xref: /linux/arch/riscv/include/asm/barrier.h (revision 27c8f12e972d3647e9d759d7cafd4c34fa513432)
1 /* SPDX-License-Identifier: GPL-2.0-only */
2 /*
3  * Based on arch/arm/include/asm/barrier.h
4  *
5  * Copyright (C) 2012 ARM Ltd.
6  * Copyright (C) 2013 Regents of the University of California
7  * Copyright (C) 2017 SiFive
8  */
9 
10 #ifndef _ASM_RISCV_BARRIER_H
11 #define _ASM_RISCV_BARRIER_H
12 
13 #ifndef __ASSEMBLY__
14 #include <asm/cmpxchg.h>
15 #include <asm/fence.h>
16 
17 #define nop()		__asm__ __volatile__ ("nop")
18 #define __nops(n)	".rept	" #n "\nnop\n.endr\n"
19 #define nops(n)		__asm__ __volatile__ (__nops(n))
20 
21 
22 /* These barriers need to enforce ordering on both devices or memory. */
23 #define __mb()		RISCV_FENCE(iorw, iorw)
24 #define __rmb()		RISCV_FENCE(ir, ir)
25 #define __wmb()		RISCV_FENCE(ow, ow)
26 
27 /* These barriers do not need to enforce ordering on devices, just memory. */
28 #define __smp_mb()	RISCV_FENCE(rw, rw)
29 #define __smp_rmb()	RISCV_FENCE(r, r)
30 #define __smp_wmb()	RISCV_FENCE(w, w)
31 
32 /*
33  * This is a very specific barrier: it's currently only used in two places in
34  * the kernel, both in the scheduler.  See include/linux/spinlock.h for the two
35  * orderings it guarantees, but the "critical section is RCsc" guarantee
36  * mandates a barrier on RISC-V.  The sequence looks like:
37  *
38  *    lr.aq lock
39  *    sc    lock <= LOCKED
40  *    smp_mb__after_spinlock()
41  *    // critical section
42  *    lr    lock
43  *    sc.rl lock <= UNLOCKED
44  *
45  * The AQ/RL pair provides a RCpc critical section, but there's not really any
46  * way we can take advantage of that here because the ordering is only enforced
47  * on that one lock.  Thus, we're just doing a full fence.
48  *
49  * Since we allow writeX to be called from preemptive regions we need at least
50  * an "o" in the predecessor set to ensure device writes are visible before the
51  * task is marked as available for scheduling on a new hart.  While I don't see
52  * any concrete reason we need a full IO fence, it seems safer to just upgrade
53  * this in order to avoid any IO crossing a scheduling boundary.  In both
54  * instances the scheduler pairs this with an mb(), so nothing is necessary on
55  * the new hart.
56  */
57 #define smp_mb__after_spinlock()	RISCV_FENCE(iorw, iorw)
58 
59 #define __smp_store_release(p, v)					\
60 do {									\
61 	compiletime_assert_atomic_type(*p);				\
62 	RISCV_FENCE(rw, w);						\
63 	WRITE_ONCE(*p, v);						\
64 } while (0)
65 
66 #define __smp_load_acquire(p)						\
67 ({									\
68 	typeof(*p) ___p1 = READ_ONCE(*p);				\
69 	compiletime_assert_atomic_type(*p);				\
70 	RISCV_FENCE(r, rw);						\
71 	___p1;								\
72 })
73 
74 #ifdef CONFIG_RISCV_ISA_ZAWRS
75 #define smp_cond_load_relaxed(ptr, cond_expr) ({			\
76 	typeof(ptr) __PTR = (ptr);					\
77 	__unqual_scalar_typeof(*ptr) VAL;				\
78 	for (;;) {							\
79 		VAL = READ_ONCE(*__PTR);				\
80 		if (cond_expr)						\
81 			break;						\
82 		__cmpwait_relaxed(ptr, VAL);				\
83 	}								\
84 	(typeof(*ptr))VAL;						\
85 })
86 #endif
87 
88 #include <asm-generic/barrier.h>
89 
90 #endif /* __ASSEMBLY__ */
91 
92 #endif /* _ASM_RISCV_BARRIER_H */
93