b.liu | e958203 | 2025-04-17 19:18:16 +0800 | [diff] [blame^] | 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 | |
| 17 | #define RISCV_FENCE(p, s) \ |
| 18 | __asm__ __volatile__ ("fence " #p "," #s : : : "memory") |
| 19 | |
| 20 | /* These barriers need to enforce ordering on both devices or memory. */ |
| 21 | #define mb() RISCV_FENCE(iorw,iorw) |
| 22 | #define rmb() RISCV_FENCE(ir,ir) |
| 23 | #define wmb() RISCV_FENCE(ow,ow) |
| 24 | |
| 25 | /* These barriers do not need to enforce ordering on devices, just memory. */ |
| 26 | #define __smp_mb() RISCV_FENCE(rw,rw) |
| 27 | #define __smp_rmb() RISCV_FENCE(r,r) |
| 28 | #define __smp_wmb() RISCV_FENCE(w,w) |
| 29 | |
| 30 | #define __smp_store_release(p, v) \ |
| 31 | do { \ |
| 32 | compiletime_assert_atomic_type(*p); \ |
| 33 | RISCV_FENCE(rw,w); \ |
| 34 | WRITE_ONCE(*p, v); \ |
| 35 | } while (0) |
| 36 | |
| 37 | #define __smp_load_acquire(p) \ |
| 38 | ({ \ |
| 39 | typeof(*p) ___p1 = READ_ONCE(*p); \ |
| 40 | compiletime_assert_atomic_type(*p); \ |
| 41 | RISCV_FENCE(r,rw); \ |
| 42 | ___p1; \ |
| 43 | }) |
| 44 | |
| 45 | /* |
| 46 | * This is a very specific barrier: it's currently only used in two places in |
| 47 | * the kernel, both in the scheduler. See include/linux/spinlock.h for the two |
| 48 | * orderings it guarantees, but the "critical section is RCsc" guarantee |
| 49 | * mandates a barrier on RISC-V. The sequence looks like: |
| 50 | * |
| 51 | * lr.aq lock |
| 52 | * sc lock <= LOCKED |
| 53 | * smp_mb__after_spinlock() |
| 54 | * // critical section |
| 55 | * lr lock |
| 56 | * sc.rl lock <= UNLOCKED |
| 57 | * |
| 58 | * The AQ/RL pair provides a RCpc critical section, but there's not really any |
| 59 | * way we can take advantage of that here because the ordering is only enforced |
| 60 | * on that one lock. Thus, we're just doing a full fence. |
| 61 | * |
| 62 | * Since we allow writeX to be called from preemptive regions we need at least |
| 63 | * an "o" in the predecessor set to ensure device writes are visible before the |
| 64 | * task is marked as available for scheduling on a new hart. While I don't see |
| 65 | * any concrete reason we need a full IO fence, it seems safer to just upgrade |
| 66 | * this in order to avoid any IO crossing a scheduling boundary. In both |
| 67 | * instances the scheduler pairs this with an mb(), so nothing is necessary on |
| 68 | * the new hart. |
| 69 | */ |
| 70 | #define smp_mb__after_spinlock() RISCV_FENCE(iorw,iorw) |
| 71 | |
| 72 | #include <asm-generic/barrier.h> |
| 73 | |
| 74 | #endif /* __ASSEMBLY__ */ |
| 75 | |
| 76 | #endif /* _ASM_RISCV_BARRIER_H */ |