1 /* 2 * Copyright 2023 Advanced Micro Devices, Inc. 3 * 4 * Permission is hereby granted, free of charge, to any person obtaining a 5 * copy of this software and associated documentation files (the "Software"), 6 * to deal in the Software without restriction, including without limitation 7 * the rights to use, copy, modify, merge, publish, distribute, sublicense, 8 * and/or sell copies of the Software, and to permit persons to whom the 9 * Software is furnished to do so, subject to the following conditions: 10 * 11 * The above copyright notice and this permission notice shall be included in 12 * all copies or substantial portions of the Software. 13 * 14 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR 15 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, 16 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL 17 * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR 18 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, 19 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR 20 * OTHER DEALINGS IN THE SOFTWARE. 21 * 22 */ 23 #include <linux/delay.h> 24 #include <linux/kernel.h> 25 #include <linux/firmware.h> 26 #include <linux/module.h> 27 #include <linux/pci.h> 28 #include "amdgpu.h" 29 #include "amdgpu_gfx.h" 30 #include "amdgpu_psp.h" 31 #include "amdgpu_smu.h" 32 #include "imu_v12_0.h" 33 #include "soc24.h" 34 #include "nvd.h" 35 36 #include "gc/gc_12_0_0_offset.h" 37 #include "gc/gc_12_0_0_sh_mask.h" 38 #include "soc24_enum.h" 39 #include "ivsrcid/gfx/irqsrcs_gfx_12_0_0.h" 40 41 #include "soc15.h" 42 #include "clearstate_gfx12.h" 43 #include "v12_structs.h" 44 #include "gfx_v12_0.h" 45 #include "nbif_v6_3_1.h" 46 #include "mes_v12_0.h" 47 #include "mes_userqueue.h" 48 #include "amdgpu_userq_fence.h" 49 50 #define GFX12_NUM_GFX_RINGS 1 51 #define GFX12_MEC_HPD_SIZE 2048 52 53 #define RLCG_UCODE_LOADING_START_ADDRESS 0x00002000L 54 55 #define regCP_GFX_MQD_CONTROL_DEFAULT 0x00000100 56 #define regCP_GFX_HQD_VMID_DEFAULT 0x00000000 57 #define regCP_GFX_HQD_QUEUE_PRIORITY_DEFAULT 0x00000000 58 #define regCP_GFX_HQD_QUANTUM_DEFAULT 0x00000a01 59 #define regCP_GFX_HQD_CNTL_DEFAULT 0x00f00000 60 #define regCP_RB_DOORBELL_CONTROL_DEFAULT 0x00000000 61 #define regCP_GFX_HQD_RPTR_DEFAULT 0x00000000 62 63 #define regCP_HQD_EOP_CONTROL_DEFAULT 0x00000006 64 #define regCP_HQD_PQ_DOORBELL_CONTROL_DEFAULT 0x00000000 65 #define regCP_MQD_CONTROL_DEFAULT 0x00000100 66 #define regCP_HQD_PQ_CONTROL_DEFAULT 0x00308509 67 #define regCP_HQD_PQ_DOORBELL_CONTROL_DEFAULT 0x00000000 68 #define regCP_HQD_PQ_RPTR_DEFAULT 0x00000000 69 #define regCP_HQD_PERSISTENT_STATE_DEFAULT 0x0be05501 70 #define regCP_HQD_IB_CONTROL_DEFAULT 0x00300000 71 72 73 MODULE_FIRMWARE("amdgpu/gc_12_0_0_pfp.bin"); 74 MODULE_FIRMWARE("amdgpu/gc_12_0_0_me.bin"); 75 MODULE_FIRMWARE("amdgpu/gc_12_0_0_mec.bin"); 76 MODULE_FIRMWARE("amdgpu/gc_12_0_0_rlc.bin"); 77 MODULE_FIRMWARE("amdgpu/gc_12_0_0_toc.bin"); 78 MODULE_FIRMWARE("amdgpu/gc_12_0_1_pfp.bin"); 79 MODULE_FIRMWARE("amdgpu/gc_12_0_1_me.bin"); 80 MODULE_FIRMWARE("amdgpu/gc_12_0_1_mec.bin"); 81 MODULE_FIRMWARE("amdgpu/gc_12_0_1_rlc.bin"); 82 MODULE_FIRMWARE("amdgpu/gc_12_0_1_rlc_kicker.bin"); 83 MODULE_FIRMWARE("amdgpu/gc_12_0_1_toc.bin"); 84 85 static const struct amdgpu_hwip_reg_entry gc_reg_list_12_0[] = { 86 SOC15_REG_ENTRY_STR(GC, 0, regGRBM_STATUS), 87 SOC15_REG_ENTRY_STR(GC, 0, regGRBM_STATUS2), 88 SOC15_REG_ENTRY_STR(GC, 0, regGRBM_STATUS3), 89 SOC15_REG_ENTRY_STR(GC, 0, regCP_STALLED_STAT1), 90 SOC15_REG_ENTRY_STR(GC, 0, regCP_STALLED_STAT2), 91 SOC15_REG_ENTRY_STR(GC, 0, regCP_STALLED_STAT3), 92 SOC15_REG_ENTRY_STR(GC, 0, regCP_CPC_STALLED_STAT1), 93 SOC15_REG_ENTRY_STR(GC, 0, regCP_CPF_STALLED_STAT1), 94 SOC15_REG_ENTRY_STR(GC, 0, regCP_BUSY_STAT), 95 SOC15_REG_ENTRY_STR(GC, 0, regCP_CPC_BUSY_STAT), 96 SOC15_REG_ENTRY_STR(GC, 0, regCP_CPF_BUSY_STAT), 97 SOC15_REG_ENTRY_STR(GC, 0, regCP_CPC_BUSY_STAT2), 98 SOC15_REG_ENTRY_STR(GC, 0, regCP_CPF_BUSY_STAT2), 99 SOC15_REG_ENTRY_STR(GC, 0, regCP_CPF_STATUS), 100 SOC15_REG_ENTRY_STR(GC, 0, regCP_GFX_ERROR), 101 SOC15_REG_ENTRY_STR(GC, 0, regCP_GFX_HPD_STATUS0), 102 SOC15_REG_ENTRY_STR(GC, 0, regCP_RB_BASE), 103 SOC15_REG_ENTRY_STR(GC, 0, regCP_RB_RPTR), 104 SOC15_REG_ENTRY_STR(GC, 0, regCP_RB_WPTR), 105 SOC15_REG_ENTRY_STR(GC, 0, regCP_RB0_BASE), 106 SOC15_REG_ENTRY_STR(GC, 0, regCP_RB0_RPTR), 107 SOC15_REG_ENTRY_STR(GC, 0, regCP_RB0_WPTR), 108 SOC15_REG_ENTRY_STR(GC, 0, regCP_IB1_CMD_BUFSZ), 109 SOC15_REG_ENTRY_STR(GC, 0, regCP_IB2_CMD_BUFSZ), 110 SOC15_REG_ENTRY_STR(GC, 0, regCP_IB1_BASE_LO), 111 SOC15_REG_ENTRY_STR(GC, 0, regCP_IB1_BASE_HI), 112 SOC15_REG_ENTRY_STR(GC, 0, regCP_IB1_BUFSZ), 113 SOC15_REG_ENTRY_STR(GC, 0, regCP_IB2_BASE_LO), 114 SOC15_REG_ENTRY_STR(GC, 0, regCP_IB2_BASE_HI), 115 SOC15_REG_ENTRY_STR(GC, 0, regCP_IB2_BUFSZ), 116 SOC15_REG_ENTRY_STR(GC, 0, regCPF_UTCL1_STATUS), 117 SOC15_REG_ENTRY_STR(GC, 0, regCPC_UTCL1_STATUS), 118 SOC15_REG_ENTRY_STR(GC, 0, regCPG_UTCL1_STATUS), 119 SOC15_REG_ENTRY_STR(GC, 0, regIA_UTCL1_STATUS), 120 SOC15_REG_ENTRY_STR(GC, 0, regIA_UTCL1_STATUS_2), 121 SOC15_REG_ENTRY_STR(GC, 0, regPA_CL_CNTL_STATUS), 122 SOC15_REG_ENTRY_STR(GC, 0, regRMI_UTCL1_STATUS), 123 SOC15_REG_ENTRY_STR(GC, 0, regSQC_CACHES), 124 SOC15_REG_ENTRY_STR(GC, 0, regSQG_STATUS), 125 SOC15_REG_ENTRY_STR(GC, 0, regWD_UTCL1_STATUS), 126 SOC15_REG_ENTRY_STR(GC, 0, regGCVM_L2_PROTECTION_FAULT_CNTL), 127 SOC15_REG_ENTRY_STR(GC, 0, regGCVM_L2_PROTECTION_FAULT_STATUS_LO32), 128 SOC15_REG_ENTRY_STR(GC, 0, regGCVM_L2_PROTECTION_FAULT_STATUS_HI32), 129 SOC15_REG_ENTRY_STR(GC, 0, regCP_DEBUG), 130 SOC15_REG_ENTRY_STR(GC, 0, regCP_MEC_CNTL), 131 SOC15_REG_ENTRY_STR(GC, 0, regCP_MES_CNTL), 132 SOC15_REG_ENTRY_STR(GC, 0, regCP_MES_INSTR_PNTR), 133 SOC15_REG_ENTRY_STR(GC, 0, regCP_ME_INSTR_PNTR), 134 SOC15_REG_ENTRY_STR(GC, 0, regCP_PFP_INSTR_PNTR), 135 SOC15_REG_ENTRY_STR(GC, 0, regCP_CPC_STATUS), 136 SOC15_REG_ENTRY_STR(GC, 0, regCP_GFX_RS64_INSTR_PNTR0), 137 SOC15_REG_ENTRY_STR(GC, 0, regCP_GFX_RS64_INSTR_PNTR1), 138 SOC15_REG_ENTRY_STR(GC, 0, regCP_MEC_RS64_INSTR_PNTR), 139 /* cp header registers */ 140 SOC15_REG_ENTRY_STR(GC, 0, regCP_MES_HEADER_DUMP), 141 SOC15_REG_ENTRY_STR(GC, 0, regCP_MES_HEADER_DUMP), 142 SOC15_REG_ENTRY_STR(GC, 0, regCP_MES_HEADER_DUMP), 143 SOC15_REG_ENTRY_STR(GC, 0, regCP_MES_HEADER_DUMP), 144 SOC15_REG_ENTRY_STR(GC, 0, regCP_MES_HEADER_DUMP), 145 SOC15_REG_ENTRY_STR(GC, 0, regCP_MES_HEADER_DUMP), 146 SOC15_REG_ENTRY_STR(GC, 0, regCP_MES_HEADER_DUMP), 147 SOC15_REG_ENTRY_STR(GC, 0, regCP_MES_HEADER_DUMP), 148 /* SE status registers */ 149 SOC15_REG_ENTRY_STR(GC, 0, regGRBM_STATUS_SE0), 150 SOC15_REG_ENTRY_STR(GC, 0, regGRBM_STATUS_SE1), 151 SOC15_REG_ENTRY_STR(GC, 0, regGRBM_STATUS_SE2), 152 SOC15_REG_ENTRY_STR(GC, 0, regGRBM_STATUS_SE3) 153 }; 154 155 static const struct amdgpu_hwip_reg_entry gc_cp_reg_list_12[] = { 156 /* compute registers */ 157 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_VMID), 158 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_PERSISTENT_STATE), 159 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_PIPE_PRIORITY), 160 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_QUEUE_PRIORITY), 161 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_QUANTUM), 162 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_PQ_BASE), 163 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_PQ_BASE_HI), 164 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_PQ_RPTR), 165 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_PQ_WPTR_POLL_ADDR), 166 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_PQ_WPTR_POLL_ADDR_HI), 167 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_PQ_DOORBELL_CONTROL), 168 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_PQ_CONTROL), 169 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_IB_BASE_ADDR), 170 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_IB_BASE_ADDR_HI), 171 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_IB_RPTR), 172 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_IB_CONTROL), 173 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_DEQUEUE_REQUEST), 174 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_EOP_BASE_ADDR), 175 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_EOP_BASE_ADDR_HI), 176 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_EOP_CONTROL), 177 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_EOP_RPTR), 178 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_EOP_WPTR), 179 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_EOP_EVENTS), 180 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_CTX_SAVE_BASE_ADDR_LO), 181 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_CTX_SAVE_BASE_ADDR_HI), 182 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_CTX_SAVE_CONTROL), 183 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_CNTL_STACK_OFFSET), 184 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_CNTL_STACK_SIZE), 185 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_WG_STATE_OFFSET), 186 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_CTX_SAVE_SIZE), 187 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_GDS_RESOURCE_STATE), 188 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_ERROR), 189 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_EOP_WPTR_MEM), 190 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_PQ_WPTR_LO), 191 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_PQ_WPTR_HI), 192 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_SUSPEND_CNTL_STACK_OFFSET), 193 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_SUSPEND_CNTL_STACK_DW_CNT), 194 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_SUSPEND_WG_STATE_OFFSET), 195 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_DEQUEUE_STATUS), 196 /* cp header registers */ 197 SOC15_REG_ENTRY_STR(GC, 0, regCP_MEC_ME1_HEADER_DUMP), 198 SOC15_REG_ENTRY_STR(GC, 0, regCP_MEC_ME1_HEADER_DUMP), 199 SOC15_REG_ENTRY_STR(GC, 0, regCP_MEC_ME1_HEADER_DUMP), 200 SOC15_REG_ENTRY_STR(GC, 0, regCP_MEC_ME1_HEADER_DUMP), 201 SOC15_REG_ENTRY_STR(GC, 0, regCP_MEC_ME1_HEADER_DUMP), 202 SOC15_REG_ENTRY_STR(GC, 0, regCP_MEC_ME1_HEADER_DUMP), 203 SOC15_REG_ENTRY_STR(GC, 0, regCP_MEC_ME1_HEADER_DUMP), 204 SOC15_REG_ENTRY_STR(GC, 0, regCP_MEC_ME1_HEADER_DUMP), 205 }; 206 207 static const struct amdgpu_hwip_reg_entry gc_gfx_queue_reg_list_12[] = { 208 /* gfx queue registers */ 209 SOC15_REG_ENTRY_STR(GC, 0, regCP_GFX_HQD_ACTIVE), 210 SOC15_REG_ENTRY_STR(GC, 0, regCP_GFX_HQD_VMID), 211 SOC15_REG_ENTRY_STR(GC, 0, regCP_GFX_HQD_QUEUE_PRIORITY), 212 SOC15_REG_ENTRY_STR(GC, 0, regCP_GFX_HQD_QUANTUM), 213 SOC15_REG_ENTRY_STR(GC, 0, regCP_GFX_HQD_BASE), 214 SOC15_REG_ENTRY_STR(GC, 0, regCP_GFX_HQD_BASE_HI), 215 SOC15_REG_ENTRY_STR(GC, 0, regCP_GFX_HQD_OFFSET), 216 SOC15_REG_ENTRY_STR(GC, 0, regCP_GFX_HQD_CNTL), 217 SOC15_REG_ENTRY_STR(GC, 0, regCP_GFX_HQD_CSMD_RPTR), 218 SOC15_REG_ENTRY_STR(GC, 0, regCP_GFX_HQD_WPTR), 219 SOC15_REG_ENTRY_STR(GC, 0, regCP_GFX_HQD_WPTR_HI), 220 SOC15_REG_ENTRY_STR(GC, 0, regCP_GFX_HQD_DEQUEUE_REQUEST), 221 SOC15_REG_ENTRY_STR(GC, 0, regCP_GFX_HQD_MAPPED), 222 SOC15_REG_ENTRY_STR(GC, 0, regCP_GFX_HQD_QUE_MGR_CONTROL), 223 SOC15_REG_ENTRY_STR(GC, 0, regCP_GFX_HQD_HQ_CONTROL0), 224 SOC15_REG_ENTRY_STR(GC, 0, regCP_GFX_HQD_HQ_STATUS0), 225 SOC15_REG_ENTRY_STR(GC, 0, regCP_GFX_MQD_BASE_ADDR), 226 SOC15_REG_ENTRY_STR(GC, 0, regCP_GFX_MQD_BASE_ADDR_HI), 227 SOC15_REG_ENTRY_STR(GC, 0, regCP_RB_WPTR_POLL_ADDR_LO), 228 SOC15_REG_ENTRY_STR(GC, 0, regCP_RB_WPTR_POLL_ADDR_HI), 229 SOC15_REG_ENTRY_STR(GC, 0, regCP_RB_RPTR), 230 SOC15_REG_ENTRY_STR(GC, 0, regCP_IB1_BASE_LO), 231 SOC15_REG_ENTRY_STR(GC, 0, regCP_IB1_BASE_HI), 232 SOC15_REG_ENTRY_STR(GC, 0, regCP_IB1_CMD_BUFSZ), 233 SOC15_REG_ENTRY_STR(GC, 0, regCP_IB1_BUFSZ), 234 /* cp header registers */ 235 SOC15_REG_ENTRY_STR(GC, 0, regCP_PFP_HEADER_DUMP), 236 SOC15_REG_ENTRY_STR(GC, 0, regCP_PFP_HEADER_DUMP), 237 SOC15_REG_ENTRY_STR(GC, 0, regCP_PFP_HEADER_DUMP), 238 SOC15_REG_ENTRY_STR(GC, 0, regCP_PFP_HEADER_DUMP), 239 SOC15_REG_ENTRY_STR(GC, 0, regCP_PFP_HEADER_DUMP), 240 SOC15_REG_ENTRY_STR(GC, 0, regCP_PFP_HEADER_DUMP), 241 SOC15_REG_ENTRY_STR(GC, 0, regCP_PFP_HEADER_DUMP), 242 SOC15_REG_ENTRY_STR(GC, 0, regCP_PFP_HEADER_DUMP), 243 SOC15_REG_ENTRY_STR(GC, 0, regCP_ME_HEADER_DUMP), 244 SOC15_REG_ENTRY_STR(GC, 0, regCP_ME_HEADER_DUMP), 245 SOC15_REG_ENTRY_STR(GC, 0, regCP_ME_HEADER_DUMP), 246 SOC15_REG_ENTRY_STR(GC, 0, regCP_ME_HEADER_DUMP), 247 SOC15_REG_ENTRY_STR(GC, 0, regCP_ME_HEADER_DUMP), 248 SOC15_REG_ENTRY_STR(GC, 0, regCP_ME_HEADER_DUMP), 249 SOC15_REG_ENTRY_STR(GC, 0, regCP_ME_HEADER_DUMP), 250 SOC15_REG_ENTRY_STR(GC, 0, regCP_ME_HEADER_DUMP), 251 }; 252 253 static const struct soc15_reg_golden golden_settings_gc_12_0_rev0[] = { 254 SOC15_REG_GOLDEN_VALUE(GC, 0, regDB_MEM_CONFIG, 0x0000000f, 0x0000000f), 255 SOC15_REG_GOLDEN_VALUE(GC, 0, regCB_HW_CONTROL_1, 0x03000000, 0x03000000), 256 SOC15_REG_GOLDEN_VALUE(GC, 0, regGL2C_CTRL5, 0x00000070, 0x00000020) 257 }; 258 259 static const struct soc15_reg_golden golden_settings_gc_12_0[] = { 260 SOC15_REG_GOLDEN_VALUE(GC, 0, regDB_MEM_CONFIG, 0x00008000, 0x00008000), 261 }; 262 263 #define DEFAULT_SH_MEM_CONFIG \ 264 ((SH_MEM_ADDRESS_MODE_64 << SH_MEM_CONFIG__ADDRESS_MODE__SHIFT) | \ 265 (SH_MEM_ALIGNMENT_MODE_UNALIGNED << SH_MEM_CONFIG__ALIGNMENT_MODE__SHIFT) | \ 266 (3 << SH_MEM_CONFIG__INITIAL_INST_PREFETCH__SHIFT)) 267 268 static void gfx_v12_0_disable_gpa_mode(struct amdgpu_device *adev); 269 static void gfx_v12_0_set_ring_funcs(struct amdgpu_device *adev); 270 static void gfx_v12_0_set_irq_funcs(struct amdgpu_device *adev); 271 static void gfx_v12_0_set_rlc_funcs(struct amdgpu_device *adev); 272 static void gfx_v12_0_set_mqd_funcs(struct amdgpu_device *adev); 273 static void gfx_v12_0_set_imu_funcs(struct amdgpu_device *adev); 274 static int gfx_v12_0_get_cu_info(struct amdgpu_device *adev, 275 struct amdgpu_cu_info *cu_info); 276 static uint64_t gfx_v12_0_get_gpu_clock_counter(struct amdgpu_device *adev); 277 static void gfx_v12_0_select_se_sh(struct amdgpu_device *adev, u32 se_num, 278 u32 sh_num, u32 instance, int xcc_id); 279 static u32 gfx_v12_0_get_wgp_active_bitmap_per_sh(struct amdgpu_device *adev); 280 281 static void gfx_v12_0_ring_emit_wreg(struct amdgpu_ring *ring, uint32_t reg, 282 uint32_t val); 283 static int gfx_v12_0_wait_for_rlc_autoload_complete(struct amdgpu_device *adev); 284 static void gfx_v12_0_ring_invalidate_tlbs(struct amdgpu_ring *ring, 285 uint16_t pasid, uint32_t flush_type, 286 bool all_hub, uint8_t dst_sel); 287 static void gfx_v12_0_set_safe_mode(struct amdgpu_device *adev, int xcc_id); 288 static void gfx_v12_0_unset_safe_mode(struct amdgpu_device *adev, int xcc_id); 289 static void gfx_v12_0_update_perf_clk(struct amdgpu_device *adev, 290 bool enable); 291 292 static void gfx_v12_0_kiq_set_resources(struct amdgpu_ring *kiq_ring, 293 uint64_t queue_mask) 294 { 295 amdgpu_ring_write(kiq_ring, PACKET3(PACKET3_SET_RESOURCES, 6)); 296 amdgpu_ring_write(kiq_ring, PACKET3_SET_RESOURCES_VMID_MASK(0) | 297 PACKET3_SET_RESOURCES_QUEUE_TYPE(0)); /* vmid_mask:0 queue_type:0 (KIQ) */ 298 amdgpu_ring_write(kiq_ring, lower_32_bits(queue_mask)); /* queue mask lo */ 299 amdgpu_ring_write(kiq_ring, upper_32_bits(queue_mask)); /* queue mask hi */ 300 amdgpu_ring_write(kiq_ring, 0); /* gws mask lo */ 301 amdgpu_ring_write(kiq_ring, 0); /* gws mask hi */ 302 amdgpu_ring_write(kiq_ring, 0); /* oac mask */ 303 amdgpu_ring_write(kiq_ring, 0); 304 } 305 306 static void gfx_v12_0_kiq_map_queues(struct amdgpu_ring *kiq_ring, 307 struct amdgpu_ring *ring) 308 { 309 uint64_t mqd_addr = amdgpu_bo_gpu_offset(ring->mqd_obj); 310 uint64_t wptr_addr = ring->wptr_gpu_addr; 311 uint32_t me = 0, eng_sel = 0; 312 313 switch (ring->funcs->type) { 314 case AMDGPU_RING_TYPE_COMPUTE: 315 me = 1; 316 eng_sel = 0; 317 break; 318 case AMDGPU_RING_TYPE_GFX: 319 me = 0; 320 eng_sel = 4; 321 break; 322 case AMDGPU_RING_TYPE_MES: 323 me = 2; 324 eng_sel = 5; 325 break; 326 default: 327 WARN_ON(1); 328 } 329 330 amdgpu_ring_write(kiq_ring, PACKET3(PACKET3_MAP_QUEUES, 5)); 331 /* Q_sel:0, vmid:0, vidmem: 1, engine:0, num_Q:1*/ 332 amdgpu_ring_write(kiq_ring, /* Q_sel: 0, vmid: 0, engine: 0, num_Q: 1 */ 333 PACKET3_MAP_QUEUES_QUEUE_SEL(0) | /* Queue_Sel */ 334 PACKET3_MAP_QUEUES_VMID(0) | /* VMID */ 335 PACKET3_MAP_QUEUES_QUEUE(ring->queue) | 336 PACKET3_MAP_QUEUES_PIPE(ring->pipe) | 337 PACKET3_MAP_QUEUES_ME((me)) | 338 PACKET3_MAP_QUEUES_QUEUE_TYPE(0) | /*queue_type: normal compute queue */ 339 PACKET3_MAP_QUEUES_ALLOC_FORMAT(0) | /* alloc format: all_on_one_pipe */ 340 PACKET3_MAP_QUEUES_ENGINE_SEL(eng_sel) | 341 PACKET3_MAP_QUEUES_NUM_QUEUES(1)); /* num_queues: must be 1 */ 342 amdgpu_ring_write(kiq_ring, PACKET3_MAP_QUEUES_DOORBELL_OFFSET(ring->doorbell_index)); 343 amdgpu_ring_write(kiq_ring, lower_32_bits(mqd_addr)); 344 amdgpu_ring_write(kiq_ring, upper_32_bits(mqd_addr)); 345 amdgpu_ring_write(kiq_ring, lower_32_bits(wptr_addr)); 346 amdgpu_ring_write(kiq_ring, upper_32_bits(wptr_addr)); 347 } 348 349 static void gfx_v12_0_kiq_unmap_queues(struct amdgpu_ring *kiq_ring, 350 struct amdgpu_ring *ring, 351 enum amdgpu_unmap_queues_action action, 352 u64 gpu_addr, u64 seq) 353 { 354 struct amdgpu_device *adev = kiq_ring->adev; 355 uint32_t eng_sel = ring->funcs->type == AMDGPU_RING_TYPE_GFX ? 4 : 0; 356 357 if (adev->enable_mes && !adev->gfx.kiq[0].ring.sched.ready) { 358 amdgpu_mes_unmap_legacy_queue(adev, ring, action, 359 gpu_addr, seq, 0); 360 return; 361 } 362 363 amdgpu_ring_write(kiq_ring, PACKET3(PACKET3_UNMAP_QUEUES, 4)); 364 amdgpu_ring_write(kiq_ring, /* Q_sel: 0, vmid: 0, engine: 0, num_Q: 1 */ 365 PACKET3_UNMAP_QUEUES_ACTION(action) | 366 PACKET3_UNMAP_QUEUES_QUEUE_SEL(0) | 367 PACKET3_UNMAP_QUEUES_ENGINE_SEL(eng_sel) | 368 PACKET3_UNMAP_QUEUES_NUM_QUEUES(1)); 369 amdgpu_ring_write(kiq_ring, 370 PACKET3_UNMAP_QUEUES_DOORBELL_OFFSET0(ring->doorbell_index)); 371 372 if (action == PREEMPT_QUEUES_NO_UNMAP) { 373 amdgpu_ring_write(kiq_ring, lower_32_bits(gpu_addr)); 374 amdgpu_ring_write(kiq_ring, upper_32_bits(gpu_addr)); 375 amdgpu_ring_write(kiq_ring, seq); 376 } else { 377 amdgpu_ring_write(kiq_ring, 0); 378 amdgpu_ring_write(kiq_ring, 0); 379 amdgpu_ring_write(kiq_ring, 0); 380 } 381 } 382 383 static void gfx_v12_0_kiq_query_status(struct amdgpu_ring *kiq_ring, 384 struct amdgpu_ring *ring, 385 u64 addr, u64 seq) 386 { 387 uint32_t eng_sel = ring->funcs->type == AMDGPU_RING_TYPE_GFX ? 4 : 0; 388 389 amdgpu_ring_write(kiq_ring, PACKET3(PACKET3_QUERY_STATUS, 5)); 390 amdgpu_ring_write(kiq_ring, 391 PACKET3_QUERY_STATUS_CONTEXT_ID(0) | 392 PACKET3_QUERY_STATUS_INTERRUPT_SEL(0) | 393 PACKET3_QUERY_STATUS_COMMAND(2)); 394 amdgpu_ring_write(kiq_ring, /* Q_sel: 0, vmid: 0, engine: 0, num_Q: 1 */ 395 PACKET3_QUERY_STATUS_DOORBELL_OFFSET(ring->doorbell_index) | 396 PACKET3_QUERY_STATUS_ENG_SEL(eng_sel)); 397 amdgpu_ring_write(kiq_ring, lower_32_bits(addr)); 398 amdgpu_ring_write(kiq_ring, upper_32_bits(addr)); 399 amdgpu_ring_write(kiq_ring, lower_32_bits(seq)); 400 amdgpu_ring_write(kiq_ring, upper_32_bits(seq)); 401 } 402 403 static void gfx_v12_0_kiq_invalidate_tlbs(struct amdgpu_ring *kiq_ring, 404 uint16_t pasid, 405 uint32_t flush_type, 406 bool all_hub) 407 { 408 gfx_v12_0_ring_invalidate_tlbs(kiq_ring, pasid, flush_type, all_hub, 1); 409 } 410 411 static const struct kiq_pm4_funcs gfx_v12_0_kiq_pm4_funcs = { 412 .kiq_set_resources = gfx_v12_0_kiq_set_resources, 413 .kiq_map_queues = gfx_v12_0_kiq_map_queues, 414 .kiq_unmap_queues = gfx_v12_0_kiq_unmap_queues, 415 .kiq_query_status = gfx_v12_0_kiq_query_status, 416 .kiq_invalidate_tlbs = gfx_v12_0_kiq_invalidate_tlbs, 417 .set_resources_size = 8, 418 .map_queues_size = 7, 419 .unmap_queues_size = 6, 420 .query_status_size = 7, 421 .invalidate_tlbs_size = 2, 422 }; 423 424 static void gfx_v12_0_set_kiq_pm4_funcs(struct amdgpu_device *adev) 425 { 426 adev->gfx.kiq[0].pmf = &gfx_v12_0_kiq_pm4_funcs; 427 } 428 429 static void gfx_v12_0_wait_reg_mem(struct amdgpu_ring *ring, int eng_sel, 430 int mem_space, int opt, uint32_t addr0, 431 uint32_t addr1, uint32_t ref, 432 uint32_t mask, uint32_t inv) 433 { 434 amdgpu_ring_write(ring, PACKET3(PACKET3_WAIT_REG_MEM, 5)); 435 amdgpu_ring_write(ring, 436 /* memory (1) or register (0) */ 437 (WAIT_REG_MEM_MEM_SPACE(mem_space) | 438 WAIT_REG_MEM_OPERATION(opt) | /* wait */ 439 WAIT_REG_MEM_FUNCTION(3) | /* equal */ 440 WAIT_REG_MEM_ENGINE(eng_sel))); 441 442 if (mem_space) 443 BUG_ON(addr0 & 0x3); /* Dword align */ 444 amdgpu_ring_write(ring, addr0); 445 amdgpu_ring_write(ring, addr1); 446 amdgpu_ring_write(ring, ref); 447 amdgpu_ring_write(ring, mask); 448 amdgpu_ring_write(ring, inv); /* poll interval */ 449 } 450 451 static int gfx_v12_0_ring_test_ring(struct amdgpu_ring *ring) 452 { 453 struct amdgpu_device *adev = ring->adev; 454 uint32_t scratch = SOC15_REG_OFFSET(GC, 0, regSCRATCH_REG0); 455 uint32_t tmp = 0; 456 unsigned i; 457 int r; 458 459 WREG32(scratch, 0xCAFEDEAD); 460 r = amdgpu_ring_alloc(ring, 5); 461 if (r) { 462 drm_err(adev_to_drm(adev), 463 "cp failed to lock ring %d (%d).\n", 464 ring->idx, r); 465 return r; 466 } 467 468 if (ring->funcs->type == AMDGPU_RING_TYPE_KIQ) { 469 gfx_v12_0_ring_emit_wreg(ring, scratch, 0xDEADBEEF); 470 } else { 471 amdgpu_ring_write(ring, PACKET3(PACKET3_SET_UCONFIG_REG, 1)); 472 amdgpu_ring_write(ring, scratch - 473 PACKET3_SET_UCONFIG_REG_START); 474 amdgpu_ring_write(ring, 0xDEADBEEF); 475 } 476 amdgpu_ring_commit(ring); 477 478 for (i = 0; i < adev->usec_timeout; i++) { 479 tmp = RREG32(scratch); 480 if (tmp == 0xDEADBEEF) 481 break; 482 if (amdgpu_emu_mode == 1) 483 msleep(1); 484 else 485 udelay(1); 486 } 487 488 if (i >= adev->usec_timeout) 489 r = -ETIMEDOUT; 490 return r; 491 } 492 493 static int gfx_v12_0_ring_test_ib(struct amdgpu_ring *ring, long timeout) 494 { 495 struct amdgpu_device *adev = ring->adev; 496 struct amdgpu_ib ib; 497 struct dma_fence *f = NULL; 498 unsigned index; 499 uint64_t gpu_addr; 500 uint32_t *cpu_ptr; 501 long r; 502 503 /* MES KIQ fw hasn't indirect buffer support for now */ 504 if (adev->enable_mes_kiq && 505 ring->funcs->type == AMDGPU_RING_TYPE_KIQ) 506 return 0; 507 508 memset(&ib, 0, sizeof(ib)); 509 510 r = amdgpu_device_wb_get(adev, &index); 511 if (r) 512 return r; 513 514 gpu_addr = adev->wb.gpu_addr + (index * 4); 515 adev->wb.wb[index] = cpu_to_le32(0xCAFEDEAD); 516 cpu_ptr = &adev->wb.wb[index]; 517 518 r = amdgpu_ib_get(adev, NULL, 16, AMDGPU_IB_POOL_DIRECT, &ib); 519 if (r) { 520 drm_err(adev_to_drm(adev), "failed to get ib (%ld).\n", r); 521 goto err1; 522 } 523 524 ib.ptr[0] = PACKET3(PACKET3_WRITE_DATA, 3); 525 ib.ptr[1] = WRITE_DATA_DST_SEL(5) | WR_CONFIRM; 526 ib.ptr[2] = lower_32_bits(gpu_addr); 527 ib.ptr[3] = upper_32_bits(gpu_addr); 528 ib.ptr[4] = 0xDEADBEEF; 529 ib.length_dw = 5; 530 531 r = amdgpu_ib_schedule(ring, 1, &ib, NULL, &f); 532 if (r) 533 goto err2; 534 535 r = dma_fence_wait_timeout(f, false, timeout); 536 if (r == 0) { 537 r = -ETIMEDOUT; 538 goto err2; 539 } else if (r < 0) { 540 goto err2; 541 } 542 543 if (le32_to_cpu(*cpu_ptr) == 0xDEADBEEF) 544 r = 0; 545 else 546 r = -EINVAL; 547 err2: 548 amdgpu_ib_free(&ib, NULL); 549 dma_fence_put(f); 550 err1: 551 amdgpu_device_wb_free(adev, index); 552 return r; 553 } 554 555 static void gfx_v12_0_free_microcode(struct amdgpu_device *adev) 556 { 557 amdgpu_ucode_release(&adev->gfx.pfp_fw); 558 amdgpu_ucode_release(&adev->gfx.me_fw); 559 amdgpu_ucode_release(&adev->gfx.rlc_fw); 560 amdgpu_ucode_release(&adev->gfx.mec_fw); 561 562 kfree(adev->gfx.rlc.register_list_format); 563 } 564 565 static int gfx_v12_0_init_toc_microcode(struct amdgpu_device *adev, const char *ucode_prefix) 566 { 567 const struct psp_firmware_header_v1_0 *toc_hdr; 568 int err = 0; 569 570 err = amdgpu_ucode_request(adev, &adev->psp.toc_fw, 571 AMDGPU_UCODE_REQUIRED, 572 "amdgpu/%s_toc.bin", ucode_prefix); 573 if (err) 574 goto out; 575 576 toc_hdr = (const struct psp_firmware_header_v1_0 *)adev->psp.toc_fw->data; 577 adev->psp.toc.fw_version = le32_to_cpu(toc_hdr->header.ucode_version); 578 adev->psp.toc.feature_version = le32_to_cpu(toc_hdr->sos.fw_version); 579 adev->psp.toc.size_bytes = le32_to_cpu(toc_hdr->header.ucode_size_bytes); 580 adev->psp.toc.start_addr = (uint8_t *)toc_hdr + 581 le32_to_cpu(toc_hdr->header.ucode_array_offset_bytes); 582 return 0; 583 out: 584 amdgpu_ucode_release(&adev->psp.toc_fw); 585 return err; 586 } 587 588 static int gfx_v12_0_init_microcode(struct amdgpu_device *adev) 589 { 590 char ucode_prefix[30]; 591 int err; 592 const struct rlc_firmware_header_v2_0 *rlc_hdr; 593 uint16_t version_major; 594 uint16_t version_minor; 595 596 DRM_DEBUG("\n"); 597 598 amdgpu_ucode_ip_version_decode(adev, GC_HWIP, ucode_prefix, sizeof(ucode_prefix)); 599 600 err = amdgpu_ucode_request(adev, &adev->gfx.pfp_fw, 601 AMDGPU_UCODE_REQUIRED, 602 "amdgpu/%s_pfp.bin", ucode_prefix); 603 if (err) 604 goto out; 605 amdgpu_gfx_cp_init_microcode(adev, AMDGPU_UCODE_ID_CP_RS64_PFP); 606 amdgpu_gfx_cp_init_microcode(adev, AMDGPU_UCODE_ID_CP_RS64_PFP_P0_STACK); 607 608 err = amdgpu_ucode_request(adev, &adev->gfx.me_fw, 609 AMDGPU_UCODE_REQUIRED, 610 "amdgpu/%s_me.bin", ucode_prefix); 611 if (err) 612 goto out; 613 amdgpu_gfx_cp_init_microcode(adev, AMDGPU_UCODE_ID_CP_RS64_ME); 614 amdgpu_gfx_cp_init_microcode(adev, AMDGPU_UCODE_ID_CP_RS64_ME_P0_STACK); 615 616 if (!amdgpu_sriov_vf(adev)) { 617 if (amdgpu_is_kicker_fw(adev)) 618 err = amdgpu_ucode_request(adev, &adev->gfx.rlc_fw, 619 AMDGPU_UCODE_REQUIRED, 620 "amdgpu/%s_rlc_kicker.bin", ucode_prefix); 621 else 622 err = amdgpu_ucode_request(adev, &adev->gfx.rlc_fw, 623 AMDGPU_UCODE_REQUIRED, 624 "amdgpu/%s_rlc.bin", ucode_prefix); 625 if (err) 626 goto out; 627 rlc_hdr = (const struct rlc_firmware_header_v2_0 *)adev->gfx.rlc_fw->data; 628 version_major = le16_to_cpu(rlc_hdr->header.header_version_major); 629 version_minor = le16_to_cpu(rlc_hdr->header.header_version_minor); 630 err = amdgpu_gfx_rlc_init_microcode(adev, version_major, version_minor); 631 if (err) 632 goto out; 633 } 634 635 err = amdgpu_ucode_request(adev, &adev->gfx.mec_fw, 636 AMDGPU_UCODE_REQUIRED, 637 "amdgpu/%s_mec.bin", ucode_prefix); 638 if (err) 639 goto out; 640 amdgpu_gfx_cp_init_microcode(adev, AMDGPU_UCODE_ID_CP_RS64_MEC); 641 amdgpu_gfx_cp_init_microcode(adev, AMDGPU_UCODE_ID_CP_RS64_MEC_P0_STACK); 642 amdgpu_gfx_cp_init_microcode(adev, AMDGPU_UCODE_ID_CP_RS64_MEC_P1_STACK); 643 644 if (adev->firmware.load_type == AMDGPU_FW_LOAD_RLC_BACKDOOR_AUTO) 645 err = gfx_v12_0_init_toc_microcode(adev, ucode_prefix); 646 647 /* only one MEC for gfx 12 */ 648 adev->gfx.mec2_fw = NULL; 649 650 if (adev->gfx.imu.funcs) { 651 if (adev->gfx.imu.funcs->init_microcode) { 652 err = adev->gfx.imu.funcs->init_microcode(adev); 653 if (err) 654 dev_err(adev->dev, "Failed to load imu firmware!\n"); 655 } 656 } 657 658 out: 659 if (err) { 660 amdgpu_ucode_release(&adev->gfx.pfp_fw); 661 amdgpu_ucode_release(&adev->gfx.me_fw); 662 amdgpu_ucode_release(&adev->gfx.rlc_fw); 663 amdgpu_ucode_release(&adev->gfx.mec_fw); 664 } 665 666 return err; 667 } 668 669 static u32 gfx_v12_0_get_csb_size(struct amdgpu_device *adev) 670 { 671 u32 count = 0; 672 const struct cs_section_def *sect = NULL; 673 const struct cs_extent_def *ext = NULL; 674 675 count += 1; 676 677 for (sect = gfx12_cs_data; sect->section != NULL; ++sect) { 678 if (sect->id == SECT_CONTEXT) { 679 for (ext = sect->section; ext->extent != NULL; ++ext) 680 count += 2 + ext->reg_count; 681 } else 682 return 0; 683 } 684 685 return count; 686 } 687 688 static void gfx_v12_0_get_csb_buffer(struct amdgpu_device *adev, u32 *buffer) 689 { 690 u32 count = 0, clustercount = 0, i; 691 const struct cs_section_def *sect = NULL; 692 const struct cs_extent_def *ext = NULL; 693 694 if (adev->gfx.rlc.cs_data == NULL) 695 return; 696 if (buffer == NULL) 697 return; 698 699 count += 1; 700 701 for (sect = adev->gfx.rlc.cs_data; sect->section != NULL; ++sect) { 702 if (sect->id == SECT_CONTEXT) { 703 for (ext = sect->section; ext->extent != NULL; ++ext) { 704 clustercount++; 705 buffer[count++] = ext->reg_count; 706 buffer[count++] = ext->reg_index; 707 708 for (i = 0; i < ext->reg_count; i++) 709 buffer[count++] = cpu_to_le32(ext->extent[i]); 710 } 711 } else 712 return; 713 } 714 715 buffer[0] = clustercount; 716 } 717 718 static void gfx_v12_0_rlc_fini(struct amdgpu_device *adev) 719 { 720 /* clear state block */ 721 amdgpu_bo_free_kernel(&adev->gfx.rlc.clear_state_obj, 722 &adev->gfx.rlc.clear_state_gpu_addr, 723 (void **)&adev->gfx.rlc.cs_ptr); 724 725 /* jump table block */ 726 amdgpu_bo_free_kernel(&adev->gfx.rlc.cp_table_obj, 727 &adev->gfx.rlc.cp_table_gpu_addr, 728 (void **)&adev->gfx.rlc.cp_table_ptr); 729 } 730 731 static void gfx_v12_0_init_rlcg_reg_access_ctrl(struct amdgpu_device *adev) 732 { 733 struct amdgpu_rlcg_reg_access_ctrl *reg_access_ctrl; 734 735 reg_access_ctrl = &adev->gfx.rlc.reg_access_ctrl[0]; 736 reg_access_ctrl->scratch_reg0 = SOC15_REG_OFFSET(GC, 0, regSCRATCH_REG0); 737 reg_access_ctrl->scratch_reg1 = SOC15_REG_OFFSET(GC, 0, regSCRATCH_REG1); 738 reg_access_ctrl->scratch_reg2 = SOC15_REG_OFFSET(GC, 0, regSCRATCH_REG2); 739 reg_access_ctrl->scratch_reg3 = SOC15_REG_OFFSET(GC, 0, regSCRATCH_REG3); 740 reg_access_ctrl->grbm_cntl = SOC15_REG_OFFSET(GC, 0, regGRBM_GFX_CNTL); 741 reg_access_ctrl->grbm_idx = SOC15_REG_OFFSET(GC, 0, regGRBM_GFX_INDEX); 742 reg_access_ctrl->spare_int = SOC15_REG_OFFSET(GC, 0, regRLC_SPARE_INT_0); 743 adev->gfx.rlc.rlcg_reg_access_supported = true; 744 } 745 746 static int gfx_v12_0_rlc_init(struct amdgpu_device *adev) 747 { 748 const struct cs_section_def *cs_data; 749 int r; 750 751 adev->gfx.rlc.cs_data = gfx12_cs_data; 752 753 cs_data = adev->gfx.rlc.cs_data; 754 755 if (cs_data) { 756 /* init clear state block */ 757 r = amdgpu_gfx_rlc_init_csb(adev); 758 if (r) 759 return r; 760 } 761 762 /* init spm vmid with 0xf */ 763 if (adev->gfx.rlc.funcs->update_spm_vmid) 764 adev->gfx.rlc.funcs->update_spm_vmid(adev, 0, NULL, 0xf); 765 766 return 0; 767 } 768 769 static void gfx_v12_0_mec_fini(struct amdgpu_device *adev) 770 { 771 amdgpu_bo_free_kernel(&adev->gfx.mec.hpd_eop_obj, NULL, NULL); 772 amdgpu_bo_free_kernel(&adev->gfx.mec.mec_fw_obj, NULL, NULL); 773 amdgpu_bo_free_kernel(&adev->gfx.mec.mec_fw_data_obj, NULL, NULL); 774 } 775 776 static void gfx_v12_0_me_init(struct amdgpu_device *adev) 777 { 778 bitmap_zero(adev->gfx.me.queue_bitmap, AMDGPU_MAX_GFX_QUEUES); 779 780 amdgpu_gfx_graphics_queue_acquire(adev); 781 } 782 783 static int gfx_v12_0_mec_init(struct amdgpu_device *adev) 784 { 785 int r; 786 u32 *hpd; 787 size_t mec_hpd_size; 788 789 bitmap_zero(adev->gfx.mec_bitmap[0].queue_bitmap, AMDGPU_MAX_COMPUTE_QUEUES); 790 791 /* take ownership of the relevant compute queues */ 792 amdgpu_gfx_compute_queue_acquire(adev); 793 mec_hpd_size = adev->gfx.num_compute_rings * GFX12_MEC_HPD_SIZE; 794 795 if (mec_hpd_size) { 796 r = amdgpu_bo_create_reserved(adev, mec_hpd_size, PAGE_SIZE, 797 AMDGPU_GEM_DOMAIN_GTT, 798 &adev->gfx.mec.hpd_eop_obj, 799 &adev->gfx.mec.hpd_eop_gpu_addr, 800 (void **)&hpd); 801 if (r) { 802 dev_warn(adev->dev, "(%d) create HDP EOP bo failed\n", r); 803 gfx_v12_0_mec_fini(adev); 804 return r; 805 } 806 807 memset(hpd, 0, mec_hpd_size); 808 809 amdgpu_bo_kunmap(adev->gfx.mec.hpd_eop_obj); 810 amdgpu_bo_unreserve(adev->gfx.mec.hpd_eop_obj); 811 } 812 813 return 0; 814 } 815 816 static uint32_t wave_read_ind(struct amdgpu_device *adev, uint32_t wave, uint32_t address) 817 { 818 WREG32_SOC15(GC, 0, regSQ_IND_INDEX, 819 (wave << SQ_IND_INDEX__WAVE_ID__SHIFT) | 820 (address << SQ_IND_INDEX__INDEX__SHIFT)); 821 return RREG32_SOC15(GC, 0, regSQ_IND_DATA); 822 } 823 824 static void wave_read_regs(struct amdgpu_device *adev, uint32_t wave, 825 uint32_t thread, uint32_t regno, 826 uint32_t num, uint32_t *out) 827 { 828 WREG32_SOC15(GC, 0, regSQ_IND_INDEX, 829 (wave << SQ_IND_INDEX__WAVE_ID__SHIFT) | 830 (regno << SQ_IND_INDEX__INDEX__SHIFT) | 831 (thread << SQ_IND_INDEX__WORKITEM_ID__SHIFT) | 832 (SQ_IND_INDEX__AUTO_INCR_MASK)); 833 while (num--) 834 *(out++) = RREG32_SOC15(GC, 0, regSQ_IND_DATA); 835 } 836 837 static void gfx_v12_0_read_wave_data(struct amdgpu_device *adev, 838 uint32_t xcc_id, 839 uint32_t simd, uint32_t wave, 840 uint32_t *dst, int *no_fields) 841 { 842 /* in gfx12 the SIMD_ID is specified as part of the INSTANCE 843 * field when performing a select_se_sh so it should be 844 * zero here */ 845 WARN_ON(simd != 0); 846 847 /* type 4 wave data */ 848 dst[(*no_fields)++] = 4; 849 dst[(*no_fields)++] = wave_read_ind(adev, wave, ixSQ_WAVE_STATUS); 850 dst[(*no_fields)++] = wave_read_ind(adev, wave, ixSQ_WAVE_PC_LO); 851 dst[(*no_fields)++] = wave_read_ind(adev, wave, ixSQ_WAVE_PC_HI); 852 dst[(*no_fields)++] = wave_read_ind(adev, wave, ixSQ_WAVE_EXEC_LO); 853 dst[(*no_fields)++] = wave_read_ind(adev, wave, ixSQ_WAVE_EXEC_HI); 854 dst[(*no_fields)++] = wave_read_ind(adev, wave, ixSQ_WAVE_HW_ID1); 855 dst[(*no_fields)++] = wave_read_ind(adev, wave, ixSQ_WAVE_HW_ID2); 856 dst[(*no_fields)++] = wave_read_ind(adev, wave, ixSQ_WAVE_GPR_ALLOC); 857 dst[(*no_fields)++] = wave_read_ind(adev, wave, ixSQ_WAVE_LDS_ALLOC); 858 dst[(*no_fields)++] = wave_read_ind(adev, wave, ixSQ_WAVE_IB_STS); 859 dst[(*no_fields)++] = wave_read_ind(adev, wave, ixSQ_WAVE_IB_STS2); 860 dst[(*no_fields)++] = wave_read_ind(adev, wave, ixSQ_WAVE_IB_DBG1); 861 dst[(*no_fields)++] = wave_read_ind(adev, wave, ixSQ_WAVE_M0); 862 dst[(*no_fields)++] = wave_read_ind(adev, wave, ixSQ_WAVE_MODE); 863 dst[(*no_fields)++] = wave_read_ind(adev, wave, ixSQ_WAVE_STATE_PRIV); 864 dst[(*no_fields)++] = wave_read_ind(adev, wave, ixSQ_WAVE_EXCP_FLAG_PRIV); 865 dst[(*no_fields)++] = wave_read_ind(adev, wave, ixSQ_WAVE_EXCP_FLAG_USER); 866 dst[(*no_fields)++] = wave_read_ind(adev, wave, ixSQ_WAVE_TRAP_CTRL); 867 dst[(*no_fields)++] = wave_read_ind(adev, wave, ixSQ_WAVE_ACTIVE); 868 dst[(*no_fields)++] = wave_read_ind(adev, wave, ixSQ_WAVE_VALID_AND_IDLE); 869 dst[(*no_fields)++] = wave_read_ind(adev, wave, ixSQ_WAVE_DVGPR_ALLOC_LO); 870 dst[(*no_fields)++] = wave_read_ind(adev, wave, ixSQ_WAVE_DVGPR_ALLOC_HI); 871 dst[(*no_fields)++] = wave_read_ind(adev, wave, ixSQ_WAVE_SCHED_MODE); 872 } 873 874 static void gfx_v12_0_read_wave_sgprs(struct amdgpu_device *adev, 875 uint32_t xcc_id, uint32_t simd, 876 uint32_t wave, uint32_t start, 877 uint32_t size, uint32_t *dst) 878 { 879 WARN_ON(simd != 0); 880 881 wave_read_regs( 882 adev, wave, 0, start + SQIND_WAVE_SGPRS_OFFSET, size, 883 dst); 884 } 885 886 static void gfx_v12_0_read_wave_vgprs(struct amdgpu_device *adev, 887 uint32_t xcc_id, uint32_t simd, 888 uint32_t wave, uint32_t thread, 889 uint32_t start, uint32_t size, 890 uint32_t *dst) 891 { 892 wave_read_regs( 893 adev, wave, thread, 894 start + SQIND_WAVE_VGPRS_OFFSET, size, dst); 895 } 896 897 static void gfx_v12_0_select_me_pipe_q(struct amdgpu_device *adev, 898 u32 me, u32 pipe, u32 q, u32 vm, u32 xcc_id) 899 { 900 soc24_grbm_select(adev, me, pipe, q, vm); 901 } 902 903 /* all sizes are in bytes */ 904 #define MQD_SHADOW_BASE_SIZE 73728 905 #define MQD_SHADOW_BASE_ALIGNMENT 256 906 #define MQD_FWWORKAREA_SIZE 484 907 #define MQD_FWWORKAREA_ALIGNMENT 256 908 909 static void gfx_v12_0_get_gfx_shadow_info_nocheck(struct amdgpu_device *adev, 910 struct amdgpu_gfx_shadow_info *shadow_info) 911 { 912 /* for gfx */ 913 shadow_info->shadow_size = MQD_SHADOW_BASE_SIZE; 914 shadow_info->shadow_alignment = MQD_SHADOW_BASE_ALIGNMENT; 915 shadow_info->csa_size = MQD_FWWORKAREA_SIZE; 916 shadow_info->csa_alignment = MQD_FWWORKAREA_ALIGNMENT; 917 /* for compute */ 918 shadow_info->eop_size = GFX12_MEC_HPD_SIZE; 919 shadow_info->eop_alignment = 256; 920 } 921 922 static int gfx_v12_0_get_gfx_shadow_info(struct amdgpu_device *adev, 923 struct amdgpu_gfx_shadow_info *shadow_info, 924 bool skip_check) 925 { 926 if (adev->gfx.cp_gfx_shadow || skip_check) { 927 gfx_v12_0_get_gfx_shadow_info_nocheck(adev, shadow_info); 928 return 0; 929 } 930 931 memset(shadow_info, 0, sizeof(struct amdgpu_gfx_shadow_info)); 932 return -EINVAL; 933 } 934 935 static const struct amdgpu_gfx_funcs gfx_v12_0_gfx_funcs = { 936 .get_gpu_clock_counter = &gfx_v12_0_get_gpu_clock_counter, 937 .select_se_sh = &gfx_v12_0_select_se_sh, 938 .read_wave_data = &gfx_v12_0_read_wave_data, 939 .read_wave_sgprs = &gfx_v12_0_read_wave_sgprs, 940 .read_wave_vgprs = &gfx_v12_0_read_wave_vgprs, 941 .select_me_pipe_q = &gfx_v12_0_select_me_pipe_q, 942 .update_perfmon_mgcg = &gfx_v12_0_update_perf_clk, 943 .get_gfx_shadow_info = &gfx_v12_0_get_gfx_shadow_info, 944 .get_hdp_flush_mask = &amdgpu_gfx_get_hdp_flush_mask, 945 }; 946 947 static int gfx_v12_0_gpu_early_init(struct amdgpu_device *adev) 948 { 949 950 switch (amdgpu_ip_version(adev, GC_HWIP, 0)) { 951 case IP_VERSION(12, 0, 0): 952 case IP_VERSION(12, 0, 1): 953 adev->gfx.config.max_hw_contexts = 8; 954 adev->gfx.config.sc_prim_fifo_size_frontend = 0x20; 955 adev->gfx.config.sc_prim_fifo_size_backend = 0x100; 956 adev->gfx.config.sc_hiz_tile_fifo_size = 0; 957 adev->gfx.config.sc_earlyz_tile_fifo_size = 0x4C0; 958 break; 959 default: 960 BUG(); 961 break; 962 } 963 964 return 0; 965 } 966 967 static int gfx_v12_0_gfx_ring_init(struct amdgpu_device *adev, int ring_id, 968 int me, int pipe, int queue) 969 { 970 int r; 971 struct amdgpu_ring *ring; 972 unsigned int irq_type; 973 974 ring = &adev->gfx.gfx_ring[ring_id]; 975 976 ring->me = me; 977 ring->pipe = pipe; 978 ring->queue = queue; 979 980 ring->ring_obj = NULL; 981 ring->use_doorbell = true; 982 983 if (!ring_id) 984 ring->doorbell_index = adev->doorbell_index.gfx_ring0 << 1; 985 else 986 ring->doorbell_index = adev->doorbell_index.gfx_ring1 << 1; 987 ring->vm_hub = AMDGPU_GFXHUB(0); 988 sprintf(ring->name, "gfx_%d.%d.%d", ring->me, ring->pipe, ring->queue); 989 990 irq_type = AMDGPU_CP_IRQ_GFX_ME0_PIPE0_EOP + ring->pipe; 991 r = amdgpu_ring_init(adev, ring, 1024, &adev->gfx.eop_irq, irq_type, 992 AMDGPU_RING_PRIO_DEFAULT, NULL); 993 if (r) 994 return r; 995 return 0; 996 } 997 998 static int gfx_v12_0_compute_ring_init(struct amdgpu_device *adev, int ring_id, 999 int mec, int pipe, int queue) 1000 { 1001 int r; 1002 unsigned irq_type; 1003 struct amdgpu_ring *ring; 1004 unsigned int hw_prio; 1005 1006 ring = &adev->gfx.compute_ring[ring_id]; 1007 1008 /* mec0 is me1 */ 1009 ring->me = mec + 1; 1010 ring->pipe = pipe; 1011 ring->queue = queue; 1012 1013 ring->ring_obj = NULL; 1014 ring->use_doorbell = true; 1015 ring->doorbell_index = (adev->doorbell_index.mec_ring0 + ring_id) << 1; 1016 ring->eop_gpu_addr = adev->gfx.mec.hpd_eop_gpu_addr 1017 + (ring_id * GFX12_MEC_HPD_SIZE); 1018 ring->vm_hub = AMDGPU_GFXHUB(0); 1019 sprintf(ring->name, "comp_%d.%d.%d", ring->me, ring->pipe, ring->queue); 1020 1021 irq_type = AMDGPU_CP_IRQ_COMPUTE_MEC1_PIPE0_EOP 1022 + ((ring->me - 1) * adev->gfx.mec.num_pipe_per_mec) 1023 + ring->pipe; 1024 hw_prio = amdgpu_gfx_is_high_priority_compute_queue(adev, ring) ? 1025 AMDGPU_GFX_PIPE_PRIO_HIGH : AMDGPU_GFX_PIPE_PRIO_NORMAL; 1026 /* type-2 packets are deprecated on MEC, use type-3 instead */ 1027 r = amdgpu_ring_init(adev, ring, 1024, &adev->gfx.eop_irq, irq_type, 1028 hw_prio, NULL); 1029 if (r) 1030 return r; 1031 1032 return 0; 1033 } 1034 1035 static struct { 1036 SOC24_FIRMWARE_ID id; 1037 unsigned int offset; 1038 unsigned int size; 1039 unsigned int size_x16; 1040 } rlc_autoload_info[SOC24_FIRMWARE_ID_MAX]; 1041 1042 #define RLC_TOC_OFFSET_DWUNIT 8 1043 #define RLC_SIZE_MULTIPLE 1024 1044 #define RLC_TOC_UMF_SIZE_inM 23ULL 1045 #define RLC_TOC_FORMAT_API 165ULL 1046 1047 static void gfx_v12_0_parse_rlc_toc(struct amdgpu_device *adev, void *rlc_toc) 1048 { 1049 RLC_TABLE_OF_CONTENT_V2 *ucode = rlc_toc; 1050 1051 while (ucode && (ucode->id > SOC24_FIRMWARE_ID_INVALID)) { 1052 rlc_autoload_info[ucode->id].id = ucode->id; 1053 rlc_autoload_info[ucode->id].offset = 1054 ucode->offset * RLC_TOC_OFFSET_DWUNIT * 4; 1055 rlc_autoload_info[ucode->id].size = 1056 ucode->size_x16 ? ucode->size * RLC_SIZE_MULTIPLE * 4 : 1057 ucode->size * 4; 1058 ucode++; 1059 } 1060 } 1061 1062 static uint32_t gfx_v12_0_calc_toc_total_size(struct amdgpu_device *adev) 1063 { 1064 uint32_t total_size = 0; 1065 SOC24_FIRMWARE_ID id; 1066 1067 gfx_v12_0_parse_rlc_toc(adev, adev->psp.toc.start_addr); 1068 1069 for (id = SOC24_FIRMWARE_ID_RLC_G_UCODE; id < SOC24_FIRMWARE_ID_MAX; id++) 1070 total_size += rlc_autoload_info[id].size; 1071 1072 /* In case the offset in rlc toc ucode is aligned */ 1073 if (total_size < rlc_autoload_info[SOC24_FIRMWARE_ID_MAX-1].offset) 1074 total_size = rlc_autoload_info[SOC24_FIRMWARE_ID_MAX-1].offset + 1075 rlc_autoload_info[SOC24_FIRMWARE_ID_MAX-1].size; 1076 if (total_size < (RLC_TOC_UMF_SIZE_inM << 20)) 1077 total_size = RLC_TOC_UMF_SIZE_inM << 20; 1078 1079 return total_size; 1080 } 1081 1082 static int gfx_v12_0_rlc_autoload_buffer_init(struct amdgpu_device *adev) 1083 { 1084 int r; 1085 uint32_t total_size; 1086 1087 total_size = gfx_v12_0_calc_toc_total_size(adev); 1088 1089 r = amdgpu_bo_create_reserved(adev, total_size, 64 * 1024, 1090 AMDGPU_GEM_DOMAIN_VRAM, 1091 &adev->gfx.rlc.rlc_autoload_bo, 1092 &adev->gfx.rlc.rlc_autoload_gpu_addr, 1093 (void **)&adev->gfx.rlc.rlc_autoload_ptr); 1094 1095 if (r) { 1096 dev_err(adev->dev, "(%d) failed to create fw autoload bo\n", r); 1097 return r; 1098 } 1099 1100 return 0; 1101 } 1102 1103 static void gfx_v12_0_rlc_backdoor_autoload_copy_ucode(struct amdgpu_device *adev, 1104 SOC24_FIRMWARE_ID id, 1105 const void *fw_data, 1106 uint32_t fw_size) 1107 { 1108 uint32_t toc_offset; 1109 uint32_t toc_fw_size; 1110 char *ptr = adev->gfx.rlc.rlc_autoload_ptr; 1111 1112 if (id <= SOC24_FIRMWARE_ID_INVALID || id >= SOC24_FIRMWARE_ID_MAX) 1113 return; 1114 1115 toc_offset = rlc_autoload_info[id].offset; 1116 toc_fw_size = rlc_autoload_info[id].size; 1117 1118 if (fw_size == 0) 1119 fw_size = toc_fw_size; 1120 1121 if (fw_size > toc_fw_size) 1122 fw_size = toc_fw_size; 1123 1124 memcpy(ptr + toc_offset, fw_data, fw_size); 1125 1126 if (fw_size < toc_fw_size) 1127 memset(ptr + toc_offset + fw_size, 0, toc_fw_size - fw_size); 1128 } 1129 1130 static void 1131 gfx_v12_0_rlc_backdoor_autoload_copy_toc_ucode(struct amdgpu_device *adev) 1132 { 1133 void *data; 1134 uint32_t size; 1135 uint32_t *toc_ptr; 1136 1137 data = adev->psp.toc.start_addr; 1138 size = rlc_autoload_info[SOC24_FIRMWARE_ID_RLC_TOC].size; 1139 1140 toc_ptr = (uint32_t *)data + size / 4 - 2; 1141 *toc_ptr = (RLC_TOC_FORMAT_API << 24) | 0x1; 1142 1143 gfx_v12_0_rlc_backdoor_autoload_copy_ucode(adev, SOC24_FIRMWARE_ID_RLC_TOC, 1144 data, size); 1145 } 1146 1147 static void 1148 gfx_v12_0_rlc_backdoor_autoload_copy_gfx_ucode(struct amdgpu_device *adev) 1149 { 1150 const __le32 *fw_data; 1151 uint32_t fw_size; 1152 const struct gfx_firmware_header_v2_0 *cpv2_hdr; 1153 const struct rlc_firmware_header_v2_0 *rlc_hdr; 1154 const struct rlc_firmware_header_v2_1 *rlcv21_hdr; 1155 const struct rlc_firmware_header_v2_2 *rlcv22_hdr; 1156 uint16_t version_major, version_minor; 1157 1158 /* pfp ucode */ 1159 cpv2_hdr = (const struct gfx_firmware_header_v2_0 *) 1160 adev->gfx.pfp_fw->data; 1161 /* instruction */ 1162 fw_data = (const __le32 *)(adev->gfx.pfp_fw->data + 1163 le32_to_cpu(cpv2_hdr->ucode_offset_bytes)); 1164 fw_size = le32_to_cpu(cpv2_hdr->ucode_size_bytes); 1165 gfx_v12_0_rlc_backdoor_autoload_copy_ucode(adev, SOC24_FIRMWARE_ID_RS64_PFP, 1166 fw_data, fw_size); 1167 /* data */ 1168 fw_data = (const __le32 *)(adev->gfx.pfp_fw->data + 1169 le32_to_cpu(cpv2_hdr->data_offset_bytes)); 1170 fw_size = le32_to_cpu(cpv2_hdr->data_size_bytes); 1171 gfx_v12_0_rlc_backdoor_autoload_copy_ucode(adev, SOC24_FIRMWARE_ID_RS64_PFP_P0_STACK, 1172 fw_data, fw_size); 1173 gfx_v12_0_rlc_backdoor_autoload_copy_ucode(adev, SOC24_FIRMWARE_ID_RS64_PFP_P1_STACK, 1174 fw_data, fw_size); 1175 /* me ucode */ 1176 cpv2_hdr = (const struct gfx_firmware_header_v2_0 *) 1177 adev->gfx.me_fw->data; 1178 /* instruction */ 1179 fw_data = (const __le32 *)(adev->gfx.me_fw->data + 1180 le32_to_cpu(cpv2_hdr->ucode_offset_bytes)); 1181 fw_size = le32_to_cpu(cpv2_hdr->ucode_size_bytes); 1182 gfx_v12_0_rlc_backdoor_autoload_copy_ucode(adev, SOC24_FIRMWARE_ID_RS64_ME, 1183 fw_data, fw_size); 1184 /* data */ 1185 fw_data = (const __le32 *)(adev->gfx.me_fw->data + 1186 le32_to_cpu(cpv2_hdr->data_offset_bytes)); 1187 fw_size = le32_to_cpu(cpv2_hdr->data_size_bytes); 1188 gfx_v12_0_rlc_backdoor_autoload_copy_ucode(adev, SOC24_FIRMWARE_ID_RS64_ME_P0_STACK, 1189 fw_data, fw_size); 1190 gfx_v12_0_rlc_backdoor_autoload_copy_ucode(adev, SOC24_FIRMWARE_ID_RS64_ME_P1_STACK, 1191 fw_data, fw_size); 1192 /* mec ucode */ 1193 cpv2_hdr = (const struct gfx_firmware_header_v2_0 *) 1194 adev->gfx.mec_fw->data; 1195 /* instruction */ 1196 fw_data = (const __le32 *) (adev->gfx.mec_fw->data + 1197 le32_to_cpu(cpv2_hdr->ucode_offset_bytes)); 1198 fw_size = le32_to_cpu(cpv2_hdr->ucode_size_bytes); 1199 gfx_v12_0_rlc_backdoor_autoload_copy_ucode(adev, SOC24_FIRMWARE_ID_RS64_MEC, 1200 fw_data, fw_size); 1201 /* data */ 1202 fw_data = (const __le32 *) (adev->gfx.mec_fw->data + 1203 le32_to_cpu(cpv2_hdr->data_offset_bytes)); 1204 fw_size = le32_to_cpu(cpv2_hdr->data_size_bytes); 1205 gfx_v12_0_rlc_backdoor_autoload_copy_ucode(adev, SOC24_FIRMWARE_ID_RS64_MEC_P0_STACK, 1206 fw_data, fw_size); 1207 gfx_v12_0_rlc_backdoor_autoload_copy_ucode(adev, SOC24_FIRMWARE_ID_RS64_MEC_P1_STACK, 1208 fw_data, fw_size); 1209 gfx_v12_0_rlc_backdoor_autoload_copy_ucode(adev, SOC24_FIRMWARE_ID_RS64_MEC_P2_STACK, 1210 fw_data, fw_size); 1211 gfx_v12_0_rlc_backdoor_autoload_copy_ucode(adev, SOC24_FIRMWARE_ID_RS64_MEC_P3_STACK, 1212 fw_data, fw_size); 1213 1214 /* rlc ucode */ 1215 rlc_hdr = (const struct rlc_firmware_header_v2_0 *) 1216 adev->gfx.rlc_fw->data; 1217 fw_data = (const __le32 *)(adev->gfx.rlc_fw->data + 1218 le32_to_cpu(rlc_hdr->header.ucode_array_offset_bytes)); 1219 fw_size = le32_to_cpu(rlc_hdr->header.ucode_size_bytes); 1220 gfx_v12_0_rlc_backdoor_autoload_copy_ucode(adev, SOC24_FIRMWARE_ID_RLC_G_UCODE, 1221 fw_data, fw_size); 1222 1223 version_major = le16_to_cpu(rlc_hdr->header.header_version_major); 1224 version_minor = le16_to_cpu(rlc_hdr->header.header_version_minor); 1225 if (version_major == 2) { 1226 if (version_minor >= 1) { 1227 rlcv21_hdr = (const struct rlc_firmware_header_v2_1 *)adev->gfx.rlc_fw->data; 1228 1229 fw_data = (const __le32 *)(adev->gfx.rlc_fw->data + 1230 le32_to_cpu(rlcv21_hdr->save_restore_list_gpm_offset_bytes)); 1231 fw_size = le32_to_cpu(rlcv21_hdr->save_restore_list_gpm_size_bytes); 1232 gfx_v12_0_rlc_backdoor_autoload_copy_ucode(adev, SOC24_FIRMWARE_ID_RLCG_SCRATCH, 1233 fw_data, fw_size); 1234 1235 fw_data = (const __le32 *)(adev->gfx.rlc_fw->data + 1236 le32_to_cpu(rlcv21_hdr->save_restore_list_srm_offset_bytes)); 1237 fw_size = le32_to_cpu(rlcv21_hdr->save_restore_list_srm_size_bytes); 1238 gfx_v12_0_rlc_backdoor_autoload_copy_ucode(adev, SOC24_FIRMWARE_ID_RLC_SRM_ARAM, 1239 fw_data, fw_size); 1240 } 1241 if (version_minor >= 2) { 1242 rlcv22_hdr = (const struct rlc_firmware_header_v2_2 *)adev->gfx.rlc_fw->data; 1243 1244 fw_data = (const __le32 *)(adev->gfx.rlc_fw->data + 1245 le32_to_cpu(rlcv22_hdr->rlc_iram_ucode_offset_bytes)); 1246 fw_size = le32_to_cpu(rlcv22_hdr->rlc_iram_ucode_size_bytes); 1247 gfx_v12_0_rlc_backdoor_autoload_copy_ucode(adev, SOC24_FIRMWARE_ID_RLX6_UCODE, 1248 fw_data, fw_size); 1249 1250 fw_data = (const __le32 *)(adev->gfx.rlc_fw->data + 1251 le32_to_cpu(rlcv22_hdr->rlc_dram_ucode_offset_bytes)); 1252 fw_size = le32_to_cpu(rlcv22_hdr->rlc_dram_ucode_size_bytes); 1253 gfx_v12_0_rlc_backdoor_autoload_copy_ucode(adev, SOC24_FIRMWARE_ID_RLX6_DRAM_BOOT, 1254 fw_data, fw_size); 1255 } 1256 } 1257 } 1258 1259 static void 1260 gfx_v12_0_rlc_backdoor_autoload_copy_sdma_ucode(struct amdgpu_device *adev) 1261 { 1262 const __le32 *fw_data; 1263 uint32_t fw_size; 1264 const struct sdma_firmware_header_v3_0 *sdma_hdr; 1265 1266 sdma_hdr = (const struct sdma_firmware_header_v3_0 *) 1267 adev->sdma.instance[0].fw->data; 1268 fw_data = (const __le32 *) (adev->sdma.instance[0].fw->data + 1269 le32_to_cpu(sdma_hdr->ucode_offset_bytes)); 1270 fw_size = le32_to_cpu(sdma_hdr->ucode_size_bytes); 1271 1272 gfx_v12_0_rlc_backdoor_autoload_copy_ucode(adev, SOC24_FIRMWARE_ID_SDMA_UCODE_TH0, 1273 fw_data, fw_size); 1274 } 1275 1276 static void 1277 gfx_v12_0_rlc_backdoor_autoload_copy_mes_ucode(struct amdgpu_device *adev) 1278 { 1279 const __le32 *fw_data; 1280 unsigned fw_size; 1281 const struct mes_firmware_header_v1_0 *mes_hdr; 1282 int pipe, ucode_id, data_id; 1283 1284 for (pipe = 0; pipe < 2; pipe++) { 1285 if (pipe == 0) { 1286 ucode_id = SOC24_FIRMWARE_ID_RS64_MES_P0; 1287 data_id = SOC24_FIRMWARE_ID_RS64_MES_P0_STACK; 1288 } else { 1289 ucode_id = SOC24_FIRMWARE_ID_RS64_MES_P1; 1290 data_id = SOC24_FIRMWARE_ID_RS64_MES_P1_STACK; 1291 } 1292 1293 mes_hdr = (const struct mes_firmware_header_v1_0 *) 1294 adev->mes.fw[pipe]->data; 1295 1296 fw_data = (const __le32 *)(adev->mes.fw[pipe]->data + 1297 le32_to_cpu(mes_hdr->mes_ucode_offset_bytes)); 1298 fw_size = le32_to_cpu(mes_hdr->mes_ucode_size_bytes); 1299 1300 gfx_v12_0_rlc_backdoor_autoload_copy_ucode(adev, ucode_id, fw_data, fw_size); 1301 1302 fw_data = (const __le32 *)(adev->mes.fw[pipe]->data + 1303 le32_to_cpu(mes_hdr->mes_ucode_data_offset_bytes)); 1304 fw_size = le32_to_cpu(mes_hdr->mes_ucode_data_size_bytes); 1305 1306 gfx_v12_0_rlc_backdoor_autoload_copy_ucode(adev, data_id, fw_data, fw_size); 1307 } 1308 } 1309 1310 static int gfx_v12_0_rlc_backdoor_autoload_enable(struct amdgpu_device *adev) 1311 { 1312 uint32_t rlc_g_offset, rlc_g_size; 1313 uint64_t gpu_addr; 1314 uint32_t data; 1315 1316 /* RLC autoload sequence 2: copy ucode */ 1317 gfx_v12_0_rlc_backdoor_autoload_copy_sdma_ucode(adev); 1318 gfx_v12_0_rlc_backdoor_autoload_copy_gfx_ucode(adev); 1319 gfx_v12_0_rlc_backdoor_autoload_copy_mes_ucode(adev); 1320 gfx_v12_0_rlc_backdoor_autoload_copy_toc_ucode(adev); 1321 1322 rlc_g_offset = rlc_autoload_info[SOC24_FIRMWARE_ID_RLC_G_UCODE].offset; 1323 rlc_g_size = rlc_autoload_info[SOC24_FIRMWARE_ID_RLC_G_UCODE].size; 1324 gpu_addr = adev->gfx.rlc.rlc_autoload_gpu_addr + rlc_g_offset - adev->gmc.vram_start; 1325 1326 WREG32_SOC15(GC, 0, regGFX_IMU_RLC_BOOTLOADER_ADDR_HI, upper_32_bits(gpu_addr)); 1327 WREG32_SOC15(GC, 0, regGFX_IMU_RLC_BOOTLOADER_ADDR_LO, lower_32_bits(gpu_addr)); 1328 1329 WREG32_SOC15(GC, 0, regGFX_IMU_RLC_BOOTLOADER_SIZE, rlc_g_size); 1330 1331 if (adev->gfx.imu.funcs && (amdgpu_dpm > 0)) { 1332 /* RLC autoload sequence 3: load IMU fw */ 1333 if (adev->gfx.imu.funcs->load_microcode) 1334 adev->gfx.imu.funcs->load_microcode(adev); 1335 /* RLC autoload sequence 4 init IMU fw */ 1336 if (adev->gfx.imu.funcs->setup_imu) 1337 adev->gfx.imu.funcs->setup_imu(adev); 1338 if (adev->gfx.imu.funcs->start_imu) 1339 adev->gfx.imu.funcs->start_imu(adev); 1340 1341 /* RLC autoload sequence 5 disable gpa mode */ 1342 gfx_v12_0_disable_gpa_mode(adev); 1343 } else { 1344 /* unhalt rlc to start autoload without imu */ 1345 data = RREG32_SOC15(GC, 0, regRLC_GPM_THREAD_ENABLE); 1346 data = REG_SET_FIELD(data, RLC_GPM_THREAD_ENABLE, THREAD0_ENABLE, 1); 1347 data = REG_SET_FIELD(data, RLC_GPM_THREAD_ENABLE, THREAD1_ENABLE, 1); 1348 WREG32_SOC15(GC, 0, regRLC_GPM_THREAD_ENABLE, data); 1349 WREG32_SOC15(GC, 0, regRLC_CNTL, RLC_CNTL__RLC_ENABLE_F32_MASK); 1350 } 1351 1352 return 0; 1353 } 1354 1355 static void gfx_v12_0_alloc_ip_dump(struct amdgpu_device *adev) 1356 { 1357 uint32_t reg_count = ARRAY_SIZE(gc_reg_list_12_0); 1358 uint32_t *ptr; 1359 uint32_t inst; 1360 1361 ptr = kcalloc(reg_count, sizeof(uint32_t), GFP_KERNEL); 1362 if (!ptr) { 1363 DRM_ERROR("Failed to allocate memory for GFX IP Dump\n"); 1364 adev->gfx.ip_dump_core = NULL; 1365 } else { 1366 adev->gfx.ip_dump_core = ptr; 1367 } 1368 1369 /* Allocate memory for compute queue registers for all the instances */ 1370 reg_count = ARRAY_SIZE(gc_cp_reg_list_12); 1371 inst = adev->gfx.mec.num_mec * adev->gfx.mec.num_pipe_per_mec * 1372 adev->gfx.mec.num_queue_per_pipe; 1373 1374 ptr = kcalloc(reg_count * inst, sizeof(uint32_t), GFP_KERNEL); 1375 if (!ptr) { 1376 DRM_ERROR("Failed to allocate memory for Compute Queues IP Dump\n"); 1377 adev->gfx.ip_dump_compute_queues = NULL; 1378 } else { 1379 adev->gfx.ip_dump_compute_queues = ptr; 1380 } 1381 1382 /* Allocate memory for gfx queue registers for all the instances */ 1383 reg_count = ARRAY_SIZE(gc_gfx_queue_reg_list_12); 1384 inst = adev->gfx.me.num_me * adev->gfx.me.num_pipe_per_me * 1385 adev->gfx.me.num_queue_per_pipe; 1386 1387 ptr = kcalloc(reg_count * inst, sizeof(uint32_t), GFP_KERNEL); 1388 if (!ptr) { 1389 DRM_ERROR("Failed to allocate memory for GFX Queues IP Dump\n"); 1390 adev->gfx.ip_dump_gfx_queues = NULL; 1391 } else { 1392 adev->gfx.ip_dump_gfx_queues = ptr; 1393 } 1394 } 1395 1396 static int gfx_v12_0_sw_init(struct amdgpu_ip_block *ip_block) 1397 { 1398 int i, j, k, r, ring_id = 0; 1399 unsigned num_compute_rings; 1400 int xcc_id = 0; 1401 struct amdgpu_device *adev = ip_block->adev; 1402 int num_queue_per_pipe = 1; /* we only enable 1 KGQ per pipe */ 1403 1404 INIT_DELAYED_WORK(&adev->gfx.idle_work, amdgpu_gfx_profile_idle_work_handler); 1405 1406 switch (amdgpu_ip_version(adev, GC_HWIP, 0)) { 1407 case IP_VERSION(12, 0, 0): 1408 case IP_VERSION(12, 0, 1): 1409 adev->gfx.me.num_me = 1; 1410 adev->gfx.me.num_pipe_per_me = 1; 1411 adev->gfx.me.num_queue_per_pipe = 8; 1412 adev->gfx.mec.num_mec = 1; 1413 adev->gfx.mec.num_pipe_per_mec = 2; 1414 adev->gfx.mec.num_queue_per_pipe = 4; 1415 break; 1416 default: 1417 adev->gfx.me.num_me = 1; 1418 adev->gfx.me.num_pipe_per_me = 1; 1419 adev->gfx.me.num_queue_per_pipe = 1; 1420 adev->gfx.mec.num_mec = 1; 1421 adev->gfx.mec.num_pipe_per_mec = 4; 1422 adev->gfx.mec.num_queue_per_pipe = 8; 1423 break; 1424 } 1425 1426 switch (amdgpu_ip_version(adev, GC_HWIP, 0)) { 1427 case IP_VERSION(12, 0, 0): 1428 case IP_VERSION(12, 0, 1): 1429 if (!adev->gfx.disable_uq && 1430 adev->gfx.me_fw_version >= 2780 && 1431 adev->gfx.pfp_fw_version >= 2840 && 1432 adev->gfx.mec_fw_version >= 3050 && 1433 adev->mes.fw_version[0] >= 123) { 1434 adev->userq_funcs[AMDGPU_HW_IP_GFX] = &userq_mes_funcs; 1435 adev->userq_funcs[AMDGPU_HW_IP_COMPUTE] = &userq_mes_funcs; 1436 } 1437 break; 1438 default: 1439 break; 1440 } 1441 1442 switch (amdgpu_ip_version(adev, GC_HWIP, 0)) { 1443 case IP_VERSION(12, 0, 0): 1444 case IP_VERSION(12, 0, 1): 1445 if (adev->gfx.me_fw_version >= 2480 && 1446 adev->gfx.pfp_fw_version >= 2530 && 1447 adev->gfx.mec_fw_version >= 2680 && 1448 adev->mes.fw_version[0] >= 100) 1449 adev->gfx.enable_cleaner_shader = true; 1450 break; 1451 default: 1452 adev->gfx.enable_cleaner_shader = false; 1453 break; 1454 } 1455 1456 if (adev->gfx.num_compute_rings) { 1457 /* recalculate compute rings to use based on hardware configuration */ 1458 num_compute_rings = (adev->gfx.mec.num_pipe_per_mec * 1459 adev->gfx.mec.num_queue_per_pipe) / 2; 1460 adev->gfx.num_compute_rings = min(adev->gfx.num_compute_rings, 1461 num_compute_rings); 1462 } 1463 1464 /* EOP Event */ 1465 r = amdgpu_irq_add_id(adev, SOC21_IH_CLIENTID_GRBM_CP, 1466 GFX_12_0_0__SRCID__CP_EOP_INTERRUPT, 1467 &adev->gfx.eop_irq); 1468 if (r) 1469 return r; 1470 1471 /* Bad opcode Event */ 1472 r = amdgpu_irq_add_id(adev, SOC21_IH_CLIENTID_GRBM_CP, 1473 GFX_12_0_0__SRCID__CP_BAD_OPCODE_ERROR, 1474 &adev->gfx.bad_op_irq); 1475 if (r) 1476 return r; 1477 1478 /* Privileged reg */ 1479 r = amdgpu_irq_add_id(adev, SOC21_IH_CLIENTID_GRBM_CP, 1480 GFX_12_0_0__SRCID__CP_PRIV_REG_FAULT, 1481 &adev->gfx.priv_reg_irq); 1482 if (r) 1483 return r; 1484 1485 /* Privileged inst */ 1486 r = amdgpu_irq_add_id(adev, SOC21_IH_CLIENTID_GRBM_CP, 1487 GFX_12_0_0__SRCID__CP_PRIV_INSTR_FAULT, 1488 &adev->gfx.priv_inst_irq); 1489 if (r) 1490 return r; 1491 1492 adev->gfx.gfx_current_status = AMDGPU_GFX_NORMAL_MODE; 1493 1494 gfx_v12_0_me_init(adev); 1495 1496 r = gfx_v12_0_rlc_init(adev); 1497 if (r) { 1498 dev_err(adev->dev, "Failed to init rlc BOs!\n"); 1499 return r; 1500 } 1501 1502 r = gfx_v12_0_mec_init(adev); 1503 if (r) { 1504 dev_err(adev->dev, "Failed to init MEC BOs!\n"); 1505 return r; 1506 } 1507 1508 if (adev->gfx.num_gfx_rings) { 1509 /* set up the gfx ring */ 1510 for (i = 0; i < adev->gfx.me.num_me; i++) { 1511 for (j = 0; j < num_queue_per_pipe; j++) { 1512 for (k = 0; k < adev->gfx.me.num_pipe_per_me; k++) { 1513 if (!amdgpu_gfx_is_me_queue_enabled(adev, i, k, j)) 1514 continue; 1515 1516 r = gfx_v12_0_gfx_ring_init(adev, ring_id, 1517 i, k, j); 1518 if (r) 1519 return r; 1520 ring_id++; 1521 } 1522 } 1523 } 1524 } 1525 1526 if (adev->gfx.num_compute_rings) { 1527 ring_id = 0; 1528 /* set up the compute queues - allocate horizontally across pipes */ 1529 for (i = 0; i < adev->gfx.mec.num_mec; ++i) { 1530 for (j = 0; j < adev->gfx.mec.num_queue_per_pipe; j++) { 1531 for (k = 0; k < adev->gfx.mec.num_pipe_per_mec; k++) { 1532 if (!amdgpu_gfx_is_mec_queue_enabled(adev, 1533 0, i, k, j)) 1534 continue; 1535 1536 r = gfx_v12_0_compute_ring_init(adev, ring_id, 1537 i, k, j); 1538 if (r) 1539 return r; 1540 1541 ring_id++; 1542 } 1543 } 1544 } 1545 } 1546 1547 adev->gfx.gfx_supported_reset = 1548 amdgpu_get_soft_full_reset_mask(&adev->gfx.gfx_ring[0]); 1549 adev->gfx.compute_supported_reset = 1550 amdgpu_get_soft_full_reset_mask(&adev->gfx.compute_ring[0]); 1551 switch (amdgpu_ip_version(adev, GC_HWIP, 0)) { 1552 case IP_VERSION(12, 0, 0): 1553 case IP_VERSION(12, 0, 1): 1554 if ((adev->gfx.me_fw_version >= 2660) && 1555 (adev->gfx.mec_fw_version >= 2920) && 1556 !amdgpu_sriov_vf(adev) && 1557 !adev->debug_disable_gpu_ring_reset) { 1558 adev->gfx.compute_supported_reset |= AMDGPU_RESET_TYPE_PER_QUEUE; 1559 adev->gfx.gfx_supported_reset |= AMDGPU_RESET_TYPE_PER_QUEUE; 1560 } 1561 break; 1562 default: 1563 break; 1564 } 1565 1566 if (!adev->enable_mes_kiq) { 1567 r = amdgpu_gfx_kiq_init(adev, GFX12_MEC_HPD_SIZE, 0); 1568 if (r) { 1569 dev_err(adev->dev, "Failed to init KIQ BOs!\n"); 1570 return r; 1571 } 1572 1573 r = amdgpu_gfx_kiq_init_ring(adev, xcc_id); 1574 if (r) 1575 return r; 1576 } 1577 1578 r = amdgpu_gfx_mqd_sw_init(adev, sizeof(struct v12_compute_mqd), 0); 1579 if (r) 1580 return r; 1581 1582 /* allocate visible FB for rlc auto-loading fw */ 1583 if (adev->firmware.load_type == AMDGPU_FW_LOAD_RLC_BACKDOOR_AUTO) { 1584 r = gfx_v12_0_rlc_autoload_buffer_init(adev); 1585 if (r) 1586 return r; 1587 } 1588 1589 r = gfx_v12_0_gpu_early_init(adev); 1590 if (r) 1591 return r; 1592 1593 gfx_v12_0_alloc_ip_dump(adev); 1594 1595 r = amdgpu_gfx_sysfs_init(adev); 1596 if (r) 1597 return r; 1598 1599 return 0; 1600 } 1601 1602 static void gfx_v12_0_pfp_fini(struct amdgpu_device *adev) 1603 { 1604 amdgpu_bo_free_kernel(&adev->gfx.pfp.pfp_fw_obj, 1605 &adev->gfx.pfp.pfp_fw_gpu_addr, 1606 (void **)&adev->gfx.pfp.pfp_fw_ptr); 1607 1608 amdgpu_bo_free_kernel(&adev->gfx.pfp.pfp_fw_data_obj, 1609 &adev->gfx.pfp.pfp_fw_data_gpu_addr, 1610 (void **)&adev->gfx.pfp.pfp_fw_data_ptr); 1611 } 1612 1613 static void gfx_v12_0_me_fini(struct amdgpu_device *adev) 1614 { 1615 amdgpu_bo_free_kernel(&adev->gfx.me.me_fw_obj, 1616 &adev->gfx.me.me_fw_gpu_addr, 1617 (void **)&adev->gfx.me.me_fw_ptr); 1618 1619 amdgpu_bo_free_kernel(&adev->gfx.me.me_fw_data_obj, 1620 &adev->gfx.me.me_fw_data_gpu_addr, 1621 (void **)&adev->gfx.me.me_fw_data_ptr); 1622 } 1623 1624 static void gfx_v12_0_rlc_autoload_buffer_fini(struct amdgpu_device *adev) 1625 { 1626 amdgpu_bo_free_kernel(&adev->gfx.rlc.rlc_autoload_bo, 1627 &adev->gfx.rlc.rlc_autoload_gpu_addr, 1628 (void **)&adev->gfx.rlc.rlc_autoload_ptr); 1629 } 1630 1631 static int gfx_v12_0_sw_fini(struct amdgpu_ip_block *ip_block) 1632 { 1633 int i; 1634 struct amdgpu_device *adev = ip_block->adev; 1635 1636 for (i = 0; i < adev->gfx.num_gfx_rings; i++) 1637 amdgpu_ring_fini(&adev->gfx.gfx_ring[i]); 1638 for (i = 0; i < adev->gfx.num_compute_rings; i++) 1639 amdgpu_ring_fini(&adev->gfx.compute_ring[i]); 1640 1641 amdgpu_gfx_mqd_sw_fini(adev, 0); 1642 1643 if (!adev->enable_mes_kiq) { 1644 amdgpu_gfx_kiq_free_ring(&adev->gfx.kiq[0].ring); 1645 amdgpu_gfx_kiq_fini(adev, 0); 1646 } 1647 1648 gfx_v12_0_pfp_fini(adev); 1649 gfx_v12_0_me_fini(adev); 1650 gfx_v12_0_rlc_fini(adev); 1651 gfx_v12_0_mec_fini(adev); 1652 1653 if (adev->firmware.load_type == AMDGPU_FW_LOAD_RLC_BACKDOOR_AUTO) 1654 gfx_v12_0_rlc_autoload_buffer_fini(adev); 1655 1656 gfx_v12_0_free_microcode(adev); 1657 1658 amdgpu_gfx_sysfs_fini(adev); 1659 1660 kfree(adev->gfx.ip_dump_core); 1661 kfree(adev->gfx.ip_dump_compute_queues); 1662 kfree(adev->gfx.ip_dump_gfx_queues); 1663 1664 return 0; 1665 } 1666 1667 static void gfx_v12_0_select_se_sh(struct amdgpu_device *adev, u32 se_num, 1668 u32 sh_num, u32 instance, int xcc_id) 1669 { 1670 u32 data; 1671 1672 if (instance == 0xffffffff) 1673 data = REG_SET_FIELD(0, GRBM_GFX_INDEX, 1674 INSTANCE_BROADCAST_WRITES, 1); 1675 else 1676 data = REG_SET_FIELD(0, GRBM_GFX_INDEX, INSTANCE_INDEX, 1677 instance); 1678 1679 if (se_num == 0xffffffff) 1680 data = REG_SET_FIELD(data, GRBM_GFX_INDEX, SE_BROADCAST_WRITES, 1681 1); 1682 else 1683 data = REG_SET_FIELD(data, GRBM_GFX_INDEX, SE_INDEX, se_num); 1684 1685 if (sh_num == 0xffffffff) 1686 data = REG_SET_FIELD(data, GRBM_GFX_INDEX, SA_BROADCAST_WRITES, 1687 1); 1688 else 1689 data = REG_SET_FIELD(data, GRBM_GFX_INDEX, SA_INDEX, sh_num); 1690 1691 WREG32_SOC15(GC, 0, regGRBM_GFX_INDEX, data); 1692 } 1693 1694 static u32 gfx_v12_0_get_sa_active_bitmap(struct amdgpu_device *adev) 1695 { 1696 u32 gc_disabled_sa_mask, gc_user_disabled_sa_mask, sa_mask; 1697 1698 gc_disabled_sa_mask = RREG32_SOC15(GC, 0, regGRBM_CC_GC_SA_UNIT_DISABLE); 1699 gc_disabled_sa_mask = REG_GET_FIELD(gc_disabled_sa_mask, 1700 GRBM_CC_GC_SA_UNIT_DISABLE, 1701 SA_DISABLE); 1702 gc_user_disabled_sa_mask = RREG32_SOC15(GC, 0, regGRBM_GC_USER_SA_UNIT_DISABLE); 1703 gc_user_disabled_sa_mask = REG_GET_FIELD(gc_user_disabled_sa_mask, 1704 GRBM_GC_USER_SA_UNIT_DISABLE, 1705 SA_DISABLE); 1706 sa_mask = amdgpu_gfx_create_bitmask(adev->gfx.config.max_sh_per_se * 1707 adev->gfx.config.max_shader_engines); 1708 1709 return sa_mask & (~(gc_disabled_sa_mask | gc_user_disabled_sa_mask)); 1710 } 1711 1712 static u32 gfx_v12_0_get_rb_active_bitmap(struct amdgpu_device *adev) 1713 { 1714 u32 gc_disabled_rb_mask, gc_user_disabled_rb_mask; 1715 u32 rb_mask; 1716 1717 gc_disabled_rb_mask = RREG32_SOC15(GC, 0, regCC_RB_BACKEND_DISABLE); 1718 gc_disabled_rb_mask = REG_GET_FIELD(gc_disabled_rb_mask, 1719 CC_RB_BACKEND_DISABLE, 1720 BACKEND_DISABLE); 1721 gc_user_disabled_rb_mask = RREG32_SOC15(GC, 0, regGC_USER_RB_BACKEND_DISABLE); 1722 gc_user_disabled_rb_mask = REG_GET_FIELD(gc_user_disabled_rb_mask, 1723 GC_USER_RB_BACKEND_DISABLE, 1724 BACKEND_DISABLE); 1725 rb_mask = amdgpu_gfx_create_bitmask(adev->gfx.config.max_backends_per_se * 1726 adev->gfx.config.max_shader_engines); 1727 1728 return rb_mask & (~(gc_disabled_rb_mask | gc_user_disabled_rb_mask)); 1729 } 1730 1731 static void gfx_v12_0_setup_rb(struct amdgpu_device *adev) 1732 { 1733 u32 rb_bitmap_per_sa; 1734 u32 rb_bitmap_width_per_sa; 1735 u32 max_sa; 1736 u32 active_sa_bitmap; 1737 u32 global_active_rb_bitmap; 1738 u32 active_rb_bitmap = 0; 1739 u32 i; 1740 1741 /* query sa bitmap from SA_UNIT_DISABLE registers */ 1742 active_sa_bitmap = gfx_v12_0_get_sa_active_bitmap(adev); 1743 /* query rb bitmap from RB_BACKEND_DISABLE registers */ 1744 global_active_rb_bitmap = gfx_v12_0_get_rb_active_bitmap(adev); 1745 1746 /* generate active rb bitmap according to active sa bitmap */ 1747 max_sa = adev->gfx.config.max_shader_engines * 1748 adev->gfx.config.max_sh_per_se; 1749 rb_bitmap_width_per_sa = adev->gfx.config.max_backends_per_se / 1750 adev->gfx.config.max_sh_per_se; 1751 rb_bitmap_per_sa = amdgpu_gfx_create_bitmask(rb_bitmap_width_per_sa); 1752 1753 for (i = 0; i < max_sa; i++) { 1754 if (active_sa_bitmap & (1 << i)) 1755 active_rb_bitmap |= (rb_bitmap_per_sa << (i * rb_bitmap_width_per_sa)); 1756 } 1757 1758 active_rb_bitmap &= global_active_rb_bitmap; 1759 adev->gfx.config.backend_enable_mask = active_rb_bitmap; 1760 adev->gfx.config.num_rbs = hweight32(active_rb_bitmap); 1761 } 1762 1763 #define LDS_APP_BASE 0x1 1764 #define SCRATCH_APP_BASE 0x2 1765 1766 static void gfx_v12_0_init_compute_vmid(struct amdgpu_device *adev) 1767 { 1768 int i; 1769 uint32_t sh_mem_bases; 1770 uint32_t data; 1771 1772 /* 1773 * Configure apertures: 1774 * LDS: 0x60000000'00000000 - 0x60000001'00000000 (4GB) 1775 * Scratch: 0x60000001'00000000 - 0x60000002'00000000 (4GB) 1776 * GPUVM: 0x60010000'00000000 - 0x60020000'00000000 (1TB) 1777 */ 1778 sh_mem_bases = (LDS_APP_BASE << SH_MEM_BASES__SHARED_BASE__SHIFT) | 1779 SCRATCH_APP_BASE; 1780 1781 mutex_lock(&adev->srbm_mutex); 1782 for (i = adev->vm_manager.first_kfd_vmid; i < AMDGPU_NUM_VMID; i++) { 1783 soc24_grbm_select(adev, 0, 0, 0, i); 1784 /* CP and shaders */ 1785 WREG32_SOC15(GC, 0, regSH_MEM_CONFIG, DEFAULT_SH_MEM_CONFIG); 1786 WREG32_SOC15(GC, 0, regSH_MEM_BASES, sh_mem_bases); 1787 1788 /* Enable trap for each kfd vmid. */ 1789 data = RREG32_SOC15(GC, 0, regSPI_GDBG_PER_VMID_CNTL); 1790 data = REG_SET_FIELD(data, SPI_GDBG_PER_VMID_CNTL, TRAP_EN, 1); 1791 WREG32_SOC15(GC, 0, regSPI_GDBG_PER_VMID_CNTL, data); 1792 } 1793 soc24_grbm_select(adev, 0, 0, 0, 0); 1794 mutex_unlock(&adev->srbm_mutex); 1795 } 1796 1797 static void gfx_v12_0_tcp_harvest(struct amdgpu_device *adev) 1798 { 1799 /* TODO: harvest feature to be added later. */ 1800 } 1801 1802 static void gfx_v12_0_get_tcc_info(struct amdgpu_device *adev) 1803 { 1804 } 1805 1806 static void gfx_v12_0_constants_init(struct amdgpu_device *adev) 1807 { 1808 u32 tmp; 1809 int i; 1810 1811 if (!amdgpu_sriov_vf(adev)) 1812 WREG32_FIELD15_PREREG(GC, 0, GRBM_CNTL, READ_TIMEOUT, 0xff); 1813 1814 gfx_v12_0_setup_rb(adev); 1815 gfx_v12_0_get_cu_info(adev, &adev->gfx.cu_info); 1816 gfx_v12_0_get_tcc_info(adev); 1817 adev->gfx.config.pa_sc_tile_steering_override = 0; 1818 1819 /* XXX SH_MEM regs */ 1820 /* where to put LDS, scratch, GPUVM in FSA64 space */ 1821 mutex_lock(&adev->srbm_mutex); 1822 for (i = 0; i < adev->vm_manager.id_mgr[AMDGPU_GFXHUB(0)].num_ids; i++) { 1823 soc24_grbm_select(adev, 0, 0, 0, i); 1824 /* CP and shaders */ 1825 WREG32_SOC15(GC, 0, regSH_MEM_CONFIG, DEFAULT_SH_MEM_CONFIG); 1826 if (i != 0) { 1827 tmp = REG_SET_FIELD(0, SH_MEM_BASES, PRIVATE_BASE, 1828 (adev->gmc.private_aperture_start >> 48)); 1829 tmp = REG_SET_FIELD(tmp, SH_MEM_BASES, SHARED_BASE, 1830 (adev->gmc.shared_aperture_start >> 48)); 1831 WREG32_SOC15(GC, 0, regSH_MEM_BASES, tmp); 1832 } 1833 } 1834 soc24_grbm_select(adev, 0, 0, 0, 0); 1835 1836 mutex_unlock(&adev->srbm_mutex); 1837 1838 gfx_v12_0_init_compute_vmid(adev); 1839 } 1840 1841 static u32 gfx_v12_0_get_cpg_int_cntl(struct amdgpu_device *adev, 1842 int me, int pipe) 1843 { 1844 if (me != 0) 1845 return 0; 1846 1847 switch (pipe) { 1848 case 0: 1849 return SOC15_REG_OFFSET(GC, 0, regCP_INT_CNTL_RING0); 1850 default: 1851 return 0; 1852 } 1853 } 1854 1855 static u32 gfx_v12_0_get_cpc_int_cntl(struct amdgpu_device *adev, 1856 int me, int pipe) 1857 { 1858 /* 1859 * amdgpu controls only the first MEC. That's why this function only 1860 * handles the setting of interrupts for this specific MEC. All other 1861 * pipes' interrupts are set by amdkfd. 1862 */ 1863 if (me != 1) 1864 return 0; 1865 1866 switch (pipe) { 1867 case 0: 1868 return SOC15_REG_OFFSET(GC, 0, regCP_ME1_PIPE0_INT_CNTL); 1869 case 1: 1870 return SOC15_REG_OFFSET(GC, 0, regCP_ME1_PIPE1_INT_CNTL); 1871 default: 1872 return 0; 1873 } 1874 } 1875 1876 static void gfx_v12_0_enable_gui_idle_interrupt(struct amdgpu_device *adev, 1877 bool enable) 1878 { 1879 u32 tmp, cp_int_cntl_reg; 1880 int i, j; 1881 1882 if (amdgpu_sriov_vf(adev)) 1883 return; 1884 1885 for (i = 0; i < adev->gfx.me.num_me; i++) { 1886 for (j = 0; j < adev->gfx.me.num_pipe_per_me; j++) { 1887 cp_int_cntl_reg = gfx_v12_0_get_cpg_int_cntl(adev, i, j); 1888 1889 if (cp_int_cntl_reg) { 1890 tmp = RREG32_SOC15_IP(GC, cp_int_cntl_reg); 1891 tmp = REG_SET_FIELD(tmp, CP_INT_CNTL_RING0, CNTX_BUSY_INT_ENABLE, 1892 enable ? 1 : 0); 1893 tmp = REG_SET_FIELD(tmp, CP_INT_CNTL_RING0, CNTX_EMPTY_INT_ENABLE, 1894 enable ? 1 : 0); 1895 tmp = REG_SET_FIELD(tmp, CP_INT_CNTL_RING0, CMP_BUSY_INT_ENABLE, 1896 enable ? 1 : 0); 1897 tmp = REG_SET_FIELD(tmp, CP_INT_CNTL_RING0, GFX_IDLE_INT_ENABLE, 1898 enable ? 1 : 0); 1899 WREG32_SOC15_IP(GC, cp_int_cntl_reg, tmp); 1900 } 1901 } 1902 } 1903 } 1904 1905 static int gfx_v12_0_init_csb(struct amdgpu_device *adev) 1906 { 1907 adev->gfx.rlc.funcs->get_csb_buffer(adev, adev->gfx.rlc.cs_ptr); 1908 1909 WREG32_SOC15(GC, 0, regRLC_CSIB_ADDR_HI, 1910 adev->gfx.rlc.clear_state_gpu_addr >> 32); 1911 WREG32_SOC15(GC, 0, regRLC_CSIB_ADDR_LO, 1912 adev->gfx.rlc.clear_state_gpu_addr & 0xfffffffc); 1913 WREG32_SOC15(GC, 0, regRLC_CSIB_LENGTH, adev->gfx.rlc.clear_state_size); 1914 1915 return 0; 1916 } 1917 1918 static void gfx_v12_0_rlc_stop(struct amdgpu_device *adev) 1919 { 1920 u32 tmp = RREG32_SOC15(GC, 0, regRLC_CNTL); 1921 1922 tmp = REG_SET_FIELD(tmp, RLC_CNTL, RLC_ENABLE_F32, 0); 1923 WREG32_SOC15(GC, 0, regRLC_CNTL, tmp); 1924 } 1925 1926 static void gfx_v12_0_rlc_reset(struct amdgpu_device *adev) 1927 { 1928 WREG32_FIELD15_PREREG(GC, 0, GRBM_SOFT_RESET, SOFT_RESET_RLC, 1); 1929 udelay(50); 1930 WREG32_FIELD15_PREREG(GC, 0, GRBM_SOFT_RESET, SOFT_RESET_RLC, 0); 1931 udelay(50); 1932 } 1933 1934 static void gfx_v12_0_rlc_smu_handshake_cntl(struct amdgpu_device *adev, 1935 bool enable) 1936 { 1937 uint32_t rlc_pg_cntl; 1938 1939 rlc_pg_cntl = RREG32_SOC15(GC, 0, regRLC_PG_CNTL); 1940 1941 if (!enable) { 1942 /* RLC_PG_CNTL[23] = 0 (default) 1943 * RLC will wait for handshake acks with SMU 1944 * GFXOFF will be enabled 1945 * RLC_PG_CNTL[23] = 1 1946 * RLC will not issue any message to SMU 1947 * hence no handshake between SMU & RLC 1948 * GFXOFF will be disabled 1949 */ 1950 rlc_pg_cntl |= RLC_PG_CNTL__SMU_HANDSHAKE_DISABLE_MASK; 1951 } else 1952 rlc_pg_cntl &= ~RLC_PG_CNTL__SMU_HANDSHAKE_DISABLE_MASK; 1953 WREG32_SOC15(GC, 0, regRLC_PG_CNTL, rlc_pg_cntl); 1954 } 1955 1956 static void gfx_v12_0_rlc_start(struct amdgpu_device *adev) 1957 { 1958 /* TODO: enable rlc & smu handshake until smu 1959 * and gfxoff feature works as expected */ 1960 if (!(amdgpu_pp_feature_mask & PP_GFXOFF_MASK)) 1961 gfx_v12_0_rlc_smu_handshake_cntl(adev, false); 1962 1963 WREG32_FIELD15_PREREG(GC, 0, RLC_CNTL, RLC_ENABLE_F32, 1); 1964 udelay(50); 1965 } 1966 1967 static void gfx_v12_0_rlc_enable_srm(struct amdgpu_device *adev) 1968 { 1969 uint32_t tmp; 1970 1971 /* enable Save Restore Machine */ 1972 tmp = RREG32(SOC15_REG_OFFSET(GC, 0, regRLC_SRM_CNTL)); 1973 tmp |= RLC_SRM_CNTL__AUTO_INCR_ADDR_MASK; 1974 tmp |= RLC_SRM_CNTL__SRM_ENABLE_MASK; 1975 WREG32(SOC15_REG_OFFSET(GC, 0, regRLC_SRM_CNTL), tmp); 1976 } 1977 1978 static void gfx_v12_0_load_rlcg_microcode(struct amdgpu_device *adev) 1979 { 1980 const struct rlc_firmware_header_v2_0 *hdr; 1981 const __le32 *fw_data; 1982 unsigned i, fw_size; 1983 1984 hdr = (const struct rlc_firmware_header_v2_0 *)adev->gfx.rlc_fw->data; 1985 fw_data = (const __le32 *)(adev->gfx.rlc_fw->data + 1986 le32_to_cpu(hdr->header.ucode_array_offset_bytes)); 1987 fw_size = le32_to_cpu(hdr->header.ucode_size_bytes) / 4; 1988 1989 WREG32_SOC15(GC, 0, regRLC_GPM_UCODE_ADDR, 1990 RLCG_UCODE_LOADING_START_ADDRESS); 1991 1992 for (i = 0; i < fw_size; i++) 1993 WREG32_SOC15(GC, 0, regRLC_GPM_UCODE_DATA, 1994 le32_to_cpup(fw_data++)); 1995 1996 WREG32_SOC15(GC, 0, regRLC_GPM_UCODE_ADDR, adev->gfx.rlc_fw_version); 1997 } 1998 1999 static void gfx_v12_0_load_rlc_iram_dram_microcode(struct amdgpu_device *adev) 2000 { 2001 const struct rlc_firmware_header_v2_2 *hdr; 2002 const __le32 *fw_data; 2003 unsigned i, fw_size; 2004 u32 tmp; 2005 2006 hdr = (const struct rlc_firmware_header_v2_2 *)adev->gfx.rlc_fw->data; 2007 2008 fw_data = (const __le32 *)(adev->gfx.rlc_fw->data + 2009 le32_to_cpu(hdr->rlc_iram_ucode_offset_bytes)); 2010 fw_size = le32_to_cpu(hdr->rlc_iram_ucode_size_bytes) / 4; 2011 2012 WREG32_SOC15(GC, 0, regRLC_LX6_IRAM_ADDR, 0); 2013 2014 for (i = 0; i < fw_size; i++) { 2015 if ((amdgpu_emu_mode == 1) && (i % 100 == 99)) 2016 msleep(1); 2017 WREG32_SOC15(GC, 0, regRLC_LX6_IRAM_DATA, 2018 le32_to_cpup(fw_data++)); 2019 } 2020 2021 WREG32_SOC15(GC, 0, regRLC_LX6_IRAM_ADDR, adev->gfx.rlc_fw_version); 2022 2023 fw_data = (const __le32 *)(adev->gfx.rlc_fw->data + 2024 le32_to_cpu(hdr->rlc_dram_ucode_offset_bytes)); 2025 fw_size = le32_to_cpu(hdr->rlc_dram_ucode_size_bytes) / 4; 2026 2027 WREG32_SOC15(GC, 0, regRLC_LX6_DRAM_ADDR, 0); 2028 for (i = 0; i < fw_size; i++) { 2029 if ((amdgpu_emu_mode == 1) && (i % 100 == 99)) 2030 msleep(1); 2031 WREG32_SOC15(GC, 0, regRLC_LX6_DRAM_DATA, 2032 le32_to_cpup(fw_data++)); 2033 } 2034 2035 WREG32_SOC15(GC, 0, regRLC_LX6_IRAM_ADDR, adev->gfx.rlc_fw_version); 2036 2037 tmp = RREG32_SOC15(GC, 0, regRLC_LX6_CNTL); 2038 tmp = REG_SET_FIELD(tmp, RLC_LX6_CNTL, PDEBUG_ENABLE, 1); 2039 tmp = REG_SET_FIELD(tmp, RLC_LX6_CNTL, BRESET, 0); 2040 WREG32_SOC15(GC, 0, regRLC_LX6_CNTL, tmp); 2041 } 2042 2043 static int gfx_v12_0_rlc_load_microcode(struct amdgpu_device *adev) 2044 { 2045 const struct rlc_firmware_header_v2_0 *hdr; 2046 uint16_t version_major; 2047 uint16_t version_minor; 2048 2049 if (!adev->gfx.rlc_fw) 2050 return -EINVAL; 2051 2052 hdr = (const struct rlc_firmware_header_v2_0 *)adev->gfx.rlc_fw->data; 2053 amdgpu_ucode_print_rlc_hdr(&hdr->header); 2054 2055 version_major = le16_to_cpu(hdr->header.header_version_major); 2056 version_minor = le16_to_cpu(hdr->header.header_version_minor); 2057 2058 if (version_major == 2) { 2059 gfx_v12_0_load_rlcg_microcode(adev); 2060 if (amdgpu_dpm == 1) { 2061 if (version_minor >= 2) 2062 gfx_v12_0_load_rlc_iram_dram_microcode(adev); 2063 } 2064 2065 return 0; 2066 } 2067 2068 return -EINVAL; 2069 } 2070 2071 static int gfx_v12_0_rlc_resume(struct amdgpu_device *adev) 2072 { 2073 int r; 2074 2075 if (adev->firmware.load_type == AMDGPU_FW_LOAD_PSP) { 2076 gfx_v12_0_init_csb(adev); 2077 2078 if (!amdgpu_sriov_vf(adev)) /* enable RLC SRM */ 2079 gfx_v12_0_rlc_enable_srm(adev); 2080 } else { 2081 if (amdgpu_sriov_vf(adev)) { 2082 gfx_v12_0_init_csb(adev); 2083 return 0; 2084 } 2085 2086 adev->gfx.rlc.funcs->stop(adev); 2087 2088 /* disable CG */ 2089 WREG32_SOC15(GC, 0, regRLC_CGCG_CGLS_CTRL, 0); 2090 2091 /* disable PG */ 2092 WREG32_SOC15(GC, 0, regRLC_PG_CNTL, 0); 2093 2094 if (adev->firmware.load_type == AMDGPU_FW_LOAD_DIRECT) { 2095 /* legacy rlc firmware loading */ 2096 r = gfx_v12_0_rlc_load_microcode(adev); 2097 if (r) 2098 return r; 2099 } 2100 2101 gfx_v12_0_init_csb(adev); 2102 2103 adev->gfx.rlc.funcs->start(adev); 2104 } 2105 2106 return 0; 2107 } 2108 2109 static void gfx_v12_0_config_gfx_rs64(struct amdgpu_device *adev) 2110 { 2111 const struct gfx_firmware_header_v2_0 *pfp_hdr; 2112 const struct gfx_firmware_header_v2_0 *me_hdr; 2113 const struct gfx_firmware_header_v2_0 *mec_hdr; 2114 uint32_t pipe_id, tmp; 2115 2116 mec_hdr = (const struct gfx_firmware_header_v2_0 *) 2117 adev->gfx.mec_fw->data; 2118 me_hdr = (const struct gfx_firmware_header_v2_0 *) 2119 adev->gfx.me_fw->data; 2120 pfp_hdr = (const struct gfx_firmware_header_v2_0 *) 2121 adev->gfx.pfp_fw->data; 2122 2123 /* config pfp program start addr */ 2124 for (pipe_id = 0; pipe_id < 2; pipe_id++) { 2125 soc24_grbm_select(adev, 0, pipe_id, 0, 0); 2126 WREG32_SOC15(GC, 0, regCP_PFP_PRGRM_CNTR_START, 2127 (pfp_hdr->ucode_start_addr_hi << 30) | 2128 (pfp_hdr->ucode_start_addr_lo >> 2)); 2129 WREG32_SOC15(GC, 0, regCP_PFP_PRGRM_CNTR_START_HI, 2130 pfp_hdr->ucode_start_addr_hi >> 2); 2131 } 2132 soc24_grbm_select(adev, 0, 0, 0, 0); 2133 2134 /* reset pfp pipe */ 2135 tmp = RREG32_SOC15(GC, 0, regCP_ME_CNTL); 2136 tmp = REG_SET_FIELD(tmp, CP_ME_CNTL, PFP_PIPE0_RESET, 1); 2137 tmp = REG_SET_FIELD(tmp, CP_ME_CNTL, PFP_PIPE1_RESET, 1); 2138 WREG32_SOC15(GC, 0, regCP_ME_CNTL, tmp); 2139 2140 /* clear pfp pipe reset */ 2141 tmp = REG_SET_FIELD(tmp, CP_ME_CNTL, PFP_PIPE0_RESET, 0); 2142 tmp = REG_SET_FIELD(tmp, CP_ME_CNTL, PFP_PIPE1_RESET, 0); 2143 WREG32_SOC15(GC, 0, regCP_ME_CNTL, tmp); 2144 2145 /* config me program start addr */ 2146 for (pipe_id = 0; pipe_id < 2; pipe_id++) { 2147 soc24_grbm_select(adev, 0, pipe_id, 0, 0); 2148 WREG32_SOC15(GC, 0, regCP_ME_PRGRM_CNTR_START, 2149 (me_hdr->ucode_start_addr_hi << 30) | 2150 (me_hdr->ucode_start_addr_lo >> 2)); 2151 WREG32_SOC15(GC, 0, regCP_ME_PRGRM_CNTR_START_HI, 2152 me_hdr->ucode_start_addr_hi>>2); 2153 } 2154 soc24_grbm_select(adev, 0, 0, 0, 0); 2155 2156 /* reset me pipe */ 2157 tmp = RREG32_SOC15(GC, 0, regCP_ME_CNTL); 2158 tmp = REG_SET_FIELD(tmp, CP_ME_CNTL, ME_PIPE0_RESET, 1); 2159 tmp = REG_SET_FIELD(tmp, CP_ME_CNTL, ME_PIPE1_RESET, 1); 2160 WREG32_SOC15(GC, 0, regCP_ME_CNTL, tmp); 2161 2162 /* clear me pipe reset */ 2163 tmp = REG_SET_FIELD(tmp, CP_ME_CNTL, ME_PIPE0_RESET, 0); 2164 tmp = REG_SET_FIELD(tmp, CP_ME_CNTL, ME_PIPE1_RESET, 0); 2165 WREG32_SOC15(GC, 0, regCP_ME_CNTL, tmp); 2166 2167 /* config mec program start addr */ 2168 for (pipe_id = 0; pipe_id < 4; pipe_id++) { 2169 soc24_grbm_select(adev, 1, pipe_id, 0, 0); 2170 WREG32_SOC15(GC, 0, regCP_MEC_RS64_PRGRM_CNTR_START, 2171 mec_hdr->ucode_start_addr_lo >> 2 | 2172 mec_hdr->ucode_start_addr_hi << 30); 2173 WREG32_SOC15(GC, 0, regCP_MEC_RS64_PRGRM_CNTR_START_HI, 2174 mec_hdr->ucode_start_addr_hi >> 2); 2175 } 2176 soc24_grbm_select(adev, 0, 0, 0, 0); 2177 2178 /* reset mec pipe */ 2179 tmp = RREG32_SOC15(GC, 0, regCP_MEC_RS64_CNTL); 2180 tmp = REG_SET_FIELD(tmp, CP_MEC_RS64_CNTL, MEC_PIPE0_RESET, 1); 2181 tmp = REG_SET_FIELD(tmp, CP_MEC_RS64_CNTL, MEC_PIPE1_RESET, 1); 2182 tmp = REG_SET_FIELD(tmp, CP_MEC_RS64_CNTL, MEC_PIPE2_RESET, 1); 2183 tmp = REG_SET_FIELD(tmp, CP_MEC_RS64_CNTL, MEC_PIPE3_RESET, 1); 2184 WREG32_SOC15(GC, 0, regCP_MEC_RS64_CNTL, tmp); 2185 2186 /* clear mec pipe reset */ 2187 tmp = REG_SET_FIELD(tmp, CP_MEC_RS64_CNTL, MEC_PIPE0_RESET, 0); 2188 tmp = REG_SET_FIELD(tmp, CP_MEC_RS64_CNTL, MEC_PIPE1_RESET, 0); 2189 tmp = REG_SET_FIELD(tmp, CP_MEC_RS64_CNTL, MEC_PIPE2_RESET, 0); 2190 tmp = REG_SET_FIELD(tmp, CP_MEC_RS64_CNTL, MEC_PIPE3_RESET, 0); 2191 WREG32_SOC15(GC, 0, regCP_MEC_RS64_CNTL, tmp); 2192 } 2193 2194 static void gfx_v12_0_set_pfp_ucode_start_addr(struct amdgpu_device *adev) 2195 { 2196 const struct gfx_firmware_header_v2_0 *cp_hdr; 2197 unsigned pipe_id, tmp; 2198 2199 cp_hdr = (const struct gfx_firmware_header_v2_0 *) 2200 adev->gfx.pfp_fw->data; 2201 mutex_lock(&adev->srbm_mutex); 2202 for (pipe_id = 0; pipe_id < adev->gfx.me.num_pipe_per_me; pipe_id++) { 2203 soc24_grbm_select(adev, 0, pipe_id, 0, 0); 2204 WREG32_SOC15(GC, 0, regCP_PFP_PRGRM_CNTR_START, 2205 (cp_hdr->ucode_start_addr_hi << 30) | 2206 (cp_hdr->ucode_start_addr_lo >> 2)); 2207 WREG32_SOC15(GC, 0, regCP_PFP_PRGRM_CNTR_START_HI, 2208 cp_hdr->ucode_start_addr_hi>>2); 2209 2210 /* 2211 * Program CP_ME_CNTL to reset given PIPE to take 2212 * effect of CP_PFP_PRGRM_CNTR_START. 2213 */ 2214 tmp = RREG32_SOC15(GC, 0, regCP_ME_CNTL); 2215 if (pipe_id == 0) 2216 tmp = REG_SET_FIELD(tmp, CP_ME_CNTL, 2217 PFP_PIPE0_RESET, 1); 2218 else 2219 tmp = REG_SET_FIELD(tmp, CP_ME_CNTL, 2220 PFP_PIPE1_RESET, 1); 2221 WREG32_SOC15(GC, 0, regCP_ME_CNTL, tmp); 2222 2223 /* Clear pfp pipe0 reset bit. */ 2224 if (pipe_id == 0) 2225 tmp = REG_SET_FIELD(tmp, CP_ME_CNTL, 2226 PFP_PIPE0_RESET, 0); 2227 else 2228 tmp = REG_SET_FIELD(tmp, CP_ME_CNTL, 2229 PFP_PIPE1_RESET, 0); 2230 WREG32_SOC15(GC, 0, regCP_ME_CNTL, tmp); 2231 } 2232 soc24_grbm_select(adev, 0, 0, 0, 0); 2233 mutex_unlock(&adev->srbm_mutex); 2234 } 2235 2236 static void gfx_v12_0_set_me_ucode_start_addr(struct amdgpu_device *adev) 2237 { 2238 const struct gfx_firmware_header_v2_0 *cp_hdr; 2239 unsigned pipe_id, tmp; 2240 2241 cp_hdr = (const struct gfx_firmware_header_v2_0 *) 2242 adev->gfx.me_fw->data; 2243 mutex_lock(&adev->srbm_mutex); 2244 for (pipe_id = 0; pipe_id < adev->gfx.me.num_pipe_per_me; pipe_id++) { 2245 soc24_grbm_select(adev, 0, pipe_id, 0, 0); 2246 WREG32_SOC15(GC, 0, regCP_ME_PRGRM_CNTR_START, 2247 (cp_hdr->ucode_start_addr_hi << 30) | 2248 (cp_hdr->ucode_start_addr_lo >> 2) ); 2249 WREG32_SOC15(GC, 0, regCP_ME_PRGRM_CNTR_START_HI, 2250 cp_hdr->ucode_start_addr_hi>>2); 2251 2252 /* 2253 * Program CP_ME_CNTL to reset given PIPE to take 2254 * effect of CP_ME_PRGRM_CNTR_START. 2255 */ 2256 tmp = RREG32_SOC15(GC, 0, regCP_ME_CNTL); 2257 if (pipe_id == 0) 2258 tmp = REG_SET_FIELD(tmp, CP_ME_CNTL, 2259 ME_PIPE0_RESET, 1); 2260 else 2261 tmp = REG_SET_FIELD(tmp, CP_ME_CNTL, 2262 ME_PIPE1_RESET, 1); 2263 WREG32_SOC15(GC, 0, regCP_ME_CNTL, tmp); 2264 2265 /* Clear pfp pipe0 reset bit. */ 2266 if (pipe_id == 0) 2267 tmp = REG_SET_FIELD(tmp, CP_ME_CNTL, 2268 ME_PIPE0_RESET, 0); 2269 else 2270 tmp = REG_SET_FIELD(tmp, CP_ME_CNTL, 2271 ME_PIPE1_RESET, 0); 2272 WREG32_SOC15(GC, 0, regCP_ME_CNTL, tmp); 2273 } 2274 soc24_grbm_select(adev, 0, 0, 0, 0); 2275 mutex_unlock(&adev->srbm_mutex); 2276 } 2277 2278 static void gfx_v12_0_set_mec_ucode_start_addr(struct amdgpu_device *adev) 2279 { 2280 const struct gfx_firmware_header_v2_0 *cp_hdr; 2281 unsigned pipe_id; 2282 2283 cp_hdr = (const struct gfx_firmware_header_v2_0 *) 2284 adev->gfx.mec_fw->data; 2285 mutex_lock(&adev->srbm_mutex); 2286 for (pipe_id = 0; pipe_id < adev->gfx.mec.num_pipe_per_mec; pipe_id++) { 2287 soc24_grbm_select(adev, 1, pipe_id, 0, 0); 2288 WREG32_SOC15(GC, 0, regCP_MEC_RS64_PRGRM_CNTR_START, 2289 cp_hdr->ucode_start_addr_lo >> 2 | 2290 cp_hdr->ucode_start_addr_hi << 30); 2291 WREG32_SOC15(GC, 0, regCP_MEC_RS64_PRGRM_CNTR_START_HI, 2292 cp_hdr->ucode_start_addr_hi >> 2); 2293 } 2294 soc24_grbm_select(adev, 0, 0, 0, 0); 2295 mutex_unlock(&adev->srbm_mutex); 2296 } 2297 2298 static int gfx_v12_0_wait_for_rlc_autoload_complete(struct amdgpu_device *adev) 2299 { 2300 uint32_t cp_status; 2301 uint32_t bootload_status; 2302 int i; 2303 2304 for (i = 0; i < adev->usec_timeout; i++) { 2305 cp_status = RREG32_SOC15(GC, 0, regCP_STAT); 2306 bootload_status = RREG32_SOC15(GC, 0, regRLC_RLCS_BOOTLOAD_STATUS); 2307 2308 if ((cp_status == 0) && 2309 (REG_GET_FIELD(bootload_status, 2310 RLC_RLCS_BOOTLOAD_STATUS, BOOTLOAD_COMPLETE) == 1)) { 2311 break; 2312 } 2313 udelay(1); 2314 if (amdgpu_emu_mode) 2315 msleep(10); 2316 } 2317 2318 if (i >= adev->usec_timeout) { 2319 dev_err(adev->dev, "rlc autoload: gc ucode autoload timeout\n"); 2320 return -ETIMEDOUT; 2321 } 2322 2323 if (adev->firmware.load_type == AMDGPU_FW_LOAD_RLC_BACKDOOR_AUTO) { 2324 gfx_v12_0_set_pfp_ucode_start_addr(adev); 2325 gfx_v12_0_set_me_ucode_start_addr(adev); 2326 gfx_v12_0_set_mec_ucode_start_addr(adev); 2327 } 2328 2329 return 0; 2330 } 2331 2332 static int gfx_v12_0_cp_gfx_enable(struct amdgpu_device *adev, bool enable) 2333 { 2334 int i; 2335 u32 tmp = RREG32_SOC15(GC, 0, regCP_ME_CNTL); 2336 2337 tmp = REG_SET_FIELD(tmp, CP_ME_CNTL, ME_HALT, enable ? 0 : 1); 2338 tmp = REG_SET_FIELD(tmp, CP_ME_CNTL, PFP_HALT, enable ? 0 : 1); 2339 WREG32_SOC15(GC, 0, regCP_ME_CNTL, tmp); 2340 2341 for (i = 0; i < adev->usec_timeout; i++) { 2342 if (RREG32_SOC15(GC, 0, regCP_STAT) == 0) 2343 break; 2344 udelay(1); 2345 } 2346 2347 if (i >= adev->usec_timeout) 2348 DRM_ERROR("failed to %s cp gfx\n", enable ? "unhalt" : "halt"); 2349 2350 return 0; 2351 } 2352 2353 static int gfx_v12_0_cp_gfx_load_pfp_microcode_rs64(struct amdgpu_device *adev) 2354 { 2355 int r; 2356 const struct gfx_firmware_header_v2_0 *pfp_hdr; 2357 const __le32 *fw_ucode, *fw_data; 2358 unsigned i, pipe_id, fw_ucode_size, fw_data_size; 2359 uint32_t tmp; 2360 uint32_t usec_timeout = 50000; /* wait for 50ms */ 2361 2362 pfp_hdr = (const struct gfx_firmware_header_v2_0 *) 2363 adev->gfx.pfp_fw->data; 2364 2365 amdgpu_ucode_print_gfx_hdr(&pfp_hdr->header); 2366 2367 /* instruction */ 2368 fw_ucode = (const __le32 *)(adev->gfx.pfp_fw->data + 2369 le32_to_cpu(pfp_hdr->ucode_offset_bytes)); 2370 fw_ucode_size = le32_to_cpu(pfp_hdr->ucode_size_bytes); 2371 /* data */ 2372 fw_data = (const __le32 *)(adev->gfx.pfp_fw->data + 2373 le32_to_cpu(pfp_hdr->data_offset_bytes)); 2374 fw_data_size = le32_to_cpu(pfp_hdr->data_size_bytes); 2375 2376 /* 64kb align */ 2377 r = amdgpu_bo_create_reserved(adev, fw_ucode_size, 2378 64 * 1024, AMDGPU_GEM_DOMAIN_VRAM, 2379 &adev->gfx.pfp.pfp_fw_obj, 2380 &adev->gfx.pfp.pfp_fw_gpu_addr, 2381 (void **)&adev->gfx.pfp.pfp_fw_ptr); 2382 if (r) { 2383 dev_err(adev->dev, "(%d) failed to create pfp ucode fw bo\n", r); 2384 gfx_v12_0_pfp_fini(adev); 2385 return r; 2386 } 2387 2388 r = amdgpu_bo_create_reserved(adev, fw_data_size, 2389 64 * 1024, AMDGPU_GEM_DOMAIN_VRAM, 2390 &adev->gfx.pfp.pfp_fw_data_obj, 2391 &adev->gfx.pfp.pfp_fw_data_gpu_addr, 2392 (void **)&adev->gfx.pfp.pfp_fw_data_ptr); 2393 if (r) { 2394 dev_err(adev->dev, "(%d) failed to create pfp data fw bo\n", r); 2395 gfx_v12_0_pfp_fini(adev); 2396 return r; 2397 } 2398 2399 memcpy(adev->gfx.pfp.pfp_fw_ptr, fw_ucode, fw_ucode_size); 2400 memcpy(adev->gfx.pfp.pfp_fw_data_ptr, fw_data, fw_data_size); 2401 2402 amdgpu_bo_kunmap(adev->gfx.pfp.pfp_fw_obj); 2403 amdgpu_bo_kunmap(adev->gfx.pfp.pfp_fw_data_obj); 2404 amdgpu_bo_unreserve(adev->gfx.pfp.pfp_fw_obj); 2405 amdgpu_bo_unreserve(adev->gfx.pfp.pfp_fw_data_obj); 2406 2407 if (amdgpu_emu_mode == 1) 2408 amdgpu_device_flush_hdp(adev, NULL); 2409 2410 WREG32_SOC15(GC, 0, regCP_PFP_IC_BASE_LO, 2411 lower_32_bits(adev->gfx.pfp.pfp_fw_gpu_addr)); 2412 WREG32_SOC15(GC, 0, regCP_PFP_IC_BASE_HI, 2413 upper_32_bits(adev->gfx.pfp.pfp_fw_gpu_addr)); 2414 2415 tmp = RREG32_SOC15(GC, 0, regCP_PFP_IC_BASE_CNTL); 2416 tmp = REG_SET_FIELD(tmp, CP_PFP_IC_BASE_CNTL, VMID, 0); 2417 tmp = REG_SET_FIELD(tmp, CP_PFP_IC_BASE_CNTL, CACHE_POLICY, 0); 2418 tmp = REG_SET_FIELD(tmp, CP_PFP_IC_BASE_CNTL, EXE_DISABLE, 0); 2419 WREG32_SOC15(GC, 0, regCP_PFP_IC_BASE_CNTL, tmp); 2420 2421 /* 2422 * Programming any of the CP_PFP_IC_BASE registers 2423 * forces invalidation of the ME L1 I$. Wait for the 2424 * invalidation complete 2425 */ 2426 for (i = 0; i < usec_timeout; i++) { 2427 tmp = RREG32_SOC15(GC, 0, regCP_PFP_IC_OP_CNTL); 2428 if (1 == REG_GET_FIELD(tmp, CP_PFP_IC_OP_CNTL, 2429 INVALIDATE_CACHE_COMPLETE)) 2430 break; 2431 udelay(1); 2432 } 2433 2434 if (i >= usec_timeout) { 2435 dev_err(adev->dev, "failed to invalidate instruction cache\n"); 2436 return -EINVAL; 2437 } 2438 2439 /* Prime the L1 instruction caches */ 2440 tmp = RREG32_SOC15(GC, 0, regCP_PFP_IC_OP_CNTL); 2441 tmp = REG_SET_FIELD(tmp, CP_PFP_IC_OP_CNTL, PRIME_ICACHE, 1); 2442 WREG32_SOC15(GC, 0, regCP_PFP_IC_OP_CNTL, tmp); 2443 /* Waiting for cache primed*/ 2444 for (i = 0; i < usec_timeout; i++) { 2445 tmp = RREG32_SOC15(GC, 0, regCP_PFP_IC_OP_CNTL); 2446 if (1 == REG_GET_FIELD(tmp, CP_PFP_IC_OP_CNTL, 2447 ICACHE_PRIMED)) 2448 break; 2449 udelay(1); 2450 } 2451 2452 if (i >= usec_timeout) { 2453 dev_err(adev->dev, "failed to prime instruction cache\n"); 2454 return -EINVAL; 2455 } 2456 2457 mutex_lock(&adev->srbm_mutex); 2458 for (pipe_id = 0; pipe_id < adev->gfx.me.num_pipe_per_me; pipe_id++) { 2459 soc24_grbm_select(adev, 0, pipe_id, 0, 0); 2460 2461 WREG32_SOC15(GC, 0, regCP_GFX_RS64_DC_BASE0_LO, 2462 lower_32_bits(adev->gfx.pfp.pfp_fw_data_gpu_addr)); 2463 WREG32_SOC15(GC, 0, regCP_GFX_RS64_DC_BASE0_HI, 2464 upper_32_bits(adev->gfx.pfp.pfp_fw_data_gpu_addr)); 2465 } 2466 soc24_grbm_select(adev, 0, 0, 0, 0); 2467 mutex_unlock(&adev->srbm_mutex); 2468 2469 tmp = RREG32_SOC15(GC, 0, regCP_GFX_RS64_DC_BASE_CNTL); 2470 tmp = REG_SET_FIELD(tmp, CP_GFX_RS64_DC_BASE_CNTL, VMID, 0); 2471 tmp = REG_SET_FIELD(tmp, CP_GFX_RS64_DC_BASE_CNTL, CACHE_POLICY, 0); 2472 WREG32_SOC15(GC, 0, regCP_GFX_RS64_DC_BASE_CNTL, tmp); 2473 2474 /* Invalidate the data caches */ 2475 tmp = RREG32_SOC15(GC, 0, regCP_GFX_RS64_DC_OP_CNTL); 2476 tmp = REG_SET_FIELD(tmp, CP_GFX_RS64_DC_OP_CNTL, INVALIDATE_DCACHE, 1); 2477 WREG32_SOC15(GC, 0, regCP_GFX_RS64_DC_OP_CNTL, tmp); 2478 2479 for (i = 0; i < usec_timeout; i++) { 2480 tmp = RREG32_SOC15(GC, 0, regCP_GFX_RS64_DC_OP_CNTL); 2481 if (1 == REG_GET_FIELD(tmp, CP_GFX_RS64_DC_OP_CNTL, 2482 INVALIDATE_DCACHE_COMPLETE)) 2483 break; 2484 udelay(1); 2485 } 2486 2487 if (i >= usec_timeout) { 2488 dev_err(adev->dev, "failed to invalidate RS64 data cache\n"); 2489 return -EINVAL; 2490 } 2491 2492 gfx_v12_0_set_pfp_ucode_start_addr(adev); 2493 2494 return 0; 2495 } 2496 2497 static int gfx_v12_0_cp_gfx_load_me_microcode_rs64(struct amdgpu_device *adev) 2498 { 2499 int r; 2500 const struct gfx_firmware_header_v2_0 *me_hdr; 2501 const __le32 *fw_ucode, *fw_data; 2502 unsigned i, pipe_id, fw_ucode_size, fw_data_size; 2503 uint32_t tmp; 2504 uint32_t usec_timeout = 50000; /* wait for 50ms */ 2505 2506 me_hdr = (const struct gfx_firmware_header_v2_0 *) 2507 adev->gfx.me_fw->data; 2508 2509 amdgpu_ucode_print_gfx_hdr(&me_hdr->header); 2510 2511 /* instruction */ 2512 fw_ucode = (const __le32 *)(adev->gfx.me_fw->data + 2513 le32_to_cpu(me_hdr->ucode_offset_bytes)); 2514 fw_ucode_size = le32_to_cpu(me_hdr->ucode_size_bytes); 2515 /* data */ 2516 fw_data = (const __le32 *)(adev->gfx.me_fw->data + 2517 le32_to_cpu(me_hdr->data_offset_bytes)); 2518 fw_data_size = le32_to_cpu(me_hdr->data_size_bytes); 2519 2520 /* 64kb align*/ 2521 r = amdgpu_bo_create_reserved(adev, fw_ucode_size, 2522 64 * 1024, AMDGPU_GEM_DOMAIN_VRAM, 2523 &adev->gfx.me.me_fw_obj, 2524 &adev->gfx.me.me_fw_gpu_addr, 2525 (void **)&adev->gfx.me.me_fw_ptr); 2526 if (r) { 2527 dev_err(adev->dev, "(%d) failed to create me ucode bo\n", r); 2528 gfx_v12_0_me_fini(adev); 2529 return r; 2530 } 2531 2532 r = amdgpu_bo_create_reserved(adev, fw_data_size, 2533 64 * 1024, AMDGPU_GEM_DOMAIN_VRAM, 2534 &adev->gfx.me.me_fw_data_obj, 2535 &adev->gfx.me.me_fw_data_gpu_addr, 2536 (void **)&adev->gfx.me.me_fw_data_ptr); 2537 if (r) { 2538 dev_err(adev->dev, "(%d) failed to create me data bo\n", r); 2539 gfx_v12_0_me_fini(adev); 2540 return r; 2541 } 2542 2543 memcpy(adev->gfx.me.me_fw_ptr, fw_ucode, fw_ucode_size); 2544 memcpy(adev->gfx.me.me_fw_data_ptr, fw_data, fw_data_size); 2545 2546 amdgpu_bo_kunmap(adev->gfx.me.me_fw_obj); 2547 amdgpu_bo_kunmap(adev->gfx.me.me_fw_data_obj); 2548 amdgpu_bo_unreserve(adev->gfx.me.me_fw_obj); 2549 amdgpu_bo_unreserve(adev->gfx.me.me_fw_data_obj); 2550 2551 if (amdgpu_emu_mode == 1) 2552 amdgpu_device_flush_hdp(adev, NULL); 2553 2554 WREG32_SOC15(GC, 0, regCP_ME_IC_BASE_LO, 2555 lower_32_bits(adev->gfx.me.me_fw_gpu_addr)); 2556 WREG32_SOC15(GC, 0, regCP_ME_IC_BASE_HI, 2557 upper_32_bits(adev->gfx.me.me_fw_gpu_addr)); 2558 2559 tmp = RREG32_SOC15(GC, 0, regCP_ME_IC_BASE_CNTL); 2560 tmp = REG_SET_FIELD(tmp, CP_ME_IC_BASE_CNTL, VMID, 0); 2561 tmp = REG_SET_FIELD(tmp, CP_ME_IC_BASE_CNTL, CACHE_POLICY, 0); 2562 tmp = REG_SET_FIELD(tmp, CP_ME_IC_BASE_CNTL, EXE_DISABLE, 0); 2563 WREG32_SOC15(GC, 0, regCP_ME_IC_BASE_CNTL, tmp); 2564 2565 /* 2566 * Programming any of the CP_ME_IC_BASE registers 2567 * forces invalidation of the ME L1 I$. Wait for the 2568 * invalidation complete 2569 */ 2570 for (i = 0; i < usec_timeout; i++) { 2571 tmp = RREG32_SOC15(GC, 0, regCP_ME_IC_OP_CNTL); 2572 if (1 == REG_GET_FIELD(tmp, CP_ME_IC_OP_CNTL, 2573 INVALIDATE_CACHE_COMPLETE)) 2574 break; 2575 udelay(1); 2576 } 2577 2578 if (i >= usec_timeout) { 2579 dev_err(adev->dev, "failed to invalidate instruction cache\n"); 2580 return -EINVAL; 2581 } 2582 2583 /* Prime the instruction caches */ 2584 tmp = RREG32_SOC15(GC, 0, regCP_ME_IC_OP_CNTL); 2585 tmp = REG_SET_FIELD(tmp, CP_ME_IC_OP_CNTL, PRIME_ICACHE, 1); 2586 WREG32_SOC15(GC, 0, regCP_ME_IC_OP_CNTL, tmp); 2587 2588 /* Waiting for instruction cache primed*/ 2589 for (i = 0; i < usec_timeout; i++) { 2590 tmp = RREG32_SOC15(GC, 0, regCP_ME_IC_OP_CNTL); 2591 if (1 == REG_GET_FIELD(tmp, CP_ME_IC_OP_CNTL, 2592 ICACHE_PRIMED)) 2593 break; 2594 udelay(1); 2595 } 2596 2597 if (i >= usec_timeout) { 2598 dev_err(adev->dev, "failed to prime instruction cache\n"); 2599 return -EINVAL; 2600 } 2601 2602 mutex_lock(&adev->srbm_mutex); 2603 for (pipe_id = 0; pipe_id < adev->gfx.me.num_pipe_per_me; pipe_id++) { 2604 soc24_grbm_select(adev, 0, pipe_id, 0, 0); 2605 2606 WREG32_SOC15(GC, 0, regCP_GFX_RS64_DC_BASE1_LO, 2607 lower_32_bits(adev->gfx.me.me_fw_data_gpu_addr)); 2608 WREG32_SOC15(GC, 0, regCP_GFX_RS64_DC_BASE1_HI, 2609 upper_32_bits(adev->gfx.me.me_fw_data_gpu_addr)); 2610 } 2611 soc24_grbm_select(adev, 0, 0, 0, 0); 2612 mutex_unlock(&adev->srbm_mutex); 2613 2614 tmp = RREG32_SOC15(GC, 0, regCP_GFX_RS64_DC_BASE_CNTL); 2615 tmp = REG_SET_FIELD(tmp, CP_GFX_RS64_DC_BASE_CNTL, VMID, 0); 2616 tmp = REG_SET_FIELD(tmp, CP_GFX_RS64_DC_BASE_CNTL, CACHE_POLICY, 0); 2617 WREG32_SOC15(GC, 0, regCP_GFX_RS64_DC_BASE_CNTL, tmp); 2618 2619 /* Invalidate the data caches */ 2620 tmp = RREG32_SOC15(GC, 0, regCP_GFX_RS64_DC_OP_CNTL); 2621 tmp = REG_SET_FIELD(tmp, CP_GFX_RS64_DC_OP_CNTL, INVALIDATE_DCACHE, 1); 2622 WREG32_SOC15(GC, 0, regCP_GFX_RS64_DC_OP_CNTL, tmp); 2623 2624 for (i = 0; i < usec_timeout; i++) { 2625 tmp = RREG32_SOC15(GC, 0, regCP_GFX_RS64_DC_OP_CNTL); 2626 if (1 == REG_GET_FIELD(tmp, CP_GFX_RS64_DC_OP_CNTL, 2627 INVALIDATE_DCACHE_COMPLETE)) 2628 break; 2629 udelay(1); 2630 } 2631 2632 if (i >= usec_timeout) { 2633 dev_err(adev->dev, "failed to invalidate RS64 data cache\n"); 2634 return -EINVAL; 2635 } 2636 2637 gfx_v12_0_set_me_ucode_start_addr(adev); 2638 2639 return 0; 2640 } 2641 2642 static int gfx_v12_0_cp_gfx_load_microcode(struct amdgpu_device *adev) 2643 { 2644 int r; 2645 2646 if (!adev->gfx.me_fw || !adev->gfx.pfp_fw) 2647 return -EINVAL; 2648 2649 gfx_v12_0_cp_gfx_enable(adev, false); 2650 2651 r = gfx_v12_0_cp_gfx_load_pfp_microcode_rs64(adev); 2652 if (r) { 2653 dev_err(adev->dev, "(%d) failed to load pfp fw\n", r); 2654 return r; 2655 } 2656 2657 r = gfx_v12_0_cp_gfx_load_me_microcode_rs64(adev); 2658 if (r) { 2659 dev_err(adev->dev, "(%d) failed to load me fw\n", r); 2660 return r; 2661 } 2662 2663 return 0; 2664 } 2665 2666 static int gfx_v12_0_cp_gfx_start(struct amdgpu_device *adev) 2667 { 2668 /* init the CP */ 2669 WREG32_SOC15(GC, 0, regCP_MAX_CONTEXT, 2670 adev->gfx.config.max_hw_contexts - 1); 2671 WREG32_SOC15(GC, 0, regCP_DEVICE_ID, 1); 2672 2673 if (!amdgpu_async_gfx_ring) 2674 gfx_v12_0_cp_gfx_enable(adev, true); 2675 2676 return 0; 2677 } 2678 2679 static void gfx_v12_0_cp_gfx_switch_pipe(struct amdgpu_device *adev, 2680 CP_PIPE_ID pipe) 2681 { 2682 u32 tmp; 2683 2684 tmp = RREG32_SOC15(GC, 0, regGRBM_GFX_CNTL); 2685 tmp = REG_SET_FIELD(tmp, GRBM_GFX_CNTL, PIPEID, pipe); 2686 2687 WREG32_SOC15(GC, 0, regGRBM_GFX_CNTL, tmp); 2688 } 2689 2690 static void gfx_v12_0_cp_gfx_set_doorbell(struct amdgpu_device *adev, 2691 struct amdgpu_ring *ring) 2692 { 2693 u32 tmp; 2694 2695 tmp = RREG32_SOC15(GC, 0, regCP_RB_DOORBELL_CONTROL); 2696 if (ring->use_doorbell) { 2697 tmp = REG_SET_FIELD(tmp, CP_RB_DOORBELL_CONTROL, 2698 DOORBELL_OFFSET, ring->doorbell_index); 2699 tmp = REG_SET_FIELD(tmp, CP_RB_DOORBELL_CONTROL, 2700 DOORBELL_EN, 1); 2701 } else { 2702 tmp = REG_SET_FIELD(tmp, CP_RB_DOORBELL_CONTROL, 2703 DOORBELL_EN, 0); 2704 } 2705 WREG32_SOC15(GC, 0, regCP_RB_DOORBELL_CONTROL, tmp); 2706 2707 tmp = REG_SET_FIELD(0, CP_RB_DOORBELL_RANGE_LOWER, 2708 DOORBELL_RANGE_LOWER, ring->doorbell_index); 2709 WREG32_SOC15(GC, 0, regCP_RB_DOORBELL_RANGE_LOWER, tmp); 2710 2711 WREG32_SOC15(GC, 0, regCP_RB_DOORBELL_RANGE_UPPER, 2712 CP_RB_DOORBELL_RANGE_UPPER__DOORBELL_RANGE_UPPER_MASK); 2713 } 2714 2715 static int gfx_v12_0_cp_gfx_resume(struct amdgpu_device *adev) 2716 { 2717 struct amdgpu_ring *ring; 2718 u32 tmp; 2719 u32 rb_bufsz; 2720 u64 rb_addr, rptr_addr, wptr_gpu_addr; 2721 2722 /* Set the write pointer delay */ 2723 WREG32_SOC15(GC, 0, regCP_RB_WPTR_DELAY, 0); 2724 2725 /* set the RB to use vmid 0 */ 2726 WREG32_SOC15(GC, 0, regCP_RB_VMID, 0); 2727 2728 /* Init gfx ring 0 for pipe 0 */ 2729 mutex_lock(&adev->srbm_mutex); 2730 gfx_v12_0_cp_gfx_switch_pipe(adev, PIPE_ID0); 2731 2732 /* Set ring buffer size */ 2733 ring = &adev->gfx.gfx_ring[0]; 2734 rb_bufsz = order_base_2(ring->ring_size / 8); 2735 tmp = REG_SET_FIELD(0, CP_RB0_CNTL, RB_BUFSZ, rb_bufsz); 2736 tmp = REG_SET_FIELD(tmp, CP_RB0_CNTL, RB_BLKSZ, rb_bufsz - 2); 2737 WREG32_SOC15(GC, 0, regCP_RB0_CNTL, tmp); 2738 2739 /* Initialize the ring buffer's write pointers */ 2740 ring->wptr = 0; 2741 WREG32_SOC15(GC, 0, regCP_RB0_WPTR, lower_32_bits(ring->wptr)); 2742 WREG32_SOC15(GC, 0, regCP_RB0_WPTR_HI, upper_32_bits(ring->wptr)); 2743 2744 /* set the wb address whether it's enabled or not */ 2745 rptr_addr = ring->rptr_gpu_addr; 2746 WREG32_SOC15(GC, 0, regCP_RB0_RPTR_ADDR, lower_32_bits(rptr_addr)); 2747 WREG32_SOC15(GC, 0, regCP_RB0_RPTR_ADDR_HI, upper_32_bits(rptr_addr) & 2748 CP_RB_RPTR_ADDR_HI__RB_RPTR_ADDR_HI_MASK); 2749 2750 wptr_gpu_addr = ring->wptr_gpu_addr; 2751 WREG32_SOC15(GC, 0, regCP_RB_WPTR_POLL_ADDR_LO, 2752 lower_32_bits(wptr_gpu_addr)); 2753 WREG32_SOC15(GC, 0, regCP_RB_WPTR_POLL_ADDR_HI, 2754 upper_32_bits(wptr_gpu_addr)); 2755 2756 mdelay(1); 2757 WREG32_SOC15(GC, 0, regCP_RB0_CNTL, tmp); 2758 2759 rb_addr = ring->gpu_addr >> 8; 2760 WREG32_SOC15(GC, 0, regCP_RB0_BASE, rb_addr); 2761 WREG32_SOC15(GC, 0, regCP_RB0_BASE_HI, upper_32_bits(rb_addr)); 2762 2763 WREG32_SOC15(GC, 0, regCP_RB_ACTIVE, 1); 2764 2765 gfx_v12_0_cp_gfx_set_doorbell(adev, ring); 2766 mutex_unlock(&adev->srbm_mutex); 2767 2768 /* Switch to pipe 0 */ 2769 mutex_lock(&adev->srbm_mutex); 2770 gfx_v12_0_cp_gfx_switch_pipe(adev, PIPE_ID0); 2771 mutex_unlock(&adev->srbm_mutex); 2772 2773 /* start the ring */ 2774 gfx_v12_0_cp_gfx_start(adev); 2775 return 0; 2776 } 2777 2778 static void gfx_v12_0_cp_compute_enable(struct amdgpu_device *adev, bool enable) 2779 { 2780 u32 data; 2781 2782 data = RREG32_SOC15(GC, 0, regCP_MEC_RS64_CNTL); 2783 data = REG_SET_FIELD(data, CP_MEC_RS64_CNTL, MEC_INVALIDATE_ICACHE, 2784 enable ? 0 : 1); 2785 data = REG_SET_FIELD(data, CP_MEC_RS64_CNTL, MEC_PIPE0_RESET, 2786 enable ? 0 : 1); 2787 data = REG_SET_FIELD(data, CP_MEC_RS64_CNTL, MEC_PIPE1_RESET, 2788 enable ? 0 : 1); 2789 data = REG_SET_FIELD(data, CP_MEC_RS64_CNTL, MEC_PIPE2_RESET, 2790 enable ? 0 : 1); 2791 data = REG_SET_FIELD(data, CP_MEC_RS64_CNTL, MEC_PIPE3_RESET, 2792 enable ? 0 : 1); 2793 data = REG_SET_FIELD(data, CP_MEC_RS64_CNTL, MEC_PIPE0_ACTIVE, 2794 enable ? 1 : 0); 2795 data = REG_SET_FIELD(data, CP_MEC_RS64_CNTL, MEC_PIPE1_ACTIVE, 2796 enable ? 1 : 0); 2797 data = REG_SET_FIELD(data, CP_MEC_RS64_CNTL, MEC_PIPE2_ACTIVE, 2798 enable ? 1 : 0); 2799 data = REG_SET_FIELD(data, CP_MEC_RS64_CNTL, MEC_PIPE3_ACTIVE, 2800 enable ? 1 : 0); 2801 data = REG_SET_FIELD(data, CP_MEC_RS64_CNTL, MEC_HALT, 2802 enable ? 0 : 1); 2803 WREG32_SOC15(GC, 0, regCP_MEC_RS64_CNTL, data); 2804 2805 adev->gfx.kiq[0].ring.sched.ready = enable; 2806 2807 udelay(50); 2808 } 2809 2810 static int gfx_v12_0_cp_compute_load_microcode_rs64(struct amdgpu_device *adev) 2811 { 2812 const struct gfx_firmware_header_v2_0 *mec_hdr; 2813 const __le32 *fw_ucode, *fw_data; 2814 u32 tmp, fw_ucode_size, fw_data_size; 2815 u32 i, usec_timeout = 50000; /* Wait for 50 ms */ 2816 u32 *fw_ucode_ptr, *fw_data_ptr; 2817 int r; 2818 2819 if (!adev->gfx.mec_fw) 2820 return -EINVAL; 2821 2822 gfx_v12_0_cp_compute_enable(adev, false); 2823 2824 mec_hdr = (const struct gfx_firmware_header_v2_0 *)adev->gfx.mec_fw->data; 2825 amdgpu_ucode_print_gfx_hdr(&mec_hdr->header); 2826 2827 fw_ucode = (const __le32 *) (adev->gfx.mec_fw->data + 2828 le32_to_cpu(mec_hdr->ucode_offset_bytes)); 2829 fw_ucode_size = le32_to_cpu(mec_hdr->ucode_size_bytes); 2830 2831 fw_data = (const __le32 *) (adev->gfx.mec_fw->data + 2832 le32_to_cpu(mec_hdr->data_offset_bytes)); 2833 fw_data_size = le32_to_cpu(mec_hdr->data_size_bytes); 2834 2835 r = amdgpu_bo_create_reserved(adev, fw_ucode_size, 2836 64 * 1024, AMDGPU_GEM_DOMAIN_VRAM, 2837 &adev->gfx.mec.mec_fw_obj, 2838 &adev->gfx.mec.mec_fw_gpu_addr, 2839 (void **)&fw_ucode_ptr); 2840 if (r) { 2841 dev_err(adev->dev, "(%d) failed to create mec fw ucode bo\n", r); 2842 gfx_v12_0_mec_fini(adev); 2843 return r; 2844 } 2845 2846 r = amdgpu_bo_create_reserved(adev, 2847 ALIGN(fw_data_size, 64 * 1024) * 2848 adev->gfx.mec.num_pipe_per_mec, 2849 64 * 1024, AMDGPU_GEM_DOMAIN_VRAM, 2850 &adev->gfx.mec.mec_fw_data_obj, 2851 &adev->gfx.mec.mec_fw_data_gpu_addr, 2852 (void **)&fw_data_ptr); 2853 if (r) { 2854 dev_err(adev->dev, "(%d) failed to create mec fw ucode bo\n", r); 2855 gfx_v12_0_mec_fini(adev); 2856 return r; 2857 } 2858 2859 memcpy(fw_ucode_ptr, fw_ucode, fw_ucode_size); 2860 for (i = 0; i < adev->gfx.mec.num_pipe_per_mec; i++) { 2861 memcpy(fw_data_ptr + i * ALIGN(fw_data_size, 64 * 1024) / 4, fw_data, fw_data_size); 2862 } 2863 2864 amdgpu_bo_kunmap(adev->gfx.mec.mec_fw_obj); 2865 amdgpu_bo_kunmap(adev->gfx.mec.mec_fw_data_obj); 2866 amdgpu_bo_unreserve(adev->gfx.mec.mec_fw_obj); 2867 amdgpu_bo_unreserve(adev->gfx.mec.mec_fw_data_obj); 2868 2869 tmp = RREG32_SOC15(GC, 0, regCP_CPC_IC_BASE_CNTL); 2870 tmp = REG_SET_FIELD(tmp, CP_CPC_IC_BASE_CNTL, VMID, 0); 2871 tmp = REG_SET_FIELD(tmp, CP_CPC_IC_BASE_CNTL, EXE_DISABLE, 0); 2872 tmp = REG_SET_FIELD(tmp, CP_CPC_IC_BASE_CNTL, CACHE_POLICY, 0); 2873 WREG32_SOC15(GC, 0, regCP_CPC_IC_BASE_CNTL, tmp); 2874 2875 tmp = RREG32_SOC15(GC, 0, regCP_MEC_DC_BASE_CNTL); 2876 tmp = REG_SET_FIELD(tmp, CP_MEC_DC_BASE_CNTL, VMID, 0); 2877 tmp = REG_SET_FIELD(tmp, CP_MEC_DC_BASE_CNTL, CACHE_POLICY, 0); 2878 WREG32_SOC15(GC, 0, regCP_MEC_DC_BASE_CNTL, tmp); 2879 2880 mutex_lock(&adev->srbm_mutex); 2881 for (i = 0; i < adev->gfx.mec.num_pipe_per_mec; i++) { 2882 soc24_grbm_select(adev, 1, i, 0, 0); 2883 2884 WREG32_SOC15(GC, 0, regCP_MEC_MDBASE_LO, 2885 lower_32_bits(adev->gfx.mec.mec_fw_data_gpu_addr + 2886 i * ALIGN(fw_data_size, 64 * 1024))); 2887 WREG32_SOC15(GC, 0, regCP_MEC_MDBASE_HI, 2888 upper_32_bits(adev->gfx.mec.mec_fw_data_gpu_addr + 2889 i * ALIGN(fw_data_size, 64 * 1024))); 2890 2891 WREG32_SOC15(GC, 0, regCP_CPC_IC_BASE_LO, 2892 lower_32_bits(adev->gfx.mec.mec_fw_gpu_addr)); 2893 WREG32_SOC15(GC, 0, regCP_CPC_IC_BASE_HI, 2894 upper_32_bits(adev->gfx.mec.mec_fw_gpu_addr)); 2895 } 2896 mutex_unlock(&adev->srbm_mutex); 2897 soc24_grbm_select(adev, 0, 0, 0, 0); 2898 2899 /* Trigger an invalidation of the L1 instruction caches */ 2900 tmp = RREG32_SOC15(GC, 0, regCP_MEC_DC_OP_CNTL); 2901 tmp = REG_SET_FIELD(tmp, CP_MEC_DC_OP_CNTL, INVALIDATE_DCACHE, 1); 2902 WREG32_SOC15(GC, 0, regCP_MEC_DC_OP_CNTL, tmp); 2903 2904 /* Wait for invalidation complete */ 2905 for (i = 0; i < usec_timeout; i++) { 2906 tmp = RREG32_SOC15(GC, 0, regCP_MEC_DC_OP_CNTL); 2907 if (1 == REG_GET_FIELD(tmp, CP_MEC_DC_OP_CNTL, 2908 INVALIDATE_DCACHE_COMPLETE)) 2909 break; 2910 udelay(1); 2911 } 2912 2913 if (i >= usec_timeout) { 2914 dev_err(adev->dev, "failed to invalidate instruction cache\n"); 2915 return -EINVAL; 2916 } 2917 2918 /* Trigger an invalidation of the L1 instruction caches */ 2919 tmp = RREG32_SOC15(GC, 0, regCP_CPC_IC_OP_CNTL); 2920 tmp = REG_SET_FIELD(tmp, CP_CPC_IC_OP_CNTL, INVALIDATE_CACHE, 1); 2921 WREG32_SOC15(GC, 0, regCP_CPC_IC_OP_CNTL, tmp); 2922 2923 /* Wait for invalidation complete */ 2924 for (i = 0; i < usec_timeout; i++) { 2925 tmp = RREG32_SOC15(GC, 0, regCP_CPC_IC_OP_CNTL); 2926 if (1 == REG_GET_FIELD(tmp, CP_CPC_IC_OP_CNTL, 2927 INVALIDATE_CACHE_COMPLETE)) 2928 break; 2929 udelay(1); 2930 } 2931 2932 if (i >= usec_timeout) { 2933 dev_err(adev->dev, "failed to invalidate instruction cache\n"); 2934 return -EINVAL; 2935 } 2936 2937 gfx_v12_0_set_mec_ucode_start_addr(adev); 2938 2939 return 0; 2940 } 2941 2942 static void gfx_v12_0_kiq_setting(struct amdgpu_ring *ring) 2943 { 2944 uint32_t tmp; 2945 struct amdgpu_device *adev = ring->adev; 2946 2947 /* tell RLC which is KIQ queue */ 2948 tmp = RREG32_SOC15(GC, 0, regRLC_CP_SCHEDULERS); 2949 tmp &= 0xffffff00; 2950 tmp |= (ring->me << 5) | (ring->pipe << 3) | (ring->queue); 2951 WREG32_SOC15(GC, 0, regRLC_CP_SCHEDULERS, tmp | 0x80); 2952 } 2953 2954 static void gfx_v12_0_cp_set_doorbell_range(struct amdgpu_device *adev) 2955 { 2956 /* set graphics engine doorbell range */ 2957 WREG32_SOC15(GC, 0, regCP_RB_DOORBELL_RANGE_LOWER, 2958 (adev->doorbell_index.gfx_ring0 * 2) << 2); 2959 WREG32_SOC15(GC, 0, regCP_RB_DOORBELL_RANGE_UPPER, 2960 (adev->doorbell_index.gfx_userqueue_end * 2) << 2); 2961 2962 /* set compute engine doorbell range */ 2963 WREG32_SOC15(GC, 0, regCP_MEC_DOORBELL_RANGE_LOWER, 2964 (adev->doorbell_index.kiq * 2) << 2); 2965 WREG32_SOC15(GC, 0, regCP_MEC_DOORBELL_RANGE_UPPER, 2966 (adev->doorbell_index.userqueue_end * 2) << 2); 2967 } 2968 2969 static int gfx_v12_0_gfx_mqd_init(struct amdgpu_device *adev, void *m, 2970 struct amdgpu_mqd_prop *prop) 2971 { 2972 struct v12_gfx_mqd *mqd = m; 2973 uint64_t hqd_gpu_addr, wb_gpu_addr; 2974 uint32_t tmp; 2975 uint32_t rb_bufsz; 2976 2977 /* set up gfx hqd wptr */ 2978 mqd->cp_gfx_hqd_wptr = 0; 2979 mqd->cp_gfx_hqd_wptr_hi = 0; 2980 2981 /* set the pointer to the MQD */ 2982 mqd->cp_mqd_base_addr = prop->mqd_gpu_addr & 0xfffffffc; 2983 mqd->cp_mqd_base_addr_hi = upper_32_bits(prop->mqd_gpu_addr); 2984 2985 /* set up mqd control */ 2986 tmp = regCP_GFX_MQD_CONTROL_DEFAULT; 2987 tmp = REG_SET_FIELD(tmp, CP_GFX_MQD_CONTROL, VMID, 0); 2988 tmp = REG_SET_FIELD(tmp, CP_GFX_MQD_CONTROL, PRIV_STATE, 1); 2989 tmp = REG_SET_FIELD(tmp, CP_GFX_MQD_CONTROL, CACHE_POLICY, 0); 2990 mqd->cp_gfx_mqd_control = tmp; 2991 2992 /* set up gfx_hqd_vimd with 0x0 to indicate the ring buffer's vmid */ 2993 tmp = regCP_GFX_HQD_VMID_DEFAULT; 2994 tmp = REG_SET_FIELD(tmp, CP_GFX_HQD_VMID, VMID, 0); 2995 mqd->cp_gfx_hqd_vmid = 0; 2996 2997 /* set up default queue priority level 2998 * 0x0 = low priority, 0x1 = high priority */ 2999 tmp = regCP_GFX_HQD_QUEUE_PRIORITY_DEFAULT; 3000 tmp = REG_SET_FIELD(tmp, CP_GFX_HQD_QUEUE_PRIORITY, PRIORITY_LEVEL, 0); 3001 mqd->cp_gfx_hqd_queue_priority = tmp; 3002 3003 /* set up time quantum */ 3004 tmp = regCP_GFX_HQD_QUANTUM_DEFAULT; 3005 tmp = REG_SET_FIELD(tmp, CP_GFX_HQD_QUANTUM, QUANTUM_EN, 1); 3006 mqd->cp_gfx_hqd_quantum = tmp; 3007 3008 /* set up gfx hqd base. this is similar as CP_RB_BASE */ 3009 hqd_gpu_addr = prop->hqd_base_gpu_addr >> 8; 3010 mqd->cp_gfx_hqd_base = hqd_gpu_addr; 3011 mqd->cp_gfx_hqd_base_hi = upper_32_bits(hqd_gpu_addr); 3012 3013 /* set up hqd_rptr_addr/_hi, similar as CP_RB_RPTR */ 3014 wb_gpu_addr = prop->rptr_gpu_addr; 3015 mqd->cp_gfx_hqd_rptr_addr = wb_gpu_addr & 0xfffffffc; 3016 mqd->cp_gfx_hqd_rptr_addr_hi = 3017 upper_32_bits(wb_gpu_addr) & 0xffff; 3018 3019 /* set up rb_wptr_poll addr */ 3020 wb_gpu_addr = prop->wptr_gpu_addr; 3021 mqd->cp_rb_wptr_poll_addr_lo = wb_gpu_addr & 0xfffffffc; 3022 mqd->cp_rb_wptr_poll_addr_hi = upper_32_bits(wb_gpu_addr) & 0xffff; 3023 3024 /* set up the gfx_hqd_control, similar as CP_RB0_CNTL */ 3025 rb_bufsz = order_base_2(prop->queue_size / 4) - 1; 3026 tmp = regCP_GFX_HQD_CNTL_DEFAULT; 3027 tmp = REG_SET_FIELD(tmp, CP_GFX_HQD_CNTL, RB_BUFSZ, rb_bufsz); 3028 tmp = REG_SET_FIELD(tmp, CP_GFX_HQD_CNTL, RB_BLKSZ, rb_bufsz - 2); 3029 #ifdef __BIG_ENDIAN 3030 tmp = REG_SET_FIELD(tmp, CP_GFX_HQD_CNTL, BUF_SWAP, 1); 3031 #endif 3032 if (prop->tmz_queue) 3033 tmp = REG_SET_FIELD(tmp, CP_GFX_HQD_CNTL, TMZ_MATCH, 1); 3034 if (!prop->kernel_queue) 3035 tmp = REG_SET_FIELD(tmp, CP_GFX_HQD_CNTL, RB_NON_PRIV, 1); 3036 mqd->cp_gfx_hqd_cntl = tmp; 3037 3038 /* set up cp_doorbell_control */ 3039 tmp = regCP_RB_DOORBELL_CONTROL_DEFAULT; 3040 if (prop->use_doorbell) { 3041 tmp = REG_SET_FIELD(tmp, CP_RB_DOORBELL_CONTROL, 3042 DOORBELL_OFFSET, prop->doorbell_index); 3043 tmp = REG_SET_FIELD(tmp, CP_RB_DOORBELL_CONTROL, 3044 DOORBELL_EN, 1); 3045 } else 3046 tmp = REG_SET_FIELD(tmp, CP_RB_DOORBELL_CONTROL, 3047 DOORBELL_EN, 0); 3048 mqd->cp_rb_doorbell_control = tmp; 3049 3050 /* reset read and write pointers, similar to CP_RB0_WPTR/_RPTR */ 3051 mqd->cp_gfx_hqd_rptr = regCP_GFX_HQD_RPTR_DEFAULT; 3052 3053 /* active the queue */ 3054 mqd->cp_gfx_hqd_active = 1; 3055 3056 /* set gfx UQ items */ 3057 mqd->shadow_base_lo = lower_32_bits(prop->shadow_addr); 3058 mqd->shadow_base_hi = upper_32_bits(prop->shadow_addr); 3059 mqd->fw_work_area_base_lo = lower_32_bits(prop->csa_addr); 3060 mqd->fw_work_area_base_hi = upper_32_bits(prop->csa_addr); 3061 mqd->fence_address_lo = lower_32_bits(prop->fence_address); 3062 mqd->fence_address_hi = upper_32_bits(prop->fence_address); 3063 3064 return 0; 3065 } 3066 3067 static int gfx_v12_0_kgq_init_queue(struct amdgpu_ring *ring, bool reset) 3068 { 3069 struct amdgpu_device *adev = ring->adev; 3070 struct v12_gfx_mqd *mqd = ring->mqd_ptr; 3071 int mqd_idx = ring - &adev->gfx.gfx_ring[0]; 3072 3073 if (!reset && !amdgpu_in_reset(adev) && !adev->in_suspend) { 3074 memset((void *)mqd, 0, sizeof(*mqd)); 3075 mutex_lock(&adev->srbm_mutex); 3076 soc24_grbm_select(adev, ring->me, ring->pipe, ring->queue, 0); 3077 amdgpu_ring_init_mqd(ring); 3078 soc24_grbm_select(adev, 0, 0, 0, 0); 3079 mutex_unlock(&adev->srbm_mutex); 3080 if (adev->gfx.me.mqd_backup[mqd_idx]) 3081 memcpy_fromio(adev->gfx.me.mqd_backup[mqd_idx], mqd, sizeof(*mqd)); 3082 } else { 3083 /* restore mqd with the backup copy */ 3084 if (adev->gfx.me.mqd_backup[mqd_idx]) 3085 memcpy_toio(mqd, adev->gfx.me.mqd_backup[mqd_idx], sizeof(*mqd)); 3086 /* reset the ring */ 3087 ring->wptr = 0; 3088 atomic64_set((atomic64_t *)ring->wptr_cpu_addr, 0); 3089 amdgpu_ring_clear_ring(ring); 3090 } 3091 3092 return 0; 3093 } 3094 3095 static int gfx_v12_0_cp_async_gfx_ring_resume(struct amdgpu_device *adev) 3096 { 3097 int i, r; 3098 3099 for (i = 0; i < adev->gfx.num_gfx_rings; i++) { 3100 r = gfx_v12_0_kgq_init_queue(&adev->gfx.gfx_ring[i], false); 3101 if (r) 3102 return r; 3103 } 3104 3105 r = amdgpu_gfx_enable_kgq(adev, 0); 3106 if (r) 3107 return r; 3108 3109 return gfx_v12_0_cp_gfx_start(adev); 3110 } 3111 3112 static void gfx_v12_0_compute_mqd_set_cu_mask(struct amdgpu_device *adev, 3113 struct v12_compute_mqd *mqd, 3114 struct amdgpu_mqd_prop *prop) 3115 { 3116 uint32_t se_mask[8] = {0}; 3117 uint32_t wa_mask; 3118 bool has_wa_flag = prop->cu_flags & (AMDGPU_UPDATE_FLAG_DBG_WA_ENABLE | 3119 AMDGPU_UPDATE_FLAG_DBG_WA_DISABLE); 3120 3121 if (!has_wa_flag && (!prop->cu_mask || !prop->cu_mask_count)) 3122 return; 3123 3124 if (has_wa_flag) { 3125 wa_mask = (prop->cu_flags & AMDGPU_UPDATE_FLAG_DBG_WA_ENABLE) ? 3126 0xffff : 0xffffffff; 3127 mqd->compute_static_thread_mgmt_se0 = wa_mask; 3128 mqd->compute_static_thread_mgmt_se1 = wa_mask; 3129 mqd->compute_static_thread_mgmt_se2 = wa_mask; 3130 mqd->compute_static_thread_mgmt_se3 = wa_mask; 3131 return; 3132 } 3133 3134 amdgpu_gfx_mqd_symmetrically_map_cu_mask(adev, prop->cu_mask, 3135 prop->cu_mask_count, se_mask); 3136 3137 mqd->compute_static_thread_mgmt_se0 = se_mask[0]; 3138 mqd->compute_static_thread_mgmt_se1 = se_mask[1]; 3139 mqd->compute_static_thread_mgmt_se2 = se_mask[2]; 3140 mqd->compute_static_thread_mgmt_se3 = se_mask[3]; 3141 } 3142 3143 static int gfx_v12_0_compute_mqd_init(struct amdgpu_device *adev, void *m, 3144 struct amdgpu_mqd_prop *prop) 3145 { 3146 struct v12_compute_mqd *mqd = m; 3147 uint64_t hqd_gpu_addr, wb_gpu_addr, eop_base_addr; 3148 uint32_t tmp; 3149 3150 mqd->header = 0xC0310800; 3151 mqd->compute_pipelinestat_enable = 0x00000001; 3152 mqd->compute_static_thread_mgmt_se0 = 0xffffffff; 3153 mqd->compute_static_thread_mgmt_se1 = 0xffffffff; 3154 mqd->compute_static_thread_mgmt_se2 = 0xffffffff; 3155 mqd->compute_static_thread_mgmt_se3 = 0xffffffff; 3156 mqd->compute_misc_reserved = 0x00000007; 3157 3158 eop_base_addr = prop->eop_gpu_addr >> 8; 3159 mqd->cp_hqd_eop_base_addr_lo = eop_base_addr; 3160 mqd->cp_hqd_eop_base_addr_hi = upper_32_bits(eop_base_addr); 3161 3162 /* set the EOP size, register value is 2^(EOP_SIZE+1) dwords */ 3163 tmp = regCP_HQD_EOP_CONTROL_DEFAULT; 3164 tmp = REG_SET_FIELD(tmp, CP_HQD_EOP_CONTROL, EOP_SIZE, 3165 (order_base_2(GFX12_MEC_HPD_SIZE / 4) - 1)); 3166 3167 mqd->cp_hqd_eop_control = tmp; 3168 3169 /* enable doorbell? */ 3170 tmp = regCP_HQD_PQ_DOORBELL_CONTROL_DEFAULT; 3171 3172 if (prop->use_doorbell) { 3173 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_DOORBELL_CONTROL, 3174 DOORBELL_OFFSET, prop->doorbell_index); 3175 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_DOORBELL_CONTROL, 3176 DOORBELL_EN, 1); 3177 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_DOORBELL_CONTROL, 3178 DOORBELL_SOURCE, 0); 3179 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_DOORBELL_CONTROL, 3180 DOORBELL_HIT, 0); 3181 } else { 3182 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_DOORBELL_CONTROL, 3183 DOORBELL_EN, 0); 3184 } 3185 3186 mqd->cp_hqd_pq_doorbell_control = tmp; 3187 3188 /* disable the queue if it's active */ 3189 mqd->cp_hqd_dequeue_request = 0; 3190 mqd->cp_hqd_pq_rptr = 0; 3191 mqd->cp_hqd_pq_wptr_lo = 0; 3192 mqd->cp_hqd_pq_wptr_hi = 0; 3193 3194 /* set the pointer to the MQD */ 3195 mqd->cp_mqd_base_addr_lo = prop->mqd_gpu_addr & 0xfffffffc; 3196 mqd->cp_mqd_base_addr_hi = upper_32_bits(prop->mqd_gpu_addr); 3197 3198 /* set MQD vmid to 0 */ 3199 tmp = regCP_MQD_CONTROL_DEFAULT; 3200 tmp = REG_SET_FIELD(tmp, CP_MQD_CONTROL, VMID, 0); 3201 mqd->cp_mqd_control = tmp; 3202 3203 /* set the pointer to the HQD, this is similar CP_RB0_BASE/_HI */ 3204 hqd_gpu_addr = prop->hqd_base_gpu_addr >> 8; 3205 mqd->cp_hqd_pq_base_lo = hqd_gpu_addr; 3206 mqd->cp_hqd_pq_base_hi = upper_32_bits(hqd_gpu_addr); 3207 3208 /* set up the HQD, this is similar to CP_RB0_CNTL */ 3209 tmp = regCP_HQD_PQ_CONTROL_DEFAULT; 3210 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_CONTROL, QUEUE_SIZE, 3211 (order_base_2(prop->queue_size / 4) - 1)); 3212 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_CONTROL, RPTR_BLOCK_SIZE, 3213 (order_base_2(AMDGPU_GPU_PAGE_SIZE / 4) - 1)); 3214 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_CONTROL, UNORD_DISPATCH, 1); 3215 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_CONTROL, TUNNEL_DISPATCH, 0); 3216 if (prop->kernel_queue) { 3217 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_CONTROL, PRIV_STATE, 1); 3218 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_CONTROL, KMD_QUEUE, 1); 3219 } 3220 if (prop->tmz_queue) 3221 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_CONTROL, TMZ, 1); 3222 mqd->cp_hqd_pq_control = tmp; 3223 3224 /* set the wb address whether it's enabled or not */ 3225 wb_gpu_addr = prop->rptr_gpu_addr; 3226 mqd->cp_hqd_pq_rptr_report_addr_lo = wb_gpu_addr & 0xfffffffc; 3227 mqd->cp_hqd_pq_rptr_report_addr_hi = 3228 upper_32_bits(wb_gpu_addr) & 0xffff; 3229 3230 /* only used if CP_PQ_WPTR_POLL_CNTL.CP_PQ_WPTR_POLL_CNTL__EN_MASK=1 */ 3231 wb_gpu_addr = prop->wptr_gpu_addr; 3232 mqd->cp_hqd_pq_wptr_poll_addr_lo = wb_gpu_addr & 0xfffffffc; 3233 mqd->cp_hqd_pq_wptr_poll_addr_hi = upper_32_bits(wb_gpu_addr) & 0xffff; 3234 3235 tmp = 0; 3236 /* enable the doorbell if requested */ 3237 if (prop->use_doorbell) { 3238 tmp = regCP_HQD_PQ_DOORBELL_CONTROL_DEFAULT; 3239 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_DOORBELL_CONTROL, 3240 DOORBELL_OFFSET, prop->doorbell_index); 3241 3242 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_DOORBELL_CONTROL, 3243 DOORBELL_EN, 1); 3244 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_DOORBELL_CONTROL, 3245 DOORBELL_SOURCE, 0); 3246 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_DOORBELL_CONTROL, 3247 DOORBELL_HIT, 0); 3248 } 3249 3250 mqd->cp_hqd_pq_doorbell_control = tmp; 3251 3252 /* reset read and write pointers, similar to CP_RB0_WPTR/_RPTR */ 3253 mqd->cp_hqd_pq_rptr = regCP_HQD_PQ_RPTR_DEFAULT; 3254 3255 /* set the vmid for the queue */ 3256 mqd->cp_hqd_vmid = 0; 3257 3258 tmp = regCP_HQD_PERSISTENT_STATE_DEFAULT; 3259 tmp = REG_SET_FIELD(tmp, CP_HQD_PERSISTENT_STATE, PRELOAD_SIZE, 0x55); 3260 mqd->cp_hqd_persistent_state = tmp; 3261 3262 /* set MIN_IB_AVAIL_SIZE */ 3263 tmp = regCP_HQD_IB_CONTROL_DEFAULT; 3264 tmp = REG_SET_FIELD(tmp, CP_HQD_IB_CONTROL, MIN_IB_AVAIL_SIZE, 3); 3265 mqd->cp_hqd_ib_control = tmp; 3266 3267 /* set static priority for a compute queue/ring */ 3268 mqd->cp_hqd_pipe_priority = prop->hqd_pipe_priority; 3269 mqd->cp_hqd_queue_priority = prop->hqd_queue_priority; 3270 3271 mqd->cp_hqd_active = prop->hqd_active; 3272 3273 /* set UQ fenceaddress */ 3274 mqd->fence_address_lo = lower_32_bits(prop->fence_address); 3275 mqd->fence_address_hi = upper_32_bits(prop->fence_address); 3276 /* set CU mask */ 3277 gfx_v12_0_compute_mqd_set_cu_mask(adev, mqd, prop); 3278 3279 return 0; 3280 } 3281 3282 static int gfx_v12_0_kiq_init_register(struct amdgpu_ring *ring) 3283 { 3284 struct amdgpu_device *adev = ring->adev; 3285 struct v12_compute_mqd *mqd = ring->mqd_ptr; 3286 int j; 3287 3288 /* inactivate the queue */ 3289 if (amdgpu_sriov_vf(adev)) 3290 WREG32_SOC15(GC, 0, regCP_HQD_ACTIVE, 0); 3291 3292 /* disable wptr polling */ 3293 WREG32_FIELD15_PREREG(GC, 0, CP_PQ_WPTR_POLL_CNTL, EN, 0); 3294 3295 /* write the EOP addr */ 3296 WREG32_SOC15(GC, 0, regCP_HQD_EOP_BASE_ADDR, 3297 mqd->cp_hqd_eop_base_addr_lo); 3298 WREG32_SOC15(GC, 0, regCP_HQD_EOP_BASE_ADDR_HI, 3299 mqd->cp_hqd_eop_base_addr_hi); 3300 3301 /* set the EOP size, register value is 2^(EOP_SIZE+1) dwords */ 3302 WREG32_SOC15(GC, 0, regCP_HQD_EOP_CONTROL, 3303 mqd->cp_hqd_eop_control); 3304 3305 /* enable doorbell? */ 3306 WREG32_SOC15(GC, 0, regCP_HQD_PQ_DOORBELL_CONTROL, 3307 mqd->cp_hqd_pq_doorbell_control); 3308 3309 /* disable the queue if it's active */ 3310 if (RREG32_SOC15(GC, 0, regCP_HQD_ACTIVE) & 1) { 3311 WREG32_SOC15(GC, 0, regCP_HQD_DEQUEUE_REQUEST, 1); 3312 for (j = 0; j < adev->usec_timeout; j++) { 3313 if (!(RREG32_SOC15(GC, 0, regCP_HQD_ACTIVE) & 1)) 3314 break; 3315 udelay(1); 3316 } 3317 WREG32_SOC15(GC, 0, regCP_HQD_DEQUEUE_REQUEST, 3318 mqd->cp_hqd_dequeue_request); 3319 WREG32_SOC15(GC, 0, regCP_HQD_PQ_RPTR, 3320 mqd->cp_hqd_pq_rptr); 3321 WREG32_SOC15(GC, 0, regCP_HQD_PQ_WPTR_LO, 3322 mqd->cp_hqd_pq_wptr_lo); 3323 WREG32_SOC15(GC, 0, regCP_HQD_PQ_WPTR_HI, 3324 mqd->cp_hqd_pq_wptr_hi); 3325 } 3326 3327 /* set the pointer to the MQD */ 3328 WREG32_SOC15(GC, 0, regCP_MQD_BASE_ADDR, 3329 mqd->cp_mqd_base_addr_lo); 3330 WREG32_SOC15(GC, 0, regCP_MQD_BASE_ADDR_HI, 3331 mqd->cp_mqd_base_addr_hi); 3332 3333 /* set MQD vmid to 0 */ 3334 WREG32_SOC15(GC, 0, regCP_MQD_CONTROL, 3335 mqd->cp_mqd_control); 3336 3337 /* set the pointer to the HQD, this is similar CP_RB0_BASE/_HI */ 3338 WREG32_SOC15(GC, 0, regCP_HQD_PQ_BASE, 3339 mqd->cp_hqd_pq_base_lo); 3340 WREG32_SOC15(GC, 0, regCP_HQD_PQ_BASE_HI, 3341 mqd->cp_hqd_pq_base_hi); 3342 3343 /* set up the HQD, this is similar to CP_RB0_CNTL */ 3344 WREG32_SOC15(GC, 0, regCP_HQD_PQ_CONTROL, 3345 mqd->cp_hqd_pq_control); 3346 3347 /* set the wb address whether it's enabled or not */ 3348 WREG32_SOC15(GC, 0, regCP_HQD_PQ_RPTR_REPORT_ADDR, 3349 mqd->cp_hqd_pq_rptr_report_addr_lo); 3350 WREG32_SOC15(GC, 0, regCP_HQD_PQ_RPTR_REPORT_ADDR_HI, 3351 mqd->cp_hqd_pq_rptr_report_addr_hi); 3352 3353 /* only used if CP_PQ_WPTR_POLL_CNTL.CP_PQ_WPTR_POLL_CNTL__EN_MASK=1 */ 3354 WREG32_SOC15(GC, 0, regCP_HQD_PQ_WPTR_POLL_ADDR, 3355 mqd->cp_hqd_pq_wptr_poll_addr_lo); 3356 WREG32_SOC15(GC, 0, regCP_HQD_PQ_WPTR_POLL_ADDR_HI, 3357 mqd->cp_hqd_pq_wptr_poll_addr_hi); 3358 3359 /* enable the doorbell if requested */ 3360 if (ring->use_doorbell) { 3361 WREG32_SOC15(GC, 0, regCP_MEC_DOORBELL_RANGE_LOWER, 3362 (adev->doorbell_index.kiq * 2) << 2); 3363 WREG32_SOC15(GC, 0, regCP_MEC_DOORBELL_RANGE_UPPER, 3364 (adev->doorbell_index.userqueue_end * 2) << 2); 3365 } 3366 3367 WREG32_SOC15(GC, 0, regCP_HQD_PQ_DOORBELL_CONTROL, 3368 mqd->cp_hqd_pq_doorbell_control); 3369 3370 /* reset read and write pointers, similar to CP_RB0_WPTR/_RPTR */ 3371 WREG32_SOC15(GC, 0, regCP_HQD_PQ_WPTR_LO, 3372 mqd->cp_hqd_pq_wptr_lo); 3373 WREG32_SOC15(GC, 0, regCP_HQD_PQ_WPTR_HI, 3374 mqd->cp_hqd_pq_wptr_hi); 3375 3376 /* set the vmid for the queue */ 3377 WREG32_SOC15(GC, 0, regCP_HQD_VMID, mqd->cp_hqd_vmid); 3378 3379 WREG32_SOC15(GC, 0, regCP_HQD_PERSISTENT_STATE, 3380 mqd->cp_hqd_persistent_state); 3381 3382 /* activate the queue */ 3383 WREG32_SOC15(GC, 0, regCP_HQD_ACTIVE, 3384 mqd->cp_hqd_active); 3385 3386 if (ring->use_doorbell) 3387 WREG32_FIELD15_PREREG(GC, 0, CP_PQ_STATUS, DOORBELL_ENABLE, 1); 3388 3389 return 0; 3390 } 3391 3392 static int gfx_v12_0_kiq_init_queue(struct amdgpu_ring *ring) 3393 { 3394 struct amdgpu_device *adev = ring->adev; 3395 struct v12_compute_mqd *mqd = ring->mqd_ptr; 3396 int mqd_idx = AMDGPU_MAX_COMPUTE_RINGS; 3397 3398 gfx_v12_0_kiq_setting(ring); 3399 3400 if (amdgpu_in_reset(adev)) { /* for GPU_RESET case */ 3401 /* reset MQD to a clean status */ 3402 if (adev->gfx.mec.mqd_backup[mqd_idx]) 3403 memcpy(mqd, adev->gfx.mec.mqd_backup[mqd_idx], sizeof(*mqd)); 3404 3405 /* reset ring buffer */ 3406 ring->wptr = 0; 3407 amdgpu_ring_clear_ring(ring); 3408 3409 mutex_lock(&adev->srbm_mutex); 3410 soc24_grbm_select(adev, ring->me, ring->pipe, ring->queue, 0); 3411 gfx_v12_0_kiq_init_register(ring); 3412 soc24_grbm_select(adev, 0, 0, 0, 0); 3413 mutex_unlock(&adev->srbm_mutex); 3414 } else { 3415 memset((void *)mqd, 0, sizeof(*mqd)); 3416 if (amdgpu_sriov_vf(adev) && adev->in_suspend) 3417 amdgpu_ring_clear_ring(ring); 3418 mutex_lock(&adev->srbm_mutex); 3419 soc24_grbm_select(adev, ring->me, ring->pipe, ring->queue, 0); 3420 amdgpu_ring_init_mqd(ring); 3421 gfx_v12_0_kiq_init_register(ring); 3422 soc24_grbm_select(adev, 0, 0, 0, 0); 3423 mutex_unlock(&adev->srbm_mutex); 3424 3425 if (adev->gfx.mec.mqd_backup[mqd_idx]) 3426 memcpy(adev->gfx.mec.mqd_backup[mqd_idx], mqd, sizeof(*mqd)); 3427 } 3428 3429 return 0; 3430 } 3431 3432 static int gfx_v12_0_kcq_init_queue(struct amdgpu_ring *ring, bool reset) 3433 { 3434 struct amdgpu_device *adev = ring->adev; 3435 struct v12_compute_mqd *mqd = ring->mqd_ptr; 3436 int mqd_idx = ring - &adev->gfx.compute_ring[0]; 3437 3438 if (!reset && !amdgpu_in_reset(adev) && !adev->in_suspend) { 3439 memset((void *)mqd, 0, sizeof(*mqd)); 3440 mutex_lock(&adev->srbm_mutex); 3441 soc24_grbm_select(adev, ring->me, ring->pipe, ring->queue, 0); 3442 amdgpu_ring_init_mqd(ring); 3443 soc24_grbm_select(adev, 0, 0, 0, 0); 3444 mutex_unlock(&adev->srbm_mutex); 3445 3446 if (adev->gfx.mec.mqd_backup[mqd_idx]) 3447 memcpy_fromio(adev->gfx.mec.mqd_backup[mqd_idx], mqd, sizeof(*mqd)); 3448 } else { 3449 /* restore MQD to a clean status */ 3450 if (adev->gfx.mec.mqd_backup[mqd_idx]) 3451 memcpy_toio(mqd, adev->gfx.mec.mqd_backup[mqd_idx], sizeof(*mqd)); 3452 /* reset ring buffer */ 3453 ring->wptr = 0; 3454 atomic64_set((atomic64_t *)ring->wptr_cpu_addr, 0); 3455 amdgpu_ring_clear_ring(ring); 3456 } 3457 3458 return 0; 3459 } 3460 3461 static int gfx_v12_0_kiq_resume(struct amdgpu_device *adev) 3462 { 3463 gfx_v12_0_kiq_init_queue(&adev->gfx.kiq[0].ring); 3464 adev->gfx.kiq[0].ring.sched.ready = true; 3465 return 0; 3466 } 3467 3468 static int gfx_v12_0_kcq_resume(struct amdgpu_device *adev) 3469 { 3470 int i, r; 3471 3472 if (!amdgpu_async_gfx_ring) 3473 gfx_v12_0_cp_compute_enable(adev, true); 3474 3475 for (i = 0; i < adev->gfx.num_compute_rings; i++) { 3476 r = gfx_v12_0_kcq_init_queue(&adev->gfx.compute_ring[i], false); 3477 if (r) 3478 return r; 3479 } 3480 3481 return amdgpu_gfx_enable_kcq(adev, 0); 3482 } 3483 3484 static int gfx_v12_0_cp_resume(struct amdgpu_device *adev) 3485 { 3486 int r, i; 3487 struct amdgpu_ring *ring; 3488 3489 if (!(adev->flags & AMD_IS_APU)) 3490 gfx_v12_0_enable_gui_idle_interrupt(adev, false); 3491 3492 if (adev->firmware.load_type == AMDGPU_FW_LOAD_DIRECT) { 3493 /* legacy firmware loading */ 3494 r = gfx_v12_0_cp_gfx_load_microcode(adev); 3495 if (r) 3496 return r; 3497 3498 r = gfx_v12_0_cp_compute_load_microcode_rs64(adev); 3499 if (r) 3500 return r; 3501 } 3502 3503 gfx_v12_0_cp_set_doorbell_range(adev); 3504 3505 if (amdgpu_async_gfx_ring) { 3506 gfx_v12_0_cp_compute_enable(adev, true); 3507 gfx_v12_0_cp_gfx_enable(adev, true); 3508 } 3509 3510 if (adev->enable_mes_kiq && adev->mes.kiq_hw_init) 3511 r = amdgpu_mes_kiq_hw_init(adev, 0); 3512 else 3513 r = gfx_v12_0_kiq_resume(adev); 3514 if (r) 3515 return r; 3516 3517 r = gfx_v12_0_kcq_resume(adev); 3518 if (r) 3519 return r; 3520 3521 if (!amdgpu_async_gfx_ring) { 3522 r = gfx_v12_0_cp_gfx_resume(adev); 3523 if (r) 3524 return r; 3525 } else { 3526 r = gfx_v12_0_cp_async_gfx_ring_resume(adev); 3527 if (r) 3528 return r; 3529 } 3530 3531 for (i = 0; i < adev->gfx.num_gfx_rings; i++) { 3532 ring = &adev->gfx.gfx_ring[i]; 3533 r = amdgpu_ring_test_helper(ring); 3534 if (r) 3535 return r; 3536 } 3537 3538 for (i = 0; i < adev->gfx.num_compute_rings; i++) { 3539 ring = &adev->gfx.compute_ring[i]; 3540 r = amdgpu_ring_test_helper(ring); 3541 if (r) 3542 return r; 3543 } 3544 3545 return 0; 3546 } 3547 3548 static void gfx_v12_0_cp_enable(struct amdgpu_device *adev, bool enable) 3549 { 3550 gfx_v12_0_cp_gfx_enable(adev, enable); 3551 gfx_v12_0_cp_compute_enable(adev, enable); 3552 } 3553 3554 static int gfx_v12_0_gfxhub_enable(struct amdgpu_device *adev) 3555 { 3556 int r; 3557 bool value; 3558 3559 r = adev->gfxhub.funcs->gart_enable(adev); 3560 if (r) 3561 return r; 3562 3563 amdgpu_device_flush_hdp(adev, NULL); 3564 3565 value = amdgpu_vm_fault_stop != AMDGPU_VM_FAULT_STOP_ALWAYS; 3566 3567 adev->gfxhub.funcs->set_fault_enable_default(adev, value); 3568 /* TODO investigate why this and the hdp flush above is needed, 3569 * are we missing a flush somewhere else? */ 3570 adev->gmc.gmc_funcs->flush_gpu_tlb(adev, 0, AMDGPU_GFXHUB(0), 0); 3571 3572 return 0; 3573 } 3574 3575 static int get_gb_addr_config(struct amdgpu_device *adev) 3576 { 3577 u32 gb_addr_config; 3578 3579 gb_addr_config = RREG32_SOC15(GC, 0, regGB_ADDR_CONFIG); 3580 if (gb_addr_config == 0) 3581 return -EINVAL; 3582 3583 adev->gfx.config.gb_addr_config_fields.num_pkrs = 3584 1 << REG_GET_FIELD(gb_addr_config, GB_ADDR_CONFIG, NUM_PKRS); 3585 3586 adev->gfx.config.gb_addr_config = gb_addr_config; 3587 3588 adev->gfx.config.gb_addr_config_fields.num_pipes = 1 << 3589 REG_GET_FIELD(adev->gfx.config.gb_addr_config, 3590 GB_ADDR_CONFIG, NUM_PIPES); 3591 3592 adev->gfx.config.max_tile_pipes = 3593 adev->gfx.config.gb_addr_config_fields.num_pipes; 3594 3595 adev->gfx.config.gb_addr_config_fields.max_compress_frags = 1 << 3596 REG_GET_FIELD(adev->gfx.config.gb_addr_config, 3597 GB_ADDR_CONFIG, MAX_COMPRESSED_FRAGS); 3598 adev->gfx.config.gb_addr_config_fields.num_rb_per_se = 1 << 3599 REG_GET_FIELD(adev->gfx.config.gb_addr_config, 3600 GB_ADDR_CONFIG, NUM_RB_PER_SE); 3601 adev->gfx.config.gb_addr_config_fields.num_se = 1 << 3602 REG_GET_FIELD(adev->gfx.config.gb_addr_config, 3603 GB_ADDR_CONFIG, NUM_SHADER_ENGINES); 3604 adev->gfx.config.gb_addr_config_fields.pipe_interleave_size = 1 << (8 + 3605 REG_GET_FIELD(adev->gfx.config.gb_addr_config, 3606 GB_ADDR_CONFIG, PIPE_INTERLEAVE_SIZE)); 3607 3608 return 0; 3609 } 3610 3611 static void gfx_v12_0_disable_gpa_mode(struct amdgpu_device *adev) 3612 { 3613 uint32_t data; 3614 3615 data = RREG32_SOC15(GC, 0, regCPC_PSP_DEBUG); 3616 data |= CPC_PSP_DEBUG__GPA_OVERRIDE_MASK; 3617 WREG32_SOC15(GC, 0, regCPC_PSP_DEBUG, data); 3618 3619 data = RREG32_SOC15(GC, 0, regCPG_PSP_DEBUG); 3620 data |= CPG_PSP_DEBUG__GPA_OVERRIDE_MASK; 3621 WREG32_SOC15(GC, 0, regCPG_PSP_DEBUG, data); 3622 } 3623 3624 static void gfx_v12_0_init_golden_registers(struct amdgpu_device *adev) 3625 { 3626 if (amdgpu_sriov_vf(adev)) 3627 return; 3628 3629 switch (amdgpu_ip_version(adev, GC_HWIP, 0)) { 3630 case IP_VERSION(12, 0, 0): 3631 case IP_VERSION(12, 0, 1): 3632 soc15_program_register_sequence(adev, 3633 golden_settings_gc_12_0, 3634 (const u32)ARRAY_SIZE(golden_settings_gc_12_0)); 3635 3636 if (adev->rev_id == 0) 3637 soc15_program_register_sequence(adev, 3638 golden_settings_gc_12_0_rev0, 3639 (const u32)ARRAY_SIZE(golden_settings_gc_12_0_rev0)); 3640 break; 3641 default: 3642 break; 3643 } 3644 } 3645 3646 static int gfx_v12_0_hw_init(struct amdgpu_ip_block *ip_block) 3647 { 3648 int r; 3649 struct amdgpu_device *adev = ip_block->adev; 3650 3651 if (adev->firmware.load_type == AMDGPU_FW_LOAD_RLC_BACKDOOR_AUTO) { 3652 if (adev->gfx.imu.funcs && (amdgpu_dpm > 0)) { 3653 /* RLC autoload sequence 1: Program rlc ram */ 3654 if (adev->gfx.imu.funcs->program_rlc_ram) 3655 adev->gfx.imu.funcs->program_rlc_ram(adev); 3656 } 3657 /* rlc autoload firmware */ 3658 r = gfx_v12_0_rlc_backdoor_autoload_enable(adev); 3659 if (r) 3660 return r; 3661 } else { 3662 if (adev->firmware.load_type == AMDGPU_FW_LOAD_DIRECT) { 3663 if (adev->gfx.imu.funcs && (amdgpu_dpm > 0)) { 3664 if (adev->gfx.imu.funcs->load_microcode) 3665 adev->gfx.imu.funcs->load_microcode(adev); 3666 if (adev->gfx.imu.funcs->setup_imu) 3667 adev->gfx.imu.funcs->setup_imu(adev); 3668 if (adev->gfx.imu.funcs->start_imu) 3669 adev->gfx.imu.funcs->start_imu(adev); 3670 } 3671 3672 /* disable gpa mode in backdoor loading */ 3673 gfx_v12_0_disable_gpa_mode(adev); 3674 } 3675 } 3676 3677 if ((adev->firmware.load_type == AMDGPU_FW_LOAD_RLC_BACKDOOR_AUTO) || 3678 (adev->firmware.load_type == AMDGPU_FW_LOAD_PSP)) { 3679 r = gfx_v12_0_wait_for_rlc_autoload_complete(adev); 3680 if (r) { 3681 dev_err(adev->dev, "(%d) failed to wait rlc autoload complete\n", r); 3682 return r; 3683 } 3684 } 3685 3686 if (!amdgpu_emu_mode) 3687 gfx_v12_0_init_golden_registers(adev); 3688 3689 adev->gfx.is_poweron = true; 3690 3691 if (get_gb_addr_config(adev)) 3692 drm_warn(adev_to_drm(adev), "Invalid gb_addr_config !\n"); 3693 3694 if (adev->firmware.load_type == AMDGPU_FW_LOAD_PSP) 3695 gfx_v12_0_config_gfx_rs64(adev); 3696 3697 r = gfx_v12_0_gfxhub_enable(adev); 3698 if (r) 3699 return r; 3700 3701 if ((adev->firmware.load_type == AMDGPU_FW_LOAD_DIRECT || 3702 adev->firmware.load_type == AMDGPU_FW_LOAD_RLC_BACKDOOR_AUTO) && 3703 (amdgpu_dpm == 1)) { 3704 /** 3705 * For gfx 12, rlc firmware loading relies on smu firmware is 3706 * loaded firstly, so in direct type, it has to load smc ucode 3707 * here before rlc. 3708 */ 3709 r = amdgpu_pm_load_smu_firmware(adev, NULL); 3710 if (r) 3711 return r; 3712 } 3713 3714 gfx_v12_0_constants_init(adev); 3715 3716 if (adev->nbio.funcs->gc_doorbell_init) 3717 adev->nbio.funcs->gc_doorbell_init(adev); 3718 3719 r = gfx_v12_0_rlc_resume(adev); 3720 if (r) 3721 return r; 3722 3723 /* 3724 * init golden registers and rlc resume may override some registers, 3725 * reconfig them here 3726 */ 3727 gfx_v12_0_tcp_harvest(adev); 3728 3729 r = gfx_v12_0_cp_resume(adev); 3730 if (r) 3731 return r; 3732 3733 return r; 3734 } 3735 3736 static int gfx_v12_0_set_userq_eop_interrupts(struct amdgpu_device *adev, 3737 bool enable) 3738 { 3739 unsigned int irq_type; 3740 int m, p, r; 3741 3742 if (adev->userq_funcs[AMDGPU_HW_IP_GFX]) { 3743 for (m = 0; m < adev->gfx.me.num_me; m++) { 3744 for (p = 0; p < adev->gfx.me.num_pipe_per_me; p++) { 3745 irq_type = AMDGPU_CP_IRQ_GFX_ME0_PIPE0_EOP + p; 3746 if (enable) 3747 r = amdgpu_irq_get(adev, &adev->gfx.eop_irq, 3748 irq_type); 3749 else 3750 r = amdgpu_irq_put(adev, &adev->gfx.eop_irq, 3751 irq_type); 3752 if (r) 3753 return r; 3754 } 3755 } 3756 } 3757 3758 if (adev->userq_funcs[AMDGPU_HW_IP_COMPUTE]) { 3759 for (m = 0; m < adev->gfx.mec.num_mec; ++m) { 3760 for (p = 0; p < adev->gfx.mec.num_pipe_per_mec; p++) { 3761 irq_type = AMDGPU_CP_IRQ_COMPUTE_MEC1_PIPE0_EOP 3762 + (m * adev->gfx.mec.num_pipe_per_mec) 3763 + p; 3764 if (enable) 3765 r = amdgpu_irq_get(adev, &adev->gfx.eop_irq, 3766 irq_type); 3767 else 3768 r = amdgpu_irq_put(adev, &adev->gfx.eop_irq, 3769 irq_type); 3770 if (r) 3771 return r; 3772 } 3773 } 3774 } 3775 3776 return 0; 3777 } 3778 3779 static int gfx_v12_0_hw_fini(struct amdgpu_ip_block *ip_block) 3780 { 3781 struct amdgpu_device *adev = ip_block->adev; 3782 uint32_t tmp; 3783 3784 cancel_delayed_work_sync(&adev->gfx.idle_work); 3785 3786 amdgpu_irq_put(adev, &adev->gfx.priv_reg_irq, 0); 3787 amdgpu_irq_put(adev, &adev->gfx.priv_inst_irq, 0); 3788 amdgpu_irq_put(adev, &adev->gfx.bad_op_irq, 0); 3789 gfx_v12_0_set_userq_eop_interrupts(adev, false); 3790 3791 if (!adev->no_hw_access) { 3792 if (amdgpu_async_gfx_ring) { 3793 if (amdgpu_gfx_disable_kgq(adev, 0)) 3794 DRM_ERROR("KGQ disable failed\n"); 3795 } 3796 3797 if (amdgpu_gfx_disable_kcq(adev, 0)) 3798 DRM_ERROR("KCQ disable failed\n"); 3799 3800 amdgpu_mes_kiq_hw_fini(adev, 0); 3801 } 3802 3803 if (amdgpu_sriov_vf(adev)) { 3804 gfx_v12_0_cp_gfx_enable(adev, false); 3805 /* Program KIQ position of RLC_CP_SCHEDULERS during destroy */ 3806 tmp = RREG32_SOC15(GC, 0, regRLC_CP_SCHEDULERS); 3807 tmp &= 0xffffff00; 3808 WREG32_SOC15(GC, 0, regRLC_CP_SCHEDULERS, tmp); 3809 3810 return 0; 3811 } 3812 gfx_v12_0_cp_enable(adev, false); 3813 gfx_v12_0_enable_gui_idle_interrupt(adev, false); 3814 3815 adev->gfxhub.funcs->gart_disable(adev); 3816 3817 adev->gfx.is_poweron = false; 3818 3819 return 0; 3820 } 3821 3822 static int gfx_v12_0_suspend(struct amdgpu_ip_block *ip_block) 3823 { 3824 return gfx_v12_0_hw_fini(ip_block); 3825 } 3826 3827 static int gfx_v12_0_resume(struct amdgpu_ip_block *ip_block) 3828 { 3829 return gfx_v12_0_hw_init(ip_block); 3830 } 3831 3832 static bool gfx_v12_0_is_idle(struct amdgpu_ip_block *ip_block) 3833 { 3834 struct amdgpu_device *adev = ip_block->adev; 3835 3836 if (REG_GET_FIELD(RREG32_SOC15(GC, 0, regGRBM_STATUS), 3837 GRBM_STATUS, GUI_ACTIVE)) 3838 return false; 3839 else 3840 return true; 3841 } 3842 3843 static int gfx_v12_0_wait_for_idle(struct amdgpu_ip_block *ip_block) 3844 { 3845 unsigned i; 3846 u32 tmp; 3847 struct amdgpu_device *adev = ip_block->adev; 3848 3849 for (i = 0; i < adev->usec_timeout; i++) { 3850 /* read MC_STATUS */ 3851 tmp = RREG32_SOC15(GC, 0, regGRBM_STATUS) & 3852 GRBM_STATUS__GUI_ACTIVE_MASK; 3853 3854 if (!REG_GET_FIELD(tmp, GRBM_STATUS, GUI_ACTIVE)) 3855 return 0; 3856 udelay(1); 3857 } 3858 return -ETIMEDOUT; 3859 } 3860 3861 static uint64_t gfx_v12_0_get_gpu_clock_counter(struct amdgpu_device *adev) 3862 { 3863 uint64_t clock = 0; 3864 3865 if (adev->smuio.funcs && 3866 adev->smuio.funcs->get_gpu_clock_counter) 3867 clock = adev->smuio.funcs->get_gpu_clock_counter(adev); 3868 else 3869 dev_warn(adev->dev, "query gpu clock counter is not supported\n"); 3870 3871 return clock; 3872 } 3873 3874 static int gfx_v12_0_early_init(struct amdgpu_ip_block *ip_block) 3875 { 3876 struct amdgpu_device *adev = ip_block->adev; 3877 3878 switch (amdgpu_user_queue) { 3879 case -1: 3880 case 0: 3881 default: 3882 adev->gfx.disable_kq = false; 3883 adev->gfx.disable_uq = true; 3884 break; 3885 case 1: 3886 adev->gfx.disable_kq = false; 3887 adev->gfx.disable_uq = false; 3888 break; 3889 case 2: 3890 adev->gfx.disable_kq = true; 3891 adev->gfx.disable_uq = false; 3892 break; 3893 } 3894 3895 adev->gfx.funcs = &gfx_v12_0_gfx_funcs; 3896 3897 if (adev->gfx.disable_kq) { 3898 adev->gfx.num_gfx_rings = 0; 3899 adev->gfx.num_compute_rings = 0; 3900 } else { 3901 adev->gfx.num_gfx_rings = GFX12_NUM_GFX_RINGS; 3902 adev->gfx.num_compute_rings = min(amdgpu_gfx_get_num_kcq(adev), 3903 AMDGPU_MAX_COMPUTE_RINGS); 3904 } 3905 3906 gfx_v12_0_set_kiq_pm4_funcs(adev); 3907 gfx_v12_0_set_ring_funcs(adev); 3908 gfx_v12_0_set_irq_funcs(adev); 3909 gfx_v12_0_set_rlc_funcs(adev); 3910 gfx_v12_0_set_mqd_funcs(adev); 3911 gfx_v12_0_set_imu_funcs(adev); 3912 3913 gfx_v12_0_init_rlcg_reg_access_ctrl(adev); 3914 3915 return gfx_v12_0_init_microcode(adev); 3916 } 3917 3918 static int gfx_v12_0_late_init(struct amdgpu_ip_block *ip_block) 3919 { 3920 struct amdgpu_device *adev = ip_block->adev; 3921 int r; 3922 3923 r = amdgpu_irq_get(adev, &adev->gfx.priv_reg_irq, 0); 3924 if (r) 3925 return r; 3926 3927 r = amdgpu_irq_get(adev, &adev->gfx.priv_inst_irq, 0); 3928 if (r) 3929 return r; 3930 3931 r = amdgpu_irq_get(adev, &adev->gfx.bad_op_irq, 0); 3932 if (r) 3933 return r; 3934 3935 r = gfx_v12_0_set_userq_eop_interrupts(adev, true); 3936 if (r) 3937 return r; 3938 3939 return 0; 3940 } 3941 3942 static bool gfx_v12_0_is_rlc_enabled(struct amdgpu_device *adev) 3943 { 3944 uint32_t rlc_cntl; 3945 3946 /* if RLC is not enabled, do nothing */ 3947 rlc_cntl = RREG32_SOC15(GC, 0, regRLC_CNTL); 3948 return (REG_GET_FIELD(rlc_cntl, RLC_CNTL, RLC_ENABLE_F32)) ? true : false; 3949 } 3950 3951 static void gfx_v12_0_set_safe_mode(struct amdgpu_device *adev, 3952 int xcc_id) 3953 { 3954 uint32_t data; 3955 unsigned i; 3956 3957 data = RLC_SAFE_MODE__CMD_MASK; 3958 data |= (1 << RLC_SAFE_MODE__MESSAGE__SHIFT); 3959 3960 WREG32_SOC15(GC, 0, regRLC_SAFE_MODE, data); 3961 3962 /* wait for RLC_SAFE_MODE */ 3963 for (i = 0; i < adev->usec_timeout; i++) { 3964 if (!REG_GET_FIELD(RREG32_SOC15(GC, 0, regRLC_SAFE_MODE), 3965 RLC_SAFE_MODE, CMD)) 3966 break; 3967 udelay(1); 3968 } 3969 } 3970 3971 static void gfx_v12_0_unset_safe_mode(struct amdgpu_device *adev, 3972 int xcc_id) 3973 { 3974 WREG32_SOC15(GC, 0, regRLC_SAFE_MODE, RLC_SAFE_MODE__CMD_MASK); 3975 } 3976 3977 static void gfx_v12_0_update_perf_clk(struct amdgpu_device *adev, 3978 bool enable) 3979 { 3980 uint32_t def, data; 3981 3982 if (!(adev->cg_flags & AMD_CG_SUPPORT_GFX_PERF_CLK)) 3983 return; 3984 3985 def = data = RREG32_SOC15(GC, 0, regRLC_CGTT_MGCG_OVERRIDE); 3986 3987 if (enable) 3988 data &= ~RLC_CGTT_MGCG_OVERRIDE__PERFMON_CLOCK_STATE_MASK; 3989 else 3990 data |= RLC_CGTT_MGCG_OVERRIDE__PERFMON_CLOCK_STATE_MASK; 3991 3992 if (def != data) 3993 WREG32_SOC15(GC, 0, regRLC_CGTT_MGCG_OVERRIDE, data); 3994 } 3995 3996 static void gfx_v12_0_update_spm_vmid(struct amdgpu_device *adev, 3997 int xcc_id, 3998 struct amdgpu_ring *ring, 3999 unsigned vmid) 4000 { 4001 u32 reg, data; 4002 4003 reg = SOC15_REG_OFFSET(GC, 0, regRLC_SPM_MC_CNTL); 4004 if (amdgpu_sriov_is_pp_one_vf(adev)) 4005 data = RREG32_NO_KIQ(reg); 4006 else 4007 data = RREG32(reg); 4008 4009 data &= ~RLC_SPM_MC_CNTL__RLC_SPM_VMID_MASK; 4010 data |= (vmid & RLC_SPM_MC_CNTL__RLC_SPM_VMID_MASK) << RLC_SPM_MC_CNTL__RLC_SPM_VMID__SHIFT; 4011 4012 if (amdgpu_sriov_is_pp_one_vf(adev)) 4013 WREG32_SOC15_NO_KIQ(GC, 0, regRLC_SPM_MC_CNTL, data); 4014 else 4015 WREG32_SOC15(GC, 0, regRLC_SPM_MC_CNTL, data); 4016 4017 if (ring 4018 && amdgpu_sriov_is_pp_one_vf(adev) 4019 && ((ring->funcs->type == AMDGPU_RING_TYPE_GFX) 4020 || (ring->funcs->type == AMDGPU_RING_TYPE_COMPUTE))) { 4021 uint32_t reg = SOC15_REG_OFFSET(GC, 0, regRLC_SPM_MC_CNTL); 4022 amdgpu_ring_emit_wreg(ring, reg, data); 4023 } 4024 } 4025 4026 static const struct amdgpu_rlc_funcs gfx_v12_0_rlc_funcs = { 4027 .is_rlc_enabled = gfx_v12_0_is_rlc_enabled, 4028 .set_safe_mode = gfx_v12_0_set_safe_mode, 4029 .unset_safe_mode = gfx_v12_0_unset_safe_mode, 4030 .init = gfx_v12_0_rlc_init, 4031 .get_csb_size = gfx_v12_0_get_csb_size, 4032 .get_csb_buffer = gfx_v12_0_get_csb_buffer, 4033 .resume = gfx_v12_0_rlc_resume, 4034 .stop = gfx_v12_0_rlc_stop, 4035 .reset = gfx_v12_0_rlc_reset, 4036 .start = gfx_v12_0_rlc_start, 4037 .update_spm_vmid = gfx_v12_0_update_spm_vmid, 4038 }; 4039 4040 #if 0 4041 static void gfx_v12_cntl_power_gating(struct amdgpu_device *adev, bool enable) 4042 { 4043 /* TODO */ 4044 } 4045 4046 static void gfx_v12_cntl_pg(struct amdgpu_device *adev, bool enable) 4047 { 4048 /* TODO */ 4049 } 4050 #endif 4051 4052 static int gfx_v12_0_set_powergating_state(struct amdgpu_ip_block *ip_block, 4053 enum amd_powergating_state state) 4054 { 4055 struct amdgpu_device *adev = ip_block->adev; 4056 bool enable = (state == AMD_PG_STATE_GATE); 4057 4058 if (amdgpu_sriov_vf(adev)) 4059 return 0; 4060 4061 switch (amdgpu_ip_version(adev, GC_HWIP, 0)) { 4062 case IP_VERSION(12, 0, 0): 4063 case IP_VERSION(12, 0, 1): 4064 amdgpu_gfx_off_ctrl(adev, enable); 4065 break; 4066 default: 4067 break; 4068 } 4069 4070 return 0; 4071 } 4072 4073 static void gfx_v12_0_update_coarse_grain_clock_gating(struct amdgpu_device *adev, 4074 bool enable) 4075 { 4076 uint32_t def, data; 4077 4078 if (!(adev->cg_flags & 4079 (AMD_CG_SUPPORT_GFX_CGCG | 4080 AMD_CG_SUPPORT_GFX_CGLS | 4081 AMD_CG_SUPPORT_GFX_3D_CGCG | 4082 AMD_CG_SUPPORT_GFX_3D_CGLS))) 4083 return; 4084 4085 if (enable) { 4086 def = data = RREG32_SOC15(GC, 0, regRLC_CGTT_MGCG_OVERRIDE); 4087 4088 /* unset CGCG override */ 4089 if (adev->cg_flags & AMD_CG_SUPPORT_GFX_CGCG) 4090 data &= ~RLC_CGTT_MGCG_OVERRIDE__GFXIP_CGCG_OVERRIDE_MASK; 4091 if (adev->cg_flags & AMD_CG_SUPPORT_GFX_CGLS) 4092 data &= ~RLC_CGTT_MGCG_OVERRIDE__GFXIP_CGLS_OVERRIDE_MASK; 4093 if (adev->cg_flags & AMD_CG_SUPPORT_GFX_3D_CGCG || 4094 adev->cg_flags & AMD_CG_SUPPORT_GFX_3D_CGLS) 4095 data &= ~RLC_CGTT_MGCG_OVERRIDE__GFXIP_GFX3D_CG_OVERRIDE_MASK; 4096 4097 /* update CGCG override bits */ 4098 if (def != data) 4099 WREG32_SOC15(GC, 0, regRLC_CGTT_MGCG_OVERRIDE, data); 4100 4101 /* enable cgcg FSM(0x0000363F) */ 4102 def = data = RREG32_SOC15(GC, 0, regRLC_CGCG_CGLS_CTRL); 4103 4104 if (adev->cg_flags & AMD_CG_SUPPORT_GFX_CGCG) { 4105 data &= ~RLC_CGCG_CGLS_CTRL__CGCG_GFX_IDLE_THRESHOLD_MASK; 4106 data |= (0x36 << RLC_CGCG_CGLS_CTRL__CGCG_GFX_IDLE_THRESHOLD__SHIFT) | 4107 RLC_CGCG_CGLS_CTRL__CGCG_EN_MASK; 4108 } 4109 4110 if (adev->cg_flags & AMD_CG_SUPPORT_GFX_CGLS) { 4111 data &= ~RLC_CGCG_CGLS_CTRL__CGLS_REP_COMPANSAT_DELAY_MASK; 4112 data |= (0x000F << RLC_CGCG_CGLS_CTRL__CGLS_REP_COMPANSAT_DELAY__SHIFT) | 4113 RLC_CGCG_CGLS_CTRL__CGLS_EN_MASK; 4114 } 4115 4116 if (def != data) 4117 WREG32_SOC15(GC, 0, regRLC_CGCG_CGLS_CTRL, data); 4118 4119 /* Program RLC_CGCG_CGLS_CTRL_3D */ 4120 def = data = RREG32_SOC15(GC, 0, regRLC_CGCG_CGLS_CTRL_3D); 4121 4122 if (adev->cg_flags & AMD_CG_SUPPORT_GFX_3D_CGCG) { 4123 data &= ~RLC_CGCG_CGLS_CTRL_3D__CGCG_GFX_IDLE_THRESHOLD_MASK; 4124 data |= (0x36 << RLC_CGCG_CGLS_CTRL_3D__CGCG_GFX_IDLE_THRESHOLD__SHIFT) | 4125 RLC_CGCG_CGLS_CTRL_3D__CGCG_EN_MASK; 4126 } 4127 4128 if (adev->cg_flags & AMD_CG_SUPPORT_GFX_3D_CGLS) { 4129 data &= ~RLC_CGCG_CGLS_CTRL_3D__CGLS_REP_COMPANSAT_DELAY_MASK; 4130 data |= (0xf << RLC_CGCG_CGLS_CTRL_3D__CGLS_REP_COMPANSAT_DELAY__SHIFT) | 4131 RLC_CGCG_CGLS_CTRL_3D__CGLS_EN_MASK; 4132 } 4133 4134 if (def != data) 4135 WREG32_SOC15(GC, 0, regRLC_CGCG_CGLS_CTRL_3D, data); 4136 4137 /* set IDLE_POLL_COUNT(0x00900100) */ 4138 def = data = RREG32_SOC15(GC, 0, regCP_RB_WPTR_POLL_CNTL); 4139 4140 data &= ~(CP_RB_WPTR_POLL_CNTL__POLL_FREQUENCY_MASK | CP_RB_WPTR_POLL_CNTL__IDLE_POLL_COUNT_MASK); 4141 data |= (0x0100 << CP_RB_WPTR_POLL_CNTL__POLL_FREQUENCY__SHIFT) | 4142 (0x0090 << CP_RB_WPTR_POLL_CNTL__IDLE_POLL_COUNT__SHIFT); 4143 4144 if (def != data) 4145 WREG32_SOC15(GC, 0, regCP_RB_WPTR_POLL_CNTL, data); 4146 4147 data = RREG32_SOC15(GC, 0, regCP_INT_CNTL); 4148 data = REG_SET_FIELD(data, CP_INT_CNTL, CNTX_BUSY_INT_ENABLE, 1); 4149 data = REG_SET_FIELD(data, CP_INT_CNTL, CNTX_EMPTY_INT_ENABLE, 1); 4150 data = REG_SET_FIELD(data, CP_INT_CNTL, CMP_BUSY_INT_ENABLE, 1); 4151 data = REG_SET_FIELD(data, CP_INT_CNTL, GFX_IDLE_INT_ENABLE, 1); 4152 WREG32_SOC15(GC, 0, regCP_INT_CNTL, data); 4153 4154 data = RREG32_SOC15(GC, 0, regSDMA0_RLC_CGCG_CTRL); 4155 data = REG_SET_FIELD(data, SDMA0_RLC_CGCG_CTRL, CGCG_INT_ENABLE, 1); 4156 WREG32_SOC15(GC, 0, regSDMA0_RLC_CGCG_CTRL, data); 4157 4158 /* Some ASICs only have one SDMA instance, not need to configure SDMA1 */ 4159 if (adev->sdma.num_instances > 1) { 4160 data = RREG32_SOC15(GC, 0, regSDMA1_RLC_CGCG_CTRL); 4161 data = REG_SET_FIELD(data, SDMA1_RLC_CGCG_CTRL, CGCG_INT_ENABLE, 1); 4162 WREG32_SOC15(GC, 0, regSDMA1_RLC_CGCG_CTRL, data); 4163 } 4164 } else { 4165 /* Program RLC_CGCG_CGLS_CTRL */ 4166 def = data = RREG32_SOC15(GC, 0, regRLC_CGCG_CGLS_CTRL); 4167 4168 if (adev->cg_flags & AMD_CG_SUPPORT_GFX_CGCG) 4169 data &= ~RLC_CGCG_CGLS_CTRL__CGCG_EN_MASK; 4170 4171 if (adev->cg_flags & AMD_CG_SUPPORT_GFX_CGLS) 4172 data &= ~RLC_CGCG_CGLS_CTRL__CGLS_EN_MASK; 4173 4174 if (def != data) 4175 WREG32_SOC15(GC, 0, regRLC_CGCG_CGLS_CTRL, data); 4176 4177 /* Program RLC_CGCG_CGLS_CTRL_3D */ 4178 def = data = RREG32_SOC15(GC, 0, regRLC_CGCG_CGLS_CTRL_3D); 4179 4180 if (adev->cg_flags & AMD_CG_SUPPORT_GFX_3D_CGCG) 4181 data &= ~RLC_CGCG_CGLS_CTRL_3D__CGCG_EN_MASK; 4182 if (adev->cg_flags & AMD_CG_SUPPORT_GFX_3D_CGLS) 4183 data &= ~RLC_CGCG_CGLS_CTRL_3D__CGLS_EN_MASK; 4184 4185 if (def != data) 4186 WREG32_SOC15(GC, 0, regRLC_CGCG_CGLS_CTRL_3D, data); 4187 } 4188 } 4189 4190 static void gfx_v12_0_update_medium_grain_clock_gating(struct amdgpu_device *adev, 4191 bool enable) 4192 { 4193 uint32_t data, def; 4194 if (!(adev->cg_flags & (AMD_CG_SUPPORT_GFX_MGCG | AMD_CG_SUPPORT_GFX_MGLS))) 4195 return; 4196 4197 /* It is disabled by HW by default */ 4198 if (enable) { 4199 if (adev->cg_flags & AMD_CG_SUPPORT_GFX_MGCG) { 4200 /* 1 - RLC_CGTT_MGCG_OVERRIDE */ 4201 def = data = RREG32_SOC15(GC, 0, regRLC_CGTT_MGCG_OVERRIDE); 4202 4203 data &= ~(RLC_CGTT_MGCG_OVERRIDE__GRBM_CGTT_SCLK_OVERRIDE_MASK | 4204 RLC_CGTT_MGCG_OVERRIDE__RLC_CGTT_SCLK_OVERRIDE_MASK | 4205 RLC_CGTT_MGCG_OVERRIDE__GFXIP_MGCG_OVERRIDE_MASK); 4206 4207 if (def != data) 4208 WREG32_SOC15(GC, 0, regRLC_CGTT_MGCG_OVERRIDE, data); 4209 } 4210 } else { 4211 if (adev->cg_flags & AMD_CG_SUPPORT_GFX_MGCG) { 4212 def = data = RREG32_SOC15(GC, 0, regRLC_CGTT_MGCG_OVERRIDE); 4213 4214 data |= (RLC_CGTT_MGCG_OVERRIDE__RLC_CGTT_SCLK_OVERRIDE_MASK | 4215 RLC_CGTT_MGCG_OVERRIDE__GRBM_CGTT_SCLK_OVERRIDE_MASK | 4216 RLC_CGTT_MGCG_OVERRIDE__GFXIP_MGCG_OVERRIDE_MASK); 4217 4218 if (def != data) 4219 WREG32_SOC15(GC, 0, regRLC_CGTT_MGCG_OVERRIDE, data); 4220 } 4221 } 4222 } 4223 4224 static void gfx_v12_0_update_repeater_fgcg(struct amdgpu_device *adev, 4225 bool enable) 4226 { 4227 uint32_t def, data; 4228 4229 if (!(adev->cg_flags & AMD_CG_SUPPORT_REPEATER_FGCG)) 4230 return; 4231 4232 def = data = RREG32_SOC15(GC, 0, regRLC_CGTT_MGCG_OVERRIDE); 4233 4234 if (enable) 4235 data &= ~(RLC_CGTT_MGCG_OVERRIDE__GFXIP_REPEATER_FGCG_OVERRIDE_MASK | 4236 RLC_CGTT_MGCG_OVERRIDE__RLC_REPEATER_FGCG_OVERRIDE_MASK); 4237 else 4238 data |= RLC_CGTT_MGCG_OVERRIDE__GFXIP_REPEATER_FGCG_OVERRIDE_MASK | 4239 RLC_CGTT_MGCG_OVERRIDE__RLC_REPEATER_FGCG_OVERRIDE_MASK; 4240 4241 if (def != data) 4242 WREG32_SOC15(GC, 0, regRLC_CGTT_MGCG_OVERRIDE, data); 4243 } 4244 4245 static void gfx_v12_0_update_sram_fgcg(struct amdgpu_device *adev, 4246 bool enable) 4247 { 4248 uint32_t def, data; 4249 4250 if (!(adev->cg_flags & AMD_CG_SUPPORT_GFX_FGCG)) 4251 return; 4252 4253 def = data = RREG32_SOC15(GC, 0, regRLC_CGTT_MGCG_OVERRIDE); 4254 4255 if (enable) 4256 data &= ~RLC_CGTT_MGCG_OVERRIDE__GFXIP_FGCG_OVERRIDE_MASK; 4257 else 4258 data |= RLC_CGTT_MGCG_OVERRIDE__GFXIP_FGCG_OVERRIDE_MASK; 4259 4260 if (def != data) 4261 WREG32_SOC15(GC, 0, regRLC_CGTT_MGCG_OVERRIDE, data); 4262 } 4263 4264 static int gfx_v12_0_update_gfx_clock_gating(struct amdgpu_device *adev, 4265 bool enable) 4266 { 4267 amdgpu_gfx_rlc_enter_safe_mode(adev, 0); 4268 4269 gfx_v12_0_update_coarse_grain_clock_gating(adev, enable); 4270 4271 gfx_v12_0_update_medium_grain_clock_gating(adev, enable); 4272 4273 gfx_v12_0_update_repeater_fgcg(adev, enable); 4274 4275 gfx_v12_0_update_sram_fgcg(adev, enable); 4276 4277 gfx_v12_0_update_perf_clk(adev, enable); 4278 4279 if (adev->cg_flags & 4280 (AMD_CG_SUPPORT_GFX_MGCG | 4281 AMD_CG_SUPPORT_GFX_CGLS | 4282 AMD_CG_SUPPORT_GFX_CGCG | 4283 AMD_CG_SUPPORT_GFX_3D_CGCG | 4284 AMD_CG_SUPPORT_GFX_3D_CGLS)) 4285 gfx_v12_0_enable_gui_idle_interrupt(adev, enable); 4286 4287 amdgpu_gfx_rlc_exit_safe_mode(adev, 0); 4288 4289 return 0; 4290 } 4291 4292 static int gfx_v12_0_set_clockgating_state(struct amdgpu_ip_block *ip_block, 4293 enum amd_clockgating_state state) 4294 { 4295 struct amdgpu_device *adev = ip_block->adev; 4296 4297 if (amdgpu_sriov_vf(adev)) 4298 return 0; 4299 4300 switch (amdgpu_ip_version(adev, GC_HWIP, 0)) { 4301 case IP_VERSION(12, 0, 0): 4302 case IP_VERSION(12, 0, 1): 4303 gfx_v12_0_update_gfx_clock_gating(adev, 4304 state == AMD_CG_STATE_GATE); 4305 break; 4306 default: 4307 break; 4308 } 4309 4310 return 0; 4311 } 4312 4313 static void gfx_v12_0_get_clockgating_state(struct amdgpu_ip_block *ip_block, u64 *flags) 4314 { 4315 struct amdgpu_device *adev = ip_block->adev; 4316 int data; 4317 4318 /* AMD_CG_SUPPORT_GFX_MGCG */ 4319 data = RREG32_SOC15(GC, 0, regRLC_CGTT_MGCG_OVERRIDE); 4320 if (!(data & RLC_CGTT_MGCG_OVERRIDE__GFXIP_MGCG_OVERRIDE_MASK)) 4321 *flags |= AMD_CG_SUPPORT_GFX_MGCG; 4322 4323 /* AMD_CG_SUPPORT_REPEATER_FGCG */ 4324 if (!(data & RLC_CGTT_MGCG_OVERRIDE__GFXIP_REPEATER_FGCG_OVERRIDE_MASK)) 4325 *flags |= AMD_CG_SUPPORT_REPEATER_FGCG; 4326 4327 /* AMD_CG_SUPPORT_GFX_FGCG */ 4328 if (!(data & RLC_CGTT_MGCG_OVERRIDE__GFXIP_FGCG_OVERRIDE_MASK)) 4329 *flags |= AMD_CG_SUPPORT_GFX_FGCG; 4330 4331 /* AMD_CG_SUPPORT_GFX_PERF_CLK */ 4332 if (!(data & RLC_CGTT_MGCG_OVERRIDE__PERFMON_CLOCK_STATE_MASK)) 4333 *flags |= AMD_CG_SUPPORT_GFX_PERF_CLK; 4334 4335 /* AMD_CG_SUPPORT_GFX_CGCG */ 4336 data = RREG32_SOC15(GC, 0, regRLC_CGCG_CGLS_CTRL); 4337 if (data & RLC_CGCG_CGLS_CTRL__CGCG_EN_MASK) 4338 *flags |= AMD_CG_SUPPORT_GFX_CGCG; 4339 4340 /* AMD_CG_SUPPORT_GFX_CGLS */ 4341 if (data & RLC_CGCG_CGLS_CTRL__CGLS_EN_MASK) 4342 *flags |= AMD_CG_SUPPORT_GFX_CGLS; 4343 4344 /* AMD_CG_SUPPORT_GFX_3D_CGCG */ 4345 data = RREG32_SOC15(GC, 0, regRLC_CGCG_CGLS_CTRL_3D); 4346 if (data & RLC_CGCG_CGLS_CTRL_3D__CGCG_EN_MASK) 4347 *flags |= AMD_CG_SUPPORT_GFX_3D_CGCG; 4348 4349 /* AMD_CG_SUPPORT_GFX_3D_CGLS */ 4350 if (data & RLC_CGCG_CGLS_CTRL_3D__CGLS_EN_MASK) 4351 *flags |= AMD_CG_SUPPORT_GFX_3D_CGLS; 4352 } 4353 4354 static u64 gfx_v12_0_ring_get_rptr_gfx(struct amdgpu_ring *ring) 4355 { 4356 /* gfx12 is 32bit rptr*/ 4357 return *(uint32_t *)ring->rptr_cpu_addr; 4358 } 4359 4360 static u64 gfx_v12_0_ring_get_wptr_gfx(struct amdgpu_ring *ring) 4361 { 4362 struct amdgpu_device *adev = ring->adev; 4363 u64 wptr; 4364 4365 /* XXX check if swapping is necessary on BE */ 4366 if (ring->use_doorbell) { 4367 wptr = atomic64_read((atomic64_t *)ring->wptr_cpu_addr); 4368 } else { 4369 wptr = RREG32_SOC15(GC, 0, regCP_RB0_WPTR); 4370 wptr += (u64)RREG32_SOC15(GC, 0, regCP_RB0_WPTR_HI) << 32; 4371 } 4372 4373 return wptr; 4374 } 4375 4376 static void gfx_v12_0_ring_set_wptr_gfx(struct amdgpu_ring *ring) 4377 { 4378 struct amdgpu_device *adev = ring->adev; 4379 4380 if (ring->use_doorbell) { 4381 /* XXX check if swapping is necessary on BE */ 4382 atomic64_set((atomic64_t *)ring->wptr_cpu_addr, 4383 ring->wptr); 4384 WDOORBELL64(ring->doorbell_index, ring->wptr); 4385 } else { 4386 WREG32_SOC15(GC, 0, regCP_RB0_WPTR, 4387 lower_32_bits(ring->wptr)); 4388 WREG32_SOC15(GC, 0, regCP_RB0_WPTR_HI, 4389 upper_32_bits(ring->wptr)); 4390 } 4391 } 4392 4393 static u64 gfx_v12_0_ring_get_rptr_compute(struct amdgpu_ring *ring) 4394 { 4395 /* gfx12 hardware is 32bit rptr */ 4396 return *(uint32_t *)ring->rptr_cpu_addr; 4397 } 4398 4399 static u64 gfx_v12_0_ring_get_wptr_compute(struct amdgpu_ring *ring) 4400 { 4401 u64 wptr; 4402 4403 /* XXX check if swapping is necessary on BE */ 4404 if (ring->use_doorbell) 4405 wptr = atomic64_read((atomic64_t *)ring->wptr_cpu_addr); 4406 else 4407 BUG(); 4408 return wptr; 4409 } 4410 4411 static void gfx_v12_0_ring_set_wptr_compute(struct amdgpu_ring *ring) 4412 { 4413 struct amdgpu_device *adev = ring->adev; 4414 4415 /* XXX check if swapping is necessary on BE */ 4416 if (ring->use_doorbell) { 4417 atomic64_set((atomic64_t *)ring->wptr_cpu_addr, 4418 ring->wptr); 4419 WDOORBELL64(ring->doorbell_index, ring->wptr); 4420 } else { 4421 BUG(); /* only DOORBELL method supported on gfx12 now */ 4422 } 4423 } 4424 4425 static void gfx_v12_0_ring_emit_hdp_flush(struct amdgpu_ring *ring) 4426 { 4427 struct amdgpu_device *adev = ring->adev; 4428 u32 ref_and_mask, reg_mem_engine; 4429 4430 if (!adev->gfx.funcs->get_hdp_flush_mask) { 4431 dev_err(adev->dev, "%s: gfx hdp flush is not supported.\n", __func__); 4432 return; 4433 } 4434 4435 adev->gfx.funcs->get_hdp_flush_mask(ring, &ref_and_mask, ®_mem_engine); 4436 gfx_v12_0_wait_reg_mem(ring, reg_mem_engine, 0, 1, 4437 adev->nbio.funcs->get_hdp_flush_req_offset(adev), 4438 adev->nbio.funcs->get_hdp_flush_done_offset(adev), 4439 ref_and_mask, ref_and_mask, 0x20); 4440 } 4441 4442 static void gfx_v12_0_ring_emit_ib_gfx(struct amdgpu_ring *ring, 4443 struct amdgpu_job *job, 4444 struct amdgpu_ib *ib, 4445 uint32_t flags) 4446 { 4447 unsigned vmid = AMDGPU_JOB_GET_VMID(job); 4448 u32 header, control = 0; 4449 4450 header = PACKET3(PACKET3_INDIRECT_BUFFER, 2); 4451 4452 control |= ib->length_dw | (vmid << 24); 4453 4454 amdgpu_ring_write(ring, header); 4455 BUG_ON(ib->gpu_addr & 0x3); /* Dword align */ 4456 amdgpu_ring_write(ring, 4457 #ifdef __BIG_ENDIAN 4458 (2 << 0) | 4459 #endif 4460 lower_32_bits(ib->gpu_addr)); 4461 amdgpu_ring_write(ring, upper_32_bits(ib->gpu_addr)); 4462 amdgpu_ring_write(ring, control); 4463 } 4464 4465 static void gfx_v12_0_ring_emit_ib_compute(struct amdgpu_ring *ring, 4466 struct amdgpu_job *job, 4467 struct amdgpu_ib *ib, 4468 uint32_t flags) 4469 { 4470 unsigned vmid = AMDGPU_JOB_GET_VMID(job); 4471 u32 control = INDIRECT_BUFFER_VALID | ib->length_dw | (vmid << 24); 4472 4473 amdgpu_ring_write(ring, PACKET3(PACKET3_INDIRECT_BUFFER, 2)); 4474 BUG_ON(ib->gpu_addr & 0x3); /* Dword align */ 4475 amdgpu_ring_write(ring, 4476 #ifdef __BIG_ENDIAN 4477 (2 << 0) | 4478 #endif 4479 lower_32_bits(ib->gpu_addr)); 4480 amdgpu_ring_write(ring, upper_32_bits(ib->gpu_addr)); 4481 amdgpu_ring_write(ring, control); 4482 } 4483 4484 static void gfx_v12_0_ring_emit_fence(struct amdgpu_ring *ring, u64 addr, 4485 u64 seq, unsigned flags) 4486 { 4487 bool write64bit = flags & AMDGPU_FENCE_FLAG_64BIT; 4488 bool int_sel = flags & AMDGPU_FENCE_FLAG_INT; 4489 4490 /* RELEASE_MEM - flush caches, send int */ 4491 amdgpu_ring_write(ring, PACKET3(PACKET3_RELEASE_MEM, 6)); 4492 amdgpu_ring_write(ring, (PACKET3_RELEASE_MEM_GCR_SEQ | 4493 PACKET3_RELEASE_MEM_GCR_GL2_WB | 4494 PACKET3_RELEASE_MEM_CACHE_POLICY(3) | 4495 PACKET3_RELEASE_MEM_EVENT_TYPE(CACHE_FLUSH_AND_INV_TS_EVENT) | 4496 PACKET3_RELEASE_MEM_EVENT_INDEX(5))); 4497 amdgpu_ring_write(ring, (PACKET3_RELEASE_MEM_DATA_SEL(write64bit ? 2 : 1) | 4498 PACKET3_RELEASE_MEM_INT_SEL(int_sel ? 2 : 0))); 4499 4500 /* 4501 * the address should be Qword aligned if 64bit write, Dword 4502 * aligned if only send 32bit data low (discard data high) 4503 */ 4504 if (write64bit) 4505 BUG_ON(addr & 0x7); 4506 else 4507 BUG_ON(addr & 0x3); 4508 amdgpu_ring_write(ring, lower_32_bits(addr)); 4509 amdgpu_ring_write(ring, upper_32_bits(addr)); 4510 amdgpu_ring_write(ring, lower_32_bits(seq)); 4511 amdgpu_ring_write(ring, upper_32_bits(seq)); 4512 amdgpu_ring_write(ring, 0); 4513 } 4514 4515 static void gfx_v12_0_ring_emit_pipeline_sync(struct amdgpu_ring *ring) 4516 { 4517 int usepfp = (ring->funcs->type == AMDGPU_RING_TYPE_GFX); 4518 uint32_t seq = ring->fence_drv.sync_seq; 4519 uint64_t addr = ring->fence_drv.gpu_addr; 4520 4521 gfx_v12_0_wait_reg_mem(ring, usepfp, 1, 0, lower_32_bits(addr), 4522 upper_32_bits(addr), seq, 0xffffffff, 4); 4523 } 4524 4525 static void gfx_v12_0_ring_invalidate_tlbs(struct amdgpu_ring *ring, 4526 uint16_t pasid, uint32_t flush_type, 4527 bool all_hub, uint8_t dst_sel) 4528 { 4529 amdgpu_ring_write(ring, PACKET3(PACKET3_INVALIDATE_TLBS, 0)); 4530 amdgpu_ring_write(ring, 4531 PACKET3_INVALIDATE_TLBS_DST_SEL(dst_sel) | 4532 PACKET3_INVALIDATE_TLBS_ALL_HUB(all_hub) | 4533 PACKET3_INVALIDATE_TLBS_PASID(pasid) | 4534 PACKET3_INVALIDATE_TLBS_FLUSH_TYPE(flush_type)); 4535 } 4536 4537 static void gfx_v12_0_ring_emit_vm_flush(struct amdgpu_ring *ring, 4538 unsigned vmid, uint64_t pd_addr) 4539 { 4540 amdgpu_gmc_emit_flush_gpu_tlb(ring, vmid, pd_addr); 4541 4542 /* compute doesn't have PFP */ 4543 if (ring->funcs->type == AMDGPU_RING_TYPE_GFX) { 4544 /* sync PFP to ME, otherwise we might get invalid PFP reads */ 4545 amdgpu_ring_write(ring, PACKET3(PACKET3_PFP_SYNC_ME, 0)); 4546 amdgpu_ring_write(ring, 0x0); 4547 } 4548 } 4549 4550 static void gfx_v12_0_ring_emit_fence_kiq(struct amdgpu_ring *ring, u64 addr, 4551 u64 seq, unsigned int flags) 4552 { 4553 struct amdgpu_device *adev = ring->adev; 4554 4555 /* we only allocate 32bit for each seq wb address */ 4556 BUG_ON(flags & AMDGPU_FENCE_FLAG_64BIT); 4557 4558 /* write fence seq to the "addr" */ 4559 amdgpu_ring_write(ring, PACKET3(PACKET3_WRITE_DATA, 3)); 4560 amdgpu_ring_write(ring, (WRITE_DATA_ENGINE_SEL(0) | 4561 WRITE_DATA_DST_SEL(5) | WR_CONFIRM)); 4562 amdgpu_ring_write(ring, lower_32_bits(addr)); 4563 amdgpu_ring_write(ring, upper_32_bits(addr)); 4564 amdgpu_ring_write(ring, lower_32_bits(seq)); 4565 4566 if (flags & AMDGPU_FENCE_FLAG_INT) { 4567 /* set register to trigger INT */ 4568 amdgpu_ring_write(ring, PACKET3(PACKET3_WRITE_DATA, 3)); 4569 amdgpu_ring_write(ring, (WRITE_DATA_ENGINE_SEL(0) | 4570 WRITE_DATA_DST_SEL(0) | WR_CONFIRM)); 4571 amdgpu_ring_write(ring, SOC15_REG_OFFSET(GC, 0, regCPC_INT_STATUS)); 4572 amdgpu_ring_write(ring, 0); 4573 amdgpu_ring_write(ring, 0x20000000); /* src_id is 178 */ 4574 } 4575 } 4576 4577 static void gfx_v12_0_ring_emit_cntxcntl(struct amdgpu_ring *ring, 4578 uint32_t flags) 4579 { 4580 uint32_t dw2 = 0; 4581 4582 dw2 |= 0x80000000; /* set load_enable otherwise this package is just NOPs */ 4583 if (flags & AMDGPU_HAVE_CTX_SWITCH) { 4584 /* set load_global_config & load_global_uconfig */ 4585 dw2 |= 0x8001; 4586 /* set load_cs_sh_regs */ 4587 dw2 |= 0x01000000; 4588 /* set load_per_context_state & load_gfx_sh_regs for GFX */ 4589 dw2 |= 0x10002; 4590 } 4591 4592 amdgpu_ring_write(ring, PACKET3(PACKET3_CONTEXT_CONTROL, 1)); 4593 amdgpu_ring_write(ring, dw2); 4594 amdgpu_ring_write(ring, 0); 4595 } 4596 4597 static unsigned gfx_v12_0_ring_emit_init_cond_exec(struct amdgpu_ring *ring, 4598 uint64_t addr) 4599 { 4600 unsigned ret; 4601 4602 amdgpu_ring_write(ring, PACKET3(PACKET3_COND_EXEC, 3)); 4603 amdgpu_ring_write(ring, lower_32_bits(addr)); 4604 amdgpu_ring_write(ring, upper_32_bits(addr)); 4605 /* discard following DWs if *cond_exec_gpu_addr==0 */ 4606 amdgpu_ring_write(ring, 0); 4607 ret = ring->wptr & ring->buf_mask; 4608 /* patch dummy value later */ 4609 amdgpu_ring_write(ring, 0); 4610 4611 return ret; 4612 } 4613 4614 static int gfx_v12_0_ring_preempt_ib(struct amdgpu_ring *ring) 4615 { 4616 int i, r = 0; 4617 struct amdgpu_device *adev = ring->adev; 4618 struct amdgpu_kiq *kiq = &adev->gfx.kiq[0]; 4619 struct amdgpu_ring *kiq_ring = &kiq->ring; 4620 unsigned long flags; 4621 4622 if (adev->enable_mes) 4623 return -EINVAL; 4624 4625 if (!kiq->pmf || !kiq->pmf->kiq_unmap_queues) 4626 return -EINVAL; 4627 4628 spin_lock_irqsave(&kiq->ring_lock, flags); 4629 4630 if (amdgpu_ring_alloc(kiq_ring, kiq->pmf->unmap_queues_size)) { 4631 spin_unlock_irqrestore(&kiq->ring_lock, flags); 4632 return -ENOMEM; 4633 } 4634 4635 /* assert preemption condition */ 4636 amdgpu_ring_set_preempt_cond_exec(ring, false); 4637 4638 /* assert IB preemption, emit the trailing fence */ 4639 kiq->pmf->kiq_unmap_queues(kiq_ring, ring, PREEMPT_QUEUES_NO_UNMAP, 4640 ring->trail_fence_gpu_addr, 4641 ++ring->trail_seq); 4642 amdgpu_ring_commit(kiq_ring); 4643 4644 spin_unlock_irqrestore(&kiq->ring_lock, flags); 4645 4646 /* poll the trailing fence */ 4647 for (i = 0; i < adev->usec_timeout; i++) { 4648 if (ring->trail_seq == 4649 le32_to_cpu(*(ring->trail_fence_cpu_addr))) 4650 break; 4651 udelay(1); 4652 } 4653 4654 if (i >= adev->usec_timeout) { 4655 r = -EINVAL; 4656 DRM_ERROR("ring %d failed to preempt ib\n", ring->idx); 4657 } 4658 4659 /* deassert preemption condition */ 4660 amdgpu_ring_set_preempt_cond_exec(ring, true); 4661 return r; 4662 } 4663 4664 static void gfx_v12_0_ring_emit_rreg(struct amdgpu_ring *ring, uint32_t reg, 4665 uint32_t reg_val_offs) 4666 { 4667 struct amdgpu_device *adev = ring->adev; 4668 4669 amdgpu_ring_write(ring, PACKET3(PACKET3_COPY_DATA, 4)); 4670 amdgpu_ring_write(ring, 0 | /* src: register*/ 4671 (5 << 8) | /* dst: memory */ 4672 (1 << 20)); /* write confirm */ 4673 amdgpu_ring_write(ring, reg); 4674 amdgpu_ring_write(ring, 0); 4675 amdgpu_ring_write(ring, lower_32_bits(adev->wb.gpu_addr + 4676 reg_val_offs * 4)); 4677 amdgpu_ring_write(ring, upper_32_bits(adev->wb.gpu_addr + 4678 reg_val_offs * 4)); 4679 } 4680 4681 static void gfx_v12_0_ring_emit_wreg(struct amdgpu_ring *ring, 4682 uint32_t reg, 4683 uint32_t val) 4684 { 4685 uint32_t cmd = 0; 4686 4687 switch (ring->funcs->type) { 4688 case AMDGPU_RING_TYPE_GFX: 4689 cmd = WRITE_DATA_ENGINE_SEL(1) | WR_CONFIRM; 4690 break; 4691 case AMDGPU_RING_TYPE_KIQ: 4692 cmd = (1 << 16); /* no inc addr */ 4693 break; 4694 default: 4695 cmd = WR_CONFIRM; 4696 break; 4697 } 4698 amdgpu_ring_write(ring, PACKET3(PACKET3_WRITE_DATA, 3)); 4699 amdgpu_ring_write(ring, cmd); 4700 amdgpu_ring_write(ring, reg); 4701 amdgpu_ring_write(ring, 0); 4702 amdgpu_ring_write(ring, val); 4703 } 4704 4705 static void gfx_v12_0_ring_emit_reg_wait(struct amdgpu_ring *ring, uint32_t reg, 4706 uint32_t val, uint32_t mask) 4707 { 4708 gfx_v12_0_wait_reg_mem(ring, 0, 0, 0, reg, 0, val, mask, 0x20); 4709 } 4710 4711 static void gfx_v12_0_ring_emit_reg_write_reg_wait(struct amdgpu_ring *ring, 4712 uint32_t reg0, uint32_t reg1, 4713 uint32_t ref, uint32_t mask) 4714 { 4715 int usepfp = (ring->funcs->type == AMDGPU_RING_TYPE_GFX); 4716 4717 gfx_v12_0_wait_reg_mem(ring, usepfp, 0, 1, reg0, reg1, 4718 ref, mask, 0x20); 4719 } 4720 4721 static void 4722 gfx_v12_0_set_gfx_eop_interrupt_state(struct amdgpu_device *adev, 4723 uint32_t me, uint32_t pipe, 4724 enum amdgpu_interrupt_state state) 4725 { 4726 uint32_t cp_int_cntl, cp_int_cntl_reg; 4727 4728 if (!me) { 4729 switch (pipe) { 4730 case 0: 4731 cp_int_cntl_reg = SOC15_REG_OFFSET(GC, 0, regCP_INT_CNTL_RING0); 4732 break; 4733 default: 4734 DRM_DEBUG("invalid pipe %d\n", pipe); 4735 return; 4736 } 4737 } else { 4738 DRM_DEBUG("invalid me %d\n", me); 4739 return; 4740 } 4741 4742 switch (state) { 4743 case AMDGPU_IRQ_STATE_DISABLE: 4744 cp_int_cntl = RREG32_SOC15_IP(GC, cp_int_cntl_reg); 4745 cp_int_cntl = REG_SET_FIELD(cp_int_cntl, CP_INT_CNTL_RING0, 4746 TIME_STAMP_INT_ENABLE, 0); 4747 cp_int_cntl = REG_SET_FIELD(cp_int_cntl, CP_INT_CNTL_RING0, 4748 GENERIC0_INT_ENABLE, 0); 4749 WREG32_SOC15_IP(GC, cp_int_cntl_reg, cp_int_cntl); 4750 break; 4751 case AMDGPU_IRQ_STATE_ENABLE: 4752 cp_int_cntl = RREG32_SOC15_IP(GC, cp_int_cntl_reg); 4753 cp_int_cntl = REG_SET_FIELD(cp_int_cntl, CP_INT_CNTL_RING0, 4754 TIME_STAMP_INT_ENABLE, 1); 4755 cp_int_cntl = REG_SET_FIELD(cp_int_cntl, CP_INT_CNTL_RING0, 4756 GENERIC0_INT_ENABLE, 1); 4757 WREG32_SOC15_IP(GC, cp_int_cntl_reg, cp_int_cntl); 4758 break; 4759 default: 4760 break; 4761 } 4762 } 4763 4764 static void gfx_v12_0_set_compute_eop_interrupt_state(struct amdgpu_device *adev, 4765 int me, int pipe, 4766 enum amdgpu_interrupt_state state) 4767 { 4768 u32 mec_int_cntl, mec_int_cntl_reg; 4769 4770 /* 4771 * amdgpu controls only the first MEC. That's why this function only 4772 * handles the setting of interrupts for this specific MEC. All other 4773 * pipes' interrupts are set by amdkfd. 4774 */ 4775 4776 if (me == 1) { 4777 switch (pipe) { 4778 case 0: 4779 mec_int_cntl_reg = SOC15_REG_OFFSET(GC, 0, regCP_ME1_PIPE0_INT_CNTL); 4780 break; 4781 case 1: 4782 mec_int_cntl_reg = SOC15_REG_OFFSET(GC, 0, regCP_ME1_PIPE1_INT_CNTL); 4783 break; 4784 default: 4785 DRM_DEBUG("invalid pipe %d\n", pipe); 4786 return; 4787 } 4788 } else { 4789 DRM_DEBUG("invalid me %d\n", me); 4790 return; 4791 } 4792 4793 switch (state) { 4794 case AMDGPU_IRQ_STATE_DISABLE: 4795 mec_int_cntl = RREG32_SOC15_IP(GC, mec_int_cntl_reg); 4796 mec_int_cntl = REG_SET_FIELD(mec_int_cntl, CP_ME1_PIPE0_INT_CNTL, 4797 TIME_STAMP_INT_ENABLE, 0); 4798 mec_int_cntl = REG_SET_FIELD(mec_int_cntl, CP_ME1_PIPE0_INT_CNTL, 4799 GENERIC0_INT_ENABLE, 0); 4800 WREG32_SOC15_IP(GC, mec_int_cntl_reg, mec_int_cntl); 4801 break; 4802 case AMDGPU_IRQ_STATE_ENABLE: 4803 mec_int_cntl = RREG32_SOC15_IP(GC, mec_int_cntl_reg); 4804 mec_int_cntl = REG_SET_FIELD(mec_int_cntl, CP_ME1_PIPE0_INT_CNTL, 4805 TIME_STAMP_INT_ENABLE, 1); 4806 mec_int_cntl = REG_SET_FIELD(mec_int_cntl, CP_ME1_PIPE0_INT_CNTL, 4807 GENERIC0_INT_ENABLE, 1); 4808 WREG32_SOC15_IP(GC, mec_int_cntl_reg, mec_int_cntl); 4809 break; 4810 default: 4811 break; 4812 } 4813 } 4814 4815 static int gfx_v12_0_set_eop_interrupt_state(struct amdgpu_device *adev, 4816 struct amdgpu_irq_src *src, 4817 unsigned type, 4818 enum amdgpu_interrupt_state state) 4819 { 4820 switch (type) { 4821 case AMDGPU_CP_IRQ_GFX_ME0_PIPE0_EOP: 4822 gfx_v12_0_set_gfx_eop_interrupt_state(adev, 0, 0, state); 4823 break; 4824 case AMDGPU_CP_IRQ_GFX_ME0_PIPE1_EOP: 4825 gfx_v12_0_set_gfx_eop_interrupt_state(adev, 0, 1, state); 4826 break; 4827 case AMDGPU_CP_IRQ_COMPUTE_MEC1_PIPE0_EOP: 4828 gfx_v12_0_set_compute_eop_interrupt_state(adev, 1, 0, state); 4829 break; 4830 case AMDGPU_CP_IRQ_COMPUTE_MEC1_PIPE1_EOP: 4831 gfx_v12_0_set_compute_eop_interrupt_state(adev, 1, 1, state); 4832 break; 4833 case AMDGPU_CP_IRQ_COMPUTE_MEC1_PIPE2_EOP: 4834 gfx_v12_0_set_compute_eop_interrupt_state(adev, 1, 2, state); 4835 break; 4836 case AMDGPU_CP_IRQ_COMPUTE_MEC1_PIPE3_EOP: 4837 gfx_v12_0_set_compute_eop_interrupt_state(adev, 1, 3, state); 4838 break; 4839 default: 4840 break; 4841 } 4842 return 0; 4843 } 4844 4845 static int gfx_v12_0_eop_irq(struct amdgpu_device *adev, 4846 struct amdgpu_irq_src *source, 4847 struct amdgpu_iv_entry *entry) 4848 { 4849 u32 doorbell_offset = entry->src_data[0]; 4850 u8 me_id, pipe_id, queue_id; 4851 struct amdgpu_ring *ring; 4852 int i; 4853 4854 DRM_DEBUG("IH: CP EOP\n"); 4855 4856 if (adev->enable_mes && doorbell_offset) { 4857 struct amdgpu_userq_fence_driver *fence_drv = NULL; 4858 struct xarray *xa = &adev->userq_xa; 4859 unsigned long flags; 4860 4861 xa_lock_irqsave(xa, flags); 4862 fence_drv = xa_load(xa, doorbell_offset); 4863 if (fence_drv) 4864 amdgpu_userq_fence_driver_process(fence_drv); 4865 xa_unlock_irqrestore(xa, flags); 4866 } else { 4867 me_id = (entry->ring_id & 0x0c) >> 2; 4868 pipe_id = (entry->ring_id & 0x03) >> 0; 4869 queue_id = (entry->ring_id & 0x70) >> 4; 4870 4871 switch (me_id) { 4872 case 0: 4873 if (pipe_id == 0) 4874 amdgpu_fence_process(&adev->gfx.gfx_ring[0]); 4875 else 4876 amdgpu_fence_process(&adev->gfx.gfx_ring[1]); 4877 break; 4878 case 1: 4879 case 2: 4880 for (i = 0; i < adev->gfx.num_compute_rings; i++) { 4881 ring = &adev->gfx.compute_ring[i]; 4882 /* Per-queue interrupt is supported for MEC starting from VI. 4883 * The interrupt can only be enabled/disabled per pipe instead 4884 * of per queue. 4885 */ 4886 if ((ring->me == me_id) && 4887 (ring->pipe == pipe_id) && 4888 (ring->queue == queue_id)) 4889 amdgpu_fence_process(ring); 4890 } 4891 break; 4892 } 4893 } 4894 4895 return 0; 4896 } 4897 4898 static int gfx_v12_0_set_priv_reg_fault_state(struct amdgpu_device *adev, 4899 struct amdgpu_irq_src *source, 4900 unsigned int type, 4901 enum amdgpu_interrupt_state state) 4902 { 4903 u32 cp_int_cntl_reg, cp_int_cntl; 4904 int i, j; 4905 4906 switch (state) { 4907 case AMDGPU_IRQ_STATE_DISABLE: 4908 case AMDGPU_IRQ_STATE_ENABLE: 4909 for (i = 0; i < adev->gfx.me.num_me; i++) { 4910 for (j = 0; j < adev->gfx.me.num_pipe_per_me; j++) { 4911 cp_int_cntl_reg = gfx_v12_0_get_cpg_int_cntl(adev, i, j); 4912 4913 if (cp_int_cntl_reg) { 4914 cp_int_cntl = RREG32_SOC15_IP(GC, cp_int_cntl_reg); 4915 cp_int_cntl = REG_SET_FIELD(cp_int_cntl, CP_INT_CNTL_RING0, 4916 PRIV_REG_INT_ENABLE, 4917 state == AMDGPU_IRQ_STATE_ENABLE ? 1 : 0); 4918 WREG32_SOC15_IP(GC, cp_int_cntl_reg, cp_int_cntl); 4919 } 4920 } 4921 } 4922 for (i = 0; i < adev->gfx.mec.num_mec; i++) { 4923 for (j = 0; j < adev->gfx.mec.num_pipe_per_mec; j++) { 4924 /* MECs start at 1 */ 4925 cp_int_cntl_reg = gfx_v12_0_get_cpc_int_cntl(adev, i + 1, j); 4926 4927 if (cp_int_cntl_reg) { 4928 cp_int_cntl = RREG32_SOC15_IP(GC, cp_int_cntl_reg); 4929 cp_int_cntl = REG_SET_FIELD(cp_int_cntl, CP_ME1_PIPE0_INT_CNTL, 4930 PRIV_REG_INT_ENABLE, 4931 state == AMDGPU_IRQ_STATE_ENABLE ? 1 : 0); 4932 WREG32_SOC15_IP(GC, cp_int_cntl_reg, cp_int_cntl); 4933 } 4934 } 4935 } 4936 break; 4937 default: 4938 break; 4939 } 4940 4941 return 0; 4942 } 4943 4944 static int gfx_v12_0_set_bad_op_fault_state(struct amdgpu_device *adev, 4945 struct amdgpu_irq_src *source, 4946 unsigned type, 4947 enum amdgpu_interrupt_state state) 4948 { 4949 u32 cp_int_cntl_reg, cp_int_cntl; 4950 int i, j; 4951 4952 switch (state) { 4953 case AMDGPU_IRQ_STATE_DISABLE: 4954 case AMDGPU_IRQ_STATE_ENABLE: 4955 for (i = 0; i < adev->gfx.me.num_me; i++) { 4956 for (j = 0; j < adev->gfx.me.num_pipe_per_me; j++) { 4957 cp_int_cntl_reg = gfx_v12_0_get_cpg_int_cntl(adev, i, j); 4958 4959 if (cp_int_cntl_reg) { 4960 cp_int_cntl = RREG32_SOC15_IP(GC, cp_int_cntl_reg); 4961 cp_int_cntl = REG_SET_FIELD(cp_int_cntl, CP_INT_CNTL_RING0, 4962 OPCODE_ERROR_INT_ENABLE, 4963 state == AMDGPU_IRQ_STATE_ENABLE ? 1 : 0); 4964 WREG32_SOC15_IP(GC, cp_int_cntl_reg, cp_int_cntl); 4965 } 4966 } 4967 } 4968 for (i = 0; i < adev->gfx.mec.num_mec; i++) { 4969 for (j = 0; j < adev->gfx.mec.num_pipe_per_mec; j++) { 4970 /* MECs start at 1 */ 4971 cp_int_cntl_reg = gfx_v12_0_get_cpc_int_cntl(adev, i + 1, j); 4972 4973 if (cp_int_cntl_reg) { 4974 cp_int_cntl = RREG32_SOC15_IP(GC, cp_int_cntl_reg); 4975 cp_int_cntl = REG_SET_FIELD(cp_int_cntl, CP_ME1_PIPE0_INT_CNTL, 4976 OPCODE_ERROR_INT_ENABLE, 4977 state == AMDGPU_IRQ_STATE_ENABLE ? 1 : 0); 4978 WREG32_SOC15_IP(GC, cp_int_cntl_reg, cp_int_cntl); 4979 } 4980 } 4981 } 4982 break; 4983 default: 4984 break; 4985 } 4986 return 0; 4987 } 4988 4989 static int gfx_v12_0_set_priv_inst_fault_state(struct amdgpu_device *adev, 4990 struct amdgpu_irq_src *source, 4991 unsigned int type, 4992 enum amdgpu_interrupt_state state) 4993 { 4994 u32 cp_int_cntl_reg, cp_int_cntl; 4995 int i, j; 4996 4997 switch (state) { 4998 case AMDGPU_IRQ_STATE_DISABLE: 4999 case AMDGPU_IRQ_STATE_ENABLE: 5000 for (i = 0; i < adev->gfx.me.num_me; i++) { 5001 for (j = 0; j < adev->gfx.me.num_pipe_per_me; j++) { 5002 cp_int_cntl_reg = gfx_v12_0_get_cpg_int_cntl(adev, i, j); 5003 5004 if (cp_int_cntl_reg) { 5005 cp_int_cntl = RREG32_SOC15_IP(GC, cp_int_cntl_reg); 5006 cp_int_cntl = REG_SET_FIELD(cp_int_cntl, CP_INT_CNTL_RING0, 5007 PRIV_INSTR_INT_ENABLE, 5008 state == AMDGPU_IRQ_STATE_ENABLE ? 1 : 0); 5009 WREG32_SOC15_IP(GC, cp_int_cntl_reg, cp_int_cntl); 5010 } 5011 } 5012 } 5013 break; 5014 default: 5015 break; 5016 } 5017 5018 return 0; 5019 } 5020 5021 static void gfx_v12_0_handle_priv_fault(struct amdgpu_device *adev, 5022 struct amdgpu_iv_entry *entry) 5023 { 5024 u8 me_id, pipe_id, queue_id; 5025 struct amdgpu_ring *ring; 5026 int i; 5027 5028 me_id = (entry->ring_id & 0x0c) >> 2; 5029 pipe_id = (entry->ring_id & 0x03) >> 0; 5030 queue_id = (entry->ring_id & 0x70) >> 4; 5031 5032 if (!adev->gfx.disable_kq) { 5033 switch (me_id) { 5034 case 0: 5035 for (i = 0; i < adev->gfx.num_gfx_rings; i++) { 5036 ring = &adev->gfx.gfx_ring[i]; 5037 if (ring->me == me_id && ring->pipe == pipe_id && 5038 ring->queue == queue_id) 5039 drm_sched_fault(&ring->sched); 5040 } 5041 break; 5042 case 1: 5043 case 2: 5044 for (i = 0; i < adev->gfx.num_compute_rings; i++) { 5045 ring = &adev->gfx.compute_ring[i]; 5046 if (ring->me == me_id && ring->pipe == pipe_id && 5047 ring->queue == queue_id) 5048 drm_sched_fault(&ring->sched); 5049 } 5050 break; 5051 default: 5052 BUG(); 5053 break; 5054 } 5055 } 5056 } 5057 5058 static int gfx_v12_0_priv_reg_irq(struct amdgpu_device *adev, 5059 struct amdgpu_irq_src *source, 5060 struct amdgpu_iv_entry *entry) 5061 { 5062 DRM_ERROR("Illegal register access in command stream\n"); 5063 gfx_v12_0_handle_priv_fault(adev, entry); 5064 return 0; 5065 } 5066 5067 static int gfx_v12_0_bad_op_irq(struct amdgpu_device *adev, 5068 struct amdgpu_irq_src *source, 5069 struct amdgpu_iv_entry *entry) 5070 { 5071 DRM_ERROR("Illegal opcode in command stream\n"); 5072 gfx_v12_0_handle_priv_fault(adev, entry); 5073 return 0; 5074 } 5075 5076 static int gfx_v12_0_priv_inst_irq(struct amdgpu_device *adev, 5077 struct amdgpu_irq_src *source, 5078 struct amdgpu_iv_entry *entry) 5079 { 5080 DRM_ERROR("Illegal instruction in command stream\n"); 5081 gfx_v12_0_handle_priv_fault(adev, entry); 5082 return 0; 5083 } 5084 5085 static void gfx_v12_0_emit_mem_sync(struct amdgpu_ring *ring) 5086 { 5087 const unsigned int gcr_cntl = 5088 PACKET3_ACQUIRE_MEM_GCR_CNTL_GL2_INV(1) | 5089 PACKET3_ACQUIRE_MEM_GCR_CNTL_GL2_WB(1) | 5090 PACKET3_ACQUIRE_MEM_GCR_CNTL_GLM_INV(1) | 5091 PACKET3_ACQUIRE_MEM_GCR_CNTL_GLM_WB(1) | 5092 PACKET3_ACQUIRE_MEM_GCR_CNTL_GL1_INV(1) | 5093 PACKET3_ACQUIRE_MEM_GCR_CNTL_GLV_INV(1) | 5094 PACKET3_ACQUIRE_MEM_GCR_CNTL_GLK_INV(1) | 5095 PACKET3_ACQUIRE_MEM_GCR_CNTL_GLI_INV(1); 5096 5097 /* ACQUIRE_MEM - make one or more surfaces valid for use by the subsequent operations */ 5098 amdgpu_ring_write(ring, PACKET3(PACKET3_ACQUIRE_MEM, 6)); 5099 amdgpu_ring_write(ring, 0); /* CP_COHER_CNTL */ 5100 amdgpu_ring_write(ring, 0xffffffff); /* CP_COHER_SIZE */ 5101 amdgpu_ring_write(ring, 0xffffff); /* CP_COHER_SIZE_HI */ 5102 amdgpu_ring_write(ring, 0); /* CP_COHER_BASE */ 5103 amdgpu_ring_write(ring, 0); /* CP_COHER_BASE_HI */ 5104 amdgpu_ring_write(ring, 0x0000000A); /* POLL_INTERVAL */ 5105 amdgpu_ring_write(ring, gcr_cntl); /* GCR_CNTL */ 5106 } 5107 5108 static void gfx_v12_ring_insert_nop(struct amdgpu_ring *ring, uint32_t num_nop) 5109 { 5110 /* Header itself is a NOP packet */ 5111 if (num_nop == 1) { 5112 amdgpu_ring_write(ring, ring->funcs->nop); 5113 return; 5114 } 5115 5116 /* Max HW optimization till 0x3ffe, followed by remaining one NOP at a time*/ 5117 amdgpu_ring_write(ring, PACKET3(PACKET3_NOP, min(num_nop - 2, 0x3ffe))); 5118 5119 /* Header is at index 0, followed by num_nops - 1 NOP packet's */ 5120 amdgpu_ring_insert_nop(ring, num_nop - 1); 5121 } 5122 5123 static void gfx_v12_0_ring_emit_cleaner_shader(struct amdgpu_ring *ring) 5124 { 5125 /* Emit the cleaner shader */ 5126 amdgpu_ring_write(ring, PACKET3(PACKET3_RUN_CLEANER_SHADER, 0)); 5127 amdgpu_ring_write(ring, 0); /* RESERVED field, programmed to zero */ 5128 } 5129 5130 static void gfx_v12_ip_print(struct amdgpu_ip_block *ip_block, struct drm_printer *p) 5131 { 5132 struct amdgpu_device *adev = ip_block->adev; 5133 uint32_t i, j, k, reg, index = 0; 5134 uint32_t reg_count = ARRAY_SIZE(gc_reg_list_12_0); 5135 5136 if (!adev->gfx.ip_dump_core) 5137 return; 5138 5139 for (i = 0; i < reg_count; i++) 5140 drm_printf(p, "%-50s \t 0x%08x\n", 5141 gc_reg_list_12_0[i].reg_name, 5142 adev->gfx.ip_dump_core[i]); 5143 5144 /* print compute queue registers for all instances */ 5145 if (!adev->gfx.ip_dump_compute_queues) 5146 return; 5147 5148 reg_count = ARRAY_SIZE(gc_cp_reg_list_12); 5149 drm_printf(p, "\nnum_mec: %d num_pipe: %d num_queue: %d\n", 5150 adev->gfx.mec.num_mec, 5151 adev->gfx.mec.num_pipe_per_mec, 5152 adev->gfx.mec.num_queue_per_pipe); 5153 5154 for (i = 0; i < adev->gfx.mec.num_mec; i++) { 5155 for (j = 0; j < adev->gfx.mec.num_pipe_per_mec; j++) { 5156 for (k = 0; k < adev->gfx.mec.num_queue_per_pipe; k++) { 5157 drm_printf(p, "\nmec %d, pipe %d, queue %d\n", i, j, k); 5158 for (reg = 0; reg < reg_count; reg++) { 5159 drm_printf(p, "%-50s \t 0x%08x\n", 5160 gc_cp_reg_list_12[reg].reg_name, 5161 adev->gfx.ip_dump_compute_queues[index + reg]); 5162 } 5163 index += reg_count; 5164 } 5165 } 5166 } 5167 5168 /* print gfx queue registers for all instances */ 5169 if (!adev->gfx.ip_dump_gfx_queues) 5170 return; 5171 5172 index = 0; 5173 reg_count = ARRAY_SIZE(gc_gfx_queue_reg_list_12); 5174 drm_printf(p, "\nnum_me: %d num_pipe: %d num_queue: %d\n", 5175 adev->gfx.me.num_me, 5176 adev->gfx.me.num_pipe_per_me, 5177 adev->gfx.me.num_queue_per_pipe); 5178 5179 for (i = 0; i < adev->gfx.me.num_me; i++) { 5180 for (j = 0; j < adev->gfx.me.num_pipe_per_me; j++) { 5181 for (k = 0; k < adev->gfx.me.num_queue_per_pipe; k++) { 5182 drm_printf(p, "\nme %d, pipe %d, queue %d\n", i, j, k); 5183 for (reg = 0; reg < reg_count; reg++) { 5184 drm_printf(p, "%-50s \t 0x%08x\n", 5185 gc_gfx_queue_reg_list_12[reg].reg_name, 5186 adev->gfx.ip_dump_gfx_queues[index + reg]); 5187 } 5188 index += reg_count; 5189 } 5190 } 5191 } 5192 } 5193 5194 static void gfx_v12_ip_dump(struct amdgpu_ip_block *ip_block) 5195 { 5196 struct amdgpu_device *adev = ip_block->adev; 5197 uint32_t i, j, k, reg, index = 0; 5198 uint32_t reg_count = ARRAY_SIZE(gc_reg_list_12_0); 5199 5200 if (!adev->gfx.ip_dump_core) 5201 return; 5202 5203 amdgpu_gfx_off_ctrl(adev, false); 5204 for (i = 0; i < reg_count; i++) 5205 adev->gfx.ip_dump_core[i] = RREG32(SOC15_REG_ENTRY_OFFSET(gc_reg_list_12_0[i])); 5206 amdgpu_gfx_off_ctrl(adev, true); 5207 5208 /* dump compute queue registers for all instances */ 5209 if (!adev->gfx.ip_dump_compute_queues) 5210 return; 5211 5212 reg_count = ARRAY_SIZE(gc_cp_reg_list_12); 5213 amdgpu_gfx_off_ctrl(adev, false); 5214 mutex_lock(&adev->srbm_mutex); 5215 for (i = 0; i < adev->gfx.mec.num_mec; i++) { 5216 for (j = 0; j < adev->gfx.mec.num_pipe_per_mec; j++) { 5217 for (k = 0; k < adev->gfx.mec.num_queue_per_pipe; k++) { 5218 /* ME0 is for GFX so start from 1 for CP */ 5219 soc24_grbm_select(adev, adev->gfx.me.num_me + i, j, k, 0); 5220 for (reg = 0; reg < reg_count; reg++) { 5221 adev->gfx.ip_dump_compute_queues[index + reg] = 5222 RREG32(SOC15_REG_ENTRY_OFFSET( 5223 gc_cp_reg_list_12[reg])); 5224 } 5225 index += reg_count; 5226 } 5227 } 5228 } 5229 soc24_grbm_select(adev, 0, 0, 0, 0); 5230 mutex_unlock(&adev->srbm_mutex); 5231 amdgpu_gfx_off_ctrl(adev, true); 5232 5233 /* dump gfx queue registers for all instances */ 5234 if (!adev->gfx.ip_dump_gfx_queues) 5235 return; 5236 5237 index = 0; 5238 reg_count = ARRAY_SIZE(gc_gfx_queue_reg_list_12); 5239 amdgpu_gfx_off_ctrl(adev, false); 5240 mutex_lock(&adev->srbm_mutex); 5241 for (i = 0; i < adev->gfx.me.num_me; i++) { 5242 for (j = 0; j < adev->gfx.me.num_pipe_per_me; j++) { 5243 for (k = 0; k < adev->gfx.me.num_queue_per_pipe; k++) { 5244 soc24_grbm_select(adev, i, j, k, 0); 5245 5246 for (reg = 0; reg < reg_count; reg++) { 5247 adev->gfx.ip_dump_gfx_queues[index + reg] = 5248 RREG32(SOC15_REG_ENTRY_OFFSET( 5249 gc_gfx_queue_reg_list_12[reg])); 5250 } 5251 index += reg_count; 5252 } 5253 } 5254 } 5255 soc24_grbm_select(adev, 0, 0, 0, 0); 5256 mutex_unlock(&adev->srbm_mutex); 5257 amdgpu_gfx_off_ctrl(adev, true); 5258 } 5259 5260 static bool gfx_v12_pipe_reset_support(struct amdgpu_device *adev) 5261 { 5262 /* Disable the pipe reset until the CPFW fully support it.*/ 5263 dev_warn_once(adev->dev, "The CPFW hasn't support pipe reset yet.\n"); 5264 return false; 5265 } 5266 5267 static int gfx_v12_reset_gfx_pipe(struct amdgpu_ring *ring) 5268 { 5269 struct amdgpu_device *adev = ring->adev; 5270 uint32_t reset_pipe = 0, clean_pipe = 0; 5271 int r; 5272 5273 if (!gfx_v12_pipe_reset_support(adev)) 5274 return -EOPNOTSUPP; 5275 5276 gfx_v12_0_set_safe_mode(adev, 0); 5277 mutex_lock(&adev->srbm_mutex); 5278 soc24_grbm_select(adev, ring->me, ring->pipe, ring->queue, 0); 5279 5280 switch (ring->pipe) { 5281 case 0: 5282 reset_pipe = REG_SET_FIELD(reset_pipe, CP_ME_CNTL, 5283 PFP_PIPE0_RESET, 1); 5284 reset_pipe = REG_SET_FIELD(reset_pipe, CP_ME_CNTL, 5285 ME_PIPE0_RESET, 1); 5286 clean_pipe = REG_SET_FIELD(clean_pipe, CP_ME_CNTL, 5287 PFP_PIPE0_RESET, 0); 5288 clean_pipe = REG_SET_FIELD(clean_pipe, CP_ME_CNTL, 5289 ME_PIPE0_RESET, 0); 5290 break; 5291 case 1: 5292 reset_pipe = REG_SET_FIELD(reset_pipe, CP_ME_CNTL, 5293 PFP_PIPE1_RESET, 1); 5294 reset_pipe = REG_SET_FIELD(reset_pipe, CP_ME_CNTL, 5295 ME_PIPE1_RESET, 1); 5296 clean_pipe = REG_SET_FIELD(clean_pipe, CP_ME_CNTL, 5297 PFP_PIPE1_RESET, 0); 5298 clean_pipe = REG_SET_FIELD(clean_pipe, CP_ME_CNTL, 5299 ME_PIPE1_RESET, 0); 5300 break; 5301 default: 5302 break; 5303 } 5304 5305 WREG32_SOC15(GC, 0, regCP_ME_CNTL, reset_pipe); 5306 WREG32_SOC15(GC, 0, regCP_ME_CNTL, clean_pipe); 5307 5308 r = (RREG32(SOC15_REG_OFFSET(GC, 0, regCP_GFX_RS64_INSTR_PNTR1)) << 2) - 5309 RS64_FW_UC_START_ADDR_LO; 5310 soc24_grbm_select(adev, 0, 0, 0, 0); 5311 mutex_unlock(&adev->srbm_mutex); 5312 gfx_v12_0_unset_safe_mode(adev, 0); 5313 5314 dev_info(adev->dev, "The ring %s pipe reset: %s\n", ring->name, 5315 r == 0 ? "successfully" : "failed"); 5316 /* Sometimes the ME start pc counter can't cache correctly, so the 5317 * PC check only as a reference and pipe reset result rely on the 5318 * later ring test. 5319 */ 5320 return 0; 5321 } 5322 5323 static int gfx_v12_0_reset_kgq(struct amdgpu_ring *ring, 5324 unsigned int vmid, 5325 struct amdgpu_fence *timedout_fence) 5326 { 5327 struct amdgpu_device *adev = ring->adev; 5328 bool use_mmio = false; 5329 int r; 5330 5331 amdgpu_ring_reset_helper_begin(ring, timedout_fence); 5332 5333 r = amdgpu_mes_reset_legacy_queue(ring->adev, ring, vmid, use_mmio, 0); 5334 if (r) { 5335 dev_warn(adev->dev, "reset via MES failed and try pipe reset %d\n", r); 5336 r = gfx_v12_reset_gfx_pipe(ring); 5337 if (r) 5338 return r; 5339 } 5340 5341 if (use_mmio) { 5342 r = gfx_v12_0_kgq_init_queue(ring, true); 5343 if (r) { 5344 dev_err(adev->dev, "failed to init kgq\n"); 5345 return r; 5346 } 5347 5348 r = amdgpu_mes_map_legacy_queue(adev, ring, 0); 5349 if (r) { 5350 dev_err(adev->dev, "failed to remap kgq\n"); 5351 return r; 5352 } 5353 } 5354 5355 return amdgpu_ring_reset_helper_end(ring, timedout_fence); 5356 } 5357 5358 static int gfx_v12_0_reset_compute_pipe(struct amdgpu_ring *ring) 5359 { 5360 struct amdgpu_device *adev = ring->adev; 5361 uint32_t reset_pipe = 0, clean_pipe = 0; 5362 int r = 0; 5363 5364 if (!gfx_v12_pipe_reset_support(adev)) 5365 return -EOPNOTSUPP; 5366 5367 gfx_v12_0_set_safe_mode(adev, 0); 5368 mutex_lock(&adev->srbm_mutex); 5369 soc24_grbm_select(adev, ring->me, ring->pipe, ring->queue, 0); 5370 5371 reset_pipe = RREG32_SOC15(GC, 0, regCP_MEC_RS64_CNTL); 5372 clean_pipe = reset_pipe; 5373 5374 if (adev->gfx.rs64_enable) { 5375 switch (ring->pipe) { 5376 case 0: 5377 reset_pipe = REG_SET_FIELD(reset_pipe, CP_MEC_RS64_CNTL, 5378 MEC_PIPE0_RESET, 1); 5379 clean_pipe = REG_SET_FIELD(clean_pipe, CP_MEC_RS64_CNTL, 5380 MEC_PIPE0_RESET, 0); 5381 break; 5382 case 1: 5383 reset_pipe = REG_SET_FIELD(reset_pipe, CP_MEC_RS64_CNTL, 5384 MEC_PIPE1_RESET, 1); 5385 clean_pipe = REG_SET_FIELD(clean_pipe, CP_MEC_RS64_CNTL, 5386 MEC_PIPE1_RESET, 0); 5387 break; 5388 case 2: 5389 reset_pipe = REG_SET_FIELD(reset_pipe, CP_MEC_RS64_CNTL, 5390 MEC_PIPE2_RESET, 1); 5391 clean_pipe = REG_SET_FIELD(clean_pipe, CP_MEC_RS64_CNTL, 5392 MEC_PIPE2_RESET, 0); 5393 break; 5394 case 3: 5395 reset_pipe = REG_SET_FIELD(reset_pipe, CP_MEC_RS64_CNTL, 5396 MEC_PIPE3_RESET, 1); 5397 clean_pipe = REG_SET_FIELD(clean_pipe, CP_MEC_RS64_CNTL, 5398 MEC_PIPE3_RESET, 0); 5399 break; 5400 default: 5401 break; 5402 } 5403 WREG32_SOC15(GC, 0, regCP_MEC_RS64_CNTL, reset_pipe); 5404 WREG32_SOC15(GC, 0, regCP_MEC_RS64_CNTL, clean_pipe); 5405 r = (RREG32_SOC15(GC, 0, regCP_MEC_RS64_INSTR_PNTR) << 2) - 5406 RS64_FW_UC_START_ADDR_LO; 5407 } else { 5408 switch (ring->pipe) { 5409 case 0: 5410 reset_pipe = REG_SET_FIELD(reset_pipe, CP_MEC_CNTL, 5411 MEC_ME1_PIPE0_RESET, 1); 5412 clean_pipe = REG_SET_FIELD(clean_pipe, CP_MEC_CNTL, 5413 MEC_ME1_PIPE0_RESET, 0); 5414 break; 5415 case 1: 5416 reset_pipe = REG_SET_FIELD(reset_pipe, CP_MEC_CNTL, 5417 MEC_ME1_PIPE1_RESET, 1); 5418 clean_pipe = REG_SET_FIELD(clean_pipe, CP_MEC_CNTL, 5419 MEC_ME1_PIPE1_RESET, 0); 5420 break; 5421 default: 5422 break; 5423 } 5424 WREG32_SOC15(GC, 0, regCP_MEC_CNTL, reset_pipe); 5425 WREG32_SOC15(GC, 0, regCP_MEC_CNTL, clean_pipe); 5426 /* Doesn't find the F32 MEC instruction pointer register, and suppose 5427 * the driver won't run into the F32 mode. 5428 */ 5429 } 5430 5431 soc24_grbm_select(adev, 0, 0, 0, 0); 5432 mutex_unlock(&adev->srbm_mutex); 5433 gfx_v12_0_unset_safe_mode(adev, 0); 5434 5435 dev_info(adev->dev, "The ring %s pipe resets: %s\n", ring->name, 5436 r == 0 ? "successfully" : "failed"); 5437 /* Need the ring test to verify the pipe reset result.*/ 5438 return 0; 5439 } 5440 5441 static int gfx_v12_0_reset_kcq(struct amdgpu_ring *ring, 5442 unsigned int vmid, 5443 struct amdgpu_fence *timedout_fence) 5444 { 5445 struct amdgpu_device *adev = ring->adev; 5446 int r; 5447 5448 amdgpu_ring_reset_helper_begin(ring, timedout_fence); 5449 5450 r = amdgpu_mes_reset_legacy_queue(ring->adev, ring, vmid, true, 0); 5451 if (r) { 5452 dev_warn(adev->dev, "fail(%d) to reset kcq and try pipe reset\n", r); 5453 r = gfx_v12_0_reset_compute_pipe(ring); 5454 if (r) 5455 return r; 5456 } 5457 5458 r = gfx_v12_0_kcq_init_queue(ring, true); 5459 if (r) { 5460 dev_err(adev->dev, "failed to init kcq\n"); 5461 return r; 5462 } 5463 r = amdgpu_mes_map_legacy_queue(adev, ring, 0); 5464 if (r) { 5465 dev_err(adev->dev, "failed to remap kcq\n"); 5466 return r; 5467 } 5468 5469 return amdgpu_ring_reset_helper_end(ring, timedout_fence); 5470 } 5471 5472 static void gfx_v12_0_ring_begin_use(struct amdgpu_ring *ring) 5473 { 5474 amdgpu_gfx_profile_ring_begin_use(ring); 5475 5476 amdgpu_gfx_enforce_isolation_ring_begin_use(ring); 5477 } 5478 5479 static void gfx_v12_0_ring_end_use(struct amdgpu_ring *ring) 5480 { 5481 amdgpu_gfx_profile_ring_end_use(ring); 5482 5483 amdgpu_gfx_enforce_isolation_ring_end_use(ring); 5484 } 5485 5486 static const struct amd_ip_funcs gfx_v12_0_ip_funcs = { 5487 .name = "gfx_v12_0", 5488 .early_init = gfx_v12_0_early_init, 5489 .late_init = gfx_v12_0_late_init, 5490 .sw_init = gfx_v12_0_sw_init, 5491 .sw_fini = gfx_v12_0_sw_fini, 5492 .hw_init = gfx_v12_0_hw_init, 5493 .hw_fini = gfx_v12_0_hw_fini, 5494 .suspend = gfx_v12_0_suspend, 5495 .resume = gfx_v12_0_resume, 5496 .is_idle = gfx_v12_0_is_idle, 5497 .wait_for_idle = gfx_v12_0_wait_for_idle, 5498 .set_clockgating_state = gfx_v12_0_set_clockgating_state, 5499 .set_powergating_state = gfx_v12_0_set_powergating_state, 5500 .get_clockgating_state = gfx_v12_0_get_clockgating_state, 5501 .dump_ip_state = gfx_v12_ip_dump, 5502 .print_ip_state = gfx_v12_ip_print, 5503 }; 5504 5505 static const struct amdgpu_ring_funcs gfx_v12_0_ring_funcs_gfx = { 5506 .type = AMDGPU_RING_TYPE_GFX, 5507 .align_mask = 0xff, 5508 .nop = PACKET3(PACKET3_NOP, 0x3FFF), 5509 .support_64bit_ptrs = true, 5510 .secure_submission_supported = true, 5511 .get_rptr = gfx_v12_0_ring_get_rptr_gfx, 5512 .get_wptr = gfx_v12_0_ring_get_wptr_gfx, 5513 .set_wptr = gfx_v12_0_ring_set_wptr_gfx, 5514 .emit_frame_size = /* totally 242 maximum if 16 IBs */ 5515 5 + /* COND_EXEC */ 5516 7 + /* PIPELINE_SYNC */ 5517 SOC15_FLUSH_GPU_TLB_NUM_WREG * 5 + 5518 SOC15_FLUSH_GPU_TLB_NUM_REG_WAIT * 7 + 5519 2 + /* VM_FLUSH */ 5520 8 + /* FENCE for VM_FLUSH */ 5521 5 + /* COND_EXEC */ 5522 7 + /* HDP_flush */ 5523 4 + /* VGT_flush */ 5524 31 + /* DE_META */ 5525 3 + /* CNTX_CTRL */ 5526 5 + /* HDP_INVL */ 5527 8 + 8 + /* FENCE x2 */ 5528 8 + /* gfx_v12_0_emit_mem_sync */ 5529 2, /* gfx_v12_0_ring_emit_cleaner_shader */ 5530 .emit_ib_size = 4, /* gfx_v12_0_ring_emit_ib_gfx */ 5531 .emit_ib = gfx_v12_0_ring_emit_ib_gfx, 5532 .emit_fence = gfx_v12_0_ring_emit_fence, 5533 .emit_pipeline_sync = gfx_v12_0_ring_emit_pipeline_sync, 5534 .emit_vm_flush = gfx_v12_0_ring_emit_vm_flush, 5535 .emit_hdp_flush = gfx_v12_0_ring_emit_hdp_flush, 5536 .test_ring = gfx_v12_0_ring_test_ring, 5537 .test_ib = gfx_v12_0_ring_test_ib, 5538 .insert_nop = gfx_v12_ring_insert_nop, 5539 .pad_ib = amdgpu_ring_generic_pad_ib, 5540 .emit_cntxcntl = gfx_v12_0_ring_emit_cntxcntl, 5541 .init_cond_exec = gfx_v12_0_ring_emit_init_cond_exec, 5542 .preempt_ib = gfx_v12_0_ring_preempt_ib, 5543 .emit_wreg = gfx_v12_0_ring_emit_wreg, 5544 .emit_reg_wait = gfx_v12_0_ring_emit_reg_wait, 5545 .emit_reg_write_reg_wait = gfx_v12_0_ring_emit_reg_write_reg_wait, 5546 .emit_mem_sync = gfx_v12_0_emit_mem_sync, 5547 .reset = gfx_v12_0_reset_kgq, 5548 .emit_cleaner_shader = gfx_v12_0_ring_emit_cleaner_shader, 5549 .begin_use = gfx_v12_0_ring_begin_use, 5550 .end_use = gfx_v12_0_ring_end_use, 5551 }; 5552 5553 static const struct amdgpu_ring_funcs gfx_v12_0_ring_funcs_compute = { 5554 .type = AMDGPU_RING_TYPE_COMPUTE, 5555 .align_mask = 0xff, 5556 .nop = PACKET3(PACKET3_NOP, 0x3FFF), 5557 .support_64bit_ptrs = true, 5558 .get_rptr = gfx_v12_0_ring_get_rptr_compute, 5559 .get_wptr = gfx_v12_0_ring_get_wptr_compute, 5560 .set_wptr = gfx_v12_0_ring_set_wptr_compute, 5561 .emit_frame_size = 5562 7 + /* gfx_v12_0_ring_emit_hdp_flush */ 5563 5 + /* hdp invalidate */ 5564 7 + /* gfx_v12_0_ring_emit_pipeline_sync */ 5565 SOC15_FLUSH_GPU_TLB_NUM_WREG * 5 + 5566 SOC15_FLUSH_GPU_TLB_NUM_REG_WAIT * 7 + 5567 2 + /* gfx_v12_0_ring_emit_vm_flush */ 5568 8 + 8 + 8 + /* gfx_v12_0_ring_emit_fence x3 for user fence, vm fence */ 5569 8 + /* gfx_v12_0_emit_mem_sync */ 5570 2, /* gfx_v12_0_ring_emit_cleaner_shader */ 5571 .emit_ib_size = 7, /* gfx_v12_0_ring_emit_ib_compute */ 5572 .emit_ib = gfx_v12_0_ring_emit_ib_compute, 5573 .emit_fence = gfx_v12_0_ring_emit_fence, 5574 .emit_pipeline_sync = gfx_v12_0_ring_emit_pipeline_sync, 5575 .emit_vm_flush = gfx_v12_0_ring_emit_vm_flush, 5576 .emit_hdp_flush = gfx_v12_0_ring_emit_hdp_flush, 5577 .test_ring = gfx_v12_0_ring_test_ring, 5578 .test_ib = gfx_v12_0_ring_test_ib, 5579 .insert_nop = gfx_v12_ring_insert_nop, 5580 .pad_ib = amdgpu_ring_generic_pad_ib, 5581 .emit_wreg = gfx_v12_0_ring_emit_wreg, 5582 .emit_reg_wait = gfx_v12_0_ring_emit_reg_wait, 5583 .emit_reg_write_reg_wait = gfx_v12_0_ring_emit_reg_write_reg_wait, 5584 .emit_mem_sync = gfx_v12_0_emit_mem_sync, 5585 .reset = gfx_v12_0_reset_kcq, 5586 .emit_cleaner_shader = gfx_v12_0_ring_emit_cleaner_shader, 5587 .begin_use = gfx_v12_0_ring_begin_use, 5588 .end_use = gfx_v12_0_ring_end_use, 5589 }; 5590 5591 static const struct amdgpu_ring_funcs gfx_v12_0_ring_funcs_kiq = { 5592 .type = AMDGPU_RING_TYPE_KIQ, 5593 .align_mask = 0xff, 5594 .nop = PACKET3(PACKET3_NOP, 0x3FFF), 5595 .support_64bit_ptrs = true, 5596 .get_rptr = gfx_v12_0_ring_get_rptr_compute, 5597 .get_wptr = gfx_v12_0_ring_get_wptr_compute, 5598 .set_wptr = gfx_v12_0_ring_set_wptr_compute, 5599 .emit_frame_size = 5600 7 + /* gfx_v12_0_ring_emit_hdp_flush */ 5601 5 + /*hdp invalidate */ 5602 7 + /* gfx_v12_0_ring_emit_pipeline_sync */ 5603 SOC15_FLUSH_GPU_TLB_NUM_WREG * 5 + 5604 SOC15_FLUSH_GPU_TLB_NUM_REG_WAIT * 7 + 5605 2 + /* gfx_v12_0_ring_emit_vm_flush */ 5606 8 + 8 + 8, /* gfx_v12_0_ring_emit_fence_kiq x3 for user fence, vm fence */ 5607 .emit_ib_size = 7, /* gfx_v12_0_ring_emit_ib_compute */ 5608 .emit_ib = gfx_v12_0_ring_emit_ib_compute, 5609 .emit_fence = gfx_v12_0_ring_emit_fence_kiq, 5610 .test_ring = gfx_v12_0_ring_test_ring, 5611 .test_ib = gfx_v12_0_ring_test_ib, 5612 .insert_nop = amdgpu_ring_insert_nop, 5613 .pad_ib = amdgpu_ring_generic_pad_ib, 5614 .emit_rreg = gfx_v12_0_ring_emit_rreg, 5615 .emit_wreg = gfx_v12_0_ring_emit_wreg, 5616 .emit_reg_wait = gfx_v12_0_ring_emit_reg_wait, 5617 .emit_reg_write_reg_wait = gfx_v12_0_ring_emit_reg_write_reg_wait, 5618 .emit_hdp_flush = gfx_v12_0_ring_emit_hdp_flush, 5619 }; 5620 5621 static void gfx_v12_0_set_ring_funcs(struct amdgpu_device *adev) 5622 { 5623 int i; 5624 5625 adev->gfx.kiq[0].ring.funcs = &gfx_v12_0_ring_funcs_kiq; 5626 5627 for (i = 0; i < adev->gfx.num_gfx_rings; i++) 5628 adev->gfx.gfx_ring[i].funcs = &gfx_v12_0_ring_funcs_gfx; 5629 5630 for (i = 0; i < adev->gfx.num_compute_rings; i++) 5631 adev->gfx.compute_ring[i].funcs = &gfx_v12_0_ring_funcs_compute; 5632 } 5633 5634 static const struct amdgpu_irq_src_funcs gfx_v12_0_eop_irq_funcs = { 5635 .set = gfx_v12_0_set_eop_interrupt_state, 5636 .process = gfx_v12_0_eop_irq, 5637 }; 5638 5639 static const struct amdgpu_irq_src_funcs gfx_v12_0_priv_reg_irq_funcs = { 5640 .set = gfx_v12_0_set_priv_reg_fault_state, 5641 .process = gfx_v12_0_priv_reg_irq, 5642 }; 5643 5644 static const struct amdgpu_irq_src_funcs gfx_v12_0_bad_op_irq_funcs = { 5645 .set = gfx_v12_0_set_bad_op_fault_state, 5646 .process = gfx_v12_0_bad_op_irq, 5647 }; 5648 5649 static const struct amdgpu_irq_src_funcs gfx_v12_0_priv_inst_irq_funcs = { 5650 .set = gfx_v12_0_set_priv_inst_fault_state, 5651 .process = gfx_v12_0_priv_inst_irq, 5652 }; 5653 5654 static void gfx_v12_0_set_irq_funcs(struct amdgpu_device *adev) 5655 { 5656 adev->gfx.eop_irq.num_types = AMDGPU_CP_IRQ_LAST; 5657 adev->gfx.eop_irq.funcs = &gfx_v12_0_eop_irq_funcs; 5658 5659 adev->gfx.priv_reg_irq.num_types = 1; 5660 adev->gfx.priv_reg_irq.funcs = &gfx_v12_0_priv_reg_irq_funcs; 5661 5662 adev->gfx.bad_op_irq.num_types = 1; 5663 adev->gfx.bad_op_irq.funcs = &gfx_v12_0_bad_op_irq_funcs; 5664 5665 adev->gfx.priv_inst_irq.num_types = 1; 5666 adev->gfx.priv_inst_irq.funcs = &gfx_v12_0_priv_inst_irq_funcs; 5667 } 5668 5669 static void gfx_v12_0_set_imu_funcs(struct amdgpu_device *adev) 5670 { 5671 if (adev->flags & AMD_IS_APU) 5672 adev->gfx.imu.mode = MISSION_MODE; 5673 else 5674 adev->gfx.imu.mode = DEBUG_MODE; 5675 5676 adev->gfx.imu.funcs = &gfx_v12_0_imu_funcs; 5677 } 5678 5679 static void gfx_v12_0_set_rlc_funcs(struct amdgpu_device *adev) 5680 { 5681 adev->gfx.rlc.funcs = &gfx_v12_0_rlc_funcs; 5682 } 5683 5684 static void gfx_v12_0_set_mqd_funcs(struct amdgpu_device *adev) 5685 { 5686 /* set gfx eng mqd */ 5687 adev->mqds[AMDGPU_HW_IP_GFX].mqd_size = 5688 sizeof(struct v12_gfx_mqd); 5689 adev->mqds[AMDGPU_HW_IP_GFX].init_mqd = 5690 gfx_v12_0_gfx_mqd_init; 5691 /* set compute eng mqd */ 5692 adev->mqds[AMDGPU_HW_IP_COMPUTE].mqd_size = 5693 sizeof(struct v12_compute_mqd); 5694 adev->mqds[AMDGPU_HW_IP_COMPUTE].init_mqd = 5695 gfx_v12_0_compute_mqd_init; 5696 } 5697 5698 static void gfx_v12_0_set_user_wgp_inactive_bitmap_per_sh(struct amdgpu_device *adev, 5699 u32 bitmap) 5700 { 5701 u32 data; 5702 5703 if (!bitmap) 5704 return; 5705 5706 data = bitmap << GC_USER_SHADER_ARRAY_CONFIG__INACTIVE_WGPS__SHIFT; 5707 data &= GC_USER_SHADER_ARRAY_CONFIG__INACTIVE_WGPS_MASK; 5708 5709 WREG32_SOC15(GC, 0, regGC_USER_SHADER_ARRAY_CONFIG, data); 5710 } 5711 5712 static u32 gfx_v12_0_get_wgp_active_bitmap_per_sh(struct amdgpu_device *adev) 5713 { 5714 u32 data, wgp_bitmask; 5715 data = RREG32_SOC15(GC, 0, regCC_GC_SHADER_ARRAY_CONFIG); 5716 data |= RREG32_SOC15(GC, 0, regGC_USER_SHADER_ARRAY_CONFIG); 5717 5718 data &= CC_GC_SHADER_ARRAY_CONFIG__INACTIVE_WGPS_MASK; 5719 data >>= CC_GC_SHADER_ARRAY_CONFIG__INACTIVE_WGPS__SHIFT; 5720 5721 wgp_bitmask = 5722 amdgpu_gfx_create_bitmask(adev->gfx.config.max_cu_per_sh >> 1); 5723 5724 return (~data) & wgp_bitmask; 5725 } 5726 5727 static u32 gfx_v12_0_get_cu_active_bitmap_per_sh(struct amdgpu_device *adev) 5728 { 5729 u32 wgp_idx, wgp_active_bitmap; 5730 u32 cu_bitmap_per_wgp, cu_active_bitmap; 5731 5732 wgp_active_bitmap = gfx_v12_0_get_wgp_active_bitmap_per_sh(adev); 5733 cu_active_bitmap = 0; 5734 5735 for (wgp_idx = 0; wgp_idx < 16; wgp_idx++) { 5736 /* if there is one WGP enabled, it means 2 CUs will be enabled */ 5737 cu_bitmap_per_wgp = 3 << (2 * wgp_idx); 5738 if (wgp_active_bitmap & (1 << wgp_idx)) 5739 cu_active_bitmap |= cu_bitmap_per_wgp; 5740 } 5741 5742 return cu_active_bitmap; 5743 } 5744 5745 static int gfx_v12_0_get_cu_info(struct amdgpu_device *adev, 5746 struct amdgpu_cu_info *cu_info) 5747 { 5748 int i, j, k, counter, active_cu_number = 0; 5749 u32 mask, bitmap; 5750 unsigned disable_masks[8 * 2]; 5751 5752 if (!adev || !cu_info) 5753 return -EINVAL; 5754 5755 amdgpu_gfx_parse_disable_cu(adev, disable_masks, 8, 2); 5756 5757 mutex_lock(&adev->grbm_idx_mutex); 5758 for (i = 0; i < adev->gfx.config.max_shader_engines; i++) { 5759 for (j = 0; j < adev->gfx.config.max_sh_per_se; j++) { 5760 bitmap = i * adev->gfx.config.max_sh_per_se + j; 5761 if (!((gfx_v12_0_get_sa_active_bitmap(adev) >> bitmap) & 1)) 5762 continue; 5763 mask = 1; 5764 counter = 0; 5765 gfx_v12_0_select_se_sh(adev, i, j, 0xffffffff, 0); 5766 if (i < 8 && j < 2) 5767 gfx_v12_0_set_user_wgp_inactive_bitmap_per_sh( 5768 adev, disable_masks[i * 2 + j]); 5769 bitmap = gfx_v12_0_get_cu_active_bitmap_per_sh(adev); 5770 5771 /** 5772 * GFX12 could support more than 4 SEs, while the bitmap 5773 * in cu_info struct is 4x4 and ioctl interface struct 5774 * drm_amdgpu_info_device should keep stable. 5775 * So we use last two columns of bitmap to store cu mask for 5776 * SEs 4 to 7, the layout of the bitmap is as below: 5777 * SE0: {SH0,SH1} --> {bitmap[0][0], bitmap[0][1]} 5778 * SE1: {SH0,SH1} --> {bitmap[1][0], bitmap[1][1]} 5779 * SE2: {SH0,SH1} --> {bitmap[2][0], bitmap[2][1]} 5780 * SE3: {SH0,SH1} --> {bitmap[3][0], bitmap[3][1]} 5781 * SE4: {SH0,SH1} --> {bitmap[0][2], bitmap[0][3]} 5782 * SE5: {SH0,SH1} --> {bitmap[1][2], bitmap[1][3]} 5783 * SE6: {SH0,SH1} --> {bitmap[2][2], bitmap[2][3]} 5784 * SE7: {SH0,SH1} --> {bitmap[3][2], bitmap[3][3]} 5785 */ 5786 cu_info->bitmap[0][i % 4][j + (i / 4) * 2] = bitmap; 5787 5788 for (k = 0; k < adev->gfx.config.max_cu_per_sh; k++) { 5789 if (bitmap & mask) 5790 counter++; 5791 5792 mask <<= 1; 5793 } 5794 active_cu_number += counter; 5795 } 5796 } 5797 gfx_v12_0_select_se_sh(adev, 0xffffffff, 0xffffffff, 0xffffffff, 0); 5798 mutex_unlock(&adev->grbm_idx_mutex); 5799 5800 cu_info->number = active_cu_number; 5801 cu_info->simd_per_cu = NUM_SIMD_PER_CU; 5802 5803 return 0; 5804 } 5805 5806 const struct amdgpu_ip_block_version gfx_v12_0_ip_block = { 5807 .type = AMD_IP_BLOCK_TYPE_GFX, 5808 .major = 12, 5809 .minor = 0, 5810 .rev = 0, 5811 .funcs = &gfx_v12_0_ip_funcs, 5812 }; 5813