xref: /linux/arch/riscv/include/asm/barrier.h (revision 2da68a77)
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 
15 #define nop()		__asm__ __volatile__ ("nop")
16 #define __nops(n)	".rept	" #n "\nnop\n.endr\n"
17 #define nops(n)		__asm__ __volatile__ (__nops(n))
18 
19 #define RISCV_FENCE(p, s) \
20 	__asm__ __volatile__ ("fence " #p "," #s : : : "memory")
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 #define __smp_store_release(p, v)					\
33 do {									\
34 	compiletime_assert_atomic_type(*p);				\
35 	RISCV_FENCE(rw,w);						\
36 	WRITE_ONCE(*p, v);						\
37 } while (0)
38 
39 #define __smp_load_acquire(p)						\
40 ({									\
41 	typeof(*p) ___p1 = READ_ONCE(*p);				\
42 	compiletime_assert_atomic_type(*p);				\
43 	RISCV_FENCE(r,rw);						\
44 	___p1;								\
45 })
46 
47 /*
48  * This is a very specific barrier: it's currently only used in two places in
49  * the kernel, both in the scheduler.  See include/linux/spinlock.h for the two
50  * orderings it guarantees, but the "critical section is RCsc" guarantee
51  * mandates a barrier on RISC-V.  The sequence looks like:
52  *
53  *    lr.aq lock
54  *    sc    lock <= LOCKED
55  *    smp_mb__after_spinlock()
56  *    // critical section
57  *    lr    lock
58  *    sc.rl lock <= UNLOCKED
59  *
60  * The AQ/RL pair provides a RCpc critical section, but there's not really any
61  * way we can take advantage of that here because the ordering is only enforced
62  * on that one lock.  Thus, we're just doing a full fence.
63  *
64  * Since we allow writeX to be called from preemptive regions we need at least
65  * an "o" in the predecessor set to ensure device writes are visible before the
66  * task is marked as available for scheduling on a new hart.  While I don't see
67  * any concrete reason we need a full IO fence, it seems safer to just upgrade
68  * this in order to avoid any IO crossing a scheduling boundary.  In both
69  * instances the scheduler pairs this with an mb(), so nothing is necessary on
70  * the new hart.
71  */
72 #define smp_mb__after_spinlock()	RISCV_FENCE(iorw,iorw)
73 
74 #include <asm-generic/barrier.h>
75 
76 #endif /* __ASSEMBLY__ */
77 
78 #endif /* _ASM_RISCV_BARRIER_H */
79