1/* $NetBSD: amdgpu_sdma_v4_0.c,v 1.3 2021/12/19 12:21:29 riastradh Exp $ */ 2 3/* 4 * Copyright 2016 Advanced Micro Devices, Inc. 5 * 6 * Permission is hereby granted, free of charge, to any person obtaining a 7 * copy of this software and associated documentation files (the "Software"), 8 * to deal in the Software without restriction, including without limitation 9 * the rights to use, copy, modify, merge, publish, distribute, sublicense, 10 * and/or sell copies of the Software, and to permit persons to whom the 11 * Software is furnished to do so, subject to the following conditions: 12 * 13 * The above copyright notice and this permission notice shall be included in 14 * all copies or substantial portions of the Software. 15 * 16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR 17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, 18 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL 19 * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR 20 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, 21 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR 22 * OTHER DEALINGS IN THE SOFTWARE. 23 * 24 */ 25 26#include <sys/cdefs.h> 27__KERNEL_RCSID(0, "$NetBSD: amdgpu_sdma_v4_0.c,v 1.3 2021/12/19 12:21:29 riastradh Exp $"); 28 29#include <linux/delay.h> 30#include <linux/firmware.h> 31#include <linux/module.h> 32#include <linux/pci.h> 33 34#include "amdgpu.h" 35#include "amdgpu_ucode.h" 36#include "amdgpu_trace.h" 37 38#include "sdma0/sdma0_4_2_offset.h" 39#include "sdma0/sdma0_4_2_sh_mask.h" 40#include "sdma1/sdma1_4_2_offset.h" 41#include "sdma1/sdma1_4_2_sh_mask.h" 42#include "sdma2/sdma2_4_2_2_offset.h" 43#include "sdma2/sdma2_4_2_2_sh_mask.h" 44#include "sdma3/sdma3_4_2_2_offset.h" 45#include "sdma3/sdma3_4_2_2_sh_mask.h" 46#include "sdma4/sdma4_4_2_2_offset.h" 47#include "sdma4/sdma4_4_2_2_sh_mask.h" 48#include "sdma5/sdma5_4_2_2_offset.h" 49#include "sdma5/sdma5_4_2_2_sh_mask.h" 50#include "sdma6/sdma6_4_2_2_offset.h" 51#include "sdma6/sdma6_4_2_2_sh_mask.h" 52#include "sdma7/sdma7_4_2_2_offset.h" 53#include "sdma7/sdma7_4_2_2_sh_mask.h" 54#include "hdp/hdp_4_0_offset.h" 55#include "sdma0/sdma0_4_1_default.h" 56 57#include "soc15_common.h" 58#include "soc15.h" 59#include "vega10_sdma_pkt_open.h" 60 61#include "ivsrcid/sdma0/irqsrcs_sdma0_4_0.h" 62#include "ivsrcid/sdma1/irqsrcs_sdma1_4_0.h" 63 64#include "amdgpu_ras.h" 65 66#include <linux/nbsd-namespace.h> 67 68MODULE_FIRMWARE("amdgpu/vega10_sdma.bin"); 69MODULE_FIRMWARE("amdgpu/vega10_sdma1.bin"); 70MODULE_FIRMWARE("amdgpu/vega12_sdma.bin"); 71MODULE_FIRMWARE("amdgpu/vega12_sdma1.bin"); 72MODULE_FIRMWARE("amdgpu/vega20_sdma.bin"); 73MODULE_FIRMWARE("amdgpu/vega20_sdma1.bin"); 74MODULE_FIRMWARE("amdgpu/raven_sdma.bin"); 75MODULE_FIRMWARE("amdgpu/picasso_sdma.bin"); 76MODULE_FIRMWARE("amdgpu/raven2_sdma.bin"); 77MODULE_FIRMWARE("amdgpu/arcturus_sdma.bin"); 78MODULE_FIRMWARE("amdgpu/renoir_sdma.bin"); 79 80#define SDMA0_POWER_CNTL__ON_OFF_CONDITION_HOLD_TIME_MASK 0x000000F8L 81#define SDMA0_POWER_CNTL__ON_OFF_STATUS_DURATION_TIME_MASK 0xFC000000L 82 83#define WREG32_SDMA(instance, offset, value) \ 84 WREG32(sdma_v4_0_get_reg_offset(adev, (instance), (offset)), value) 85#define RREG32_SDMA(instance, offset) \ 86 RREG32(sdma_v4_0_get_reg_offset(adev, (instance), (offset))) 87 88static void sdma_v4_0_set_ring_funcs(struct amdgpu_device *adev); 89static void sdma_v4_0_set_buffer_funcs(struct amdgpu_device *adev); 90static void sdma_v4_0_set_vm_pte_funcs(struct amdgpu_device *adev); 91static void sdma_v4_0_set_irq_funcs(struct amdgpu_device *adev); 92static void sdma_v4_0_set_ras_funcs(struct amdgpu_device *adev); 93 94static const struct soc15_reg_golden golden_settings_sdma_4[] = { 95 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_CHICKEN_BITS, 0xfe931f07, 0x02831d07), 96 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_CLK_CTRL, 0xff000ff0, 0x3f000100), 97 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_GFX_IB_CNTL, 0x800f0100, 0x00000100), 98 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_GFX_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000), 99 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_PAGE_IB_CNTL, 0x800f0100, 0x00000100), 100 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_PAGE_RB_WPTR_POLL_CNTL, 0x0000fff0, 0x00403000), 101 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_POWER_CNTL, 0x003ff006, 0x0003c000), 102 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_RLC0_IB_CNTL, 0x800f0100, 0x00000100), 103 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_RLC0_RB_WPTR_POLL_CNTL, 0x0000fff0, 0x00403000), 104 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_RLC1_IB_CNTL, 0x800f0100, 0x00000100), 105 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_RLC1_RB_WPTR_POLL_CNTL, 0x0000fff0, 0x00403000), 106 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_UTCL1_PAGE, 0x000003ff, 0x000003c0), 107 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_UTCL1_WATERMK, 0xfc000000, 0x00000000), 108 SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_CLK_CTRL, 0xffffffff, 0x3f000100), 109 SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_GFX_IB_CNTL, 0x800f0100, 0x00000100), 110 SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_GFX_RB_WPTR_POLL_CNTL, 0x0000fff0, 0x00403000), 111 SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_PAGE_IB_CNTL, 0x800f0100, 0x00000100), 112 SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_PAGE_RB_WPTR_POLL_CNTL, 0x0000fff0, 0x00403000), 113 SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_POWER_CNTL, 0x003ff000, 0x0003c000), 114 SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_RLC0_IB_CNTL, 0x800f0100, 0x00000100), 115 SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_RLC0_RB_WPTR_POLL_CNTL, 0x0000fff0, 0x00403000), 116 SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_RLC1_IB_CNTL, 0x800f0100, 0x00000100), 117 SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_RLC1_RB_WPTR_POLL_CNTL, 0x0000fff0, 0x00403000), 118 SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_UTCL1_PAGE, 0x000003ff, 0x000003c0), 119 SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_UTCL1_WATERMK, 0xfc000000, 0x00000000) 120}; 121 122static const struct soc15_reg_golden golden_settings_sdma_vg10[] = { 123 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_GB_ADDR_CONFIG, 0x0018773f, 0x00104002), 124 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_GB_ADDR_CONFIG_READ, 0x0018773f, 0x00104002), 125 SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_CHICKEN_BITS, 0xfe931f07, 0x02831d07), 126 SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_GB_ADDR_CONFIG, 0x0018773f, 0x00104002), 127 SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_GB_ADDR_CONFIG_READ, 0x0018773f, 0x00104002) 128}; 129 130static const struct soc15_reg_golden golden_settings_sdma_vg12[] = { 131 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_GB_ADDR_CONFIG, 0x0018773f, 0x00104001), 132 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_GB_ADDR_CONFIG_READ, 0x0018773f, 0x00104001), 133 SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_CHICKEN_BITS, 0xfe931f07, 0x02831d07), 134 SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_GB_ADDR_CONFIG, 0x0018773f, 0x00104001), 135 SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_GB_ADDR_CONFIG_READ, 0x0018773f, 0x00104001) 136}; 137 138static const struct soc15_reg_golden golden_settings_sdma_4_1[] = { 139 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_CHICKEN_BITS, 0xfe931f07, 0x02831d07), 140 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_CLK_CTRL, 0xffffffff, 0x3f000100), 141 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_GFX_IB_CNTL, 0x800f0111, 0x00000100), 142 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_GFX_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000), 143 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_POWER_CNTL, 0xfc3fffff, 0x40000051), 144 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_RLC0_IB_CNTL, 0x800f0111, 0x00000100), 145 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_RLC0_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000), 146 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_RLC1_IB_CNTL, 0x800f0111, 0x00000100), 147 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_RLC1_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000), 148 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_UTCL1_PAGE, 0x000003ff, 0x000003c0), 149 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_UTCL1_WATERMK, 0xfc000000, 0x00000000) 150}; 151 152static const struct soc15_reg_golden golden_settings_sdma0_4_2_init[] = { 153 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_RLC0_RB_WPTR_POLL_CNTL, 0xfffffff0, 0x00403000), 154}; 155 156static const struct soc15_reg_golden golden_settings_sdma0_4_2[] = 157{ 158 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_CHICKEN_BITS, 0xfe931f07, 0x02831f07), 159 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_CLK_CTRL, 0xffffffff, 0x3f000100), 160 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_GB_ADDR_CONFIG, 0x0000773f, 0x00004002), 161 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_GB_ADDR_CONFIG_READ, 0x0000773f, 0x00004002), 162 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_GFX_RB_RPTR_ADDR_LO, 0xfffffffd, 0x00000001), 163 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_GFX_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000), 164 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_PAGE_RB_RPTR_ADDR_LO, 0xfffffffd, 0x00000001), 165 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_PAGE_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000), 166 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_RD_BURST_CNTL, 0x0000000f, 0x00000003), 167 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_RLC0_RB_RPTR_ADDR_LO, 0xfffffffd, 0x00000001), 168 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_RLC0_RB_WPTR_POLL_CNTL, 0xfffffff0, 0x00403000), 169 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_RLC1_RB_RPTR_ADDR_LO, 0xfffffffd, 0x00000001), 170 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_RLC1_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000), 171 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_RLC2_RB_RPTR_ADDR_LO, 0xfffffffd, 0x00000001), 172 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_RLC2_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000), 173 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_RLC3_RB_RPTR_ADDR_LO, 0xfffffffd, 0x00000001), 174 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_RLC3_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000), 175 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_RLC4_RB_RPTR_ADDR_LO, 0xfffffffd, 0x00000001), 176 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_RLC4_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000), 177 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_RLC5_RB_RPTR_ADDR_LO, 0xfffffffd, 0x00000001), 178 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_RLC5_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000), 179 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_RLC6_RB_RPTR_ADDR_LO, 0xfffffffd, 0x00000001), 180 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_RLC6_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000), 181 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_RLC7_RB_RPTR_ADDR_LO, 0xfffffffd, 0x00000001), 182 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_RLC7_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000), 183 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_UTCL1_PAGE, 0x000003ff, 0x000003c0), 184}; 185 186static const struct soc15_reg_golden golden_settings_sdma1_4_2[] = { 187 SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_CHICKEN_BITS, 0xfe931f07, 0x02831f07), 188 SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_CLK_CTRL, 0xffffffff, 0x3f000100), 189 SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_GB_ADDR_CONFIG, 0x0000773f, 0x00004002), 190 SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_GB_ADDR_CONFIG_READ, 0x0000773f, 0x00004002), 191 SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_GFX_RB_RPTR_ADDR_LO, 0xfffffffd, 0x00000001), 192 SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_GFX_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000), 193 SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_PAGE_RB_RPTR_ADDR_LO, 0xfffffffd, 0x00000001), 194 SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_PAGE_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000), 195 SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_RD_BURST_CNTL, 0x0000000f, 0x00000003), 196 SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_RLC0_RB_RPTR_ADDR_LO, 0xfffffffd, 0x00000001), 197 SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_RLC0_RB_WPTR_POLL_CNTL, 0xfffffff0, 0x00403000), 198 SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_RLC1_RB_RPTR_ADDR_LO, 0xfffffffd, 0x00000001), 199 SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_RLC1_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000), 200 SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_RLC2_RB_RPTR_ADDR_LO, 0xfffffffd, 0x00000001), 201 SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_RLC2_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000), 202 SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_RLC3_RB_RPTR_ADDR_LO, 0xfffffffd, 0x00000001), 203 SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_RLC3_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000), 204 SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_RLC4_RB_RPTR_ADDR_LO, 0xfffffffd, 0x00000001), 205 SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_RLC4_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000), 206 SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_RLC5_RB_RPTR_ADDR_LO, 0xfffffffd, 0x00000001), 207 SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_RLC5_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000), 208 SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_RLC6_RB_RPTR_ADDR_LO, 0xfffffffd, 0x00000001), 209 SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_RLC6_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000), 210 SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_RLC7_RB_RPTR_ADDR_LO, 0xfffffffd, 0x00000001), 211 SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_RLC7_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000), 212 SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_UTCL1_PAGE, 0x000003ff, 0x000003c0), 213}; 214 215static const struct soc15_reg_golden golden_settings_sdma_rv1[] = 216{ 217 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_GB_ADDR_CONFIG, 0x0018773f, 0x00000002), 218 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_GB_ADDR_CONFIG_READ, 0x0018773f, 0x00000002) 219}; 220 221static const struct soc15_reg_golden golden_settings_sdma_rv2[] = 222{ 223 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_GB_ADDR_CONFIG, 0x0018773f, 0x00003001), 224 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_GB_ADDR_CONFIG_READ, 0x0018773f, 0x00003001) 225}; 226 227static const struct soc15_reg_golden golden_settings_sdma_arct[] = 228{ 229 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_CHICKEN_BITS, 0xfe931f07, 0x02831f07), 230 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_GB_ADDR_CONFIG, 0x0000773f, 0x00004002), 231 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_GB_ADDR_CONFIG_READ, 0x0000773f, 0x00004002), 232 SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_CHICKEN_BITS, 0xfe931f07, 0x02831f07), 233 SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_GB_ADDR_CONFIG, 0x0000773f, 0x00004002), 234 SOC15_REG_GOLDEN_VALUE(SDMA1, 0, mmSDMA1_GB_ADDR_CONFIG_READ, 0x0000773f, 0x00004002), 235 SOC15_REG_GOLDEN_VALUE(SDMA2, 0, mmSDMA2_CHICKEN_BITS, 0xfe931f07, 0x02831f07), 236 SOC15_REG_GOLDEN_VALUE(SDMA2, 0, mmSDMA2_GB_ADDR_CONFIG, 0x0000773f, 0x00004002), 237 SOC15_REG_GOLDEN_VALUE(SDMA2, 0, mmSDMA2_GB_ADDR_CONFIG_READ, 0x0000773f, 0x00004002), 238 SOC15_REG_GOLDEN_VALUE(SDMA3, 0, mmSDMA3_CHICKEN_BITS, 0xfe931f07, 0x02831f07), 239 SOC15_REG_GOLDEN_VALUE(SDMA3, 0, mmSDMA3_GB_ADDR_CONFIG, 0x0000773f, 0x00004002), 240 SOC15_REG_GOLDEN_VALUE(SDMA3, 0, mmSDMA3_GB_ADDR_CONFIG_READ, 0x0000773f, 0x00004002), 241 SOC15_REG_GOLDEN_VALUE(SDMA4, 0, mmSDMA4_CHICKEN_BITS, 0xfe931f07, 0x02831f07), 242 SOC15_REG_GOLDEN_VALUE(SDMA4, 0, mmSDMA4_GB_ADDR_CONFIG, 0x0000773f, 0x00004002), 243 SOC15_REG_GOLDEN_VALUE(SDMA4, 0, mmSDMA4_GB_ADDR_CONFIG_READ, 0x0000773f, 0x00004002), 244 SOC15_REG_GOLDEN_VALUE(SDMA5, 0, mmSDMA5_CHICKEN_BITS, 0xfe931f07, 0x02831f07), 245 SOC15_REG_GOLDEN_VALUE(SDMA5, 0, mmSDMA5_GB_ADDR_CONFIG, 0x0000773f, 0x00004002), 246 SOC15_REG_GOLDEN_VALUE(SDMA5, 0, mmSDMA5_GB_ADDR_CONFIG_READ, 0x0000773f, 0x00004002), 247 SOC15_REG_GOLDEN_VALUE(SDMA6, 0, mmSDMA6_CHICKEN_BITS, 0xfe931f07, 0x02831f07), 248 SOC15_REG_GOLDEN_VALUE(SDMA6, 0, mmSDMA6_GB_ADDR_CONFIG, 0x0000773f, 0x00004002), 249 SOC15_REG_GOLDEN_VALUE(SDMA6, 0, mmSDMA6_GB_ADDR_CONFIG_READ, 0x0000773f, 0x00004002), 250 SOC15_REG_GOLDEN_VALUE(SDMA7, 0, mmSDMA7_CHICKEN_BITS, 0xfe931f07, 0x02831f07), 251 SOC15_REG_GOLDEN_VALUE(SDMA7, 0, mmSDMA7_GB_ADDR_CONFIG, 0x0000773f, 0x00004002), 252 SOC15_REG_GOLDEN_VALUE(SDMA7, 0, mmSDMA7_GB_ADDR_CONFIG_READ, 0x0000773f, 0x00004002) 253}; 254 255static const struct soc15_reg_golden golden_settings_sdma_4_3[] = { 256 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_CHICKEN_BITS, 0xfe931f07, 0x02831f07), 257 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_CLK_CTRL, 0xffffffff, 0x3f000100), 258 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_GB_ADDR_CONFIG, 0x0018773f, 0x00000002), 259 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_GB_ADDR_CONFIG_READ, 0x0018773f, 0x00000002), 260 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_GFX_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000), 261 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_POWER_CNTL, 0x003fff07, 0x40000051), 262 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_RLC0_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000), 263 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_RLC1_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000), 264 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_UTCL1_PAGE, 0x000003ff, 0x000003c0), 265 SOC15_REG_GOLDEN_VALUE(SDMA0, 0, mmSDMA0_UTCL1_WATERMK, 0xfc000000, 0x03fbe1fe) 266}; 267 268static const struct soc15_ras_field_entry sdma_v4_0_ras_fields[] = { 269 { "SDMA_UCODE_BUF_SED", SOC15_REG_ENTRY(SDMA0, 0, mmSDMA0_EDC_COUNTER), 270 SOC15_REG_FIELD(SDMA0_EDC_COUNTER, SDMA_UCODE_BUF_SED), 271 0, 0, 272 }, 273 { "SDMA_RB_CMD_BUF_SED", SOC15_REG_ENTRY(SDMA0, 0, mmSDMA0_EDC_COUNTER), 274 SOC15_REG_FIELD(SDMA0_EDC_COUNTER, SDMA_RB_CMD_BUF_SED), 275 0, 0, 276 }, 277 { "SDMA_IB_CMD_BUF_SED", SOC15_REG_ENTRY(SDMA0, 0, mmSDMA0_EDC_COUNTER), 278 SOC15_REG_FIELD(SDMA0_EDC_COUNTER, SDMA_IB_CMD_BUF_SED), 279 0, 0, 280 }, 281 { "SDMA_UTCL1_RD_FIFO_SED", SOC15_REG_ENTRY(SDMA0, 0, mmSDMA0_EDC_COUNTER), 282 SOC15_REG_FIELD(SDMA0_EDC_COUNTER, SDMA_UTCL1_RD_FIFO_SED), 283 0, 0, 284 }, 285 { "SDMA_UTCL1_RDBST_FIFO_SED", SOC15_REG_ENTRY(SDMA0, 0, mmSDMA0_EDC_COUNTER), 286 SOC15_REG_FIELD(SDMA0_EDC_COUNTER, SDMA_UTCL1_RDBST_FIFO_SED), 287 0, 0, 288 }, 289 { "SDMA_DATA_LUT_FIFO_SED", SOC15_REG_ENTRY(SDMA0, 0, mmSDMA0_EDC_COUNTER), 290 SOC15_REG_FIELD(SDMA0_EDC_COUNTER, SDMA_DATA_LUT_FIFO_SED), 291 0, 0, 292 }, 293 { "SDMA_MBANK_DATA_BUF0_SED", SOC15_REG_ENTRY(SDMA0, 0, mmSDMA0_EDC_COUNTER), 294 SOC15_REG_FIELD(SDMA0_EDC_COUNTER, SDMA_MBANK_DATA_BUF0_SED), 295 0, 0, 296 }, 297 { "SDMA_MBANK_DATA_BUF1_SED", SOC15_REG_ENTRY(SDMA0, 0, mmSDMA0_EDC_COUNTER), 298 SOC15_REG_FIELD(SDMA0_EDC_COUNTER, SDMA_MBANK_DATA_BUF1_SED), 299 0, 0, 300 }, 301 { "SDMA_MBANK_DATA_BUF2_SED", SOC15_REG_ENTRY(SDMA0, 0, mmSDMA0_EDC_COUNTER), 302 SOC15_REG_FIELD(SDMA0_EDC_COUNTER, SDMA_MBANK_DATA_BUF2_SED), 303 0, 0, 304 }, 305 { "SDMA_MBANK_DATA_BUF3_SED", SOC15_REG_ENTRY(SDMA0, 0, mmSDMA0_EDC_COUNTER), 306 SOC15_REG_FIELD(SDMA0_EDC_COUNTER, SDMA_MBANK_DATA_BUF3_SED), 307 0, 0, 308 }, 309 { "SDMA_MBANK_DATA_BUF4_SED", SOC15_REG_ENTRY(SDMA0, 0, mmSDMA0_EDC_COUNTER), 310 SOC15_REG_FIELD(SDMA0_EDC_COUNTER, SDMA_MBANK_DATA_BUF4_SED), 311 0, 0, 312 }, 313 { "SDMA_MBANK_DATA_BUF5_SED", SOC15_REG_ENTRY(SDMA0, 0, mmSDMA0_EDC_COUNTER), 314 SOC15_REG_FIELD(SDMA0_EDC_COUNTER, SDMA_MBANK_DATA_BUF5_SED), 315 0, 0, 316 }, 317 { "SDMA_MBANK_DATA_BUF6_SED", SOC15_REG_ENTRY(SDMA0, 0, mmSDMA0_EDC_COUNTER), 318 SOC15_REG_FIELD(SDMA0_EDC_COUNTER, SDMA_MBANK_DATA_BUF6_SED), 319 0, 0, 320 }, 321 { "SDMA_MBANK_DATA_BUF7_SED", SOC15_REG_ENTRY(SDMA0, 0, mmSDMA0_EDC_COUNTER), 322 SOC15_REG_FIELD(SDMA0_EDC_COUNTER, SDMA_MBANK_DATA_BUF7_SED), 323 0, 0, 324 }, 325 { "SDMA_MBANK_DATA_BUF8_SED", SOC15_REG_ENTRY(SDMA0, 0, mmSDMA0_EDC_COUNTER), 326 SOC15_REG_FIELD(SDMA0_EDC_COUNTER, SDMA_MBANK_DATA_BUF8_SED), 327 0, 0, 328 }, 329 { "SDMA_MBANK_DATA_BUF9_SED", SOC15_REG_ENTRY(SDMA0, 0, mmSDMA0_EDC_COUNTER), 330 SOC15_REG_FIELD(SDMA0_EDC_COUNTER, SDMA_MBANK_DATA_BUF9_SED), 331 0, 0, 332 }, 333 { "SDMA_MBANK_DATA_BUF10_SED", SOC15_REG_ENTRY(SDMA0, 0, mmSDMA0_EDC_COUNTER), 334 SOC15_REG_FIELD(SDMA0_EDC_COUNTER, SDMA_MBANK_DATA_BUF10_SED), 335 0, 0, 336 }, 337 { "SDMA_MBANK_DATA_BUF11_SED", SOC15_REG_ENTRY(SDMA0, 0, mmSDMA0_EDC_COUNTER), 338 SOC15_REG_FIELD(SDMA0_EDC_COUNTER, SDMA_MBANK_DATA_BUF11_SED), 339 0, 0, 340 }, 341 { "SDMA_MBANK_DATA_BUF12_SED", SOC15_REG_ENTRY(SDMA0, 0, mmSDMA0_EDC_COUNTER), 342 SOC15_REG_FIELD(SDMA0_EDC_COUNTER, SDMA_MBANK_DATA_BUF12_SED), 343 0, 0, 344 }, 345 { "SDMA_MBANK_DATA_BUF13_SED", SOC15_REG_ENTRY(SDMA0, 0, mmSDMA0_EDC_COUNTER), 346 SOC15_REG_FIELD(SDMA0_EDC_COUNTER, SDMA_MBANK_DATA_BUF13_SED), 347 0, 0, 348 }, 349 { "SDMA_MBANK_DATA_BUF14_SED", SOC15_REG_ENTRY(SDMA0, 0, mmSDMA0_EDC_COUNTER), 350 SOC15_REG_FIELD(SDMA0_EDC_COUNTER, SDMA_MBANK_DATA_BUF14_SED), 351 0, 0, 352 }, 353 { "SDMA_MBANK_DATA_BUF15_SED", SOC15_REG_ENTRY(SDMA0, 0, mmSDMA0_EDC_COUNTER), 354 SOC15_REG_FIELD(SDMA0_EDC_COUNTER, SDMA_MBANK_DATA_BUF15_SED), 355 0, 0, 356 }, 357 { "SDMA_SPLIT_DAT_BUF_SED", SOC15_REG_ENTRY(SDMA0, 0, mmSDMA0_EDC_COUNTER), 358 SOC15_REG_FIELD(SDMA0_EDC_COUNTER, SDMA_SPLIT_DAT_BUF_SED), 359 0, 0, 360 }, 361 { "SDMA_MC_WR_ADDR_FIFO_SED", SOC15_REG_ENTRY(SDMA0, 0, mmSDMA0_EDC_COUNTER), 362 SOC15_REG_FIELD(SDMA0_EDC_COUNTER, SDMA_MC_WR_ADDR_FIFO_SED), 363 0, 0, 364 }, 365}; 366 367static u32 sdma_v4_0_get_reg_offset(struct amdgpu_device *adev, 368 u32 instance, u32 offset) 369{ 370 switch (instance) { 371 case 0: 372 return (adev->reg_offset[SDMA0_HWIP][0][0] + offset); 373 case 1: 374 return (adev->reg_offset[SDMA1_HWIP][0][0] + offset); 375 case 2: 376 return (adev->reg_offset[SDMA2_HWIP][0][1] + offset); 377 case 3: 378 return (adev->reg_offset[SDMA3_HWIP][0][1] + offset); 379 case 4: 380 return (adev->reg_offset[SDMA4_HWIP][0][1] + offset); 381 case 5: 382 return (adev->reg_offset[SDMA5_HWIP][0][1] + offset); 383 case 6: 384 return (adev->reg_offset[SDMA6_HWIP][0][1] + offset); 385 case 7: 386 return (adev->reg_offset[SDMA7_HWIP][0][1] + offset); 387 default: 388 break; 389 } 390 return 0; 391} 392 393static unsigned sdma_v4_0_seq_to_irq_id(int seq_num) 394{ 395 switch (seq_num) { 396 case 0: 397 return SOC15_IH_CLIENTID_SDMA0; 398 case 1: 399 return SOC15_IH_CLIENTID_SDMA1; 400 case 2: 401 return SOC15_IH_CLIENTID_SDMA2; 402 case 3: 403 return SOC15_IH_CLIENTID_SDMA3; 404 case 4: 405 return SOC15_IH_CLIENTID_SDMA4; 406 case 5: 407 return SOC15_IH_CLIENTID_SDMA5; 408 case 6: 409 return SOC15_IH_CLIENTID_SDMA6; 410 case 7: 411 return SOC15_IH_CLIENTID_SDMA7; 412 default: 413 break; 414 } 415 return -EINVAL; 416} 417 418static int sdma_v4_0_irq_id_to_seq(unsigned client_id) 419{ 420 switch (client_id) { 421 case SOC15_IH_CLIENTID_SDMA0: 422 return 0; 423 case SOC15_IH_CLIENTID_SDMA1: 424 return 1; 425 case SOC15_IH_CLIENTID_SDMA2: 426 return 2; 427 case SOC15_IH_CLIENTID_SDMA3: 428 return 3; 429 case SOC15_IH_CLIENTID_SDMA4: 430 return 4; 431 case SOC15_IH_CLIENTID_SDMA5: 432 return 5; 433 case SOC15_IH_CLIENTID_SDMA6: 434 return 6; 435 case SOC15_IH_CLIENTID_SDMA7: 436 return 7; 437 default: 438 break; 439 } 440 return -EINVAL; 441} 442 443static void sdma_v4_0_init_golden_registers(struct amdgpu_device *adev) 444{ 445 switch (adev->asic_type) { 446 case CHIP_VEGA10: 447 soc15_program_register_sequence(adev, 448 golden_settings_sdma_4, 449 ARRAY_SIZE(golden_settings_sdma_4)); 450 soc15_program_register_sequence(adev, 451 golden_settings_sdma_vg10, 452 ARRAY_SIZE(golden_settings_sdma_vg10)); 453 break; 454 case CHIP_VEGA12: 455 soc15_program_register_sequence(adev, 456 golden_settings_sdma_4, 457 ARRAY_SIZE(golden_settings_sdma_4)); 458 soc15_program_register_sequence(adev, 459 golden_settings_sdma_vg12, 460 ARRAY_SIZE(golden_settings_sdma_vg12)); 461 break; 462 case CHIP_VEGA20: 463 soc15_program_register_sequence(adev, 464 golden_settings_sdma0_4_2_init, 465 ARRAY_SIZE(golden_settings_sdma0_4_2_init)); 466 soc15_program_register_sequence(adev, 467 golden_settings_sdma0_4_2, 468 ARRAY_SIZE(golden_settings_sdma0_4_2)); 469 soc15_program_register_sequence(adev, 470 golden_settings_sdma1_4_2, 471 ARRAY_SIZE(golden_settings_sdma1_4_2)); 472 break; 473 case CHIP_ARCTURUS: 474 soc15_program_register_sequence(adev, 475 golden_settings_sdma_arct, 476 ARRAY_SIZE(golden_settings_sdma_arct)); 477 break; 478 case CHIP_RAVEN: 479 soc15_program_register_sequence(adev, 480 golden_settings_sdma_4_1, 481 ARRAY_SIZE(golden_settings_sdma_4_1)); 482 if (adev->rev_id >= 8) 483 soc15_program_register_sequence(adev, 484 golden_settings_sdma_rv2, 485 ARRAY_SIZE(golden_settings_sdma_rv2)); 486 else 487 soc15_program_register_sequence(adev, 488 golden_settings_sdma_rv1, 489 ARRAY_SIZE(golden_settings_sdma_rv1)); 490 break; 491 case CHIP_RENOIR: 492 soc15_program_register_sequence(adev, 493 golden_settings_sdma_4_3, 494 ARRAY_SIZE(golden_settings_sdma_4_3)); 495 break; 496 default: 497 break; 498 } 499} 500 501static int sdma_v4_0_init_inst_ctx(struct amdgpu_sdma_instance *sdma_inst) 502{ 503 int err = 0; 504 const struct sdma_firmware_header_v1_0 *hdr; 505 506 err = amdgpu_ucode_validate(sdma_inst->fw); 507 if (err) 508 return err; 509 510 hdr = (const struct sdma_firmware_header_v1_0 *)sdma_inst->fw->data; 511 sdma_inst->fw_version = le32_to_cpu(hdr->header.ucode_version); 512 sdma_inst->feature_version = le32_to_cpu(hdr->ucode_feature_version); 513 514 if (sdma_inst->feature_version >= 20) 515 sdma_inst->burst_nop = true; 516 517 return 0; 518} 519 520static void sdma_v4_0_destroy_inst_ctx(struct amdgpu_device *adev) 521{ 522 int i; 523 524 for (i = 0; i < adev->sdma.num_instances; i++) { 525 if (adev->sdma.instance[i].fw != NULL) 526 release_firmware(adev->sdma.instance[i].fw); 527 528 /* arcturus shares the same FW memory across 529 all SDMA isntances */ 530 if (adev->asic_type == CHIP_ARCTURUS) 531 break; 532 } 533 534 memset((void*)adev->sdma.instance, 0, 535 sizeof(struct amdgpu_sdma_instance) * AMDGPU_MAX_SDMA_INSTANCES); 536} 537 538/** 539 * sdma_v4_0_init_microcode - load ucode images from disk 540 * 541 * @adev: amdgpu_device pointer 542 * 543 * Use the firmware interface to load the ucode images into 544 * the driver (not loaded into hw). 545 * Returns 0 on success, error on failure. 546 */ 547 548// emulation only, won't work on real chip 549// vega10 real chip need to use PSP to load firmware 550static int sdma_v4_0_init_microcode(struct amdgpu_device *adev) 551{ 552 const char *chip_name; 553 char fw_name[30]; 554 int err = 0, i; 555 struct amdgpu_firmware_info *info = NULL; 556 const struct common_firmware_header *header = NULL; 557 558 DRM_DEBUG("\n"); 559 560 switch (adev->asic_type) { 561 case CHIP_VEGA10: 562 chip_name = "vega10"; 563 break; 564 case CHIP_VEGA12: 565 chip_name = "vega12"; 566 break; 567 case CHIP_VEGA20: 568 chip_name = "vega20"; 569 break; 570 case CHIP_RAVEN: 571 if (adev->rev_id >= 8) 572 chip_name = "raven2"; 573 else if (adev->pdev->device == 0x15d8) 574 chip_name = "picasso"; 575 else 576 chip_name = "raven"; 577 break; 578 case CHIP_ARCTURUS: 579 chip_name = "arcturus"; 580 break; 581 case CHIP_RENOIR: 582 chip_name = "renoir"; 583 break; 584 default: 585 BUG(); 586 } 587 588 snprintf(fw_name, sizeof(fw_name), "amdgpu/%s_sdma.bin", chip_name); 589 590 err = request_firmware(&adev->sdma.instance[0].fw, fw_name, adev->dev); 591 if (err) 592 goto out; 593 594 err = sdma_v4_0_init_inst_ctx(&adev->sdma.instance[0]); 595 if (err) 596 goto out; 597 598 for (i = 1; i < adev->sdma.num_instances; i++) { 599 if (adev->asic_type == CHIP_ARCTURUS) { 600 /* Acturus will leverage the same FW memory 601 for every SDMA instance */ 602 memcpy((void*)&adev->sdma.instance[i], 603 (void*)&adev->sdma.instance[0], 604 sizeof(struct amdgpu_sdma_instance)); 605 } 606 else { 607 snprintf(fw_name, sizeof(fw_name), "amdgpu/%s_sdma%d.bin", chip_name, i); 608 609 err = request_firmware(&adev->sdma.instance[i].fw, fw_name, adev->dev); 610 if (err) 611 goto out; 612 613 err = sdma_v4_0_init_inst_ctx(&adev->sdma.instance[i]); 614 if (err) 615 goto out; 616 } 617 } 618 619 DRM_DEBUG("psp_load == '%s'\n", 620 adev->firmware.load_type == AMDGPU_FW_LOAD_PSP ? "true" : "false"); 621 622 if (adev->firmware.load_type == AMDGPU_FW_LOAD_PSP) { 623 for (i = 0; i < adev->sdma.num_instances; i++) { 624 info = &adev->firmware.ucode[AMDGPU_UCODE_ID_SDMA0 + i]; 625 info->ucode_id = AMDGPU_UCODE_ID_SDMA0 + i; 626 info->fw = adev->sdma.instance[i].fw; 627 header = (const struct common_firmware_header *)info->fw->data; 628 adev->firmware.fw_size += 629 ALIGN(le32_to_cpu(header->ucode_size_bytes), PAGE_SIZE); 630 } 631 } 632 633out: 634 if (err) { 635 DRM_ERROR("sdma_v4_0: Failed to load firmware \"%s\"\n", fw_name); 636 sdma_v4_0_destroy_inst_ctx(adev); 637 } 638 return err; 639} 640 641/** 642 * sdma_v4_0_ring_get_rptr - get the current read pointer 643 * 644 * @ring: amdgpu ring pointer 645 * 646 * Get the current rptr from the hardware (VEGA10+). 647 */ 648static uint64_t sdma_v4_0_ring_get_rptr(struct amdgpu_ring *ring) 649{ 650 volatile u64 *rptr; 651 652 /* XXX check if swapping is necessary on BE */ 653 rptr = ((volatile u64 *)&ring->adev->wb.wb[ring->rptr_offs]); 654 655 DRM_DEBUG("rptr before shift == 0x%016"PRIx64"\n", *rptr); 656 return ((*rptr) >> 2); 657} 658 659/** 660 * sdma_v4_0_ring_get_wptr - get the current write pointer 661 * 662 * @ring: amdgpu ring pointer 663 * 664 * Get the current wptr from the hardware (VEGA10+). 665 */ 666static uint64_t sdma_v4_0_ring_get_wptr(struct amdgpu_ring *ring) 667{ 668 struct amdgpu_device *adev = ring->adev; 669 u64 wptr; 670 671 if (ring->use_doorbell) { 672 /* XXX check if swapping is necessary on BE */ 673 wptr = READ_ONCE(*((volatile u64 *)&adev->wb.wb[ring->wptr_offs])); 674 DRM_DEBUG("wptr/doorbell before shift == 0x%016"PRIx64"\n", wptr); 675 } else { 676 wptr = RREG32_SDMA(ring->me, mmSDMA0_GFX_RB_WPTR_HI); 677 wptr = wptr << 32; 678 wptr |= RREG32_SDMA(ring->me, mmSDMA0_GFX_RB_WPTR); 679 DRM_DEBUG("wptr before shift [%i] wptr == 0x%016"PRIx64"\n", 680 ring->me, wptr); 681 } 682 683 return wptr >> 2; 684} 685 686/** 687 * sdma_v4_0_ring_set_wptr - commit the write pointer 688 * 689 * @ring: amdgpu ring pointer 690 * 691 * Write the wptr back to the hardware (VEGA10+). 692 */ 693static void sdma_v4_0_ring_set_wptr(struct amdgpu_ring *ring) 694{ 695 struct amdgpu_device *adev = ring->adev; 696 697 DRM_DEBUG("Setting write pointer\n"); 698 if (ring->use_doorbell) { 699 volatile u64 *wb = (volatile u64 *)&adev->wb.wb[ring->wptr_offs]; 700 701 DRM_DEBUG("Using doorbell -- " 702 "wptr_offs == 0x%08x " 703 "lower_32_bits(ring->wptr) << 2 == 0x%08x " 704 "upper_32_bits(ring->wptr) << 2 == 0x%08x\n", 705 ring->wptr_offs, 706 lower_32_bits(ring->wptr << 2), 707 upper_32_bits(ring->wptr << 2)); 708 /* XXX check if swapping is necessary on BE */ 709 WRITE_ONCE(*wb, (ring->wptr << 2)); 710 DRM_DEBUG("calling WDOORBELL64(0x%08x, 0x%016"PRIx64")\n", 711 ring->doorbell_index, ring->wptr << 2); 712 WDOORBELL64(ring->doorbell_index, ring->wptr << 2); 713 } else { 714 DRM_DEBUG("Not using doorbell -- " 715 "mmSDMA%i_GFX_RB_WPTR == 0x%08x " 716 "mmSDMA%i_GFX_RB_WPTR_HI == 0x%08x\n", 717 ring->me, 718 lower_32_bits(ring->wptr << 2), 719 ring->me, 720 upper_32_bits(ring->wptr << 2)); 721 WREG32_SDMA(ring->me, mmSDMA0_GFX_RB_WPTR, 722 lower_32_bits(ring->wptr << 2)); 723 WREG32_SDMA(ring->me, mmSDMA0_GFX_RB_WPTR_HI, 724 upper_32_bits(ring->wptr << 2)); 725 } 726} 727 728/** 729 * sdma_v4_0_page_ring_get_wptr - get the current write pointer 730 * 731 * @ring: amdgpu ring pointer 732 * 733 * Get the current wptr from the hardware (VEGA10+). 734 */ 735static uint64_t sdma_v4_0_page_ring_get_wptr(struct amdgpu_ring *ring) 736{ 737 struct amdgpu_device *adev = ring->adev; 738 u64 wptr; 739 740 if (ring->use_doorbell) { 741 /* XXX check if swapping is necessary on BE */ 742 wptr = READ_ONCE(*((volatile u64 *)&adev->wb.wb[ring->wptr_offs])); 743 } else { 744 wptr = RREG32_SDMA(ring->me, mmSDMA0_PAGE_RB_WPTR_HI); 745 wptr = wptr << 32; 746 wptr |= RREG32_SDMA(ring->me, mmSDMA0_PAGE_RB_WPTR); 747 } 748 749 return wptr >> 2; 750} 751 752/** 753 * sdma_v4_0_ring_set_wptr - commit the write pointer 754 * 755 * @ring: amdgpu ring pointer 756 * 757 * Write the wptr back to the hardware (VEGA10+). 758 */ 759static void sdma_v4_0_page_ring_set_wptr(struct amdgpu_ring *ring) 760{ 761 struct amdgpu_device *adev = ring->adev; 762 763 if (ring->use_doorbell) { 764 volatile u64 *wb = (volatile u64 *)&adev->wb.wb[ring->wptr_offs]; 765 766 /* XXX check if swapping is necessary on BE */ 767 WRITE_ONCE(*wb, (ring->wptr << 2)); 768 WDOORBELL64(ring->doorbell_index, ring->wptr << 2); 769 } else { 770 uint64_t wptr = ring->wptr << 2; 771 772 WREG32_SDMA(ring->me, mmSDMA0_PAGE_RB_WPTR, 773 lower_32_bits(wptr)); 774 WREG32_SDMA(ring->me, mmSDMA0_PAGE_RB_WPTR_HI, 775 upper_32_bits(wptr)); 776 } 777} 778 779static void sdma_v4_0_ring_insert_nop(struct amdgpu_ring *ring, uint32_t count) 780{ 781 struct amdgpu_sdma_instance *sdma = amdgpu_sdma_get_instance_from_ring(ring); 782 int i; 783 784 for (i = 0; i < count; i++) 785 if (sdma && sdma->burst_nop && (i == 0)) 786 amdgpu_ring_write(ring, ring->funcs->nop | 787 SDMA_PKT_NOP_HEADER_COUNT(count - 1)); 788 else 789 amdgpu_ring_write(ring, ring->funcs->nop); 790} 791 792/** 793 * sdma_v4_0_ring_emit_ib - Schedule an IB on the DMA engine 794 * 795 * @ring: amdgpu ring pointer 796 * @ib: IB object to schedule 797 * 798 * Schedule an IB in the DMA ring (VEGA10). 799 */ 800static void sdma_v4_0_ring_emit_ib(struct amdgpu_ring *ring, 801 struct amdgpu_job *job, 802 struct amdgpu_ib *ib, 803 uint32_t flags) 804{ 805 unsigned vmid = AMDGPU_JOB_GET_VMID(job); 806 807 /* IB packet must end on a 8 DW boundary */ 808 sdma_v4_0_ring_insert_nop(ring, (2 - lower_32_bits(ring->wptr)) & 7); 809 810 amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_INDIRECT) | 811 SDMA_PKT_INDIRECT_HEADER_VMID(vmid & 0xf)); 812 /* base must be 32 byte aligned */ 813 amdgpu_ring_write(ring, lower_32_bits(ib->gpu_addr) & 0xffffffe0); 814 amdgpu_ring_write(ring, upper_32_bits(ib->gpu_addr)); 815 amdgpu_ring_write(ring, ib->length_dw); 816 amdgpu_ring_write(ring, 0); 817 amdgpu_ring_write(ring, 0); 818 819} 820 821static void sdma_v4_0_wait_reg_mem(struct amdgpu_ring *ring, 822 int mem_space, int hdp, 823 uint32_t addr0, uint32_t addr1, 824 uint32_t ref, uint32_t mask, 825 uint32_t inv) 826{ 827 amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_POLL_REGMEM) | 828 SDMA_PKT_POLL_REGMEM_HEADER_HDP_FLUSH(hdp) | 829 SDMA_PKT_POLL_REGMEM_HEADER_MEM_POLL(mem_space) | 830 SDMA_PKT_POLL_REGMEM_HEADER_FUNC(3)); /* == */ 831 if (mem_space) { 832 /* memory */ 833 amdgpu_ring_write(ring, addr0); 834 amdgpu_ring_write(ring, addr1); 835 } else { 836 /* registers */ 837 amdgpu_ring_write(ring, addr0 << 2); 838 amdgpu_ring_write(ring, addr1 << 2); 839 } 840 amdgpu_ring_write(ring, ref); /* reference */ 841 amdgpu_ring_write(ring, mask); /* mask */ 842 amdgpu_ring_write(ring, SDMA_PKT_POLL_REGMEM_DW5_RETRY_COUNT(0xfff) | 843 SDMA_PKT_POLL_REGMEM_DW5_INTERVAL(inv)); /* retry count, poll interval */ 844} 845 846/** 847 * sdma_v4_0_ring_emit_hdp_flush - emit an hdp flush on the DMA ring 848 * 849 * @ring: amdgpu ring pointer 850 * 851 * Emit an hdp flush packet on the requested DMA ring. 852 */ 853static void sdma_v4_0_ring_emit_hdp_flush(struct amdgpu_ring *ring) 854{ 855 struct amdgpu_device *adev = ring->adev; 856 u32 ref_and_mask = 0; 857 const struct nbio_hdp_flush_reg *nbio_hf_reg = adev->nbio.hdp_flush_reg; 858 859 ref_and_mask = nbio_hf_reg->ref_and_mask_sdma0 << ring->me; 860 861 sdma_v4_0_wait_reg_mem(ring, 0, 1, 862 adev->nbio.funcs->get_hdp_flush_done_offset(adev), 863 adev->nbio.funcs->get_hdp_flush_req_offset(adev), 864 ref_and_mask, ref_and_mask, 10); 865} 866 867/** 868 * sdma_v4_0_ring_emit_fence - emit a fence on the DMA ring 869 * 870 * @ring: amdgpu ring pointer 871 * @fence: amdgpu fence object 872 * 873 * Add a DMA fence packet to the ring to write 874 * the fence seq number and DMA trap packet to generate 875 * an interrupt if needed (VEGA10). 876 */ 877static void sdma_v4_0_ring_emit_fence(struct amdgpu_ring *ring, u64 addr, u64 seq, 878 unsigned flags) 879{ 880 bool write64bit = flags & AMDGPU_FENCE_FLAG_64BIT; 881 /* write the fence */ 882 amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_FENCE)); 883 /* zero in first two bits */ 884 BUG_ON(addr & 0x3); 885 amdgpu_ring_write(ring, lower_32_bits(addr)); 886 amdgpu_ring_write(ring, upper_32_bits(addr)); 887 amdgpu_ring_write(ring, lower_32_bits(seq)); 888 889 /* optionally write high bits as well */ 890 if (write64bit) { 891 addr += 4; 892 amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_FENCE)); 893 /* zero in first two bits */ 894 BUG_ON(addr & 0x3); 895 amdgpu_ring_write(ring, lower_32_bits(addr)); 896 amdgpu_ring_write(ring, upper_32_bits(addr)); 897 amdgpu_ring_write(ring, upper_32_bits(seq)); 898 } 899 900 /* generate an interrupt */ 901 amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_TRAP)); 902 amdgpu_ring_write(ring, SDMA_PKT_TRAP_INT_CONTEXT_INT_CONTEXT(0)); 903} 904 905 906/** 907 * sdma_v4_0_gfx_stop - stop the gfx async dma engines 908 * 909 * @adev: amdgpu_device pointer 910 * 911 * Stop the gfx async dma ring buffers (VEGA10). 912 */ 913static void sdma_v4_0_gfx_stop(struct amdgpu_device *adev) 914{ 915 struct amdgpu_ring *sdma[AMDGPU_MAX_SDMA_INSTANCES]; 916 u32 rb_cntl, ib_cntl; 917 int i, unset = 0; 918 919 for (i = 0; i < adev->sdma.num_instances; i++) { 920 sdma[i] = &adev->sdma.instance[i].ring; 921 922 if ((adev->mman.buffer_funcs_ring == sdma[i]) && unset != 1) { 923 amdgpu_ttm_set_buffer_funcs_status(adev, false); 924 unset = 1; 925 } 926 927 rb_cntl = RREG32_SDMA(i, mmSDMA0_GFX_RB_CNTL); 928 rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RB_ENABLE, 0); 929 WREG32_SDMA(i, mmSDMA0_GFX_RB_CNTL, rb_cntl); 930 ib_cntl = RREG32_SDMA(i, mmSDMA0_GFX_IB_CNTL); 931 ib_cntl = REG_SET_FIELD(ib_cntl, SDMA0_GFX_IB_CNTL, IB_ENABLE, 0); 932 WREG32_SDMA(i, mmSDMA0_GFX_IB_CNTL, ib_cntl); 933 934 sdma[i]->sched.ready = false; 935 } 936} 937 938/** 939 * sdma_v4_0_rlc_stop - stop the compute async dma engines 940 * 941 * @adev: amdgpu_device pointer 942 * 943 * Stop the compute async dma queues (VEGA10). 944 */ 945static void sdma_v4_0_rlc_stop(struct amdgpu_device *adev) 946{ 947 /* XXX todo */ 948} 949 950/** 951 * sdma_v4_0_page_stop - stop the page async dma engines 952 * 953 * @adev: amdgpu_device pointer 954 * 955 * Stop the page async dma ring buffers (VEGA10). 956 */ 957static void sdma_v4_0_page_stop(struct amdgpu_device *adev) 958{ 959 struct amdgpu_ring *sdma[AMDGPU_MAX_SDMA_INSTANCES]; 960 u32 rb_cntl, ib_cntl; 961 int i; 962 bool unset = false; 963 964 for (i = 0; i < adev->sdma.num_instances; i++) { 965 sdma[i] = &adev->sdma.instance[i].page; 966 967 if ((adev->mman.buffer_funcs_ring == sdma[i]) && 968 (unset == false)) { 969 amdgpu_ttm_set_buffer_funcs_status(adev, false); 970 unset = true; 971 } 972 973 rb_cntl = RREG32_SDMA(i, mmSDMA0_PAGE_RB_CNTL); 974 rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_PAGE_RB_CNTL, 975 RB_ENABLE, 0); 976 WREG32_SDMA(i, mmSDMA0_PAGE_RB_CNTL, rb_cntl); 977 ib_cntl = RREG32_SDMA(i, mmSDMA0_PAGE_IB_CNTL); 978 ib_cntl = REG_SET_FIELD(ib_cntl, SDMA0_PAGE_IB_CNTL, 979 IB_ENABLE, 0); 980 WREG32_SDMA(i, mmSDMA0_PAGE_IB_CNTL, ib_cntl); 981 982 sdma[i]->sched.ready = false; 983 } 984} 985 986/** 987 * sdma_v_0_ctx_switch_enable - stop the async dma engines context switch 988 * 989 * @adev: amdgpu_device pointer 990 * @enable: enable/disable the DMA MEs context switch. 991 * 992 * Halt or unhalt the async dma engines context switch (VEGA10). 993 */ 994static void sdma_v4_0_ctx_switch_enable(struct amdgpu_device *adev, bool enable) 995{ 996 u32 f32_cntl, phase_quantum = 0; 997 int i; 998 999 if (amdgpu_sdma_phase_quantum) { 1000 unsigned value = amdgpu_sdma_phase_quantum; 1001 unsigned unit = 0; 1002 1003 while (value > (SDMA0_PHASE0_QUANTUM__VALUE_MASK >> 1004 SDMA0_PHASE0_QUANTUM__VALUE__SHIFT)) { 1005 value = (value + 1) >> 1; 1006 unit++; 1007 } 1008 if (unit > (SDMA0_PHASE0_QUANTUM__UNIT_MASK >> 1009 SDMA0_PHASE0_QUANTUM__UNIT__SHIFT)) { 1010 value = (SDMA0_PHASE0_QUANTUM__VALUE_MASK >> 1011 SDMA0_PHASE0_QUANTUM__VALUE__SHIFT); 1012 unit = (SDMA0_PHASE0_QUANTUM__UNIT_MASK >> 1013 SDMA0_PHASE0_QUANTUM__UNIT__SHIFT); 1014 WARN_ONCE(1, 1015 "clamping sdma_phase_quantum to %uK clock cycles\n", 1016 value << unit); 1017 } 1018 phase_quantum = 1019 value << SDMA0_PHASE0_QUANTUM__VALUE__SHIFT | 1020 unit << SDMA0_PHASE0_QUANTUM__UNIT__SHIFT; 1021 } 1022 1023 for (i = 0; i < adev->sdma.num_instances; i++) { 1024 f32_cntl = RREG32_SDMA(i, mmSDMA0_CNTL); 1025 f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_CNTL, 1026 AUTO_CTXSW_ENABLE, enable ? 1 : 0); 1027 if (enable && amdgpu_sdma_phase_quantum) { 1028 WREG32_SDMA(i, mmSDMA0_PHASE0_QUANTUM, phase_quantum); 1029 WREG32_SDMA(i, mmSDMA0_PHASE1_QUANTUM, phase_quantum); 1030 WREG32_SDMA(i, mmSDMA0_PHASE2_QUANTUM, phase_quantum); 1031 } 1032 WREG32_SDMA(i, mmSDMA0_CNTL, f32_cntl); 1033 } 1034 1035} 1036 1037/** 1038 * sdma_v4_0_enable - stop the async dma engines 1039 * 1040 * @adev: amdgpu_device pointer 1041 * @enable: enable/disable the DMA MEs. 1042 * 1043 * Halt or unhalt the async dma engines (VEGA10). 1044 */ 1045static void sdma_v4_0_enable(struct amdgpu_device *adev, bool enable) 1046{ 1047 u32 f32_cntl; 1048 int i; 1049 1050 if (enable == false) { 1051 sdma_v4_0_gfx_stop(adev); 1052 sdma_v4_0_rlc_stop(adev); 1053 if (adev->sdma.has_page_queue) 1054 sdma_v4_0_page_stop(adev); 1055 } 1056 1057 for (i = 0; i < adev->sdma.num_instances; i++) { 1058 f32_cntl = RREG32_SDMA(i, mmSDMA0_F32_CNTL); 1059 f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_F32_CNTL, HALT, enable ? 0 : 1); 1060 WREG32_SDMA(i, mmSDMA0_F32_CNTL, f32_cntl); 1061 } 1062} 1063 1064/** 1065 * sdma_v4_0_rb_cntl - get parameters for rb_cntl 1066 */ 1067static uint32_t sdma_v4_0_rb_cntl(struct amdgpu_ring *ring, uint32_t rb_cntl) 1068{ 1069 /* Set ring buffer size in dwords */ 1070 uint32_t rb_bufsz = order_base_2(ring->ring_size / 4); 1071 1072 rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RB_SIZE, rb_bufsz); 1073#ifdef __BIG_ENDIAN 1074 rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RB_SWAP_ENABLE, 1); 1075 rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, 1076 RPTR_WRITEBACK_SWAP_ENABLE, 1); 1077#endif 1078 return rb_cntl; 1079} 1080 1081/** 1082 * sdma_v4_0_gfx_resume - setup and start the async dma engines 1083 * 1084 * @adev: amdgpu_device pointer 1085 * @i: instance to resume 1086 * 1087 * Set up the gfx DMA ring buffers and enable them (VEGA10). 1088 * Returns 0 for success, error for failure. 1089 */ 1090static void sdma_v4_0_gfx_resume(struct amdgpu_device *adev, unsigned int i) 1091{ 1092 struct amdgpu_ring *ring = &adev->sdma.instance[i].ring; 1093 u32 rb_cntl, ib_cntl, wptr_poll_cntl; 1094 u32 wb_offset; 1095 u32 doorbell; 1096 u32 doorbell_offset; 1097 u64 wptr_gpu_addr; 1098 1099 wb_offset = (ring->rptr_offs * 4); 1100 1101 rb_cntl = RREG32_SDMA(i, mmSDMA0_GFX_RB_CNTL); 1102 rb_cntl = sdma_v4_0_rb_cntl(ring, rb_cntl); 1103 WREG32_SDMA(i, mmSDMA0_GFX_RB_CNTL, rb_cntl); 1104 1105 /* Initialize the ring buffer's read and write pointers */ 1106 WREG32_SDMA(i, mmSDMA0_GFX_RB_RPTR, 0); 1107 WREG32_SDMA(i, mmSDMA0_GFX_RB_RPTR_HI, 0); 1108 WREG32_SDMA(i, mmSDMA0_GFX_RB_WPTR, 0); 1109 WREG32_SDMA(i, mmSDMA0_GFX_RB_WPTR_HI, 0); 1110 1111 /* set the wb address whether it's enabled or not */ 1112 WREG32_SDMA(i, mmSDMA0_GFX_RB_RPTR_ADDR_HI, 1113 upper_32_bits(adev->wb.gpu_addr + wb_offset) & 0xFFFFFFFF); 1114 WREG32_SDMA(i, mmSDMA0_GFX_RB_RPTR_ADDR_LO, 1115 lower_32_bits(adev->wb.gpu_addr + wb_offset) & 0xFFFFFFFC); 1116 1117 rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, 1118 RPTR_WRITEBACK_ENABLE, 1); 1119 1120 WREG32_SDMA(i, mmSDMA0_GFX_RB_BASE, ring->gpu_addr >> 8); 1121 WREG32_SDMA(i, mmSDMA0_GFX_RB_BASE_HI, ring->gpu_addr >> 40); 1122 1123 ring->wptr = 0; 1124 1125 /* before programing wptr to a less value, need set minor_ptr_update first */ 1126 WREG32_SDMA(i, mmSDMA0_GFX_MINOR_PTR_UPDATE, 1); 1127 1128 doorbell = RREG32_SDMA(i, mmSDMA0_GFX_DOORBELL); 1129 doorbell_offset = RREG32_SDMA(i, mmSDMA0_GFX_DOORBELL_OFFSET); 1130 1131 doorbell = REG_SET_FIELD(doorbell, SDMA0_GFX_DOORBELL, ENABLE, 1132 ring->use_doorbell); 1133 doorbell_offset = REG_SET_FIELD(doorbell_offset, 1134 SDMA0_GFX_DOORBELL_OFFSET, 1135 OFFSET, ring->doorbell_index); 1136 WREG32_SDMA(i, mmSDMA0_GFX_DOORBELL, doorbell); 1137 WREG32_SDMA(i, mmSDMA0_GFX_DOORBELL_OFFSET, doorbell_offset); 1138 1139 sdma_v4_0_ring_set_wptr(ring); 1140 1141 /* set minor_ptr_update to 0 after wptr programed */ 1142 WREG32_SDMA(i, mmSDMA0_GFX_MINOR_PTR_UPDATE, 0); 1143 1144 /* setup the wptr shadow polling */ 1145 wptr_gpu_addr = adev->wb.gpu_addr + (ring->wptr_offs * 4); 1146 WREG32_SDMA(i, mmSDMA0_GFX_RB_WPTR_POLL_ADDR_LO, 1147 lower_32_bits(wptr_gpu_addr)); 1148 WREG32_SDMA(i, mmSDMA0_GFX_RB_WPTR_POLL_ADDR_HI, 1149 upper_32_bits(wptr_gpu_addr)); 1150 wptr_poll_cntl = RREG32_SDMA(i, mmSDMA0_GFX_RB_WPTR_POLL_CNTL); 1151 wptr_poll_cntl = REG_SET_FIELD(wptr_poll_cntl, 1152 SDMA0_GFX_RB_WPTR_POLL_CNTL, 1153 F32_POLL_ENABLE, amdgpu_sriov_vf(adev)? 1 : 0); 1154 WREG32_SDMA(i, mmSDMA0_GFX_RB_WPTR_POLL_CNTL, wptr_poll_cntl); 1155 1156 /* enable DMA RB */ 1157 rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RB_ENABLE, 1); 1158 WREG32_SDMA(i, mmSDMA0_GFX_RB_CNTL, rb_cntl); 1159 1160 ib_cntl = RREG32_SDMA(i, mmSDMA0_GFX_IB_CNTL); 1161 ib_cntl = REG_SET_FIELD(ib_cntl, SDMA0_GFX_IB_CNTL, IB_ENABLE, 1); 1162#ifdef __BIG_ENDIAN 1163 ib_cntl = REG_SET_FIELD(ib_cntl, SDMA0_GFX_IB_CNTL, IB_SWAP_ENABLE, 1); 1164#endif 1165 /* enable DMA IBs */ 1166 WREG32_SDMA(i, mmSDMA0_GFX_IB_CNTL, ib_cntl); 1167 1168 ring->sched.ready = true; 1169} 1170 1171/** 1172 * sdma_v4_0_page_resume - setup and start the async dma engines 1173 * 1174 * @adev: amdgpu_device pointer 1175 * @i: instance to resume 1176 * 1177 * Set up the page DMA ring buffers and enable them (VEGA10). 1178 * Returns 0 for success, error for failure. 1179 */ 1180static void sdma_v4_0_page_resume(struct amdgpu_device *adev, unsigned int i) 1181{ 1182 struct amdgpu_ring *ring = &adev->sdma.instance[i].page; 1183 u32 rb_cntl, ib_cntl, wptr_poll_cntl; 1184 u32 wb_offset; 1185 u32 doorbell; 1186 u32 doorbell_offset; 1187 u64 wptr_gpu_addr; 1188 1189 wb_offset = (ring->rptr_offs * 4); 1190 1191 rb_cntl = RREG32_SDMA(i, mmSDMA0_PAGE_RB_CNTL); 1192 rb_cntl = sdma_v4_0_rb_cntl(ring, rb_cntl); 1193 WREG32_SDMA(i, mmSDMA0_PAGE_RB_CNTL, rb_cntl); 1194 1195 /* Initialize the ring buffer's read and write pointers */ 1196 WREG32_SDMA(i, mmSDMA0_PAGE_RB_RPTR, 0); 1197 WREG32_SDMA(i, mmSDMA0_PAGE_RB_RPTR_HI, 0); 1198 WREG32_SDMA(i, mmSDMA0_PAGE_RB_WPTR, 0); 1199 WREG32_SDMA(i, mmSDMA0_PAGE_RB_WPTR_HI, 0); 1200 1201 /* set the wb address whether it's enabled or not */ 1202 WREG32_SDMA(i, mmSDMA0_PAGE_RB_RPTR_ADDR_HI, 1203 upper_32_bits(adev->wb.gpu_addr + wb_offset) & 0xFFFFFFFF); 1204 WREG32_SDMA(i, mmSDMA0_PAGE_RB_RPTR_ADDR_LO, 1205 lower_32_bits(adev->wb.gpu_addr + wb_offset) & 0xFFFFFFFC); 1206 1207 rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_PAGE_RB_CNTL, 1208 RPTR_WRITEBACK_ENABLE, 1); 1209 1210 WREG32_SDMA(i, mmSDMA0_PAGE_RB_BASE, ring->gpu_addr >> 8); 1211 WREG32_SDMA(i, mmSDMA0_PAGE_RB_BASE_HI, ring->gpu_addr >> 40); 1212 1213 ring->wptr = 0; 1214 1215 /* before programing wptr to a less value, need set minor_ptr_update first */ 1216 WREG32_SDMA(i, mmSDMA0_PAGE_MINOR_PTR_UPDATE, 1); 1217 1218 doorbell = RREG32_SDMA(i, mmSDMA0_PAGE_DOORBELL); 1219 doorbell_offset = RREG32_SDMA(i, mmSDMA0_PAGE_DOORBELL_OFFSET); 1220 1221 doorbell = REG_SET_FIELD(doorbell, SDMA0_PAGE_DOORBELL, ENABLE, 1222 ring->use_doorbell); 1223 doorbell_offset = REG_SET_FIELD(doorbell_offset, 1224 SDMA0_PAGE_DOORBELL_OFFSET, 1225 OFFSET, ring->doorbell_index); 1226 WREG32_SDMA(i, mmSDMA0_PAGE_DOORBELL, doorbell); 1227 WREG32_SDMA(i, mmSDMA0_PAGE_DOORBELL_OFFSET, doorbell_offset); 1228 1229 /* paging queue doorbell range is setup at sdma_v4_0_gfx_resume */ 1230 sdma_v4_0_page_ring_set_wptr(ring); 1231 1232 /* set minor_ptr_update to 0 after wptr programed */ 1233 WREG32_SDMA(i, mmSDMA0_PAGE_MINOR_PTR_UPDATE, 0); 1234 1235 /* setup the wptr shadow polling */ 1236 wptr_gpu_addr = adev->wb.gpu_addr + (ring->wptr_offs * 4); 1237 WREG32_SDMA(i, mmSDMA0_PAGE_RB_WPTR_POLL_ADDR_LO, 1238 lower_32_bits(wptr_gpu_addr)); 1239 WREG32_SDMA(i, mmSDMA0_PAGE_RB_WPTR_POLL_ADDR_HI, 1240 upper_32_bits(wptr_gpu_addr)); 1241 wptr_poll_cntl = RREG32_SDMA(i, mmSDMA0_PAGE_RB_WPTR_POLL_CNTL); 1242 wptr_poll_cntl = REG_SET_FIELD(wptr_poll_cntl, 1243 SDMA0_PAGE_RB_WPTR_POLL_CNTL, 1244 F32_POLL_ENABLE, amdgpu_sriov_vf(adev)? 1 : 0); 1245 WREG32_SDMA(i, mmSDMA0_PAGE_RB_WPTR_POLL_CNTL, wptr_poll_cntl); 1246 1247 /* enable DMA RB */ 1248 rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_PAGE_RB_CNTL, RB_ENABLE, 1); 1249 WREG32_SDMA(i, mmSDMA0_PAGE_RB_CNTL, rb_cntl); 1250 1251 ib_cntl = RREG32_SDMA(i, mmSDMA0_PAGE_IB_CNTL); 1252 ib_cntl = REG_SET_FIELD(ib_cntl, SDMA0_PAGE_IB_CNTL, IB_ENABLE, 1); 1253#ifdef __BIG_ENDIAN 1254 ib_cntl = REG_SET_FIELD(ib_cntl, SDMA0_PAGE_IB_CNTL, IB_SWAP_ENABLE, 1); 1255#endif 1256 /* enable DMA IBs */ 1257 WREG32_SDMA(i, mmSDMA0_PAGE_IB_CNTL, ib_cntl); 1258 1259 ring->sched.ready = true; 1260} 1261 1262static void 1263sdma_v4_1_update_power_gating(struct amdgpu_device *adev, bool enable) 1264{ 1265 uint32_t def, data; 1266 1267 if (enable && (adev->pg_flags & AMD_PG_SUPPORT_SDMA)) { 1268 /* enable idle interrupt */ 1269 def = data = RREG32(SOC15_REG_OFFSET(SDMA0, 0, mmSDMA0_CNTL)); 1270 data |= SDMA0_CNTL__CTXEMPTY_INT_ENABLE_MASK; 1271 1272 if (data != def) 1273 WREG32(SOC15_REG_OFFSET(SDMA0, 0, mmSDMA0_CNTL), data); 1274 } else { 1275 /* disable idle interrupt */ 1276 def = data = RREG32(SOC15_REG_OFFSET(SDMA0, 0, mmSDMA0_CNTL)); 1277 data &= ~SDMA0_CNTL__CTXEMPTY_INT_ENABLE_MASK; 1278 if (data != def) 1279 WREG32(SOC15_REG_OFFSET(SDMA0, 0, mmSDMA0_CNTL), data); 1280 } 1281} 1282 1283static void sdma_v4_1_init_power_gating(struct amdgpu_device *adev) 1284{ 1285 uint32_t def, data; 1286 1287 /* Enable HW based PG. */ 1288 def = data = RREG32(SOC15_REG_OFFSET(SDMA0, 0, mmSDMA0_POWER_CNTL)); 1289 data |= SDMA0_POWER_CNTL__PG_CNTL_ENABLE_MASK; 1290 if (data != def) 1291 WREG32(SOC15_REG_OFFSET(SDMA0, 0, mmSDMA0_POWER_CNTL), data); 1292 1293 /* enable interrupt */ 1294 def = data = RREG32(SOC15_REG_OFFSET(SDMA0, 0, mmSDMA0_CNTL)); 1295 data |= SDMA0_CNTL__CTXEMPTY_INT_ENABLE_MASK; 1296 if (data != def) 1297 WREG32(SOC15_REG_OFFSET(SDMA0, 0, mmSDMA0_CNTL), data); 1298 1299 /* Configure hold time to filter in-valid power on/off request. Use default right now */ 1300 def = data = RREG32(SOC15_REG_OFFSET(SDMA0, 0, mmSDMA0_POWER_CNTL)); 1301 data &= ~SDMA0_POWER_CNTL__ON_OFF_CONDITION_HOLD_TIME_MASK; 1302 data |= (mmSDMA0_POWER_CNTL_DEFAULT & SDMA0_POWER_CNTL__ON_OFF_CONDITION_HOLD_TIME_MASK); 1303 /* Configure switch time for hysteresis purpose. Use default right now */ 1304 data &= ~SDMA0_POWER_CNTL__ON_OFF_STATUS_DURATION_TIME_MASK; 1305 data |= (mmSDMA0_POWER_CNTL_DEFAULT & SDMA0_POWER_CNTL__ON_OFF_STATUS_DURATION_TIME_MASK); 1306 if(data != def) 1307 WREG32(SOC15_REG_OFFSET(SDMA0, 0, mmSDMA0_POWER_CNTL), data); 1308} 1309 1310static void sdma_v4_0_init_pg(struct amdgpu_device *adev) 1311{ 1312 if (!(adev->pg_flags & AMD_PG_SUPPORT_SDMA)) 1313 return; 1314 1315 switch (adev->asic_type) { 1316 case CHIP_RAVEN: 1317 case CHIP_RENOIR: 1318 sdma_v4_1_init_power_gating(adev); 1319 sdma_v4_1_update_power_gating(adev, true); 1320 break; 1321 default: 1322 break; 1323 } 1324} 1325 1326/** 1327 * sdma_v4_0_rlc_resume - setup and start the async dma engines 1328 * 1329 * @adev: amdgpu_device pointer 1330 * 1331 * Set up the compute DMA queues and enable them (VEGA10). 1332 * Returns 0 for success, error for failure. 1333 */ 1334static int sdma_v4_0_rlc_resume(struct amdgpu_device *adev) 1335{ 1336 sdma_v4_0_init_pg(adev); 1337 1338 return 0; 1339} 1340 1341/** 1342 * sdma_v4_0_load_microcode - load the sDMA ME ucode 1343 * 1344 * @adev: amdgpu_device pointer 1345 * 1346 * Loads the sDMA0/1 ucode. 1347 * Returns 0 for success, -EINVAL if the ucode is not available. 1348 */ 1349static int sdma_v4_0_load_microcode(struct amdgpu_device *adev) 1350{ 1351 const struct sdma_firmware_header_v1_0 *hdr; 1352 const __le32 *fw_data; 1353 u32 fw_size; 1354 int i, j; 1355 1356 /* halt the MEs */ 1357 sdma_v4_0_enable(adev, false); 1358 1359 for (i = 0; i < adev->sdma.num_instances; i++) { 1360 if (!adev->sdma.instance[i].fw) 1361 return -EINVAL; 1362 1363 hdr = (const struct sdma_firmware_header_v1_0 *)adev->sdma.instance[i].fw->data; 1364 amdgpu_ucode_print_sdma_hdr(&hdr->header); 1365 fw_size = le32_to_cpu(hdr->header.ucode_size_bytes) / 4; 1366 1367 fw_data = (const __le32 *) 1368 (adev->sdma.instance[i].fw->data + 1369 le32_to_cpu(hdr->header.ucode_array_offset_bytes)); 1370 1371 WREG32_SDMA(i, mmSDMA0_UCODE_ADDR, 0); 1372 1373 for (j = 0; j < fw_size; j++) 1374 WREG32_SDMA(i, mmSDMA0_UCODE_DATA, 1375 le32_to_cpup(fw_data++)); 1376 1377 WREG32_SDMA(i, mmSDMA0_UCODE_ADDR, 1378 adev->sdma.instance[i].fw_version); 1379 } 1380 1381 return 0; 1382} 1383 1384/** 1385 * sdma_v4_0_start - setup and start the async dma engines 1386 * 1387 * @adev: amdgpu_device pointer 1388 * 1389 * Set up the DMA engines and enable them (VEGA10). 1390 * Returns 0 for success, error for failure. 1391 */ 1392static int sdma_v4_0_start(struct amdgpu_device *adev) 1393{ 1394 struct amdgpu_ring *ring; 1395 int i, r = 0; 1396 1397 if (amdgpu_sriov_vf(adev)) { 1398 sdma_v4_0_ctx_switch_enable(adev, false); 1399 sdma_v4_0_enable(adev, false); 1400 } else { 1401 1402 if (adev->firmware.load_type != AMDGPU_FW_LOAD_PSP) { 1403 r = sdma_v4_0_load_microcode(adev); 1404 if (r) 1405 return r; 1406 } 1407 1408 /* unhalt the MEs */ 1409 sdma_v4_0_enable(adev, true); 1410 /* enable sdma ring preemption */ 1411 sdma_v4_0_ctx_switch_enable(adev, true); 1412 } 1413 1414 /* start the gfx rings and rlc compute queues */ 1415 for (i = 0; i < adev->sdma.num_instances; i++) { 1416 uint32_t temp; 1417 1418 WREG32_SDMA(i, mmSDMA0_SEM_WAIT_FAIL_TIMER_CNTL, 0); 1419 sdma_v4_0_gfx_resume(adev, i); 1420 if (adev->sdma.has_page_queue) 1421 sdma_v4_0_page_resume(adev, i); 1422 1423 /* set utc l1 enable flag always to 1 */ 1424 temp = RREG32_SDMA(i, mmSDMA0_CNTL); 1425 temp = REG_SET_FIELD(temp, SDMA0_CNTL, UTC_L1_ENABLE, 1); 1426 WREG32_SDMA(i, mmSDMA0_CNTL, temp); 1427 1428 if (!amdgpu_sriov_vf(adev)) { 1429 /* unhalt engine */ 1430 temp = RREG32_SDMA(i, mmSDMA0_F32_CNTL); 1431 temp = REG_SET_FIELD(temp, SDMA0_F32_CNTL, HALT, 0); 1432 WREG32_SDMA(i, mmSDMA0_F32_CNTL, temp); 1433 } 1434 } 1435 1436 if (amdgpu_sriov_vf(adev)) { 1437 sdma_v4_0_ctx_switch_enable(adev, true); 1438 sdma_v4_0_enable(adev, true); 1439 } else { 1440 r = sdma_v4_0_rlc_resume(adev); 1441 if (r) 1442 return r; 1443 } 1444 1445 for (i = 0; i < adev->sdma.num_instances; i++) { 1446 ring = &adev->sdma.instance[i].ring; 1447 1448 r = amdgpu_ring_test_helper(ring); 1449 if (r) 1450 return r; 1451 1452 if (adev->sdma.has_page_queue) { 1453 struct amdgpu_ring *page = &adev->sdma.instance[i].page; 1454 1455 r = amdgpu_ring_test_helper(page); 1456 if (r) 1457 return r; 1458 1459 if (adev->mman.buffer_funcs_ring == page) 1460 amdgpu_ttm_set_buffer_funcs_status(adev, true); 1461 } 1462 1463 if (adev->mman.buffer_funcs_ring == ring) 1464 amdgpu_ttm_set_buffer_funcs_status(adev, true); 1465 } 1466 1467 return r; 1468} 1469 1470/** 1471 * sdma_v4_0_ring_test_ring - simple async dma engine test 1472 * 1473 * @ring: amdgpu_ring structure holding ring information 1474 * 1475 * Test the DMA engine by writing using it to write an 1476 * value to memory. (VEGA10). 1477 * Returns 0 for success, error for failure. 1478 */ 1479static int sdma_v4_0_ring_test_ring(struct amdgpu_ring *ring) 1480{ 1481 struct amdgpu_device *adev = ring->adev; 1482 unsigned i; 1483 unsigned index; 1484 int r; 1485 u32 tmp; 1486 u64 gpu_addr; 1487 1488 r = amdgpu_device_wb_get(adev, &index); 1489 if (r) 1490 return r; 1491 1492 gpu_addr = adev->wb.gpu_addr + (index * 4); 1493 tmp = 0xCAFEDEAD; 1494 adev->wb.wb[index] = cpu_to_le32(tmp); 1495 1496 r = amdgpu_ring_alloc(ring, 5); 1497 if (r) 1498 goto error_free_wb; 1499 1500 amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_WRITE) | 1501 SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_WRITE_LINEAR)); 1502 amdgpu_ring_write(ring, lower_32_bits(gpu_addr)); 1503 amdgpu_ring_write(ring, upper_32_bits(gpu_addr)); 1504 amdgpu_ring_write(ring, SDMA_PKT_WRITE_UNTILED_DW_3_COUNT(0)); 1505 amdgpu_ring_write(ring, 0xDEADBEEF); 1506 amdgpu_ring_commit(ring); 1507 1508 for (i = 0; i < adev->usec_timeout; i++) { 1509 tmp = le32_to_cpu(adev->wb.wb[index]); 1510 if (tmp == 0xDEADBEEF) 1511 break; 1512 udelay(1); 1513 } 1514 1515 if (i >= adev->usec_timeout) 1516 r = -ETIMEDOUT; 1517 1518error_free_wb: 1519 amdgpu_device_wb_free(adev, index); 1520 return r; 1521} 1522 1523/** 1524 * sdma_v4_0_ring_test_ib - test an IB on the DMA engine 1525 * 1526 * @ring: amdgpu_ring structure holding ring information 1527 * 1528 * Test a simple IB in the DMA ring (VEGA10). 1529 * Returns 0 on success, error on failure. 1530 */ 1531static int sdma_v4_0_ring_test_ib(struct amdgpu_ring *ring, long timeout) 1532{ 1533 struct amdgpu_device *adev = ring->adev; 1534 struct amdgpu_ib ib; 1535 struct dma_fence *f = NULL; 1536 unsigned index; 1537 long r; 1538 u32 tmp = 0; 1539 u64 gpu_addr; 1540 1541 r = amdgpu_device_wb_get(adev, &index); 1542 if (r) 1543 return r; 1544 1545 gpu_addr = adev->wb.gpu_addr + (index * 4); 1546 tmp = 0xCAFEDEAD; 1547 adev->wb.wb[index] = cpu_to_le32(tmp); 1548 memset(&ib, 0, sizeof(ib)); 1549 r = amdgpu_ib_get(adev, NULL, 256, &ib); 1550 if (r) 1551 goto err0; 1552 1553 ib.ptr[0] = SDMA_PKT_HEADER_OP(SDMA_OP_WRITE) | 1554 SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_WRITE_LINEAR); 1555 ib.ptr[1] = lower_32_bits(gpu_addr); 1556 ib.ptr[2] = upper_32_bits(gpu_addr); 1557 ib.ptr[3] = SDMA_PKT_WRITE_UNTILED_DW_3_COUNT(0); 1558 ib.ptr[4] = 0xDEADBEEF; 1559 ib.ptr[5] = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP); 1560 ib.ptr[6] = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP); 1561 ib.ptr[7] = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP); 1562 ib.length_dw = 8; 1563 1564 r = amdgpu_ib_schedule(ring, 1, &ib, NULL, &f); 1565 if (r) 1566 goto err1; 1567 1568 r = dma_fence_wait_timeout(f, false, timeout); 1569 if (r == 0) { 1570 r = -ETIMEDOUT; 1571 goto err1; 1572 } else if (r < 0) { 1573 goto err1; 1574 } 1575 tmp = le32_to_cpu(adev->wb.wb[index]); 1576 if (tmp == 0xDEADBEEF) 1577 r = 0; 1578 else 1579 r = -EINVAL; 1580 1581err1: 1582 amdgpu_ib_free(adev, &ib, NULL); 1583 dma_fence_put(f); 1584err0: 1585 amdgpu_device_wb_free(adev, index); 1586 return r; 1587} 1588 1589 1590/** 1591 * sdma_v4_0_vm_copy_pte - update PTEs by copying them from the GART 1592 * 1593 * @ib: indirect buffer to fill with commands 1594 * @pe: addr of the page entry 1595 * @src: src addr to copy from 1596 * @count: number of page entries to update 1597 * 1598 * Update PTEs by copying them from the GART using sDMA (VEGA10). 1599 */ 1600static void sdma_v4_0_vm_copy_pte(struct amdgpu_ib *ib, 1601 uint64_t pe, uint64_t src, 1602 unsigned count) 1603{ 1604 unsigned bytes = count * 8; 1605 1606 ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_COPY) | 1607 SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_COPY_LINEAR); 1608 ib->ptr[ib->length_dw++] = bytes - 1; 1609 ib->ptr[ib->length_dw++] = 0; /* src/dst endian swap */ 1610 ib->ptr[ib->length_dw++] = lower_32_bits(src); 1611 ib->ptr[ib->length_dw++] = upper_32_bits(src); 1612 ib->ptr[ib->length_dw++] = lower_32_bits(pe); 1613 ib->ptr[ib->length_dw++] = upper_32_bits(pe); 1614 1615} 1616 1617/** 1618 * sdma_v4_0_vm_write_pte - update PTEs by writing them manually 1619 * 1620 * @ib: indirect buffer to fill with commands 1621 * @pe: addr of the page entry 1622 * @addr: dst addr to write into pe 1623 * @count: number of page entries to update 1624 * @incr: increase next addr by incr bytes 1625 * @flags: access flags 1626 * 1627 * Update PTEs by writing them manually using sDMA (VEGA10). 1628 */ 1629static void sdma_v4_0_vm_write_pte(struct amdgpu_ib *ib, uint64_t pe, 1630 uint64_t value, unsigned count, 1631 uint32_t incr) 1632{ 1633 unsigned ndw = count * 2; 1634 1635 ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_WRITE) | 1636 SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_WRITE_LINEAR); 1637 ib->ptr[ib->length_dw++] = lower_32_bits(pe); 1638 ib->ptr[ib->length_dw++] = upper_32_bits(pe); 1639 ib->ptr[ib->length_dw++] = ndw - 1; 1640 for (; ndw > 0; ndw -= 2) { 1641 ib->ptr[ib->length_dw++] = lower_32_bits(value); 1642 ib->ptr[ib->length_dw++] = upper_32_bits(value); 1643 value += incr; 1644 } 1645} 1646 1647/** 1648 * sdma_v4_0_vm_set_pte_pde - update the page tables using sDMA 1649 * 1650 * @ib: indirect buffer to fill with commands 1651 * @pe: addr of the page entry 1652 * @addr: dst addr to write into pe 1653 * @count: number of page entries to update 1654 * @incr: increase next addr by incr bytes 1655 * @flags: access flags 1656 * 1657 * Update the page tables using sDMA (VEGA10). 1658 */ 1659static void sdma_v4_0_vm_set_pte_pde(struct amdgpu_ib *ib, 1660 uint64_t pe, 1661 uint64_t addr, unsigned count, 1662 uint32_t incr, uint64_t flags) 1663{ 1664 /* for physically contiguous pages (vram) */ 1665 ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_PTEPDE); 1666 ib->ptr[ib->length_dw++] = lower_32_bits(pe); /* dst addr */ 1667 ib->ptr[ib->length_dw++] = upper_32_bits(pe); 1668 ib->ptr[ib->length_dw++] = lower_32_bits(flags); /* mask */ 1669 ib->ptr[ib->length_dw++] = upper_32_bits(flags); 1670 ib->ptr[ib->length_dw++] = lower_32_bits(addr); /* value */ 1671 ib->ptr[ib->length_dw++] = upper_32_bits(addr); 1672 ib->ptr[ib->length_dw++] = incr; /* increment size */ 1673 ib->ptr[ib->length_dw++] = 0; 1674 ib->ptr[ib->length_dw++] = count - 1; /* number of entries */ 1675} 1676 1677/** 1678 * sdma_v4_0_ring_pad_ib - pad the IB to the required number of dw 1679 * 1680 * @ib: indirect buffer to fill with padding 1681 * 1682 */ 1683static void sdma_v4_0_ring_pad_ib(struct amdgpu_ring *ring, struct amdgpu_ib *ib) 1684{ 1685 struct amdgpu_sdma_instance *sdma = amdgpu_sdma_get_instance_from_ring(ring); 1686 u32 pad_count; 1687 int i; 1688 1689 pad_count = (-ib->length_dw) & 7; 1690 for (i = 0; i < pad_count; i++) 1691 if (sdma && sdma->burst_nop && (i == 0)) 1692 ib->ptr[ib->length_dw++] = 1693 SDMA_PKT_HEADER_OP(SDMA_OP_NOP) | 1694 SDMA_PKT_NOP_HEADER_COUNT(pad_count - 1); 1695 else 1696 ib->ptr[ib->length_dw++] = 1697 SDMA_PKT_HEADER_OP(SDMA_OP_NOP); 1698} 1699 1700 1701/** 1702 * sdma_v4_0_ring_emit_pipeline_sync - sync the pipeline 1703 * 1704 * @ring: amdgpu_ring pointer 1705 * 1706 * Make sure all previous operations are completed (CIK). 1707 */ 1708static void sdma_v4_0_ring_emit_pipeline_sync(struct amdgpu_ring *ring) 1709{ 1710 uint32_t seq = ring->fence_drv.sync_seq; 1711 uint64_t addr = ring->fence_drv.gpu_addr; 1712 1713 /* wait for idle */ 1714 sdma_v4_0_wait_reg_mem(ring, 1, 0, 1715 addr & 0xfffffffc, 1716 upper_32_bits(addr) & 0xffffffff, 1717 seq, 0xffffffff, 4); 1718} 1719 1720 1721/** 1722 * sdma_v4_0_ring_emit_vm_flush - vm flush using sDMA 1723 * 1724 * @ring: amdgpu_ring pointer 1725 * @vm: amdgpu_vm pointer 1726 * 1727 * Update the page table base and flush the VM TLB 1728 * using sDMA (VEGA10). 1729 */ 1730static void sdma_v4_0_ring_emit_vm_flush(struct amdgpu_ring *ring, 1731 unsigned vmid, uint64_t pd_addr) 1732{ 1733 amdgpu_gmc_emit_flush_gpu_tlb(ring, vmid, pd_addr); 1734} 1735 1736static void sdma_v4_0_ring_emit_wreg(struct amdgpu_ring *ring, 1737 uint32_t reg, uint32_t val) 1738{ 1739 amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_SRBM_WRITE) | 1740 SDMA_PKT_SRBM_WRITE_HEADER_BYTE_EN(0xf)); 1741 amdgpu_ring_write(ring, reg); 1742 amdgpu_ring_write(ring, val); 1743} 1744 1745static void sdma_v4_0_ring_emit_reg_wait(struct amdgpu_ring *ring, uint32_t reg, 1746 uint32_t val, uint32_t mask) 1747{ 1748 sdma_v4_0_wait_reg_mem(ring, 0, 0, reg, 0, val, mask, 10); 1749} 1750 1751static bool sdma_v4_0_fw_support_paging_queue(struct amdgpu_device *adev) 1752{ 1753 uint fw_version = adev->sdma.instance[0].fw_version; 1754 1755 switch (adev->asic_type) { 1756 case CHIP_VEGA10: 1757 return fw_version >= 430; 1758 case CHIP_VEGA12: 1759 /*return fw_version >= 31;*/ 1760 return false; 1761 case CHIP_VEGA20: 1762 return fw_version >= 123; 1763 default: 1764 return false; 1765 } 1766} 1767 1768static int sdma_v4_0_early_init(void *handle) 1769{ 1770 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1771 int r; 1772 1773 if (adev->asic_type == CHIP_RAVEN || adev->asic_type == CHIP_RENOIR) 1774 adev->sdma.num_instances = 1; 1775 else if (adev->asic_type == CHIP_ARCTURUS) 1776 adev->sdma.num_instances = 8; 1777 else 1778 adev->sdma.num_instances = 2; 1779 1780 r = sdma_v4_0_init_microcode(adev); 1781 if (r) { 1782 DRM_ERROR("Failed to load sdma firmware!\n"); 1783 return r; 1784 } 1785 1786 /* TODO: Page queue breaks driver reload under SRIOV */ 1787 if ((adev->asic_type == CHIP_VEGA10) && amdgpu_sriov_vf((adev))) 1788 adev->sdma.has_page_queue = false; 1789 else if (sdma_v4_0_fw_support_paging_queue(adev)) 1790 adev->sdma.has_page_queue = true; 1791 1792 sdma_v4_0_set_ring_funcs(adev); 1793 sdma_v4_0_set_buffer_funcs(adev); 1794 sdma_v4_0_set_vm_pte_funcs(adev); 1795 sdma_v4_0_set_irq_funcs(adev); 1796 sdma_v4_0_set_ras_funcs(adev); 1797 1798 return 0; 1799} 1800 1801static int sdma_v4_0_process_ras_data_cb(struct amdgpu_device *adev, 1802 void *err_data, 1803 struct amdgpu_iv_entry *entry); 1804 1805static int sdma_v4_0_late_init(void *handle) 1806{ 1807 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1808 struct ras_ih_if ih_info = { 1809 .cb = sdma_v4_0_process_ras_data_cb, 1810 }; 1811 int i; 1812 1813 /* read back edc counter registers to clear the counters */ 1814 if (amdgpu_ras_is_supported(adev, AMDGPU_RAS_BLOCK__SDMA)) { 1815 for (i = 0; i < adev->sdma.num_instances; i++) 1816 RREG32_SDMA(i, mmSDMA0_EDC_COUNTER); 1817 } 1818 1819 if (adev->sdma.funcs && adev->sdma.funcs->ras_late_init) 1820 return adev->sdma.funcs->ras_late_init(adev, &ih_info); 1821 else 1822 return 0; 1823} 1824 1825static int sdma_v4_0_sw_init(void *handle) 1826{ 1827 struct amdgpu_ring *ring; 1828 int r, i; 1829 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1830 1831 /* SDMA trap event */ 1832 for (i = 0; i < adev->sdma.num_instances; i++) { 1833 r = amdgpu_irq_add_id(adev, sdma_v4_0_seq_to_irq_id(i), 1834 SDMA0_4_0__SRCID__SDMA_TRAP, 1835 &adev->sdma.trap_irq); 1836 if (r) 1837 return r; 1838 } 1839 1840 /* SDMA SRAM ECC event */ 1841 for (i = 0; i < adev->sdma.num_instances; i++) { 1842 r = amdgpu_irq_add_id(adev, sdma_v4_0_seq_to_irq_id(i), 1843 SDMA0_4_0__SRCID__SDMA_SRAM_ECC, 1844 &adev->sdma.ecc_irq); 1845 if (r) 1846 return r; 1847 } 1848 1849 for (i = 0; i < adev->sdma.num_instances; i++) { 1850 ring = &adev->sdma.instance[i].ring; 1851 ring->ring_obj = NULL; 1852 ring->use_doorbell = true; 1853 1854 DRM_INFO("use_doorbell being set to: [%s]\n", 1855 ring->use_doorbell?"true":"false"); 1856 1857 /* doorbell size is 2 dwords, get DWORD offset */ 1858 ring->doorbell_index = adev->doorbell_index.sdma_engine[i] << 1; 1859 1860 snprintf(ring->name, sizeof(ring->name), "sdma%d", i); 1861 r = amdgpu_ring_init(adev, ring, 1024, &adev->sdma.trap_irq, 1862 AMDGPU_SDMA_IRQ_INSTANCE0 + i); 1863 if (r) 1864 return r; 1865 1866 if (adev->sdma.has_page_queue) { 1867 ring = &adev->sdma.instance[i].page; 1868 ring->ring_obj = NULL; 1869 ring->use_doorbell = true; 1870 1871 /* paging queue use same doorbell index/routing as gfx queue 1872 * with 0x400 (4096 dwords) offset on second doorbell page 1873 */ 1874 ring->doorbell_index = adev->doorbell_index.sdma_engine[i] << 1; 1875 ring->doorbell_index += 0x400; 1876 1877 snprintf(ring->name, sizeof(ring->name), "page%d", i); 1878 r = amdgpu_ring_init(adev, ring, 1024, 1879 &adev->sdma.trap_irq, 1880 AMDGPU_SDMA_IRQ_INSTANCE0 + i); 1881 if (r) 1882 return r; 1883 } 1884 } 1885 1886 return r; 1887} 1888 1889static int sdma_v4_0_sw_fini(void *handle) 1890{ 1891 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1892 int i; 1893 1894 if (adev->sdma.funcs && adev->sdma.funcs->ras_fini) 1895 adev->sdma.funcs->ras_fini(adev); 1896 1897 for (i = 0; i < adev->sdma.num_instances; i++) { 1898 amdgpu_ring_fini(&adev->sdma.instance[i].ring); 1899 if (adev->sdma.has_page_queue) 1900 amdgpu_ring_fini(&adev->sdma.instance[i].page); 1901 } 1902 1903 sdma_v4_0_destroy_inst_ctx(adev); 1904 1905 return 0; 1906} 1907 1908static int sdma_v4_0_hw_init(void *handle) 1909{ 1910 int r; 1911 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1912 1913 if ((adev->asic_type == CHIP_RAVEN && adev->powerplay.pp_funcs && 1914 adev->powerplay.pp_funcs->set_powergating_by_smu) || 1915 (adev->asic_type == CHIP_RENOIR && !adev->in_gpu_reset)) 1916 amdgpu_dpm_set_powergating_by_smu(adev, AMD_IP_BLOCK_TYPE_SDMA, false); 1917 1918 if (!amdgpu_sriov_vf(adev)) 1919 sdma_v4_0_init_golden_registers(adev); 1920 1921 r = sdma_v4_0_start(adev); 1922 1923 return r; 1924} 1925 1926static int sdma_v4_0_hw_fini(void *handle) 1927{ 1928 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1929 int i; 1930 1931 if (amdgpu_sriov_vf(adev)) 1932 return 0; 1933 1934 for (i = 0; i < adev->sdma.num_instances; i++) { 1935 amdgpu_irq_put(adev, &adev->sdma.ecc_irq, 1936 AMDGPU_SDMA_IRQ_INSTANCE0 + i); 1937 } 1938 1939 sdma_v4_0_ctx_switch_enable(adev, false); 1940 sdma_v4_0_enable(adev, false); 1941 1942 if ((adev->asic_type == CHIP_RAVEN && adev->powerplay.pp_funcs 1943 && adev->powerplay.pp_funcs->set_powergating_by_smu) || 1944 adev->asic_type == CHIP_RENOIR) 1945 amdgpu_dpm_set_powergating_by_smu(adev, AMD_IP_BLOCK_TYPE_SDMA, true); 1946 1947 return 0; 1948} 1949 1950static int sdma_v4_0_suspend(void *handle) 1951{ 1952 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1953 1954 return sdma_v4_0_hw_fini(adev); 1955} 1956 1957static int sdma_v4_0_resume(void *handle) 1958{ 1959 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1960 1961 return sdma_v4_0_hw_init(adev); 1962} 1963 1964static bool sdma_v4_0_is_idle(void *handle) 1965{ 1966 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1967 u32 i; 1968 1969 for (i = 0; i < adev->sdma.num_instances; i++) { 1970 u32 tmp = RREG32_SDMA(i, mmSDMA0_STATUS_REG); 1971 1972 if (!(tmp & SDMA0_STATUS_REG__IDLE_MASK)) 1973 return false; 1974 } 1975 1976 return true; 1977} 1978 1979static int sdma_v4_0_wait_for_idle(void *handle) 1980{ 1981 unsigned i, j; 1982 u32 sdma[AMDGPU_MAX_SDMA_INSTANCES]; 1983 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1984 1985 for (i = 0; i < adev->usec_timeout; i++) { 1986 for (j = 0; j < adev->sdma.num_instances; j++) { 1987 sdma[j] = RREG32_SDMA(j, mmSDMA0_STATUS_REG); 1988 if (!(sdma[j] & SDMA0_STATUS_REG__IDLE_MASK)) 1989 break; 1990 } 1991 if (j == adev->sdma.num_instances) 1992 return 0; 1993 udelay(1); 1994 } 1995 return -ETIMEDOUT; 1996} 1997 1998static int sdma_v4_0_soft_reset(void *handle) 1999{ 2000 /* todo */ 2001 2002 return 0; 2003} 2004 2005static int sdma_v4_0_set_trap_irq_state(struct amdgpu_device *adev, 2006 struct amdgpu_irq_src *source, 2007 unsigned type, 2008 enum amdgpu_interrupt_state state) 2009{ 2010 u32 sdma_cntl; 2011 2012 sdma_cntl = RREG32_SDMA(type, mmSDMA0_CNTL); 2013 sdma_cntl = REG_SET_FIELD(sdma_cntl, SDMA0_CNTL, TRAP_ENABLE, 2014 state == AMDGPU_IRQ_STATE_ENABLE ? 1 : 0); 2015 WREG32_SDMA(type, mmSDMA0_CNTL, sdma_cntl); 2016 2017 return 0; 2018} 2019 2020static int sdma_v4_0_process_trap_irq(struct amdgpu_device *adev, 2021 struct amdgpu_irq_src *source, 2022 struct amdgpu_iv_entry *entry) 2023{ 2024 uint32_t instance; 2025 2026 DRM_DEBUG("IH: SDMA trap\n"); 2027 instance = sdma_v4_0_irq_id_to_seq(entry->client_id); 2028 switch (entry->ring_id) { 2029 case 0: 2030 amdgpu_fence_process(&adev->sdma.instance[instance].ring); 2031 break; 2032 case 1: 2033 if (adev->asic_type == CHIP_VEGA20) 2034 amdgpu_fence_process(&adev->sdma.instance[instance].page); 2035 break; 2036 case 2: 2037 /* XXX compute */ 2038 break; 2039 case 3: 2040 if (adev->asic_type != CHIP_VEGA20) 2041 amdgpu_fence_process(&adev->sdma.instance[instance].page); 2042 break; 2043 } 2044 return 0; 2045} 2046 2047static int sdma_v4_0_process_ras_data_cb(struct amdgpu_device *adev, 2048 void *err_data, 2049 struct amdgpu_iv_entry *entry) 2050{ 2051 int instance; 2052 2053 /* When ���Full RAS��� is enabled, the per-IP interrupt sources should 2054 * be disabled and the driver should only look for the aggregated 2055 * interrupt via sync flood 2056 */ 2057 if (amdgpu_ras_is_supported(adev, AMDGPU_RAS_BLOCK__GFX)) 2058 goto out; 2059 2060 instance = sdma_v4_0_irq_id_to_seq(entry->client_id); 2061 if (instance < 0) 2062 goto out; 2063 2064 amdgpu_sdma_process_ras_data_cb(adev, err_data, entry); 2065 2066out: 2067 return AMDGPU_RAS_SUCCESS; 2068} 2069 2070static int sdma_v4_0_process_illegal_inst_irq(struct amdgpu_device *adev, 2071 struct amdgpu_irq_src *source, 2072 struct amdgpu_iv_entry *entry) 2073{ 2074 int instance; 2075 2076 DRM_ERROR("Illegal instruction in SDMA command stream\n"); 2077 2078 instance = sdma_v4_0_irq_id_to_seq(entry->client_id); 2079 if (instance < 0) 2080 return 0; 2081 2082 switch (entry->ring_id) { 2083 case 0: 2084 drm_sched_fault(&adev->sdma.instance[instance].ring.sched); 2085 break; 2086 } 2087 return 0; 2088} 2089 2090static int sdma_v4_0_set_ecc_irq_state(struct amdgpu_device *adev, 2091 struct amdgpu_irq_src *source, 2092 unsigned type, 2093 enum amdgpu_interrupt_state state) 2094{ 2095 u32 sdma_edc_config; 2096 2097 sdma_edc_config = RREG32_SDMA(type, mmSDMA0_EDC_CONFIG); 2098 sdma_edc_config = REG_SET_FIELD(sdma_edc_config, SDMA0_EDC_CONFIG, ECC_INT_ENABLE, 2099 state == AMDGPU_IRQ_STATE_ENABLE ? 1 : 0); 2100 WREG32_SDMA(type, mmSDMA0_EDC_CONFIG, sdma_edc_config); 2101 2102 return 0; 2103} 2104 2105static void sdma_v4_0_update_medium_grain_clock_gating( 2106 struct amdgpu_device *adev, 2107 bool enable) 2108{ 2109 uint32_t data, def; 2110 int i; 2111 2112 if (enable && (adev->cg_flags & AMD_CG_SUPPORT_SDMA_MGCG)) { 2113 for (i = 0; i < adev->sdma.num_instances; i++) { 2114 def = data = RREG32_SDMA(i, mmSDMA0_CLK_CTRL); 2115 data &= ~(SDMA0_CLK_CTRL__SOFT_OVERRIDE7_MASK | 2116 SDMA0_CLK_CTRL__SOFT_OVERRIDE6_MASK | 2117 SDMA0_CLK_CTRL__SOFT_OVERRIDE5_MASK | 2118 SDMA0_CLK_CTRL__SOFT_OVERRIDE4_MASK | 2119 SDMA0_CLK_CTRL__SOFT_OVERRIDE3_MASK | 2120 SDMA0_CLK_CTRL__SOFT_OVERRIDE2_MASK | 2121 SDMA0_CLK_CTRL__SOFT_OVERRIDE1_MASK | 2122 SDMA0_CLK_CTRL__SOFT_OVERRIDE0_MASK); 2123 if (def != data) 2124 WREG32_SDMA(i, mmSDMA0_CLK_CTRL, data); 2125 } 2126 } else { 2127 for (i = 0; i < adev->sdma.num_instances; i++) { 2128 def = data = RREG32_SDMA(i, mmSDMA0_CLK_CTRL); 2129 data |= (SDMA0_CLK_CTRL__SOFT_OVERRIDE7_MASK | 2130 SDMA0_CLK_CTRL__SOFT_OVERRIDE6_MASK | 2131 SDMA0_CLK_CTRL__SOFT_OVERRIDE5_MASK | 2132 SDMA0_CLK_CTRL__SOFT_OVERRIDE4_MASK | 2133 SDMA0_CLK_CTRL__SOFT_OVERRIDE3_MASK | 2134 SDMA0_CLK_CTRL__SOFT_OVERRIDE2_MASK | 2135 SDMA0_CLK_CTRL__SOFT_OVERRIDE1_MASK | 2136 SDMA0_CLK_CTRL__SOFT_OVERRIDE0_MASK); 2137 if (def != data) 2138 WREG32_SDMA(i, mmSDMA0_CLK_CTRL, data); 2139 } 2140 } 2141} 2142 2143 2144static void sdma_v4_0_update_medium_grain_light_sleep( 2145 struct amdgpu_device *adev, 2146 bool enable) 2147{ 2148 uint32_t data, def; 2149 int i; 2150 2151 if (enable && (adev->cg_flags & AMD_CG_SUPPORT_SDMA_LS)) { 2152 for (i = 0; i < adev->sdma.num_instances; i++) { 2153 /* 1-not override: enable sdma mem light sleep */ 2154 def = data = RREG32_SDMA(0, mmSDMA0_POWER_CNTL); 2155 data |= SDMA0_POWER_CNTL__MEM_POWER_OVERRIDE_MASK; 2156 if (def != data) 2157 WREG32_SDMA(0, mmSDMA0_POWER_CNTL, data); 2158 } 2159 } else { 2160 for (i = 0; i < adev->sdma.num_instances; i++) { 2161 /* 0-override:disable sdma mem light sleep */ 2162 def = data = RREG32_SDMA(0, mmSDMA0_POWER_CNTL); 2163 data &= ~SDMA0_POWER_CNTL__MEM_POWER_OVERRIDE_MASK; 2164 if (def != data) 2165 WREG32_SDMA(0, mmSDMA0_POWER_CNTL, data); 2166 } 2167 } 2168} 2169 2170static int sdma_v4_0_set_clockgating_state(void *handle, 2171 enum amd_clockgating_state state) 2172{ 2173 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 2174 2175 if (amdgpu_sriov_vf(adev)) 2176 return 0; 2177 2178 switch (adev->asic_type) { 2179 case CHIP_VEGA10: 2180 case CHIP_VEGA12: 2181 case CHIP_VEGA20: 2182 case CHIP_RAVEN: 2183 case CHIP_ARCTURUS: 2184 case CHIP_RENOIR: 2185 sdma_v4_0_update_medium_grain_clock_gating(adev, 2186 state == AMD_CG_STATE_GATE); 2187 sdma_v4_0_update_medium_grain_light_sleep(adev, 2188 state == AMD_CG_STATE_GATE); 2189 break; 2190 default: 2191 break; 2192 } 2193 return 0; 2194} 2195 2196static int sdma_v4_0_set_powergating_state(void *handle, 2197 enum amd_powergating_state state) 2198{ 2199 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 2200 2201 switch (adev->asic_type) { 2202 case CHIP_RAVEN: 2203 sdma_v4_1_update_power_gating(adev, 2204 state == AMD_PG_STATE_GATE ? true : false); 2205 break; 2206 default: 2207 break; 2208 } 2209 2210 return 0; 2211} 2212 2213static void sdma_v4_0_get_clockgating_state(void *handle, u32 *flags) 2214{ 2215 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 2216 int data; 2217 2218 if (amdgpu_sriov_vf(adev)) 2219 *flags = 0; 2220 2221 /* AMD_CG_SUPPORT_SDMA_MGCG */ 2222 data = RREG32(SOC15_REG_OFFSET(SDMA0, 0, mmSDMA0_CLK_CTRL)); 2223 if (!(data & SDMA0_CLK_CTRL__SOFT_OVERRIDE7_MASK)) 2224 *flags |= AMD_CG_SUPPORT_SDMA_MGCG; 2225 2226 /* AMD_CG_SUPPORT_SDMA_LS */ 2227 data = RREG32(SOC15_REG_OFFSET(SDMA0, 0, mmSDMA0_POWER_CNTL)); 2228 if (data & SDMA0_POWER_CNTL__MEM_POWER_OVERRIDE_MASK) 2229 *flags |= AMD_CG_SUPPORT_SDMA_LS; 2230} 2231 2232const struct amd_ip_funcs sdma_v4_0_ip_funcs = { 2233 .name = "sdma_v4_0", 2234 .early_init = sdma_v4_0_early_init, 2235 .late_init = sdma_v4_0_late_init, 2236 .sw_init = sdma_v4_0_sw_init, 2237 .sw_fini = sdma_v4_0_sw_fini, 2238 .hw_init = sdma_v4_0_hw_init, 2239 .hw_fini = sdma_v4_0_hw_fini, 2240 .suspend = sdma_v4_0_suspend, 2241 .resume = sdma_v4_0_resume, 2242 .is_idle = sdma_v4_0_is_idle, 2243 .wait_for_idle = sdma_v4_0_wait_for_idle, 2244 .soft_reset = sdma_v4_0_soft_reset, 2245 .set_clockgating_state = sdma_v4_0_set_clockgating_state, 2246 .set_powergating_state = sdma_v4_0_set_powergating_state, 2247 .get_clockgating_state = sdma_v4_0_get_clockgating_state, 2248}; 2249 2250static const struct amdgpu_ring_funcs sdma_v4_0_ring_funcs = { 2251 .type = AMDGPU_RING_TYPE_SDMA, 2252 .align_mask = 0xf, 2253 .nop = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP), 2254 .support_64bit_ptrs = true, 2255 .vmhub = AMDGPU_MMHUB_0, 2256 .get_rptr = sdma_v4_0_ring_get_rptr, 2257 .get_wptr = sdma_v4_0_ring_get_wptr, 2258 .set_wptr = sdma_v4_0_ring_set_wptr, 2259 .emit_frame_size = 2260 6 + /* sdma_v4_0_ring_emit_hdp_flush */ 2261 3 + /* hdp invalidate */ 2262 6 + /* sdma_v4_0_ring_emit_pipeline_sync */ 2263 /* sdma_v4_0_ring_emit_vm_flush */ 2264 SOC15_FLUSH_GPU_TLB_NUM_WREG * 3 + 2265 SOC15_FLUSH_GPU_TLB_NUM_REG_WAIT * 6 + 2266 10 + 10 + 10, /* sdma_v4_0_ring_emit_fence x3 for user fence, vm fence */ 2267 .emit_ib_size = 7 + 6, /* sdma_v4_0_ring_emit_ib */ 2268 .emit_ib = sdma_v4_0_ring_emit_ib, 2269 .emit_fence = sdma_v4_0_ring_emit_fence, 2270 .emit_pipeline_sync = sdma_v4_0_ring_emit_pipeline_sync, 2271 .emit_vm_flush = sdma_v4_0_ring_emit_vm_flush, 2272 .emit_hdp_flush = sdma_v4_0_ring_emit_hdp_flush, 2273 .test_ring = sdma_v4_0_ring_test_ring, 2274 .test_ib = sdma_v4_0_ring_test_ib, 2275 .insert_nop = sdma_v4_0_ring_insert_nop, 2276 .pad_ib = sdma_v4_0_ring_pad_ib, 2277 .emit_wreg = sdma_v4_0_ring_emit_wreg, 2278 .emit_reg_wait = sdma_v4_0_ring_emit_reg_wait, 2279 .emit_reg_write_reg_wait = amdgpu_ring_emit_reg_write_reg_wait_helper, 2280}; 2281 2282/* 2283 * On Arcturus, SDMA instance 5~7 has a different vmhub type(AMDGPU_MMHUB_1). 2284 * So create a individual constant ring_funcs for those instances. 2285 */ 2286static const struct amdgpu_ring_funcs sdma_v4_0_ring_funcs_2nd_mmhub = { 2287 .type = AMDGPU_RING_TYPE_SDMA, 2288 .align_mask = 0xf, 2289 .nop = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP), 2290 .support_64bit_ptrs = true, 2291 .vmhub = AMDGPU_MMHUB_1, 2292 .get_rptr = sdma_v4_0_ring_get_rptr, 2293 .get_wptr = sdma_v4_0_ring_get_wptr, 2294 .set_wptr = sdma_v4_0_ring_set_wptr, 2295 .emit_frame_size = 2296 6 + /* sdma_v4_0_ring_emit_hdp_flush */ 2297 3 + /* hdp invalidate */ 2298 6 + /* sdma_v4_0_ring_emit_pipeline_sync */ 2299 /* sdma_v4_0_ring_emit_vm_flush */ 2300 SOC15_FLUSH_GPU_TLB_NUM_WREG * 3 + 2301 SOC15_FLUSH_GPU_TLB_NUM_REG_WAIT * 6 + 2302 10 + 10 + 10, /* sdma_v4_0_ring_emit_fence x3 for user fence, vm fence */ 2303 .emit_ib_size = 7 + 6, /* sdma_v4_0_ring_emit_ib */ 2304 .emit_ib = sdma_v4_0_ring_emit_ib, 2305 .emit_fence = sdma_v4_0_ring_emit_fence, 2306 .emit_pipeline_sync = sdma_v4_0_ring_emit_pipeline_sync, 2307 .emit_vm_flush = sdma_v4_0_ring_emit_vm_flush, 2308 .emit_hdp_flush = sdma_v4_0_ring_emit_hdp_flush, 2309 .test_ring = sdma_v4_0_ring_test_ring, 2310 .test_ib = sdma_v4_0_ring_test_ib, 2311 .insert_nop = sdma_v4_0_ring_insert_nop, 2312 .pad_ib = sdma_v4_0_ring_pad_ib, 2313 .emit_wreg = sdma_v4_0_ring_emit_wreg, 2314 .emit_reg_wait = sdma_v4_0_ring_emit_reg_wait, 2315 .emit_reg_write_reg_wait = amdgpu_ring_emit_reg_write_reg_wait_helper, 2316}; 2317 2318static const struct amdgpu_ring_funcs sdma_v4_0_page_ring_funcs = { 2319 .type = AMDGPU_RING_TYPE_SDMA, 2320 .align_mask = 0xf, 2321 .nop = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP), 2322 .support_64bit_ptrs = true, 2323 .vmhub = AMDGPU_MMHUB_0, 2324 .get_rptr = sdma_v4_0_ring_get_rptr, 2325 .get_wptr = sdma_v4_0_page_ring_get_wptr, 2326 .set_wptr = sdma_v4_0_page_ring_set_wptr, 2327 .emit_frame_size = 2328 6 + /* sdma_v4_0_ring_emit_hdp_flush */ 2329 3 + /* hdp invalidate */ 2330 6 + /* sdma_v4_0_ring_emit_pipeline_sync */ 2331 /* sdma_v4_0_ring_emit_vm_flush */ 2332 SOC15_FLUSH_GPU_TLB_NUM_WREG * 3 + 2333 SOC15_FLUSH_GPU_TLB_NUM_REG_WAIT * 6 + 2334 10 + 10 + 10, /* sdma_v4_0_ring_emit_fence x3 for user fence, vm fence */ 2335 .emit_ib_size = 7 + 6, /* sdma_v4_0_ring_emit_ib */ 2336 .emit_ib = sdma_v4_0_ring_emit_ib, 2337 .emit_fence = sdma_v4_0_ring_emit_fence, 2338 .emit_pipeline_sync = sdma_v4_0_ring_emit_pipeline_sync, 2339 .emit_vm_flush = sdma_v4_0_ring_emit_vm_flush, 2340 .emit_hdp_flush = sdma_v4_0_ring_emit_hdp_flush, 2341 .test_ring = sdma_v4_0_ring_test_ring, 2342 .test_ib = sdma_v4_0_ring_test_ib, 2343 .insert_nop = sdma_v4_0_ring_insert_nop, 2344 .pad_ib = sdma_v4_0_ring_pad_ib, 2345 .emit_wreg = sdma_v4_0_ring_emit_wreg, 2346 .emit_reg_wait = sdma_v4_0_ring_emit_reg_wait, 2347 .emit_reg_write_reg_wait = amdgpu_ring_emit_reg_write_reg_wait_helper, 2348}; 2349 2350static const struct amdgpu_ring_funcs sdma_v4_0_page_ring_funcs_2nd_mmhub = { 2351 .type = AMDGPU_RING_TYPE_SDMA, 2352 .align_mask = 0xf, 2353 .nop = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP), 2354 .support_64bit_ptrs = true, 2355 .vmhub = AMDGPU_MMHUB_1, 2356 .get_rptr = sdma_v4_0_ring_get_rptr, 2357 .get_wptr = sdma_v4_0_page_ring_get_wptr, 2358 .set_wptr = sdma_v4_0_page_ring_set_wptr, 2359 .emit_frame_size = 2360 6 + /* sdma_v4_0_ring_emit_hdp_flush */ 2361 3 + /* hdp invalidate */ 2362 6 + /* sdma_v4_0_ring_emit_pipeline_sync */ 2363 /* sdma_v4_0_ring_emit_vm_flush */ 2364 SOC15_FLUSH_GPU_TLB_NUM_WREG * 3 + 2365 SOC15_FLUSH_GPU_TLB_NUM_REG_WAIT * 6 + 2366 10 + 10 + 10, /* sdma_v4_0_ring_emit_fence x3 for user fence, vm fence */ 2367 .emit_ib_size = 7 + 6, /* sdma_v4_0_ring_emit_ib */ 2368 .emit_ib = sdma_v4_0_ring_emit_ib, 2369 .emit_fence = sdma_v4_0_ring_emit_fence, 2370 .emit_pipeline_sync = sdma_v4_0_ring_emit_pipeline_sync, 2371 .emit_vm_flush = sdma_v4_0_ring_emit_vm_flush, 2372 .emit_hdp_flush = sdma_v4_0_ring_emit_hdp_flush, 2373 .test_ring = sdma_v4_0_ring_test_ring, 2374 .test_ib = sdma_v4_0_ring_test_ib, 2375 .insert_nop = sdma_v4_0_ring_insert_nop, 2376 .pad_ib = sdma_v4_0_ring_pad_ib, 2377 .emit_wreg = sdma_v4_0_ring_emit_wreg, 2378 .emit_reg_wait = sdma_v4_0_ring_emit_reg_wait, 2379 .emit_reg_write_reg_wait = amdgpu_ring_emit_reg_write_reg_wait_helper, 2380}; 2381 2382static void sdma_v4_0_set_ring_funcs(struct amdgpu_device *adev) 2383{ 2384 int i; 2385 2386 for (i = 0; i < adev->sdma.num_instances; i++) { 2387 if (adev->asic_type == CHIP_ARCTURUS && i >= 5) 2388 adev->sdma.instance[i].ring.funcs = 2389 &sdma_v4_0_ring_funcs_2nd_mmhub; 2390 else 2391 adev->sdma.instance[i].ring.funcs = 2392 &sdma_v4_0_ring_funcs; 2393 adev->sdma.instance[i].ring.me = i; 2394 if (adev->sdma.has_page_queue) { 2395 if (adev->asic_type == CHIP_ARCTURUS && i >= 5) 2396 adev->sdma.instance[i].page.funcs = 2397 &sdma_v4_0_page_ring_funcs_2nd_mmhub; 2398 else 2399 adev->sdma.instance[i].page.funcs = 2400 &sdma_v4_0_page_ring_funcs; 2401 adev->sdma.instance[i].page.me = i; 2402 } 2403 } 2404} 2405 2406static const struct amdgpu_irq_src_funcs sdma_v4_0_trap_irq_funcs = { 2407 .set = sdma_v4_0_set_trap_irq_state, 2408 .process = sdma_v4_0_process_trap_irq, 2409}; 2410 2411static const struct amdgpu_irq_src_funcs sdma_v4_0_illegal_inst_irq_funcs = { 2412 .process = sdma_v4_0_process_illegal_inst_irq, 2413}; 2414 2415static const struct amdgpu_irq_src_funcs sdma_v4_0_ecc_irq_funcs = { 2416 .set = sdma_v4_0_set_ecc_irq_state, 2417 .process = amdgpu_sdma_process_ecc_irq, 2418}; 2419 2420 2421 2422static void sdma_v4_0_set_irq_funcs(struct amdgpu_device *adev) 2423{ 2424 switch (adev->sdma.num_instances) { 2425 case 1: 2426 adev->sdma.trap_irq.num_types = AMDGPU_SDMA_IRQ_INSTANCE1; 2427 adev->sdma.ecc_irq.num_types = AMDGPU_SDMA_IRQ_INSTANCE1; 2428 break; 2429 case 8: 2430 adev->sdma.trap_irq.num_types = AMDGPU_SDMA_IRQ_LAST; 2431 adev->sdma.ecc_irq.num_types = AMDGPU_SDMA_IRQ_LAST; 2432 break; 2433 case 2: 2434 default: 2435 adev->sdma.trap_irq.num_types = AMDGPU_SDMA_IRQ_INSTANCE2; 2436 adev->sdma.ecc_irq.num_types = AMDGPU_SDMA_IRQ_INSTANCE2; 2437 break; 2438 } 2439 adev->sdma.trap_irq.funcs = &sdma_v4_0_trap_irq_funcs; 2440 adev->sdma.illegal_inst_irq.funcs = &sdma_v4_0_illegal_inst_irq_funcs; 2441 adev->sdma.ecc_irq.funcs = &sdma_v4_0_ecc_irq_funcs; 2442} 2443 2444/** 2445 * sdma_v4_0_emit_copy_buffer - copy buffer using the sDMA engine 2446 * 2447 * @ring: amdgpu_ring structure holding ring information 2448 * @src_offset: src GPU address 2449 * @dst_offset: dst GPU address 2450 * @byte_count: number of bytes to xfer 2451 * 2452 * Copy GPU buffers using the DMA engine (VEGA10/12). 2453 * Used by the amdgpu ttm implementation to move pages if 2454 * registered as the asic copy callback. 2455 */ 2456static void sdma_v4_0_emit_copy_buffer(struct amdgpu_ib *ib, 2457 uint64_t src_offset, 2458 uint64_t dst_offset, 2459 uint32_t byte_count) 2460{ 2461 ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_COPY) | 2462 SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_COPY_LINEAR); 2463 ib->ptr[ib->length_dw++] = byte_count - 1; 2464 ib->ptr[ib->length_dw++] = 0; /* src/dst endian swap */ 2465 ib->ptr[ib->length_dw++] = lower_32_bits(src_offset); 2466 ib->ptr[ib->length_dw++] = upper_32_bits(src_offset); 2467 ib->ptr[ib->length_dw++] = lower_32_bits(dst_offset); 2468 ib->ptr[ib->length_dw++] = upper_32_bits(dst_offset); 2469} 2470 2471/** 2472 * sdma_v4_0_emit_fill_buffer - fill buffer using the sDMA engine 2473 * 2474 * @ring: amdgpu_ring structure holding ring information 2475 * @src_data: value to write to buffer 2476 * @dst_offset: dst GPU address 2477 * @byte_count: number of bytes to xfer 2478 * 2479 * Fill GPU buffers using the DMA engine (VEGA10/12). 2480 */ 2481static void sdma_v4_0_emit_fill_buffer(struct amdgpu_ib *ib, 2482 uint32_t src_data, 2483 uint64_t dst_offset, 2484 uint32_t byte_count) 2485{ 2486 ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_CONST_FILL); 2487 ib->ptr[ib->length_dw++] = lower_32_bits(dst_offset); 2488 ib->ptr[ib->length_dw++] = upper_32_bits(dst_offset); 2489 ib->ptr[ib->length_dw++] = src_data; 2490 ib->ptr[ib->length_dw++] = byte_count - 1; 2491} 2492 2493static const struct amdgpu_buffer_funcs sdma_v4_0_buffer_funcs = { 2494 .copy_max_bytes = 0x400000, 2495 .copy_num_dw = 7, 2496 .emit_copy_buffer = sdma_v4_0_emit_copy_buffer, 2497 2498 .fill_max_bytes = 0x400000, 2499 .fill_num_dw = 5, 2500 .emit_fill_buffer = sdma_v4_0_emit_fill_buffer, 2501}; 2502 2503static void sdma_v4_0_set_buffer_funcs(struct amdgpu_device *adev) 2504{ 2505 adev->mman.buffer_funcs = &sdma_v4_0_buffer_funcs; 2506 if (adev->sdma.has_page_queue) 2507 adev->mman.buffer_funcs_ring = &adev->sdma.instance[0].page; 2508 else 2509 adev->mman.buffer_funcs_ring = &adev->sdma.instance[0].ring; 2510} 2511 2512static const struct amdgpu_vm_pte_funcs sdma_v4_0_vm_pte_funcs = { 2513 .copy_pte_num_dw = 7, 2514 .copy_pte = sdma_v4_0_vm_copy_pte, 2515 2516 .write_pte = sdma_v4_0_vm_write_pte, 2517 .set_pte_pde = sdma_v4_0_vm_set_pte_pde, 2518}; 2519 2520static void sdma_v4_0_set_vm_pte_funcs(struct amdgpu_device *adev) 2521{ 2522 struct drm_gpu_scheduler *sched; 2523 unsigned i; 2524 2525 adev->vm_manager.vm_pte_funcs = &sdma_v4_0_vm_pte_funcs; 2526 for (i = 0; i < adev->sdma.num_instances; i++) { 2527 if (adev->sdma.has_page_queue) 2528 sched = &adev->sdma.instance[i].page.sched; 2529 else 2530 sched = &adev->sdma.instance[i].ring.sched; 2531 adev->vm_manager.vm_pte_scheds[i] = sched; 2532 } 2533 adev->vm_manager.vm_pte_num_scheds = adev->sdma.num_instances; 2534} 2535 2536static void sdma_v4_0_get_ras_error_count(uint32_t value, 2537 uint32_t instance, 2538 uint32_t *sec_count) 2539{ 2540 uint32_t i; 2541 uint32_t sec_cnt; 2542 2543 /* double bits error (multiple bits) error detection is not supported */ 2544 for (i = 0; i < ARRAY_SIZE(sdma_v4_0_ras_fields); i++) { 2545 /* the SDMA_EDC_COUNTER register in each sdma instance 2546 * shares the same sed shift_mask 2547 * */ 2548 sec_cnt = (value & 2549 sdma_v4_0_ras_fields[i].sec_count_mask) >> 2550 sdma_v4_0_ras_fields[i].sec_count_shift; 2551 if (sec_cnt) { 2552 DRM_INFO("Detected %s in SDMA%d, SED %d\n", 2553 sdma_v4_0_ras_fields[i].name, 2554 instance, sec_cnt); 2555 *sec_count += sec_cnt; 2556 } 2557 } 2558} 2559 2560static int sdma_v4_0_query_ras_error_count(struct amdgpu_device *adev, 2561 uint32_t instance, void *ras_error_status) 2562{ 2563 struct ras_err_data *err_data = (struct ras_err_data *)ras_error_status; 2564 uint32_t sec_count = 0; 2565 uint32_t reg_value = 0; 2566 2567 reg_value = RREG32_SDMA(instance, mmSDMA0_EDC_COUNTER); 2568 /* double bit error is not supported */ 2569 if (reg_value) 2570 sdma_v4_0_get_ras_error_count(reg_value, 2571 instance, &sec_count); 2572 /* err_data->ce_count should be initialized to 0 2573 * before calling into this function */ 2574 err_data->ce_count += sec_count; 2575 /* double bit error is not supported 2576 * set ue count to 0 */ 2577 err_data->ue_count = 0; 2578 2579 return 0; 2580}; 2581 2582static const struct amdgpu_sdma_ras_funcs sdma_v4_0_ras_funcs = { 2583 .ras_late_init = amdgpu_sdma_ras_late_init, 2584 .ras_fini = amdgpu_sdma_ras_fini, 2585 .query_ras_error_count = sdma_v4_0_query_ras_error_count, 2586}; 2587 2588static void sdma_v4_0_set_ras_funcs(struct amdgpu_device *adev) 2589{ 2590 switch (adev->asic_type) { 2591 case CHIP_VEGA20: 2592 case CHIP_ARCTURUS: 2593 adev->sdma.funcs = &sdma_v4_0_ras_funcs; 2594 break; 2595 default: 2596 break; 2597 } 2598} 2599 2600const struct amdgpu_ip_block_version sdma_v4_0_ip_block = { 2601 .type = AMD_IP_BLOCK_TYPE_SDMA, 2602 .major = 4, 2603 .minor = 0, 2604 .rev = 0, 2605 .funcs = &sdma_v4_0_ip_funcs, 2606}; 2607