| 1 | /* SPDX-License-Identifier: GPL-2.0-only */ |
| 2 | /* |
| 3 | * Coherency fabric: low level functions |
| 4 | * |
| 5 | * Copyright (C) 2012 Marvell |
| 6 | * |
| 7 | * Gregory CLEMENT <gregory.clement@free-electrons.com> |
| 8 | * |
| 9 | * This file implements the assembly function to add a CPU to the |
| 10 | * coherency fabric. This function is called by each of the secondary |
| 11 | * CPUs during their early boot in an SMP kernel, this why this |
| 12 | * function have to callable from assembly. It can also be called by a |
| 13 | * primary CPU from C code during its boot. |
| 14 | */ |
| 15 | |
| 16 | #include <linux/linkage.h> |
| 17 | #define ARMADA_XP_CFB_CTL_REG_OFFSET 0x0 |
| 18 | #define ARMADA_XP_CFB_CFG_REG_OFFSET 0x4 |
| 19 | |
| 20 | #include <asm/assembler.h> |
| 21 | #include <asm/cp15.h> |
| 22 | |
| 23 | .arch armv7-a |
| 24 | .text |
| 25 | /* |
| 26 | * Returns the coherency base address in r1 (r0 is untouched), or 0 if |
| 27 | * the coherency fabric is not enabled. |
| 28 | */ |
| 29 | ENTRY(ll_get_coherency_base) |
| 30 | mrc p15, 0, r1, c1, c0, 0 |
| 31 | tst r1, #CR_M @ Check MMU bit enabled |
| 32 | bne 1f |
| 33 | |
| 34 | /* |
| 35 | * MMU is disabled, use the physical address of the coherency |
| 36 | * base address, (or 0x0 if the coherency fabric is not mapped) |
| 37 | */ |
| 38 | adr r1, 3f |
| 39 | ldr r3, [r1] |
| 40 | ldr r1, [r1, r3] |
| 41 | b 2f |
| 42 | 1: |
| 43 | /* |
| 44 | * MMU is enabled, use the virtual address of the coherency |
| 45 | * base address. |
| 46 | */ |
| 47 | ldr r1, =coherency_base |
| 48 | ldr r1, [r1] |
| 49 | 2: |
| 50 | ret lr |
| 51 | ENDPROC(ll_get_coherency_base) |
| 52 | |
| 53 | /* |
| 54 | * Returns the coherency CPU mask in r3 (r0 is untouched). This |
| 55 | * coherency CPU mask can be used with the coherency fabric |
| 56 | * configuration and control registers. Note that the mask is already |
| 57 | * endian-swapped as appropriate so that the calling functions do not |
| 58 | * have to care about endianness issues while accessing the coherency |
| 59 | * fabric registers |
| 60 | */ |
| 61 | ENTRY(ll_get_coherency_cpumask) |
| 62 | mrc p15, 0, r3, cr0, cr0, 5 |
| 63 | and r3, r3, #15 |
| 64 | mov r2, #(1 << 24) |
| 65 | lsl r3, r2, r3 |
| 66 | ARM_BE8(rev r3, r3) |
| 67 | ret lr |
| 68 | ENDPROC(ll_get_coherency_cpumask) |
| 69 | |
| 70 | /* |
| 71 | * ll_add_cpu_to_smp_group(), ll_enable_coherency() and |
| 72 | * ll_disable_coherency() use the strex/ldrex instructions while the |
| 73 | * MMU can be disabled. The Armada XP SoC has an exclusive monitor |
| 74 | * that tracks transactions to Device and/or SO memory and thanks to |
| 75 | * that, exclusive transactions are functional even when the MMU is |
| 76 | * disabled. |
| 77 | */ |
| 78 | |
| 79 | ENTRY(ll_add_cpu_to_smp_group) |
| 80 | /* |
| 81 | * As r0 is not modified by ll_get_coherency_base() and |
| 82 | * ll_get_coherency_cpumask(), we use it to temporarly save lr |
| 83 | * and avoid it being modified by the branch and link |
| 84 | * calls. This function is used very early in the secondary |
| 85 | * CPU boot, and no stack is available at this point. |
| 86 | */ |
| 87 | mov r0, lr |
| 88 | bl ll_get_coherency_base |
| 89 | /* Bail out if the coherency is not enabled */ |
| 90 | cmp r1, #0 |
| 91 | reteq r0 |
| 92 | bl ll_get_coherency_cpumask |
| 93 | mov lr, r0 |
| 94 | add r0, r1, #ARMADA_XP_CFB_CFG_REG_OFFSET |
| 95 | 1: |
| 96 | ldrex r2, [r0] |
| 97 | orr r2, r2, r3 |
| 98 | strex r1, r2, [r0] |
| 99 | cmp r1, #0 |
| 100 | bne 1b |
| 101 | ret lr |
| 102 | ENDPROC(ll_add_cpu_to_smp_group) |
| 103 | |
| 104 | ENTRY(ll_enable_coherency) |
| 105 | /* |
| 106 | * As r0 is not modified by ll_get_coherency_base() and |
| 107 | * ll_get_coherency_cpumask(), we use it to temporarly save lr |
| 108 | * and avoid it being modified by the branch and link |
| 109 | * calls. This function is used very early in the secondary |
| 110 | * CPU boot, and no stack is available at this point. |
| 111 | */ |
| 112 | mov r0, lr |
| 113 | bl ll_get_coherency_base |
| 114 | /* Bail out if the coherency is not enabled */ |
| 115 | cmp r1, #0 |
| 116 | reteq r0 |
| 117 | bl ll_get_coherency_cpumask |
| 118 | mov lr, r0 |
| 119 | add r0, r1, #ARMADA_XP_CFB_CTL_REG_OFFSET |
| 120 | 1: |
| 121 | ldrex r2, [r0] |
| 122 | orr r2, r2, r3 |
| 123 | strex r1, r2, [r0] |
| 124 | cmp r1, #0 |
| 125 | bne 1b |
| 126 | dsb |
| 127 | mov r0, #0 |
| 128 | ret lr |
| 129 | ENDPROC(ll_enable_coherency) |
| 130 | |
| 131 | ENTRY(ll_disable_coherency) |
| 132 | /* |
| 133 | * As r0 is not modified by ll_get_coherency_base() and |
| 134 | * ll_get_coherency_cpumask(), we use it to temporarly save lr |
| 135 | * and avoid it being modified by the branch and link |
| 136 | * calls. This function is used very early in the secondary |
| 137 | * CPU boot, and no stack is available at this point. |
| 138 | */ |
| 139 | mov r0, lr |
| 140 | bl ll_get_coherency_base |
| 141 | /* Bail out if the coherency is not enabled */ |
| 142 | cmp r1, #0 |
| 143 | reteq r0 |
| 144 | bl ll_get_coherency_cpumask |
| 145 | mov lr, r0 |
| 146 | add r0, r1, #ARMADA_XP_CFB_CTL_REG_OFFSET |
| 147 | 1: |
| 148 | ldrex r2, [r0] |
| 149 | bic r2, r2, r3 |
| 150 | strex r1, r2, [r0] |
| 151 | cmp r1, #0 |
| 152 | bne 1b |
| 153 | dsb |
| 154 | ret lr |
| 155 | ENDPROC(ll_disable_coherency) |
| 156 | |
| 157 | .align 2 |
| 158 | 3: |
| 159 | .long coherency_phys_base - . |
| 160 | |