == cmpval) { + *p = newval; + ret = 1; + } else { + ret = 0; + } + }); + return (ret); +} + static __inline void atomic_add_32(volatile u_int32_t *p, u_int32_t val) { @@ -510,11 +791,23 @@ atomic_add_32(volatile u_int32_t *p, u_int32_t val) } static __inline void +atomic_add_64(volatile u_int64_t *p, u_int64_t val) +{ + __with_interrupts_disabled(*p += val); +} + +static __inline void atomic_subtract_32(volatile u_int32_t *p, u_int32_t val) { __with_interrupts_disabled(*p -= val); } +static __inline void +atomic_subtract_64(volatile u_int64_t *p, u_int64_t val) +{ + __with_interrupts_disabled(*p -= val); +} + static __inline uint32_t atomic_fetchadd_32(volatile uint32_t *p, uint32_t v) { @@ -528,6 +821,34 @@ atomic_fetchadd_32(volatile uint32_t *p, uint32_t v) return (value); } +static __inline uint64_t +atomic_fetchadd_64(volatile uint64_t *p, uint64_t v) +{ + uint64_t value; + + __with_interrupts_disabled( + { + value = *p; + *p += v; + }); + return (value); +} + +static __inline uint64_t +atomic_load_64(volatile uint64_t *p) +{ + uint64_t value; + + __with_interrupts_disabled(value = *p); + return (value); +} + +static __inline void +atomic_store_64(volatile uint64_t *p, uint64_t value) +{ + __with_interrupts_disabled(*p = value); +} + #else /* !_KERNEL */ static __inline u_int32_t diff --git a/sys/cddl/compat/opensolaris/kern/opensolaris_atomic.c b/sys/cddl/compat/opensolaris/kern/opensolaris_atomic.c index 4aba9ea..8a0fdf3 100644 --- a/sys/cddl/compat/opensolaris/kern/opensolaris_atomic.c +++ b/sys/cddl/compat/opensolaris/kern/opensolaris_atomic.c @@ -52,7 +52,7 @@ atomic_init(void) } #endif -#if !defined(__LP64__) && !defined(__mips_n32) +#if !defined(__LP64__) && !defined(__mips_n32) && !defined(ARM_HAVE_ATOMIC64) void atomic_add_64(volatile uint64_t *target, int64_t delta) { diff --git a/sys/cddl/compat/opensolaris/sys/atomic.h b/sys/cddl/compat/opensolaris/sys/atomic.h index f34d77e..65c9e71 100644 --- a/sys/cddl/compat/opensolaris/sys/atomic.h +++ b/sys/cddl/compat/opensolaris/sys/atomic.h @@ -36,7 +36,7 @@ atomic_cmpset_ptr((volatile uintptr_t *)(_a), (uintptr_t)(_b), (uintptr_t) (_c)) #define cas32 atomic_cmpset_32 -#if !defined(__LP64__) && !defined(__mips_n32) +#if !defined(__LP64__) && !defined(__mips_n32) && !defined(ARM_HAVE_ATOMIC64) extern void atomic_add_64(volatile uint64_t *target, int64_t delta); extern void atomic_dec_64(volatile uint64_t *target); #endif @@ -85,7 +85,7 @@ atomic_dec_32_nv(volatile uint32_t *target) return (atomic_fetchadd_32(target, -1) - 1); } -#if defined(__LP64__) || defined(__mips_n32) +#if defined(__LP64__) || defined(__mips_n32) || defined(ARM_HAVE_ATOMIC64) static __inline void atomic_dec_64(volatile uint64_t *target) { |